diff --git a/.gitattributes b/.gitattributes index 7341827dc6f7d5da96247b29aaf4dfbe7e331cdd..4e5e1e1595a290c950b93ba1c275568447b67d23 100644 --- a/.gitattributes +++ b/.gitattributes @@ -596,3 +596,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/Dissolving[[:space:]]Is[[:space:]]Amplifying_[[:space:]]Towards[[:space:]]Fine-Grained[[:space:]]Anomaly[[:space:]]Detection/0d1ddaa9-caa8-4dd2-9351-149d495d099b_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Distill[[:space:]]Gold[[:space:]]from[[:space:]]Massive[[:space:]]Ores_[[:space:]]Bi-level[[:space:]]Data[[:space:]]Pruning[[:space:]]towards[[:space:]]Efficient[[:space:]]Dataset[[:space:]]Distillation/f4366f6e-ee9e-4483-800e-3cc9ed152bdc_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Distilling[[:space:]]Diffusion[[:space:]]Models[[:space:]]into[[:space:]]Conditional[[:space:]]GANs/7a6aca0b-5aad-4819-a493-a6d4c2b7fd31_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Distilling[[:space:]]Knowledge[[:space:]]from[[:space:]]Large-Scale[[:space:]]Image[[:space:]]Models[[:space:]]for[[:space:]]Object[[:space:]]Detection/a5a39b01-2823-428c-b29a-0b036d9c3243_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Distractor-Free[[:space:]]Novel[[:space:]]View[[:space:]]Synthesis[[:space:]]via[[:space:]]Exploiting[[:space:]]Memorization[[:space:]]Effect[[:space:]]in[[:space:]]Optimization/b0f9a1b1-12b0-444d-bc6b-1b50e3076e57_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Distractors-Immune[[:space:]]Representation[[:space:]]Learning[[:space:]]with[[:space:]]Cross-modal[[:space:]]Contrastive[[:space:]]Regularization[[:space:]]for[[:space:]]Change[[:space:]]Captioning/01669dd0-b2b9-4e44-8fe5-812e59df2574_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Distributed[[:space:]]Active[[:space:]]Client[[:space:]]Selection[[:space:]]With[[:space:]]Noisy[[:space:]]Clients[[:space:]]Using[[:space:]]Model[[:space:]]Association[[:space:]]Scores/24102e0c-3413-4f79-9cc9-125f01961814_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Distributed[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]with[[:space:]]Efficient[[:space:]]Joint[[:space:]]Source[[:space:]]and[[:space:]]Task[[:space:]]Decoding/e1a0f3a6-55e7-483d-b43c-267a5ed4362a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Distribution[[:space:]]Alignment[[:space:]]for[[:space:]]Fully[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]with[[:space:]]Dynamic[[:space:]]Online[[:space:]]Data[[:space:]]Streams/c50a6abf-242e-4a7f-b483-6b9b1c9f9e9e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Distribution-Aware[[:space:]]Robust[[:space:]]Learning[[:space:]]from[[:space:]]Long-Tailed[[:space:]]Data[[:space:]]with[[:space:]]Noisy[[:space:]]Labels/b2569ce8-7eb5-419e-b936-dac7e5e92228_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Distributionally[[:space:]]Robust[[:space:]]Loss[[:space:]]for[[:space:]]Long-Tailed[[:space:]]Multi-Label[[:space:]]Image[[:space:]]Classification/3a53af20-ceaa-4cd1-97b7-25b047dad714_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Diverse[[:space:]]Text-to-3D[[:space:]]Synthesis[[:space:]]with[[:space:]]Augmented[[:space:]]Text[[:space:]]Embedding/fde25e2b-d996-44a9-9010-e892a31a015b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Divide[[:space:]]and[[:space:]]Fuse_[[:space:]]Body[[:space:]]Part[[:space:]]Mesh[[:space:]]Recovery[[:space:]]from[[:space:]]Partially[[:space:]]Visible[[:space:]]Human[[:space:]]Images/479a8d70-6b81-45c9-9a2b-41900ddb3c7a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Do[[:space:]]Generalised[[:space:]]Classifiers[[:space:]]really[[:space:]]work[[:space:]]on[[:space:]]Human[[:space:]]Drawn[[:space:]]Sketches_/bc04159d-eb34-419d-aca0-ee68db7983e1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Do[[:space:]]text-free[[:space:]]diffusion[[:space:]]models[[:space:]]learn[[:space:]]discriminative[[:space:]]visual[[:space:]]representations_/fcca39c9-690d-476f-a349-a6a59d51ad43_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dolfin_[[:space:]]Diffusion[[:space:]]Layout[[:space:]]Transformers[[:space:]]without[[:space:]]Autoencoder/2494309a-b8d6-48fc-80d5-5b94333c4006_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dolphins_[[:space:]]Multimodal[[:space:]]Language[[:space:]]Model[[:space:]]for[[:space:]]Driving/b5560cb5-15a6-4d92-8a97-922cfc13a651_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Domain[[:space:]]Generalization[[:space:]]of[[:space:]]3D[[:space:]]Object[[:space:]]Detection[[:space:]]by[[:space:]]Density-Resampling/8a5ee75d-cb46-4ce3-beac-84f2a812533e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Domain[[:space:]]Reduction[[:space:]]Strategy[[:space:]]for[[:space:]]Non-Line-of-Sight[[:space:]]Imaging/c6c2423d-8d1d-484a-8bc8-04d19b02d4cf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Domain[[:space:]]Shifting_[[:space:]]A[[:space:]]Generalized[[:space:]]Solution[[:space:]]for[[:space:]]Heterogeneous[[:space:]]Cross-Modality[[:space:]]Person[[:space:]]Re-Identification/5fc09790-84be-4878-b694-c3ea5f24bfcf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Domain-Adaptive[[:space:]]2D[[:space:]]Human[[:space:]]Pose[[:space:]]Estimation[[:space:]]via[[:space:]]Dual[[:space:]]Teachers[[:space:]]in[[:space:]]Extremely[[:space:]]Low-Light[[:space:]]Conditions/ecde83fb-5e32-46f5-9632-8c45e6166c9c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Domain-adaptive[[:space:]]Video[[:space:]]Deblurring[[:space:]]via[[:space:]]Test-time[[:space:]]Blurring/66831ba5-9fd2-4206-bfe2-0a509908f21b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DomainFusion_[[:space:]]Generalizing[[:space:]]To[[:space:]]Unseen[[:space:]]Domains[[:space:]]with[[:space:]]Latent[[:space:]]Diffusion[[:space:]]Models/0b3bd78f-1df1-4861-91a5-28398500bf06_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Domesticating[[:space:]]SAM[[:space:]]for[[:space:]]Breast[[:space:]]Ultrasound[[:space:]]Image[[:space:]]Segmentation[[:space:]]via[[:space:]]Spatial-frequency[[:space:]]Fusion[[:space:]]and[[:space:]]Uncertainty[[:space:]]Correction/febd10be-a3f6-49ba-81c4-17093c128b2c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DoubleTake_[[:space:]]Geometry[[:space:]]Guided[[:space:]]Depth[[:space:]]Estimation/c7414a9d-1d2f-4336-bda3-fbc0d507c6a8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DoughNet_[[:space:]]A[[:space:]]Visual[[:space:]]Predictive[[:space:]]Model[[:space:]]for[[:space:]]Topological[[:space:]]Manipulation[[:space:]]of[[:space:]]Deformable[[:space:]]Objects/b1b8be76-4756-4b36-861b-819f82a8f69a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Drag[[:space:]]Anything_[[:space:]]Motion[[:space:]]Control[[:space:]]for[[:space:]]Anything[[:space:]]using[[:space:]]Entity[[:space:]]Representation/609873a3-bce7-4906-af31-611a31720bde_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DragAPart_[[:space:]]Learning[[:space:]]a[[:space:]]Part-Level[[:space:]]Motion[[:space:]]Prior[[:space:]]for[[:space:]]Articulated[[:space:]]Objects/ee97cbfb-a666-4be9-b74e-b6a04fe53615_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DragVideo_[[:space:]]Interactive[[:space:]]Drag-style[[:space:]]Video[[:space:]]Editing/138bb549-991d-4f3f-aa93-4d5d920a365d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DreamDiffusion_[[:space:]]High-Quality[[:space:]]EEG-to-Image[[:space:]]Generation[[:space:]]with[[:space:]]Temporal[[:space:]]Masked[[:space:]]Signal[[:space:]]Modeling[[:space:]]and[[:space:]]CLIP[[:space:]]Alignment/6e5e3f05-0f66-4e01-a586-661c480f999d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DreamDissector_[[:space:]]Learning[[:space:]]Disentangled[[:space:]]Text-to-3D[[:space:]]Generation[[:space:]]from[[:space:]]2D[[:space:]]Diffusion[[:space:]]Priors/3b00f1b8-762a-423d-aceb-1e1df486aabc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DreamDrone_[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models[[:space:]]are[[:space:]]Zero-shot[[:space:]]Perpetual[[:space:]]View[[:space:]]Generators/572bf8a7-8648-4063-a876-5aff54663cfd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DreamMesh_[[:space:]]Jointly[[:space:]]Manipulating[[:space:]]and[[:space:]]Texturing[[:space:]]Triangle[[:space:]]Meshes[[:space:]]for[[:space:]]Text-to-3D[[:space:]]Generation/66c3d28e-791d-46a2-9966-7af31ca09b86_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DreamMotion_[[:space:]]Space-Time[[:space:]]Self-Similar[[:space:]]Score[[:space:]]Distillation[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Video[[:space:]]Editing/c477982e-df42-4b4c-99a7-758e260c15e0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DreamMover_[[:space:]]Leveraging[[:space:]]the[[:space:]]Prior[[:space:]]of[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]Image[[:space:]]Interpolation[[:space:]]with[[:space:]]Large[[:space:]]Motion/1acf9d30-7441-443d-b225-b1caa85110ce_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DreamReward_[[:space:]]Aligning[[:space:]]Human[[:space:]]Preference[[:space:]]in[[:space:]]Text-to-3D[[:space:]]Generation/b875aeb2-76e7-4156-a9d4-26a3217fb3fd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DreamSampler_[[:space:]]Unifying[[:space:]]Diffusion[[:space:]]Sampling[[:space:]]and[[:space:]]Score[[:space:]]Distillation[[:space:]]for[[:space:]]Image[[:space:]]Manipulation/a6786215-938c-4c19-a135-3cbfc3429168_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DreamScene360_[[:space:]]Unconstrained[[:space:]]Text-to-3D[[:space:]]Scene[[:space:]]Generation[[:space:]]with[[:space:]]Panoramic[[:space:]]Gaussian[[:space:]]Splatting/68b66e4d-0f5b-48bc-b23c-c86728a67574_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DreamScene_[[:space:]]3D[[:space:]]Gaussian-based[[:space:]]Text-to-3D[[:space:]]Scene[[:space:]]Generation[[:space:]]via[[:space:]]Formation[[:space:]]Pattern[[:space:]]Sampling/5abb8fa9-32e0-4d74-b34b-79a19abee05e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DreamStruct_[[:space:]]Understanding[[:space:]]Slides[[:space:]]and[[:space:]]User[[:space:]]Interfaces[[:space:]]via[[:space:]]Synthetic[[:space:]]Data[[:space:]]Generation/9a84924c-860b-480a-a11e-434289823a51_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DreamView_[[:space:]]Injecting[[:space:]]View-specific[[:space:]]Text[[:space:]]Guidance[[:space:]]into[[:space:]]Text-to-3D[[:space:]]Generation/6893e14b-955c-493c-b72d-e013a4cd0afb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DriveDreamer_[[:space:]]Towards[[:space:]]Real-world-driven[[:space:]]World[[:space:]]Models[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/b3a9b2e0-c6b3-45c2-894b-1fe6645bb2d6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DriveLM_[[:space:]]Driving[[:space:]]with[[:space:]]Graph[[:space:]]Visual[[:space:]]Question[[:space:]]Answering/9d73cce6-98b7-4306-a0e2-b71ca2dc6301_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DrivingDiffusion_[[:space:]]Layout-Guided[[:space:]]Multi-View[[:space:]]Driving[[:space:]]Scenarios[[:space:]]Video[[:space:]]Generation[[:space:]]with[[:space:]]Latent[[:space:]]Diffusion[[:space:]]Model/d3f1b68c-7ba2-4ef2-9d4d-e7cd1b55b01b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dropout[[:space:]]Mixture[[:space:]]Low-Rank[[:space:]]Adaptation[[:space:]]for[[:space:]]Visual[[:space:]]Parameters-Efficient[[:space:]]Fine-Tuning/bf1cd395-d52e-4155-b4d6-d9e3901c3d70_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dual-Camera[[:space:]]Smooth[[:space:]]Zoom[[:space:]]on[[:space:]]Mobile[[:space:]]Phones/7cf0c12d-7d0b-49f2-bbb4-9ff839914914_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dual-Decoupling[[:space:]]Learning[[:space:]]and[[:space:]]Metric-Adaptive[[:space:]]Thresholding[[:space:]]for[[:space:]]Semi-Supervised[[:space:]]Multi-Label[[:space:]]Learning/f16ca595-d4df-47f0-8d92-0041cbb98b26_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dual-Path[[:space:]]Adversarial[[:space:]]Lifting[[:space:]]for[[:space:]]Domain[[:space:]]Shift[[:space:]]Correction[[:space:]]in[[:space:]]Online[[:space:]]Test-time[[:space:]]Adaptation/9e630b6a-a313-4990-8f38-8a2e0463175b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dual-Rain_[[:space:]]Video[[:space:]]Rain[[:space:]]Removal[[:space:]]using[[:space:]]Assertive[[:space:]]and[[:space:]]Gentle[[:space:]]Teachers/864d7ca9-2e2d-4a4a-b090-69e76f77dc45_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dual-level[[:space:]]Adaptive[[:space:]]Self-Labeling[[:space:]]for[[:space:]]Novel[[:space:]]Class[[:space:]]Discovery[[:space:]]in[[:space:]]Point[[:space:]]Cloud[[:space:]]Segmentation/3892acba-2abf-43c1-a9ca-4af71e6f6e90_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dual-stage[[:space:]]Hyperspectral[[:space:]]Image[[:space:]]Classification[[:space:]]Model[[:space:]]with[[:space:]]Spectral[[:space:]]Supertoken/6c2308a7-97a5-4067-a4cd-768ed63816cc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DualBEV_[[:space:]]Unifying[[:space:]]Dual[[:space:]]View[[:space:]]Transformation[[:space:]]with[[:space:]]Probabilistic[[:space:]]Correspondences/20332652-3ac3-4bf5-9c09-d44a64357e21_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DualDn_[[:space:]]Dual-domain[[:space:]]Denoising[[:space:]]via[[:space:]]Differentiable[[:space:]]ISP/a63c7ea7-432e-416c-9357-4098920444ac_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DyFADet_[[:space:]]Dynamic[[:space:]]Feature[[:space:]]Aggregation[[:space:]]for[[:space:]]Temporal[[:space:]]Action[[:space:]]Detection/6b9c6c3e-beb8-42c8-b330-a505468e59ec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DySeT_[[:space:]]a[[:space:]]Dynamic[[:space:]]Masked[[:space:]]Self-distillation[[:space:]]Approach[[:space:]]for[[:space:]]Robust[[:space:]]Trajectory[[:space:]]Prediction/37a092bf-cdb2-4a6c-b6e4-01391d654834_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dyn-Adapter_[[:space:]]Towards[[:space:]]Disentangled[[:space:]]Representation[[:space:]]for[[:space:]]Efficient[[:space:]]Visual[[:space:]]Recognition/99b14675-328d-464e-af03-b53b21297b07_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DynMF_[[:space:]]Neural[[:space:]]Motion[[:space:]]Factorization[[:space:]]for[[:space:]]Real-time[[:space:]]Dynamic[[:space:]]View[[:space:]]Synthesis[[:space:]]with[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/726f0fb5-9bd8-4730-98a9-dd1fa94cc5f4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DynamiCrafter_[[:space:]]Animating[[:space:]]Open-domain[[:space:]]Images[[:space:]]with[[:space:]]Video[[:space:]]Diffusion[[:space:]]Priors/b4546a47-92fb-41e4-b357-35b7a2854b32_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dynamic[[:space:]]Data[[:space:]]Selection[[:space:]]for[[:space:]]Efficient[[:space:]]SSL[[:space:]]via[[:space:]]Coarse-to-Fine[[:space:]]Refinement/712a2e2c-220a-4e3c-a25f-91e8c59d68f3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dynamic[[:space:]]Guidance[[:space:]]Adversarial[[:space:]]Distillation[[:space:]]with[[:space:]]Enhanced[[:space:]]Teacher[[:space:]]Knowledge/5b77f32e-c906-448b-afcf-86bd851d16d5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dynamic[[:space:]]Neural[[:space:]]Radiance[[:space:]]Field[[:space:]]From[[:space:]]Defocused[[:space:]]Monocular[[:space:]]Video/2b201000-ab09-44c9-b10c-1f3070b8fbdf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Dynamic[[:space:]]Retraining-Updating[[:space:]]Mean[[:space:]]Teacher[[:space:]]for[[:space:]]Source-Free[[:space:]]Object[[:space:]]Detection/c1db7240-8cc2-4b1c-8f4d-8a9bd9495385_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DynoSurf_[[:space:]]Neural[[:space:]]Deformation-based[[:space:]]Temporally[[:space:]]Consistent[[:space:]]Dynamic[[:space:]]Surface[[:space:]]Reconstruction/ea87eb7e-95f7-4494-a3d0-65eadfd7678e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/DεpS_[[:space:]]Delayed[[:space:]]ε-Shrinking[[:space:]]for[[:space:]]Faster[[:space:]]Once-For-All[[:space:]]Training/185e6ed2-b0a7-4eb0-8603-1f97bf3513ed_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/E.T.[[:space:]]the[[:space:]]Exceptional[[:space:]]Trajectory_[[:space:]]Text-to-camera-trajectory[[:space:]]generation[[:space:]]with[[:space:]]character[[:space:]]awareness/e403c218-18e3-419e-8f52-622631101354_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/E3M_[[:space:]]Zero-Shot[[:space:]]Spatio-Temporal[[:space:]]Video[[:space:]]Grounding[[:space:]]with[[:space:]]Expectation-Maximization[[:space:]]Multimodal[[:space:]]Modulation/35518b33-6a9a-4eaa-a4a0-0877fc589c9b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/E3V-K5_[[:space:]]An[[:space:]]Authentic[[:space:]]Benchmark[[:space:]]for[[:space:]]Redefining[[:space:]]Video-Based[[:space:]]Energy[[:space:]]Expenditure[[:space:]]Estimation/92c524f3-a098-4ff1-8637-ca99428c38a6_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/a5a39b01-2823-428c-b29a-0b036d9c3243_content_list.json b/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/a5a39b01-2823-428c-b29a-0b036d9c3243_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ffa341ea339dad3da9c7675df5bb0ad1745aa99f --- /dev/null +++ b/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/a5a39b01-2823-428c-b29a-0b036d9c3243_content_list.json @@ -0,0 +1,1836 @@ +[ + { + "type": "text", + "text": "Distilling Knowledge from Large-Scale Image Models for Object Detection", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gang Li $^{1,2\\dagger}$ , Wenhai Wang $^{3,2}$ , Xiang Li $^{4}$ , Ziheng Li $^{5}$ , Jian Yang $^{1}$ , Jifeng Dai $^{5,2}$ , Yu Qiao $^{2}$ , and Shanshan Zhang $^{1\\boxtimes}$", + "bbox": [ + 285, + 210, + 722, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nanjing University of Science and Technology $^{2}$ OpenGVLab, Shanghai AI Laboratory $^{3}$ The Chinese University of Hong Kong $^{4}$ Nankai University $^{5}$ Tsinghua University", + "bbox": [ + 233, + 255, + 769, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Large-scale image models have made great progress in recent years, pushing the boundaries of many vision tasks, e.g., object detection. Considering that deploying large models is impractical in many scenes due to expensive computation overhead, this paper presents a new knowledge distillation method, which Distills knowledge from Large-scale Image Models for object detection (dubbed DLIM-Det). To this end, we make the following two efforts: (1) To bridge the gap between the teacher and student, we present a frozen teacher approach. Specifically, to create the teacher model via fine-tuning large models on a specific task, we freeze the pretrained backbone and only optimize the task head. This preserves the generalization capability of large models and gives rise to distinctive characteristics in the teacher. In particular, when equipped with DEtection Transformers (DETRs), the frozen teacher exhibits sparse query locations, thereby facilitating the distillation process. (2) Considering that large-scale detectors are mainly based on DETRs, we propose a Query Distillation (QD) method specifically tailored for DETRs. The QD performs knowledge distillation by leveraging the spatial positions and pair-wise relations of teacher's queries as knowledge to guide the learning of object queries of the student model. Extensive experiments are conducted on various large-scale image models with parameter sizes ranging from 200M to 1B. Our DLIM-Det improves the student with Swin-Tiny by 3.1 mAP when the DINO detector with Swin-Large is used as the teacher. Besides, even when the teacher has 30 times more parameters than the student, DLIM-Det still attains a +2.9 distillation gain.", + "bbox": [ + 261, + 321, + 743, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Object detection $\\cdot$ Knowledge distillation $\\cdot$ Large-scale image model", + "bbox": [ + 261, + 638, + 740, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 680, + 356, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It is proved that scaling up model sizes is a promising direction to make neural networks reach human performance [33, 42, 48, 52]. Recently, great progress has been achieved in large-scale image models [11, 33, 34, 41], pushing performance boundaries of many vision tasks, including object detection, semantic segmentation, etc. These large-scale models demand for high performance computing device, making them hard to deploy in commonly used hardware platforms. Therefore, Knowledge Distillation (KD) is necessary", + "bbox": [ + 212, + 710, + 787, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† This work was done while G. Li was an intern at Shanghai AI Laboratory.", + "bbox": [ + 230, + 811, + 679, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author (shanshan.zhang@njust.edu.cn).", + "bbox": [ + 230, + 825, + 575, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/832a8a9182abe72865c99a7f18a8a47ba47adcdce852c5751193373071b36ae2.jpg", + "image_caption": [ + "Fig. 1: Comparison between previous KD [3] and our DLIM-Det method for distilling large-sized teacher models. Previous methods struggle to distill knowledge from large models, whereas our method works well. Experiments are conducted on Swin Transformer [34]." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 535, + 281 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5bd748ae7395043a83f110af21c45286acdea2da855bc2bd07a5972dbf89c429.jpg", + "image_caption": [ + "Fig.2: Average Recall (AR) comparison on LVIS labels for the frozen teacher and fully fine-tuned teacher. Higher AR numbers indicate better generalization capacity." + ], + "image_footnote": [], + "bbox": [ + 549, + 157, + 784, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to extract powerful and informative knowledge from the large model (teacher) to boost the compact model (student). The teacher knowledge serves as auxiliary supervisions and helps the student achieve better performance than trained only with ground-truth labels [21, 37]. However, existing KD methods [3, 8, 15, 50, 51] for object detection mainly focus on medium- or small-sized models, leaving the distillation from large-scale models rarely explored.", + "bbox": [ + 212, + 371, + 787, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Distilling knowledge from large-scale image models is not trivial. As observed in previous works [35, 36], the discrepancy between the teacher and student becomes severe as the model size of teacher model increases, which hinders the effectiveness of KD methods. As shown in Fig. 1, when directly applying the prior KD method (e.g., DETRDistill [3]) to a detector (e.g., DINO [54]), the performance even degrades as the teacher model becomes stronger. Specifically, though the Swin-Large [34] has $2 \\times$ parameters than the Swin-Base, the latter is a better teacher and outperforms the former by $+0.5\\mathrm{mAP}$ in terms of KD gains. To address this problem, previous works [35, 39, 56] design a teacher assistant strategy, which generates multiple intermediate-sized teacher candidates to bridge the gap between the teacher and student. However, such a multi-step KD method induces additional training costs and severely affects the efficiency. In this paper, we propose that when obtaining the teacher model through fine-tuning the large-scale image backbone on a specific dataset, freezing the backbone and only optimizing the task head can empower the teacher with richer transferable knowledge compared to fully fine-tuning. Large-scale models exhibit increased generalizability due to their strong capacity and training on extensive datasets, however, there is a risk of overfitting when fine-tuned on specific datasets. Through freezing the pre-trained backbone, we maintain the generalizability of large-scale models, resulting in improved distillation performance. The teacher with pretrained large-scale backbone frozen is denoted as the frozen teacher.", + "bbox": [ + 212, + 462, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To quantitatively assess the generalizability of both the frozen and fully fine-tuned (ft.) teachers, we compute the Average Recall (AR) on the LVIS annotations [16], and the results are presented in Fig. 2. Swin-Large [34] based DINO [54] serves as the teacher model. Notably, both teachers are trained only with supervision from the 80 object categories in COCO. The LVIS dataset, while sourcing images from COCO, features", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 320, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a re-annotated and broader spectrum of 1203 categories. As results shown, the frozen teacher outperforms the fully ft. teacher by a margin of $4.6\\%$ in $\\mathrm{AR}_{900}$ , indicating that the frozen teacher generates more generalizable features. Strengthening the generalizability of the teacher model endows it with specific characteristics that facilitate the distillation process. In particular, when equipped with detection transformers, sparse query locations emerge in the frozen teacher.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Advanced large-scale detectors [5, 41, 54, 58] are predominately based on detection transformers, where object queries absorb contextual information for each instance then output predictions. Object queries from the frozen teacher exhibit sparser distribution compared to those from the fully fine-tuned teacher. This sparsity characteristic enables the student model to circumvent unnecessary investments in modeling complex query relations, particularly those leading to excessively compact query locations. To fully exploit sparse query locations from the frozen teacher, we introduce a Query Distillation method. Our Query Distillation aligns object queries between the teacher and student models based on spatial positions and relations. This method consists of two key components: Query Position Distill (QPD) and Query Relation Distill (QRD). Initially, we establish a one-to-one set matching between object queries of two models. Subsequently, the proposed QPD allows the teacher's queries to serve as auxiliary localization supervisions during student model training. Aside from QPD, we further propose Query Relation Distillation to lead the student model to mimic pair-wise relations between teacher's queries. Combining Query Distillation and the frozen teacher can significantly alleviate the learning burden for the student, thereby contributing to a more efficient knowledge distillation process.", + "bbox": [ + 212, + 237, + 789, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To conclude, we propose a new object detection knowledge distillation method, called DLIM-Det, tailored for Large-scale Image Models ranging from 200M to 1B parameters. Leveraging Query Distillation and the frozen teacher approach, DLIM-Det effectively transfers knowledge from large-scale models to compact ones. Extensive experiments on both the Transformer-based (Swin Transformer [34]) and CNN-based (InternImage [41]) models demonstrate the effectiveness of our approach. Results indicate our Query Distillation method works well for detection transformers, surpassing existing KD methods. Additionally, the frozen teacher approach bridges the gap between teacher and student, facilitating efficient distillation from large-scale models. Notably, when the teacher model is Swin-Large, DLIM-Det improves Swin-Tiny based DINO [54] by $3.1\\%$ ; even for the giant InternImage-Huge model ( $1B+$ parameters), it improves the InternImage-Tiny student by $2.9\\%$ , achieving a remarkable 56.1 mAP.", + "bbox": [ + 212, + 494, + 789, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 700, + 369, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge Distillation. As a classical model compression technique, Knowledge Distillation (KD) has been applied to many vision tasks [14, 17, 24, 25, 32, 47, 53]. In this paper, we focus on the KD for object detection. Using the intermediate features of the teacher as knowledge, feature distillation offers deep supervisions for the student [4, 15, 37, 51]. However, feature distillation is usually vulnerable to semantic misalignment between features of two models, and suffers from performance degradation when the discrepancy between two models is too large. Besides, to implement relation-based", + "bbox": [ + 212, + 734, + 789, + 842 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Distilling Knowledge from Large-Scale Image Models for Object Detection", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "distillation, previous works [8, 27, 49] design delicate modules or schemes to capture relations between different instances [8], anchors [27], patches [17], and pixels [49]. At last, the response-based KD methods are also widely investigated in previous works [8, 15, 22, 56], where the probabilistic scores are matched between two models. With the development of detection transformers (DETRs), some recently proposed KD methods [3, 7] are tailored for DETRs. DETRDistill [3] transfers the label assignment results from the teacher to student in order to stabilize the bipartite matching, while D3ETR [7] feeds the teacher queries to the student decoder and then trains the decoder using matching results from the teacher. Although these KD methods achieve decent distillation performance for medium-sized teacher models (such as ResNet-101 [20]), few of them explore the distillation from large-scale image models. In contrast to these works, we focus on distillation from large-scale models with parameters spanning from 200M to 1B. Through our method, we show that larger models lead to greater distillation gains.", + "bbox": [ + 215, + 146, + 787, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Detection. Object detection is a fundamental task in the computer vision field [28, 46]. Detection Transformers (DETRs) achieve great progress recently and gradually dominate object detection. With outstanding scalability, current mainstream large-scale detectors [5, 54, 58] typically rely on DETRs. To address the inefficiency issue of the original DETR [2], Deformable DETR [57] and Adamixer [13] introduce the local attention mechanism to replace the original global attention, resulting in faster model convergence. Some works [31, 43] attempt to insert the spatial prior into the object queries and lead each query to attend to a certain region. DN-DETR [26] designs a query denoising scheme to improve the stability of bipartite graph matching. Based on DN-DETR, DINO [54] presents a contrastive way for denoising training and achieves the state-of-the-art performance. Building on SoTA performance, we employ DINO as the default detector in this paper.", + "bbox": [ + 215, + 387, + 787, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Large-scale Image Models. Some recent works [10, 33, 34, 41, 42, 48] scale up image models to reach 3 billion parameters, demonstrating a very powerful capacity of large-scale image models. Swin Transformer [33], built on the transformer architecture, incorporates techniques like post-norm and scaled cosine attention to enhance the stability of large vision models. Beit-3 [42], adopting the ViT [10] architecture, explores scaling up models through multi-modal pretraining. InternImage [41], utilizing a CNN base, integrates deformable convolution as a core operator to establish large-scale CNN models. After acquiring large-scale pretrained models, there has been significant interest in efficiently fine-tuning these models for downstream tasks. Notably, approaches such as VPT [23] insert learnable parameters into the input space while keeping most of the pretrained model frozen. Similarly, AdaptFormer [6] introduces lightweight learnable modules into the MLP block. Lin et al. [30] investigate the pretraining tasks that best align with the frozen setting. Different from these works, our paper aims to extract transferable knowledge from large models and use it to enhance tiny-scale models via the KD.", + "bbox": [ + 215, + 613, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 320, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 143, + 316, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we provide a concise overview of knowledge distillation and the detection transformer in Sec. 3.1. Subsequently, we introduce the frozen teacher method in Sec. 3.2, followed by an explanation of Query Distillation in Sec. 3.3.", + "bbox": [ + 214, + 176, + 787, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 244, + 341, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Knowledge Distillation (KD). Standard KD enforces the student to align its probabilistic scores $(p^{(s)})$ with those of the teacher via the Kullback-Leibler (KL) divergence [21, 55]:", + "bbox": [ + 214, + 270, + 787, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K D} = p ^ {(t)} \\log \\left(\\frac {p ^ {(t)}}{p ^ {(s)}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 311, + 785, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The teacher's logits $(p^{(t)})$ serve as soft label, providing non-target classes not present in the ground truth labels. Despite being weak, this non-target class information significantly enhances the student's generalization. To amplify this effect, prior works typically apply a high temperature to soften the $p^{(t)}$ .", + "bbox": [ + 214, + 353, + 787, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detection Transformers (DETRs). In DETRs, the backbone extracts image features, which are then inputted into the transformer encoder along with positional encodings. The enhanced image features serve as the key and value, providing contextual information for the subsequent transformer decoder. The input for the decoder is N embeddings, known as object queries, denoted as $Q = \\{q_0, q_1, \\dots, q_{N-1}\\}$ . Each query will be decoded into class probability $p$ and box coordinate $b$ via the transformer decoder.", + "bbox": [ + 214, + 414, + 787, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Frozen Teacher Method", + "text_level": 1, + "bbox": [ + 215, + 527, + 423, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by the NLP domain, where scaling up language models achieves remarkable success [9, 12], some recent works scale up vision models to more than 1 billion parameters and set new records across various vision tasks [33,41,42,48]. It is natural to explore the possibility of transferring the knowledge from these large models to more compact ones using the Knowledge Distillation (KD) technique. However, empirical observations (Fig. 1) suggest that as the model size of the teacher increases, the discrepancy between the teacher and student becomes more severe, limiting the effectiveness of the KD.", + "bbox": [ + 214, + 551, + 787, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we observed that when fine-tuning large-scale pretrained models on a specific dataset to create the teacher model, freezing the backbone can endow the teacher with richer transferable knowledge compared to fully fine-tuning. The overall framework of our DLIM-Det is illustrated in Fig. 3. The pretraining-then-finetuning paradigm is the cornerstone for modern neural networks [18, 19, 45], especially for large-scale models. Pretraining on massive data is essential for achieving top performance in the downstream tasks [41]. Although fully fine-tuning the entire model may enhance task-specific performance, it also heightens the risk of overfitting and diminishes the model's generalizability. We denote the teacher produced by the frozen setting as the frozen teacher. This frozen teacher maintains the generalizability of the large-scale models. As evidenced by the Average Recall (AR) metrics on LVIS labels shown in Fig. 2, the frozen teacher exhibits an increased ability to recognize a broader array of non-target objects", + "bbox": [ + 214, + 659, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Distilling Knowledge from Large-Scale Image Models for Object Detection", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3392f4f0a86addaaf5b05910f262f3bf8e4ae3adc4bb00ce93b59776da8fc22e.jpg", + "image_caption": [ + "Fig. 3: The framework of our DLIM-Det. Freezing the backbone when fine-tuning large-sized models on a specific dataset to create the teacher. The alignment between object queries of the teacher and student is conducted based on positions and relations." + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 764, + 308 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/eef5c0334ad2072d7d74a4776cb78a2b01afe00d52da10165e06176581e1252b.jpg", + "table_caption": [ + "Table 1: The query density with different IoU thresholds." + ], + "table_footnote": [], + "table_body": "
MethodsIoU: 0.5IoU: 0.7
Swin-LIntern-LSwin-LIntern-L
Full finetuning38.133.615.213.0
Frozen21.321.87.58.3
", + "bbox": [ + 367, + 383, + 630, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "while avoiding excessive specialization on the ground truths. Enhanced generalizability endows the teacher with specific characteristics that facilitate the distillation process.", + "bbox": [ + 212, + 444, + 785, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the following, we investigate how these distinctive characteristics enhance the efficiency of the distillation process. When equipped with DETRs, we compare the spatial distributions of object queries under both the frozen and fully fine-tuned settings. We observe that the frozen teacher locates object queries more dispersely and sparsely than the fully fine-tuned teacher. To quantify the dispersion of query spatial distribution, we introduce a metric called Query Density. Query density represents the number of object queries with an Intersection over Union (IoU) greater than $\\tau$ surrounding each ground truth. We average the query density over all GT objects in the COCO 2017 validation set [29], with $\\tau$ set to 0.5 and 0.7, respectively. The results are presented in Tab. 1. As results show, the backbone-frozen strategy significantly affects the query density, reducing it by at least $1/3$ . The fully fine-tuned teacher exhibits high query density values, and each GT object is surrounded by more than 30 queries with IoUs larger than 0.5. Among these compact queries, only one query should predict the object, while the others should predict no object. One example is visualized in Fig. 5. The detector needs to model their relations, identifying the optimal localized query with high confidence and suppressing the others with low confidence. In the case of sparse query locations, an instance is surrounded by fewer object queries, making their relations much easier to model.", + "bbox": [ + 212, + 476, + 787, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When using the fully fine-tuned teacher, the student model needs to allocate unnecessary capacity to model complex query relations in order to mimic the compact query locations of the teacher. In contrast, the sparse query locations provided by the frozen teacher are easier for the student to learn, reducing the learning burden. To fully leverage the sparse query of the frozen teacher, we design a Query Distillation method, which is described in Sec. 3.3.", + "bbox": [ + 212, + 750, + 789, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 320, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6354959c5867e932eca3ac2f94820c2aeeec9ae999a93d7f0bff5f2a72c7ab3b.jpg", + "image_caption": [ + "(a) Comparing four teachers." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 390, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1b20ee9aad885e188f94c9cf6d38c82349a675b13c69d02b7e0b5e70fe9e97b.jpg", + "image_caption": [ + "Input Images" + ], + "image_footnote": [], + "bbox": [ + 413, + 154, + 485, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf6bcba23cbe5bcda067c629234c709461f6e58f39e23703a57b2986968fedf2.jpg", + "image_caption": [ + "Fig. 4: Comparison of different teacher models based on mutual information. (a) Multiple teacher models and their mutual information concerning the input $(I(X;F))$ and target $(I(Y;F))$ . The blue lines represent the process of searching for the model whose representation maximizes $I(X;F) + I(Y;F)$ . \"Full ft.\" denotes full fine-tuning. (b) The reconstructed image qualitatively reflects $I(X;F)$ ." + ], + "image_footnote": [], + "bbox": [ + 413, + 199, + 483, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ebb3b414094d4c68475fb0a664de49691afcbb1e9971a36a6d862fa42e8e289.jpg", + "image_caption": [ + "Full ft. Swin Base" + ], + "image_footnote": [], + "bbox": [ + 488, + 154, + 558, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/658a95210b226657247c29a7f847e8f092caf40756d75f1ff326d32cdb89c0f2.jpg", + "image_caption": [ + "(b) Visualization of reconstructed images." + ], + "image_footnote": [], + "bbox": [ + 488, + 199, + 558, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7cbee10677cb09278bcfc387a01642e2d492d6c382d6ee751e9a64f1d88b495d.jpg", + "image_caption": [ + "Frozen Swin Base" + ], + "image_footnote": [], + "bbox": [ + 560, + 154, + 630, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46ad419ea7f7b5e2f9298a7a50e1721f145b171da41be0bf1af21f4a7062eda3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 199, + 630, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/714a0c20640896fcaee6b9c6a61e84988747e22269c8c90446d33614d2b7eaf7.jpg", + "image_caption": [ + "Full ft. Swin Large" + ], + "image_footnote": [], + "bbox": [ + 633, + 154, + 705, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2d5547af3cbdb2d9c7ea5d47c20442454cd3b65b8982495317ab84518092c66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 199, + 705, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/77719ac267537b1b8b50deb0593299fdce25000d177c55f93b0f748a126ba929.jpg", + "image_caption": [ + "Frozen Swin Large" + ], + "image_footnote": [], + "bbox": [ + 709, + 154, + 779, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98aebe2e1a2640072dd5242b9b6bada62a6062edad234adbd6a9c5d489250c5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 199, + 779, + 243 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Taking into account the aspect of generalization and drawing inspiration from prior research [40], which introduced an optimal teacher model selection algorithm based on mutual information, we also substantiate the superiority of the frozen teacher from the perspective of the mutual information principle. Specifically, the entropy of teacher's representation $(H(F))$ can be formulated as:", + "bbox": [ + 212, + 348, + 787, + 424 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nH (F) = I (X; F) + I (Y; F) + Z, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 443, + 785, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$I(X;F)$ and $I(Y;F)$ represent task-relevant information regarding the input image $X$ and ground truth target $Y$ , respectively. $Z$ refers to nuisance, which is task-irrelevant information. Following standard practice [40, 44], we employ the detection performance (e.g. mAP) to quantify $I(Y;F)$ . To measure the mutual information concerning the input $(I(X;F))$ , we feed the intermediate features into a decoder network (typically, U-Net [38]) to reconstruct the input image $X$ . The average reconstruction loss is used to quantify $I(X;F)$ . Different values of $I(X;F)$ can be clearly distinguished based on the quality of reconstructed images, as illustrated in Fig. 4b. The theoretical proofs and training loss curves can be found in the Appendix.", + "bbox": [ + 212, + 479, + 787, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After defining $I(X;F)$ and $I(Y;F)$ , we compare base-scale and large-scale models under both full fine-tuning and frozen settings, with results depicted in Fig. 4a, where mutual information $(I(X;F)$ and $I(Y;F))$ is normalized to ensure the scale matching. Among the four teacher models, the frozen Swin-L attains the maximum of task relevant information $I(X;F) + I(Y;F)$ , achieving a delicate trade-off between $I(X;F)$ and $I(Y;F)$ . A model with high values of $I(X;F) + I(Y;F)$ can effectively suppress nuisance $Z$ and serves as an informative teacher. Though Base-scale models exhibit higher values of $I(X;F)$ compared to Large-scale ones, they lag in terms of $I(Y;F)$ , which may result in inferior KD performance. Empirical experiments (Tab. 3) support our assumption: the frozen Swin-L achieves the best distillation performance among the four teacher models. To conclude, the teacher model selection algorithm [40], grounded in mutual information, further supports our frozen teacher method.", + "bbox": [ + 212, + 619, + 789, + 801 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Distilling Knowledge from Large-Scale Image Models for Object Detection", + "bbox": [ + 282, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1 The features are from the last layer of Transformer Encoder in DETRs.", + "bbox": [ + 220, + 824, + 653, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Query Distillation", + "text_level": 1, + "bbox": [ + 215, + 146, + 382, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Query Distillation aligns the spatial positions and pair-wise relations of queries between the teacher and student, consisting of Query Relation Distill and Query Relation Distill. Initially, we establish the one-to-one matching for object queries of two models, where the cost matrix is computed based on classification and localization costs. Given the cost matrix, Hungarian Matching is used to solve the graph matching, yielding the optimal permutation function $\\hat{\\sigma}$ :", + "bbox": [ + 212, + 172, + 787, + 263 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\sigma} = \\underset {\\sigma \\in \\mathfrak {S} _ {N}} {\\arg \\min } \\sum_ {i} ^ {N} \\mathcal {L} _ {\\text {m a t c h}} \\left(q _ {i} ^ {(t)}, q _ {\\sigma (i)} ^ {(s)}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 275, + 787, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $N$ equals the number of predicted objects. Among all permutations of $N$ elements $\\mathfrak{S}_N$ , $\\hat{\\sigma}$ is the one with the lowest cost. Based on the class probability $p$ and box coordinate $b$ decoded from each object query, we define the cost between the teacher query $q^{(t)}$ and student $q^{(s)}$ as follows:", + "bbox": [ + 214, + 327, + 784, + 387 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a t c h}} \\left(q ^ {(t)}, q ^ {(s)}\\right) = \\alpha_ {1} \\mathcal {L} _ {K L} \\left(p ^ {(t)}, p ^ {(s)}\\right) + \\alpha_ {2} \\mathcal {L} _ {L 1 L o s s} \\left(b ^ {(t)}, b ^ {(s)}\\right) + \\alpha_ {3} \\mathcal {L} _ {G I o U} \\left(b ^ {(t)}, b ^ {(s)}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 397, + 785, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\alpha_{1},\\alpha_{2},\\alpha_{3}$ are set to 2,5,2 empirically.", + "bbox": [ + 214, + 431, + 519, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Query Position Distill. Instead of matching the probabilistic scores of two models, we observe that the spatial position of object queries can serve as informative knowledge. The position of an object query refers to the bounding box decoded from the query, represented by a vector: $b \\in [x, y, w, h]$ , where $x$ and $y$ denote the center coordinates, while $w$ and $h$ represent the width and height, respectively. Given the matching $\\sigma$ between object queries of two models, we formulate the procedure of Query Position Distill as follows:", + "bbox": [ + 212, + 446, + 787, + 551 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {Q P D} = \\sum_ {i} ^ {N} \\beta_ {1} \\mathcal {L} _ {L 1} \\left(b _ {i} ^ {(t)}, b _ {\\sigma (i)} ^ {(s)}\\right) + \\beta_ {2} \\mathcal {L} _ {G I o U} \\left(b _ {i} ^ {(t)}, b _ {\\sigma (i)} ^ {(s)}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 560, + 787, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\beta_{1},\\beta_{2}$ are set to 5 and 2, empirically. Through the thorough experiments (listed in Tab. 8b), we empirically find that besides the foreground queries, even poorly localized teacher queries that cover background regions still contain valuable knowledge to distill.", + "bbox": [ + 212, + 613, + 785, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to the one-to-one label assignment, only a small subset of object queries receives localization supervision in each image. For example, on the COCO dataset [29], which typically contains an average of 7 instances per image, only 7 out of the 900 queries are involved in the regression loss. This sparsity of localization supervision impairs the training of detection transformers. By the proposed Query Position Distill, the spatial positions of teacher's queries serve as additional supervisions to provide richer information than the sparse ground truth labels.", + "bbox": [ + 212, + 659, + 787, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Query Relation Distill. In previous works, relation-based KD methods have been widely investigated. These methods aim to capture various relations between different entities such as instances [8, 22], pixels [49], anchors [27] or patches [17]. To model these relations, intricate modules, e.g., GcBlock [1], and complex relation mechanisms, e.g. Rank Mimicking [27], have been designed, which not only lack interpretability but", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 320, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b27d4468abbc83abfcc0c3a1d77e4e58efdd09cd6ebb3a1e561480dcdb34e7ec.jpg", + "image_caption": [ + "Fig.5: An example to show the position, score and relation of object queries. Query relations are obtained from self-attention. Only 3 queries are visualized for a clear presentation." + ], + "image_footnote": [], + "bbox": [ + 308, + 147, + 694, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "also introduce additional training costs. In contrast, the transformer decoder of DETRs explicitly models the pair-wise relations between object queries through the self-attention. We denote by $A = [a_{ij}]_{N \\times N}$ the attention map from the self-attention, where $N$ denotes the number of queries. L1 Loss is employed to narrow the distance between $A^{(t)}$ and $A^{(s)}$ :", + "bbox": [ + 212, + 323, + 787, + 395 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {Q R D} = \\frac {1}{N} \\sum_ {i} ^ {N} \\sum_ {j} ^ {N} \\left| A _ {i, j} ^ {(t)} - A _ {\\sigma (i), \\sigma (j)} ^ {(s)} \\right|, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 396, + 785, + 436 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\sigma$ denotes the permutation function to match the queries between the teacher and student. Query relations are critical for end-to-end detection. As shown in Fig. 5, precise query relations allow the best localized query to pop out with high classification scores, while the others will be suppressed with low scores.", + "bbox": [ + 212, + 441, + 787, + 501 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The overall training loss $\\mathcal{L}_{\\text{overall}}$ is:", + "bbox": [ + 238, + 503, + 480, + 517 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o v e r a l l}} = \\gamma_ {1} \\mathcal {L} _ {\\text {t a s k}} + \\gamma_ {2} \\mathcal {L} _ {Q P D} + \\gamma_ {3} \\mathcal {L} _ {Q R D}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 527, + 785, + 544 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\gamma_{1},\\gamma_{2},\\gamma_{3}$ are set to 1, 2, 1, respectively.", + "bbox": [ + 214, + 553, + 522, + 568 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We also find that the parameters of the detection head in the teacher model serve as a source of knowledge. Even for heterogeneous DETRs, the basic blocks in their heads are consistent, including attention and feed forward network (FFN). Through empirical investigations, we prove that inheriting the parameters of these basic blocks from the teacher head to student head is convenient and beneficial.", + "bbox": [ + 212, + 569, + 787, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 665, + 356, + 683 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental settings", + "text_level": 1, + "bbox": [ + 214, + 695, + 408, + 710 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets and Backbone. We conduct the experiments on the MS COCO 2017 dataset [29], which is a commonly used object detection dataset. The dataset is split into the train2017 set with 118k images and the val2017 set with 5k images. All models are trained on the train2017 set, and evaluated on the val2017 set. Results are reported with two backbones: Swin Transformer [34] and InternImage [41]. We employ the COCO-style average precision (mAP) as the metric. DINO [54] is the default detector.", + "bbox": [ + 212, + 719, + 802, + 809 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation details. All models, except for the Intern-Huge model, are trained on $8 \\times 32\\mathrm{G}$ V100 GPUs with a batch size of 16. In contrast, the Intern-Huge models are", + "bbox": [ + 214, + 810, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Distilling Knowledge from Large-Scale Image Models for Object Detection", + "bbox": [ + 282, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/d567d56e780285c5eefd84c92336e1f8b91364bb01186da22fe23bff4ffafa94.jpg", + "table_caption": [ + "Table 2: Effectiveness of each component." + ], + "table_footnote": [], + "table_body": "
MethodsSwin TransformerInternImage
mAP\\( AP_{50} \\)\\( AP_{75} \\)mAP\\( AP_{50} \\)\\( AP_{75} \\)
Teacher: Large Scale58.177.163.556.975.861.9
Student: Tiny Scale51.369.056.053.271.558.0
Query DistillFrozen TeacherInherit Head
✓51.6 (+0.3)69.556.454.2 (+1.0)72.559.2
✓✓52.2 (+0.9)70.157.054.7 (+1.5)72.959.8
✓✓53.7 (+2.4)71.758.855.2 (+2.0)73.760.3
✓✓✓54.4 (+3.1)72.159.455.9 (+2.7)74.161.4
", + "bbox": [ + 279, + 157, + 720, + 267 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/04d18e14147b14a15caa3e6a679b05fe590f03213812e4157a88ed50afc9678f.jpg", + "table_caption": [ + "Table 3: Distillation performance comparison between Large- and Base-scale teacher models. Numbers in brackets refer to distillation gains against the original student (Swin-T, 51.3 mAP)." + ], + "table_footnote": [], + "table_body": "
TeacherFreeze BackboneDistilled Student
mAPAP50AP75
Large-scale✓54.4 (+3.1)72.1 (+3.1)59.4 (+3.4)
Base-scale✓53.5 (+2.2)71.5 (+2.5)58.5 (+2.5)
Base-scale×53.1 (+1.8)71.1 (+2.1)58.0 (+2.0)
", + "bbox": [ + 364, + 308, + 637, + 373 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "trained on 80GB A100 GPUs. Student models are trained with 12 epochs with AdamW as the optimizer. The initial learning rate is set to 0.0001, which is divided by 10 at the 11th epoch. Following the default practice of DINO [54], the number of object query is set to 900.", + "bbox": [ + 212, + 393, + 787, + 455 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 479, + 375, + 494 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In ablation studies, unless otherwise stated, we adopt the tiny-scale models with around 30M parameters as the student.", + "bbox": [ + 212, + 506, + 787, + 537 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Effects of each component. In Tab. 2, we verify the effectiveness of each component, including the frozen teacher method, Query Distillation, and head inheriting. When facing large-scale teacher models, Query Distillation produces a modest increase of $+0.3 / 1.0$ mAP for Swin/Intern, respectively. Subsequently, we replace the fully fine-tuned teacher with the frozen teacher and perform Query Distillation again. The results demonstrate that the frozen teacher significantly amplifies the distillation gains. Specifically, compared to the fully fine-tuned teacher, the frozen teacher yields an additional increase of $+0.6 / 0.5\\%$ for Swin/Intern models, respectively. Moreover, inheriting the transformer's parameters of the teacher's head produces noticeable improvements, suggesting that the head parameters of the teacher contain some general knowledge that is somewhat independent of the backbone structure. Building on head parameter inheriting, the frozen teacher outperforms the fully fine-tuned teacher by $+0.7 / 0.7\\%$ for Swin/Intern models. These experimental findings validate that the frozen teacher is easier to distill than the previously employed fully fine-tuned teacher. Finally, by combining the three components, DLIM-Det, achieves significant improvements, surpassing the student baseline by $+3.1 / 2.7\\%$ for the Swin/Intern models, respectively.", + "bbox": [ + 212, + 537, + 789, + 779 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Distillation on medium- and large-sized teacher models. In Tab. 3, we compare the distillation performance under Base- and Large-scale teacher models. With the help of the frozen teacher method, a large-scale teacher surpasses its base-scale counterpart by $+0.9$ mAP. These results affirm the core premise of this paper: larger teacher models lead", + "bbox": [ + 212, + 780, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 320, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/6748c80de52614409e1056b370a16d770a4b496a658eb3c5fd600cced7e39ea6.jpg", + "table_caption": [ + "Table 4: The frozen setting affects the detection performance." + ], + "table_footnote": [], + "table_body": "
MethodsFreeze BackboneSchedulemAP\\( AP_{50} \\)\\( AP_{75} \\)
Swin-Large×3x58.177.163.5
✓3x56.6 (-1.5)74.961.9
Intern-Large×2x57.876.563.1
✓2x55.8 (-2.0)74.461.0
", + "bbox": [ + 359, + 160, + 640, + 239 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d48f63c7b9048b56a1acc80e90c815fd4f32ba0b355eefced1ec2bcb35ecc5ee.jpg", + "table_caption": [ + "Table 5: Applying different training schedules to the Frozen and Full ft. settings as to obtain teachers with identical performance. The KD contains Query Distillation and the inheriting head." + ], + "table_footnote": [], + "table_body": "
MethodsFreeze BackboneSchedulemAP\\( AP_{50} \\)\\( AP_{75} \\)
Teacher:
T1: Swin-Large×1x56.975.662.2
T2: Swin-Large✓3x56.774.961.9
Student:
Swin-Tiny-1x51.369.056.0
KD from T1-1x53.1 (+1.8)70.957.9
KD from T2-1x54.4 (+3.1)72.159.4
", + "bbox": [ + 334, + 277, + 661, + 398 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/9adce99e3ee610a82dda73c27bccc392c7056c4ed1f757f9431439b22ddd6c06.jpg", + "table_caption": [ + "Table 6: Effectiveness of Query Position Distill (QPD) and Query Relation Distill (QRD). \"T/S\" denotes the Teacher/Student." + ], + "table_footnote": [], + "table_body": "
MethodsInternImageSwin Transformer
mAPAP50AP75mAPAP50AP75
T: Base54.673.659.355.874.460.7
S: Tiny53.271.558.051.369.056.0
+ QPD54.2 (+1.0)72.459.151.6 (+0.3)69.556.2
+ QRD54.2 (+1.0)72.959.151.8 (+0.5)70.256.5
+ QPD & QRD54.5 (+1.3)73.259.552.1 (+0.8)69.956.7
", + "bbox": [ + 334, + 436, + 661, + 523 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "to greater distillation gains. Additionally, the frozen teacher method generalizes well to base-scale models. When using the base-scale model as a teacher, the frozen teacher surpasses the fully fine-tuned teacher by $0.4\\mathrm{mAP}$ .", + "bbox": [ + 212, + 536, + 784, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fair comparison between the fully fine-tuned teacher and frozen teacher. Freezing the backbone typically results in performance degradation to some extent compared to fully fine-tuning. As shown in Tab. 4, under the same training schedule, the frozen teacher lags behind the fully fine-tuned teacher by around $2\\mathrm{mAP}$ . One might argue that, owing to lower detection performance, the frozen teacher narrows the gap between the teacher and student, thereby attaining better distillation results. In Tab. 5, we conduct experiments to challenge this perspective. Concretely, we assign long and short training schedules to the frozen and fully fine-tuned teachers, resulting in teachers with nearly identical performance. By ensuring the same gap between the teacher and student, we find that the frozen teacher surpasses the fully fine-tuned teacher by a large margin, +1.3 mAP. These results highlight the intrinsic advantages of the frozen teacher approach, such as sparse query locations.", + "bbox": [ + 212, + 582, + 785, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effectiveness of the Query Distillation. The Query Distillation is specifically designed for detection transformers, including Query Position Distill (QPD) and Query Relation Distill (QRD). We ablate each of them step by step in Tab. 6. Here, we employ the medium-sized models as the teacher, to exclude the effects of the large gap between the teacher and student. Besides the experiments on the advanced network architecture, such", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Distilling Knowledge from Large-Scale Image Models for Object Detection", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/6f2bd9657515d9a86cfe999097b125956efc8cd6c75115d41ae7615f1c3f43b9.jpg", + "table_caption": [ + "Table 7: Applying Query Position Distill (QPD) and Query Relation Distill (QRD) to classical backbones (e.g. ResNet [20])." + ], + "table_footnote": [], + "table_body": "
MethodsResnet
mAP\\( AP_{50} \\)\\( AP_{75} \\)
T: ResNet10150.067.754.4
S: ResNet5048.966.453.4
+ QPD49.5 (+0.5)67.253.9
+ QRD49.5 (+0.5)67.554.1
+ QPD & QRD49.8 (+0.9)67.654.2
", + "bbox": [ + 397, + 172, + 599, + 255 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/69d333cbabc1d690fd3d43b926d3de24b4ed52fe4f7164495a9867a4a1ad0764.jpg", + "table_caption": [ + "Table 8: Study on distillation positions and teacher query selection.", + "(a) Study on distillation positions. Teacher: InternBase (54.6 mAP); Student: Interm-Tiny (53.2 mAP)" + ], + "table_footnote": [], + "table_body": "
KD on decoder layermAP\\( AP_{50} \\)\\( AP_{75} \\)
123456
✓✓✓54.172.659.0
✓✓✓53.872.358.8
✓✓✓✓✓✓54.573.259.5
✓54.172.659.0
✓53.572.158.3
✓52.470.957.0
", + "bbox": [ + 277, + 300, + 464, + 390 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ec2c07121941d1ca4b90a0520beddd920562595d9a128313be199b5af2e20de0.jpg", + "table_caption": [ + "(b) Study on teacher query selection. FG refers to foreground; BG refers to background." + ], + "table_footnote": [], + "table_body": "
MethodsmAP\\( AP_{50} \\)\\( AP_{75} \\)
T: Intern-Base54.673.659.3
S: Intern-Tiny53.271.558.0
No Sample (900)54.272.659.1
Random Sample (300)54.573.259.5
FG queries (IoU > 0.3)54.272.559.2
BG queries (IoU < 0.3)54.272.859.2
", + "bbox": [ + 529, + 301, + 720, + 390 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/dbf7bb15c47a6c49cf1a102b8a9def62f89be5b7bba9d53cb5fe9450a986b0ec.jpg", + "table_caption": [ + "Table 9: Applying our DLIM-Det to the huge model (InternImage-Huge) with 1B parameters to verify the scalability. Numbers in brackets refer to distillation gains against the original student." + ], + "table_footnote": [], + "table_body": "
Teacher#ParamsDistilled Student
mAPAP50AP75
Intern-Huge1.08B56.1 (+2.9)74.4 (+2.9)61.5 (+3.5)
Intern-Large223M55.9 (+2.7)74.1 (+2.6)61.4 (+3.4)
", + "bbox": [ + 357, + 430, + 638, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "as Swin Transformer and InternImage (Tab. 6), we also report results on ResNet [20] in Tab. 7. Our Query Distillation method yields consistent improvements on three different backbones, at least $+0.8$ mAP. Furthermore, combining the QPD and QRD together achieves better performance than using them individually, verifying the effectiveness of each component.", + "bbox": [ + 212, + 488, + 787, + 563 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Study on the distillation positions. To investigate the impact of Query Distillation on different decoder layers in DETRs, we conducted experiments, and the results are presented in Tab. 8a. When applying Query Distillation to all 6 decoder layers, we observe the best performance, surpassing the counterparts where it was applied to the first three or last three decoder layers. Furthermore, we examine the performance when applying Query Distillation to only one decoder layer. The results show that distillation on shallow decoder layers outperforms distillation on deeper layers.", + "bbox": [ + 212, + 566, + 787, + 671 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Selection of teacher queries. Using our Query Distillation method, we further conduct experiments to investigate the impact of different subsets of teacher queries on knowledge transfer. Results are reported in Tab. 8b. First, we perform KD on all teacher queries, obtaining $+1.0\\%$ gains. Then we randomly sample 300 queries and achieve further improvements, surpassing the 900 query baseline by $+0.3\\%$ . To explore alternative query selection strategies, we define the queries whose IoUs with GT are greater than 0.3 as foreground queries, otherwise as background queries. Distilling knowledge from either pure foreground (namely IoU $>0.3$ in Tab. 8b) or pure background queries results in inferior performance compared to random sampling. This suggests that both foreground and background queries contain valuable knowledge, and distilling knowledge from a combination of them yields optimal results.", + "bbox": [ + 212, + 674, + 789, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 320, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/69db4f5e5f7319b0bb1e88225d97a7ea2ab8e2726f0d2698657535d609ca3374.jpg", + "table_caption": [ + "Table 10: Experiments on heterogeneous teacher-student pairs. Heterogeneous components are underlined. Numbers in brackets denote distillation gains against the original student." + ], + "table_footnote": [], + "table_body": "
TeacherStudentDistilled Student
mAPAP50AP75
DINO-Swin-LDeformable DETR-Swin-T49.6 (+3.5)68.3 (+3.4)53.7 (+3.9)
DINO-Intern-HDINO-Swin-T53.3 (+2.0)71.3 (+2.3)58.3 (+2.3)
", + "bbox": [ + 305, + 174, + 691, + 227 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/aa5053da2905d5598d7ac6a4b94e897acf63174865db976f100be363e1877e70.jpg", + "table_caption": [ + "Table 11: Comparison with state-of-the-art knowledge distillation methods." + ], + "table_footnote": [], + "table_body": "
MethodsSwin TransformerInternImage
mAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)mAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
Teacher: Large58.177.163.541.262.173.957.876.563.141.661.373.0
Student: Tiny51.369.056.034.554.466.153.271.558.035.556.669.3
FGD [49]51.1 (-0.2)69.255.834.353.865.652.8 (-0.4)71.157.636.456.466.8
MGD [50]51.1 (-0.2)69.855.834.954.266.154.2 (+1.0)72.859.138.257.669.4
FitNet [37]51.3 (+0.0)70.055.934.754.566.353.8 (+0.5)72.458.738.257.468.2
DETRDistill [3]51.7 (+0.4)69.456.435.054.866.054.7 (+1.5)72.859.838.558.269.4
Our DLIM-Det54.4 (+3.1)72.159.438.157.469.755.9 (+2.7)74.161.440.359.470.8
", + "bbox": [ + 272, + 248, + 725, + 358 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Scalability to huge teacher models. We apply DLIM-Det to InternImage-Huge with 1.08 billion parameters. From Tab. 9, we observe that DLIM-Det achieves satisfactory performance even when the teacher has more than 30 times the parameters of the student. Here, the student is represented by Intern-Tiny with 30M parameters, achieving a mAP of 53.2. Specifically, Intern-Huge produces a $+2.9$ distillation gain on Intern-Tiny based student and surpasses Intern-Large by 0.2 mAP (56.1 v.s. 55.9).", + "bbox": [ + 212, + 369, + 785, + 459 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Distillation on heterogeneous architectures. To verify the generalization of DLIM-Det, we conduct experiments on heterogeneous teacher and student models, including heterogeneous detectors and heterogeneous backbones. For heterogeneous detectors, with DINO [54] as the teacher and Deformable DETR [57] as the student, DLIM-Det improves the student from 46.1 to $49.6\\mathrm{mAP}$ , as reported in Tab. 10. In the case of heterogeneous backbones, where Intern-Huge/Swin-Tiny are employed as the teacher/student, DLIM-Det obtains a considerable improvement of $+2.0\\mathrm{mAP}$ . These results show the strong generalization of DLIM-Det.", + "bbox": [ + 212, + 460, + 787, + 580 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 214, + 603, + 553, + 618 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Main results on two backbones. In Tab. 11, we compare our DLIM-Det with several popular KD methods. FitNet [37] is a classical feature distillation method, which introduces adaptation layers to bridge the semantic gap between the teacher and student. However, FitNet fails to improve the performance of the transformer-based backbone (Swin Transformer), showing no improvements. MGD [50] is inspired by the masked image modeling [18] and randomly masks pixels of the student's features. It then leads the student to generate the teacher's features in the masked regions. MGD obtains a noticeable improvement of $+1.0\\%$ for InternImage, but struggles to improve the Swin Transformer. FGD [49] leverages the GcBlock [1] to capture the relations between pixels from feature maps. However, FGD does not perform well on both backbones, indicating that the hand-crafted relation mechanism may conflict with the intrinsic relation modeling of detection transformers. In fact, FGD even hurts the performance of InternImage by $-0.4\\%$ . The recently proposed DETRDistill [3] is designed for distilling DETRs. In addition to performing customized logits and feature distillation, it also distills the label", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Distilling Knowledge from Large-Scale Image Models for Object Detection", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "assignment results of teacher. Notably, our DLIM-Det outperforms it by a large margin, at least $+1.2\\%$ . We explain that the Query prior Assignment Distillation in DETRDistill serves a role akin to the De-Noising group in DINO, thereby it only yields minor improvements in our experiments. Moreover, another contribution of DETRDistill [3] is a feature imitation method, which is vulnerable to significant discrepancy between the teacher and student, as confirmed by other works [37,49,50].", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "These experimental findings suggest that while previous works can achieve decent distillation gains for medium-sized teacher, they encounter challenges in distilling large-scale teacher models. In contrast, DLIM-Det demonstrates significant distillation performance specifically tailored for large-scale models.", + "bbox": [ + 212, + 237, + 787, + 297 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/bfdfe9ead35c4d01fa735f899a614a25c2aa3b716e7bfcf20a00d766ea017cdd.jpg", + "table_caption": [ + "Table 12: Comparison with the Teacher Assistant (TA) on two backbones. The TA consists of two steps, thus the overall training time is the sum of the training time for both steps." + ], + "table_footnote": [], + "table_body": "
MethodsSwin TransformerInternImage
mAP\\( AP_{50} \\)\\( AP_{75} \\)Train TimemAP\\( AP_{50} \\)\\( AP_{75} \\)Train Time
Baseline:
Teacher: Large58.177.163.5-57.876.563.1-
Intermediate Teacher: Base55.874.460.7-54.673.659.3-
Student: Tiny51.369.056.0-53.271.558.0-
TA step1 (Large→Base)57.5 (+1.7)76.162.827.3h56.8 (+2.2)75.262.134.5h
TA setup2 (Base→Tiny)53.7 (+2.4)71.758.919.2h55.7 (+2.5)74.060.723.6h
Our DLIM-Det (Large→Tiny)54.4 (+3.1)72.159.420h55.9 (+2.7)74.161.424.8h
", + "bbox": [ + 271, + 337, + 727, + 452 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Comparison with the Teacher Assistant. The Teacher Assistant (TA) [35, 39, 56] is proved effective in reducing the gap between the teacher and student models. Following the previous work [35], we use the base-scale model with approximately 90M parameters as the TA model. The TA model is distilled from the teacher model and then serves as the teacher for the student model. The results in Tab. 12 demonstrate that our DLIM-Det, which directly distills knowledge from large-scale models, achieves superior performance and higher efficiency compared to the TA method. Specifically, by using only half of the training time, DLIM-Det outperforms Teacher Assistant by $0.7 / 0.2$ mAP on Swin/Intern models, respectively.", + "bbox": [ + 212, + 463, + 787, + 599 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 625, + 344, + 641 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce DLIM-Det, a novel knowledge distillation method specifically tailored for large-scale image models in the context of object detection. To bridge the gap between the teacher and student, we propose the frozen teacher method, which involves freezing the pretrained backbone during teacher creation. The frozen teacher preserves the generalization capacity of large-scale models and gives rise to distinctive characteristics, such as sparse query locations. Through the proposed Query Distillation method, we harness the full advantages offered by the frozen teacher, thereby facilitating the distillation process. Extensive experiments provide robust evidence that larger teacher models can indeed lead to greater distillation gains through DLIM-Det.", + "bbox": [ + 212, + 657, + 787, + 792 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. This work is partially supported by National Natural Science Foundation of China (No. 62322602), Natural Science Foundation of Jiangsu Province, China (No. BK20230033), and National Natural Science Foundation of China (No. 62172225).", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 320, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 310, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Cao, Y., Xu, J., Lin, S., Wei, F., Hu, H.: Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In: Proceedings of the IEEE/CVF international conference on computer vision workshops. pp. 0-0 (2019) 8, 13", + "2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020) 4", + "3. Chang, J., Wang, S., Xu, H.M., Chen, Z., Yang, C., Zhao, F.: Detrdistill: A universal knowledge distillation framework for detr-families. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6898-6908 (2023) 2, 4, 13, 14", + "4. Chen, G., Choi, W., Yu, X., Han, T., Chandraker, M.: Learning efficient object detection models with knowledge distillation. Advances in neural information processing systems 30 (2017) 3", + "5. Chen, Q., Chen, X., Zeng, G., Wang, J.: Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085 (2022) 3, 4", + "6. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) 4", + "7. Chen, X., Chen, J., Liu, Y., Zeng, G.: DG etr: Decoder distillation for detection transformer arXiv preprint arXiv:2211.09768 (2022) 4", + "8. Dai, X., Jiang, Z., Wu, Z., Bao, Y., Wang, Z., Liu, S., Zhou, E.: General instance distillation for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7842-7851 (2021) 2, 4, 8", + "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 5", + "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 4", + "11. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. arXiv preprint arXiv:2211.07636 (2022) 1", + "12. Floridi, L., Chiriatti, M.: Gpt-3: Its nature, scope, limits, and consequences. Minds and Machines 30, 681–694 (2020) 5", + "13. Gao, Z., Wang, L., Han, B., Guo, S.: Adamixer: A fast-converging query-based object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5364-5373 (2022) 4", + "14. Guo, G., Han, L., Wang, L., Zhang, D., Han, J.: Semantic-aware knowledge distillation with parameter-free feature uniformization. Visual Intelligence 1(1), 6 (2023) 3", + "15. Guo, J., Han, K., Wang, Y., Wu, H., Chen, X., Xu, C., Xu, C.: Distilling object detectors via decoupled features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2154-2164 (2021) 2, 3, 4", + "16. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019) 2", + "17. Hao, Z., Guo, J., Jia, D., Han, K., Tang, Y., Zhang, C., Hu, H., Wang, Y.: Learning efficient vision transformers via fine-grained manifold distillation. In: Advances in Neural Information Processing Systems (2021) 3, 4, 8" + ], + "bbox": [ + 220, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Distilling Knowledge from Large-Scale Image Models for Object Detection", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022) 5, 13", + "19. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 5", + "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 4, 12", + "21. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 2, 5", + "22. Huang, T., You, S., Wang, F., Qian, C., Xu, C.: Knowledge distillation from a stronger teacher. arXiv preprint arXiv:2205.10536 (2022) 4, 8", + "23. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) 4", + "24. Jia, Z., Sun, S., Liu, G., Liu, B.: Mssd: multi-scale self-distillation for object detection. Visual Intelligence 2(1), 8 (2024) 3", + "25. Kang, Z., Zhang, P., Zhang, X., Sun, J., Zheng, N.: Instance-conditional knowledge distillation for object detection. Advances in Neural Information Processing Systems 34, 16468-16480 (2021) 3", + "26. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022) 4", + "27. Li, G., Li, X., Wang, Y., Zhang, S., Wu, Y., Liang, D.: Knowledge distillation for object detection via rank mimicking and prediction-guided feature imitation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1306-1313 (2022) 4, 8", + "28. Li, G., Li, X., Zhang, S., Yang, J.: Towards more reliable evaluation in pedestrian detection by rethinking \"ignore regions\". Visual Intelligence 2(1), 4 (2024) 4", + "29. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 6, 8, 9", + "30. Lin, Y., Liu, Z., Zhang, Z., Hu, H., Zheng, N., Lin, S., Cao, Y.: Could giant pre-trained image models extract universal representations? Advances in Neural Information Processing Systems 35, 8332-8346 (2022) 4", + "31. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022) 4", + "32. Liu, Y., Chen, K., Liu, C., Qin, Z., Luo, Z., Wang, J.: Structured knowledge distillation for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2604-2613 (2019) 3", + "33. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12009-12019 (2022) 1, 4, 5", + "34. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) 1, 2, 3, 4, 9", + "35. Mirzadeh, S.I., Farajtabar, M., Li, A., Levine, N., Matsukawa, A., Ghasemzadeh, H.: Improved knowledge distillation via teacher assistant. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 5191-5198 (2020) 2, 14" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 218, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 318, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Park, D.Y., Cha, M.H., Kim, D., Han, B., et al.: Learning student-friendly teacher networks for knowledge distillation. Advances in neural information processing systems 34, 13292-13303 (2021) 2", + "37. Romero, A., Ballas, N., Kahou, S.E., Chassang, A., Gatta, C., Bengio, Y.: Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550 (2014) 2, 3, 13, 14", + "38. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) 7", + "39. Son, W., Na, J., Choi, J., Hwang, W.: Densely guided knowledge distillation using multiple teacher assistants. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9395-9404 (2021) 2, 14", + "40. Wang, C., Yang, Q., Huang, R., Song, S., Huang, G.: Efficient knowledge distillation from model checkpoints. Advances in Neural Information Processing Systems 35, 607-619 (2022) 7", + "41. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., et al.: Interimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778 (2022) 1, 3, 4, 5, 9", + "42. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442 (2022) 1, 4, 5", + "43. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2567-2575 (2022) 4", + "44. Wang, Y., Ni, Z., Song, S., Yang, L., Huang, G.: Revisiting locally supervised learning: an alternative to end-to-end training. arXiv preprint arXiv:2101.10832 (2021) 7", + "45. Wei, F., Gao, Y., Wu, Z., Hu, H., Lin, S.: Aligning pretraining for detection via object-level contrastive learning. Advances in Neural Information Processing Systems 34, 22682-22694 (2021) 5", + "46. Xu, T., Zhu, X.F., Wu, X.J.: Learning spatio-temporal discriminative model for affine subspace based visual object tracking. Visual Intelligence 1(1), 4 (2023) 4", + "47. Yang, C., Zhou, H., An, Z., Jiang, X., Xu, Y., Zhang, Q.: Cross-image relational knowledge distillation for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12319-12328 (2022) 3", + "48. Yang, J., Li, C., Zhang, P., Dai, X., Xiao, B., Yuan, L., Gao, J.: Focal attention for long-range interactions in vision transformers. Advances in Neural Information Processing Systems 34, 30008-30022 (2021) 1, 4, 5", + "49. Yang, Z., Li, Z., Jiang, X., Gong, Y., Yuan, Z., Zhao, D., Yuan, C.: Focal and global knowledge distillation for detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4643-4652 (2022) 4, 8, 13, 14", + "50. Yang, Z., Li, Z., Shao, M., Shi, D., Yuan, Z., Yuan, C.: Masked generative distillation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI. pp. 53-69. Springer (2022) 2, 13, 14", + "51. Yao, L., Pi, R., Xu, H., Zhang, W., Li, Z., Zhang, T.: G-detkd: towards general distillation framework for object detectors via contrastive and semantic-guided feature imitation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3591-3600 (2021) 2, 3", + "52. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021) 1" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Distilling Knowledge from Large-Scale Image Models for Object Detection", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Zeng, Z., Yang, F., Liu, H., Satoh, S.: Improving deep metric learning via self-distillation and online batch diffusion process. Visual Intelligence 2(1), 1-13 (2024) 3", + "54. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022) 2, 3, 4, 9, 10, 13", + "55. Zhang, Y., Xiang, T., Hospedales, T.M., Lu, H.: Deep mutual learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4320-4328 (2018) 5", + "56. Zheng, Z., Ye, R., Hou, Q., Ren, D., Wang, P., Zuo, W., Cheng, M.M.: Localization distillation for object detection. arXiv preprint arXiv:2204.05957 (2022) 2, 4, 14", + "57. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) 4, 13", + "58. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training. arXiv preprint arXiv:2211.12860 (2022) 3, 4" + ], + "bbox": [ + 215, + 146, + 787, + 327 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 320, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/a5a39b01-2823-428c-b29a-0b036d9c3243_model.json b/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/a5a39b01-2823-428c-b29a-0b036d9c3243_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9662d09241e51eb3e833c6647aadb880405986e5 --- /dev/null +++ b/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/a5a39b01-2823-428c-b29a-0b036d9c3243_model.json @@ -0,0 +1,2590 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.212, + 0.723, + 0.247 + ], + "angle": 0, + "content": "Gang Li\\(^{1,2\\dagger}\\), Wenhai Wang\\(^{3,2}\\), Xiang Li\\(^{4}\\), Ziheng Li\\(^{5}\\), Jian Yang\\(^{1}\\), Jifeng Dai\\(^{5,2}\\), Yu Qiao\\(^{2}\\), and Shanshan Zhang\\(^{1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.256, + 0.77, + 0.287 + ], + "angle": 0, + "content": "\\(^{1}\\)Nanjing University of Science and Technology \\(^{2}\\)OpenGVLab, Shanghai AI Laboratory \\(^{3}\\)The Chinese University of Hong Kong \\(^{4}\\)Nankai University \\(^{5}\\)Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.322, + 0.744, + 0.628 + ], + "angle": 0, + "content": "Abstract. Large-scale image models have made great progress in recent years, pushing the boundaries of many vision tasks, e.g., object detection. Considering that deploying large models is impractical in many scenes due to expensive computation overhead, this paper presents a new knowledge distillation method, which Distills knowledge from Large-scale Image Models for object detection (dubbed DLIM-Det). To this end, we make the following two efforts: (1) To bridge the gap between the teacher and student, we present a frozen teacher approach. Specifically, to create the teacher model via fine-tuning large models on a specific task, we freeze the pretrained backbone and only optimize the task head. This preserves the generalization capability of large models and gives rise to distinctive characteristics in the teacher. In particular, when equipped with DEtection Transformers (DETRs), the frozen teacher exhibits sparse query locations, thereby facilitating the distillation process. (2) Considering that large-scale detectors are mainly based on DETRs, we propose a Query Distillation (QD) method specifically tailored for DETRs. The QD performs knowledge distillation by leveraging the spatial positions and pair-wise relations of teacher's queries as knowledge to guide the learning of object queries of the student model. Extensive experiments are conducted on various large-scale image models with parameter sizes ranging from 200M to 1B. Our DLIM-Det improves the student with Swin-Tiny by 3.1 mAP when the DINO detector with Swin-Large is used as the teacher. Besides, even when the teacher has 30 times more parameters than the student, DLIM-Det still attains a +2.9 distillation gain." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.64, + 0.741, + 0.655 + ], + "angle": 0, + "content": "Keywords: Object detection \\(\\cdot\\) Knowledge distillation \\(\\cdot\\) Large-scale image model" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.681, + 0.357, + 0.697 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.804 + ], + "angle": 0, + "content": "It is proved that scaling up model sizes is a promising direction to make neural networks reach human performance [33, 42, 48, 52]. Recently, great progress has been achieved in large-scale image models [11, 33, 34, 41], pushing performance boundaries of many vision tasks, including object detection, semantic segmentation, etc. These large-scale models demand for high performance computing device, making them hard to deploy in commonly used hardware platforms. Therefore, Knowledge Distillation (KD) is necessary" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.812, + 0.681, + 0.826 + ], + "angle": 0, + "content": "† This work was done while G. Li was an intern at Shanghai AI Laboratory." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.576, + 0.84 + ], + "angle": 0, + "content": "Corresponding author (shanshan.zhang@njust.edu.cn)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.321, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.536, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.289, + 0.54, + 0.359 + ], + "angle": 0, + "content": "Fig. 1: Comparison between previous KD [3] and our DLIM-Det method for distilling large-sized teacher models. Previous methods struggle to distill knowledge from large models, whereas our method works well. Experiments are conducted on Swin Transformer [34]." + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.159, + 0.785, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.289, + 0.788, + 0.358 + ], + "angle": 0, + "content": "Fig.2: Average Recall (AR) comparison on LVIS labels for the frozen teacher and fully fine-tuned teacher. Higher AR numbers indicate better generalization capacity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.462 + ], + "angle": 0, + "content": "to extract powerful and informative knowledge from the large model (teacher) to boost the compact model (student). The teacher knowledge serves as auxiliary supervisions and helps the student achieve better performance than trained only with ground-truth labels [21, 37]. However, existing KD methods [3, 8, 15, 50, 51] for object detection mainly focus on medium- or small-sized models, leaving the distillation from large-scale models rarely explored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.789, + 0.765 + ], + "angle": 0, + "content": "Distilling knowledge from large-scale image models is not trivial. As observed in previous works [35, 36], the discrepancy between the teacher and student becomes severe as the model size of teacher model increases, which hinders the effectiveness of KD methods. As shown in Fig. 1, when directly applying the prior KD method (e.g., DETRDistill [3]) to a detector (e.g., DINO [54]), the performance even degrades as the teacher model becomes stronger. Specifically, though the Swin-Large [34] has \\(2 \\times\\) parameters than the Swin-Base, the latter is a better teacher and outperforms the former by \\(+0.5\\mathrm{mAP}\\) in terms of KD gains. To address this problem, previous works [35, 39, 56] design a teacher assistant strategy, which generates multiple intermediate-sized teacher candidates to bridge the gap between the teacher and student. However, such a multi-step KD method induces additional training costs and severely affects the efficiency. In this paper, we propose that when obtaining the teacher model through fine-tuning the large-scale image backbone on a specific dataset, freezing the backbone and only optimizing the task head can empower the teacher with richer transferable knowledge compared to fully fine-tuning. Large-scale models exhibit increased generalizability due to their strong capacity and training on extensive datasets, however, there is a risk of overfitting when fine-tuned on specific datasets. Through freezing the pre-trained backbone, we maintain the generalizability of large-scale models, resulting in improved distillation performance. The teacher with pretrained large-scale backbone frozen is denoted as the frozen teacher." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.84 + ], + "angle": 0, + "content": "To quantitatively assess the generalizability of both the frozen and fully fine-tuned (ft.) teachers, we compute the Average Recall (AR) on the LVIS annotations [16], and the results are presented in Fig. 2. Swin-Large [34] based DINO [54] serves as the teacher model. Notably, both teachers are trained only with supervision from the 80 object categories in COCO. The LVIS dataset, while sourcing images from COCO, features" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "a re-annotated and broader spectrum of 1203 categories. As results shown, the frozen teacher outperforms the fully ft. teacher by a margin of \\(4.6\\%\\) in \\(\\mathrm{AR}_{900}\\), indicating that the frozen teacher generates more generalizable features. Strengthening the generalizability of the teacher model endows it with specific characteristics that facilitate the distillation process. In particular, when equipped with detection transformers, sparse query locations emerge in the frozen teacher." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.79, + 0.494 + ], + "angle": 0, + "content": "Advanced large-scale detectors [5, 41, 54, 58] are predominately based on detection transformers, where object queries absorb contextual information for each instance then output predictions. Object queries from the frozen teacher exhibit sparser distribution compared to those from the fully fine-tuned teacher. This sparsity characteristic enables the student model to circumvent unnecessary investments in modeling complex query relations, particularly those leading to excessively compact query locations. To fully exploit sparse query locations from the frozen teacher, we introduce a Query Distillation method. Our Query Distillation aligns object queries between the teacher and student models based on spatial positions and relations. This method consists of two key components: Query Position Distill (QPD) and Query Relation Distill (QRD). Initially, we establish a one-to-one set matching between object queries of two models. Subsequently, the proposed QPD allows the teacher's queries to serve as auxiliary localization supervisions during student model training. Aside from QPD, we further propose Query Relation Distillation to lead the student model to mimic pair-wise relations between teacher's queries. Combining Query Distillation and the frozen teacher can significantly alleviate the learning burden for the student, thereby contributing to a more efficient knowledge distillation process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.79, + 0.678 + ], + "angle": 0, + "content": "To conclude, we propose a new object detection knowledge distillation method, called DLIM-Det, tailored for Large-scale Image Models ranging from 200M to 1B parameters. Leveraging Query Distillation and the frozen teacher approach, DLIM-Det effectively transfers knowledge from large-scale models to compact ones. Extensive experiments on both the Transformer-based (Swin Transformer [34]) and CNN-based (InternImage [41]) models demonstrate the effectiveness of our approach. Results indicate our Query Distillation method works well for detection transformers, surpassing existing KD methods. Additionally, the frozen teacher approach bridges the gap between teacher and student, facilitating efficient distillation from large-scale models. Notably, when the teacher model is Swin-Large, DLIM-Det improves Swin-Tiny based DINO [54] by \\(3.1\\%\\); even for the giant InternImage-Huge model (\\(1B+\\) parameters), it improves the InternImage-Tiny student by \\(2.9\\%\\), achieving a remarkable 56.1 mAP." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.701, + 0.37, + 0.717 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.79, + 0.843 + ], + "angle": 0, + "content": "Knowledge Distillation. As a classical model compression technique, Knowledge Distillation (KD) has been applied to many vision tasks [14, 17, 24, 25, 32, 47, 53]. In this paper, we focus on the KD for object detection. Using the intermediate features of the teacher as knowledge, feature distillation offers deep supervisions for the student [4, 15, 37, 51]. However, feature distillation is usually vulnerable to semantic misalignment between features of two models, and suffers from performance degradation when the discrepancy between two models is too large. Besides, to implement relation-based" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.321, + 0.128 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "distillation, previous works [8, 27, 49] design delicate modules or schemes to capture relations between different instances [8], anchors [27], patches [17], and pixels [49]. At last, the response-based KD methods are also widely investigated in previous works [8, 15, 22, 56], where the probabilistic scores are matched between two models. With the development of detection transformers (DETRs), some recently proposed KD methods [3, 7] are tailored for DETRs. DETRDistill [3] transfers the label assignment results from the teacher to student in order to stabilize the bipartite matching, while D3ETR [7] feeds the teacher queries to the student decoder and then trains the decoder using matching results from the teacher. Although these KD methods achieve decent distillation performance for medium-sized teacher models (such as ResNet-101 [20]), few of them explore the distillation from large-scale image models. In contrast to these works, we focus on distillation from large-scale models with parameters spanning from 200M to 1B. Through our method, we show that larger models lead to greater distillation gains." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.388, + 0.788, + 0.57 + ], + "angle": 0, + "content": "Object Detection. Object detection is a fundamental task in the computer vision field [28, 46]. Detection Transformers (DETRs) achieve great progress recently and gradually dominate object detection. With outstanding scalability, current mainstream large-scale detectors [5, 54, 58] typically rely on DETRs. To address the inefficiency issue of the original DETR [2], Deformable DETR [57] and Adamixer [13] introduce the local attention mechanism to replace the original global attention, resulting in faster model convergence. Some works [31, 43] attempt to insert the spatial prior into the object queries and lead each query to attend to a certain region. DN-DETR [26] designs a query denoising scheme to improve the stability of bipartite graph matching. Based on DN-DETR, DINO [54] presents a contrastive way for denoising training and achieves the state-of-the-art performance. Building on SoTA performance, we employ DINO as the default detector in this paper." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.614, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Large-scale Image Models. Some recent works [10, 33, 34, 41, 42, 48] scale up image models to reach 3 billion parameters, demonstrating a very powerful capacity of large-scale image models. Swin Transformer [33], built on the transformer architecture, incorporates techniques like post-norm and scaled cosine attention to enhance the stability of large vision models. Beit-3 [42], adopting the ViT [10] architecture, explores scaling up models through multi-modal pretraining. InternImage [41], utilizing a CNN base, integrates deformable convolution as a core operator to establish large-scale CNN models. After acquiring large-scale pretrained models, there has been significant interest in efficiently fine-tuning these models for downstream tasks. Notably, approaches such as VPT [23] insert learnable parameters into the input space while keeping most of the pretrained model frozen. Similarly, AdaptFormer [6] introduces lightweight learnable modules into the MLP block. Lin et al. [30] investigate the pretraining tasks that best align with the frozen setting. Different from these works, our paper aims to extract transferable knowledge from large models and use it to enhance tiny-scale models via the KD." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.317, + 0.161 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.178, + 0.788, + 0.223 + ], + "angle": 0, + "content": "In this section, we provide a concise overview of knowledge distillation and the detection transformer in Sec. 3.1. Subsequently, we introduce the frozen teacher method in Sec. 3.2, followed by an explanation of Query Distillation in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.246, + 0.342, + 0.261 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.271, + 0.788, + 0.302 + ], + "angle": 0, + "content": "Knowledge Distillation (KD). Standard KD enforces the student to align its probabilistic scores \\((p^{(s)})\\) with those of the teacher via the Kullback-Leibler (KL) divergence [21, 55]:" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.313, + 0.787, + 0.347 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K D} = p ^ {(t)} \\log \\left(\\frac {p ^ {(t)}}{p ^ {(s)}}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.354, + 0.788, + 0.413 + ], + "angle": 0, + "content": "The teacher's logits \\((p^{(t)})\\) serve as soft label, providing non-target classes not present in the ground truth labels. Despite being weak, this non-target class information significantly enhances the student's generalization. To amplify this effect, prior works typically apply a high temperature to soften the \\(p^{(t)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.415, + 0.788, + 0.505 + ], + "angle": 0, + "content": "Detection Transformers (DETRs). In DETRs, the backbone extracts image features, which are then inputted into the transformer encoder along with positional encodings. The enhanced image features serve as the key and value, providing contextual information for the subsequent transformer decoder. The input for the decoder is N embeddings, known as object queries, denoted as \\( Q = \\{q_0, q_1, \\dots, q_{N-1}\\} \\). Each query will be decoded into class probability \\( p \\) and box coordinate \\( b \\) via the transformer decoder." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.528, + 0.424, + 0.542 + ], + "angle": 0, + "content": "3.2 Frozen Teacher Method" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.553, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Inspired by the NLP domain, where scaling up language models achieves remarkable success [9, 12], some recent works scale up vision models to more than 1 billion parameters and set new records across various vision tasks [33,41,42,48]. It is natural to explore the possibility of transferring the knowledge from these large models to more compact ones using the Knowledge Distillation (KD) technique. However, empirical observations (Fig. 1) suggest that as the model size of the teacher increases, the discrepancy between the teacher and student becomes more severe, limiting the effectiveness of the KD." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.66, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we observed that when fine-tuning large-scale pretrained models on a specific dataset to create the teacher model, freezing the backbone can endow the teacher with richer transferable knowledge compared to fully fine-tuning. The overall framework of our DLIM-Det is illustrated in Fig. 3. The pretraining-then-finetuning paradigm is the cornerstone for modern neural networks [18, 19, 45], especially for large-scale models. Pretraining on massive data is essential for achieving top performance in the downstream tasks [41]. Although fully fine-tuning the entire model may enhance task-specific performance, it also heightens the risk of overfitting and diminishes the model's generalizability. We denote the teacher produced by the frozen setting as the frozen teacher. This frozen teacher maintains the generalizability of the large-scale models. As evidenced by the Average Recall (AR) metrics on LVIS labels shown in Fig. 2, the frozen teacher exhibits an increased ability to recognize a broader array of non-target objects" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.321, + 0.128 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.145, + 0.766, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.311, + 0.788, + 0.354 + ], + "angle": 0, + "content": "Fig. 3: The framework of our DLIM-Det. Freezing the backbone when fine-tuning large-sized models on a specific dataset to create the teacher. The alignment between object queries of the teacher and student is conducted based on positions and relations." + }, + { + "type": "table_caption", + "bbox": [ + 0.331, + 0.37, + 0.673, + 0.384 + ], + "angle": 0, + "content": "Table 1: The query density with different IoU thresholds." + }, + { + "type": "table", + "bbox": [ + 0.368, + 0.384, + 0.632, + 0.435 + ], + "angle": 0, + "content": "
MethodsIoU: 0.5IoU: 0.7
Swin-LIntern-LSwin-LIntern-L
Full finetuning38.133.615.213.0
Frozen21.321.87.58.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.787, + 0.476 + ], + "angle": 0, + "content": "while avoiding excessive specialization on the ground truths. Enhanced generalizability endows the teacher with specific characteristics that facilitate the distillation process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.789, + 0.748 + ], + "angle": 0, + "content": "In the following, we investigate how these distinctive characteristics enhance the efficiency of the distillation process. When equipped with DETRs, we compare the spatial distributions of object queries under both the frozen and fully fine-tuned settings. We observe that the frozen teacher locates object queries more dispersely and sparsely than the fully fine-tuned teacher. To quantify the dispersion of query spatial distribution, we introduce a metric called Query Density. Query density represents the number of object queries with an Intersection over Union (IoU) greater than \\(\\tau\\) surrounding each ground truth. We average the query density over all GT objects in the COCO 2017 validation set [29], with \\(\\tau\\) set to 0.5 and 0.7, respectively. The results are presented in Tab. 1. As results show, the backbone-frozen strategy significantly affects the query density, reducing it by at least \\(1/3\\). The fully fine-tuned teacher exhibits high query density values, and each GT object is surrounded by more than 30 queries with IoUs larger than 0.5. Among these compact queries, only one query should predict the object, while the others should predict no object. One example is visualized in Fig. 5. The detector needs to model their relations, identifying the optimal localized query with high confidence and suppressing the others with low confidence. In the case of sparse query locations, an instance is surrounded by fewer object queries, making their relations much easier to model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.79, + 0.84 + ], + "angle": 0, + "content": "When using the fully fine-tuned teacher, the student model needs to allocate unnecessary capacity to model complex query relations in order to mimic the compact query locations of the teacher. In contrast, the sparse query locations provided by the frozen teacher are easier for the student to learn, reducing the learning burden. To fully leverage the sparse query of the frozen teacher, we design a Query Distillation method, which is described in Sec. 3.3." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.391, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.248, + 0.372, + 0.259 + ], + "angle": 0, + "content": "(a) Comparing four teachers." + }, + { + "type": "image_caption", + "bbox": [ + 0.414, + 0.145, + 0.486, + 0.154 + ], + "angle": 0, + "content": "Input Images" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.155, + 0.486, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.2, + 0.485, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.145, + 0.555, + 0.154 + ], + "angle": 0, + "content": "Full ft. Swin Base" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.155, + 0.559, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.2, + 0.559, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.145, + 0.631, + 0.154 + ], + "angle": 0, + "content": "Frozen Swin Base" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.155, + 0.632, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.2, + 0.632, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.145, + 0.701, + 0.154 + ], + "angle": 0, + "content": "Full ft. Swin Large" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.155, + 0.706, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.2, + 0.706, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.145, + 0.778, + 0.154 + ], + "angle": 0, + "content": "Frozen Swin Large" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.155, + 0.78, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.2, + 0.78, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.247, + 0.694, + 0.258 + ], + "angle": 0, + "content": "(b) Visualization of reconstructed images." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.26, + 0.787, + 0.329 + ], + "angle": 0, + "content": "Fig. 4: Comparison of different teacher models based on mutual information. (a) Multiple teacher models and their mutual information concerning the input \\((I(X;F))\\) and target \\((I(Y;F))\\). The blue lines represent the process of searching for the model whose representation maximizes \\(I(X;F) + I(Y;F)\\). \"Full ft.\" denotes full fine-tuning. (b) The reconstructed image qualitatively reflects \\(I(X;F)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.788, + 0.425 + ], + "angle": 0, + "content": "Taking into account the aspect of generalization and drawing inspiration from prior research [40], which introduced an optimal teacher model selection algorithm based on mutual information, we also substantiate the superiority of the frozen teacher from the perspective of the mutual information principle. Specifically, the entropy of teacher's representation \\((H(F))\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.444, + 0.787, + 0.461 + ], + "angle": 0, + "content": "\\[\nH (F) = I (X; F) + I (Y; F) + Z, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.789, + 0.618 + ], + "angle": 0, + "content": "\\(I(X;F)\\) and \\(I(Y;F)\\) represent task-relevant information regarding the input image \\(X\\) and ground truth target \\(Y\\), respectively. \\(Z\\) refers to nuisance, which is task-irrelevant information. Following standard practice [40, 44], we employ the detection performance (e.g. mAP) to quantify \\(I(Y;F)\\). To measure the mutual information concerning the input \\((I(X;F))\\), we feed the intermediate features into a decoder network (typically, U-Net [38]) to reconstruct the input image \\(X\\). The average reconstruction loss is used to quantify \\(I(X;F)\\). Different values of \\(I(X;F)\\) can be clearly distinguished based on the quality of reconstructed images, as illustrated in Fig. 4b. The theoretical proofs and training loss curves can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.79, + 0.803 + ], + "angle": 0, + "content": "After defining \\( I(X;F) \\) and \\( I(Y;F) \\), we compare base-scale and large-scale models under both full fine-tuning and frozen settings, with results depicted in Fig. 4a, where mutual information \\( (I(X;F) \\) and \\( I(Y;F)) \\) is normalized to ensure the scale matching. Among the four teacher models, the frozen Swin-L attains the maximum of task relevant information \\( I(X;F) + I(Y;F) \\), achieving a delicate trade-off between \\( I(X;F) \\) and \\( I(Y;F) \\). A model with high values of \\( I(X;F) + I(Y;F) \\) can effectively suppress nuisance \\( Z \\) and serves as an informative teacher. Though Base-scale models exhibit higher values of \\( I(X;F) \\) compared to Large-scale ones, they lag in terms of \\( I(Y;F) \\), which may result in inferior KD performance. Empirical experiments (Tab. 3) support our assumption: the frozen Swin-L achieves the best distillation performance among the four teacher models. To conclude, the teacher model selection algorithm [40], grounded in mutual information, further supports our frozen teacher method." + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.825, + 0.654, + 0.84 + ], + "angle": 0, + "content": "1 The features are from the last layer of Transformer Encoder in DETRs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.321, + 0.128 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.383, + 0.162 + ], + "angle": 0, + "content": "3.3 Query Distillation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.264 + ], + "angle": 0, + "content": "Query Distillation aligns the spatial positions and pair-wise relations of queries between the teacher and student, consisting of Query Relation Distill and Query Relation Distill. Initially, we establish the one-to-one matching for object queries of two models, where the cost matrix is computed based on classification and localization costs. Given the cost matrix, Hungarian Matching is used to solve the graph matching, yielding the optimal permutation function \\(\\hat{\\sigma}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.276, + 0.788, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\hat {\\sigma} = \\underset {\\sigma \\in \\mathfrak {S} _ {N}} {\\arg \\min } \\sum_ {i} ^ {N} \\mathcal {L} _ {\\text {m a t c h}} \\left(q _ {i} ^ {(t)}, q _ {\\sigma (i)} ^ {(s)}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.328, + 0.785, + 0.388 + ], + "angle": 0, + "content": "where \\(N\\) equals the number of predicted objects. Among all permutations of \\(N\\) elements \\(\\mathfrak{S}_N\\), \\(\\hat{\\sigma}\\) is the one with the lowest cost. Based on the class probability \\(p\\) and box coordinate \\(b\\) decoded from each object query, we define the cost between the teacher query \\(q^{(t)}\\) and student \\(q^{(s)}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.398, + 0.787, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a t c h}} \\left(q ^ {(t)}, q ^ {(s)}\\right) = \\alpha_ {1} \\mathcal {L} _ {K L} \\left(p ^ {(t)}, p ^ {(s)}\\right) + \\alpha_ {2} \\mathcal {L} _ {L 1 L o s s} \\left(b ^ {(t)}, b ^ {(s)}\\right) + \\alpha_ {3} \\mathcal {L} _ {G I o U} \\left(b ^ {(t)}, b ^ {(s)}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.432, + 0.521, + 0.446 + ], + "angle": 0, + "content": "where \\(\\alpha_{1},\\alpha_{2},\\alpha_{3}\\) are set to 2,5,2 empirically." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Query Position Distill. Instead of matching the probabilistic scores of two models, we observe that the spatial position of object queries can serve as informative knowledge. The position of an object query refers to the bounding box decoded from the query, represented by a vector: \\( b \\in [x, y, w, h] \\), where \\( x \\) and \\( y \\) denote the center coordinates, while \\( w \\) and \\( h \\) represent the width and height, respectively. Given the matching \\( \\sigma \\) between object queries of two models, we formulate the procedure of Query Position Distill as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.561, + 0.788, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {Q P D} = \\sum_ {i} ^ {N} \\beta_ {1} \\mathcal {L} _ {L 1} \\left(b _ {i} ^ {(t)}, b _ {\\sigma (i)} ^ {(s)}\\right) + \\beta_ {2} \\mathcal {L} _ {G I o U} \\left(b _ {i} ^ {(t)}, b _ {\\sigma (i)} ^ {(s)}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.658 + ], + "angle": 0, + "content": "where \\(\\beta_{1},\\beta_{2}\\) are set to 5 and 2, empirically. Through the thorough experiments (listed in Tab. 8b), we empirically find that besides the foreground queries, even poorly localized teacher queries that cover background regions still contain valuable knowledge to distill." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Due to the one-to-one label assignment, only a small subset of object queries receives localization supervision in each image. For example, on the COCO dataset [29], which typically contains an average of 7 instances per image, only 7 out of the 900 queries are involved in the regression loss. This sparsity of localization supervision impairs the training of detection transformers. By the proposed Query Position Distill, the spatial positions of teacher's queries serve as additional supervisions to provide richer information than the sparse ground truth labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Query Relation Distill. In previous works, relation-based KD methods have been widely investigated. These methods aim to capture various relations between different entities such as instances [8, 22], pixels [49], anchors [27] or patches [17]. To model these relations, intricate modules, e.g., GcBlock [1], and complex relation mechanisms, e.g. Rank Mimicking [27], have been designed, which not only lack interpretability but" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.148, + 0.696, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.268, + 0.788, + 0.296 + ], + "angle": 0, + "content": "Fig.5: An example to show the position, score and relation of object queries. Query relations are obtained from self-attention. Only 3 queries are visualized for a clear presentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.396 + ], + "angle": 0, + "content": "also introduce additional training costs. In contrast, the transformer decoder of DETRs explicitly models the pair-wise relations between object queries through the self-attention. We denote by \\( A = [a_{ij}]_{N \\times N} \\) the attention map from the self-attention, where \\( N \\) denotes the number of queries. L1 Loss is employed to narrow the distance between \\( A^{(t)} \\) and \\( A^{(s)} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.397, + 0.787, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {Q R D} = \\frac {1}{N} \\sum_ {i} ^ {N} \\sum_ {j} ^ {N} \\left| A _ {i, j} ^ {(t)} - A _ {\\sigma (i), \\sigma (j)} ^ {(s)} \\right|, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.502 + ], + "angle": 0, + "content": "where \\(\\sigma\\) denotes the permutation function to match the queries between the teacher and student. Query relations are critical for end-to-end detection. As shown in Fig. 5, precise query relations allow the best localized query to pop out with high classification scores, while the others will be suppressed with low scores." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.504, + 0.481, + 0.518 + ], + "angle": 0, + "content": "The overall training loss \\(\\mathcal{L}_{\\text{overall}}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.528, + 0.787, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o v e r a l l}} = \\gamma_ {1} \\mathcal {L} _ {\\text {t a s k}} + \\gamma_ {2} \\mathcal {L} _ {Q P D} + \\gamma_ {3} \\mathcal {L} _ {Q R D}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.554, + 0.524, + 0.569 + ], + "angle": 0, + "content": "where \\(\\gamma_{1},\\gamma_{2},\\gamma_{3}\\) are set to 1, 2, 1, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.645 + ], + "angle": 0, + "content": "We also find that the parameters of the detection head in the teacher model serve as a source of knowledge. Even for heterogeneous DETRs, the basic blocks in their heads are consistent, including attention and feed forward network (FFN). Through empirical investigations, we prove that inheriting the parameters of these basic blocks from the teacher head to student head is convenient and beneficial." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.666, + 0.357, + 0.684 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.696, + 0.41, + 0.712 + ], + "angle": 0, + "content": "4.1 Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.803, + 0.81 + ], + "angle": 0, + "content": "Datasets and Backbone. We conduct the experiments on the MS COCO 2017 dataset [29], which is a commonly used object detection dataset. The dataset is split into the train2017 set with 118k images and the val2017 set with 5k images. All models are trained on the train2017 set, and evaluated on the val2017 set. Results are reported with two backbones: Swin Transformer [34] and InternImage [41]. We employ the COCO-style average precision (mAP) as the metric. DINO [54] is the default detector." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Implementation details. All models, except for the Intern-Huge model, are trained on \\(8 \\times 32\\mathrm{G}\\) V100 GPUs with a batch size of 16. In contrast, the Intern-Huge models are" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.321, + 0.128 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.373, + 0.145, + 0.629, + 0.159 + ], + "angle": 0, + "content": "Table 2: Effectiveness of each component." + }, + { + "type": "table", + "bbox": [ + 0.28, + 0.159, + 0.722, + 0.268 + ], + "angle": 0, + "content": "
MethodsSwin TransformerInternImage
mAP\\( AP_{50} \\)\\( AP_{75} \\)mAP\\( AP_{50} \\)\\( AP_{75} \\)
Teacher: Large Scale58.177.163.556.975.861.9
Student: Tiny Scale51.369.056.053.271.558.0
Query DistillFrozen TeacherInherit Head
✓51.6 (+0.3)69.556.454.2 (+1.0)72.559.2
✓✓52.2 (+0.9)70.157.054.7 (+1.5)72.959.8
✓✓53.7 (+2.4)71.758.855.2 (+2.0)73.760.3
✓✓✓54.4 (+3.1)72.159.455.9 (+2.7)74.161.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.278, + 0.789, + 0.307 + ], + "angle": 0, + "content": "Table 3: Distillation performance comparison between Large- and Base-scale teacher models. Numbers in brackets refer to distillation gains against the original student (Swin-T, 51.3 mAP)." + }, + { + "type": "table", + "bbox": [ + 0.365, + 0.309, + 0.638, + 0.374 + ], + "angle": 0, + "content": "
TeacherFreeze BackboneDistilled Student
mAPAP50AP75
Large-scale✓54.4 (+3.1)72.1 (+3.1)59.4 (+3.4)
Base-scale✓53.5 (+2.2)71.5 (+2.5)58.5 (+2.5)
Base-scale×53.1 (+1.8)71.1 (+2.1)58.0 (+2.0)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.789, + 0.456 + ], + "angle": 0, + "content": "trained on 80GB A100 GPUs. Student models are trained with 12 epochs with AdamW as the optimizer. The initial learning rate is set to 0.0001, which is divided by 10 at the 11th epoch. Following the default practice of DINO [54], the number of object query is set to 900." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.481, + 0.377, + 0.495 + ], + "angle": 0, + "content": "4.2 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.789, + 0.538 + ], + "angle": 0, + "content": "In ablation studies, unless otherwise stated, we adopt the tiny-scale models with around 30M parameters as the student." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.79, + 0.78 + ], + "angle": 0, + "content": "Effects of each component. In Tab. 2, we verify the effectiveness of each component, including the frozen teacher method, Query Distillation, and head inheriting. When facing large-scale teacher models, Query Distillation produces a modest increase of \\(+0.3 / 1.0\\) mAP for Swin/Intern, respectively. Subsequently, we replace the fully fine-tuned teacher with the frozen teacher and perform Query Distillation again. The results demonstrate that the frozen teacher significantly amplifies the distillation gains. Specifically, compared to the fully fine-tuned teacher, the frozen teacher yields an additional increase of \\(+0.6 / 0.5\\%\\) for Swin/Intern models, respectively. Moreover, inheriting the transformer's parameters of the teacher's head produces noticeable improvements, suggesting that the head parameters of the teacher contain some general knowledge that is somewhat independent of the backbone structure. Building on head parameter inheriting, the frozen teacher outperforms the fully fine-tuned teacher by \\(+0.7 / 0.7\\%\\) for Swin/Intern models. These experimental findings validate that the frozen teacher is easier to distill than the previously employed fully fine-tuned teacher. Finally, by combining the three components, DLIM-Det, achieves significant improvements, surpassing the student baseline by \\(+3.1 / 2.7\\%\\) for the Swin/Intern models, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Distillation on medium- and large-sized teacher models. In Tab. 3, we compare the distillation performance under Base- and Large-scale teacher models. With the help of the frozen teacher method, a large-scale teacher surpasses its base-scale counterpart by \\(+0.9\\) mAP. These results affirm the core premise of this paper: larger teacher models lead" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.317, + 0.145, + 0.684, + 0.159 + ], + "angle": 0, + "content": "Table 4: The frozen setting affects the detection performance." + }, + { + "type": "table", + "bbox": [ + 0.361, + 0.161, + 0.642, + 0.24 + ], + "angle": 0, + "content": "
MethodsFreeze BackboneSchedulemAP\\( AP_{50} \\)\\( AP_{75} \\)
Swin-Large×3x58.177.163.5
✓3x56.6 (-1.5)74.961.9
Intern-Large×2x57.876.563.1
✓2x55.8 (-2.0)74.461.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.245, + 0.784, + 0.273 + ], + "angle": 0, + "content": "Table 5: Applying different training schedules to the Frozen and Full ft. settings as to obtain teachers with identical performance. The KD contains Query Distillation and the inheriting head." + }, + { + "type": "table", + "bbox": [ + 0.336, + 0.279, + 0.663, + 0.4 + ], + "angle": 0, + "content": "
MethodsFreeze BackboneSchedulemAP\\( AP_{50} \\)\\( AP_{75} \\)
Teacher:
T1: Swin-Large×1x56.975.662.2
T2: Swin-Large✓3x56.774.961.9
Student:
Swin-Tiny-1x51.369.056.0
KD from T1-1x53.1 (+1.8)70.957.9
KD from T2-1x54.4 (+3.1)72.159.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.411, + 0.787, + 0.437 + ], + "angle": 0, + "content": "Table 6: Effectiveness of Query Position Distill (QPD) and Query Relation Distill (QRD). \"T/S\" denotes the Teacher/Student." + }, + { + "type": "table", + "bbox": [ + 0.335, + 0.438, + 0.663, + 0.524 + ], + "angle": 0, + "content": "
MethodsInternImageSwin Transformer
mAPAP50AP75mAPAP50AP75
T: Base54.673.659.355.874.460.7
S: Tiny53.271.558.051.369.056.0
+ QPD54.2 (+1.0)72.459.151.6 (+0.3)69.556.2
+ QRD54.2 (+1.0)72.959.151.8 (+0.5)70.256.5
+ QPD & QRD54.5 (+1.3)73.259.552.1 (+0.8)69.956.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.785, + 0.581 + ], + "angle": 0, + "content": "to greater distillation gains. Additionally, the frozen teacher method generalizes well to base-scale models. When using the base-scale model as a teacher, the frozen teacher surpasses the fully fine-tuned teacher by \\(0.4\\mathrm{mAP}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.764 + ], + "angle": 0, + "content": "Fair comparison between the fully fine-tuned teacher and frozen teacher. Freezing the backbone typically results in performance degradation to some extent compared to fully fine-tuning. As shown in Tab. 4, under the same training schedule, the frozen teacher lags behind the fully fine-tuned teacher by around \\(2\\mathrm{mAP}\\). One might argue that, owing to lower detection performance, the frozen teacher narrows the gap between the teacher and student, thereby attaining better distillation results. In Tab. 5, we conduct experiments to challenge this perspective. Concretely, we assign long and short training schedules to the frozen and fully fine-tuned teachers, resulting in teachers with nearly identical performance. By ensuring the same gap between the teacher and student, we find that the frozen teacher surpasses the fully fine-tuned teacher by a large margin, +1.3 mAP. These results highlight the intrinsic advantages of the frozen teacher approach, such as sparse query locations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effectiveness of the Query Distillation. The Query Distillation is specifically designed for detection transformers, including Query Position Distill (QPD) and Query Relation Distill (QRD). We ablate each of them step by step in Tab. 6. Here, we employ the medium-sized models as the teacher, to exclude the effects of the large gap between the teacher and student. Besides the experiments on the advanced network architecture, such" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.321, + 0.128 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.172 + ], + "angle": 0, + "content": "Table 7: Applying Query Position Distill (QPD) and Query Relation Distill (QRD) to classical backbones (e.g. ResNet [20])." + }, + { + "type": "table", + "bbox": [ + 0.398, + 0.173, + 0.601, + 0.256 + ], + "angle": 0, + "content": "
MethodsResnet
mAP\\( AP_{50} \\)\\( AP_{75} \\)
T: ResNet10150.067.754.4
S: ResNet5048.966.453.4
+ QPD49.5 (+0.5)67.253.9
+ QRD49.5 (+0.5)67.554.1
+ QPD & QRD49.8 (+0.9)67.654.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.299, + 0.267, + 0.7, + 0.28 + ], + "angle": 0, + "content": "Table 8: Study on distillation positions and teacher query selection." + }, + { + "type": "table_caption", + "bbox": [ + 0.251, + 0.281, + 0.495, + 0.302 + ], + "angle": 0, + "content": "(a) Study on distillation positions. Teacher: InternBase (54.6 mAP); Student: Interm-Tiny (53.2 mAP)" + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.301, + 0.465, + 0.391 + ], + "angle": 0, + "content": "
KD on decoder layermAP\\( AP_{50} \\)\\( AP_{75} \\)
123456
✓✓✓54.172.659.0
✓✓✓53.872.358.8
✓✓✓✓✓✓54.573.259.5
✓54.172.659.0
✓53.572.158.3
✓52.470.957.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.507, + 0.281, + 0.747, + 0.302 + ], + "angle": 0, + "content": "(b) Study on teacher query selection. FG refers to foreground; BG refers to background." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.303, + 0.721, + 0.391 + ], + "angle": 0, + "content": "
MethodsmAP\\( AP_{50} \\)\\( AP_{75} \\)
T: Intern-Base54.673.659.3
S: Intern-Tiny53.271.558.0
No Sample (900)54.272.659.1
Random Sample (300)54.573.259.5
FG queries (IoU > 0.3)54.272.559.2
BG queries (IoU < 0.3)54.272.859.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.43 + ], + "angle": 0, + "content": "Table 9: Applying our DLIM-Det to the huge model (InternImage-Huge) with 1B parameters to verify the scalability. Numbers in brackets refer to distillation gains against the original student." + }, + { + "type": "table", + "bbox": [ + 0.358, + 0.431, + 0.64, + 0.481 + ], + "angle": 0, + "content": "
Teacher#ParamsDistilled Student
mAPAP50AP75
Intern-Huge1.08B56.1 (+2.9)74.4 (+2.9)61.5 (+3.5)
Intern-Large223M55.9 (+2.7)74.1 (+2.6)61.4 (+3.4)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.789, + 0.564 + ], + "angle": 0, + "content": "as Swin Transformer and InternImage (Tab. 6), we also report results on ResNet [20] in Tab. 7. Our Query Distillation method yields consistent improvements on three different backbones, at least \\(+0.8\\) mAP. Furthermore, combining the QPD and QRD together achieves better performance than using them individually, verifying the effectiveness of each component." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.789, + 0.672 + ], + "angle": 0, + "content": "Study on the distillation positions. To investigate the impact of Query Distillation on different decoder layers in DETRs, we conducted experiments, and the results are presented in Tab. 8a. When applying Query Distillation to all 6 decoder layers, we observe the best performance, surpassing the counterparts where it was applied to the first three or last three decoder layers. Furthermore, we examine the performance when applying Query Distillation to only one decoder layer. The results show that distillation on shallow decoder layers outperforms distillation on deeper layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Selection of teacher queries. Using our Query Distillation method, we further conduct experiments to investigate the impact of different subsets of teacher queries on knowledge transfer. Results are reported in Tab. 8b. First, we perform KD on all teacher queries, obtaining \\(+1.0\\%\\) gains. Then we randomly sample 300 queries and achieve further improvements, surpassing the 900 query baseline by \\(+0.3\\%\\). To explore alternative query selection strategies, we define the queries whose IoUs with GT are greater than 0.3 as foreground queries, otherwise as background queries. Distilling knowledge from either pure foreground (namely IoU \\(>0.3\\) in Tab. 8b) or pure background queries results in inferior performance compared to random sampling. This suggests that both foreground and background queries contain valuable knowledge, and distilling knowledge from a combination of them yields optimal results." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 10: Experiments on heterogeneous teacher-student pairs. Heterogeneous components are underlined. Numbers in brackets denote distillation gains against the original student." + }, + { + "type": "table", + "bbox": [ + 0.307, + 0.175, + 0.692, + 0.228 + ], + "angle": 0, + "content": "
TeacherStudentDistilled Student
mAPAP50AP75
DINO-Swin-LDeformable DETR-Swin-T49.6 (+3.5)68.3 (+3.4)53.7 (+3.9)
DINO-Intern-HDINO-Swin-T53.3 (+2.0)71.3 (+2.3)58.3 (+2.3)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.276, + 0.235, + 0.723, + 0.248 + ], + "angle": 0, + "content": "Table 11: Comparison with state-of-the-art knowledge distillation methods." + }, + { + "type": "table", + "bbox": [ + 0.273, + 0.249, + 0.726, + 0.359 + ], + "angle": 0, + "content": "
MethodsSwin TransformerInternImage
mAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)mAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
Teacher: Large58.177.163.541.262.173.957.876.563.141.661.373.0
Student: Tiny51.369.056.034.554.466.153.271.558.035.556.669.3
FGD [49]51.1 (-0.2)69.255.834.353.865.652.8 (-0.4)71.157.636.456.466.8
MGD [50]51.1 (-0.2)69.855.834.954.266.154.2 (+1.0)72.859.138.257.669.4
FitNet [37]51.3 (+0.0)70.055.934.754.566.353.8 (+0.5)72.458.738.257.468.2
DETRDistill [3]51.7 (+0.4)69.456.435.054.866.054.7 (+1.5)72.859.838.558.269.4
Our DLIM-Det54.4 (+3.1)72.159.438.157.469.755.9 (+2.7)74.161.440.359.470.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.787, + 0.46 + ], + "angle": 0, + "content": "Scalability to huge teacher models. We apply DLIM-Det to InternImage-Huge with 1.08 billion parameters. From Tab. 9, we observe that DLIM-Det achieves satisfactory performance even when the teacher has more than 30 times the parameters of the student. Here, the student is represented by Intern-Tiny with 30M parameters, achieving a mAP of 53.2. Specifically, Intern-Huge produces a \\(+2.9\\) distillation gain on Intern-Tiny based student and surpasses Intern-Large by 0.2 mAP (56.1 v.s. 55.9)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Distillation on heterogeneous architectures. To verify the generalization of DLIM-Det, we conduct experiments on heterogeneous teacher and student models, including heterogeneous detectors and heterogeneous backbones. For heterogeneous detectors, with DINO [54] as the teacher and Deformable DETR [57] as the student, DLIM-Det improves the student from 46.1 to \\(49.6\\mathrm{mAP}\\), as reported in Tab. 10. In the case of heterogeneous backbones, where Intern-Huge/Swin-Tiny are employed as the teacher/student, DLIM-Det obtains a considerable improvement of \\(+2.0\\mathrm{mAP}\\). These results show the strong generalization of DLIM-Det." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.604, + 0.555, + 0.619 + ], + "angle": 0, + "content": "4.3 Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Main results on two backbones. In Tab. 11, we compare our DLIM-Det with several popular KD methods. FitNet [37] is a classical feature distillation method, which introduces adaptation layers to bridge the semantic gap between the teacher and student. However, FitNet fails to improve the performance of the transformer-based backbone (Swin Transformer), showing no improvements. MGD [50] is inspired by the masked image modeling [18] and randomly masks pixels of the student's features. It then leads the student to generate the teacher's features in the masked regions. MGD obtains a noticeable improvement of \\(+1.0\\%\\) for InternImage, but struggles to improve the Swin Transformer. FGD [49] leverages the GcBlock [1] to capture the relations between pixels from feature maps. However, FGD does not perform well on both backbones, indicating that the hand-crafted relation mechanism may conflict with the intrinsic relation modeling of detection transformers. In fact, FGD even hurts the performance of InternImage by \\(-0.4\\%\\). The recently proposed DETRDistill [3] is designed for distilling DETRs. In addition to performing customized logits and feature distillation, it also distills the label" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.321, + 0.128 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "assignment results of teacher. Notably, our DLIM-Det outperforms it by a large margin, at least \\(+1.2\\%\\). We explain that the Query prior Assignment Distillation in DETRDistill serves a role akin to the De-Noising group in DINO, thereby it only yields minor improvements in our experiments. Moreover, another contribution of DETRDistill [3] is a feature imitation method, which is vulnerable to significant discrepancy between the teacher and student, as confirmed by other works [37,49,50]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.299 + ], + "angle": 0, + "content": "These experimental findings suggest that while previous works can achieve decent distillation gains for medium-sized teacher, they encounter challenges in distilling large-scale teacher models. In contrast, DLIM-Det demonstrates significant distillation performance specifically tailored for large-scale models." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.336 + ], + "angle": 0, + "content": "Table 12: Comparison with the Teacher Assistant (TA) on two backbones. The TA consists of two steps, thus the overall training time is the sum of the training time for both steps." + }, + { + "type": "table", + "bbox": [ + 0.272, + 0.338, + 0.728, + 0.453 + ], + "angle": 0, + "content": "
MethodsSwin TransformerInternImage
mAP\\( AP_{50} \\)\\( AP_{75} \\)Train TimemAP\\( AP_{50} \\)\\( AP_{75} \\)Train Time
Baseline:
Teacher: Large58.177.163.5-57.876.563.1-
Intermediate Teacher: Base55.874.460.7-54.673.659.3-
Student: Tiny51.369.056.0-53.271.558.0-
TA step1 (Large→Base)57.5 (+1.7)76.162.827.3h56.8 (+2.2)75.262.134.5h
TA setup2 (Base→Tiny)53.7 (+2.4)71.758.919.2h55.7 (+2.5)74.060.723.6h
Our DLIM-Det (Large→Tiny)54.4 (+3.1)72.159.420h55.9 (+2.7)74.161.424.8h
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.601 + ], + "angle": 0, + "content": "Comparison with the Teacher Assistant. The Teacher Assistant (TA) [35, 39, 56] is proved effective in reducing the gap between the teacher and student models. Following the previous work [35], we use the base-scale model with approximately 90M parameters as the TA model. The TA model is distilled from the teacher model and then serves as the teacher for the student model. The results in Tab. 12 demonstrate that our DLIM-Det, which directly distills knowledge from large-scale models, achieves superior performance and higher efficiency compared to the TA method. Specifically, by using only half of the training time, DLIM-Det outperforms Teacher Assistant by \\(0.7 / 0.2\\) mAP on Swin/Intern models, respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.625, + 0.345, + 0.642 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.794 + ], + "angle": 0, + "content": "In this paper, we introduce DLIM-Det, a novel knowledge distillation method specifically tailored for large-scale image models in the context of object detection. To bridge the gap between the teacher and student, we propose the frozen teacher method, which involves freezing the pretrained backbone during teacher creation. The frozen teacher preserves the generalization capacity of large-scale models and gives rise to distinctive characteristics, such as sparse query locations. Through the proposed Query Distillation method, we harness the full advantages offered by the frozen teacher, thereby facilitating the distillation process. Extensive experiments provide robust evidence that larger teacher models can indeed lead to greater distillation gains through DLIM-Det." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Acknowledgement. This work is partially supported by National Natural Science Foundation of China (No. 62322602), Natural Science Foundation of Jiangsu Province, China (No. BK20230033), and National Natural Science Foundation of China (No. 62172225)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.312, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.221 + ], + "angle": 0, + "content": "1. Cao, Y., Xu, J., Lin, S., Wei, F., Hu, H.: Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In: Proceedings of the IEEE/CVF international conference on computer vision workshops. pp. 0-0 (2019) 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.787, + 0.277 + ], + "angle": 0, + "content": "2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.786, + 0.319 + ], + "angle": 0, + "content": "3. Chang, J., Wang, S., Xu, H.M., Chen, Z., Yang, C., Zhao, F.: Detrdistill: A universal knowledge distillation framework for detr-families. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6898-6908 (2023) 2, 4, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.786, + 0.361 + ], + "angle": 0, + "content": "4. Chen, G., Choi, W., Yu, X., Han, T., Chandraker, M.: Learning efficient object detection models with knowledge distillation. Advances in neural information processing systems 30 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.786, + 0.39 + ], + "angle": 0, + "content": "5. Chen, Q., Chen, X., Zeng, G., Wang, J.: Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085 (2022) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.786, + 0.432 + ], + "angle": 0, + "content": "6. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.786, + 0.459 + ], + "angle": 0, + "content": "7. Chen, X., Chen, J., Liu, Y., Zeng, G.: DG etr: Decoder distillation for detection transformer arXiv preprint arXiv:2211.09768 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.786, + 0.503 + ], + "angle": 0, + "content": "8. Dai, X., Jiang, Z., Wu, Z., Bao, Y., Wang, Z., Liu, S., Zhou, E.: General instance distillation for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7842-7851 (2021) 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.786, + 0.531 + ], + "angle": 0, + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.532, + 0.787, + 0.573 + ], + "angle": 0, + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.574, + 0.787, + 0.614 + ], + "angle": 0, + "content": "11. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. arXiv preprint arXiv:2211.07636 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.616, + 0.787, + 0.643 + ], + "angle": 0, + "content": "12. Floridi, L., Chiriatti, M.: Gpt-3: Its nature, scope, limits, and consequences. Minds and Machines 30, 681–694 (2020) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.645, + 0.787, + 0.685 + ], + "angle": 0, + "content": "13. Gao, Z., Wang, L., Han, B., Guo, S.: Adamixer: A fast-converging query-based object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5364-5373 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.687, + 0.787, + 0.713 + ], + "angle": 0, + "content": "14. Guo, G., Han, L., Wang, L., Zhang, D., Han, J.: Semantic-aware knowledge distillation with parameter-free feature uniformization. Visual Intelligence 1(1), 6 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.715, + 0.787, + 0.755 + ], + "angle": 0, + "content": "15. Guo, J., Han, K., Wang, Y., Wu, H., Chen, X., Xu, C., Xu, C.: Distilling object detectors via decoupled features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2154-2164 (2021) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.757, + 0.787, + 0.797 + ], + "angle": 0, + "content": "16. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "17. Hao, Z., Guo, J., Jia, D., Han, K., Tang, Y., Zhang, C., Hu, H., Wang, Y.: Learning efficient vision transformers via fine-grained manifold distillation. In: Advances in Neural Information Processing Systems (2021) 3, 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.179, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.219, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.32, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "18. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022) 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "19. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.271 + ], + "angle": 0, + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.299 + ], + "angle": 0, + "content": "21. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.327 + ], + "angle": 0, + "content": "22. Huang, T., You, S., Wang, F., Qian, C., Xu, C.: Knowledge distillation from a stronger teacher. arXiv preprint arXiv:2205.10536 (2022) 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.367 + ], + "angle": 0, + "content": "23. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.369, + 0.787, + 0.396 + ], + "angle": 0, + "content": "24. Jia, Z., Sun, S., Liu, G., Liu, B.: Mssd: multi-scale self-distillation for object detection. Visual Intelligence 2(1), 8 (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.437 + ], + "angle": 0, + "content": "25. Kang, Z., Zhang, P., Zhang, X., Sun, J., Zheng, N.: Instance-conditional knowledge distillation for object detection. Advances in Neural Information Processing Systems 34, 16468-16480 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.438, + 0.787, + 0.481 + ], + "angle": 0, + "content": "26. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "27. Li, G., Li, X., Wang, Y., Zhang, S., Wu, Y., Liang, D.: Knowledge distillation for object detection via rank mimicking and prediction-guided feature imitation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1306-1313 (2022) 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.55 + ], + "angle": 0, + "content": "28. Li, G., Li, X., Zhang, S., Yang, J.: Towards more reliable evaluation in pedestrian detection by rethinking \"ignore regions\". Visual Intelligence 2(1), 4 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "29. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 6, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "30. Lin, Y., Liu, Z., Zhang, Z., Hu, H., Zheng, N., Lin, S., Cao, Y.: Could giant pre-trained image models extract universal representations? Advances in Neural Information Processing Systems 35, 8332-8346 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.674 + ], + "angle": 0, + "content": "31. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "32. Liu, Y., Chen, K., Liu, C., Qin, Z., Luo, Z., Wang, J.: Structured knowledge distillation for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2604-2613 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "33. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12009-12019 (2022) 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "34. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) 1, 2, 3, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "35. Mirzadeh, S.I., Farajtabar, M., Li, A., Levine, N., Matsukawa, A., Ghasemzadeh, H.: Improved knowledge distillation via teacher assistant. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 5191-5198 (2020) 2, 14" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.188 + ], + "angle": 0, + "content": "36. Park, D.Y., Cha, M.H., Kim, D., Han, B., et al.: Learning student-friendly teacher networks for knowledge distillation. Advances in neural information processing systems 34, 13292-13303 (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.217 + ], + "angle": 0, + "content": "37. Romero, A., Ballas, N., Kahou, S.E., Chassang, A., Gatta, C., Bengio, Y.: Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550 (2014) 2, 3, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.788, + 0.272 + ], + "angle": 0, + "content": "38. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.788, + 0.313 + ], + "angle": 0, + "content": "39. Son, W., Na, J., Choi, J., Hwang, W.: Densely guided knowledge distillation using multiple teacher assistants. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9395-9404 (2021) 2, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.788, + 0.354 + ], + "angle": 0, + "content": "40. Wang, C., Yang, Q., Huang, R., Song, S., Huang, G.: Efficient knowledge distillation from model checkpoints. Advances in Neural Information Processing Systems 35, 607-619 (2022) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.355, + 0.788, + 0.397 + ], + "angle": 0, + "content": "41. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., et al.: Interimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778 (2022) 1, 3, 4, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.788, + 0.438 + ], + "angle": 0, + "content": "42. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442 (2022) 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.788, + 0.479 + ], + "angle": 0, + "content": "43. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2567-2575 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.788, + 0.508 + ], + "angle": 0, + "content": "44. Wang, Y., Ni, Z., Song, S., Yang, L., Huang, G.: Revisiting locally supervised learning: an alternative to end-to-end training. arXiv preprint arXiv:2101.10832 (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.788, + 0.548 + ], + "angle": 0, + "content": "45. Wei, F., Gao, Y., Wu, Z., Hu, H., Lin, S.: Aligning pretraining for detection via object-level contrastive learning. Advances in Neural Information Processing Systems 34, 22682-22694 (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.788, + 0.577 + ], + "angle": 0, + "content": "46. Xu, T., Zhu, X.F., Wu, X.J.: Learning spatio-temporal discriminative model for affine subspace based visual object tracking. Visual Intelligence 1(1), 4 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.788, + 0.618 + ], + "angle": 0, + "content": "47. Yang, C., Zhou, H., An, Z., Jiang, X., Xu, Y., Zhang, Q.: Cross-image relational knowledge distillation for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12319-12328 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.788, + 0.659 + ], + "angle": 0, + "content": "48. Yang, J., Li, C., Zhang, P., Dai, X., Xiao, B., Yuan, L., Gao, J.: Focal attention for long-range interactions in vision transformers. Advances in Neural Information Processing Systems 34, 30008-30022 (2021) 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.788, + 0.702 + ], + "angle": 0, + "content": "49. Yang, Z., Li, Z., Jiang, X., Gong, Y., Yuan, Z., Zhao, D., Yuan, C.: Focal and global knowledge distillation for detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4643-4652 (2022) 4, 8, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.788, + 0.743 + ], + "angle": 0, + "content": "50. Yang, Z., Li, Z., Shao, M., Shi, D., Yuan, Z., Yuan, C.: Masked generative distillation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI. pp. 53-69. Springer (2022) 2, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.788, + 0.797 + ], + "angle": 0, + "content": "51. Yao, L., Pi, R., Xu, H., Zhang, W., Li, Z., Zhang, T.: G-detkd: towards general distillation framework for object detectors via contrastive and semantic-guided feature imitation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3591-3600 (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "52. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021) 1" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.321, + 0.128 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "53. Zeng, Z., Yang, F., Liu, H., Satoh, S.: Improving deep metric learning via self-distillation and online batch diffusion process. Visual Intelligence 2(1), 1-13 (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.788, + 0.217 + ], + "angle": 0, + "content": "54. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022) 2, 3, 4, 9, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "55. Zhang, Y., Xiang, T., Hospedales, T.M., Lu, H.: Deep mutual learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4320-4328 (2018) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.273 + ], + "angle": 0, + "content": "56. Zheng, Z., Ye, R., Hou, Q., Ren, D., Wang, P., Zuo, W., Cheng, M.M.: Localization distillation for object detection. arXiv preprint arXiv:2204.05957 (2022) 2, 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "57. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "58. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training. arXiv preprint arXiv:2211.12860 (2022) 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/a5a39b01-2823-428c-b29a-0b036d9c3243_origin.pdf b/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/a5a39b01-2823-428c-b29a-0b036d9c3243_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5bd78a1bd4c99c00e69232d51761df967ff3de79 --- /dev/null +++ b/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/a5a39b01-2823-428c-b29a-0b036d9c3243_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9184556201c1c0148793444903c78daf2ff1bd9947feb8a474ef3fb0740fd347 +size 1031692 diff --git a/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/full.md b/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..957468c1aa29605efc3ff655f65cf57c64adbf34 --- /dev/null +++ b/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/full.md @@ -0,0 +1,328 @@ +# Distilling Knowledge from Large-Scale Image Models for Object Detection + +Gang Li $^{1,2\dagger}$ , Wenhai Wang $^{3,2}$ , Xiang Li $^{4}$ , Ziheng Li $^{5}$ , Jian Yang $^{1}$ , Jifeng Dai $^{5,2}$ , Yu Qiao $^{2}$ , and Shanshan Zhang $^{1\boxtimes}$ + +$^{1}$ Nanjing University of Science and Technology $^{2}$ OpenGVLab, Shanghai AI Laboratory $^{3}$ The Chinese University of Hong Kong $^{4}$ Nankai University $^{5}$ Tsinghua University + +Abstract. Large-scale image models have made great progress in recent years, pushing the boundaries of many vision tasks, e.g., object detection. Considering that deploying large models is impractical in many scenes due to expensive computation overhead, this paper presents a new knowledge distillation method, which Distills knowledge from Large-scale Image Models for object detection (dubbed DLIM-Det). To this end, we make the following two efforts: (1) To bridge the gap between the teacher and student, we present a frozen teacher approach. Specifically, to create the teacher model via fine-tuning large models on a specific task, we freeze the pretrained backbone and only optimize the task head. This preserves the generalization capability of large models and gives rise to distinctive characteristics in the teacher. In particular, when equipped with DEtection Transformers (DETRs), the frozen teacher exhibits sparse query locations, thereby facilitating the distillation process. (2) Considering that large-scale detectors are mainly based on DETRs, we propose a Query Distillation (QD) method specifically tailored for DETRs. The QD performs knowledge distillation by leveraging the spatial positions and pair-wise relations of teacher's queries as knowledge to guide the learning of object queries of the student model. Extensive experiments are conducted on various large-scale image models with parameter sizes ranging from 200M to 1B. Our DLIM-Det improves the student with Swin-Tiny by 3.1 mAP when the DINO detector with Swin-Large is used as the teacher. Besides, even when the teacher has 30 times more parameters than the student, DLIM-Det still attains a +2.9 distillation gain. + +Keywords: Object detection $\cdot$ Knowledge distillation $\cdot$ Large-scale image model + +# 1 Introduction + +It is proved that scaling up model sizes is a promising direction to make neural networks reach human performance [33, 42, 48, 52]. Recently, great progress has been achieved in large-scale image models [11, 33, 34, 41], pushing performance boundaries of many vision tasks, including object detection, semantic segmentation, etc. These large-scale models demand for high performance computing device, making them hard to deploy in commonly used hardware platforms. Therefore, Knowledge Distillation (KD) is necessary + +![](images/832a8a9182abe72865c99a7f18a8a47ba47adcdce852c5751193373071b36ae2.jpg) +Fig. 1: Comparison between previous KD [3] and our DLIM-Det method for distilling large-sized teacher models. Previous methods struggle to distill knowledge from large models, whereas our method works well. Experiments are conducted on Swin Transformer [34]. + +![](images/5bd748ae7395043a83f110af21c45286acdea2da855bc2bd07a5972dbf89c429.jpg) +Fig.2: Average Recall (AR) comparison on LVIS labels for the frozen teacher and fully fine-tuned teacher. Higher AR numbers indicate better generalization capacity. + +to extract powerful and informative knowledge from the large model (teacher) to boost the compact model (student). The teacher knowledge serves as auxiliary supervisions and helps the student achieve better performance than trained only with ground-truth labels [21, 37]. However, existing KD methods [3, 8, 15, 50, 51] for object detection mainly focus on medium- or small-sized models, leaving the distillation from large-scale models rarely explored. + +Distilling knowledge from large-scale image models is not trivial. As observed in previous works [35, 36], the discrepancy between the teacher and student becomes severe as the model size of teacher model increases, which hinders the effectiveness of KD methods. As shown in Fig. 1, when directly applying the prior KD method (e.g., DETRDistill [3]) to a detector (e.g., DINO [54]), the performance even degrades as the teacher model becomes stronger. Specifically, though the Swin-Large [34] has $2 \times$ parameters than the Swin-Base, the latter is a better teacher and outperforms the former by $+0.5\mathrm{mAP}$ in terms of KD gains. To address this problem, previous works [35, 39, 56] design a teacher assistant strategy, which generates multiple intermediate-sized teacher candidates to bridge the gap between the teacher and student. However, such a multi-step KD method induces additional training costs and severely affects the efficiency. In this paper, we propose that when obtaining the teacher model through fine-tuning the large-scale image backbone on a specific dataset, freezing the backbone and only optimizing the task head can empower the teacher with richer transferable knowledge compared to fully fine-tuning. Large-scale models exhibit increased generalizability due to their strong capacity and training on extensive datasets, however, there is a risk of overfitting when fine-tuned on specific datasets. Through freezing the pre-trained backbone, we maintain the generalizability of large-scale models, resulting in improved distillation performance. The teacher with pretrained large-scale backbone frozen is denoted as the frozen teacher. + +To quantitatively assess the generalizability of both the frozen and fully fine-tuned (ft.) teachers, we compute the Average Recall (AR) on the LVIS annotations [16], and the results are presented in Fig. 2. Swin-Large [34] based DINO [54] serves as the teacher model. Notably, both teachers are trained only with supervision from the 80 object categories in COCO. The LVIS dataset, while sourcing images from COCO, features + +a re-annotated and broader spectrum of 1203 categories. As results shown, the frozen teacher outperforms the fully ft. teacher by a margin of $4.6\%$ in $\mathrm{AR}_{900}$ , indicating that the frozen teacher generates more generalizable features. Strengthening the generalizability of the teacher model endows it with specific characteristics that facilitate the distillation process. In particular, when equipped with detection transformers, sparse query locations emerge in the frozen teacher. + +Advanced large-scale detectors [5, 41, 54, 58] are predominately based on detection transformers, where object queries absorb contextual information for each instance then output predictions. Object queries from the frozen teacher exhibit sparser distribution compared to those from the fully fine-tuned teacher. This sparsity characteristic enables the student model to circumvent unnecessary investments in modeling complex query relations, particularly those leading to excessively compact query locations. To fully exploit sparse query locations from the frozen teacher, we introduce a Query Distillation method. Our Query Distillation aligns object queries between the teacher and student models based on spatial positions and relations. This method consists of two key components: Query Position Distill (QPD) and Query Relation Distill (QRD). Initially, we establish a one-to-one set matching between object queries of two models. Subsequently, the proposed QPD allows the teacher's queries to serve as auxiliary localization supervisions during student model training. Aside from QPD, we further propose Query Relation Distillation to lead the student model to mimic pair-wise relations between teacher's queries. Combining Query Distillation and the frozen teacher can significantly alleviate the learning burden for the student, thereby contributing to a more efficient knowledge distillation process. + +To conclude, we propose a new object detection knowledge distillation method, called DLIM-Det, tailored for Large-scale Image Models ranging from 200M to 1B parameters. Leveraging Query Distillation and the frozen teacher approach, DLIM-Det effectively transfers knowledge from large-scale models to compact ones. Extensive experiments on both the Transformer-based (Swin Transformer [34]) and CNN-based (InternImage [41]) models demonstrate the effectiveness of our approach. Results indicate our Query Distillation method works well for detection transformers, surpassing existing KD methods. Additionally, the frozen teacher approach bridges the gap between teacher and student, facilitating efficient distillation from large-scale models. Notably, when the teacher model is Swin-Large, DLIM-Det improves Swin-Tiny based DINO [54] by $3.1\%$ ; even for the giant InternImage-Huge model ( $1B+$ parameters), it improves the InternImage-Tiny student by $2.9\%$ , achieving a remarkable 56.1 mAP. + +# 2 Related Work + +Knowledge Distillation. As a classical model compression technique, Knowledge Distillation (KD) has been applied to many vision tasks [14, 17, 24, 25, 32, 47, 53]. In this paper, we focus on the KD for object detection. Using the intermediate features of the teacher as knowledge, feature distillation offers deep supervisions for the student [4, 15, 37, 51]. However, feature distillation is usually vulnerable to semantic misalignment between features of two models, and suffers from performance degradation when the discrepancy between two models is too large. Besides, to implement relation-based + +distillation, previous works [8, 27, 49] design delicate modules or schemes to capture relations between different instances [8], anchors [27], patches [17], and pixels [49]. At last, the response-based KD methods are also widely investigated in previous works [8, 15, 22, 56], where the probabilistic scores are matched between two models. With the development of detection transformers (DETRs), some recently proposed KD methods [3, 7] are tailored for DETRs. DETRDistill [3] transfers the label assignment results from the teacher to student in order to stabilize the bipartite matching, while D3ETR [7] feeds the teacher queries to the student decoder and then trains the decoder using matching results from the teacher. Although these KD methods achieve decent distillation performance for medium-sized teacher models (such as ResNet-101 [20]), few of them explore the distillation from large-scale image models. In contrast to these works, we focus on distillation from large-scale models with parameters spanning from 200M to 1B. Through our method, we show that larger models lead to greater distillation gains. + +Object Detection. Object detection is a fundamental task in the computer vision field [28, 46]. Detection Transformers (DETRs) achieve great progress recently and gradually dominate object detection. With outstanding scalability, current mainstream large-scale detectors [5, 54, 58] typically rely on DETRs. To address the inefficiency issue of the original DETR [2], Deformable DETR [57] and Adamixer [13] introduce the local attention mechanism to replace the original global attention, resulting in faster model convergence. Some works [31, 43] attempt to insert the spatial prior into the object queries and lead each query to attend to a certain region. DN-DETR [26] designs a query denoising scheme to improve the stability of bipartite graph matching. Based on DN-DETR, DINO [54] presents a contrastive way for denoising training and achieves the state-of-the-art performance. Building on SoTA performance, we employ DINO as the default detector in this paper. + +Large-scale Image Models. Some recent works [10, 33, 34, 41, 42, 48] scale up image models to reach 3 billion parameters, demonstrating a very powerful capacity of large-scale image models. Swin Transformer [33], built on the transformer architecture, incorporates techniques like post-norm and scaled cosine attention to enhance the stability of large vision models. Beit-3 [42], adopting the ViT [10] architecture, explores scaling up models through multi-modal pretraining. InternImage [41], utilizing a CNN base, integrates deformable convolution as a core operator to establish large-scale CNN models. After acquiring large-scale pretrained models, there has been significant interest in efficiently fine-tuning these models for downstream tasks. Notably, approaches such as VPT [23] insert learnable parameters into the input space while keeping most of the pretrained model frozen. Similarly, AdaptFormer [6] introduces lightweight learnable modules into the MLP block. Lin et al. [30] investigate the pretraining tasks that best align with the frozen setting. Different from these works, our paper aims to extract transferable knowledge from large models and use it to enhance tiny-scale models via the KD. + +# 3 Method + +In this section, we provide a concise overview of knowledge distillation and the detection transformer in Sec. 3.1. Subsequently, we introduce the frozen teacher method in Sec. 3.2, followed by an explanation of Query Distillation in Sec. 3.3. + +# 3.1 Preliminary + +Knowledge Distillation (KD). Standard KD enforces the student to align its probabilistic scores $(p^{(s)})$ with those of the teacher via the Kullback-Leibler (KL) divergence [21, 55]: + +$$ +\mathcal {L} _ {K D} = p ^ {(t)} \log \left(\frac {p ^ {(t)}}{p ^ {(s)}}\right). \tag {1} +$$ + +The teacher's logits $(p^{(t)})$ serve as soft label, providing non-target classes not present in the ground truth labels. Despite being weak, this non-target class information significantly enhances the student's generalization. To amplify this effect, prior works typically apply a high temperature to soften the $p^{(t)}$ . + +Detection Transformers (DETRs). In DETRs, the backbone extracts image features, which are then inputted into the transformer encoder along with positional encodings. The enhanced image features serve as the key and value, providing contextual information for the subsequent transformer decoder. The input for the decoder is N embeddings, known as object queries, denoted as $Q = \{q_0, q_1, \dots, q_{N-1}\}$ . Each query will be decoded into class probability $p$ and box coordinate $b$ via the transformer decoder. + +# 3.2 Frozen Teacher Method + +Inspired by the NLP domain, where scaling up language models achieves remarkable success [9, 12], some recent works scale up vision models to more than 1 billion parameters and set new records across various vision tasks [33,41,42,48]. It is natural to explore the possibility of transferring the knowledge from these large models to more compact ones using the Knowledge Distillation (KD) technique. However, empirical observations (Fig. 1) suggest that as the model size of the teacher increases, the discrepancy between the teacher and student becomes more severe, limiting the effectiveness of the KD. + +In this paper, we observed that when fine-tuning large-scale pretrained models on a specific dataset to create the teacher model, freezing the backbone can endow the teacher with richer transferable knowledge compared to fully fine-tuning. The overall framework of our DLIM-Det is illustrated in Fig. 3. The pretraining-then-finetuning paradigm is the cornerstone for modern neural networks [18, 19, 45], especially for large-scale models. Pretraining on massive data is essential for achieving top performance in the downstream tasks [41]. Although fully fine-tuning the entire model may enhance task-specific performance, it also heightens the risk of overfitting and diminishes the model's generalizability. We denote the teacher produced by the frozen setting as the frozen teacher. This frozen teacher maintains the generalizability of the large-scale models. As evidenced by the Average Recall (AR) metrics on LVIS labels shown in Fig. 2, the frozen teacher exhibits an increased ability to recognize a broader array of non-target objects + +![](images/3392f4f0a86addaaf5b05910f262f3bf8e4ae3adc4bb00ce93b59776da8fc22e.jpg) +Fig. 3: The framework of our DLIM-Det. Freezing the backbone when fine-tuning large-sized models on a specific dataset to create the teacher. The alignment between object queries of the teacher and student is conducted based on positions and relations. + +Table 1: The query density with different IoU thresholds. + +
MethodsIoU: 0.5IoU: 0.7
Swin-LIntern-LSwin-LIntern-L
Full finetuning38.133.615.213.0
Frozen21.321.87.58.3
+ +while avoiding excessive specialization on the ground truths. Enhanced generalizability endows the teacher with specific characteristics that facilitate the distillation process. + +In the following, we investigate how these distinctive characteristics enhance the efficiency of the distillation process. When equipped with DETRs, we compare the spatial distributions of object queries under both the frozen and fully fine-tuned settings. We observe that the frozen teacher locates object queries more dispersely and sparsely than the fully fine-tuned teacher. To quantify the dispersion of query spatial distribution, we introduce a metric called Query Density. Query density represents the number of object queries with an Intersection over Union (IoU) greater than $\tau$ surrounding each ground truth. We average the query density over all GT objects in the COCO 2017 validation set [29], with $\tau$ set to 0.5 and 0.7, respectively. The results are presented in Tab. 1. As results show, the backbone-frozen strategy significantly affects the query density, reducing it by at least $1/3$ . The fully fine-tuned teacher exhibits high query density values, and each GT object is surrounded by more than 30 queries with IoUs larger than 0.5. Among these compact queries, only one query should predict the object, while the others should predict no object. One example is visualized in Fig. 5. The detector needs to model their relations, identifying the optimal localized query with high confidence and suppressing the others with low confidence. In the case of sparse query locations, an instance is surrounded by fewer object queries, making their relations much easier to model. + +When using the fully fine-tuned teacher, the student model needs to allocate unnecessary capacity to model complex query relations in order to mimic the compact query locations of the teacher. In contrast, the sparse query locations provided by the frozen teacher are easier for the student to learn, reducing the learning burden. To fully leverage the sparse query of the frozen teacher, we design a Query Distillation method, which is described in Sec. 3.3. + +![](images/6354959c5867e932eca3ac2f94820c2aeeec9ae999a93d7f0bff5f2a72c7ab3b.jpg) +(a) Comparing four teachers. + +![](images/b1b20ee9aad885e188f94c9cf6d38c82349a675b13c69d02b7e0b5e70fe9e97b.jpg) +Input Images + +![](images/bf6bcba23cbe5bcda067c629234c709461f6e58f39e23703a57b2986968fedf2.jpg) +Fig. 4: Comparison of different teacher models based on mutual information. (a) Multiple teacher models and their mutual information concerning the input $(I(X;F))$ and target $(I(Y;F))$ . The blue lines represent the process of searching for the model whose representation maximizes $I(X;F) + I(Y;F)$ . "Full ft." denotes full fine-tuning. (b) The reconstructed image qualitatively reflects $I(X;F)$ . + +![](images/5ebb3b414094d4c68475fb0a664de49691afcbb1e9971a36a6d862fa42e8e289.jpg) +Full ft. Swin Base + +![](images/658a95210b226657247c29a7f847e8f092caf40756d75f1ff326d32cdb89c0f2.jpg) +(b) Visualization of reconstructed images. + +![](images/7cbee10677cb09278bcfc387a01642e2d492d6c382d6ee751e9a64f1d88b495d.jpg) +Frozen Swin Base + +![](images/46ad419ea7f7b5e2f9298a7a50e1721f145b171da41be0bf1af21f4a7062eda3.jpg) + +![](images/714a0c20640896fcaee6b9c6a61e84988747e22269c8c90446d33614d2b7eaf7.jpg) +Full ft. Swin Large + +![](images/f2d5547af3cbdb2d9c7ea5d47c20442454cd3b65b8982495317ab84518092c66.jpg) + +![](images/77719ac267537b1b8b50deb0593299fdce25000d177c55f93b0f748a126ba929.jpg) +Frozen Swin Large + +![](images/98aebe2e1a2640072dd5242b9b6bada62a6062edad234adbd6a9c5d489250c5b.jpg) + +Taking into account the aspect of generalization and drawing inspiration from prior research [40], which introduced an optimal teacher model selection algorithm based on mutual information, we also substantiate the superiority of the frozen teacher from the perspective of the mutual information principle. Specifically, the entropy of teacher's representation $(H(F))$ can be formulated as: + +$$ +H (F) = I (X; F) + I (Y; F) + Z, \tag {2} +$$ + +$I(X;F)$ and $I(Y;F)$ represent task-relevant information regarding the input image $X$ and ground truth target $Y$ , respectively. $Z$ refers to nuisance, which is task-irrelevant information. Following standard practice [40, 44], we employ the detection performance (e.g. mAP) to quantify $I(Y;F)$ . To measure the mutual information concerning the input $(I(X;F))$ , we feed the intermediate features into a decoder network (typically, U-Net [38]) to reconstruct the input image $X$ . The average reconstruction loss is used to quantify $I(X;F)$ . Different values of $I(X;F)$ can be clearly distinguished based on the quality of reconstructed images, as illustrated in Fig. 4b. The theoretical proofs and training loss curves can be found in the Appendix. + +After defining $I(X;F)$ and $I(Y;F)$ , we compare base-scale and large-scale models under both full fine-tuning and frozen settings, with results depicted in Fig. 4a, where mutual information $(I(X;F)$ and $I(Y;F))$ is normalized to ensure the scale matching. Among the four teacher models, the frozen Swin-L attains the maximum of task relevant information $I(X;F) + I(Y;F)$ , achieving a delicate trade-off between $I(X;F)$ and $I(Y;F)$ . A model with high values of $I(X;F) + I(Y;F)$ can effectively suppress nuisance $Z$ and serves as an informative teacher. Though Base-scale models exhibit higher values of $I(X;F)$ compared to Large-scale ones, they lag in terms of $I(Y;F)$ , which may result in inferior KD performance. Empirical experiments (Tab. 3) support our assumption: the frozen Swin-L achieves the best distillation performance among the four teacher models. To conclude, the teacher model selection algorithm [40], grounded in mutual information, further supports our frozen teacher method. + +# 3.3 Query Distillation + +Query Distillation aligns the spatial positions and pair-wise relations of queries between the teacher and student, consisting of Query Relation Distill and Query Relation Distill. Initially, we establish the one-to-one matching for object queries of two models, where the cost matrix is computed based on classification and localization costs. Given the cost matrix, Hungarian Matching is used to solve the graph matching, yielding the optimal permutation function $\hat{\sigma}$ : + +$$ +\hat {\sigma} = \underset {\sigma \in \mathfrak {S} _ {N}} {\arg \min } \sum_ {i} ^ {N} \mathcal {L} _ {\text {m a t c h}} \left(q _ {i} ^ {(t)}, q _ {\sigma (i)} ^ {(s)}\right), \tag {3} +$$ + +where $N$ equals the number of predicted objects. Among all permutations of $N$ elements $\mathfrak{S}_N$ , $\hat{\sigma}$ is the one with the lowest cost. Based on the class probability $p$ and box coordinate $b$ decoded from each object query, we define the cost between the teacher query $q^{(t)}$ and student $q^{(s)}$ as follows: + +$$ +\mathcal {L} _ {\text {m a t c h}} \left(q ^ {(t)}, q ^ {(s)}\right) = \alpha_ {1} \mathcal {L} _ {K L} \left(p ^ {(t)}, p ^ {(s)}\right) + \alpha_ {2} \mathcal {L} _ {L 1 L o s s} \left(b ^ {(t)}, b ^ {(s)}\right) + \alpha_ {3} \mathcal {L} _ {G I o U} \left(b ^ {(t)}, b ^ {(s)}\right), \tag {4} +$$ + +where $\alpha_{1},\alpha_{2},\alpha_{3}$ are set to 2,5,2 empirically. + +Query Position Distill. Instead of matching the probabilistic scores of two models, we observe that the spatial position of object queries can serve as informative knowledge. The position of an object query refers to the bounding box decoded from the query, represented by a vector: $b \in [x, y, w, h]$ , where $x$ and $y$ denote the center coordinates, while $w$ and $h$ represent the width and height, respectively. Given the matching $\sigma$ between object queries of two models, we formulate the procedure of Query Position Distill as follows: + +$$ +\mathcal {L} _ {Q P D} = \sum_ {i} ^ {N} \beta_ {1} \mathcal {L} _ {L 1} \left(b _ {i} ^ {(t)}, b _ {\sigma (i)} ^ {(s)}\right) + \beta_ {2} \mathcal {L} _ {G I o U} \left(b _ {i} ^ {(t)}, b _ {\sigma (i)} ^ {(s)}\right), \tag {5} +$$ + +where $\beta_{1},\beta_{2}$ are set to 5 and 2, empirically. Through the thorough experiments (listed in Tab. 8b), we empirically find that besides the foreground queries, even poorly localized teacher queries that cover background regions still contain valuable knowledge to distill. + +Due to the one-to-one label assignment, only a small subset of object queries receives localization supervision in each image. For example, on the COCO dataset [29], which typically contains an average of 7 instances per image, only 7 out of the 900 queries are involved in the regression loss. This sparsity of localization supervision impairs the training of detection transformers. By the proposed Query Position Distill, the spatial positions of teacher's queries serve as additional supervisions to provide richer information than the sparse ground truth labels. + +Query Relation Distill. In previous works, relation-based KD methods have been widely investigated. These methods aim to capture various relations between different entities such as instances [8, 22], pixels [49], anchors [27] or patches [17]. To model these relations, intricate modules, e.g., GcBlock [1], and complex relation mechanisms, e.g. Rank Mimicking [27], have been designed, which not only lack interpretability but + +![](images/b27d4468abbc83abfcc0c3a1d77e4e58efdd09cd6ebb3a1e561480dcdb34e7ec.jpg) +Fig.5: An example to show the position, score and relation of object queries. Query relations are obtained from self-attention. Only 3 queries are visualized for a clear presentation. + +also introduce additional training costs. In contrast, the transformer decoder of DETRs explicitly models the pair-wise relations between object queries through the self-attention. We denote by $A = [a_{ij}]_{N \times N}$ the attention map from the self-attention, where $N$ denotes the number of queries. L1 Loss is employed to narrow the distance between $A^{(t)}$ and $A^{(s)}$ : + +$$ +\mathcal {L} _ {Q R D} = \frac {1}{N} \sum_ {i} ^ {N} \sum_ {j} ^ {N} \left| A _ {i, j} ^ {(t)} - A _ {\sigma (i), \sigma (j)} ^ {(s)} \right|, \tag {6} +$$ + +where $\sigma$ denotes the permutation function to match the queries between the teacher and student. Query relations are critical for end-to-end detection. As shown in Fig. 5, precise query relations allow the best localized query to pop out with high classification scores, while the others will be suppressed with low scores. + +The overall training loss $\mathcal{L}_{\text{overall}}$ is: + +$$ +\mathcal {L} _ {\text {o v e r a l l}} = \gamma_ {1} \mathcal {L} _ {\text {t a s k}} + \gamma_ {2} \mathcal {L} _ {Q P D} + \gamma_ {3} \mathcal {L} _ {Q R D}, \tag {7} +$$ + +where $\gamma_{1},\gamma_{2},\gamma_{3}$ are set to 1, 2, 1, respectively. + +We also find that the parameters of the detection head in the teacher model serve as a source of knowledge. Even for heterogeneous DETRs, the basic blocks in their heads are consistent, including attention and feed forward network (FFN). Through empirical investigations, we prove that inheriting the parameters of these basic blocks from the teacher head to student head is convenient and beneficial. + +# 4 Experiments + +# 4.1 Experimental settings + +Datasets and Backbone. We conduct the experiments on the MS COCO 2017 dataset [29], which is a commonly used object detection dataset. The dataset is split into the train2017 set with 118k images and the val2017 set with 5k images. All models are trained on the train2017 set, and evaluated on the val2017 set. Results are reported with two backbones: Swin Transformer [34] and InternImage [41]. We employ the COCO-style average precision (mAP) as the metric. DINO [54] is the default detector. + +Implementation details. All models, except for the Intern-Huge model, are trained on $8 \times 32\mathrm{G}$ V100 GPUs with a batch size of 16. In contrast, the Intern-Huge models are + +Table 2: Effectiveness of each component. + +
MethodsSwin TransformerInternImage
mAP\( AP_{50} \)\( AP_{75} \)mAP\( AP_{50} \)\( AP_{75} \)
Teacher: Large Scale58.177.163.556.975.861.9
Student: Tiny Scale51.369.056.053.271.558.0
Query DistillFrozen TeacherInherit Head
✓51.6 (+0.3)69.556.454.2 (+1.0)72.559.2
✓✓52.2 (+0.9)70.157.054.7 (+1.5)72.959.8
✓✓53.7 (+2.4)71.758.855.2 (+2.0)73.760.3
✓✓✓54.4 (+3.1)72.159.455.9 (+2.7)74.161.4
+ +Table 3: Distillation performance comparison between Large- and Base-scale teacher models. Numbers in brackets refer to distillation gains against the original student (Swin-T, 51.3 mAP). + +
TeacherFreeze BackboneDistilled Student
mAPAP50AP75
Large-scale✓54.4 (+3.1)72.1 (+3.1)59.4 (+3.4)
Base-scale✓53.5 (+2.2)71.5 (+2.5)58.5 (+2.5)
Base-scale×53.1 (+1.8)71.1 (+2.1)58.0 (+2.0)
+ +trained on 80GB A100 GPUs. Student models are trained with 12 epochs with AdamW as the optimizer. The initial learning rate is set to 0.0001, which is divided by 10 at the 11th epoch. Following the default practice of DINO [54], the number of object query is set to 900. + +# 4.2 Ablation Studies + +In ablation studies, unless otherwise stated, we adopt the tiny-scale models with around 30M parameters as the student. + +Effects of each component. In Tab. 2, we verify the effectiveness of each component, including the frozen teacher method, Query Distillation, and head inheriting. When facing large-scale teacher models, Query Distillation produces a modest increase of $+0.3 / 1.0$ mAP for Swin/Intern, respectively. Subsequently, we replace the fully fine-tuned teacher with the frozen teacher and perform Query Distillation again. The results demonstrate that the frozen teacher significantly amplifies the distillation gains. Specifically, compared to the fully fine-tuned teacher, the frozen teacher yields an additional increase of $+0.6 / 0.5\%$ for Swin/Intern models, respectively. Moreover, inheriting the transformer's parameters of the teacher's head produces noticeable improvements, suggesting that the head parameters of the teacher contain some general knowledge that is somewhat independent of the backbone structure. Building on head parameter inheriting, the frozen teacher outperforms the fully fine-tuned teacher by $+0.7 / 0.7\%$ for Swin/Intern models. These experimental findings validate that the frozen teacher is easier to distill than the previously employed fully fine-tuned teacher. Finally, by combining the three components, DLIM-Det, achieves significant improvements, surpassing the student baseline by $+3.1 / 2.7\%$ for the Swin/Intern models, respectively. + +Distillation on medium- and large-sized teacher models. In Tab. 3, we compare the distillation performance under Base- and Large-scale teacher models. With the help of the frozen teacher method, a large-scale teacher surpasses its base-scale counterpart by $+0.9$ mAP. These results affirm the core premise of this paper: larger teacher models lead + +Table 4: The frozen setting affects the detection performance. + +
MethodsFreeze BackboneSchedulemAP\( AP_{50} \)\( AP_{75} \)
Swin-Large×3x58.177.163.5
✓3x56.6 (-1.5)74.961.9
Intern-Large×2x57.876.563.1
✓2x55.8 (-2.0)74.461.0
+ +Table 5: Applying different training schedules to the Frozen and Full ft. settings as to obtain teachers with identical performance. The KD contains Query Distillation and the inheriting head. + +
MethodsFreeze BackboneSchedulemAP\( AP_{50} \)\( AP_{75} \)
Teacher:
T1: Swin-Large×1x56.975.662.2
T2: Swin-Large✓3x56.774.961.9
Student:
Swin-Tiny-1x51.369.056.0
KD from T1-1x53.1 (+1.8)70.957.9
KD from T2-1x54.4 (+3.1)72.159.4
+ +Table 6: Effectiveness of Query Position Distill (QPD) and Query Relation Distill (QRD). "T/S" denotes the Teacher/Student. + +
MethodsInternImageSwin Transformer
mAPAP50AP75mAPAP50AP75
T: Base54.673.659.355.874.460.7
S: Tiny53.271.558.051.369.056.0
+ QPD54.2 (+1.0)72.459.151.6 (+0.3)69.556.2
+ QRD54.2 (+1.0)72.959.151.8 (+0.5)70.256.5
+ QPD & QRD54.5 (+1.3)73.259.552.1 (+0.8)69.956.7
+ +to greater distillation gains. Additionally, the frozen teacher method generalizes well to base-scale models. When using the base-scale model as a teacher, the frozen teacher surpasses the fully fine-tuned teacher by $0.4\mathrm{mAP}$ . + +Fair comparison between the fully fine-tuned teacher and frozen teacher. Freezing the backbone typically results in performance degradation to some extent compared to fully fine-tuning. As shown in Tab. 4, under the same training schedule, the frozen teacher lags behind the fully fine-tuned teacher by around $2\mathrm{mAP}$ . One might argue that, owing to lower detection performance, the frozen teacher narrows the gap between the teacher and student, thereby attaining better distillation results. In Tab. 5, we conduct experiments to challenge this perspective. Concretely, we assign long and short training schedules to the frozen and fully fine-tuned teachers, resulting in teachers with nearly identical performance. By ensuring the same gap between the teacher and student, we find that the frozen teacher surpasses the fully fine-tuned teacher by a large margin, +1.3 mAP. These results highlight the intrinsic advantages of the frozen teacher approach, such as sparse query locations. + +Effectiveness of the Query Distillation. The Query Distillation is specifically designed for detection transformers, including Query Position Distill (QPD) and Query Relation Distill (QRD). We ablate each of them step by step in Tab. 6. Here, we employ the medium-sized models as the teacher, to exclude the effects of the large gap between the teacher and student. Besides the experiments on the advanced network architecture, such + +Table 7: Applying Query Position Distill (QPD) and Query Relation Distill (QRD) to classical backbones (e.g. ResNet [20]). + +
MethodsResnet
mAP\( AP_{50} \)\( AP_{75} \)
T: ResNet10150.067.754.4
S: ResNet5048.966.453.4
+ QPD49.5 (+0.5)67.253.9
+ QRD49.5 (+0.5)67.554.1
+ QPD & QRD49.8 (+0.9)67.654.2
+ +Table 8: Study on distillation positions and teacher query selection. +(a) Study on distillation positions. Teacher: InternBase (54.6 mAP); Student: Interm-Tiny (53.2 mAP) + +
KD on decoder layermAP\( AP_{50} \)\( AP_{75} \)
123456
✓✓✓54.172.659.0
✓✓✓53.872.358.8
✓✓✓✓✓✓54.573.259.5
✓54.172.659.0
✓53.572.158.3
✓52.470.957.0
+ +(b) Study on teacher query selection. FG refers to foreground; BG refers to background. + +
MethodsmAP\( AP_{50} \)\( AP_{75} \)
T: Intern-Base54.673.659.3
S: Intern-Tiny53.271.558.0
No Sample (900)54.272.659.1
Random Sample (300)54.573.259.5
FG queries (IoU > 0.3)54.272.559.2
BG queries (IoU < 0.3)54.272.859.2
+ +Table 9: Applying our DLIM-Det to the huge model (InternImage-Huge) with 1B parameters to verify the scalability. Numbers in brackets refer to distillation gains against the original student. + +
Teacher#ParamsDistilled Student
mAPAP50AP75
Intern-Huge1.08B56.1 (+2.9)74.4 (+2.9)61.5 (+3.5)
Intern-Large223M55.9 (+2.7)74.1 (+2.6)61.4 (+3.4)
+ +as Swin Transformer and InternImage (Tab. 6), we also report results on ResNet [20] in Tab. 7. Our Query Distillation method yields consistent improvements on three different backbones, at least $+0.8$ mAP. Furthermore, combining the QPD and QRD together achieves better performance than using them individually, verifying the effectiveness of each component. + +Study on the distillation positions. To investigate the impact of Query Distillation on different decoder layers in DETRs, we conducted experiments, and the results are presented in Tab. 8a. When applying Query Distillation to all 6 decoder layers, we observe the best performance, surpassing the counterparts where it was applied to the first three or last three decoder layers. Furthermore, we examine the performance when applying Query Distillation to only one decoder layer. The results show that distillation on shallow decoder layers outperforms distillation on deeper layers. + +Selection of teacher queries. Using our Query Distillation method, we further conduct experiments to investigate the impact of different subsets of teacher queries on knowledge transfer. Results are reported in Tab. 8b. First, we perform KD on all teacher queries, obtaining $+1.0\%$ gains. Then we randomly sample 300 queries and achieve further improvements, surpassing the 900 query baseline by $+0.3\%$ . To explore alternative query selection strategies, we define the queries whose IoUs with GT are greater than 0.3 as foreground queries, otherwise as background queries. Distilling knowledge from either pure foreground (namely IoU $>0.3$ in Tab. 8b) or pure background queries results in inferior performance compared to random sampling. This suggests that both foreground and background queries contain valuable knowledge, and distilling knowledge from a combination of them yields optimal results. + +Table 10: Experiments on heterogeneous teacher-student pairs. Heterogeneous components are underlined. Numbers in brackets denote distillation gains against the original student. + +
TeacherStudentDistilled Student
mAPAP50AP75
DINO-Swin-LDeformable DETR-Swin-T49.6 (+3.5)68.3 (+3.4)53.7 (+3.9)
DINO-Intern-HDINO-Swin-T53.3 (+2.0)71.3 (+2.3)58.3 (+2.3)
+ +Table 11: Comparison with state-of-the-art knowledge distillation methods. + +
MethodsSwin TransformerInternImage
mAP\( AP_{50} \)\( AP_{75} \)\( AP_S \)\( AP_M \)\( AP_L \)mAP\( AP_{50} \)\( AP_{75} \)\( AP_S \)\( AP_M \)\( AP_L \)
Teacher: Large58.177.163.541.262.173.957.876.563.141.661.373.0
Student: Tiny51.369.056.034.554.466.153.271.558.035.556.669.3
FGD [49]51.1 (-0.2)69.255.834.353.865.652.8 (-0.4)71.157.636.456.466.8
MGD [50]51.1 (-0.2)69.855.834.954.266.154.2 (+1.0)72.859.138.257.669.4
FitNet [37]51.3 (+0.0)70.055.934.754.566.353.8 (+0.5)72.458.738.257.468.2
DETRDistill [3]51.7 (+0.4)69.456.435.054.866.054.7 (+1.5)72.859.838.558.269.4
Our DLIM-Det54.4 (+3.1)72.159.438.157.469.755.9 (+2.7)74.161.440.359.470.8
+ +Scalability to huge teacher models. We apply DLIM-Det to InternImage-Huge with 1.08 billion parameters. From Tab. 9, we observe that DLIM-Det achieves satisfactory performance even when the teacher has more than 30 times the parameters of the student. Here, the student is represented by Intern-Tiny with 30M parameters, achieving a mAP of 53.2. Specifically, Intern-Huge produces a $+2.9$ distillation gain on Intern-Tiny based student and surpasses Intern-Large by 0.2 mAP (56.1 v.s. 55.9). + +Distillation on heterogeneous architectures. To verify the generalization of DLIM-Det, we conduct experiments on heterogeneous teacher and student models, including heterogeneous detectors and heterogeneous backbones. For heterogeneous detectors, with DINO [54] as the teacher and Deformable DETR [57] as the student, DLIM-Det improves the student from 46.1 to $49.6\mathrm{mAP}$ , as reported in Tab. 10. In the case of heterogeneous backbones, where Intern-Huge/Swin-Tiny are employed as the teacher/student, DLIM-Det obtains a considerable improvement of $+2.0\mathrm{mAP}$ . These results show the strong generalization of DLIM-Det. + +# 4.3 Comparison with State-of-the-art Methods + +Main results on two backbones. In Tab. 11, we compare our DLIM-Det with several popular KD methods. FitNet [37] is a classical feature distillation method, which introduces adaptation layers to bridge the semantic gap between the teacher and student. However, FitNet fails to improve the performance of the transformer-based backbone (Swin Transformer), showing no improvements. MGD [50] is inspired by the masked image modeling [18] and randomly masks pixels of the student's features. It then leads the student to generate the teacher's features in the masked regions. MGD obtains a noticeable improvement of $+1.0\%$ for InternImage, but struggles to improve the Swin Transformer. FGD [49] leverages the GcBlock [1] to capture the relations between pixels from feature maps. However, FGD does not perform well on both backbones, indicating that the hand-crafted relation mechanism may conflict with the intrinsic relation modeling of detection transformers. In fact, FGD even hurts the performance of InternImage by $-0.4\%$ . The recently proposed DETRDistill [3] is designed for distilling DETRs. In addition to performing customized logits and feature distillation, it also distills the label + +assignment results of teacher. Notably, our DLIM-Det outperforms it by a large margin, at least $+1.2\%$ . We explain that the Query prior Assignment Distillation in DETRDistill serves a role akin to the De-Noising group in DINO, thereby it only yields minor improvements in our experiments. Moreover, another contribution of DETRDistill [3] is a feature imitation method, which is vulnerable to significant discrepancy between the teacher and student, as confirmed by other works [37,49,50]. + +These experimental findings suggest that while previous works can achieve decent distillation gains for medium-sized teacher, they encounter challenges in distilling large-scale teacher models. In contrast, DLIM-Det demonstrates significant distillation performance specifically tailored for large-scale models. + +Table 12: Comparison with the Teacher Assistant (TA) on two backbones. The TA consists of two steps, thus the overall training time is the sum of the training time for both steps. + +
MethodsSwin TransformerInternImage
mAP\( AP_{50} \)\( AP_{75} \)Train TimemAP\( AP_{50} \)\( AP_{75} \)Train Time
Baseline:
Teacher: Large58.177.163.5-57.876.563.1-
Intermediate Teacher: Base55.874.460.7-54.673.659.3-
Student: Tiny51.369.056.0-53.271.558.0-
TA step1 (Large→Base)57.5 (+1.7)76.162.827.3h56.8 (+2.2)75.262.134.5h
TA setup2 (Base→Tiny)53.7 (+2.4)71.758.919.2h55.7 (+2.5)74.060.723.6h
Our DLIM-Det (Large→Tiny)54.4 (+3.1)72.159.420h55.9 (+2.7)74.161.424.8h
+ +Comparison with the Teacher Assistant. The Teacher Assistant (TA) [35, 39, 56] is proved effective in reducing the gap between the teacher and student models. Following the previous work [35], we use the base-scale model with approximately 90M parameters as the TA model. The TA model is distilled from the teacher model and then serves as the teacher for the student model. The results in Tab. 12 demonstrate that our DLIM-Det, which directly distills knowledge from large-scale models, achieves superior performance and higher efficiency compared to the TA method. Specifically, by using only half of the training time, DLIM-Det outperforms Teacher Assistant by $0.7 / 0.2$ mAP on Swin/Intern models, respectively. + +# 5 Conclusion + +In this paper, we introduce DLIM-Det, a novel knowledge distillation method specifically tailored for large-scale image models in the context of object detection. To bridge the gap between the teacher and student, we propose the frozen teacher method, which involves freezing the pretrained backbone during teacher creation. The frozen teacher preserves the generalization capacity of large-scale models and gives rise to distinctive characteristics, such as sparse query locations. Through the proposed Query Distillation method, we harness the full advantages offered by the frozen teacher, thereby facilitating the distillation process. Extensive experiments provide robust evidence that larger teacher models can indeed lead to greater distillation gains through DLIM-Det. + +Acknowledgement. This work is partially supported by National Natural Science Foundation of China (No. 62322602), Natural Science Foundation of Jiangsu Province, China (No. BK20230033), and National Natural Science Foundation of China (No. 62172225). + +# References + +1. Cao, Y., Xu, J., Lin, S., Wei, F., Hu, H.: Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In: Proceedings of the IEEE/CVF international conference on computer vision workshops. pp. 0-0 (2019) 8, 13 +2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020) 4 +3. Chang, J., Wang, S., Xu, H.M., Chen, Z., Yang, C., Zhao, F.: Detrdistill: A universal knowledge distillation framework for detr-families. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6898-6908 (2023) 2, 4, 13, 14 +4. Chen, G., Choi, W., Yu, X., Han, T., Chandraker, M.: Learning efficient object detection models with knowledge distillation. Advances in neural information processing systems 30 (2017) 3 +5. Chen, Q., Chen, X., Zeng, G., Wang, J.: Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085 (2022) 3, 4 +6. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) 4 +7. Chen, X., Chen, J., Liu, Y., Zeng, G.: DG etr: Decoder distillation for detection transformer arXiv preprint arXiv:2211.09768 (2022) 4 +8. Dai, X., Jiang, Z., Wu, Z., Bao, Y., Wang, Z., Liu, S., Zhou, E.: General instance distillation for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7842-7851 (2021) 2, 4, 8 +9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 5 +10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 4 +11. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. arXiv preprint arXiv:2211.07636 (2022) 1 +12. Floridi, L., Chiriatti, M.: Gpt-3: Its nature, scope, limits, and consequences. Minds and Machines 30, 681–694 (2020) 5 +13. Gao, Z., Wang, L., Han, B., Guo, S.: Adamixer: A fast-converging query-based object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5364-5373 (2022) 4 +14. Guo, G., Han, L., Wang, L., Zhang, D., Han, J.: Semantic-aware knowledge distillation with parameter-free feature uniformization. Visual Intelligence 1(1), 6 (2023) 3 +15. Guo, J., Han, K., Wang, Y., Wu, H., Chen, X., Xu, C., Xu, C.: Distilling object detectors via decoupled features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2154-2164 (2021) 2, 3, 4 +16. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019) 2 +17. Hao, Z., Guo, J., Jia, D., Han, K., Tang, Y., Zhang, C., Hu, H., Wang, Y.: Learning efficient vision transformers via fine-grained manifold distillation. In: Advances in Neural Information Processing Systems (2021) 3, 4, 8 + +18. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022) 5, 13 +19. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 5 +20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 4, 12 +21. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 2, 5 +22. Huang, T., You, S., Wang, F., Qian, C., Xu, C.: Knowledge distillation from a stronger teacher. arXiv preprint arXiv:2205.10536 (2022) 4, 8 +23. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) 4 +24. Jia, Z., Sun, S., Liu, G., Liu, B.: Mssd: multi-scale self-distillation for object detection. Visual Intelligence 2(1), 8 (2024) 3 +25. Kang, Z., Zhang, P., Zhang, X., Sun, J., Zheng, N.: Instance-conditional knowledge distillation for object detection. Advances in Neural Information Processing Systems 34, 16468-16480 (2021) 3 +26. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022) 4 +27. Li, G., Li, X., Wang, Y., Zhang, S., Wu, Y., Liang, D.: Knowledge distillation for object detection via rank mimicking and prediction-guided feature imitation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1306-1313 (2022) 4, 8 +28. Li, G., Li, X., Zhang, S., Yang, J.: Towards more reliable evaluation in pedestrian detection by rethinking "ignore regions". Visual Intelligence 2(1), 4 (2024) 4 +29. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 6, 8, 9 +30. Lin, Y., Liu, Z., Zhang, Z., Hu, H., Zheng, N., Lin, S., Cao, Y.: Could giant pre-trained image models extract universal representations? Advances in Neural Information Processing Systems 35, 8332-8346 (2022) 4 +31. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022) 4 +32. Liu, Y., Chen, K., Liu, C., Qin, Z., Luo, Z., Wang, J.: Structured knowledge distillation for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2604-2613 (2019) 3 +33. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12009-12019 (2022) 1, 4, 5 +34. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) 1, 2, 3, 4, 9 +35. Mirzadeh, S.I., Farajtabar, M., Li, A., Levine, N., Matsukawa, A., Ghasemzadeh, H.: Improved knowledge distillation via teacher assistant. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 5191-5198 (2020) 2, 14 + +36. Park, D.Y., Cha, M.H., Kim, D., Han, B., et al.: Learning student-friendly teacher networks for knowledge distillation. Advances in neural information processing systems 34, 13292-13303 (2021) 2 +37. Romero, A., Ballas, N., Kahou, S.E., Chassang, A., Gatta, C., Bengio, Y.: Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550 (2014) 2, 3, 13, 14 +38. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) 7 +39. Son, W., Na, J., Choi, J., Hwang, W.: Densely guided knowledge distillation using multiple teacher assistants. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9395-9404 (2021) 2, 14 +40. Wang, C., Yang, Q., Huang, R., Song, S., Huang, G.: Efficient knowledge distillation from model checkpoints. Advances in Neural Information Processing Systems 35, 607-619 (2022) 7 +41. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., et al.: Interimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778 (2022) 1, 3, 4, 5, 9 +42. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442 (2022) 1, 4, 5 +43. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2567-2575 (2022) 4 +44. Wang, Y., Ni, Z., Song, S., Yang, L., Huang, G.: Revisiting locally supervised learning: an alternative to end-to-end training. arXiv preprint arXiv:2101.10832 (2021) 7 +45. Wei, F., Gao, Y., Wu, Z., Hu, H., Lin, S.: Aligning pretraining for detection via object-level contrastive learning. Advances in Neural Information Processing Systems 34, 22682-22694 (2021) 5 +46. Xu, T., Zhu, X.F., Wu, X.J.: Learning spatio-temporal discriminative model for affine subspace based visual object tracking. Visual Intelligence 1(1), 4 (2023) 4 +47. Yang, C., Zhou, H., An, Z., Jiang, X., Xu, Y., Zhang, Q.: Cross-image relational knowledge distillation for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12319-12328 (2022) 3 +48. Yang, J., Li, C., Zhang, P., Dai, X., Xiao, B., Yuan, L., Gao, J.: Focal attention for long-range interactions in vision transformers. Advances in Neural Information Processing Systems 34, 30008-30022 (2021) 1, 4, 5 +49. Yang, Z., Li, Z., Jiang, X., Gong, Y., Yuan, Z., Zhao, D., Yuan, C.: Focal and global knowledge distillation for detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4643-4652 (2022) 4, 8, 13, 14 +50. Yang, Z., Li, Z., Shao, M., Shi, D., Yuan, Z., Yuan, C.: Masked generative distillation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI. pp. 53-69. Springer (2022) 2, 13, 14 +51. Yao, L., Pi, R., Xu, H., Zhang, W., Li, Z., Zhang, T.: G-detkd: towards general distillation framework for object detectors via contrastive and semantic-guided feature imitation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3591-3600 (2021) 2, 3 +52. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021) 1 + +53. Zeng, Z., Yang, F., Liu, H., Satoh, S.: Improving deep metric learning via self-distillation and online batch diffusion process. Visual Intelligence 2(1), 1-13 (2024) 3 +54. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022) 2, 3, 4, 9, 10, 13 +55. Zhang, Y., Xiang, T., Hospedales, T.M., Lu, H.: Deep mutual learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4320-4328 (2018) 5 +56. Zheng, Z., Ye, R., Hou, Q., Ren, D., Wang, P., Zuo, W., Cheng, M.M.: Localization distillation for object detection. arXiv preprint arXiv:2204.05957 (2022) 2, 4, 14 +57. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) 4, 13 +58. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training. arXiv preprint arXiv:2211.12860 (2022) 3, 4 \ No newline at end of file diff --git a/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/images.zip b/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ac4a153e88320de4da370af128c92e0e505aca22 --- /dev/null +++ b/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:098f90463d2b0f4f95cbe381fd38c0ee830134335da6b39d425ced9dad7b7949 +size 518809 diff --git a/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/layout.json b/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0c2f71feb95f5df368626137d953e82d957349e6 --- /dev/null +++ b/2024/Distilling Knowledge from Large-Scale Image Models for Object Detection/layout.json @@ -0,0 +1,10110 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "text", + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "spans": [ + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "text", + "content": "Gang Li" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "text", + "content": ", Wenhai Wang" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "inline_equation", + "content": "^{3,2}" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "text", + "content": ", Xiang Li" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "text", + "content": ", Ziheng Li" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "text", + "content": ", Jian Yang" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "text", + "content": ", Jifeng Dai" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "inline_equation", + "content": "^{5,2}" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "text", + "content": ", Yu Qiao" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "text", + "content": ", and Shanshan Zhang" + }, + { + "bbox": [ + 175, + 167, + 442, + 195 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 143, + 202, + 471, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 202, + 471, + 227 + ], + "spans": [ + { + "bbox": [ + 143, + 202, + 471, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 202, + 471, + 227 + ], + "type": "text", + "content": "Nanjing University of Science and Technology " + }, + { + "bbox": [ + 143, + 202, + 471, + 227 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 202, + 471, + 227 + ], + "type": "text", + "content": "OpenGVLab, Shanghai AI Laboratory " + }, + { + "bbox": [ + 143, + 202, + 471, + 227 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 143, + 202, + 471, + 227 + ], + "type": "text", + "content": "The Chinese University of Hong Kong " + }, + { + "bbox": [ + 143, + 202, + 471, + 227 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 143, + 202, + 471, + 227 + ], + "type": "text", + "content": "Nankai University " + }, + { + "bbox": [ + 143, + 202, + 471, + 227 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 143, + 202, + 471, + 227 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 255, + 455, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 255, + 455, + 497 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 455, + 497 + ], + "type": "text", + "content": "Abstract. Large-scale image models have made great progress in recent years, pushing the boundaries of many vision tasks, e.g., object detection. Considering that deploying large models is impractical in many scenes due to expensive computation overhead, this paper presents a new knowledge distillation method, which Distills knowledge from Large-scale Image Models for object detection (dubbed DLIM-Det). To this end, we make the following two efforts: (1) To bridge the gap between the teacher and student, we present a frozen teacher approach. Specifically, to create the teacher model via fine-tuning large models on a specific task, we freeze the pretrained backbone and only optimize the task head. This preserves the generalization capability of large models and gives rise to distinctive characteristics in the teacher. In particular, when equipped with DEtection Transformers (DETRs), the frozen teacher exhibits sparse query locations, thereby facilitating the distillation process. (2) Considering that large-scale detectors are mainly based on DETRs, we propose a Query Distillation (QD) method specifically tailored for DETRs. The QD performs knowledge distillation by leveraging the spatial positions and pair-wise relations of teacher's queries as knowledge to guide the learning of object queries of the student model. Extensive experiments are conducted on various large-scale image models with parameter sizes ranging from 200M to 1B. Our DLIM-Det improves the student with Swin-Tiny by 3.1 mAP when the DINO detector with Swin-Large is used as the teacher. Besides, even when the teacher has 30 times more parameters than the student, DLIM-Det still attains a +2.9 distillation gain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 506, + 453, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 506, + 453, + 518 + ], + "spans": [ + { + "bbox": [ + 160, + 506, + 453, + 518 + ], + "type": "text", + "content": "Keywords: Object detection " + }, + { + "bbox": [ + 160, + 506, + 453, + 518 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 506, + 453, + 518 + ], + "type": "text", + "content": " Knowledge distillation " + }, + { + "bbox": [ + 160, + 506, + 453, + 518 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 506, + 453, + 518 + ], + "type": "text", + "content": " Large-scale image model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 539, + 218, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 218, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 218, + 552 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "type": "text", + "content": "It is proved that scaling up model sizes is a promising direction to make neural networks reach human performance [33, 42, 48, 52]. Recently, great progress has been achieved in large-scale image models [11, 33, 34, 41], pushing performance boundaries of many vision tasks, including object detection, semantic segmentation, etc. These large-scale models demand for high performance computing device, making them hard to deploy in commonly used hardware platforms. Therefore, Knowledge Distillation (KD) is necessary" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 643, + 416, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 643, + 416, + 654 + ], + "spans": [ + { + "bbox": [ + 141, + 643, + 416, + 654 + ], + "type": "text", + "content": "† This work was done while G. Li was an intern at Shanghai AI Laboratory." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 654, + 352, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 352, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 352, + 665 + ], + "type": "text", + "content": "Corresponding author (shanshan.zhang@njust.edu.cn)." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 328, + 223 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 328, + 223 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 328, + 223 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 328, + 223 + ], + "type": "image", + "image_path": "832a8a9182abe72865c99a7f18a8a47ba47adcdce852c5751193373071b36ae2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 228, + 330, + 284 + ], + "lines": [ + { + "bbox": [ + 131, + 228, + 330, + 284 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 330, + 284 + ], + "type": "text", + "content": "Fig. 1: Comparison between previous KD [3] and our DLIM-Det method for distilling large-sized teacher models. Previous methods struggle to distill knowledge from large models, whereas our method works well. Experiments are conducted on Swin Transformer [34]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 336, + 125, + 480, + 222 + ], + "blocks": [ + { + "bbox": [ + 336, + 125, + 480, + 222 + ], + "lines": [ + { + "bbox": [ + 336, + 125, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 336, + 125, + 480, + 222 + ], + "type": "image", + "image_path": "5bd748ae7395043a83f110af21c45286acdea2da855bc2bd07a5972dbf89c429.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 228, + 482, + 283 + ], + "lines": [ + { + "bbox": [ + 335, + 228, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 335, + 228, + 482, + 283 + ], + "type": "text", + "content": "Fig.2: Average Recall (AR) comparison on LVIS labels for the frozen teacher and fully fine-tuned teacher. Higher AR numbers indicate better generalization capacity." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 294, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 365 + ], + "type": "text", + "content": "to extract powerful and informative knowledge from the large model (teacher) to boost the compact model (student). The teacher knowledge serves as auxiliary supervisions and helps the student achieve better performance than trained only with ground-truth labels [21, 37]. However, existing KD methods [3, 8, 15, 50, 51] for object detection mainly focus on medium- or small-sized models, leaving the distillation from large-scale models rarely explored." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 366, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 605 + ], + "type": "text", + "content": "Distilling knowledge from large-scale image models is not trivial. As observed in previous works [35, 36], the discrepancy between the teacher and student becomes severe as the model size of teacher model increases, which hinders the effectiveness of KD methods. As shown in Fig. 1, when directly applying the prior KD method (e.g., DETRDistill [3]) to a detector (e.g., DINO [54]), the performance even degrades as the teacher model becomes stronger. Specifically, though the Swin-Large [34] has " + }, + { + "bbox": [ + 130, + 366, + 482, + 605 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 130, + 366, + 482, + 605 + ], + "type": "text", + "content": " parameters than the Swin-Base, the latter is a better teacher and outperforms the former by " + }, + { + "bbox": [ + 130, + 366, + 482, + 605 + ], + "type": "inline_equation", + "content": "+0.5\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 366, + 482, + 605 + ], + "type": "text", + "content": " in terms of KD gains. To address this problem, previous works [35, 39, 56] design a teacher assistant strategy, which generates multiple intermediate-sized teacher candidates to bridge the gap between the teacher and student. However, such a multi-step KD method induces additional training costs and severely affects the efficiency. In this paper, we propose that when obtaining the teacher model through fine-tuning the large-scale image backbone on a specific dataset, freezing the backbone and only optimizing the task head can empower the teacher with richer transferable knowledge compared to fully fine-tuning. Large-scale models exhibit increased generalizability due to their strong capacity and training on extensive datasets, however, there is a risk of overfitting when fine-tuned on specific datasets. Through freezing the pre-trained backbone, we maintain the generalizability of large-scale models, resulting in improved distillation performance. The teacher with pretrained large-scale backbone frozen is denoted as the frozen teacher." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "To quantitatively assess the generalizability of both the frozen and fully fine-tuned (ft.) teachers, we compute the Average Recall (AR) on the LVIS annotations [16], and the results are presented in Fig. 2. Swin-Large [34] based DINO [54] serves as the teacher model. Notably, both teachers are trained only with supervision from the 80 object categories in COCO. The LVIS dataset, while sourcing images from COCO, features" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 196, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 196, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 196, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "a re-annotated and broader spectrum of 1203 categories. As results shown, the frozen teacher outperforms the fully ft. teacher by a margin of " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_{900}" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": ", indicating that the frozen teacher generates more generalizable features. Strengthening the generalizability of the teacher model endows it with specific characteristics that facilitate the distillation process. In particular, when equipped with detection transformers, sparse query locations emerge in the frozen teacher." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 483, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 483, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 483, + 391 + ], + "type": "text", + "content": "Advanced large-scale detectors [5, 41, 54, 58] are predominately based on detection transformers, where object queries absorb contextual information for each instance then output predictions. Object queries from the frozen teacher exhibit sparser distribution compared to those from the fully fine-tuned teacher. This sparsity characteristic enables the student model to circumvent unnecessary investments in modeling complex query relations, particularly those leading to excessively compact query locations. To fully exploit sparse query locations from the frozen teacher, we introduce a Query Distillation method. Our Query Distillation aligns object queries between the teacher and student models based on spatial positions and relations. This method consists of two key components: Query Position Distill (QPD) and Query Relation Distill (QRD). Initially, we establish a one-to-one set matching between object queries of two models. Subsequently, the proposed QPD allows the teacher's queries to serve as auxiliary localization supervisions during student model training. Aside from QPD, we further propose Query Relation Distillation to lead the student model to mimic pair-wise relations between teacher's queries. Combining Query Distillation and the frozen teacher can significantly alleviate the learning burden for the student, thereby contributing to a more efficient knowledge distillation process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 483, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 483, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 483, + 536 + ], + "type": "text", + "content": "To conclude, we propose a new object detection knowledge distillation method, called DLIM-Det, tailored for Large-scale Image Models ranging from 200M to 1B parameters. Leveraging Query Distillation and the frozen teacher approach, DLIM-Det effectively transfers knowledge from large-scale models to compact ones. Extensive experiments on both the Transformer-based (Swin Transformer [34]) and CNN-based (InternImage [41]) models demonstrate the effectiveness of our approach. Results indicate our Query Distillation method works well for detection transformers, surpassing existing KD methods. Additionally, the frozen teacher approach bridges the gap between teacher and student, facilitating efficient distillation from large-scale models. Notably, when the teacher model is Swin-Large, DLIM-Det improves Swin-Tiny based DINO [54] by " + }, + { + "bbox": [ + 130, + 392, + 483, + 536 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 130, + 392, + 483, + 536 + ], + "type": "text", + "content": "; even for the giant InternImage-Huge model (" + }, + { + "bbox": [ + 130, + 392, + 483, + 536 + ], + "type": "inline_equation", + "content": "1B+" + }, + { + "bbox": [ + 130, + 392, + 483, + 536 + ], + "type": "text", + "content": " parameters), it improves the InternImage-Tiny student by " + }, + { + "bbox": [ + 130, + 392, + 483, + 536 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 130, + 392, + 483, + 536 + ], + "type": "text", + "content": ", achieving a remarkable 56.1 mAP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 555, + 226, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 555, + 226, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 555, + 226, + 567 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 483, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 483, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 483, + 667 + ], + "type": "text", + "content": "Knowledge Distillation. As a classical model compression technique, Knowledge Distillation (KD) has been applied to many vision tasks [14, 17, 24, 25, 32, 47, 53]. In this paper, we focus on the KD for object detection. Using the intermediate features of the teacher as knowledge, feature distillation offers deep supervisions for the student [4, 15, 37, 51]. However, feature distillation is usually vulnerable to semantic misalignment between features of two models, and suffers from performance degradation when the discrepancy between two models is too large. Besides, to implement relation-based" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 272 + ], + "type": "text", + "content": "distillation, previous works [8, 27, 49] design delicate modules or schemes to capture relations between different instances [8], anchors [27], patches [17], and pixels [49]. At last, the response-based KD methods are also widely investigated in previous works [8, 15, 22, 56], where the probabilistic scores are matched between two models. With the development of detection transformers (DETRs), some recently proposed KD methods [3, 7] are tailored for DETRs. DETRDistill [3] transfers the label assignment results from the teacher to student in order to stabilize the bipartite matching, while D3ETR [7] feeds the teacher queries to the student decoder and then trains the decoder using matching results from the teacher. Although these KD methods achieve decent distillation performance for medium-sized teacher models (such as ResNet-101 [20]), few of them explore the distillation from large-scale image models. In contrast to these works, we focus on distillation from large-scale models with parameters spanning from 200M to 1B. Through our method, we show that larger models lead to greater distillation gains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 307, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 482, + 451 + ], + "type": "text", + "content": "Object Detection. Object detection is a fundamental task in the computer vision field [28, 46]. Detection Transformers (DETRs) achieve great progress recently and gradually dominate object detection. With outstanding scalability, current mainstream large-scale detectors [5, 54, 58] typically rely on DETRs. To address the inefficiency issue of the original DETR [2], Deformable DETR [57] and Adamixer [13] introduce the local attention mechanism to replace the original global attention, resulting in faster model convergence. Some works [31, 43] attempt to insert the spatial prior into the object queries and lead each query to attend to a certain region. DN-DETR [26] designs a query denoising scheme to improve the stability of bipartite graph matching. Based on DN-DETR, DINO [54] presents a contrastive way for denoising training and achieves the state-of-the-art performance. Building on SoTA performance, we employ DINO as the default detector in this paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 486, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 482, + 665 + ], + "type": "text", + "content": "Large-scale Image Models. Some recent works [10, 33, 34, 41, 42, 48] scale up image models to reach 3 billion parameters, demonstrating a very powerful capacity of large-scale image models. Swin Transformer [33], built on the transformer architecture, incorporates techniques like post-norm and scaled cosine attention to enhance the stability of large vision models. Beit-3 [42], adopting the ViT [10] architecture, explores scaling up models through multi-modal pretraining. InternImage [41], utilizing a CNN base, integrates deformable convolution as a core operator to establish large-scale CNN models. After acquiring large-scale pretrained models, there has been significant interest in efficiently fine-tuning these models for downstream tasks. Notably, approaches such as VPT [23] insert learnable parameters into the input space while keeping most of the pretrained model frozen. Similarly, AdaptFormer [6] introduces lightweight learnable modules into the MLP block. Lin et al. [30] investigate the pretraining tasks that best align with the frozen setting. Different from these works, our paper aims to extract transferable knowledge from large models and use it to enhance tiny-scale models via the KD." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 194, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 194, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 194, + 127 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 140, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 140, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 131, + 140, + 482, + 176 + ], + "type": "text", + "content": "In this section, we provide a concise overview of knowledge distillation and the detection transformer in Sec. 3.1. Subsequently, we introduce the frozen teacher method in Sec. 3.2, followed by an explanation of Query Distillation in Sec. 3.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 209, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 209, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 209, + 206 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 214, + 482, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 214, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 214, + 482, + 239 + ], + "type": "text", + "content": "Knowledge Distillation (KD). Standard KD enforces the student to align its probabilistic scores " + }, + { + "bbox": [ + 131, + 214, + 482, + 239 + ], + "type": "inline_equation", + "content": "(p^{(s)})" + }, + { + "bbox": [ + 131, + 214, + 482, + 239 + ], + "type": "text", + "content": " with those of the teacher via the Kullback-Leibler (KL) divergence [21, 55]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 259, + 247, + 481, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 247, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 259, + 247, + 481, + 274 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K D} = p ^ {(t)} \\log \\left(\\frac {p ^ {(t)}}{p ^ {(s)}}\\right). \\tag {1}", + "image_path": "bc0cc160cb2a7089a6a579b4a0bf179e37bd3ed708a12b2c8487463f9d0a95a6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 280, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 280, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 280, + 482, + 327 + ], + "type": "text", + "content": "The teacher's logits " + }, + { + "bbox": [ + 131, + 280, + 482, + 327 + ], + "type": "inline_equation", + "content": "(p^{(t)})" + }, + { + "bbox": [ + 131, + 280, + 482, + 327 + ], + "type": "text", + "content": " serve as soft label, providing non-target classes not present in the ground truth labels. Despite being weak, this non-target class information significantly enhances the student's generalization. To amplify this effect, prior works typically apply a high temperature to soften the " + }, + { + "bbox": [ + 131, + 280, + 482, + 327 + ], + "type": "inline_equation", + "content": "p^{(t)}" + }, + { + "bbox": [ + 131, + 280, + 482, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 328, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 328, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 328, + 482, + 399 + ], + "type": "text", + "content": "Detection Transformers (DETRs). In DETRs, the backbone extracts image features, which are then inputted into the transformer encoder along with positional encodings. The enhanced image features serve as the key and value, providing contextual information for the subsequent transformer decoder. The input for the decoder is N embeddings, known as object queries, denoted as " + }, + { + "bbox": [ + 131, + 328, + 482, + 399 + ], + "type": "inline_equation", + "content": "Q = \\{q_0, q_1, \\dots, q_{N-1}\\}" + }, + { + "bbox": [ + 131, + 328, + 482, + 399 + ], + "type": "text", + "content": ". Each query will be decoded into class probability " + }, + { + "bbox": [ + 131, + 328, + 482, + 399 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 328, + 482, + 399 + ], + "type": "text", + "content": " and box coordinate " + }, + { + "bbox": [ + 131, + 328, + 482, + 399 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 131, + 328, + 482, + 399 + ], + "type": "text", + "content": " via the transformer decoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 418, + 259, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 259, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 259, + 429 + ], + "type": "text", + "content": "3.2 Frozen Teacher Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 482, + 521 + ], + "type": "text", + "content": "Inspired by the NLP domain, where scaling up language models achieves remarkable success [9, 12], some recent works scale up vision models to more than 1 billion parameters and set new records across various vision tasks [33,41,42,48]. It is natural to explore the possibility of transferring the knowledge from these large models to more compact ones using the Knowledge Distillation (KD) technique. However, empirical observations (Fig. 1) suggest that as the model size of the teacher increases, the discrepancy between the teacher and student becomes more severe, limiting the effectiveness of the KD." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 522, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we observed that when fine-tuning large-scale pretrained models on a specific dataset to create the teacher model, freezing the backbone can endow the teacher with richer transferable knowledge compared to fully fine-tuning. The overall framework of our DLIM-Det is illustrated in Fig. 3. The pretraining-then-finetuning paradigm is the cornerstone for modern neural networks [18, 19, 45], especially for large-scale models. Pretraining on massive data is essential for achieving top performance in the downstream tasks [41]. Although fully fine-tuning the entire model may enhance task-specific performance, it also heightens the risk of overfitting and diminishes the model's generalizability. We denote the teacher produced by the frozen setting as the frozen teacher. This frozen teacher maintains the generalizability of the large-scale models. As evidenced by the Average Recall (AR) metrics on LVIS labels shown in Fig. 2, the frozen teacher exhibits an increased ability to recognize a broader array of non-target objects" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 468, + 244 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 468, + 244 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 468, + 244 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 468, + 244 + ], + "type": "image", + "image_path": "3392f4f0a86addaaf5b05910f262f3bf8e4ae3adc4bb00ce93b59776da8fc22e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 246, + 482, + 280 + ], + "lines": [ + { + "bbox": [ + 132, + 246, + 482, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 482, + 280 + ], + "type": "text", + "content": "Fig. 3: The framework of our DLIM-Det. Freezing the backbone when fine-tuning large-sized models on a specific dataset to create the teacher. The alignment between object queries of the teacher and student is conducted based on positions and relations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 225, + 304, + 386, + 344 + ], + "blocks": [ + { + "bbox": [ + 202, + 293, + 411, + 304 + ], + "lines": [ + { + "bbox": [ + 202, + 293, + 411, + 304 + ], + "spans": [ + { + "bbox": [ + 202, + 293, + 411, + 304 + ], + "type": "text", + "content": "Table 1: The query density with different IoU thresholds." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 225, + 304, + 386, + 344 + ], + "lines": [ + { + "bbox": [ + 225, + 304, + 386, + 344 + ], + "spans": [ + { + "bbox": [ + 225, + 304, + 386, + 344 + ], + "type": "table", + "html": "
MethodsIoU: 0.5IoU: 0.7
Swin-LIntern-LSwin-LIntern-L
Full finetuning38.133.615.213.0
Frozen21.321.87.58.3
", + "image_path": "eef5c0334ad2072d7d74a4776cb78a2b01afe00d52da10165e06176581e1252b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 352, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 481, + 376 + ], + "type": "text", + "content": "while avoiding excessive specialization on the ground truths. Enhanced generalizability endows the teacher with specific characteristics that facilitate the distillation process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 377, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 592 + ], + "type": "text", + "content": "In the following, we investigate how these distinctive characteristics enhance the efficiency of the distillation process. When equipped with DETRs, we compare the spatial distributions of object queries under both the frozen and fully fine-tuned settings. We observe that the frozen teacher locates object queries more dispersely and sparsely than the fully fine-tuned teacher. To quantify the dispersion of query spatial distribution, we introduce a metric called Query Density. Query density represents the number of object queries with an Intersection over Union (IoU) greater than " + }, + { + "bbox": [ + 130, + 377, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 377, + 482, + 592 + ], + "type": "text", + "content": " surrounding each ground truth. We average the query density over all GT objects in the COCO 2017 validation set [29], with " + }, + { + "bbox": [ + 130, + 377, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 377, + 482, + 592 + ], + "type": "text", + "content": " set to 0.5 and 0.7, respectively. The results are presented in Tab. 1. As results show, the backbone-frozen strategy significantly affects the query density, reducing it by at least " + }, + { + "bbox": [ + 130, + 377, + 482, + 592 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 130, + 377, + 482, + 592 + ], + "type": "text", + "content": ". The fully fine-tuned teacher exhibits high query density values, and each GT object is surrounded by more than 30 queries with IoUs larger than 0.5. Among these compact queries, only one query should predict the object, while the others should predict no object. One example is visualized in Fig. 5. The detector needs to model their relations, identifying the optimal localized query with high confidence and suppressing the others with low confidence. In the case of sparse query locations, an instance is surrounded by fewer object queries, making their relations much easier to model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 483, + 665 + ], + "type": "text", + "content": "When using the fully fine-tuned teacher, the student model needs to allocate unnecessary capacity to model complex query relations in order to mimic the compact query locations of the teacher. In contrast, the sparse query locations provided by the frozen teacher are easier for the student to learn, reducing the learning burden. To fully leverage the sparse query of the frozen teacher, we design a Query Distillation method, which is described in Sec. 3.3." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 239, + 194 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 239, + 194 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 239, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 239, + 194 + ], + "type": "image", + "image_path": "6354959c5867e932eca3ac2f94820c2aeeec9ae999a93d7f0bff5f2a72c7ab3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 196, + 227, + 205 + ], + "lines": [ + { + "bbox": [ + 144, + 196, + 227, + 205 + ], + "spans": [ + { + "bbox": [ + 144, + 196, + 227, + 205 + ], + "type": "text", + "content": "(a) Comparing four teachers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 122, + 297, + 156 + ], + "blocks": [ + { + "bbox": [ + 253, + 114, + 297, + 121 + ], + "lines": [ + { + "bbox": [ + 253, + 114, + 297, + 121 + ], + "spans": [ + { + "bbox": [ + 253, + 114, + 297, + 121 + ], + "type": "text", + "content": "Input Images" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 253, + 122, + 297, + 156 + ], + "lines": [ + { + "bbox": [ + 253, + 122, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 253, + 122, + 297, + 156 + ], + "type": "image", + "image_path": "b1b20ee9aad885e188f94c9cf6d38c82349a675b13c69d02b7e0b5e70fe9e97b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 253, + 158, + 296, + 193 + ], + "blocks": [ + { + "bbox": [ + 253, + 158, + 296, + 193 + ], + "lines": [ + { + "bbox": [ + 253, + 158, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 253, + 158, + 296, + 193 + ], + "type": "image", + "image_path": "bf6bcba23cbe5bcda067c629234c709461f6e58f39e23703a57b2986968fedf2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 205, + 481, + 260 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 481, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 481, + 260 + ], + "type": "text", + "content": "Fig. 4: Comparison of different teacher models based on mutual information. (a) Multiple teacher models and their mutual information concerning the input " + }, + { + "bbox": [ + 130, + 205, + 481, + 260 + ], + "type": "inline_equation", + "content": "(I(X;F))" + }, + { + "bbox": [ + 130, + 205, + 481, + 260 + ], + "type": "text", + "content": " and target " + }, + { + "bbox": [ + 130, + 205, + 481, + 260 + ], + "type": "inline_equation", + "content": "(I(Y;F))" + }, + { + "bbox": [ + 130, + 205, + 481, + 260 + ], + "type": "text", + "content": ". The blue lines represent the process of searching for the model whose representation maximizes " + }, + { + "bbox": [ + 130, + 205, + 481, + 260 + ], + "type": "inline_equation", + "content": "I(X;F) + I(Y;F)" + }, + { + "bbox": [ + 130, + 205, + 481, + 260 + ], + "type": "text", + "content": ". \"Full ft.\" denotes full fine-tuning. (b) The reconstructed image qualitatively reflects " + }, + { + "bbox": [ + 130, + 205, + 481, + 260 + ], + "type": "inline_equation", + "content": "I(X;F)" + }, + { + "bbox": [ + 130, + 205, + 481, + 260 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 299, + 122, + 342, + 156 + ], + "blocks": [ + { + "bbox": [ + 299, + 114, + 339, + 121 + ], + "lines": [ + { + "bbox": [ + 299, + 114, + 339, + 121 + ], + "spans": [ + { + "bbox": [ + 299, + 114, + 339, + 121 + ], + "type": "text", + "content": "Full ft. Swin Base" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 122, + 342, + 156 + ], + "lines": [ + { + "bbox": [ + 299, + 122, + 342, + 156 + ], + "spans": [ + { + "bbox": [ + 299, + 122, + 342, + 156 + ], + "type": "image", + "image_path": "5ebb3b414094d4c68475fb0a664de49691afcbb1e9971a36a6d862fa42e8e289.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 299, + 158, + 342, + 193 + ], + "blocks": [ + { + "bbox": [ + 299, + 158, + 342, + 193 + ], + "lines": [ + { + "bbox": [ + 299, + 158, + 342, + 193 + ], + "spans": [ + { + "bbox": [ + 299, + 158, + 342, + 193 + ], + "type": "image", + "image_path": "658a95210b226657247c29a7f847e8f092caf40756d75f1ff326d32cdb89c0f2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 195, + 424, + 204 + ], + "lines": [ + { + "bbox": [ + 305, + 195, + 424, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 195, + 424, + 204 + ], + "type": "text", + "content": "(b) Visualization of reconstructed images." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 343, + 122, + 386, + 156 + ], + "blocks": [ + { + "bbox": [ + 345, + 114, + 386, + 121 + ], + "lines": [ + { + "bbox": [ + 345, + 114, + 386, + 121 + ], + "spans": [ + { + "bbox": [ + 345, + 114, + 386, + 121 + ], + "type": "text", + "content": "Frozen Swin Base" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 122, + 386, + 156 + ], + "lines": [ + { + "bbox": [ + 343, + 122, + 386, + 156 + ], + "spans": [ + { + "bbox": [ + 343, + 122, + 386, + 156 + ], + "type": "image", + "image_path": "7cbee10677cb09278bcfc387a01642e2d492d6c382d6ee751e9a64f1d88b495d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 343, + 158, + 386, + 193 + ], + "blocks": [ + { + "bbox": [ + 343, + 158, + 386, + 193 + ], + "lines": [ + { + "bbox": [ + 343, + 158, + 386, + 193 + ], + "spans": [ + { + "bbox": [ + 343, + 158, + 386, + 193 + ], + "type": "image", + "image_path": "46ad419ea7f7b5e2f9298a7a50e1721f145b171da41be0bf1af21f4a7062eda3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 388, + 122, + 432, + 156 + ], + "blocks": [ + { + "bbox": [ + 389, + 114, + 429, + 121 + ], + "lines": [ + { + "bbox": [ + 389, + 114, + 429, + 121 + ], + "spans": [ + { + "bbox": [ + 389, + 114, + 429, + 121 + ], + "type": "text", + "content": "Full ft. Swin Large" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 122, + 432, + 156 + ], + "lines": [ + { + "bbox": [ + 388, + 122, + 432, + 156 + ], + "spans": [ + { + "bbox": [ + 388, + 122, + 432, + 156 + ], + "type": "image", + "image_path": "714a0c20640896fcaee6b9c6a61e84988747e22269c8c90446d33614d2b7eaf7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 388, + 158, + 432, + 193 + ], + "blocks": [ + { + "bbox": [ + 388, + 158, + 432, + 193 + ], + "lines": [ + { + "bbox": [ + 388, + 158, + 432, + 193 + ], + "spans": [ + { + "bbox": [ + 388, + 158, + 432, + 193 + ], + "type": "image", + "image_path": "f2d5547af3cbdb2d9c7ea5d47c20442454cd3b65b8982495317ab84518092c66.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 434, + 122, + 477, + 156 + ], + "blocks": [ + { + "bbox": [ + 435, + 114, + 476, + 121 + ], + "lines": [ + { + "bbox": [ + 435, + 114, + 476, + 121 + ], + "spans": [ + { + "bbox": [ + 435, + 114, + 476, + 121 + ], + "type": "text", + "content": "Frozen Swin Large" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 434, + 122, + 477, + 156 + ], + "lines": [ + { + "bbox": [ + 434, + 122, + 477, + 156 + ], + "spans": [ + { + "bbox": [ + 434, + 122, + 477, + 156 + ], + "type": "image", + "image_path": "77719ac267537b1b8b50deb0593299fdce25000d177c55f93b0f748a126ba929.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 434, + 158, + 477, + 193 + ], + "blocks": [ + { + "bbox": [ + 434, + 158, + 477, + 193 + ], + "lines": [ + { + "bbox": [ + 434, + 158, + 477, + 193 + ], + "spans": [ + { + "bbox": [ + 434, + 158, + 477, + 193 + ], + "type": "image", + "image_path": "98aebe2e1a2640072dd5242b9b6bada62a6062edad234adbd6a9c5d489250c5b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 276, + 482, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 482, + 336 + ], + "type": "text", + "content": "Taking into account the aspect of generalization and drawing inspiration from prior research [40], which introduced an optimal teacher model selection algorithm based on mutual information, we also substantiate the superiority of the frozen teacher from the perspective of the mutual information principle. Specifically, the entropy of teacher's representation " + }, + { + "bbox": [ + 130, + 276, + 482, + 336 + ], + "type": "inline_equation", + "content": "(H(F))" + }, + { + "bbox": [ + 130, + 276, + 482, + 336 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 235, + 351, + 481, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 351, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 235, + 351, + 481, + 365 + ], + "type": "interline_equation", + "content": "H (F) = I (X; F) + I (Y; F) + Z, \\tag {2}", + "image_path": "05b77a8e31c7490f6ca0a6e02c38aacafa54190a3a78560f562bbfc38ec25fab.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "inline_equation", + "content": "I(X;F)" + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "inline_equation", + "content": "I(Y;F)" + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "text", + "content": " represent task-relevant information regarding the input image " + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "text", + "content": " and ground truth target " + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "text", + "content": " refers to nuisance, which is task-irrelevant information. Following standard practice [40, 44], we employ the detection performance (e.g. mAP) to quantify " + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "inline_equation", + "content": "I(Y;F)" + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "text", + "content": ". To measure the mutual information concerning the input " + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "inline_equation", + "content": "(I(X;F))" + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "text", + "content": ", we feed the intermediate features into a decoder network (typically, U-Net [38]) to reconstruct the input image " + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "text", + "content": ". The average reconstruction loss is used to quantify " + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "inline_equation", + "content": "I(X;F)" + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "text", + "content": ". Different values of " + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "inline_equation", + "content": "I(X;F)" + }, + { + "bbox": [ + 130, + 380, + 482, + 489 + ], + "type": "text", + "content": " can be clearly distinguished based on the quality of reconstructed images, as illustrated in Fig. 4b. The theoretical proofs and training loss curves can be found in the Appendix." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "text", + "content": "After defining " + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "inline_equation", + "content": "I(X;F)" + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "inline_equation", + "content": "I(Y;F)" + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "text", + "content": ", we compare base-scale and large-scale models under both full fine-tuning and frozen settings, with results depicted in Fig. 4a, where mutual information " + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "inline_equation", + "content": "(I(X;F)" + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "inline_equation", + "content": "I(Y;F))" + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "text", + "content": " is normalized to ensure the scale matching. Among the four teacher models, the frozen Swin-L attains the maximum of task relevant information " + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "inline_equation", + "content": "I(X;F) + I(Y;F)" + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "text", + "content": ", achieving a delicate trade-off between " + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "inline_equation", + "content": "I(X;F)" + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "inline_equation", + "content": "I(Y;F)" + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "text", + "content": ". A model with high values of " + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "inline_equation", + "content": "I(X;F) + I(Y;F)" + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "text", + "content": " can effectively suppress nuisance " + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "text", + "content": " and serves as an informative teacher. Though Base-scale models exhibit higher values of " + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "inline_equation", + "content": "I(X;F)" + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "text", + "content": " compared to Large-scale ones, they lag in terms of " + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "inline_equation", + "content": "I(Y;F)" + }, + { + "bbox": [ + 130, + 491, + 483, + 635 + ], + "type": "text", + "content": ", which may result in inferior KD performance. Empirical experiments (Tab. 3) support our assumption: the frozen Swin-L achieves the best distillation performance among the four teacher models. To conclude, the teacher model selection algorithm [40], grounded in mutual information, further supports our frozen teacher method." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 448, + 102 + ], + "type": "text", + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 653, + 400, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 400, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 400, + 665 + ], + "type": "text", + "content": "1 The features are from the last layer of Transformer Encoder in DETRs." + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 234, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 234, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 234, + 128 + ], + "type": "text", + "content": "3.3 Query Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "content": "Query Distillation aligns the spatial positions and pair-wise relations of queries between the teacher and student, consisting of Query Relation Distill and Query Relation Distill. Initially, we establish the one-to-one matching for object queries of two models, where the cost matrix is computed based on classification and localization costs. Given the cost matrix, Hungarian Matching is used to solve the graph matching, yielding the optimal permutation function " + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}" + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 230, + 218, + 482, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 218, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 230, + 218, + 482, + 251 + ], + "type": "interline_equation", + "content": "\\hat {\\sigma} = \\underset {\\sigma \\in \\mathfrak {S} _ {N}} {\\arg \\min } \\sum_ {i} ^ {N} \\mathcal {L} _ {\\text {m a t c h}} \\left(q _ {i} ^ {(t)}, q _ {\\sigma (i)} ^ {(s)}\\right), \\tag {3}", + "image_path": "5515b4ea40e834f7bf147b19468e8b3c16ab627f2897600b472ec414ca192e25.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "text", + "content": " equals the number of predicted objects. Among all permutations of " + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "text", + "content": " elements " + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "inline_equation", + "content": "\\mathfrak{S}_N" + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}" + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "text", + "content": " is the one with the lowest cost. Based on the class probability " + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "text", + "content": " and box coordinate " + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "text", + "content": " decoded from each object query, we define the cost between the teacher query " + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "inline_equation", + "content": "q^{(t)}" + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "text", + "content": " and student " + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "inline_equation", + "content": "q^{(s)}" + }, + { + "bbox": [ + 131, + 259, + 480, + 307 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 315, + 481, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 341 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a t c h}} \\left(q ^ {(t)}, q ^ {(s)}\\right) = \\alpha_ {1} \\mathcal {L} _ {K L} \\left(p ^ {(t)}, p ^ {(s)}\\right) + \\alpha_ {2} \\mathcal {L} _ {L 1 L o s s} \\left(b ^ {(t)}, b ^ {(s)}\\right) + \\alpha_ {3} \\mathcal {L} _ {G I o U} \\left(b ^ {(t)}, b ^ {(s)}\\right), \\tag {4}", + "image_path": "0070b435ca5d0d1406267a7025eb4f582a490fa84d348b867a8bd43086fe21a5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 342, + 318, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 342, + 318, + 353 + ], + "spans": [ + { + "bbox": [ + 131, + 342, + 318, + 353 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 342, + 318, + 353 + ], + "type": "inline_equation", + "content": "\\alpha_{1},\\alpha_{2},\\alpha_{3}" + }, + { + "bbox": [ + 131, + 342, + 318, + 353 + ], + "type": "text", + "content": " are set to 2,5,2 empirically." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": "Query Position Distill. Instead of matching the probabilistic scores of two models, we observe that the spatial position of object queries can serve as informative knowledge. The position of an object query refers to the bounding box decoded from the query, represented by a vector: " + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "inline_equation", + "content": "b \\in [x, y, w, h]" + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": " denote the center coordinates, while " + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": " represent the width and height, respectively. Given the matching " + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": " between object queries of two models, we formulate the procedure of Query Position Distill as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 192, + 444, + 482, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 444, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 192, + 444, + 482, + 476 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {Q P D} = \\sum_ {i} ^ {N} \\beta_ {1} \\mathcal {L} _ {L 1} \\left(b _ {i} ^ {(t)}, b _ {\\sigma (i)} ^ {(s)}\\right) + \\beta_ {2} \\mathcal {L} _ {G I o U} \\left(b _ {i} ^ {(t)}, b _ {\\sigma (i)} ^ {(s)}\\right), \\tag {5}", + "image_path": "47864f83c261a76781ab99316707ad33600bd21d604f5207f8299396733e109c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 486, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 486, + 481, + 521 + ], + "type": "inline_equation", + "content": "\\beta_{1},\\beta_{2}" + }, + { + "bbox": [ + 130, + 486, + 481, + 521 + ], + "type": "text", + "content": " are set to 5 and 2, empirically. Through the thorough experiments (listed in Tab. 8b), we empirically find that besides the foreground queries, even poorly localized teacher queries that cover background regions still contain valuable knowledge to distill." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 605 + ], + "type": "text", + "content": "Due to the one-to-one label assignment, only a small subset of object queries receives localization supervision in each image. For example, on the COCO dataset [29], which typically contains an average of 7 instances per image, only 7 out of the 900 queries are involved in the regression loss. This sparsity of localization supervision impairs the training of detection transformers. By the proposed Query Position Distill, the spatial positions of teacher's queries serve as additional supervisions to provide richer information than the sparse ground truth labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Query Relation Distill. In previous works, relation-based KD methods have been widely investigated. These methods aim to capture various relations between different entities such as instances [8, 22], pixels [49], anchors [27] or patches [17]. To model these relations, intricate modules, e.g., GcBlock [1], and complex relation mechanisms, e.g. Rank Mimicking [27], have been designed, which not only lack interpretability but" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 189, + 117, + 425, + 209 + ], + "blocks": [ + { + "bbox": [ + 189, + 117, + 425, + 209 + ], + "lines": [ + { + "bbox": [ + 189, + 117, + 425, + 209 + ], + "spans": [ + { + "bbox": [ + 189, + 117, + 425, + 209 + ], + "type": "image", + "image_path": "b27d4468abbc83abfcc0c3a1d77e4e58efdd09cd6ebb3a1e561480dcdb34e7ec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 212, + 482, + 234 + ], + "lines": [ + { + "bbox": [ + 131, + 212, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 131, + 212, + 482, + 234 + ], + "type": "text", + "content": "Fig.5: An example to show the position, score and relation of object queries. Query relations are obtained from self-attention. Only 3 queries are visualized for a clear presentation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 256, + 482, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 313 + ], + "type": "text", + "content": "also introduce additional training costs. In contrast, the transformer decoder of DETRs explicitly models the pair-wise relations between object queries through the self-attention. We denote by " + }, + { + "bbox": [ + 130, + 256, + 482, + 313 + ], + "type": "inline_equation", + "content": "A = [a_{ij}]_{N \\times N}" + }, + { + "bbox": [ + 130, + 256, + 482, + 313 + ], + "type": "text", + "content": " the attention map from the self-attention, where " + }, + { + "bbox": [ + 130, + 256, + 482, + 313 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 256, + 482, + 313 + ], + "type": "text", + "content": " denotes the number of queries. L1 Loss is employed to narrow the distance between " + }, + { + "bbox": [ + 130, + 256, + 482, + 313 + ], + "type": "inline_equation", + "content": "A^{(t)}" + }, + { + "bbox": [ + 130, + 256, + 482, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 256, + 482, + 313 + ], + "type": "inline_equation", + "content": "A^{(s)}" + }, + { + "bbox": [ + 130, + 256, + 482, + 313 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 223, + 314, + 481, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 314, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 223, + 314, + 481, + 346 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {Q R D} = \\frac {1}{N} \\sum_ {i} ^ {N} \\sum_ {j} ^ {N} \\left| A _ {i, j} ^ {(t)} - A _ {\\sigma (i), \\sigma (j)} ^ {(s)} \\right|, \\tag {6}", + "image_path": "2d79ecce8726c4f7dbfa86239e9d15c4ef52ca2a37b4224aa927a316188d530c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 350, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 397 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 350, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 350, + 482, + 397 + ], + "type": "text", + "content": " denotes the permutation function to match the queries between the teacher and student. Query relations are critical for end-to-end detection. As shown in Fig. 5, precise query relations allow the best localized query to pop out with high classification scores, while the others will be suppressed with low scores." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 399, + 294, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 399, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 146, + 399, + 294, + 410 + ], + "type": "text", + "content": "The overall training loss " + }, + { + "bbox": [ + 146, + 399, + 294, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{overall}}" + }, + { + "bbox": [ + 146, + 399, + 294, + 410 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 418, + 481, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 418, + 481, + 431 + ], + "spans": [ + { + "bbox": [ + 216, + 418, + 481, + 431 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o v e r a l l}} = \\gamma_ {1} \\mathcal {L} _ {\\text {t a s k}} + \\gamma_ {2} \\mathcal {L} _ {Q P D} + \\gamma_ {3} \\mathcal {L} _ {Q R D}, \\tag {7}", + "image_path": "a89ffdbef09a3e7f7ce2521b996ec46209f30f6731f3d1a43d8149270424f746.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 438, + 320, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 438, + 320, + 450 + ], + "spans": [ + { + "bbox": [ + 131, + 438, + 320, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 438, + 320, + 450 + ], + "type": "inline_equation", + "content": "\\gamma_{1},\\gamma_{2},\\gamma_{3}" + }, + { + "bbox": [ + 131, + 438, + 320, + 450 + ], + "type": "text", + "content": " are set to 1, 2, 1, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 451, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 510 + ], + "type": "text", + "content": "We also find that the parameters of the detection head in the teacher model serve as a source of knowledge. Even for heterogeneous DETRs, the basic blocks in their heads are consistent, including attention and feed forward network (FFN). Through empirical investigations, we prove that inheriting the parameters of these basic blocks from the teacher head to student head is convenient and beneficial." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 527, + 218, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 527, + 218, + 541 + ], + "spans": [ + { + "bbox": [ + 131, + 527, + 218, + 541 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 551, + 250, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 551, + 250, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 551, + 250, + 563 + ], + "type": "text", + "content": "4.1 Experimental settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 570, + 491, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 491, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 491, + 641 + ], + "type": "text", + "content": "Datasets and Backbone. We conduct the experiments on the MS COCO 2017 dataset [29], which is a commonly used object detection dataset. The dataset is split into the train2017 set with 118k images and the val2017 set with 5k images. All models are trained on the train2017 set, and evaluated on the val2017 set. Results are reported with two backbones: Swin Transformer [34] and InternImage [41]. We employ the COCO-style average precision (mAP) as the metric. DINO [54] is the default detector." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "text", + "content": "Implementation details. All models, except for the Intern-Huge model, are trained on " + }, + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "inline_equation", + "content": "8 \\times 32\\mathrm{G}" + }, + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "text", + "content": " V100 GPUs with a batch size of 16. In contrast, the Intern-Huge models are" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 448, + 102 + ], + "type": "text", + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 171, + 125, + 441, + 212 + ], + "blocks": [ + { + "bbox": [ + 228, + 114, + 384, + 125 + ], + "lines": [ + { + "bbox": [ + 228, + 114, + 384, + 125 + ], + "spans": [ + { + "bbox": [ + 228, + 114, + 384, + 125 + ], + "type": "text", + "content": "Table 2: Effectiveness of each component." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 171, + 125, + 441, + 212 + ], + "lines": [ + { + "bbox": [ + 171, + 125, + 441, + 212 + ], + "spans": [ + { + "bbox": [ + 171, + 125, + 441, + 212 + ], + "type": "table", + "html": "
MethodsSwin TransformerInternImage
mAP\\( AP_{50} \\)\\( AP_{75} \\)mAP\\( AP_{50} \\)\\( AP_{75} \\)
Teacher: Large Scale58.177.163.556.975.861.9
Student: Tiny Scale51.369.056.053.271.558.0
Query DistillFrozen TeacherInherit Head
✓51.6 (+0.3)69.556.454.2 (+1.0)72.559.2
✓✓52.2 (+0.9)70.157.054.7 (+1.5)72.959.8
✓✓53.7 (+2.4)71.758.855.2 (+2.0)73.760.3
✓✓✓54.4 (+3.1)72.159.455.9 (+2.7)74.161.4
", + "image_path": "d567d56e780285c5eefd84c92336e1f8b91364bb01186da22fe23bff4ffafa94.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 223, + 244, + 390, + 296 + ], + "blocks": [ + { + "bbox": [ + 132, + 220, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 132, + 220, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 482, + 243 + ], + "type": "text", + "content": "Table 3: Distillation performance comparison between Large- and Base-scale teacher models. Numbers in brackets refer to distillation gains against the original student (Swin-T, 51.3 mAP)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 223, + 244, + 390, + 296 + ], + "lines": [ + { + "bbox": [ + 223, + 244, + 390, + 296 + ], + "spans": [ + { + "bbox": [ + 223, + 244, + 390, + 296 + ], + "type": "table", + "html": "
TeacherFreeze BackboneDistilled Student
mAPAP50AP75
Large-scale✓54.4 (+3.1)72.1 (+3.1)59.4 (+3.4)
Base-scale✓53.5 (+2.2)71.5 (+2.5)58.5 (+2.5)
Base-scale×53.1 (+1.8)71.1 (+2.1)58.0 (+2.0)
", + "image_path": "04d18e14147b14a15caa3e6a679b05fe590f03213812e4157a88ed50afc9678f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 312, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 361 + ], + "type": "text", + "content": "trained on 80GB A100 GPUs. Student models are trained with 12 epochs with AdamW as the optimizer. The initial learning rate is set to 0.0001, which is divided by 10 at the 11th epoch. Following the default practice of DINO [54], the number of object query is set to 900." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 380, + 230, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 230, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 230, + 392 + ], + "type": "text", + "content": "4.2 Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 401, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 426 + ], + "type": "text", + "content": "In ablation studies, unless otherwise stated, we adopt the tiny-scale models with around 30M parameters as the student." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 483, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 483, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 483, + 617 + ], + "type": "text", + "content": "Effects of each component. In Tab. 2, we verify the effectiveness of each component, including the frozen teacher method, Query Distillation, and head inheriting. When facing large-scale teacher models, Query Distillation produces a modest increase of " + }, + { + "bbox": [ + 130, + 426, + 483, + 617 + ], + "type": "inline_equation", + "content": "+0.3 / 1.0" + }, + { + "bbox": [ + 130, + 426, + 483, + 617 + ], + "type": "text", + "content": " mAP for Swin/Intern, respectively. Subsequently, we replace the fully fine-tuned teacher with the frozen teacher and perform Query Distillation again. The results demonstrate that the frozen teacher significantly amplifies the distillation gains. Specifically, compared to the fully fine-tuned teacher, the frozen teacher yields an additional increase of " + }, + { + "bbox": [ + 130, + 426, + 483, + 617 + ], + "type": "inline_equation", + "content": "+0.6 / 0.5\\%" + }, + { + "bbox": [ + 130, + 426, + 483, + 617 + ], + "type": "text", + "content": " for Swin/Intern models, respectively. Moreover, inheriting the transformer's parameters of the teacher's head produces noticeable improvements, suggesting that the head parameters of the teacher contain some general knowledge that is somewhat independent of the backbone structure. Building on head parameter inheriting, the frozen teacher outperforms the fully fine-tuned teacher by " + }, + { + "bbox": [ + 130, + 426, + 483, + 617 + ], + "type": "inline_equation", + "content": "+0.7 / 0.7\\%" + }, + { + "bbox": [ + 130, + 426, + 483, + 617 + ], + "type": "text", + "content": " for Swin/Intern models. These experimental findings validate that the frozen teacher is easier to distill than the previously employed fully fine-tuned teacher. Finally, by combining the three components, DLIM-Det, achieves significant improvements, surpassing the student baseline by " + }, + { + "bbox": [ + 130, + 426, + 483, + 617 + ], + "type": "inline_equation", + "content": "+3.1 / 2.7\\%" + }, + { + "bbox": [ + 130, + 426, + 483, + 617 + ], + "type": "text", + "content": " for the Swin/Intern models, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": "Distillation on medium- and large-sized teacher models. In Tab. 3, we compare the distillation performance under Base- and Large-scale teacher models. With the help of the frozen teacher method, a large-scale teacher surpasses its base-scale counterpart by " + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "inline_equation", + "content": "+0.9" + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": " mAP. These results affirm the core premise of this paper: larger teacher models lead" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 220, + 127, + 392, + 190 + ], + "blocks": [ + { + "bbox": [ + 194, + 114, + 418, + 125 + ], + "lines": [ + { + "bbox": [ + 194, + 114, + 418, + 125 + ], + "spans": [ + { + "bbox": [ + 194, + 114, + 418, + 125 + ], + "type": "text", + "content": "Table 4: The frozen setting affects the detection performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 220, + 127, + 392, + 190 + ], + "lines": [ + { + "bbox": [ + 220, + 127, + 392, + 190 + ], + "spans": [ + { + "bbox": [ + 220, + 127, + 392, + 190 + ], + "type": "table", + "html": "
MethodsFreeze BackboneSchedulemAP\\( AP_{50} \\)\\( AP_{75} \\)
Swin-Large×3x58.177.163.5
✓3x56.6 (-1.5)74.961.9
Intern-Large×2x57.876.563.1
✓2x55.8 (-2.0)74.461.0
", + "image_path": "6748c80de52614409e1056b370a16d770a4b496a658eb3c5fd600cced7e39ea6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 205, + 220, + 405, + 316 + ], + "blocks": [ + { + "bbox": [ + 131, + 194, + 479, + 216 + ], + "lines": [ + { + "bbox": [ + 131, + 194, + 479, + 216 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 479, + 216 + ], + "type": "text", + "content": "Table 5: Applying different training schedules to the Frozen and Full ft. settings as to obtain teachers with identical performance. The KD contains Query Distillation and the inheriting head." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 205, + 220, + 405, + 316 + ], + "lines": [ + { + "bbox": [ + 205, + 220, + 405, + 316 + ], + "spans": [ + { + "bbox": [ + 205, + 220, + 405, + 316 + ], + "type": "table", + "html": "
MethodsFreeze BackboneSchedulemAP\\( AP_{50} \\)\\( AP_{75} \\)
Teacher:
T1: Swin-Large×1x56.975.662.2
T2: Swin-Large✓3x56.774.961.9
Student:
Swin-Tiny-1x51.369.056.0
KD from T1-1x53.1 (+1.8)70.957.9
KD from T2-1x54.4 (+3.1)72.159.4
", + "image_path": "d48f63c7b9048b56a1acc80e90c815fd4f32ba0b355eefced1ec2bcb35ecc5ee.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 205, + 346, + 405, + 415 + ], + "blocks": [ + { + "bbox": [ + 131, + 325, + 481, + 346 + ], + "lines": [ + { + "bbox": [ + 131, + 325, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 325, + 481, + 346 + ], + "type": "text", + "content": "Table 6: Effectiveness of Query Position Distill (QPD) and Query Relation Distill (QRD). \"T/S\" denotes the Teacher/Student." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 205, + 346, + 405, + 415 + ], + "lines": [ + { + "bbox": [ + 205, + 346, + 405, + 415 + ], + "spans": [ + { + "bbox": [ + 205, + 346, + 405, + 415 + ], + "type": "table", + "html": "
MethodsInternImageSwin Transformer
mAPAP50AP75mAPAP50AP75
T: Base54.673.659.355.874.460.7
S: Tiny53.271.558.051.369.056.0
+ QPD54.2 (+1.0)72.459.151.6 (+0.3)69.556.2
+ QRD54.2 (+1.0)72.959.151.8 (+0.5)70.256.5
+ QPD & QRD54.5 (+1.3)73.259.552.1 (+0.8)69.956.7
", + "image_path": "9adce99e3ee610a82dda73c27bccc392c7056c4ed1f757f9431439b22ddd6c06.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 425, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 480, + 460 + ], + "type": "text", + "content": "to greater distillation gains. Additionally, the frozen teacher method generalizes well to base-scale models. When using the base-scale model as a teacher, the frozen teacher surpasses the fully fine-tuned teacher by " + }, + { + "bbox": [ + 130, + 425, + 480, + 460 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 425, + 480, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 461, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 605 + ], + "type": "text", + "content": "Fair comparison between the fully fine-tuned teacher and frozen teacher. Freezing the backbone typically results in performance degradation to some extent compared to fully fine-tuning. As shown in Tab. 4, under the same training schedule, the frozen teacher lags behind the fully fine-tuned teacher by around " + }, + { + "bbox": [ + 130, + 461, + 481, + 605 + ], + "type": "inline_equation", + "content": "2\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 461, + 481, + 605 + ], + "type": "text", + "content": ". One might argue that, owing to lower detection performance, the frozen teacher narrows the gap between the teacher and student, thereby attaining better distillation results. In Tab. 5, we conduct experiments to challenge this perspective. Concretely, we assign long and short training schedules to the frozen and fully fine-tuned teachers, resulting in teachers with nearly identical performance. By ensuring the same gap between the teacher and student, we find that the frozen teacher surpasses the fully fine-tuned teacher by a large margin, +1.3 mAP. These results highlight the intrinsic advantages of the frozen teacher approach, such as sparse query locations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Effectiveness of the Query Distillation. The Query Distillation is specifically designed for detection transformers, including Query Position Distill (QPD) and Query Relation Distill (QRD). We ablate each of them step by step in Tab. 6. Here, we employ the medium-sized models as the teacher, to exclude the effects of the large gap between the teacher and student. Besides the experiments on the advanced network architecture, such" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 243, + 137, + 367, + 202 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 136 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 136 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 136 + ], + "type": "text", + "content": "Table 7: Applying Query Position Distill (QPD) and Query Relation Distill (QRD) to classical backbones (e.g. ResNet [20])." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 243, + 137, + 367, + 202 + ], + "lines": [ + { + "bbox": [ + 243, + 137, + 367, + 202 + ], + "spans": [ + { + "bbox": [ + 243, + 137, + 367, + 202 + ], + "type": "table", + "html": "
MethodsResnet
mAP\\( AP_{50} \\)\\( AP_{75} \\)
T: ResNet10150.067.754.4
S: ResNet5048.966.453.4
+ QPD49.5 (+0.5)67.253.9
+ QRD49.5 (+0.5)67.554.1
+ QPD & QRD49.8 (+0.9)67.654.2
", + "image_path": "6f2bd9657515d9a86cfe999097b125956efc8cd6c75115d41ae7615f1c3f43b9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 170, + 238, + 284, + 309 + ], + "blocks": [ + { + "bbox": [ + 182, + 211, + 428, + 221 + ], + "lines": [ + { + "bbox": [ + 182, + 211, + 428, + 221 + ], + "spans": [ + { + "bbox": [ + 182, + 211, + 428, + 221 + ], + "type": "text", + "content": "Table 8: Study on distillation positions and teacher query selection." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 222, + 302, + 239 + ], + "lines": [ + { + "bbox": [ + 153, + 222, + 302, + 239 + ], + "spans": [ + { + "bbox": [ + 153, + 222, + 302, + 239 + ], + "type": "text", + "content": "(a) Study on distillation positions. Teacher: InternBase (54.6 mAP); Student: Interm-Tiny (53.2 mAP)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 238, + 284, + 309 + ], + "lines": [ + { + "bbox": [ + 170, + 238, + 284, + 309 + ], + "spans": [ + { + "bbox": [ + 170, + 238, + 284, + 309 + ], + "type": "table", + "html": "
KD on decoder layermAP\\( AP_{50} \\)\\( AP_{75} \\)
123456
✓✓✓54.172.659.0
✓✓✓53.872.358.8
✓✓✓✓✓✓54.573.259.5
✓54.172.659.0
✓53.572.158.3
✓52.470.957.0
", + "image_path": "69d333cbabc1d690fd3d43b926d3de24b4ed52fe4f7164495a9867a4a1ad0764.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 324, + 239, + 441, + 309 + ], + "blocks": [ + { + "bbox": [ + 310, + 222, + 457, + 239 + ], + "lines": [ + { + "bbox": [ + 310, + 222, + 457, + 239 + ], + "spans": [ + { + "bbox": [ + 310, + 222, + 457, + 239 + ], + "type": "text", + "content": "(b) Study on teacher query selection. FG refers to foreground; BG refers to background." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 239, + 441, + 309 + ], + "lines": [ + { + "bbox": [ + 324, + 239, + 441, + 309 + ], + "spans": [ + { + "bbox": [ + 324, + 239, + 441, + 309 + ], + "type": "table", + "html": "
MethodsmAP\\( AP_{50} \\)\\( AP_{75} \\)
T: Intern-Base54.673.659.3
S: Intern-Tiny53.271.558.0
No Sample (900)54.272.659.1
Random Sample (300)54.573.259.5
FG queries (IoU > 0.3)54.272.559.2
BG queries (IoU < 0.3)54.272.859.2
", + "image_path": "ec2c07121941d1ca4b90a0520beddd920562595d9a128313be199b5af2e20de0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 219, + 341, + 391, + 380 + ], + "blocks": [ + { + "bbox": [ + 130, + 318, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 340 + ], + "type": "text", + "content": "Table 9: Applying our DLIM-Det to the huge model (InternImage-Huge) with 1B parameters to verify the scalability. Numbers in brackets refer to distillation gains against the original student." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 219, + 341, + 391, + 380 + ], + "lines": [ + { + "bbox": [ + 219, + 341, + 391, + 380 + ], + "spans": [ + { + "bbox": [ + 219, + 341, + 391, + 380 + ], + "type": "table", + "html": "
Teacher#ParamsDistilled Student
mAPAP50AP75
Intern-Huge1.08B56.1 (+2.9)74.4 (+2.9)61.5 (+3.5)
Intern-Large223M55.9 (+2.7)74.1 (+2.6)61.4 (+3.4)
", + "image_path": "dbf7bb15c47a6c49cf1a102b8a9def62f89be5b7bba9d53cb5fe9450a986b0ec.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 387, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 446 + ], + "type": "text", + "content": "as Swin Transformer and InternImage (Tab. 6), we also report results on ResNet [20] in Tab. 7. Our Query Distillation method yields consistent improvements on three different backbones, at least " + }, + { + "bbox": [ + 130, + 387, + 482, + 446 + ], + "type": "inline_equation", + "content": "+0.8" + }, + { + "bbox": [ + 130, + 387, + 482, + 446 + ], + "type": "text", + "content": " mAP. Furthermore, combining the QPD and QRD together achieves better performance than using them individually, verifying the effectiveness of each component." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 449, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 532 + ], + "type": "text", + "content": "Study on the distillation positions. To investigate the impact of Query Distillation on different decoder layers in DETRs, we conducted experiments, and the results are presented in Tab. 8a. When applying Query Distillation to all 6 decoder layers, we observe the best performance, surpassing the counterparts where it was applied to the first three or last three decoder layers. Furthermore, we examine the performance when applying Query Distillation to only one decoder layer. The results show that distillation on shallow decoder layers outperforms distillation on deeper layers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 534, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 483, + 666 + ], + "type": "text", + "content": "Selection of teacher queries. Using our Query Distillation method, we further conduct experiments to investigate the impact of different subsets of teacher queries on knowledge transfer. Results are reported in Tab. 8b. First, we perform KD on all teacher queries, obtaining " + }, + { + "bbox": [ + 130, + 534, + 483, + 666 + ], + "type": "inline_equation", + "content": "+1.0\\%" + }, + { + "bbox": [ + 130, + 534, + 483, + 666 + ], + "type": "text", + "content": " gains. Then we randomly sample 300 queries and achieve further improvements, surpassing the 900 query baseline by " + }, + { + "bbox": [ + 130, + 534, + 483, + 666 + ], + "type": "inline_equation", + "content": "+0.3\\%" + }, + { + "bbox": [ + 130, + 534, + 483, + 666 + ], + "type": "text", + "content": ". To explore alternative query selection strategies, we define the queries whose IoUs with GT are greater than 0.3 as foreground queries, otherwise as background queries. Distilling knowledge from either pure foreground (namely IoU " + }, + { + "bbox": [ + 130, + 534, + 483, + 666 + ], + "type": "inline_equation", + "content": ">0.3" + }, + { + "bbox": [ + 130, + 534, + 483, + 666 + ], + "type": "text", + "content": " in Tab. 8b) or pure background queries results in inferior performance compared to random sampling. This suggests that both foreground and background queries contain valuable knowledge, and distilling knowledge from a combination of them yields optimal results." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 138, + 423, + 180 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 10: Experiments on heterogeneous teacher-student pairs. Heterogeneous components are underlined. Numbers in brackets denote distillation gains against the original student." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 138, + 423, + 180 + ], + "lines": [ + { + "bbox": [ + 187, + 138, + 423, + 180 + ], + "spans": [ + { + "bbox": [ + 187, + 138, + 423, + 180 + ], + "type": "table", + "html": "
TeacherStudentDistilled Student
mAPAP50AP75
DINO-Swin-LDeformable DETR-Swin-T49.6 (+3.5)68.3 (+3.4)53.7 (+3.9)
DINO-Intern-HDINO-Swin-T53.3 (+2.0)71.3 (+2.3)58.3 (+2.3)
", + "image_path": "69db4f5e5f7319b0bb1e88225d97a7ea2ab8e2726f0d2698657535d609ca3374.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 167, + 197, + 444, + 284 + ], + "blocks": [ + { + "bbox": [ + 168, + 186, + 442, + 196 + ], + "lines": [ + { + "bbox": [ + 168, + 186, + 442, + 196 + ], + "spans": [ + { + "bbox": [ + 168, + 186, + 442, + 196 + ], + "type": "text", + "content": "Table 11: Comparison with state-of-the-art knowledge distillation methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 197, + 444, + 284 + ], + "lines": [ + { + "bbox": [ + 167, + 197, + 444, + 284 + ], + "spans": [ + { + "bbox": [ + 167, + 197, + 444, + 284 + ], + "type": "table", + "html": "
MethodsSwin TransformerInternImage
mAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)mAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
Teacher: Large58.177.163.541.262.173.957.876.563.141.661.373.0
Student: Tiny51.369.056.034.554.466.153.271.558.035.556.669.3
FGD [49]51.1 (-0.2)69.255.834.353.865.652.8 (-0.4)71.157.636.456.466.8
MGD [50]51.1 (-0.2)69.855.834.954.266.154.2 (+1.0)72.859.138.257.669.4
FitNet [37]51.3 (+0.0)70.055.934.754.566.353.8 (+0.5)72.458.738.257.468.2
DETRDistill [3]51.7 (+0.4)69.456.435.054.866.054.7 (+1.5)72.859.838.558.269.4
Our DLIM-Det54.4 (+3.1)72.159.438.157.469.755.9 (+2.7)74.161.440.359.470.8
", + "image_path": "aa5053da2905d5598d7ac6a4b94e897acf63174865db976f100be363e1877e70.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 293, + 481, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 364 + ], + "type": "text", + "content": "Scalability to huge teacher models. We apply DLIM-Det to InternImage-Huge with 1.08 billion parameters. From Tab. 9, we observe that DLIM-Det achieves satisfactory performance even when the teacher has more than 30 times the parameters of the student. Here, the student is represented by Intern-Tiny with 30M parameters, achieving a mAP of 53.2. Specifically, Intern-Huge produces a " + }, + { + "bbox": [ + 130, + 293, + 481, + 364 + ], + "type": "inline_equation", + "content": "+2.9" + }, + { + "bbox": [ + 130, + 293, + 481, + 364 + ], + "type": "text", + "content": " distillation gain on Intern-Tiny based student and surpasses Intern-Large by 0.2 mAP (56.1 v.s. 55.9)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 365, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 460 + ], + "type": "text", + "content": "Distillation on heterogeneous architectures. To verify the generalization of DLIM-Det, we conduct experiments on heterogeneous teacher and student models, including heterogeneous detectors and heterogeneous backbones. For heterogeneous detectors, with DINO [54] as the teacher and Deformable DETR [57] as the student, DLIM-Det improves the student from 46.1 to " + }, + { + "bbox": [ + 130, + 365, + 482, + 460 + ], + "type": "inline_equation", + "content": "49.6\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 365, + 482, + 460 + ], + "type": "text", + "content": ", as reported in Tab. 10. In the case of heterogeneous backbones, where Intern-Huge/Swin-Tiny are employed as the teacher/student, DLIM-Det obtains a considerable improvement of " + }, + { + "bbox": [ + 130, + 365, + 482, + 460 + ], + "type": "inline_equation", + "content": "+2.0\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 365, + 482, + 460 + ], + "type": "text", + "content": ". These results show the strong generalization of DLIM-Det." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 478, + 339, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 478, + 339, + 490 + ], + "spans": [ + { + "bbox": [ + 131, + 478, + 339, + 490 + ], + "type": "text", + "content": "4.3 Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Main results on two backbones. In Tab. 11, we compare our DLIM-Det with several popular KD methods. FitNet [37] is a classical feature distillation method, which introduces adaptation layers to bridge the semantic gap between the teacher and student. However, FitNet fails to improve the performance of the transformer-based backbone (Swin Transformer), showing no improvements. MGD [50] is inspired by the masked image modeling [18] and randomly masks pixels of the student's features. It then leads the student to generate the teacher's features in the masked regions. MGD obtains a noticeable improvement of " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "+1.0\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " for InternImage, but struggles to improve the Swin Transformer. FGD [49] leverages the GcBlock [1] to capture the relations between pixels from feature maps. However, FGD does not perform well on both backbones, indicating that the hand-crafted relation mechanism may conflict with the intrinsic relation modeling of detection transformers. In fact, FGD even hurts the performance of InternImage by " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "-0.4\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ". The recently proposed DETRDistill [3] is designed for distilling DETRs. In addition to performing customized logits and feature distillation, it also distills the label" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "assignment results of teacher. Notably, our DLIM-Det outperforms it by a large margin, at least " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "+1.2\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": ". We explain that the Query prior Assignment Distillation in DETRDistill serves a role akin to the De-Noising group in DINO, thereby it only yields minor improvements in our experiments. Moreover, another contribution of DETRDistill [3] is a feature imitation method, which is vulnerable to significant discrepancy between the teacher and student, as confirmed by other works [37,49,50]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 236 + ], + "type": "text", + "content": "These experimental findings suggest that while previous works can achieve decent distillation gains for medium-sized teacher, they encounter challenges in distilling large-scale teacher models. In contrast, DLIM-Det demonstrates significant distillation performance specifically tailored for large-scale models." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 166, + 267, + 445, + 358 + ], + "blocks": [ + { + "bbox": [ + 130, + 243, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 266 + ], + "type": "text", + "content": "Table 12: Comparison with the Teacher Assistant (TA) on two backbones. The TA consists of two steps, thus the overall training time is the sum of the training time for both steps." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 166, + 267, + 445, + 358 + ], + "lines": [ + { + "bbox": [ + 166, + 267, + 445, + 358 + ], + "spans": [ + { + "bbox": [ + 166, + 267, + 445, + 358 + ], + "type": "table", + "html": "
MethodsSwin TransformerInternImage
mAP\\( AP_{50} \\)\\( AP_{75} \\)Train TimemAP\\( AP_{50} \\)\\( AP_{75} \\)Train Time
Baseline:
Teacher: Large58.177.163.5-57.876.563.1-
Intermediate Teacher: Base55.874.460.7-54.673.659.3-
Student: Tiny51.369.056.0-53.271.558.0-
TA step1 (Large→Base)57.5 (+1.7)76.162.827.3h56.8 (+2.2)75.262.134.5h
TA setup2 (Base→Tiny)53.7 (+2.4)71.758.919.2h55.7 (+2.5)74.060.723.6h
Our DLIM-Det (Large→Tiny)54.4 (+3.1)72.159.420h55.9 (+2.7)74.161.424.8h
", + "image_path": "bfdfe9ead35c4d01fa735f899a614a25c2aa3b716e7bfcf20a00d766ea017cdd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "content": "Comparison with the Teacher Assistant. The Teacher Assistant (TA) [35, 39, 56] is proved effective in reducing the gap between the teacher and student models. Following the previous work [35], we use the base-scale model with approximately 90M parameters as the TA model. The TA model is distilled from the teacher model and then serves as the teacher for the student model. The results in Tab. 12 demonstrate that our DLIM-Det, which directly distills knowledge from large-scale models, achieves superior performance and higher efficiency compared to the TA method. Specifically, by using only half of the training time, DLIM-Det outperforms Teacher Assistant by " + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "inline_equation", + "content": "0.7 / 0.2" + }, + { + "bbox": [ + 130, + 367, + 482, + 475 + ], + "type": "text", + "content": " mAP on Swin/Intern models, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 495, + 211, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 495, + 211, + 508 + ], + "spans": [ + { + "bbox": [ + 131, + 495, + 211, + 508 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": "In this paper, we introduce DLIM-Det, a novel knowledge distillation method specifically tailored for large-scale image models in the context of object detection. To bridge the gap between the teacher and student, we propose the frozen teacher method, which involves freezing the pretrained backbone during teacher creation. The frozen teacher preserves the generalization capacity of large-scale models and gives rise to distinctive characteristics, such as sparse query locations. Through the proposed Query Distillation method, we harness the full advantages offered by the frozen teacher, thereby facilitating the distillation process. Extensive experiments provide robust evidence that larger teacher models can indeed lead to greater distillation gains through DLIM-Det." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "Acknowledgement. This work is partially supported by National Natural Science Foundation of China (No. 62322602), Natural Science Foundation of Jiangsu Province, China (No. BK20230033), and National Natural Science Foundation of China (No. 62172225)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 175 + ], + "type": "text", + "content": "1. Cao, Y., Xu, J., Lin, S., Wei, F., Hu, H.: Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In: Proceedings of the IEEE/CVF international conference on computer vision workshops. pp. 0-0 (2019) 8, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 481, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 481, + 219 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 481, + 219 + ], + "type": "text", + "content": "2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 213-229. Springer (2020) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 220, + 481, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 481, + 252 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 481, + 252 + ], + "type": "text", + "content": "3. Chang, J., Wang, S., Xu, H.M., Chen, Z., Yang, C., Zhao, F.: Detrdistill: A universal knowledge distillation framework for detr-families. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6898-6908 (2023) 2, 4, 13, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 254, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 481, + 285 + ], + "type": "text", + "content": "4. Chen, G., Choi, W., Yu, X., Han, T., Chandraker, M.: Learning efficient object detection models with knowledge distillation. Advances in neural information processing systems 30 (2017) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 287, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 481, + 308 + ], + "type": "text", + "content": "5. Chen, Q., Chen, X., Zeng, G., Wang, J.: Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085 (2022) 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "type": "text", + "content": "6. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 481, + 363 + ], + "type": "text", + "content": "7. Chen, X., Chen, J., Liu, Y., Zeng, G.: DG etr: Decoder distillation for detection transformer arXiv preprint arXiv:2211.09768 (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 481, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 398 + ], + "type": "text", + "content": "8. Dai, X., Jiang, Z., Wu, Z., Bao, Y., Wang, Z., Liu, S., Zhou, E.: General instance distillation for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7842-7851 (2021) 2, 4, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 399, + 481, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 420 + ], + "type": "text", + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 421, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 421, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 135, + 421, + 481, + 453 + ], + "type": "text", + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 135, + 454, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 454, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 135, + 454, + 481, + 486 + ], + "type": "text", + "content": "11. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. arXiv preprint arXiv:2211.07636 (2022) 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 135, + 487, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 487, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 135, + 487, + 481, + 509 + ], + "type": "text", + "content": "12. Floridi, L., Chiriatti, M.: Gpt-3: Its nature, scope, limits, and consequences. Minds and Machines 30, 681–694 (2020) 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 135, + 510, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 510, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 135, + 510, + 481, + 542 + ], + "type": "text", + "content": "13. Gao, Z., Wang, L., Han, B., Guo, S.: Adamixer: A fast-converging query-based object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5364-5373 (2022) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 544, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 544, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 135, + 544, + 481, + 564 + ], + "type": "text", + "content": "14. Guo, G., Han, L., Wang, L., Zhang, D., Han, J.: Semantic-aware knowledge distillation with parameter-free feature uniformization. Visual Intelligence 1(1), 6 (2023) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 135, + 566, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 566, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 135, + 566, + 481, + 597 + ], + "type": "text", + "content": "15. Guo, J., Han, K., Wang, Y., Wu, H., Chen, X., Xu, C., Xu, C.: Distilling object detectors via decoupled features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2154-2164 (2021) 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 135, + 599, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 599, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 135, + 599, + 481, + 631 + ], + "type": "text", + "content": "16. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5356-5364 (2019) 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 135, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 632, + 481, + 665 + ], + "type": "text", + "content": "17. Hao, Z., Guo, J., Jia, D., Han, K., Tang, Y., Zhang, C., Hu, H., Wang, Y.: Learning efficient vision transformers via fine-grained manifold distillation. In: Advances in Neural Information Processing Systems (2021) 3, 4, 8" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "18. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022) 5, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "19. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 214 + ], + "type": "text", + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 4, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 236 + ], + "type": "text", + "content": "21. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 258 + ], + "type": "text", + "content": "22. Huang, T., You, S., Wang, F., Qian, C., Xu, C.: Knowledge distillation from a stronger teacher. arXiv preprint arXiv:2205.10536 (2022) 4, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 290 + ], + "type": "text", + "content": "23. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 292, + 481, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 292, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 292, + 481, + 313 + ], + "type": "text", + "content": "24. Jia, Z., Sun, S., Liu, G., Liu, B.: Mssd: multi-scale self-distillation for object detection. Visual Intelligence 2(1), 8 (2024) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "25. Kang, Z., Zhang, P., Zhang, X., Sun, J., Zheng, N.: Instance-conditional knowledge distillation for object detection. Advances in Neural Information Processing Systems 34, 16468-16480 (2021) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 346, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 481, + 380 + ], + "type": "text", + "content": "26. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "27. Li, G., Li, X., Wang, Y., Zhang, S., Wu, Y., Liang, D.: Knowledge distillation for object detection via rank mimicking and prediction-guided feature imitation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1306-1313 (2022) 4, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "text", + "content": "28. Li, G., Li, X., Zhang, S., Yang, J.: Towards more reliable evaluation in pedestrian detection by rethinking \"ignore regions\". Visual Intelligence 2(1), 4 (2024) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "text", + "content": "29. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 6, 8, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "30. Lin, Y., Liu, Z., Zhang, Z., Hu, H., Zheng, N., Lin, S., Cao, Y.: Could giant pre-trained image models extract universal representations? Advances in Neural Information Processing Systems 35, 8332-8346 (2022) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "31. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "32. Liu, Y., Chen, K., Liu, C., Qin, Z., Luo, Z., Wang, J.: Structured knowledge distillation for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2604-2613 (2019) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "33. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12009-12019 (2022) 1, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "34. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) 1, 2, 3, 4, 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "35. Mirzadeh, S.I., Farajtabar, M., Li, A., Levine, N., Matsukawa, A., Ghasemzadeh, H.: Improved knowledge distillation via teacher assistant. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 5191-5198 (2020) 2, 14" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 195, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 195, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 195, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 148 + ], + "type": "text", + "content": "36. Park, D.Y., Cha, M.H., Kim, D., Han, B., et al.: Learning student-friendly teacher networks for knowledge distillation. Advances in neural information processing systems 34, 13292-13303 (2021) 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 171 + ], + "type": "text", + "content": "37. Romero, A., Ballas, N., Kahou, S.E., Chassang, A., Gatta, C., Bengio, Y.: Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550 (2014) 2, 3, 13, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 482, + 215 + ], + "type": "text", + "content": "38. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 482, + 247 + ], + "type": "text", + "content": "39. Son, W., Na, J., Choi, J., Hwang, W.: Densely guided knowledge distillation using multiple teacher assistants. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9395-9404 (2021) 2, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 482, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 482, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 482, + 280 + ], + "type": "text", + "content": "40. Wang, C., Yang, Q., Huang, R., Song, S., Huang, G.: Efficient knowledge distillation from model checkpoints. Advances in Neural Information Processing Systems 35, 607-619 (2022) 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 482, + 314 + ], + "type": "text", + "content": "41. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., et al.: Interimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778 (2022) 1, 3, 4, 5, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 482, + 346 + ], + "type": "text", + "content": "42. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442 (2022) 1, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 482, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 482, + 379 + ], + "type": "text", + "content": "43. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2567-2575 (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 482, + 402 + ], + "type": "text", + "content": "44. Wang, Y., Ni, Z., Song, S., Yang, L., Huang, G.: Revisiting locally supervised learning: an alternative to end-to-end training. arXiv preprint arXiv:2101.10832 (2021) 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 482, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 482, + 434 + ], + "type": "text", + "content": "45. Wei, F., Gao, Y., Wu, Z., Hu, H., Lin, S.: Aligning pretraining for detection via object-level contrastive learning. Advances in Neural Information Processing Systems 34, 22682-22694 (2021) 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 434, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 482, + 456 + ], + "type": "text", + "content": "46. Xu, T., Zhu, X.F., Wu, X.J.: Learning spatio-temporal discriminative model for affine subspace based visual object tracking. Visual Intelligence 1(1), 4 (2023) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 482, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 482, + 489 + ], + "type": "text", + "content": "47. Yang, C., Zhou, H., An, Z., Jiang, X., Xu, Y., Zhang, Q.: Cross-image relational knowledge distillation for semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12319-12328 (2022) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 482, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 482, + 521 + ], + "type": "text", + "content": "48. Yang, J., Li, C., Zhang, P., Dai, X., Xiao, B., Yuan, L., Gao, J.: Focal attention for long-range interactions in vision transformers. Advances in Neural Information Processing Systems 34, 30008-30022 (2021) 1, 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 482, + 555 + ], + "type": "text", + "content": "49. Yang, Z., Li, Z., Jiang, X., Gong, Y., Yuan, Z., Zhao, D., Yuan, C.: Focal and global knowledge distillation for detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4643-4652 (2022) 4, 8, 13, 14" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 482, + 588 + ], + "type": "text", + "content": "50. Yang, Z., Li, Z., Shao, M., Shi, D., Yuan, Z., Yuan, C.: Masked generative distillation. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI. pp. 53-69. Springer (2022) 2, 13, 14" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 482, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 482, + 631 + ], + "type": "text", + "content": "51. Yao, L., Pi, R., Xu, H., Zhang, W., Li, Z., Zhang, T.: G-detkd: towards general distillation framework for object detectors via contrastive and semantic-guided feature imitation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3591-3600 (2021) 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 482, + 665 + ], + "type": "text", + "content": "52. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021) 1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "Distilling Knowledge from Large-Scale Image Models for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 259 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "type": "text", + "content": "53. Zeng, Z., Yang, F., Liu, H., Satoh, S.: Improving deep metric learning via self-distillation and online batch diffusion process. Visual Intelligence 2(1), 1-13 (2024) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 482, + 171 + ], + "type": "text", + "content": "54. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022) 2, 3, 4, 9, 10, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "55. Zhang, Y., Xiang, T., Hospedales, T.M., Lu, H.: Deep mutual learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4320-4328 (2018) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "56. Zheng, Z., Ye, R., Hou, Q., Ren, D., Wang, P., Zuo, W., Cheng, M.M.: Localization distillation for object detection. arXiv preprint arXiv:2204.05957 (2022) 2, 4, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "57. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) 4, 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "58. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training. arXiv preprint arXiv:2211.12860 (2022) 3, 4" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 196, + 101 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/b0f9a1b1-12b0-444d-bc6b-1b50e3076e57_content_list.json b/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/b0f9a1b1-12b0-444d-bc6b-1b50e3076e57_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..62d8ea5dba806a60d9f3ff5704418a7caa7b3b85 --- /dev/null +++ b/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/b0f9a1b1-12b0-444d-bc6b-1b50e3076e57_content_list.json @@ -0,0 +1,1656 @@ +[ + { + "type": "text", + "text": "Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yukun Wang $^{1}$ , Kunhong Li $^{1}$ , Minglin Chen $^{1}$ , Longguang Wang $^{1}$ , Shunbo Zhou $^{2}$ , Kaiwen Xue $^{2}$ , and Yulan Guo $^{1}$", + "bbox": [ + 246, + 210, + 754, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Shenzhen Campus of Sun Yat-sen University, Sun Yat-sen University {wangyk59,likh25,chenmlin8}@mail2.sysu.edu.cn guoyulan@sysu.edu.cn", + "bbox": [ + 222, + 253, + 777, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Huawei Cloud Computing Technical Innovation Dept., Huawei Cloud Computing Technologies Co., Ltd.", + "bbox": [ + 223, + 296, + 776, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Neural Radiance Fields (NeRF) and 3D Gaussian Splitting (3DGS) have greatly advanced novel view synthesis, which is capable of photo-realistic rendering. However, these methods require the foundational assumption of the static scene (e.g., consistent lighting condition and persistent object positions), which is often violated in real-world scenarios. In this study, we introduce MemE, an unsupervised plug-and-play module, to achieve high-quality novel view synthesis in noisy input scenarios. MemE leverages the inherent property in parameter optimization, known as the memorization effect to achieve distractor filtering and can be easily combined with NeRF or 3DGS. Furthermore, MemE is applicable in environments both with and without distractors, significantly enhancing the adaptability of NeRF and 3DGS across diverse input scenarios. Extensive experiments show that our methods (i.e., MemE-NeRF and MemE-3DGS) achieve state-of-the-art performance on both real and synthetic noisy scenes. We will release our code for further research at https://github.com/Yukun66/MemE.", + "bbox": [ + 259, + 359, + 743, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D reconstruction $\\cdot$ Unsupervised noise modeling $\\cdot$ Memorization effect", + "bbox": [ + 259, + 595, + 740, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 648, + 375, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, the emergence of methods like Neural Radiance Fields (NeRF) [4,36] and 3D Gaussian Splatting (3DGS) [20] has greatly propelled the development of novel view synthesis. Additionally, the application of novel view synthesis in fields such as robotics, 3D reconstruction, and AR/VR is gradually being explored. However, these methods are built upon certain critical assumptions, including consistent lighting conditions and persistent object positions. These assumptions are frequently violated in real-world scenarios, which leads to artifacts, significant degradation in rendering quality. This study tackles the specific challenge encountered in novel view synthesis: reconstructing the static and clean", + "bbox": [ + 212, + 679, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\boxtimes$ Corresponding author.", + "bbox": [ + 230, + 823, + 400, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2bad65dcaaad2d1b27fe879cba36185daf0054fa09ce77754447584c07811a95.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 150, + 781, + 304 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/516c99d15bb5c83ee3ea6d17b311edda59c98166d6cf99c7ca4a7c93b8b53a7f.jpg", + "image_caption": [ + "(a) Mip-NeRF 360", + "(b) MemE-NeRF (Ours)", + "Fig. 1: Memorization Effect in Optimization. The purity (clean scene) and distractor pixels correspond to the blue and red bars in the histograms, respectively. In the training of NeRF and 3DGS, we observed the phenomenon of memorization effect. This phenomenon refers to the parameters initially fitting purity and excluding distractor pixels in the early stages of training, gradually overfitting to distractor pixels in the later stages, as evidenced by the rendered image and histogram in (a). Our method leverages the memorization effect to effectively prevent distractors from overfitted, maintaining the position of the red bars unchanged as much as possible, as demonstrated in (b)." + ], + "image_footnote": [], + "bbox": [ + 225, + 323, + 781, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scene from noisy images containing distractors, which disrupt the assumption of the static scene.", + "bbox": [ + 212, + 611, + 787, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To comprehend the impact of distractors in novel view synthesis, we analyze the optimization process of NeRF and 3DGS. Specifically, we exemplify this analysis using Mip-NeRF 360 [4], as depicted in Fig. 1a. Our observations indicate that during the initial stages of optimization (e.g., the 5000th step), the image exhibits blurring, yet the scene remains relatively clean. However, as optimization progresses, the image sharpens, concomitant with the emergence of distractors. Further analyzing the accumulated loss distribution, Mip-NeRF 360 primarily focuses on learning purity pixels (of the clean scenes) and leaves most distractor pixels out of the learning process at the early stage, as evidenced by the minimal changes in the histogram of distractor pixels (red bars). This observation, namely memorization effect [2, 42], enables a clear distinction between purity and distractor pixels. Subsequently, we can promote the fitting of purity pixels and suppress the learning of distractor pixels throughout the Mip-NeRF", + "bbox": [ + 212, + 643, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "360 training process to easily achieve sharp and clean view rendering in a noisy input scenario.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the above insight, we propose MemE, an unsupervised plug-and-play method that can be incorporated with NeRF and 3DGS, to achieve high-quality novel view synthesis in noisy input scenarios. MemE is based on the discovery of Memorization Effect in NeRF and 3DGS optimization, and the core of MemE is to distinguish distractor pixels without prior knowledge and exclude them throughout the optimization process. Specifically, our MemE begins with our Hierarchical Residual Measurement module to quantify the residuals between the rendered results and the noisy input images, outputting residual vectors for other modules. Subsequently, within our Purity-Distractor Mixture Modeling module, we employ a multi-level Gaussian Mixture Model (GMM) to model the residual distribution of both purity and distractor pixels according to the residual vectors. Finally, we calculate a purity attention weight for NeRF/3DGS optimization. With that attention, we can ensure that the distractor pixels remains unlearned throughout the parameters optimization process, preventing the parameters from overfitting to distractor pixels and thereby achieving the distractor-free novel view synthesis, as shown in Fig. 1b.", + "bbox": [ + 212, + 176, + 787, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In comparison to prior works, our approach offers distinct advantages: 1) It is solely grounded in the inherent property of NeRF and 3DGS training, i.e., memorization effect, and conducted in an unsupervised manner. 2) Our MemE is capable of processing training data, whether it contains distractors or not, without the need for prior identification. 3) Our method operates independently of pre-trained models or parameters, ensuring simplicity and effectiveness, thus enabling seamless integration as a plug-and-play solution. We incorporate MemE with NeRF (referred to as MemE-NeRF) and 3DGS (referred to as MemE-3DGS) to demonstrate the universality of MemE.", + "bbox": [ + 212, + 417, + 787, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows,", + "bbox": [ + 238, + 554, + 591, + 569 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first work to exploit the memorization effect in NeRF and 3DGS optimization process. Further, we propose MemE-NeRF and MemE-3DGS to achieve high-quality novel view synthesis in the noisy input scenario in an unsupervised manner.", + "- We propose MemE, a simple yet plug-and-play method to filter distractor pixels. MemE can adapt to scenes both with and without distractors simultaneously, which can enhance the versatility of NeRF/3DGS in various types of input scenarios.", + "- We conduct extensive experiments and the results show that our method achieves state-of-the-art performance on both real and synthetic noisy scenes." + ], + "bbox": [ + 223, + 579, + 782, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 752, + 387, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Novel View Synthesis", + "text_level": 1, + "bbox": [ + 214, + 784, + 441, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF) [36] has made a breakthrough in novel view synthesis, which is capable of synthesizing photo-realistic images at arbitrary", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "views. NeRF represents a 3D scene as a continuous radiance field, and synthesizes images with differentiable volumetric rendering. Mip-NeRF [3] and Zip-NeRF [5] improve the rendering quality by reducing the aliasing artifacts. $\\mathrm{NeRF}++$ [57] and Mip-NeRF 360 [4] are proposed to extend NeRF to $360^{\\circ}$ unbounded scene [7]. Several works accelerate NeRF by parameterizing the radiance fields with high-efficient data structures, such as multiresolution hash table [38], Octree [11,56], and orthogonal planes [6]. Recently, 3D Gaussian Splatting [20] has emerged as another promising paradigm for novel view synthesis, which synthesizes the image by rendering a set of learnable Gaussian points.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite great development in novel view synthesis, the aforementioned methods are limited to static scenes with constant light conditions. They are struggling with multi-view images that contain distractors. Our method aims to extend existing methods to distractor-free novel view synthesis.", + "bbox": [ + 212, + 282, + 787, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 View Synthesis from Noisy Inputs", + "text_level": 1, + "bbox": [ + 214, + 363, + 545, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, efforts are made to remove noises during NeRF reconstruction, such as training data noise [15, 35, 39, 43, 45, 50], deblurring [9, 21, 22, 33, 51], artifact removal [12, 23, 52], dynamic distractor elimination [34, 40, 41, 44, 55], and scene inpainting [37, 49, 53, 54]. To address transient occluders, NeRF-W [34] employs multilayer perceptrons (MLPs) modeling for static and transient components separately. Methods for removing movable objects in large-scale street scenes [40, 44] have utilized semantic segmentation models to explicitly filter out dynamic outliers, such as cars and pedestrians. Recent advancements have extended NeRF to video data [26, 27, 48], introducing an auxiliary temporal MLP for separately modeling dynamic and static scenes. RobustNeRF [41] processes discrete images from casual viewpoints, ranking training pixel errors to identify and eliminate distractors using hand-crafted ratios, which may yield suboptimal results.", + "bbox": [ + 212, + 387, + 787, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most of the aforementioned methods do not consider the case of distractor noise, while the most related counterpart of our method is RobustNeRF [41]. Compared to RobustNeRF, our method is capable of scenes with or without distractors, with no need for analysis and artificial prior on the distractor content of the scene in advance. Our method can achieve similar results to Mip-NeRF 360 and 3DGS in distractor-free scenes and does not require special adjustment of hyperparameters.", + "bbox": [ + 212, + 569, + 787, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Noisy Label Learning", + "text_level": 1, + "bbox": [ + 214, + 695, + 441, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Numerous methods [1, 14, 16, 17, 19, 24, 28-32, 47, 58] have investigated the challenge of label noise in the field of image classification due to the difficulty in data labeling [42]. Prior arts on the memorization effect [1, 2, 42] indicates that neural networks tend first to learn simple and clean samples, and then gradually memorize noisy labeled samples. Based on this, several efforts are devoted to combat label noise by selecting clean samples from noisy dataset during model training, thus avoiding model from overfitting to those noisy samples. Co-teaching [14] retains two networks, and each network selects a portion of samples with small-loss", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/75e2e49fee8494af2307f02d549e5e1e0790b8680c021a6fbd7b7dbe0243591e.jpg", + "image_caption": [ + "Fig. 2: The Framework of MemE. Our MemE integrates three primary components: Hierarchical Residual Measurement (Sec. 3.1), Purity-Distractor Mixture Modeling (Sec. 3.2), and Purity Attention Generation (Sec. 3.3). Hierarchical Residual Measurement module generates residual vectors $\\{\\pmb{\\Gamma}_n\\}$ and residual maps $\\{\\pmb{\\epsilon}_n\\}$ , which are then utilized by the subsequent modules. Purity-Distractor Mixture Modeling module fits the multi-level Gaussian Mixture Model (GMM) according to $\\{\\pmb{\\Gamma}_n\\}$ . Finally, the Purity Attention Generation module creates attention weight $\\mathbf{M}$ to exclude distractors selectively." + ], + "image_footnote": [], + "bbox": [ + 222, + 150, + 777, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and feeds them to another network. Note that above methods discard intensive noisy samples, which inevitably incurs information loss, limiting their performance especially when the noise ratio is high. To remedy this, DivideMix [24] employs Gaussian mixture models to split training dataset into labeled (clean) and unlabeled (noisy) sets, and then adopts MixMatch to make full use of those noisy samples. Besides, LongReMix [8] also uses MixMatch strategy and augments high-confidence samples to achieve superior robustness.", + "bbox": [ + 212, + 488, + 787, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 618, + 330, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present a novel plug-and-play method (namely MemE), which is aimed at enhancing existing novel view synthesis methods when multi-view images contain discriminators. Existing novel view synthesis methods represent a 3D scene as a parameterized model $\\mathbf{S}_{\\theta}$ , while facilitating a differentiable rendering technique $\\mathcal{R}(\\cdot ,\\cdot)$ to render an image $\\hat{\\mathbf{I}}$ as following:", + "bbox": [ + 212, + 650, + 787, + 727 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {I}} = \\mathcal {R} (\\mathbf {S} _ {\\theta}, \\pi), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 449, + 739, + 785, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi$ is the image pose. Concretely, Neural Radiance Fields [36] leverages the Multilayer Perceptron (MLP) and volumetric rendering, whereas 3D Gaussian Splatting [20] is built upon a collection of learnable Gaussian points and splatting rasterization. Because multi-view images may contain distractors, we employ the proposed MemE to generate an attention weight $\\mathbf{M}$ to indicate the distractor", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization", + "bbox": [ + 243, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pixels. Therefore, existing methods with our MemE can achieve distractor-free novel view synthesis using the following loss:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) = \\sum_ {r} \\mathbf {M} _ {r} | | \\hat {\\mathbf {I}} _ {r} - \\mathbf {I} _ {r} | | _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 188, + 785, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{I}}_r$ , $\\mathbf{I}_r$ and $\\mathbf{M}_r$ are the rendered color, ground-truth color, and attention weight at pixel $r$ , respectively.", + "bbox": [ + 212, + 231, + 782, + 261 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The framework of our MemE is depicted in Fig. 2. Our method consists of three components: hierarchical residual measurement (Sec. 3.1), purity-distractor mixture modeling (Sec. 3.2), and purity attention generation (Sec. 3.3). The hierarchical residual measurement module quantifies the residuals between noisy input $\\mathbf{I}$ and rendered result $\\hat{\\mathbf{I}}$ . The purity-distractor mixture modeling module fits the distribution of residuals. The purity attention generation module generates a noisy-excluded attention mask $\\mathbf{M}$ for loss adjusting. By composing the three simple modules, MemE exploits the memorization effect existing in NeRF and 3DGS, and models the purity-distractor distribution of the scene in an unsupervised manner. The resulting mask $\\mathbf{M}$ can effectively help NeRF and 3DGS focus only on purity and ignore distractors.", + "bbox": [ + 212, + 262, + 787, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Hierarchical Residual Measurement", + "text_level": 1, + "bbox": [ + 215, + 450, + 555, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MemE identifies which parts of the scene are prioritized for learning by calculating residuals of rendering results and noisy input. According to the observed memorization effect, the areas learned first more likely belong to purity, whereas distractors are more likely to be left out of training in the early stage. Further, we employ residual measurement on multiple scales to provide more comprehensive local information hierarchically.", + "bbox": [ + 212, + 474, + 787, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given a noisy input patch $\\mathbf{I} \\in \\mathbb{R}^{H \\times W}$ and a rendering result patch $\\hat{\\mathbf{I}} \\in \\mathbb{R}^{H \\times W}$ , the Hierarchical Residual Measurement module eventually outputs hierarchical residual vectors and residual maps to the other modules. Specifically, the residual map between $\\hat{\\mathbf{I}}$ and $\\mathbf{I}$ is first calculated by the L2 norms, noted as $\\epsilon$ . Then a residual pyramid $\\{\\epsilon_n\\}$ is generated by pixel-wise and tile-wise operations according to $\\epsilon$ . Finally, $\\{\\epsilon_n\\}$ are simply flattened to vectors $\\{\\Gamma_n\\}$ :", + "bbox": [ + 212, + 566, + 787, + 657 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\epsilon} _ {n} = \\tau_ {n} \\odot \\boldsymbol {\\epsilon}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 457, + 670, + 784, + 686 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Gamma} _ {n} = \\nu \\odot \\boldsymbol {\\epsilon} _ {n}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 452, + 712, + 784, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\nu$ is the vectorization operation to flatten the patch, $\\tau_{n}$ represents the pixel-wise and tile-wise operations, and $\\odot$ represents the operation deployment.", + "bbox": [ + 212, + 734, + 782, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The pixel-wise operations are the average convolutions with kernels of $\\mathbb{R}^{i\\times i}$ , $i\\in \\mathcal{I}$ , stride 1 and padding, where $\\mathcal{I}$ refers to the set of pixel-wise scales. As a result, the multi-level residuals $\\epsilon_{i}$ are with spatial size $\\mathbb{R}^{H\\times W}$ . The tile-wise operations are the average pooling with kernels of $\\mathbb{R}^{j\\times j}$ , $j\\in \\mathcal{J}$ , stride $j$ and no padding, where $\\mathcal{J}$ refers to the set of tile-wise scales. Thus, the multi-level", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "residuals $\\pmb{\\epsilon}_{j}$ are with spatial size $\\mathbb{R}^{H / j\\times W / j}$ . In conclusion, the outputs of Hierarchical Residual Measurement are multi-level residual maps $\\{\\pmb{\\epsilon}_n\\}$ and residual vectors $\\{\\pmb{\\Gamma}_n\\}$ , where $n\\in \\mathcal{I}\\cup \\mathcal{J}$ .", + "bbox": [ + 212, + 145, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Purity-Distractor Mixture Modeling", + "text_level": 1, + "bbox": [ + 214, + 212, + 562, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the residual vectors $\\{\\pmb{\\Gamma}_n\\}$ from the Hierarchical Residual Measurement module, it is necessary to use a model to fit the distribution of purity and distractors in the NeRF and 3DGS parameter optimization process, so that MemE can use the model to identify purity and distractors in the next module, and calculate the mask to adjust training loss.", + "bbox": [ + 212, + 236, + 782, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the Purity-Distractor Mixture Modeling module, a one-dimensional and two-component GMM is utilized to fit the residual distributions of purity and distractor pixels because the residual distribution tends to be bi-modal in the early training stage. The probability density function (pdf) of a mixture model of $K$ components, applied to a residual vector $\\pmb{\\Gamma}_{n}$ , is defined as:", + "bbox": [ + 212, + 311, + 784, + 387 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\boldsymbol {\\Gamma} _ {n}\\right) = \\sum_ {k = 1} ^ {K} \\lambda_ {k} \\cdot \\mathcal {N} \\left(\\boldsymbol {\\Gamma} _ {n} \\mid \\mu_ {k}, \\sigma_ {k} ^ {2}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 398, + 784, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda_{k}$ represents the mixing coefficients for the convex combination of each pdf $\\mathcal{N}(\\pmb{\\Gamma}_n\\mid \\mu_k,\\sigma_k^2)$ . The fitting of GMM is conducted through the Expectation-Maximization (EM) algorithm [10]. As the residual vectors $\\{\\pmb{\\Gamma}_n\\}$ are with multi-level, the GMM is implemented with each level respectively, and the output is a set of pdfs expressed as $\\{p(\\pmb{\\Gamma}_n)\\}$ .", + "bbox": [ + 212, + 446, + 782, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, to achieve robust training, we update $\\{p(\\boldsymbol{\\Gamma}_n)\\}$ every $T$ steps. That is, we accumulate $T$ residual vectors $\\{\\boldsymbol{\\Gamma}_n^t | t \\in \\{0,1,\\dots,T-1\\}\\}$ for each level, and compute $\\{p(\\cup \\boldsymbol{\\Gamma}_n^t)\\}$ all at once. For convenience, we continue to use the symbol $\\{p(\\boldsymbol{\\Gamma}_n)\\}$ instead of $\\{p(\\cup \\boldsymbol{\\Gamma}_n^t)\\}$ to denote the output pdfs in subsequent sections. And at the very beginning $T$ steps, we do not employ our MemE for training.", + "bbox": [ + 212, + 523, + 784, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Purity Attention Generation", + "text_level": 1, + "bbox": [ + 214, + 633, + 498, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Utilized the residual maps $\\{\\pmb{\\epsilon}_n\\}$ and pdf sets $\\{p(\\pmb{\\Gamma}_n)\\}$ obtained from two former modules, Purity Attention Generation outputs an attention map $\\mathbf{M}$ adjusting loss function to constraint view synthesis model only focuses on the purity and repel distractors out of optimization procedure.", + "bbox": [ + 212, + 657, + 782, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, for the residual maps processed through tile-wise operations, we upsample their spatial size to the original size, i.e., $\\mathbb{R}^{H\\times W}$ , using nearest interpolation. Then, for spatially aligned residual maps $\\{\\pmb {\\epsilon}_n\\}$ , we have corresponding pdfs $\\{p(\\pmb {\\Gamma}_n)\\}$ , and we compute adaptive thresholds $\\{\\rho_n\\}$ according to $\\{p(\\pmb {\\Gamma}_n)\\}$ . Next, the binary masks $\\{\\mathbf{M}_n|\\mathbf{M}_n\\triangleq \\pmb {\\epsilon}_n\\leq \\rho_n\\}$ are generated according to $\\{\\rho_{n}\\}$ . Finally, $\\{\\mathbf{M}_n\\}$ are fused into a purity attention map $\\mathbf{M}$ .", + "bbox": [ + 212, + 719, + 782, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptive Threshold. We apply a threshold to the residual map to derive the mask at each level. Since our destination is to repel distractors from the", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization", + "bbox": [ + 243, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "optimization process, we hope this threshold is: 1) smaller than the most of residuals of distractors in the total optimization process; 2) slightly larger than most of the residuals of purity at the later stage of optimization to preserve most purity and make sure to exclude the distractors as many as possible; and 3) starting from a moderate value and gradually decrease as the optimization progresses because purity pixels may also have large residual values in the early stages of optimization.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Therefore, in each scale, we calculate a threshold from the position lying between the two means of $p(\\pmb{\\Gamma}_n)$ , i.e., $\\mu_1(\\pmb{\\Gamma}_n)$ and $\\mu_2(\\pmb{\\Gamma}_n)$ . There are three reasons as follows: 1) the means indicate the most concentrated residual magnitude of the purity and distractors; 2) the $\\mu_1(\\pmb{\\Gamma}_n)$ is the ideal value of our threshold at the end of optimization; and 3) the $\\mu_2(\\pmb{\\Gamma}_n)$ provide a moderate starting point lying between the ideal value $\\mu_1(\\pmb{\\Gamma}_n)$ and the unconstrained value, i.e., 1. Further, we design adaptive thresholds that exponentially move from $\\mu_2(\\pmb{\\Gamma}_n)$ to $\\mu_1(\\pmb{\\Gamma}_n)$ :", + "bbox": [ + 212, + 255, + 787, + 363 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {n} = \\alpha \\cdot \\mu_ {1} (\\boldsymbol {\\Gamma} _ {n}) + (1 - \\alpha) \\cdot \\mu_ {2} (\\boldsymbol {\\Gamma} _ {n}), \\quad \\alpha = 1 - e ^ {\\left(- \\omega \\cdot \\frac {s}{N}\\right)} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 378, + 785, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\omega$ controls the moving velocity, $s$ and $N$ are current training steps and total steps, and $\\rho_{n}$ is the threshold for generating the binary mask $\\mathbf{M}_n$ , where 0 refers to the prediction as distractors and 1 refers to the prediction as purity. Collecting the masks of different scales, we have a set of binary masks $\\{\\mathbf{M}_n\\}$ .", + "bbox": [ + 212, + 412, + 787, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mask Fusion. After the binary masks $\\{\\mathbf{M}_n\\}$ are generated, we further fuse the multi-level masks into a single purity attention map $\\mathbf{M}$ . These masks contain information at multiple scales and are fused in this module. We split the $\\{\\mathbf{M}_n\\}$ into pixel-wise operated subset $\\{\\mathbf{M}_i|i\\in \\mathcal{I}\\}$ and tile-wise operated subset $\\{\\mathbf{M}_j|j\\in \\mathcal{J}\\}$ . Generally, the masks in $\\{\\mathbf{M}_i\\}$ are fine-grained and the masks in $\\{\\mathbf{M}_j\\}$ are coarse grained. Within the $\\{\\mathbf{M}_i\\}$ , the mask with smaller $i$ is detailed but noisy; while the mask with larger $i$ is noise-filtered but blurry. As such, we first fuse the $\\{\\mathbf{M}_i\\}$ with OR operation, noted as $\\oplus$ , to combine their advantages including detailed and noise-filtered, generating the fused pixel-wise mask $\\mathbf{M}^p$ :", + "bbox": [ + 212, + 476, + 789, + 612 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} ^ {p} = \\oplus \\left\\{\\mathbf {M} _ {i} \\mid i \\in \\mathcal {I} \\right\\}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 428, + 630, + 785, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The OR operation removes the noise of value 0 in binary masks while preserving the noise of value 1. As such we combine the hierarchical tile-wise masks $\\{\\mathbf{M}_j\\}$ providing more comprehensive local information to generate the purity attention map $\\mathbf{M}$ :", + "bbox": [ + 212, + 662, + 787, + 724 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\frac {1}{\\mathcal {D}} \\left(\\mathbf {M} ^ {p} + \\sum_ {j} \\left(\\mathbf {M} _ {j}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 732, + 785, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ means the total number of masks, including $\\{\\mathbf{M}_j\\}$ and $\\mathbf{M}^p$ . We use $\\mathcal{I} = \\{1,3,5,7\\}$ and $\\mathcal{J} = \\{4,8,16\\}$ in our method. Therefore, the elements of purity attention map $\\mathbf{M}$ are discrete values belongs to the set $\\{0,0.25,0.5,0.75,1\\}$ .", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 145, + 374, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 176, + 333, + 189 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "RobustNeRF Dataset. The RobustNeRF dataset [41] comprises four different indoor scenes. Within each scene, there exists a sub-training set containing distractors and a sub-testing set without distractors. Additionally, in the BabyYoda scene, there is an additional sub-validation set devoid of distractors, featuring views distinct from those in the training set.", + "bbox": [ + 212, + 200, + 782, + 273 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Kubric Dataset. The Kubric dataset introduced in [55], is a synthetic dataset comprising five scenes, generated using the Kubric dataset generator [13]. Each scene comprises an empty room with a texture-less background and several static objects. In each frame, there are distractors present with continuous motion.", + "bbox": [ + 212, + 276, + 782, + 334 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Phototourism Dataset. The Phototourism dataset [18] is a typical outdoor scene dataset, which comprises images collected from the internet, thus containing numerous distractors such as tourists and illumination changes. We randomly select six scenes from the Phototourism dataset to assess the performance of our MemE.", + "bbox": [ + 212, + 337, + 782, + 409 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Metrics and Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 435, + 560, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We report PSNR, SSIM, and LPIPS to quantitatively evaluate the proposed MemE-NeRF and MemE-3DGS. MemE-NeRF is trained with 64 or 16 training patches per step for every scene. To ensure a fair comparison with RobustNeRF reported in [41], where training was conducted with 64 patches per step, we also train RobustNeRF with 16 patches (denoted as RobustNeRF*). MemE-NeRF is based on Mip-NeRF 360 [4] implemented using JAX, while MemE-3DGS is based on 3DGS [20] implemented using PyTorch. Experiments are conducted on 4 NVIDIA RTX-3090 GPUs for 64 patches and 1 for 16 patches. All models are trained with images containing distractors while tested on distractor-free images.", + "bbox": [ + 212, + 460, + 782, + 593 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Evaluation", + "text_level": 1, + "bbox": [ + 215, + 618, + 349, + 632 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "RobustNeRF Dataset. On the RobustNeRF dataset, our MemE outperforms the counterparts, as shown in Table 1 and Fig. 3. MemE-NeRF surpasses RobustNeRF across most metrics, while MemE-3DGS significantly enhances the performance of 3DGS in scenes containing distractors. Notably, our MemE-NeRF achieves comparable or even higher PSNR than RobustNeRF, despite being trained with fewer patches.", + "bbox": [ + 212, + 643, + 782, + 729 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, our MemE-NeRF accurately removes distractors while preserving more details. In Fig. 3 BabyYoda, MemE-NeRF effectively eliminates distractors and retains the detailed patterns of the table. In contrast, although RobustNeRF* removes distractors, it introduces incorrect floaters and misses details of the table. In Fig. 3 Statue, both MemE-NeRF and MemE-3DGS successfully remove distractors. However, RobustNeRF* is susceptible to contamination by distractors, leading to missing details and floaters. Compared", + "bbox": [ + 212, + 734, + 782, + 837 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/68225af0af38e6ba430639b969488bf753fe35ee67af11515f395582c004967c.jpg", + "table_caption": [ + "Table 1: Quantitative comparison on the RobustNeRF dataset." + ], + "table_footnote": [ + "the number of patches in each training step." + ], + "table_body": "
Methodsp†StatueAndroidCrabBabyYoda
LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑
Mip-NeRF 360 (L2) [4]640.360.6619.090.400.6519.350.270.7725.730.310.7522.97
Mip-NeRF 360 (L1) [4]640.300.7219.550.400.6619.380.220.7926.690.220.8026.15
Mip-NeRF 360 (Ch.) [4]640.300.7319.640.400.6619.530.210.8027.720.220.8025.22
D2NeRF [55]-0.480.4919.090.430.5720.610.420.6821.180.440.6517.32
RobustNeRF [41]640.280.7520.890.310.6521.720.210.8130.750.200.8330.87
MemE-NeRF (Ours)640.230.7321.990.300.6821.740.200.8330.640.250.8331.36
RobustNeRF* [41]160.330.6621.900.310.6721.400.240.8230.580.310.7929.76
MemE-NeRF (Ours)160.310.6821.810.300.6821.540.230.8330.610.290.8131.08
3DGS [20]-0.200.7620.220.250.6821.540.190.8925.790.190.9128.12
MemE-3DGS (Ours)-0.140.8121.740.180.7322.720.190.9025.810.190.9128.40
", + "bbox": [ + 222, + 171, + 781, + 311 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/91f3b21e872f9dc506acbc673f69f00d709faa67f8d6530c561a500fe6eedd99.jpg", + "image_caption": [ + "Fig.3: Qualitative Comparison on the RobustNeRF Dataset. Compared to the baselines, i.e., Mip-NeRF 360 and 3DGS, our MemE-NeRF and MemE-3DGS not only efficiently eliminate distractors but also retain a higher level of detail. In comparison with RobustNeRF*, MemE models demonstrate superior performance in removing distractors, leading to a reduction in artifacts and an enhancement in detail preservation." + ], + "image_footnote": [], + "bbox": [ + 225, + 339, + 782, + 599 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "with 3DGS, MemE-3DGS exhibits significant improvements in both qualitative and quantitative comparisons. 3DGS is prone to overfitting distractors, resulting in incorrect view synthesis ( Fig. 3 Statue and Crab) or artifacts in the rendered images ( Fig. 3 Android and Baby Yoda). It is noteworthy that RobustNeRF leads to significant performance degradation in clean scene training, whereas the integration of our MemE is non-intrusive to existing NeRF and 3DGS frameworks, as demonstrated in [41] and Sec. 4.4.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c22a64d089710108902eb54bbf5f3c0676c27a34ad5b49378efc2ea91f322045.jpg", + "table_caption": [ + "Table 2: Quantitative comparison on the Kubric dataset." + ], + "table_footnote": [], + "table_body": "
MethodsCarCarsBagChairsPillow
LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑
NeRF-W [34].218.81424.23.243.87324.51.139.79120.65.150.68123.77.088.93528.24
NSFF [25].200.80624.90.620.37610.29.108.89225.62.682.28412.82.782.3434.55
NeuralDiff [46].065.95231.89.098.92125.93.117.91029.02.112.72224.42.565.65220.09
D2NeRF [55].062.97534.27.090.95326.27.076.97934.14.095.70724.63.076.97936.58
RobustNeRF [41].013.98837.73.063.95726.31.006.99541.82.007.99241.23.018.99038.95
RobustNeRF* [41].021.97340.83.020.98643.70.027.98331.62.009.99346.46.026.98644.23
MemE-NeRF (Ours).021.97340.86.019.98744.34.003.99748.79.007.99447.87.016.98945.97
3DGS [20].037.96638.36.027.98341.85.005.99545.94.011.99144.76.034.98341.09
MemE-3DGS (Ours).032.96840.28.015.98945.40.002.99749.11.008.99447.82.016.98944.29
", + "bbox": [ + 218, + 171, + 782, + 287 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3dcb2d4a3a60e5096771e39f238c7d03a2e1e2717f75daca3e44cee8f220723e.jpg", + "image_caption": [ + "Fig. 4: Qualitative Comparison on the Kubric Dataset. Our MemE effectively manages scenes with distractors that exhibit continuous motion across successive frames, resulting in better distractor elimination and static object preserving." + ], + "image_footnote": [], + "bbox": [ + 225, + 311, + 779, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kubric Dataset. Compared with previous works, our MemE achieves consistent improvement in terms of all scenes and all metrics, as shown in Table 2 and Fig. 4. Note that the distractors in the Kubric dataset exhibit continuous motion between different frames, increasing the probability of overlap, particularly for objects with simple textures. Leveraging the inherent memorization", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c81c7c924afc35e6c52105fd2dd0c2072fbb89e60a259fb2e2f78feb1fb0b00e.jpg", + "image_caption": [ + "Fig.5: Quality Results of 3DGS and MemE-3DGS(Ours) on the Phototourism Dataset. The pairs of tri-column figures illustrate the test view images and the rendering images of 3DGS and MemE-3DGS, respectively. Our MemE-3DGS effectively eliminates artifacts caused by distractors and significant illumination variations." + ], + "image_footnote": [], + "bbox": [ + 220, + 148, + 777, + 431 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "effect in view synthesis models, our MemE utilizes GMM to adaptively perform customized fitting for each scene. Consequently, our MemE enhances the performance of NeRF and 3DGS to achieve SoTA results, regardless of whether the inputs are temporally continuous or discrete, and remains robust to inputs with varying distractor ratios.", + "bbox": [ + 212, + 537, + 787, + 612 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Phototourism Dataset. The Phototourism dataset remains challenging for existing novel view synthesis methods due to the diverse sources of images. This variability results in differences in distractors, illumination, and even camera equipment across images. Nevertheless, our MemE-3DGS achieves satisfactory view rendering results on this dataset, as shown in Fig. 5. Notably, severe lighting variations, as well as distractors, can significantly impact the quality of rendered images of 3DGS. However, our MemE is capable of effectively mitigating the influence of lighting variations and distractors.", + "bbox": [ + 212, + 614, + 787, + 736 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 763, + 390, + 779 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct ablation studies on the BabyYoda scene within the RobustNeRF dataset. This scene comprises subsets designated for training, testing, and validation. Specifically, the training subset contains distractors, while the testing and", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f020e8c7a457d2652675ed60f0ad88e5eb0611ca046a1da214bb4495e601874a.jpg", + "image_caption": [ + "(a) Adaptability analysis of MemE-NeRF." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 452, + 248 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0ab7dc82ac893e40afde025aa2dadc98da1f9728abb0aa2164ff435b205fb31d.jpg", + "image_caption": [ + "(b) Adaptability analysis of MemE-3DGS.", + "Fig. 6: Adaptability Analysis of MemE on NeRF and 3DGS in Distractor-Free Scenes. MemE demonstrates comparable performance to Mip-NeRF and 3DGS in scenes devoid of distractors. Its adaptability across various scenarios, regardless of the presence of distractors, enhances the versatility of NeRF/3DGS frameworks." + ], + "image_footnote": [], + "bbox": [ + 544, + 143, + 764, + 250 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "validation subsets are devoid of distractors. Additionally, the validation subset encompasses different views from those in the training set.", + "bbox": [ + 212, + 366, + 782, + 396 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Adaptability of MemE. While MemE performs admirably in scenes with distractors, we sought to evaluate its performance in scenarios where the presence of distractors is unknown. We conducted a comparison between Mip-NeRF 360, RobustNeRF*, and 3DGS with our MemE on the distractor-free BabyYoda scene. The PSNR curve during training, as depicted in Fig. 6, demonstrates that our MemE-NeRF achieves performance very similar to Mip-NeRF 360 without experiencing significant performance loss. In contrast, RobustNeRF* suffers from considerable performance degradation in distractor-free scenarios. Similarly, MemE-3DGS exhibits a similar trend compared to 3DGS. Thus, MemE demonstrates adaptability in adjusting the focus of NeRF/3DGS in different types of inputs. That is, in scenes with distractors, MemE can effectively and adaptively eliminate them, while in distractor-free scenes, MemE can focus solely on purity without causing significant performance degradation for NeRF/3DGS. Consequently, MemE serves as a plug-and-play module capable of adaptively handling various scenarios with or without distractors.", + "bbox": [ + 212, + 398, + 787, + 625 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Mask Fusion. In our MemE, we utilize mask fusion to generate purity attention by aggregating binary masks $\\{\\mathbf{M}_n\\}$ across multiple scales. These multi-scale masks consist of fine-grained pixel-wise masks $\\{\\mathbf{M}_i\\}$ and coarse-grained tile-wise masks $\\{\\mathbf{M}_j\\}$ , with pixel-wise masks providing detailed information and tile-wise masks aggregating local information within each residual tile. We explore various strategies for integrating hierarchical masks $\\{\\mathbf{M}_n\\}$ , as detailed in Table 3. Initially, using only the most fine-grained mask $\\mathbf{M}_1$ resulted in a significant drop in PSNR metric due to its inability to aggregate local information, resulting in noise. Conversely, employing solely the pixel-wise masks $\\{\\mathbf{M}_i|i\\in \\{1,3,5,7\\}\\}$ improved performance by leveraging the OR operation to aggregate their advantages: $\\mathbf{M}_1$ providing detail and $\\{\\mathbf{M}_i|i\\in \\{3,5,7\\}\\}$ filtering noise. Additionally, we tested the performance using the MEAN operation on pixel-wise masks, denoted as $\\{\\mathbf{M}_i|i\\in \\{1,3,5,7\\}\\}^*$ in Table 3. Similarly, solely utilizing tile-wise masks also improved results by leveraging hierarchical residual information. However, our", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization", + "bbox": [ + 243, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5659d05990927762b1da5bde25c692fd89d083905555b3ac05f6d2eb8c4b6290.jpg", + "table_caption": [ + "Table 3: Quantitative Results of Ablation Study." + ], + "table_footnote": [], + "table_body": "
MethodsPSNR ↑SSIM ↑LPIPS ↓
w/o mask fusionM130.090.800.31
{Mi|i ∈ {1,3,5,7}}30.990.800.29
{Mi|i ∈ {1,3,5,7}}*30.760.810.29
{Mj|j ∈ {4,8,16}}30.930.800.30
w/o adaptive thresholds{μ1(Γn)}24.230.770.34
{μ2(Γn)}28.500.770.32
linear moving30.740.790.31
Ours31.080.810.29
", + "bbox": [ + 308, + 170, + 694, + 295 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "method, combining both pixel-wise and tile-wise masks, achieved the best performance.", + "bbox": [ + 212, + 324, + 782, + 353 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Adaptive Thresholds. In our MemE, masks are derived by applying adaptive thresholds to hierarchical residual maps $\\{\\epsilon_n\\}$ . These thresholds transition exponentially from the distractor means $\\{\\mu_2(\\Gamma_n)\\}$ to the purity means $\\{\\mu_1(\\Gamma_n)\\}$ . The means $\\mu_1$ , $\\mu_2$ of the GMM are determined from training images of a specific scene at specific scale. We evaluate various threshold-generating strategies, as detailed in Table 3. Initially, we utilize the purity means $\\{\\mu_1(\\Gamma_n)\\}$ as thresholds, aiming to attain purity and exclude distractors. However, this approach leads to a significant performance drop. This drop occurs because the purity means $\\{\\mu_1(\\Gamma_n)\\}$ decrease rapidly due to the memorization effect during the early stages. However, some areas of purity may exhibit large residuals early on, remaining unaccounted for in the training procedure. Conversely, employing the distractor mean $\\{\\mu_2(\\Gamma_n)\\}$ as thresholds introduces some distractors into the training procedure, resulting in a performance decrease. Additionally, we explore the linear threshold strategy, where the thresholds transition linearly from $\\{\\mu_2(\\Gamma_n)\\}$ to $\\{\\mu_1(\\Gamma_n)\\}$ . The ablation results demonstrate that our method achieves the best performance.", + "bbox": [ + 212, + 354, + 787, + 598 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 623, + 359, + 638 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we introduce MemE, an unsupervised plug-and-play module designed to facilitate high-quality novel view synthesis in noisy input scenarios. MemE capitalizes on the inherent property known as the memorization effect during the parameter optimization phase of view synthesis models, enabling effective distractor filtering. MemE seamlessly integrates with both NeRF and 3DGS frameworks. Notably, our work marks the first instance of leveraging the memorization effect in the training process of NeRF and 3DGS. Additionally, MemE proves to be effective in environments, irrespective of the presence of distractors, thereby substantially improving the versatility of NeRF and 3DGS across a variety of input scenarios. We evaluate MemE across three datasets to assess its effectiveness against other SoTA methods. Both quantitative and qualitative results demonstrate that our approach achieves SoTA performance.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Yangdi Lu, Guangchi Fang, and Runqing Jiang for their insightful comments and valuable discussions. This work was partially supported by the National Natural Science Foundation of China (62301601), Guangdong Basic and Applied Basic Research Foundation (2022B1515020103, 2023B1515120087), the Shenzhen Science and Technology Program (No. RCYX20200714114641140).", + "bbox": [ + 212, + 176, + 787, + 253 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 277, + 321, + 292 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arazo, E., Ortego, D., Albert, P., O'Connor, N., McGuinness, K.: Unsupervised label noise modeling and loss correction. In: ICML. pp. 312-321. PMLR (2019)", + "2. Arpit, D., Jastrzegski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M.S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., et al.: A closer look at memorization in deep networks. In: ICML. pp. 233-242 (2017)", + "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV. pp. 5855-5864 (2021)", + "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: CVPR. pp. 5470-5479 (2022)", + "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-NeRF: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)", + "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensorRF: Tensorial radiance fields. In: ECCV. pp. 333-350. Springer (2022)", + "7. Chen, M., Wang, L., Lei, Y., Dong, Z., Guo, Y.: Learning spherical radiance field for efficient $360^{\\circ}$ unbounded novel view synthesis. IEEE Transactions on Image Processing (2024)", + "8. Cordeiro, F.R., Sachdeva, R., Belagiannis, V., Reid, I., Carneiro, G.: LongReMix: Robust learning with high confidence samples in a noisy label environment. PR 133, 109013 (2023)", + "9. Dai, P., Zhang, Y., Yu, X., Lyu, X., Qi, X.: Hybrid neural rendering for large-scale scenes with motion blur. In: CVPR. pp. 154-164 (2023)", + "10. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological) 39(1), 1-22 (1977)", + "1. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR. pp. 5501-5510 (2022)", + "2. Goli, L., Reading, C., Sellán, S., Jacobson, A., Tagliasacchi, A.: Bayes' rays: Uncertainty quantification for neural radiance fields. arXiv preprint arXiv:2309.03185 (2023)", + "3. Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., et al.: Kubric: A scalable dataset generator. In: CVPR. pp. 3749-3761 (2022)", + "4. Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., Tsang, I., Sugiyama, M.: Co-teaching: Robust training of deep neural networks with extremely noisy labels. NeurIPS 31 (2018)" + ], + "bbox": [ + 225, + 310, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, X., Wang, Q.: HDR-NeRF: High dynamic range neural radiance fields. In: CVPR. pp. 18398-18408 (2022)", + "16. Jiang, R., Yan, Y., Xue, J.H., Chen, S., Wang, N., Wang, H.: Knowledge distillation meets label noise learning: Ambiguity-guided mutual label refinery. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2023)", + "17. Jiang, R., Yan, Y., Xue, J.H., Wang, B., Wang, H.: When sparse neural network meets label noise learning: A multistage learning framework. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2022)", + "18. Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. IJCV 129(2), 517-547 (2021)", + "19. Karim, N., Rizve, M.N., Rahnavard, N., Mian, A., Shah, M.: Unicon: Combating label noise through uniform selection and contrastive learning. In: CVPR. pp. 9676-9686 (2022)", + "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG 42(4) (2023)", + "21. Lee, D., Lee, M., Shin, C., Lee, S.: DP-NeRF: Deblurred neural radiance field with physical scene priors. In: CVPR. pp. 12386-12396 (2023)", + "22. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: ExBluRF: Efficient radiance fields for extreme motion blurred images. In: ICCV. pp. 17639-17648 (2023)", + "23. Levy, D., Peleg, A., Pearl, N., Rosenbaum, D., Akkaynak, D., Korman, S., Treibitz, T.: SeaThru-NeRF: Neural radiance fields in scattering media. In: CVPR. pp. 56–65 (2023)", + "24. Li, J., Socher, R., Hoi, S.C.: DivideMix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394 (2020)", + "25. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: CVPR. pp. 6498-6508 (2021)", + "26. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: DynIBaR: Neural dynamic image-based rendering. In: CVPR. pp. 4273-4284 (2023)", + "27. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: CVPR. pp. 13-23 (2023)", + "28. Lu, Y., Zhang, Y., Han, B., Cheung, Y.m., Wang, H.: Label-noise learning with intrinsically long-tailed data. In: ICCV. pp. 1369-1378 (2023)", + "29. Lu, Y., Bo, Y., He, W.: An ensemble model for combating label noise. In: Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining. pp. 608-617 (2022)", + "30. Lu, Y., Bo, Y., He, W.: Noise attention learning: Enhancing noise robustness by gradient scaling. NeurIPS 35, 23164-23177 (2022)", + "31. Lu, Y., He, W.: SELC: Self-ensemble label correction improves learning with noisy labels. arXiv preprint arXiv:2205.01156 (2022)", + "32. Lu, Y., He, W.: Learning with noisy ground truth: From 2d classification to 3d reconstruction. arXiv preprint arXiv:2406.15982 (2024)", + "33. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-NeRF: Neural radiance fields from blurry images. In: CVPR. pp. 12861–12870 (2022)", + "34. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the wild: Neural radiance fields for unconstrained photo collections. In: CVPR. pp. 7210-7219 (2021)", + "35. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRFs in the dark: High dynamic range view synthesis from noisy raw images. In: CVPR. pp. 16190–16199 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "37. Mirzaei, A., Aumentado-Armstrong, T., Derpanis, K.G., Kelly, J., Brubaker, M.A., Gilitschenski, I., Levinshtein, A.: SPIn-NeRF: Multiview segmentation and perceptual inpainting with neural radiance fields. In: CVPR. pp. 20669–20679 (2023)", + "38. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG 41(4), 1-15 (2022)", + "39. Pearl, N., Treibitz, T., Korman, S.: NAN: Noise-aware nerfs for burst-denoising. In: CVPR. pp. 12672-12681 (2022)", + "40. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: CVPR. pp. 12932-12942 (2022)", + "41. Sabour, S., Vora, S., Duckworth, D., Krasin, I., Fleet, D.J., Tagliasacchi, A.: RobustNeRF: Ignoring distractors with robust losses. In: CVPR. pp. 20626-20636 (2023)", + "42. Song, H., Kim, M., Park, D., Shin, Y., Lee, J.G.: Learning from noisy labels with deep neural networks: A survey. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2022)", + "43. Tanay, T., Leonardis, A., Maggioni, M.: Efficient view synthesis and 3d-based multi-frame denoising with multiplane feature representations. In: CVPR. pp. 20898-20907 (2023)", + "44. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-NeRF: Scalable large scene neural view synthesis. In: CVPR. pp. 8248-8258 (2022)", + "45. Teigen, A.L., Yip, M., Hamran, V.P., Skui, V., Stahl, A., Mester, R.: Removing adverse volumetric effects from trained neural radiance fields. arXiv preprint arXiv:2311.10523 (2023)", + "46. Tschernezki, V., Larlus, D., Vedaldi, A.: NeuralDiff: Segmenting 3D objects that move in egocentric videos. In: Proceedings of the IEEE International Conference on 3D Vision (3DV). pp. 910-919 (2021)", + "47. Tu, Y., Zhang, B., Li, Y., Liu, L., Li, J., Wang, Y., Wang, C., Zhao, C.R.: Learning from noisy labels with decoupled meta label purifier. In: CVPR. pp. 19934-19943 (2023)", + "48. Turki, H., Zhang, J.Y., Ferroni, F., Ramanan, D.: SUDS: Scalable urban dynamic scenes. In: CVPR. pp. 12375-12385 (2023)", + "49. Wang, D., Zhang, T., Abboud, A., Süsstrunk, S.: InpaintNeRF360: Text-guided 3d inpainting on unbounded neural radiance fields. arXiv preprint arXiv:2305.15094 (2023)", + "50. Wang, H., Xu, X., Xu, K., Lau, R.W.: Lighting up NeRF via unsupervised decomposition and enhancement. In: ICCV. pp. 12632-12641 (2023)", + "51. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle adjusted deblur neural radiance fields. In: CVPR. pp. 4170-4179 (2023)", + "52. Warburg, F., Weber, E., Tancik, M., Holynski, A., Kanazawa, A.: NeRF-busters: Removing ghostly artifacts from casually captured NeRFs. arXiv preprint arXiv:2304.10532 (2023)", + "53. Weder, S., Garcia-Hernando, G., Monszpart, A., Pollefeys, M., Brostow, G.J., Firman, M., Vicente, S.: Removing objects from neural radiance fields. In: CVPR. pp. 16528-16538 (2023)", + "54. Wei, F., Funkhouser, T., Rusinkiewicz, S.: Clutter detection and removal in 3d scenes with view-consistent inpainting. In: ICCV. pp. 18131-18141 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\\~ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS 35, 32653-32666 (2022)", + "56. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: ICCV. pp. 5752-5761 (2021)", + "57. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: NeRF++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)", + "58. Zhang, M., Zhao, X., Yao, J., Yuan, C., Huang, W.: When noisy labels meet long tail dilemmas: A representation calibration method. In: ICCV. pp. 15890-15900 (2023)" + ], + "bbox": [ + 215, + 146, + 787, + 285 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/b0f9a1b1-12b0-444d-bc6b-1b50e3076e57_model.json b/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/b0f9a1b1-12b0-444d-bc6b-1b50e3076e57_model.json new file mode 100644 index 0000000000000000000000000000000000000000..48ab0a864f43c6418c267df297020a2a5a0dd980 --- /dev/null +++ b/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/b0f9a1b1-12b0-444d-bc6b-1b50e3076e57_model.json @@ -0,0 +1,2370 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.229, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.212, + 0.756, + 0.244 + ], + "angle": 0, + "content": "Yukun Wang\\(^{1}\\), Kunhong Li\\(^{1}\\), Minglin Chen\\(^{1}\\), Longguang Wang\\(^{1}\\), Shunbo Zhou\\(^{2}\\), Kaiwen Xue\\(^{2}\\), and Yulan Guo\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.254, + 0.779, + 0.297 + ], + "angle": 0, + "content": "1 The Shenzhen Campus of Sun Yat-sen University, Sun Yat-sen University {wangyk59,likh25,chenmlin8}@mail2.sysu.edu.cn guoyulan@sysu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.297, + 0.777, + 0.325 + ], + "angle": 0, + "content": "\\(^{2}\\) Huawei Cloud Computing Technical Innovation Dept., Huawei Cloud Computing Technologies Co., Ltd." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.361, + 0.744, + 0.584 + ], + "angle": 0, + "content": "Abstract. Neural Radiance Fields (NeRF) and 3D Gaussian Splitting (3DGS) have greatly advanced novel view synthesis, which is capable of photo-realistic rendering. However, these methods require the foundational assumption of the static scene (e.g., consistent lighting condition and persistent object positions), which is often violated in real-world scenarios. In this study, we introduce MemE, an unsupervised plug-and-play module, to achieve high-quality novel view synthesis in noisy input scenarios. MemE leverages the inherent property in parameter optimization, known as the memorization effect to achieve distractor filtering and can be easily combined with NeRF or 3DGS. Furthermore, MemE is applicable in environments both with and without distractors, significantly enhancing the adaptability of NeRF and 3DGS across diverse input scenarios. Extensive experiments show that our methods (i.e., MemE-NeRF and MemE-3DGS) achieve state-of-the-art performance on both real and synthetic noisy scenes. We will release our code for further research at https://github.com/Yukun66/MemE." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.596, + 0.741, + 0.625 + ], + "angle": 0, + "content": "Keywords: 3D reconstruction \\(\\cdot\\) Unsupervised noise modeling \\(\\cdot\\) Memorization effect" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.377, + 0.667 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Recently, the emergence of methods like Neural Radiance Fields (NeRF) [4,36] and 3D Gaussian Splatting (3DGS) [20] has greatly propelled the development of novel view synthesis. Additionally, the application of novel view synthesis in fields such as robotics, 3D reconstruction, and AR/VR is gradually being explored. However, these methods are built upon certain critical assumptions, including consistent lighting conditions and persistent object positions. These assumptions are frequently violated in real-world scenarios, which leads to artifacts, significant degradation in rendering quality. This study tackles the specific challenge encountered in novel view synthesis: reconstructing the static and clean" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.824, + 0.401, + 0.841 + ], + "angle": 0, + "content": "\\(\\boxtimes\\) Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.151, + 0.782, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.31, + 0.558, + 0.322 + ], + "angle": 0, + "content": "(a) Mip-NeRF 360" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.324, + 0.782, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.43, + 0.574, + 0.443 + ], + "angle": 0, + "content": "(b) MemE-NeRF (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Fig. 1: Memorization Effect in Optimization. The purity (clean scene) and distractor pixels correspond to the blue and red bars in the histograms, respectively. In the training of NeRF and 3DGS, we observed the phenomenon of memorization effect. This phenomenon refers to the parameters initially fitting purity and excluding distractor pixels in the early stages of training, gradually overfitting to distractor pixels in the later stages, as evidenced by the rendered image and histogram in (a). Our method leverages the memorization effect to effectively prevent distractors from overfitted, maintaining the position of the red bars unchanged as much as possible, as demonstrated in (b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.789, + 0.642 + ], + "angle": 0, + "content": "scene from noisy images containing distractors, which disrupt the assumption of the static scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.844 + ], + "angle": 0, + "content": "To comprehend the impact of distractors in novel view synthesis, we analyze the optimization process of NeRF and 3DGS. Specifically, we exemplify this analysis using Mip-NeRF 360 [4], as depicted in Fig. 1a. Our observations indicate that during the initial stages of optimization (e.g., the 5000th step), the image exhibits blurring, yet the scene remains relatively clean. However, as optimization progresses, the image sharpens, concomitant with the emergence of distractors. Further analyzing the accumulated loss distribution, Mip-NeRF 360 primarily focuses on learning purity pixels (of the clean scenes) and leaves most distractor pixels out of the learning process at the early stage, as evidenced by the minimal changes in the histogram of distractor pixels (red bars). This observation, namely memorization effect [2, 42], enables a clear distinction between purity and distractor pixels. Subsequently, we can promote the fitting of purity pixels and suppress the learning of distractor pixels throughout the Mip-NeRF" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "360 training process to easily achieve sharp and clean view rendering in a noisy input scenario." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Based on the above insight, we propose MemE, an unsupervised plug-and-play method that can be incorporated with NeRF and 3DGS, to achieve high-quality novel view synthesis in noisy input scenarios. MemE is based on the discovery of Memorization Effect in NeRF and 3DGS optimization, and the core of MemE is to distinguish distractor pixels without prior knowledge and exclude them throughout the optimization process. Specifically, our MemE begins with our Hierarchical Residual Measurement module to quantify the residuals between the rendered results and the noisy input images, outputting residual vectors for other modules. Subsequently, within our Purity-Distractor Mixture Modeling module, we employ a multi-level Gaussian Mixture Model (GMM) to model the residual distribution of both purity and distractor pixels according to the residual vectors. Finally, we calculate a purity attention weight for NeRF/3DGS optimization. With that attention, we can ensure that the distractor pixels remains unlearned throughout the parameters optimization process, preventing the parameters from overfitting to distractor pixels and thereby achieving the distractor-free novel view synthesis, as shown in Fig. 1b." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.555 + ], + "angle": 0, + "content": "In comparison to prior works, our approach offers distinct advantages: 1) It is solely grounded in the inherent property of NeRF and 3DGS training, i.e., memorization effect, and conducted in an unsupervised manner. 2) Our MemE is capable of processing training data, whether it contains distractors or not, without the need for prior identification. 3) Our method operates independently of pre-trained models or parameters, ensuring simplicity and effectiveness, thus enabling seamless integration as a plug-and-play solution. We incorporate MemE with NeRF (referred to as MemE-NeRF) and 3DGS (referred to as MemE-3DGS) to demonstrate the universality of MemE." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.555, + 0.593, + 0.57 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows," + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.58, + 0.784, + 0.638 + ], + "angle": 0, + "content": "- We are the first work to exploit the memorization effect in NeRF and 3DGS optimization process. Further, we propose MemE-NeRF and MemE-3DGS to achieve high-quality novel view synthesis in the noisy input scenario in an unsupervised manner." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.64, + 0.784, + 0.697 + ], + "angle": 0, + "content": "- We propose MemE, a simple yet plug-and-play method to filter distractor pixels. MemE can adapt to scenes both with and without distractors simultaneously, which can enhance the versatility of NeRF/3DGS in various types of input scenarios." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.7, + 0.784, + 0.728 + ], + "angle": 0, + "content": "- We conduct extensive experiments and the results show that our method achieves state-of-the-art performance on both real and synthetic noisy scenes." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.58, + 0.784, + 0.728 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.753, + 0.388, + 0.769 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.785, + 0.442, + 0.8 + ], + "angle": 0, + "content": "2.1 Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF) [36] has made a breakthrough in novel view synthesis, which is capable of synthesizing photo-realistic images at arbitrary" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "views. NeRF represents a 3D scene as a continuous radiance field, and synthesizes images with differentiable volumetric rendering. Mip-NeRF [3] and Zip-NeRF [5] improve the rendering quality by reducing the aliasing artifacts. \\(\\mathrm{NeRF}++\\) [57] and Mip-NeRF 360 [4] are proposed to extend NeRF to \\(360^{\\circ}\\) unbounded scene [7]. Several works accelerate NeRF by parameterizing the radiance fields with high-efficient data structures, such as multiresolution hash table [38], Octree [11,56], and orthogonal planes [6]. Recently, 3D Gaussian Splatting [20] has emerged as another promising paradigm for novel view synthesis, which synthesizes the image by rendering a set of learnable Gaussian points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Despite great development in novel view synthesis, the aforementioned methods are limited to static scenes with constant light conditions. They are struggling with multi-view images that contain distractors. Our method aims to extend existing methods to distractor-free novel view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.364, + 0.546, + 0.38 + ], + "angle": 0, + "content": "2.2 View Synthesis from Noisy Inputs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Recently, efforts are made to remove noises during NeRF reconstruction, such as training data noise [15, 35, 39, 43, 45, 50], deblurring [9, 21, 22, 33, 51], artifact removal [12, 23, 52], dynamic distractor elimination [34, 40, 41, 44, 55], and scene inpainting [37, 49, 53, 54]. To address transient occluders, NeRF-W [34] employs multilayer perceptrons (MLPs) modeling for static and transient components separately. Methods for removing movable objects in large-scale street scenes [40, 44] have utilized semantic segmentation models to explicitly filter out dynamic outliers, such as cars and pedestrians. Recent advancements have extended NeRF to video data [26, 27, 48], introducing an auxiliary temporal MLP for separately modeling dynamic and static scenes. RobustNeRF [41] processes discrete images from casual viewpoints, ranking training pixel errors to identify and eliminate distractors using hand-crafted ratios, which may yield suboptimal results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.676 + ], + "angle": 0, + "content": "Most of the aforementioned methods do not consider the case of distractor noise, while the most related counterpart of our method is RobustNeRF [41]. Compared to RobustNeRF, our method is capable of scenes with or without distractors, with no need for analysis and artificial prior on the distractor content of the scene in advance. Our method can achieve similar results to Mip-NeRF 360 and 3DGS in distractor-free scenes and does not require special adjustment of hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.696, + 0.442, + 0.712 + ], + "angle": 0, + "content": "2.3 Noisy Label Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Numerous methods [1, 14, 16, 17, 19, 24, 28-32, 47, 58] have investigated the challenge of label noise in the field of image classification due to the difficulty in data labeling [42]. Prior arts on the memorization effect [1, 2, 42] indicates that neural networks tend first to learn simple and clean samples, and then gradually memorize noisy labeled samples. Based on this, several efforts are devoted to combat label noise by selecting clean samples from noisy dataset during model training, thus avoiding model from overfitting to those noisy samples. Co-teaching [14] retains two networks, and each network selects a portion of samples with small-loss" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.151, + 0.778, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.46 + ], + "angle": 0, + "content": "Fig. 2: The Framework of MemE. Our MemE integrates three primary components: Hierarchical Residual Measurement (Sec. 3.1), Purity-Distractor Mixture Modeling (Sec. 3.2), and Purity Attention Generation (Sec. 3.3). Hierarchical Residual Measurement module generates residual vectors \\(\\{\\pmb{\\Gamma}_n\\}\\) and residual maps \\(\\{\\pmb{\\epsilon}_n\\}\\), which are then utilized by the subsequent modules. Purity-Distractor Mixture Modeling module fits the multi-level Gaussian Mixture Model (GMM) according to \\(\\{\\pmb{\\Gamma}_n\\}\\). Finally, the Purity Attention Generation module creates attention weight \\(\\mathbf{M}\\) to exclude distractors selectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.596 + ], + "angle": 0, + "content": "and feeds them to another network. Note that above methods discard intensive noisy samples, which inevitably incurs information loss, limiting their performance especially when the noise ratio is high. To remedy this, DivideMix [24] employs Gaussian mixture models to split training dataset into labeled (clean) and unlabeled (noisy) sets, and then adopts MixMatch to make full use of those noisy samples. Besides, LongReMix [8] also uses MixMatch strategy and augments high-confidence samples to achieve superior robustness." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.331, + 0.635 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.728 + ], + "angle": 0, + "content": "In this section, we present a novel plug-and-play method (namely MemE), which is aimed at enhancing existing novel view synthesis methods when multi-view images contain discriminators. Existing novel view synthesis methods represent a 3D scene as a parameterized model \\(\\mathbf{S}_{\\theta}\\), while facilitating a differentiable rendering technique \\(\\mathcal{R}(\\cdot ,\\cdot)\\) to render an image \\(\\hat{\\mathbf{I}}\\) as following:" + }, + { + "type": "equation", + "bbox": [ + 0.45, + 0.74, + 0.787, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {I}} = \\mathcal {R} (\\mathbf {S} _ {\\theta}, \\pi), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\pi\\) is the image pose. Concretely, Neural Radiance Fields [36] leverages the Multilayer Perceptron (MLP) and volumetric rendering, whereas 3D Gaussian Splatting [20] is built upon a collection of learnable Gaussian points and splatting rasterization. Because multi-view images may contain distractors, we employ the proposed MemE to generate an attention weight \\(\\mathbf{M}\\) to indicate the distractor" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "pixels. Therefore, existing methods with our MemE can achieve distractor-free novel view synthesis using the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.189, + 0.786, + 0.219 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) = \\sum_ {r} \\mathbf {M} _ {r} | | \\hat {\\mathbf {I}} _ {r} - \\mathbf {I} _ {r} | | _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.232, + 0.784, + 0.262 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{I}}_r\\), \\(\\mathbf{I}_r\\) and \\(\\mathbf{M}_r\\) are the rendered color, ground-truth color, and attention weight at pixel \\(r\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.263, + 0.789, + 0.43 + ], + "angle": 0, + "content": "The framework of our MemE is depicted in Fig. 2. Our method consists of three components: hierarchical residual measurement (Sec. 3.1), purity-distractor mixture modeling (Sec. 3.2), and purity attention generation (Sec. 3.3). The hierarchical residual measurement module quantifies the residuals between noisy input \\(\\mathbf{I}\\) and rendered result \\(\\hat{\\mathbf{I}}\\). The purity-distractor mixture modeling module fits the distribution of residuals. The purity attention generation module generates a noisy-excluded attention mask \\(\\mathbf{M}\\) for loss adjusting. By composing the three simple modules, MemE exploits the memorization effect existing in NeRF and 3DGS, and models the purity-distractor distribution of the scene in an unsupervised manner. The resulting mask \\(\\mathbf{M}\\) can effectively help NeRF and 3DGS focus only on purity and ignore distractors." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.451, + 0.557, + 0.466 + ], + "angle": 0, + "content": "3.1 Hierarchical Residual Measurement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.567 + ], + "angle": 0, + "content": "MemE identifies which parts of the scene are prioritized for learning by calculating residuals of rendering results and noisy input. According to the observed memorization effect, the areas learned first more likely belong to purity, whereas distractors are more likely to be left out of training in the early stage. Further, we employ residual measurement on multiple scales to provide more comprehensive local information hierarchically." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Given a noisy input patch \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W}\\) and a rendering result patch \\(\\hat{\\mathbf{I}} \\in \\mathbb{R}^{H \\times W}\\), the Hierarchical Residual Measurement module eventually outputs hierarchical residual vectors and residual maps to the other modules. Specifically, the residual map between \\(\\hat{\\mathbf{I}}\\) and \\(\\mathbf{I}\\) is first calculated by the L2 norms, noted as \\(\\epsilon\\). Then a residual pyramid \\(\\{\\epsilon_n\\}\\) is generated by pixel-wise and tile-wise operations according to \\(\\epsilon\\). Finally, \\(\\{\\epsilon_n\\}\\) are simply flattened to vectors \\(\\{\\Gamma_n\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.458, + 0.671, + 0.785, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\epsilon} _ {n} = \\tau_ {n} \\odot \\boldsymbol {\\epsilon}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.454, + 0.713, + 0.785, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Gamma} _ {n} = \\nu \\odot \\boldsymbol {\\epsilon} _ {n}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.765 + ], + "angle": 0, + "content": "where \\(\\nu\\) is the vectorization operation to flatten the patch, \\(\\tau_{n}\\) represents the pixel-wise and tile-wise operations, and \\(\\odot\\) represents the operation deployment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The pixel-wise operations are the average convolutions with kernels of \\(\\mathbb{R}^{i\\times i}\\), \\(i\\in \\mathcal{I}\\), stride 1 and padding, where \\(\\mathcal{I}\\) refers to the set of pixel-wise scales. As a result, the multi-level residuals \\(\\epsilon_{i}\\) are with spatial size \\(\\mathbb{R}^{H\\times W}\\). The tile-wise operations are the average pooling with kernels of \\(\\mathbb{R}^{j\\times j}\\), \\(j\\in \\mathcal{J}\\), stride \\(j\\) and no padding, where \\(\\mathcal{J}\\) refers to the set of tile-wise scales. Thus, the multi-level" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.193 + ], + "angle": 0, + "content": "residuals \\(\\pmb{\\epsilon}_{j}\\) are with spatial size \\(\\mathbb{R}^{H / j\\times W / j}\\). In conclusion, the outputs of Hierarchical Residual Measurement are multi-level residual maps \\(\\{\\pmb{\\epsilon}_n\\}\\) and residual vectors \\(\\{\\pmb{\\Gamma}_n\\}\\), where \\(n\\in \\mathcal{I}\\cup \\mathcal{J}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.213, + 0.563, + 0.23 + ], + "angle": 0, + "content": "3.2 Purity-Distractor Mixture Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.237, + 0.784, + 0.312 + ], + "angle": 0, + "content": "Based on the residual vectors \\(\\{\\pmb{\\Gamma}_n\\}\\) from the Hierarchical Residual Measurement module, it is necessary to use a model to fit the distribution of purity and distractors in the NeRF and 3DGS parameter optimization process, so that MemE can use the model to identify purity and distractors in the next module, and calculate the mask to adjust training loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.785, + 0.388 + ], + "angle": 0, + "content": "In the Purity-Distractor Mixture Modeling module, a one-dimensional and two-component GMM is utilized to fit the residual distributions of purity and distractor pixels because the residual distribution tends to be bi-modal in the early training stage. The probability density function (pdf) of a mixture model of \\(K\\) components, applied to a residual vector \\(\\pmb{\\Gamma}_{n}\\), is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.399, + 0.785, + 0.44 + ], + "angle": 0, + "content": "\\[\np \\left(\\boldsymbol {\\Gamma} _ {n}\\right) = \\sum_ {k = 1} ^ {K} \\lambda_ {k} \\cdot \\mathcal {N} \\left(\\boldsymbol {\\Gamma} _ {n} \\mid \\mu_ {k}, \\sigma_ {k} ^ {2}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.784, + 0.524 + ], + "angle": 0, + "content": "where \\(\\lambda_{k}\\) represents the mixing coefficients for the convex combination of each pdf \\(\\mathcal{N}(\\pmb{\\Gamma}_n\\mid \\mu_k,\\sigma_k^2)\\). The fitting of GMM is conducted through the Expectation-Maximization (EM) algorithm [10]. As the residual vectors \\(\\{\\pmb{\\Gamma}_n\\}\\) are with multi-level, the GMM is implemented with each level respectively, and the output is a set of pdfs expressed as \\(\\{p(\\pmb{\\Gamma}_n)\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.785, + 0.614 + ], + "angle": 0, + "content": "Additionally, to achieve robust training, we update \\(\\{p(\\boldsymbol{\\Gamma}_n)\\}\\) every \\(T\\) steps. That is, we accumulate \\(T\\) residual vectors \\(\\{\\boldsymbol{\\Gamma}_n^t | t \\in \\{0,1,\\dots,T-1\\}\\}\\) for each level, and compute \\(\\{p(\\cup \\boldsymbol{\\Gamma}_n^t)\\}\\) all at once. For convenience, we continue to use the symbol \\(\\{p(\\boldsymbol{\\Gamma}_n)\\}\\) instead of \\(\\{p(\\cup \\boldsymbol{\\Gamma}_n^t)\\}\\) to denote the output pdfs in subsequent sections. And at the very beginning \\(T\\) steps, we do not employ our MemE for training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.635, + 0.499, + 0.65 + ], + "angle": 0, + "content": "3.3 Purity Attention Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.784, + 0.719 + ], + "angle": 0, + "content": "Utilized the residual maps \\(\\{\\pmb{\\epsilon}_n\\}\\) and pdf sets \\(\\{p(\\pmb{\\Gamma}_n)\\}\\) obtained from two former modules, Purity Attention Generation outputs an attention map \\(\\mathbf{M}\\) adjusting loss function to constraint view synthesis model only focuses on the purity and repel distractors out of optimization procedure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.81 + ], + "angle": 0, + "content": "Specifically, for the residual maps processed through tile-wise operations, we upsample their spatial size to the original size, i.e., \\(\\mathbb{R}^{H\\times W}\\), using nearest interpolation. Then, for spatially aligned residual maps \\(\\{\\pmb {\\epsilon}_n\\}\\), we have corresponding pdfs \\(\\{p(\\pmb {\\Gamma}_n)\\}\\), and we compute adaptive thresholds \\(\\{\\rho_n\\}\\) according to \\(\\{p(\\pmb {\\Gamma}_n)\\}\\). Next, the binary masks \\(\\{\\mathbf{M}_n|\\mathbf{M}_n\\triangleq \\pmb {\\epsilon}_n\\leq \\rho_n\\}\\) are generated according to \\(\\{\\rho_{n}\\}\\). Finally, \\(\\{\\mathbf{M}_n\\}\\) are fused into a purity attention map \\(\\mathbf{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Adaptive Threshold. We apply a threshold to the residual map to derive the mask at each level. Since our destination is to repel distractors from the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "optimization process, we hope this threshold is: 1) smaller than the most of residuals of distractors in the total optimization process; 2) slightly larger than most of the residuals of purity at the later stage of optimization to preserve most purity and make sure to exclude the distractors as many as possible; and 3) starting from a moderate value and gradually decrease as the optimization progresses because purity pixels may also have large residual values in the early stages of optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.789, + 0.364 + ], + "angle": 0, + "content": "Therefore, in each scale, we calculate a threshold from the position lying between the two means of \\( p(\\pmb{\\Gamma}_n) \\), i.e., \\( \\mu_1(\\pmb{\\Gamma}_n) \\) and \\( \\mu_2(\\pmb{\\Gamma}_n) \\). There are three reasons as follows: 1) the means indicate the most concentrated residual magnitude of the purity and distractors; 2) the \\( \\mu_1(\\pmb{\\Gamma}_n) \\) is the ideal value of our threshold at the end of optimization; and 3) the \\( \\mu_2(\\pmb{\\Gamma}_n) \\) provide a moderate starting point lying between the ideal value \\( \\mu_1(\\pmb{\\Gamma}_n) \\) and the unconstrained value, i.e., 1. Further, we design adaptive thresholds that exponentially move from \\( \\mu_2(\\pmb{\\Gamma}_n) \\) to \\( \\mu_1(\\pmb{\\Gamma}_n) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.379, + 0.786, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\rho_ {n} = \\alpha \\cdot \\mu_ {1} (\\boldsymbol {\\Gamma} _ {n}) + (1 - \\alpha) \\cdot \\mu_ {2} (\\boldsymbol {\\Gamma} _ {n}), \\quad \\alpha = 1 - e ^ {\\left(- \\omega \\cdot \\frac {s}{N}\\right)} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.789, + 0.475 + ], + "angle": 0, + "content": "where \\(\\omega\\) controls the moving velocity, \\(s\\) and \\(N\\) are current training steps and total steps, and \\(\\rho_{n}\\) is the threshold for generating the binary mask \\(\\mathbf{M}_n\\), where 0 refers to the prediction as distractors and 1 refers to the prediction as purity. Collecting the masks of different scales, we have a set of binary masks \\(\\{\\mathbf{M}_n\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.79, + 0.613 + ], + "angle": 0, + "content": "Mask Fusion. After the binary masks \\(\\{\\mathbf{M}_n\\}\\) are generated, we further fuse the multi-level masks into a single purity attention map \\(\\mathbf{M}\\). These masks contain information at multiple scales and are fused in this module. We split the \\(\\{\\mathbf{M}_n\\}\\) into pixel-wise operated subset \\(\\{\\mathbf{M}_i|i\\in \\mathcal{I}\\}\\) and tile-wise operated subset \\(\\{\\mathbf{M}_j|j\\in \\mathcal{J}\\}\\). Generally, the masks in \\(\\{\\mathbf{M}_i\\}\\) are fine-grained and the masks in \\(\\{\\mathbf{M}_j\\}\\) are coarse grained. Within the \\(\\{\\mathbf{M}_i\\}\\), the mask with smaller \\(i\\) is detailed but noisy; while the mask with larger \\(i\\) is noise-filtered but blurry. As such, we first fuse the \\(\\{\\mathbf{M}_i\\}\\) with OR operation, noted as \\(\\oplus\\), to combine their advantages including detailed and noise-filtered, generating the fused pixel-wise mask \\(\\mathbf{M}^p\\):" + }, + { + "type": "equation", + "bbox": [ + 0.429, + 0.631, + 0.786, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} ^ {p} = \\oplus \\left\\{\\mathbf {M} _ {i} \\mid i \\in \\mathcal {I} \\right\\}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.725 + ], + "angle": 0, + "content": "The OR operation removes the noise of value 0 in binary masks while preserving the noise of value 1. As such we combine the hierarchical tile-wise masks \\(\\{\\mathbf{M}_j\\}\\) providing more comprehensive local information to generate the purity attention map \\(\\mathbf{M}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.733, + 0.786, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\frac {1}{\\mathcal {D}} \\left(\\mathbf {M} ^ {p} + \\sum_ {j} \\left(\\mathbf {M} _ {j}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) means the total number of masks, including \\(\\{\\mathbf{M}_j\\}\\) and \\(\\mathbf{M}^p\\). We use \\(\\mathcal{I} = \\{1,3,5,7\\}\\) and \\(\\mathcal{J} = \\{4,8,16\\}\\) in our method. Therefore, the elements of purity attention map \\(\\mathbf{M}\\) are discrete values belongs to the set \\(\\{0,0.25,0.5,0.75,1\\}\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.146, + 0.375, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.177, + 0.334, + 0.19 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.201, + 0.783, + 0.274 + ], + "angle": 0, + "content": "RobustNeRF Dataset. The RobustNeRF dataset [41] comprises four different indoor scenes. Within each scene, there exists a sub-training set containing distractors and a sub-testing set without distractors. Additionally, in the BabyYoda scene, there is an additional sub-validation set devoid of distractors, featuring views distinct from those in the training set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.783, + 0.335 + ], + "angle": 0, + "content": "Kubric Dataset. The Kubric dataset introduced in [55], is a synthetic dataset comprising five scenes, generated using the Kubric dataset generator [13]. Each scene comprises an empty room with a texture-less background and several static objects. In each frame, there are distractors present with continuous motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.783, + 0.41 + ], + "angle": 0, + "content": "Phototourism Dataset. The Phototourism dataset [18] is a typical outdoor scene dataset, which comprises images collected from the internet, thus containing numerous distractors such as tourists and illumination changes. We randomly select six scenes from the Phototourism dataset to assess the performance of our MemE." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.436, + 0.561, + 0.451 + ], + "angle": 0, + "content": "4.2 Metrics and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.783, + 0.594 + ], + "angle": 0, + "content": "We report PSNR, SSIM, and LPIPS to quantitatively evaluate the proposed MemE-NeRF and MemE-3DGS. MemE-NeRF is trained with 64 or 16 training patches per step for every scene. To ensure a fair comparison with RobustNeRF reported in [41], where training was conducted with 64 patches per step, we also train RobustNeRF with 16 patches (denoted as RobustNeRF*). MemE-NeRF is based on Mip-NeRF 360 [4] implemented using JAX, while MemE-3DGS is based on 3DGS [20] implemented using PyTorch. Experiments are conducted on 4 NVIDIA RTX-3090 GPUs for 64 patches and 1 for 16 patches. All models are trained with images containing distractors while tested on distractor-free images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.351, + 0.633 + ], + "angle": 0, + "content": "4.3 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.783, + 0.73 + ], + "angle": 0, + "content": "RobustNeRF Dataset. On the RobustNeRF dataset, our MemE outperforms the counterparts, as shown in Table 1 and Fig. 3. MemE-NeRF surpasses RobustNeRF across most metrics, while MemE-3DGS significantly enhances the performance of 3DGS in scenes containing distractors. Notably, our MemE-NeRF achieves comparable or even higher PSNR than RobustNeRF, despite being trained with fewer patches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.783, + 0.838 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, our MemE-NeRF accurately removes distractors while preserving more details. In Fig. 3 BabyYoda, MemE-NeRF effectively eliminates distractors and retains the detailed patterns of the table. In contrast, although RobustNeRF* removes distractors, it introduces incorrect floaters and misses details of the table. In Fig. 3 Statue, both MemE-NeRF and MemE-3DGS successfully remove distractors. However, RobustNeRF* is susceptible to contamination by distractors, leading to missing details and floaters. Compared" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.287, + 0.145, + 0.715, + 0.16 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison on the RobustNeRF dataset." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.172, + 0.782, + 0.312 + ], + "angle": 0, + "content": "
Methodsp†StatueAndroidCrabBabyYoda
LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑
Mip-NeRF 360 (L2) [4]640.360.6619.090.400.6519.350.270.7725.730.310.7522.97
Mip-NeRF 360 (L1) [4]640.300.7219.550.400.6619.380.220.7926.690.220.8026.15
Mip-NeRF 360 (Ch.) [4]640.300.7319.640.400.6619.530.210.8027.720.220.8025.22
D2NeRF [55]-0.480.4919.090.430.5720.610.420.6821.180.440.6517.32
RobustNeRF [41]640.280.7520.890.310.6521.720.210.8130.750.200.8330.87
MemE-NeRF (Ours)640.230.7321.990.300.6821.740.200.8330.640.250.8331.36
RobustNeRF* [41]160.330.6621.900.310.6721.400.240.8230.580.310.7929.76
MemE-NeRF (Ours)160.310.6821.810.300.6821.540.230.8330.610.290.8131.08
3DGS [20]-0.200.7620.220.250.6821.540.190.8925.790.190.9128.12
MemE-3DGS (Ours)-0.140.8121.740.180.7322.720.190.9025.810.190.9128.40
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.223, + 0.312, + 0.443, + 0.323 + ], + "angle": 0, + "content": "the number of patches in each training step." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.34, + 0.784, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Fig.3: Qualitative Comparison on the RobustNeRF Dataset. Compared to the baselines, i.e., Mip-NeRF 360 and 3DGS, our MemE-NeRF and MemE-3DGS not only efficiently eliminate distractors but also retain a higher level of detail. In comparison with RobustNeRF*, MemE models demonstrate superior performance in removing distractors, leading to a reduction in artifacts and an enhancement in detail preservation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "with 3DGS, MemE-3DGS exhibits significant improvements in both qualitative and quantitative comparisons. 3DGS is prone to overfitting distractors, resulting in incorrect view synthesis ( Fig. 3 Statue and Crab) or artifacts in the rendered images ( Fig. 3 Android and Baby Yoda). It is noteworthy that RobustNeRF leads to significant performance degradation in clean scene training, whereas the integration of our MemE is non-intrusive to existing NeRF and 3DGS frameworks, as demonstrated in [41] and Sec. 4.4." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.307, + 0.145, + 0.695, + 0.16 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison on the Kubric dataset." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.172, + 0.784, + 0.289 + ], + "angle": 0, + "content": "
MethodsCarCarsBagChairsPillow
LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑
NeRF-W [34].218.81424.23.243.87324.51.139.79120.65.150.68123.77.088.93528.24
NSFF [25].200.80624.90.620.37610.29.108.89225.62.682.28412.82.782.3434.55
NeuralDiff [46].065.95231.89.098.92125.93.117.91029.02.112.72224.42.565.65220.09
D2NeRF [55].062.97534.27.090.95326.27.076.97934.14.095.70724.63.076.97936.58
RobustNeRF [41].013.98837.73.063.95726.31.006.99541.82.007.99241.23.018.99038.95
RobustNeRF* [41].021.97340.83.020.98643.70.027.98331.62.009.99346.46.026.98644.23
MemE-NeRF (Ours).021.97340.86.019.98744.34.003.99748.79.007.99447.87.016.98945.97
3DGS [20].037.96638.36.027.98341.85.005.99545.94.011.99144.76.034.98341.09
MemE-3DGS (Ours).032.96840.28.015.98945.40.002.99749.11.008.99447.82.016.98944.29
" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.313, + 0.78, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.691, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Fig. 4: Qualitative Comparison on the Kubric Dataset. Our MemE effectively manages scenes with distractors that exhibit continuous motion across successive frames, resulting in better distractor elimination and static object preserving." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Kubric Dataset. Compared with previous works, our MemE achieves consistent improvement in terms of all scenes and all metrics, as shown in Table 2 and Fig. 4. Note that the distractors in the Kubric dataset exhibit continuous motion between different frames, increasing the probability of overlap, particularly for objects with simple textures. Leveraging the inherent memorization" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.149, + 0.778, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.449, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Fig.5: Quality Results of 3DGS and MemE-3DGS(Ours) on the Phototourism Dataset. The pairs of tri-column figures illustrate the test view images and the rendering images of 3DGS and MemE-3DGS, respectively. Our MemE-3DGS effectively eliminates artifacts caused by distractors and significant illumination variations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.613 + ], + "angle": 0, + "content": "effect in view synthesis models, our MemE utilizes GMM to adaptively perform customized fitting for each scene. Consequently, our MemE enhances the performance of NeRF and 3DGS to achieve SoTA results, regardless of whether the inputs are temporally continuous or discrete, and remains robust to inputs with varying distractor ratios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.737 + ], + "angle": 0, + "content": "Phototourism Dataset. The Phototourism dataset remains challenging for existing novel view synthesis methods due to the diverse sources of images. This variability results in differences in distractors, illumination, and even camera equipment across images. Nevertheless, our MemE-3DGS achieves satisfactory view rendering results on this dataset, as shown in Fig. 5. Notably, severe lighting variations, as well as distractors, can significantly impact the quality of rendered images of 3DGS. However, our MemE is capable of effectively mitigating the influence of lighting variations and distractors." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.764, + 0.391, + 0.78 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We conduct ablation studies on the BabyYoda scene within the RobustNeRF dataset. This scene comprises subsets designated for training, testing, and validation. Specifically, the training subset contains distractors, while the testing and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.454, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.251, + 0.468, + 0.263 + ], + "angle": 0, + "content": "(a) Adaptability analysis of MemE-NeRF." + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.145, + 0.765, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.251, + 0.78, + 0.264 + ], + "angle": 0, + "content": "(b) Adaptability analysis of MemE-3DGS." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.275, + 0.789, + 0.331 + ], + "angle": 0, + "content": "Fig. 6: Adaptability Analysis of MemE on NeRF and 3DGS in Distractor-Free Scenes. MemE demonstrates comparable performance to Mip-NeRF and 3DGS in scenes devoid of distractors. Its adaptability across various scenarios, regardless of the presence of distractors, enhances the versatility of NeRF/3DGS frameworks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.784, + 0.397 + ], + "angle": 0, + "content": "validation subsets are devoid of distractors. Additionally, the validation subset encompasses different views from those in the training set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.626 + ], + "angle": 0, + "content": "Adaptability of MemE. While MemE performs admirably in scenes with distractors, we sought to evaluate its performance in scenarios where the presence of distractors is unknown. We conducted a comparison between Mip-NeRF 360, RobustNeRF*, and 3DGS with our MemE on the distractor-free BabyYoda scene. The PSNR curve during training, as depicted in Fig. 6, demonstrates that our MemE-NeRF achieves performance very similar to Mip-NeRF 360 without experiencing significant performance loss. In contrast, RobustNeRF* suffers from considerable performance degradation in distractor-free scenarios. Similarly, MemE-3DGS exhibits a similar trend compared to 3DGS. Thus, MemE demonstrates adaptability in adjusting the focus of NeRF/3DGS in different types of inputs. That is, in scenes with distractors, MemE can effectively and adaptively eliminate them, while in distractor-free scenes, MemE can focus solely on purity without causing significant performance degradation for NeRF/3DGS. Consequently, MemE serves as a plug-and-play module capable of adaptively handling various scenarios with or without distractors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Mask Fusion. In our MemE, we utilize mask fusion to generate purity attention by aggregating binary masks \\(\\{\\mathbf{M}_n\\}\\) across multiple scales. These multi-scale masks consist of fine-grained pixel-wise masks \\(\\{\\mathbf{M}_i\\}\\) and coarse-grained tile-wise masks \\(\\{\\mathbf{M}_j\\}\\), with pixel-wise masks providing detailed information and tile-wise masks aggregating local information within each residual tile. We explore various strategies for integrating hierarchical masks \\(\\{\\mathbf{M}_n\\}\\), as detailed in Table 3. Initially, using only the most fine-grained mask \\(\\mathbf{M}_1\\) resulted in a significant drop in PSNR metric due to its inability to aggregate local information, resulting in noise. Conversely, employing solely the pixel-wise masks \\(\\{\\mathbf{M}_i|i\\in \\{1,3,5,7\\}\\}\\) improved performance by leveraging the OR operation to aggregate their advantages: \\(\\mathbf{M}_1\\) providing detail and \\(\\{\\mathbf{M}_i|i\\in \\{3,5,7\\}\\}\\) filtering noise. Additionally, we tested the performance using the MEAN operation on pixel-wise masks, denoted as \\(\\{\\mathbf{M}_i|i\\in \\{1,3,5,7\\}\\}^*\\) in Table 3. Similarly, solely utilizing tile-wise masks also improved results by leveraging hierarchical residual information. However, our" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.334, + 0.145, + 0.668, + 0.16 + ], + "angle": 0, + "content": "Table 3: Quantitative Results of Ablation Study." + }, + { + "type": "table", + "bbox": [ + 0.31, + 0.171, + 0.695, + 0.296 + ], + "angle": 0, + "content": "
MethodsPSNR ↑SSIM ↑LPIPS ↓
w/o mask fusionM130.090.800.31
{Mi|i ∈ {1,3,5,7}}30.990.800.29
{Mi|i ∈ {1,3,5,7}}*30.760.810.29
{Mj|j ∈ {4,8,16}}30.930.800.30
w/o adaptive thresholds{μ1(Γn)}24.230.770.34
{μ2(Γn)}28.500.770.32
linear moving30.740.790.31
Ours31.080.810.29
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.784, + 0.354 + ], + "angle": 0, + "content": "method, combining both pixel-wise and tile-wise masks, achieved the best performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.599 + ], + "angle": 0, + "content": "Adaptive Thresholds. In our MemE, masks are derived by applying adaptive thresholds to hierarchical residual maps \\(\\{\\epsilon_n\\}\\). These thresholds transition exponentially from the distractor means \\(\\{\\mu_2(\\Gamma_n)\\}\\) to the purity means \\(\\{\\mu_1(\\Gamma_n)\\}\\). The means \\(\\mu_1\\), \\(\\mu_2\\) of the GMM are determined from training images of a specific scene at specific scale. We evaluate various threshold-generating strategies, as detailed in Table 3. Initially, we utilize the purity means \\(\\{\\mu_1(\\Gamma_n)\\}\\) as thresholds, aiming to attain purity and exclude distractors. However, this approach leads to a significant performance drop. This drop occurs because the purity means \\(\\{\\mu_1(\\Gamma_n)\\}\\) decrease rapidly due to the memorization effect during the early stages. However, some areas of purity may exhibit large residuals early on, remaining unaccounted for in the training procedure. Conversely, employing the distractor mean \\(\\{\\mu_2(\\Gamma_n)\\}\\) as thresholds introduces some distractors into the training procedure, resulting in a performance decrease. Additionally, we explore the linear threshold strategy, where the thresholds transition linearly from \\(\\{\\mu_2(\\Gamma_n)\\}\\) to \\(\\{\\mu_1(\\Gamma_n)\\}\\). The ablation results demonstrate that our method achieves the best performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.624, + 0.36, + 0.64 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this study, we introduce MemE, an unsupervised plug-and-play module designed to facilitate high-quality novel view synthesis in noisy input scenarios. MemE capitalizes on the inherent property known as the memorization effect during the parameter optimization phase of view synthesis models, enabling effective distractor filtering. MemE seamlessly integrates with both NeRF and 3DGS frameworks. Notably, our work marks the first instance of leveraging the memorization effect in the training process of NeRF and 3DGS. Additionally, MemE proves to be effective in environments, irrespective of the presence of distractors, thereby substantially improving the versatility of NeRF and 3DGS across a variety of input scenarios. We evaluate MemE across three datasets to assess its effectiveness against other SoTA methods. Both quantitative and qualitative results demonstrate that our approach achieves SoTA performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.254 + ], + "angle": 0, + "content": "We thank Yangdi Lu, Guangchi Fang, and Runqing Jiang for their insightful comments and valuable discussions. This work was partially supported by the National Natural Science Foundation of China (62301601), Guangdong Basic and Applied Basic Research Foundation (2022B1515020103, 2023B1515120087), the Shenzhen Science and Technology Program (No. RCYX20200714114641140)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.278, + 0.323, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.339 + ], + "angle": 0, + "content": "1. Arazo, E., Ortego, D., Albert, P., O'Connor, N., McGuinness, K.: Unsupervised label noise modeling and loss correction. In: ICML. pp. 312-321. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.381 + ], + "angle": 0, + "content": "2. Arpit, D., Jastrzegski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M.S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., et al.: A closer look at memorization in deep networks. In: ICML. pp. 233-242 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.382, + 0.785, + 0.422 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV. pp. 5855-5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.785, + 0.463 + ], + "angle": 0, + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: CVPR. pp. 5470-5479 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.465, + 0.785, + 0.505 + ], + "angle": 0, + "content": "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-NeRF: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.785, + 0.534 + ], + "angle": 0, + "content": "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensorRF: Tensorial radiance fields. In: ECCV. pp. 333-350. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.575 + ], + "angle": 0, + "content": "7. Chen, M., Wang, L., Lei, Y., Dong, Z., Guo, Y.: Learning spherical radiance field for efficient \\(360^{\\circ}\\) unbounded novel view synthesis. IEEE Transactions on Image Processing (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.617 + ], + "angle": 0, + "content": "8. Cordeiro, F.R., Sachdeva, R., Belagiannis, V., Reid, I., Carneiro, G.: LongReMix: Robust learning with high confidence samples in a noisy label environment. PR 133, 109013 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.785, + 0.645 + ], + "angle": 0, + "content": "9. Dai, P., Zhang, Y., Yu, X., Lyu, X., Qi, X.: Hybrid neural rendering for large-scale scenes with motion blur. In: CVPR. pp. 154-164 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.687 + ], + "angle": 0, + "content": "10. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological) 39(1), 1-22 (1977)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "1. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR. pp. 5501-5510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "2. Goli, L., Reading, C., Sellán, S., Jacobson, A., Tagliasacchi, A.: Bayes' rays: Uncertainty quantification for neural radiance fields. arXiv preprint arXiv:2309.03185 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "3. Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., et al.: Kubric: A scalable dataset generator. In: CVPR. pp. 3749-3761 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., Tsang, I., Sugiyama, M.: Co-teaching: Robust training of deep neural networks with extremely noisy labels. NeurIPS 31 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "15. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, X., Wang, Q.: HDR-NeRF: High dynamic range neural radiance fields. In: CVPR. pp. 18398-18408 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "16. Jiang, R., Yan, Y., Xue, J.H., Chen, S., Wang, N., Wang, H.: Knowledge distillation meets label noise learning: Ambiguity-guided mutual label refinery. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "17. Jiang, R., Yan, Y., Xue, J.H., Wang, B., Wang, H.: When sparse neural network meets label noise learning: A multistage learning framework. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "18. Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. IJCV 129(2), 517-547 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.301, + 0.785, + 0.341 + ], + "angle": 0, + "content": "19. Karim, N., Rizve, M.N., Rahnavard, N., Mian, A., Shah, M.: Unicon: Combating label noise through uniform selection and contrastive learning. In: CVPR. pp. 9676-9686 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG 42(4) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "21. Lee, D., Lee, M., Shin, C., Lee, S.: DP-NeRF: Deblurred neural radiance field with physical scene priors. In: CVPR. pp. 12386-12396 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "22. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: ExBluRF: Efficient radiance fields for extreme motion blurred images. In: ICCV. pp. 17639-17648 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "23. Levy, D., Peleg, A., Pearl, N., Rosenbaum, D., Akkaynak, D., Korman, S., Treibitz, T.: SeaThru-NeRF: Neural radiance fields in scattering media. In: CVPR. pp. 56–65 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "24. Li, J., Socher, R., Hoi, S.C.: DivideMix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "25. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: CVPR. pp. 6498-6508 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.549 + ], + "angle": 0, + "content": "26. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: DynIBaR: Neural dynamic image-based rendering. In: CVPR. pp. 4273-4284 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "27. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: CVPR. pp. 13-23 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "28. Lu, Y., Zhang, Y., Han, B., Cheung, Y.m., Wang, H.: Label-noise learning with intrinsically long-tailed data. In: ICCV. pp. 1369-1378 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "29. Lu, Y., Bo, Y., He, W.: An ensemble model for combating label noise. In: Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining. pp. 608-617 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "30. Lu, Y., Bo, Y., He, W.: Noise attention learning: Enhancing noise robustness by gradient scaling. NeurIPS 35, 23164-23177 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "31. Lu, Y., He, W.: SELC: Self-ensemble label correction improves learning with noisy labels. arXiv preprint arXiv:2205.01156 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "32. Lu, Y., He, W.: Learning with noisy ground truth: From 2d classification to 3d reconstruction. arXiv preprint arXiv:2406.15982 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "33. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-NeRF: Neural radiance fields from blurry images. In: CVPR. pp. 12861–12870 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "34. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the wild: Neural radiance fields for unconstrained photo collections. In: CVPR. pp. 7210-7219 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "35. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRFs in the dark: High dynamic range view synthesis from noisy raw images. In: CVPR. pp. 16190–16199 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "36. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "37. Mirzaei, A., Aumentado-Armstrong, T., Derpanis, K.G., Kelly, J., Brubaker, M.A., Gilitschenski, I., Levinshtein, A.: SPIn-NeRF: Multiview segmentation and perceptual inpainting with neural radiance fields. In: CVPR. pp. 20669–20679 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "38. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "39. Pearl, N., Treibitz, T., Korman, S.: NAN: Noise-aware nerfs for burst-denoising. In: CVPR. pp. 12672-12681 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "40. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: CVPR. pp. 12932-12942 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "41. Sabour, S., Vora, S., Duckworth, D., Krasin, I., Fleet, D.J., Tagliasacchi, A.: RobustNeRF: Ignoring distractors with robust losses. In: CVPR. pp. 20626-20636 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.397 + ], + "angle": 0, + "content": "42. Song, H., Kim, M., Park, D., Shin, Y., Lee, J.G.: Learning from noisy labels with deep neural networks: A survey. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "43. Tanay, T., Leonardis, A., Maggioni, M.: Efficient view synthesis and 3d-based multi-frame denoising with multiplane feature representations. In: CVPR. pp. 20898-20907 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "44. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-NeRF: Scalable large scene neural view synthesis. In: CVPR. pp. 8248-8258 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "45. Teigen, A.L., Yip, M., Hamran, V.P., Skui, V., Stahl, A., Mester, R.: Removing adverse volumetric effects from trained neural radiance fields. arXiv preprint arXiv:2311.10523 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "46. Tschernezki, V., Larlus, D., Vedaldi, A.: NeuralDiff: Segmenting 3D objects that move in egocentric videos. In: Proceedings of the IEEE International Conference on 3D Vision (3DV). pp. 910-919 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "47. Tu, Y., Zhang, B., Li, Y., Liu, L., Li, J., Wang, Y., Wang, C., Zhao, C.R.: Learning from noisy labels with decoupled meta label purifier. In: CVPR. pp. 19934-19943 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "48. Turki, H., Zhang, J.Y., Ferroni, F., Ramanan, D.: SUDS: Scalable urban dynamic scenes. In: CVPR. pp. 12375-12385 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "49. Wang, D., Zhang, T., Abboud, A., Süsstrunk, S.: InpaintNeRF360: Text-guided 3d inpainting on unbounded neural radiance fields. arXiv preprint arXiv:2305.15094 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.702 + ], + "angle": 0, + "content": "50. Wang, H., Xu, X., Xu, K., Lau, R.W.: Lighting up NeRF via unsupervised decomposition and enhancement. In: ICCV. pp. 12632-12641 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.729 + ], + "angle": 0, + "content": "51. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle adjusted deblur neural radiance fields. In: CVPR. pp. 4170-4179 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.771 + ], + "angle": 0, + "content": "52. Warburg, F., Weber, E., Tancik, M., Holynski, A., Kanazawa, A.: NeRF-busters: Removing ghostly artifacts from casually captured NeRFs. arXiv preprint arXiv:2304.10532 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "53. Weder, S., Garcia-Hernando, G., Monszpart, A., Pollefeys, M., Brostow, G.J., Firman, M., Vicente, S.: Removing objects from neural radiance fields. In: CVPR. pp. 16528-16538 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "54. Wei, F., Funkhouser, T., Rusinkiewicz, S.: Clutter detection and removal in 3d scenes with view-consistent inpainting. In: ICCV. pp. 18131-18141 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "55. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\\~ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS 35, 32653-32666 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.218 + ], + "angle": 0, + "content": "56. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: ICCV. pp. 5752-5761 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.788, + 0.246 + ], + "angle": 0, + "content": "57. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: NeRF++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.788, + 0.286 + ], + "angle": 0, + "content": "58. Zhang, M., Zhao, X., Yao, J., Yuan, C., Huang, W.: When noisy labels meet long tail dilemmas: A representation calibration method. In: ICCV. pp. 15890-15900 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.286 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/b0f9a1b1-12b0-444d-bc6b-1b50e3076e57_origin.pdf b/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/b0f9a1b1-12b0-444d-bc6b-1b50e3076e57_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3a4fd41da27dc92dbe8d08c9e74e2bcfe8a1ef59 --- /dev/null +++ b/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/b0f9a1b1-12b0-444d-bc6b-1b50e3076e57_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1482a5604d2dcaf253a0c4cdfd2bb3ccf5df00e552107974b0b40f316cc35b4b +size 7021425 diff --git a/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/full.md b/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..31cec41bce0461c5723c3b15b07ce3cdda2e8a02 --- /dev/null +++ b/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/full.md @@ -0,0 +1,293 @@ +# Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization + +Yukun Wang $^{1}$ , Kunhong Li $^{1}$ , Minglin Chen $^{1}$ , Longguang Wang $^{1}$ , Shunbo Zhou $^{2}$ , Kaiwen Xue $^{2}$ , and Yulan Guo $^{1}$ + +1 The Shenzhen Campus of Sun Yat-sen University, Sun Yat-sen University {wangyk59,likh25,chenmlin8}@mail2.sysu.edu.cn guoyulan@sysu.edu.cn + +$^{2}$ Huawei Cloud Computing Technical Innovation Dept., Huawei Cloud Computing Technologies Co., Ltd. + +Abstract. Neural Radiance Fields (NeRF) and 3D Gaussian Splitting (3DGS) have greatly advanced novel view synthesis, which is capable of photo-realistic rendering. However, these methods require the foundational assumption of the static scene (e.g., consistent lighting condition and persistent object positions), which is often violated in real-world scenarios. In this study, we introduce MemE, an unsupervised plug-and-play module, to achieve high-quality novel view synthesis in noisy input scenarios. MemE leverages the inherent property in parameter optimization, known as the memorization effect to achieve distractor filtering and can be easily combined with NeRF or 3DGS. Furthermore, MemE is applicable in environments both with and without distractors, significantly enhancing the adaptability of NeRF and 3DGS across diverse input scenarios. Extensive experiments show that our methods (i.e., MemE-NeRF and MemE-3DGS) achieve state-of-the-art performance on both real and synthetic noisy scenes. We will release our code for further research at https://github.com/Yukun66/MemE. + +Keywords: 3D reconstruction $\cdot$ Unsupervised noise modeling $\cdot$ Memorization effect + +# 1 Introduction + +Recently, the emergence of methods like Neural Radiance Fields (NeRF) [4,36] and 3D Gaussian Splatting (3DGS) [20] has greatly propelled the development of novel view synthesis. Additionally, the application of novel view synthesis in fields such as robotics, 3D reconstruction, and AR/VR is gradually being explored. However, these methods are built upon certain critical assumptions, including consistent lighting conditions and persistent object positions. These assumptions are frequently violated in real-world scenarios, which leads to artifacts, significant degradation in rendering quality. This study tackles the specific challenge encountered in novel view synthesis: reconstructing the static and clean + +![](images/2bad65dcaaad2d1b27fe879cba36185daf0054fa09ce77754447584c07811a95.jpg) + +![](images/516c99d15bb5c83ee3ea6d17b311edda59c98166d6cf99c7ca4a7c93b8b53a7f.jpg) +(a) Mip-NeRF 360 +(b) MemE-NeRF (Ours) +Fig. 1: Memorization Effect in Optimization. The purity (clean scene) and distractor pixels correspond to the blue and red bars in the histograms, respectively. In the training of NeRF and 3DGS, we observed the phenomenon of memorization effect. This phenomenon refers to the parameters initially fitting purity and excluding distractor pixels in the early stages of training, gradually overfitting to distractor pixels in the later stages, as evidenced by the rendered image and histogram in (a). Our method leverages the memorization effect to effectively prevent distractors from overfitted, maintaining the position of the red bars unchanged as much as possible, as demonstrated in (b). + +scene from noisy images containing distractors, which disrupt the assumption of the static scene. + +To comprehend the impact of distractors in novel view synthesis, we analyze the optimization process of NeRF and 3DGS. Specifically, we exemplify this analysis using Mip-NeRF 360 [4], as depicted in Fig. 1a. Our observations indicate that during the initial stages of optimization (e.g., the 5000th step), the image exhibits blurring, yet the scene remains relatively clean. However, as optimization progresses, the image sharpens, concomitant with the emergence of distractors. Further analyzing the accumulated loss distribution, Mip-NeRF 360 primarily focuses on learning purity pixels (of the clean scenes) and leaves most distractor pixels out of the learning process at the early stage, as evidenced by the minimal changes in the histogram of distractor pixels (red bars). This observation, namely memorization effect [2, 42], enables a clear distinction between purity and distractor pixels. Subsequently, we can promote the fitting of purity pixels and suppress the learning of distractor pixels throughout the Mip-NeRF + +360 training process to easily achieve sharp and clean view rendering in a noisy input scenario. + +Based on the above insight, we propose MemE, an unsupervised plug-and-play method that can be incorporated with NeRF and 3DGS, to achieve high-quality novel view synthesis in noisy input scenarios. MemE is based on the discovery of Memorization Effect in NeRF and 3DGS optimization, and the core of MemE is to distinguish distractor pixels without prior knowledge and exclude them throughout the optimization process. Specifically, our MemE begins with our Hierarchical Residual Measurement module to quantify the residuals between the rendered results and the noisy input images, outputting residual vectors for other modules. Subsequently, within our Purity-Distractor Mixture Modeling module, we employ a multi-level Gaussian Mixture Model (GMM) to model the residual distribution of both purity and distractor pixels according to the residual vectors. Finally, we calculate a purity attention weight for NeRF/3DGS optimization. With that attention, we can ensure that the distractor pixels remains unlearned throughout the parameters optimization process, preventing the parameters from overfitting to distractor pixels and thereby achieving the distractor-free novel view synthesis, as shown in Fig. 1b. + +In comparison to prior works, our approach offers distinct advantages: 1) It is solely grounded in the inherent property of NeRF and 3DGS training, i.e., memorization effect, and conducted in an unsupervised manner. 2) Our MemE is capable of processing training data, whether it contains distractors or not, without the need for prior identification. 3) Our method operates independently of pre-trained models or parameters, ensuring simplicity and effectiveness, thus enabling seamless integration as a plug-and-play solution. We incorporate MemE with NeRF (referred to as MemE-NeRF) and 3DGS (referred to as MemE-3DGS) to demonstrate the universality of MemE. + +Our contributions can be summarized as follows, + +- We are the first work to exploit the memorization effect in NeRF and 3DGS optimization process. Further, we propose MemE-NeRF and MemE-3DGS to achieve high-quality novel view synthesis in the noisy input scenario in an unsupervised manner. +- We propose MemE, a simple yet plug-and-play method to filter distractor pixels. MemE can adapt to scenes both with and without distractors simultaneously, which can enhance the versatility of NeRF/3DGS in various types of input scenarios. +- We conduct extensive experiments and the results show that our method achieves state-of-the-art performance on both real and synthetic noisy scenes. + +# 2 Related Work + +# 2.1 Novel View Synthesis + +Neural Radiance Fields (NeRF) [36] has made a breakthrough in novel view synthesis, which is capable of synthesizing photo-realistic images at arbitrary + +views. NeRF represents a 3D scene as a continuous radiance field, and synthesizes images with differentiable volumetric rendering. Mip-NeRF [3] and Zip-NeRF [5] improve the rendering quality by reducing the aliasing artifacts. $\mathrm{NeRF}++$ [57] and Mip-NeRF 360 [4] are proposed to extend NeRF to $360^{\circ}$ unbounded scene [7]. Several works accelerate NeRF by parameterizing the radiance fields with high-efficient data structures, such as multiresolution hash table [38], Octree [11,56], and orthogonal planes [6]. Recently, 3D Gaussian Splatting [20] has emerged as another promising paradigm for novel view synthesis, which synthesizes the image by rendering a set of learnable Gaussian points. + +Despite great development in novel view synthesis, the aforementioned methods are limited to static scenes with constant light conditions. They are struggling with multi-view images that contain distractors. Our method aims to extend existing methods to distractor-free novel view synthesis. + +# 2.2 View Synthesis from Noisy Inputs + +Recently, efforts are made to remove noises during NeRF reconstruction, such as training data noise [15, 35, 39, 43, 45, 50], deblurring [9, 21, 22, 33, 51], artifact removal [12, 23, 52], dynamic distractor elimination [34, 40, 41, 44, 55], and scene inpainting [37, 49, 53, 54]. To address transient occluders, NeRF-W [34] employs multilayer perceptrons (MLPs) modeling for static and transient components separately. Methods for removing movable objects in large-scale street scenes [40, 44] have utilized semantic segmentation models to explicitly filter out dynamic outliers, such as cars and pedestrians. Recent advancements have extended NeRF to video data [26, 27, 48], introducing an auxiliary temporal MLP for separately modeling dynamic and static scenes. RobustNeRF [41] processes discrete images from casual viewpoints, ranking training pixel errors to identify and eliminate distractors using hand-crafted ratios, which may yield suboptimal results. + +Most of the aforementioned methods do not consider the case of distractor noise, while the most related counterpart of our method is RobustNeRF [41]. Compared to RobustNeRF, our method is capable of scenes with or without distractors, with no need for analysis and artificial prior on the distractor content of the scene in advance. Our method can achieve similar results to Mip-NeRF 360 and 3DGS in distractor-free scenes and does not require special adjustment of hyperparameters. + +# 2.3 Noisy Label Learning + +Numerous methods [1, 14, 16, 17, 19, 24, 28-32, 47, 58] have investigated the challenge of label noise in the field of image classification due to the difficulty in data labeling [42]. Prior arts on the memorization effect [1, 2, 42] indicates that neural networks tend first to learn simple and clean samples, and then gradually memorize noisy labeled samples. Based on this, several efforts are devoted to combat label noise by selecting clean samples from noisy dataset during model training, thus avoiding model from overfitting to those noisy samples. Co-teaching [14] retains two networks, and each network selects a portion of samples with small-loss + +![](images/75e2e49fee8494af2307f02d549e5e1e0790b8680c021a6fbd7b7dbe0243591e.jpg) +Fig. 2: The Framework of MemE. Our MemE integrates three primary components: Hierarchical Residual Measurement (Sec. 3.1), Purity-Distractor Mixture Modeling (Sec. 3.2), and Purity Attention Generation (Sec. 3.3). Hierarchical Residual Measurement module generates residual vectors $\{\pmb{\Gamma}_n\}$ and residual maps $\{\pmb{\epsilon}_n\}$ , which are then utilized by the subsequent modules. Purity-Distractor Mixture Modeling module fits the multi-level Gaussian Mixture Model (GMM) according to $\{\pmb{\Gamma}_n\}$ . Finally, the Purity Attention Generation module creates attention weight $\mathbf{M}$ to exclude distractors selectively. + +and feeds them to another network. Note that above methods discard intensive noisy samples, which inevitably incurs information loss, limiting their performance especially when the noise ratio is high. To remedy this, DivideMix [24] employs Gaussian mixture models to split training dataset into labeled (clean) and unlabeled (noisy) sets, and then adopts MixMatch to make full use of those noisy samples. Besides, LongReMix [8] also uses MixMatch strategy and augments high-confidence samples to achieve superior robustness. + +# 3 Method + +In this section, we present a novel plug-and-play method (namely MemE), which is aimed at enhancing existing novel view synthesis methods when multi-view images contain discriminators. Existing novel view synthesis methods represent a 3D scene as a parameterized model $\mathbf{S}_{\theta}$ , while facilitating a differentiable rendering technique $\mathcal{R}(\cdot ,\cdot)$ to render an image $\hat{\mathbf{I}}$ as following: + +$$ +\hat {\mathbf {I}} = \mathcal {R} (\mathbf {S} _ {\theta}, \pi), \tag {1} +$$ + +where $\pi$ is the image pose. Concretely, Neural Radiance Fields [36] leverages the Multilayer Perceptron (MLP) and volumetric rendering, whereas 3D Gaussian Splatting [20] is built upon a collection of learnable Gaussian points and splatting rasterization. Because multi-view images may contain distractors, we employ the proposed MemE to generate an attention weight $\mathbf{M}$ to indicate the distractor + +pixels. Therefore, existing methods with our MemE can achieve distractor-free novel view synthesis using the following loss: + +$$ +\mathcal {L} (\theta) = \sum_ {r} \mathbf {M} _ {r} | | \hat {\mathbf {I}} _ {r} - \mathbf {I} _ {r} | | _ {2} ^ {2}, \tag {2} +$$ + +where $\hat{\mathbf{I}}_r$ , $\mathbf{I}_r$ and $\mathbf{M}_r$ are the rendered color, ground-truth color, and attention weight at pixel $r$ , respectively. + +The framework of our MemE is depicted in Fig. 2. Our method consists of three components: hierarchical residual measurement (Sec. 3.1), purity-distractor mixture modeling (Sec. 3.2), and purity attention generation (Sec. 3.3). The hierarchical residual measurement module quantifies the residuals between noisy input $\mathbf{I}$ and rendered result $\hat{\mathbf{I}}$ . The purity-distractor mixture modeling module fits the distribution of residuals. The purity attention generation module generates a noisy-excluded attention mask $\mathbf{M}$ for loss adjusting. By composing the three simple modules, MemE exploits the memorization effect existing in NeRF and 3DGS, and models the purity-distractor distribution of the scene in an unsupervised manner. The resulting mask $\mathbf{M}$ can effectively help NeRF and 3DGS focus only on purity and ignore distractors. + +# 3.1 Hierarchical Residual Measurement + +MemE identifies which parts of the scene are prioritized for learning by calculating residuals of rendering results and noisy input. According to the observed memorization effect, the areas learned first more likely belong to purity, whereas distractors are more likely to be left out of training in the early stage. Further, we employ residual measurement on multiple scales to provide more comprehensive local information hierarchically. + +Given a noisy input patch $\mathbf{I} \in \mathbb{R}^{H \times W}$ and a rendering result patch $\hat{\mathbf{I}} \in \mathbb{R}^{H \times W}$ , the Hierarchical Residual Measurement module eventually outputs hierarchical residual vectors and residual maps to the other modules. Specifically, the residual map between $\hat{\mathbf{I}}$ and $\mathbf{I}$ is first calculated by the L2 norms, noted as $\epsilon$ . Then a residual pyramid $\{\epsilon_n\}$ is generated by pixel-wise and tile-wise operations according to $\epsilon$ . Finally, $\{\epsilon_n\}$ are simply flattened to vectors $\{\Gamma_n\}$ : + +$$ +\boldsymbol {\epsilon} _ {n} = \tau_ {n} \odot \boldsymbol {\epsilon}, \tag {3} +$$ + +$$ +\boldsymbol {\Gamma} _ {n} = \nu \odot \boldsymbol {\epsilon} _ {n}, \tag {4} +$$ + +where $\nu$ is the vectorization operation to flatten the patch, $\tau_{n}$ represents the pixel-wise and tile-wise operations, and $\odot$ represents the operation deployment. + +The pixel-wise operations are the average convolutions with kernels of $\mathbb{R}^{i\times i}$ , $i\in \mathcal{I}$ , stride 1 and padding, where $\mathcal{I}$ refers to the set of pixel-wise scales. As a result, the multi-level residuals $\epsilon_{i}$ are with spatial size $\mathbb{R}^{H\times W}$ . The tile-wise operations are the average pooling with kernels of $\mathbb{R}^{j\times j}$ , $j\in \mathcal{J}$ , stride $j$ and no padding, where $\mathcal{J}$ refers to the set of tile-wise scales. Thus, the multi-level + +residuals $\pmb{\epsilon}_{j}$ are with spatial size $\mathbb{R}^{H / j\times W / j}$ . In conclusion, the outputs of Hierarchical Residual Measurement are multi-level residual maps $\{\pmb{\epsilon}_n\}$ and residual vectors $\{\pmb{\Gamma}_n\}$ , where $n\in \mathcal{I}\cup \mathcal{J}$ . + +# 3.2 Purity-Distractor Mixture Modeling + +Based on the residual vectors $\{\pmb{\Gamma}_n\}$ from the Hierarchical Residual Measurement module, it is necessary to use a model to fit the distribution of purity and distractors in the NeRF and 3DGS parameter optimization process, so that MemE can use the model to identify purity and distractors in the next module, and calculate the mask to adjust training loss. + +In the Purity-Distractor Mixture Modeling module, a one-dimensional and two-component GMM is utilized to fit the residual distributions of purity and distractor pixels because the residual distribution tends to be bi-modal in the early training stage. The probability density function (pdf) of a mixture model of $K$ components, applied to a residual vector $\pmb{\Gamma}_{n}$ , is defined as: + +$$ +p \left(\boldsymbol {\Gamma} _ {n}\right) = \sum_ {k = 1} ^ {K} \lambda_ {k} \cdot \mathcal {N} \left(\boldsymbol {\Gamma} _ {n} \mid \mu_ {k}, \sigma_ {k} ^ {2}\right), \tag {5} +$$ + +where $\lambda_{k}$ represents the mixing coefficients for the convex combination of each pdf $\mathcal{N}(\pmb{\Gamma}_n\mid \mu_k,\sigma_k^2)$ . The fitting of GMM is conducted through the Expectation-Maximization (EM) algorithm [10]. As the residual vectors $\{\pmb{\Gamma}_n\}$ are with multi-level, the GMM is implemented with each level respectively, and the output is a set of pdfs expressed as $\{p(\pmb{\Gamma}_n)\}$ . + +Additionally, to achieve robust training, we update $\{p(\boldsymbol{\Gamma}_n)\}$ every $T$ steps. That is, we accumulate $T$ residual vectors $\{\boldsymbol{\Gamma}_n^t | t \in \{0,1,\dots,T-1\}\}$ for each level, and compute $\{p(\cup \boldsymbol{\Gamma}_n^t)\}$ all at once. For convenience, we continue to use the symbol $\{p(\boldsymbol{\Gamma}_n)\}$ instead of $\{p(\cup \boldsymbol{\Gamma}_n^t)\}$ to denote the output pdfs in subsequent sections. And at the very beginning $T$ steps, we do not employ our MemE for training. + +# 3.3 Purity Attention Generation + +Utilized the residual maps $\{\pmb{\epsilon}_n\}$ and pdf sets $\{p(\pmb{\Gamma}_n)\}$ obtained from two former modules, Purity Attention Generation outputs an attention map $\mathbf{M}$ adjusting loss function to constraint view synthesis model only focuses on the purity and repel distractors out of optimization procedure. + +Specifically, for the residual maps processed through tile-wise operations, we upsample their spatial size to the original size, i.e., $\mathbb{R}^{H\times W}$ , using nearest interpolation. Then, for spatially aligned residual maps $\{\pmb {\epsilon}_n\}$ , we have corresponding pdfs $\{p(\pmb {\Gamma}_n)\}$ , and we compute adaptive thresholds $\{\rho_n\}$ according to $\{p(\pmb {\Gamma}_n)\}$ . Next, the binary masks $\{\mathbf{M}_n|\mathbf{M}_n\triangleq \pmb {\epsilon}_n\leq \rho_n\}$ are generated according to $\{\rho_{n}\}$ . Finally, $\{\mathbf{M}_n\}$ are fused into a purity attention map $\mathbf{M}$ . + +Adaptive Threshold. We apply a threshold to the residual map to derive the mask at each level. Since our destination is to repel distractors from the + +optimization process, we hope this threshold is: 1) smaller than the most of residuals of distractors in the total optimization process; 2) slightly larger than most of the residuals of purity at the later stage of optimization to preserve most purity and make sure to exclude the distractors as many as possible; and 3) starting from a moderate value and gradually decrease as the optimization progresses because purity pixels may also have large residual values in the early stages of optimization. + +Therefore, in each scale, we calculate a threshold from the position lying between the two means of $p(\pmb{\Gamma}_n)$ , i.e., $\mu_1(\pmb{\Gamma}_n)$ and $\mu_2(\pmb{\Gamma}_n)$ . There are three reasons as follows: 1) the means indicate the most concentrated residual magnitude of the purity and distractors; 2) the $\mu_1(\pmb{\Gamma}_n)$ is the ideal value of our threshold at the end of optimization; and 3) the $\mu_2(\pmb{\Gamma}_n)$ provide a moderate starting point lying between the ideal value $\mu_1(\pmb{\Gamma}_n)$ and the unconstrained value, i.e., 1. Further, we design adaptive thresholds that exponentially move from $\mu_2(\pmb{\Gamma}_n)$ to $\mu_1(\pmb{\Gamma}_n)$ : + +$$ +\rho_ {n} = \alpha \cdot \mu_ {1} (\boldsymbol {\Gamma} _ {n}) + (1 - \alpha) \cdot \mu_ {2} (\boldsymbol {\Gamma} _ {n}), \quad \alpha = 1 - e ^ {\left(- \omega \cdot \frac {s}{N}\right)} \tag {6} +$$ + +where $\omega$ controls the moving velocity, $s$ and $N$ are current training steps and total steps, and $\rho_{n}$ is the threshold for generating the binary mask $\mathbf{M}_n$ , where 0 refers to the prediction as distractors and 1 refers to the prediction as purity. Collecting the masks of different scales, we have a set of binary masks $\{\mathbf{M}_n\}$ . + +Mask Fusion. After the binary masks $\{\mathbf{M}_n\}$ are generated, we further fuse the multi-level masks into a single purity attention map $\mathbf{M}$ . These masks contain information at multiple scales and are fused in this module. We split the $\{\mathbf{M}_n\}$ into pixel-wise operated subset $\{\mathbf{M}_i|i\in \mathcal{I}\}$ and tile-wise operated subset $\{\mathbf{M}_j|j\in \mathcal{J}\}$ . Generally, the masks in $\{\mathbf{M}_i\}$ are fine-grained and the masks in $\{\mathbf{M}_j\}$ are coarse grained. Within the $\{\mathbf{M}_i\}$ , the mask with smaller $i$ is detailed but noisy; while the mask with larger $i$ is noise-filtered but blurry. As such, we first fuse the $\{\mathbf{M}_i\}$ with OR operation, noted as $\oplus$ , to combine their advantages including detailed and noise-filtered, generating the fused pixel-wise mask $\mathbf{M}^p$ : + +$$ +\mathbf {M} ^ {p} = \oplus \left\{\mathbf {M} _ {i} \mid i \in \mathcal {I} \right\}. \tag {7} +$$ + +The OR operation removes the noise of value 0 in binary masks while preserving the noise of value 1. As such we combine the hierarchical tile-wise masks $\{\mathbf{M}_j\}$ providing more comprehensive local information to generate the purity attention map $\mathbf{M}$ : + +$$ +\mathbf {M} = \frac {1}{\mathcal {D}} \left(\mathbf {M} ^ {p} + \sum_ {j} \left(\mathbf {M} _ {j}\right)\right), \tag {8} +$$ + +where $\mathcal{D}$ means the total number of masks, including $\{\mathbf{M}_j\}$ and $\mathbf{M}^p$ . We use $\mathcal{I} = \{1,3,5,7\}$ and $\mathcal{J} = \{4,8,16\}$ in our method. Therefore, the elements of purity attention map $\mathbf{M}$ are discrete values belongs to the set $\{0,0.25,0.5,0.75,1\}$ . + +# 4 Experiments + +# 4.1 Datasets + +RobustNeRF Dataset. The RobustNeRF dataset [41] comprises four different indoor scenes. Within each scene, there exists a sub-training set containing distractors and a sub-testing set without distractors. Additionally, in the BabyYoda scene, there is an additional sub-validation set devoid of distractors, featuring views distinct from those in the training set. + +Kubric Dataset. The Kubric dataset introduced in [55], is a synthetic dataset comprising five scenes, generated using the Kubric dataset generator [13]. Each scene comprises an empty room with a texture-less background and several static objects. In each frame, there are distractors present with continuous motion. + +Phototourism Dataset. The Phototourism dataset [18] is a typical outdoor scene dataset, which comprises images collected from the internet, thus containing numerous distractors such as tourists and illumination changes. We randomly select six scenes from the Phototourism dataset to assess the performance of our MemE. + +# 4.2 Metrics and Implementation Details + +We report PSNR, SSIM, and LPIPS to quantitatively evaluate the proposed MemE-NeRF and MemE-3DGS. MemE-NeRF is trained with 64 or 16 training patches per step for every scene. To ensure a fair comparison with RobustNeRF reported in [41], where training was conducted with 64 patches per step, we also train RobustNeRF with 16 patches (denoted as RobustNeRF*). MemE-NeRF is based on Mip-NeRF 360 [4] implemented using JAX, while MemE-3DGS is based on 3DGS [20] implemented using PyTorch. Experiments are conducted on 4 NVIDIA RTX-3090 GPUs for 64 patches and 1 for 16 patches. All models are trained with images containing distractors while tested on distractor-free images. + +# 4.3 Evaluation + +RobustNeRF Dataset. On the RobustNeRF dataset, our MemE outperforms the counterparts, as shown in Table 1 and Fig. 3. MemE-NeRF surpasses RobustNeRF across most metrics, while MemE-3DGS significantly enhances the performance of 3DGS in scenes containing distractors. Notably, our MemE-NeRF achieves comparable or even higher PSNR than RobustNeRF, despite being trained with fewer patches. + +As illustrated in Fig. 3, our MemE-NeRF accurately removes distractors while preserving more details. In Fig. 3 BabyYoda, MemE-NeRF effectively eliminates distractors and retains the detailed patterns of the table. In contrast, although RobustNeRF* removes distractors, it introduces incorrect floaters and misses details of the table. In Fig. 3 Statue, both MemE-NeRF and MemE-3DGS successfully remove distractors. However, RobustNeRF* is susceptible to contamination by distractors, leading to missing details and floaters. Compared + +Table 1: Quantitative comparison on the RobustNeRF dataset. + +
Methodsp†StatueAndroidCrabBabyYoda
LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑
Mip-NeRF 360 (L2) [4]640.360.6619.090.400.6519.350.270.7725.730.310.7522.97
Mip-NeRF 360 (L1) [4]640.300.7219.550.400.6619.380.220.7926.690.220.8026.15
Mip-NeRF 360 (Ch.) [4]640.300.7319.640.400.6619.530.210.8027.720.220.8025.22
D2NeRF [55]-0.480.4919.090.430.5720.610.420.6821.180.440.6517.32
RobustNeRF [41]640.280.7520.890.310.6521.720.210.8130.750.200.8330.87
MemE-NeRF (Ours)640.230.7321.990.300.6821.740.200.8330.640.250.8331.36
RobustNeRF* [41]160.330.6621.900.310.6721.400.240.8230.580.310.7929.76
MemE-NeRF (Ours)160.310.6821.810.300.6821.540.230.8330.610.290.8131.08
3DGS [20]-0.200.7620.220.250.6821.540.190.8925.790.190.9128.12
MemE-3DGS (Ours)-0.140.8121.740.180.7322.720.190.9025.810.190.9128.40
+ +the number of patches in each training step. + +![](images/91f3b21e872f9dc506acbc673f69f00d709faa67f8d6530c561a500fe6eedd99.jpg) +Fig.3: Qualitative Comparison on the RobustNeRF Dataset. Compared to the baselines, i.e., Mip-NeRF 360 and 3DGS, our MemE-NeRF and MemE-3DGS not only efficiently eliminate distractors but also retain a higher level of detail. In comparison with RobustNeRF*, MemE models demonstrate superior performance in removing distractors, leading to a reduction in artifacts and an enhancement in detail preservation. + +with 3DGS, MemE-3DGS exhibits significant improvements in both qualitative and quantitative comparisons. 3DGS is prone to overfitting distractors, resulting in incorrect view synthesis ( Fig. 3 Statue and Crab) or artifacts in the rendered images ( Fig. 3 Android and Baby Yoda). It is noteworthy that RobustNeRF leads to significant performance degradation in clean scene training, whereas the integration of our MemE is non-intrusive to existing NeRF and 3DGS frameworks, as demonstrated in [41] and Sec. 4.4. + +Table 2: Quantitative comparison on the Kubric dataset. + +
MethodsCarCarsBagChairsPillow
LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑
NeRF-W [34].218.81424.23.243.87324.51.139.79120.65.150.68123.77.088.93528.24
NSFF [25].200.80624.90.620.37610.29.108.89225.62.682.28412.82.782.3434.55
NeuralDiff [46].065.95231.89.098.92125.93.117.91029.02.112.72224.42.565.65220.09
D2NeRF [55].062.97534.27.090.95326.27.076.97934.14.095.70724.63.076.97936.58
RobustNeRF [41].013.98837.73.063.95726.31.006.99541.82.007.99241.23.018.99038.95
RobustNeRF* [41].021.97340.83.020.98643.70.027.98331.62.009.99346.46.026.98644.23
MemE-NeRF (Ours).021.97340.86.019.98744.34.003.99748.79.007.99447.87.016.98945.97
3DGS [20].037.96638.36.027.98341.85.005.99545.94.011.99144.76.034.98341.09
MemE-3DGS (Ours).032.96840.28.015.98945.40.002.99749.11.008.99447.82.016.98944.29
+ +![](images/3dcb2d4a3a60e5096771e39f238c7d03a2e1e2717f75daca3e44cee8f220723e.jpg) +Fig. 4: Qualitative Comparison on the Kubric Dataset. Our MemE effectively manages scenes with distractors that exhibit continuous motion across successive frames, resulting in better distractor elimination and static object preserving. + +Kubric Dataset. Compared with previous works, our MemE achieves consistent improvement in terms of all scenes and all metrics, as shown in Table 2 and Fig. 4. Note that the distractors in the Kubric dataset exhibit continuous motion between different frames, increasing the probability of overlap, particularly for objects with simple textures. Leveraging the inherent memorization + +![](images/c81c7c924afc35e6c52105fd2dd0c2072fbb89e60a259fb2e2f78feb1fb0b00e.jpg) +Fig.5: Quality Results of 3DGS and MemE-3DGS(Ours) on the Phototourism Dataset. The pairs of tri-column figures illustrate the test view images and the rendering images of 3DGS and MemE-3DGS, respectively. Our MemE-3DGS effectively eliminates artifacts caused by distractors and significant illumination variations. + +effect in view synthesis models, our MemE utilizes GMM to adaptively perform customized fitting for each scene. Consequently, our MemE enhances the performance of NeRF and 3DGS to achieve SoTA results, regardless of whether the inputs are temporally continuous or discrete, and remains robust to inputs with varying distractor ratios. + +Phototourism Dataset. The Phototourism dataset remains challenging for existing novel view synthesis methods due to the diverse sources of images. This variability results in differences in distractors, illumination, and even camera equipment across images. Nevertheless, our MemE-3DGS achieves satisfactory view rendering results on this dataset, as shown in Fig. 5. Notably, severe lighting variations, as well as distractors, can significantly impact the quality of rendered images of 3DGS. However, our MemE is capable of effectively mitigating the influence of lighting variations and distractors. + +# 4.4 Ablation Study + +We conduct ablation studies on the BabyYoda scene within the RobustNeRF dataset. This scene comprises subsets designated for training, testing, and validation. Specifically, the training subset contains distractors, while the testing and + +![](images/f020e8c7a457d2652675ed60f0ad88e5eb0611ca046a1da214bb4495e601874a.jpg) +(a) Adaptability analysis of MemE-NeRF. + +![](images/0ab7dc82ac893e40afde025aa2dadc98da1f9728abb0aa2164ff435b205fb31d.jpg) +(b) Adaptability analysis of MemE-3DGS. +Fig. 6: Adaptability Analysis of MemE on NeRF and 3DGS in Distractor-Free Scenes. MemE demonstrates comparable performance to Mip-NeRF and 3DGS in scenes devoid of distractors. Its adaptability across various scenarios, regardless of the presence of distractors, enhances the versatility of NeRF/3DGS frameworks. + +validation subsets are devoid of distractors. Additionally, the validation subset encompasses different views from those in the training set. + +Adaptability of MemE. While MemE performs admirably in scenes with distractors, we sought to evaluate its performance in scenarios where the presence of distractors is unknown. We conducted a comparison between Mip-NeRF 360, RobustNeRF*, and 3DGS with our MemE on the distractor-free BabyYoda scene. The PSNR curve during training, as depicted in Fig. 6, demonstrates that our MemE-NeRF achieves performance very similar to Mip-NeRF 360 without experiencing significant performance loss. In contrast, RobustNeRF* suffers from considerable performance degradation in distractor-free scenarios. Similarly, MemE-3DGS exhibits a similar trend compared to 3DGS. Thus, MemE demonstrates adaptability in adjusting the focus of NeRF/3DGS in different types of inputs. That is, in scenes with distractors, MemE can effectively and adaptively eliminate them, while in distractor-free scenes, MemE can focus solely on purity without causing significant performance degradation for NeRF/3DGS. Consequently, MemE serves as a plug-and-play module capable of adaptively handling various scenarios with or without distractors. + +Mask Fusion. In our MemE, we utilize mask fusion to generate purity attention by aggregating binary masks $\{\mathbf{M}_n\}$ across multiple scales. These multi-scale masks consist of fine-grained pixel-wise masks $\{\mathbf{M}_i\}$ and coarse-grained tile-wise masks $\{\mathbf{M}_j\}$ , with pixel-wise masks providing detailed information and tile-wise masks aggregating local information within each residual tile. We explore various strategies for integrating hierarchical masks $\{\mathbf{M}_n\}$ , as detailed in Table 3. Initially, using only the most fine-grained mask $\mathbf{M}_1$ resulted in a significant drop in PSNR metric due to its inability to aggregate local information, resulting in noise. Conversely, employing solely the pixel-wise masks $\{\mathbf{M}_i|i\in \{1,3,5,7\}\}$ improved performance by leveraging the OR operation to aggregate their advantages: $\mathbf{M}_1$ providing detail and $\{\mathbf{M}_i|i\in \{3,5,7\}\}$ filtering noise. Additionally, we tested the performance using the MEAN operation on pixel-wise masks, denoted as $\{\mathbf{M}_i|i\in \{1,3,5,7\}\}^*$ in Table 3. Similarly, solely utilizing tile-wise masks also improved results by leveraging hierarchical residual information. However, our + +Table 3: Quantitative Results of Ablation Study. + +
MethodsPSNR ↑SSIM ↑LPIPS ↓
w/o mask fusionM130.090.800.31
{Mi|i ∈ {1,3,5,7}}30.990.800.29
{Mi|i ∈ {1,3,5,7}}*30.760.810.29
{Mj|j ∈ {4,8,16}}30.930.800.30
w/o adaptive thresholds{μ1(Γn)}24.230.770.34
{μ2(Γn)}28.500.770.32
linear moving30.740.790.31
Ours31.080.810.29
+ +method, combining both pixel-wise and tile-wise masks, achieved the best performance. + +Adaptive Thresholds. In our MemE, masks are derived by applying adaptive thresholds to hierarchical residual maps $\{\epsilon_n\}$ . These thresholds transition exponentially from the distractor means $\{\mu_2(\Gamma_n)\}$ to the purity means $\{\mu_1(\Gamma_n)\}$ . The means $\mu_1$ , $\mu_2$ of the GMM are determined from training images of a specific scene at specific scale. We evaluate various threshold-generating strategies, as detailed in Table 3. Initially, we utilize the purity means $\{\mu_1(\Gamma_n)\}$ as thresholds, aiming to attain purity and exclude distractors. However, this approach leads to a significant performance drop. This drop occurs because the purity means $\{\mu_1(\Gamma_n)\}$ decrease rapidly due to the memorization effect during the early stages. However, some areas of purity may exhibit large residuals early on, remaining unaccounted for in the training procedure. Conversely, employing the distractor mean $\{\mu_2(\Gamma_n)\}$ as thresholds introduces some distractors into the training procedure, resulting in a performance decrease. Additionally, we explore the linear threshold strategy, where the thresholds transition linearly from $\{\mu_2(\Gamma_n)\}$ to $\{\mu_1(\Gamma_n)\}$ . The ablation results demonstrate that our method achieves the best performance. + +# 5 Conclusion + +In this study, we introduce MemE, an unsupervised plug-and-play module designed to facilitate high-quality novel view synthesis in noisy input scenarios. MemE capitalizes on the inherent property known as the memorization effect during the parameter optimization phase of view synthesis models, enabling effective distractor filtering. MemE seamlessly integrates with both NeRF and 3DGS frameworks. Notably, our work marks the first instance of leveraging the memorization effect in the training process of NeRF and 3DGS. Additionally, MemE proves to be effective in environments, irrespective of the presence of distractors, thereby substantially improving the versatility of NeRF and 3DGS across a variety of input scenarios. We evaluate MemE across three datasets to assess its effectiveness against other SoTA methods. Both quantitative and qualitative results demonstrate that our approach achieves SoTA performance. + +# Acknowledgements + +We thank Yangdi Lu, Guangchi Fang, and Runqing Jiang for their insightful comments and valuable discussions. This work was partially supported by the National Natural Science Foundation of China (62301601), Guangdong Basic and Applied Basic Research Foundation (2022B1515020103, 2023B1515120087), the Shenzhen Science and Technology Program (No. RCYX20200714114641140). + +# References + +1. Arazo, E., Ortego, D., Albert, P., O'Connor, N., McGuinness, K.: Unsupervised label noise modeling and loss correction. In: ICML. pp. 312-321. PMLR (2019) +2. Arpit, D., Jastrzegski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M.S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., et al.: A closer look at memorization in deep networks. In: ICML. pp. 233-242 (2017) +3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV. pp. 5855-5864 (2021) +4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: CVPR. pp. 5470-5479 (2022) +5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-NeRF: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023) +6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensorRF: Tensorial radiance fields. In: ECCV. pp. 333-350. Springer (2022) +7. Chen, M., Wang, L., Lei, Y., Dong, Z., Guo, Y.: Learning spherical radiance field for efficient $360^{\circ}$ unbounded novel view synthesis. IEEE Transactions on Image Processing (2024) +8. Cordeiro, F.R., Sachdeva, R., Belagiannis, V., Reid, I., Carneiro, G.: LongReMix: Robust learning with high confidence samples in a noisy label environment. PR 133, 109013 (2023) +9. Dai, P., Zhang, Y., Yu, X., Lyu, X., Qi, X.: Hybrid neural rendering for large-scale scenes with motion blur. In: CVPR. pp. 154-164 (2023) +10. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological) 39(1), 1-22 (1977) +1. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR. pp. 5501-5510 (2022) +2. Goli, L., Reading, C., Sellán, S., Jacobson, A., Tagliasacchi, A.: Bayes' rays: Uncertainty quantification for neural radiance fields. arXiv preprint arXiv:2309.03185 (2023) +3. Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., et al.: Kubric: A scalable dataset generator. In: CVPR. pp. 3749-3761 (2022) +4. Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., Tsang, I., Sugiyama, M.: Co-teaching: Robust training of deep neural networks with extremely noisy labels. NeurIPS 31 (2018) + +15. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, X., Wang, Q.: HDR-NeRF: High dynamic range neural radiance fields. In: CVPR. pp. 18398-18408 (2022) +16. Jiang, R., Yan, Y., Xue, J.H., Chen, S., Wang, N., Wang, H.: Knowledge distillation meets label noise learning: Ambiguity-guided mutual label refinery. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2023) +17. Jiang, R., Yan, Y., Xue, J.H., Wang, B., Wang, H.: When sparse neural network meets label noise learning: A multistage learning framework. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2022) +18. Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. IJCV 129(2), 517-547 (2021) +19. Karim, N., Rizve, M.N., Rahnavard, N., Mian, A., Shah, M.: Unicon: Combating label noise through uniform selection and contrastive learning. In: CVPR. pp. 9676-9686 (2022) +20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG 42(4) (2023) +21. Lee, D., Lee, M., Shin, C., Lee, S.: DP-NeRF: Deblurred neural radiance field with physical scene priors. In: CVPR. pp. 12386-12396 (2023) +22. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: ExBluRF: Efficient radiance fields for extreme motion blurred images. In: ICCV. pp. 17639-17648 (2023) +23. Levy, D., Peleg, A., Pearl, N., Rosenbaum, D., Akkaynak, D., Korman, S., Treibitz, T.: SeaThru-NeRF: Neural radiance fields in scattering media. In: CVPR. pp. 56–65 (2023) +24. Li, J., Socher, R., Hoi, S.C.: DivideMix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394 (2020) +25. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: CVPR. pp. 6498-6508 (2021) +26. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: DynIBaR: Neural dynamic image-based rendering. In: CVPR. pp. 4273-4284 (2023) +27. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: CVPR. pp. 13-23 (2023) +28. Lu, Y., Zhang, Y., Han, B., Cheung, Y.m., Wang, H.: Label-noise learning with intrinsically long-tailed data. In: ICCV. pp. 1369-1378 (2023) +29. Lu, Y., Bo, Y., He, W.: An ensemble model for combating label noise. In: Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining. pp. 608-617 (2022) +30. Lu, Y., Bo, Y., He, W.: Noise attention learning: Enhancing noise robustness by gradient scaling. NeurIPS 35, 23164-23177 (2022) +31. Lu, Y., He, W.: SELC: Self-ensemble label correction improves learning with noisy labels. arXiv preprint arXiv:2205.01156 (2022) +32. Lu, Y., He, W.: Learning with noisy ground truth: From 2d classification to 3d reconstruction. arXiv preprint arXiv:2406.15982 (2024) +33. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-NeRF: Neural radiance fields from blurry images. In: CVPR. pp. 12861–12870 (2022) +34. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the wild: Neural radiance fields for unconstrained photo collections. In: CVPR. pp. 7210-7219 (2021) +35. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRFs in the dark: High dynamic range view synthesis from noisy raw images. In: CVPR. pp. 16190–16199 (2022) + +36. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +37. Mirzaei, A., Aumentado-Armstrong, T., Derpanis, K.G., Kelly, J., Brubaker, M.A., Gilitschenski, I., Levinshtein, A.: SPIn-NeRF: Multiview segmentation and perceptual inpainting with neural radiance fields. In: CVPR. pp. 20669–20679 (2023) +38. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG 41(4), 1-15 (2022) +39. Pearl, N., Treibitz, T., Korman, S.: NAN: Noise-aware nerfs for burst-denoising. In: CVPR. pp. 12672-12681 (2022) +40. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: CVPR. pp. 12932-12942 (2022) +41. Sabour, S., Vora, S., Duckworth, D., Krasin, I., Fleet, D.J., Tagliasacchi, A.: RobustNeRF: Ignoring distractors with robust losses. In: CVPR. pp. 20626-20636 (2023) +42. Song, H., Kim, M., Park, D., Shin, Y., Lee, J.G.: Learning from noisy labels with deep neural networks: A survey. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2022) +43. Tanay, T., Leonardis, A., Maggioni, M.: Efficient view synthesis and 3d-based multi-frame denoising with multiplane feature representations. In: CVPR. pp. 20898-20907 (2023) +44. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-NeRF: Scalable large scene neural view synthesis. In: CVPR. pp. 8248-8258 (2022) +45. Teigen, A.L., Yip, M., Hamran, V.P., Skui, V., Stahl, A., Mester, R.: Removing adverse volumetric effects from trained neural radiance fields. arXiv preprint arXiv:2311.10523 (2023) +46. Tschernezki, V., Larlus, D., Vedaldi, A.: NeuralDiff: Segmenting 3D objects that move in egocentric videos. In: Proceedings of the IEEE International Conference on 3D Vision (3DV). pp. 910-919 (2021) +47. Tu, Y., Zhang, B., Li, Y., Liu, L., Li, J., Wang, Y., Wang, C., Zhao, C.R.: Learning from noisy labels with decoupled meta label purifier. In: CVPR. pp. 19934-19943 (2023) +48. Turki, H., Zhang, J.Y., Ferroni, F., Ramanan, D.: SUDS: Scalable urban dynamic scenes. In: CVPR. pp. 12375-12385 (2023) +49. Wang, D., Zhang, T., Abboud, A., Süsstrunk, S.: InpaintNeRF360: Text-guided 3d inpainting on unbounded neural radiance fields. arXiv preprint arXiv:2305.15094 (2023) +50. Wang, H., Xu, X., Xu, K., Lau, R.W.: Lighting up NeRF via unsupervised decomposition and enhancement. In: ICCV. pp. 12632-12641 (2023) +51. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle adjusted deblur neural radiance fields. In: CVPR. pp. 4170-4179 (2023) +52. Warburg, F., Weber, E., Tancik, M., Holynski, A., Kanazawa, A.: NeRF-busters: Removing ghostly artifacts from casually captured NeRFs. arXiv preprint arXiv:2304.10532 (2023) +53. Weder, S., Garcia-Hernando, G., Monszpart, A., Pollefeys, M., Brostow, G.J., Firman, M., Vicente, S.: Removing objects from neural radiance fields. In: CVPR. pp. 16528-16538 (2023) +54. Wei, F., Funkhouser, T., Rusinkiewicz, S.: Clutter detection and removal in 3d scenes with view-consistent inpainting. In: ICCV. pp. 18131-18141 (2023) + +55. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\~ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS 35, 32653-32666 (2022) +56. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: ICCV. pp. 5752-5761 (2021) +57. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: NeRF++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020) +58. Zhang, M., Zhao, X., Yao, J., Yuan, C., Huang, W.: When noisy labels meet long tail dilemmas: A representation calibration method. In: ICCV. pp. 15890-15900 (2023) \ No newline at end of file diff --git a/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/images.zip b/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ccf10d4ab0c96128aac9aff5a2d13e8d3e2b2366 --- /dev/null +++ b/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0f3a50cfb3404aedb293849d8a3698a41f52ff10f732a555c2778016e5faade +size 705468 diff --git a/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/layout.json b/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e481e2708608bdc881fd2029b1031c2a574fe3ab --- /dev/null +++ b/2024/Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization/layout.json @@ -0,0 +1,9998 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "Distractor-Free Novel View Synthesis via Exploiting Memorization Effect in Optimization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "spans": [ + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "text", + "content": "Yukun Wang" + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "text", + "content": ", Kunhong Li" + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "text", + "content": ", Minglin Chen" + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "text", + "content": ", Longguang Wang" + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "text", + "content": ", Shunbo Zhou" + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "text", + "content": ", Kaiwen Xue" + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "text", + "content": ", and Yulan Guo" + }, + { + "bbox": [ + 151, + 167, + 462, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 136, + 201, + 476, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 201, + 476, + 235 + ], + "spans": [ + { + "bbox": [ + 136, + 201, + 476, + 235 + ], + "type": "text", + "content": "1 The Shenzhen Campus of Sun Yat-sen University, Sun Yat-sen University {wangyk59,likh25,chenmlin8}@mail2.sysu.edu.cn guoyulan@sysu.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 235, + 475, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 235, + 475, + 257 + ], + "spans": [ + { + "bbox": [ + 137, + 235, + 475, + 257 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 235, + 475, + 257 + ], + "type": "text", + "content": " Huawei Cloud Computing Technical Innovation Dept., Huawei Cloud Computing Technologies Co., Ltd." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 285, + 455, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 285, + 455, + 462 + ], + "spans": [ + { + "bbox": [ + 159, + 285, + 455, + 462 + ], + "type": "text", + "content": "Abstract. Neural Radiance Fields (NeRF) and 3D Gaussian Splitting (3DGS) have greatly advanced novel view synthesis, which is capable of photo-realistic rendering. However, these methods require the foundational assumption of the static scene (e.g., consistent lighting condition and persistent object positions), which is often violated in real-world scenarios. In this study, we introduce MemE, an unsupervised plug-and-play module, to achieve high-quality novel view synthesis in noisy input scenarios. MemE leverages the inherent property in parameter optimization, known as the memorization effect to achieve distractor filtering and can be easily combined with NeRF or 3DGS. Furthermore, MemE is applicable in environments both with and without distractors, significantly enhancing the adaptability of NeRF and 3DGS across diverse input scenarios. Extensive experiments show that our methods (i.e., MemE-NeRF and MemE-3DGS) achieve state-of-the-art performance on both real and synthetic noisy scenes. We will release our code for further research at https://github.com/Yukun66/MemE." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "spans": [ + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "type": "text", + "content": "Keywords: 3D reconstruction " + }, + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "type": "text", + "content": " Unsupervised noise modeling " + }, + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "type": "text", + "content": " Memorization effect" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 538, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 647 + ], + "type": "text", + "content": "Recently, the emergence of methods like Neural Radiance Fields (NeRF) [4,36] and 3D Gaussian Splatting (3DGS) [20] has greatly propelled the development of novel view synthesis. Additionally, the application of novel view synthesis in fields such as robotics, 3D reconstruction, and AR/VR is gradually being explored. However, these methods are built upon certain critical assumptions, including consistent lighting conditions and persistent object positions. These assumptions are frequently violated in real-world scenarios, which leads to artifacts, significant degradation in rendering quality. This study tackles the specific challenge encountered in novel view synthesis: reconstructing the static and clean" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 652, + 245, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 652, + 245, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 652, + 245, + 666 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 141, + 652, + 245, + 666 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 119, + 478, + 241 + ], + "blocks": [ + { + "bbox": [ + 138, + 119, + 478, + 241 + ], + "lines": [ + { + "bbox": [ + 138, + 119, + 478, + 241 + ], + "spans": [ + { + "bbox": [ + 138, + 119, + 478, + 241 + ], + "type": "image", + "image_path": "2bad65dcaaad2d1b27fe879cba36185daf0054fa09ce77754447584c07811a95.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 256, + 478, + 338 + ], + "blocks": [ + { + "bbox": [ + 273, + 245, + 341, + 255 + ], + "lines": [ + { + "bbox": [ + 273, + 245, + 341, + 255 + ], + "spans": [ + { + "bbox": [ + 273, + 245, + 341, + 255 + ], + "type": "text", + "content": "(a) Mip-NeRF 360" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 256, + 478, + 338 + ], + "lines": [ + { + "bbox": [ + 138, + 256, + 478, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 478, + 338 + ], + "type": "image", + "image_path": "516c99d15bb5c83ee3ea6d17b311edda59c98166d6cf99c7ca4a7c93b8b53a7f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 340, + 351, + 350 + ], + "lines": [ + { + "bbox": [ + 263, + 340, + 351, + 350 + ], + "spans": [ + { + "bbox": [ + 263, + 340, + 351, + 350 + ], + "type": "text", + "content": "(b) MemE-NeRF (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "text", + "content": "Fig. 1: Memorization Effect in Optimization. The purity (clean scene) and distractor pixels correspond to the blue and red bars in the histograms, respectively. In the training of NeRF and 3DGS, we observed the phenomenon of memorization effect. This phenomenon refers to the parameters initially fitting purity and excluding distractor pixels in the early stages of training, gradually overfitting to distractor pixels in the later stages, as evidenced by the rendered image and histogram in (a). Our method leverages the memorization effect to effectively prevent distractors from overfitted, maintaining the position of the red bars unchanged as much as possible, as demonstrated in (b)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 484, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 508 + ], + "type": "text", + "content": "scene from noisy images containing distractors, which disrupt the assumption of the static scene." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "content": "To comprehend the impact of distractors in novel view synthesis, we analyze the optimization process of NeRF and 3DGS. Specifically, we exemplify this analysis using Mip-NeRF 360 [4], as depicted in Fig. 1a. Our observations indicate that during the initial stages of optimization (e.g., the 5000th step), the image exhibits blurring, yet the scene remains relatively clean. However, as optimization progresses, the image sharpens, concomitant with the emergence of distractors. Further analyzing the accumulated loss distribution, Mip-NeRF 360 primarily focuses on learning purity pixels (of the clean scenes) and leaves most distractor pixels out of the learning process at the early stage, as evidenced by the minimal changes in the histogram of distractor pixels (red bars). This observation, namely memorization effect [2, 42], enables a clear distinction between purity and distractor pixels. Subsequently, we can promote the fitting of purity pixels and suppress the learning of distractor pixels throughout the Mip-NeRF" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "360 training process to easily achieve sharp and clean view rendering in a noisy input scenario." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 331 + ], + "type": "text", + "content": "Based on the above insight, we propose MemE, an unsupervised plug-and-play method that can be incorporated with NeRF and 3DGS, to achieve high-quality novel view synthesis in noisy input scenarios. MemE is based on the discovery of Memorization Effect in NeRF and 3DGS optimization, and the core of MemE is to distinguish distractor pixels without prior knowledge and exclude them throughout the optimization process. Specifically, our MemE begins with our Hierarchical Residual Measurement module to quantify the residuals between the rendered results and the noisy input images, outputting residual vectors for other modules. Subsequently, within our Purity-Distractor Mixture Modeling module, we employ a multi-level Gaussian Mixture Model (GMM) to model the residual distribution of both purity and distractor pixels according to the residual vectors. Finally, we calculate a purity attention weight for NeRF/3DGS optimization. With that attention, we can ensure that the distractor pixels remains unlearned throughout the parameters optimization process, preventing the parameters from overfitting to distractor pixels and thereby achieving the distractor-free novel view synthesis, as shown in Fig. 1b." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 439 + ], + "type": "text", + "content": "In comparison to prior works, our approach offers distinct advantages: 1) It is solely grounded in the inherent property of NeRF and 3DGS training, i.e., memorization effect, and conducted in an unsupervised manner. 2) Our MemE is capable of processing training data, whether it contains distractors or not, without the need for prior identification. 3) Our method operates independently of pre-trained models or parameters, ensuring simplicity and effectiveness, thus enabling seamless integration as a plug-and-play solution. We incorporate MemE with NeRF (referred to as MemE-NeRF) and 3DGS (referred to as MemE-3DGS) to demonstrate the universality of MemE." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 439, + 362, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 439, + 362, + 451 + ], + "spans": [ + { + "bbox": [ + 146, + 439, + 362, + 451 + ], + "type": "text", + "content": "Our contributions can be summarized as follows," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 459, + 479, + 576 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 459, + 479, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 479, + 505 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 479, + 505 + ], + "type": "text", + "content": "- We are the first work to exploit the memorization effect in NeRF and 3DGS optimization process. Further, we propose MemE-NeRF and MemE-3DGS to achieve high-quality novel view synthesis in the noisy input scenario in an unsupervised manner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 506, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 506, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 506, + 479, + 552 + ], + "type": "text", + "content": "- We propose MemE, a simple yet plug-and-play method to filter distractor pixels. MemE can adapt to scenes both with and without distractors simultaneously, which can enhance the versatility of NeRF/3DGS in various types of input scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 554, + 479, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 554, + 479, + 576 + ], + "spans": [ + { + "bbox": [ + 137, + 554, + 479, + 576 + ], + "type": "text", + "content": "- We conduct extensive experiments and the results show that our method achieves state-of-the-art performance on both real and synthetic noisy scenes." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 596, + 237, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 596, + 237, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 596, + 237, + 609 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 621, + 270, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 621, + 270, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 270, + 633 + ], + "type": "text", + "content": "2.1 Novel View Synthesis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF) [36] has made a breakthrough in novel view synthesis, which is capable of synthesizing photo-realistic images at arbitrary" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "views. NeRF represents a 3D scene as a continuous radiance field, and synthesizes images with differentiable volumetric rendering. Mip-NeRF [3] and Zip-NeRF [5] improve the rendering quality by reducing the aliasing artifacts. " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathrm{NeRF}++" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " [57] and Mip-NeRF 360 [4] are proposed to extend NeRF to " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " unbounded scene [7]. Several works accelerate NeRF by parameterizing the radiance fields with high-efficient data structures, such as multiresolution hash table [38], Octree [11,56], and orthogonal planes [6]. Recently, 3D Gaussian Splatting [20] has emerged as another promising paradigm for novel view synthesis, which synthesizes the image by rendering a set of learnable Gaussian points." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 272 + ], + "type": "text", + "content": "Despite great development in novel view synthesis, the aforementioned methods are limited to static scenes with constant light conditions. They are struggling with multi-view images that contain distractors. Our method aims to extend existing methods to distractor-free novel view synthesis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 288, + 334, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 288, + 334, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 288, + 334, + 300 + ], + "type": "text", + "content": "2.2 View Synthesis from Noisy Inputs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 307, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 450 + ], + "type": "text", + "content": "Recently, efforts are made to remove noises during NeRF reconstruction, such as training data noise [15, 35, 39, 43, 45, 50], deblurring [9, 21, 22, 33, 51], artifact removal [12, 23, 52], dynamic distractor elimination [34, 40, 41, 44, 55], and scene inpainting [37, 49, 53, 54]. To address transient occluders, NeRF-W [34] employs multilayer perceptrons (MLPs) modeling for static and transient components separately. Methods for removing movable objects in large-scale street scenes [40, 44] have utilized semantic segmentation models to explicitly filter out dynamic outliers, such as cars and pedestrians. Recent advancements have extended NeRF to video data [26, 27, 48], introducing an auxiliary temporal MLP for separately modeling dynamic and static scenes. RobustNeRF [41] processes discrete images from casual viewpoints, ranking training pixel errors to identify and eliminate distractors using hand-crafted ratios, which may yield suboptimal results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 451, + 482, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 535 + ], + "type": "text", + "content": "Most of the aforementioned methods do not consider the case of distractor noise, while the most related counterpart of our method is RobustNeRF [41]. Compared to RobustNeRF, our method is capable of scenes with or without distractors, with no need for analysis and artificial prior on the distractor content of the scene in advance. Our method can achieve similar results to Mip-NeRF 360 and 3DGS in distractor-free scenes and does not require special adjustment of hyperparameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 551, + 270, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 551, + 270, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 551, + 270, + 563 + ], + "type": "text", + "content": "2.3 Noisy Label Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Numerous methods [1, 14, 16, 17, 19, 24, 28-32, 47, 58] have investigated the challenge of label noise in the field of image classification due to the difficulty in data labeling [42]. Prior arts on the memorization effect [1, 2, 42] indicates that neural networks tend first to learn simple and clean samples, and then gradually memorize noisy labeled samples. Based on this, several efforts are devoted to combat label noise by selecting clean samples from noisy dataset during model training, thus avoiding model from overfitting to those noisy samples. Co-teaching [14] retains two networks, and each network selects a portion of samples with small-loss" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 119, + 476, + 264 + ], + "blocks": [ + { + "bbox": [ + 136, + 119, + 476, + 264 + ], + "lines": [ + { + "bbox": [ + 136, + 119, + 476, + 264 + ], + "spans": [ + { + "bbox": [ + 136, + 119, + 476, + 264 + ], + "type": "image", + "image_path": "75e2e49fee8494af2307f02d549e5e1e0790b8680c021a6fbd7b7dbe0243591e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 275, + 482, + 364 + ], + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 364 + ], + "type": "text", + "content": "Fig. 2: The Framework of MemE. Our MemE integrates three primary components: Hierarchical Residual Measurement (Sec. 3.1), Purity-Distractor Mixture Modeling (Sec. 3.2), and Purity Attention Generation (Sec. 3.3). Hierarchical Residual Measurement module generates residual vectors " + }, + { + "bbox": [ + 130, + 275, + 482, + 364 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\Gamma}_n\\}" + }, + { + "bbox": [ + 130, + 275, + 482, + 364 + ], + "type": "text", + "content": " and residual maps " + }, + { + "bbox": [ + 130, + 275, + 482, + 364 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\epsilon}_n\\}" + }, + { + "bbox": [ + 130, + 275, + 482, + 364 + ], + "type": "text", + "content": ", which are then utilized by the subsequent modules. Purity-Distractor Mixture Modeling module fits the multi-level Gaussian Mixture Model (GMM) according to " + }, + { + "bbox": [ + 130, + 275, + 482, + 364 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\Gamma}_n\\}" + }, + { + "bbox": [ + 130, + 275, + 482, + 364 + ], + "type": "text", + "content": ". Finally, the Purity Attention Generation module creates attention weight " + }, + { + "bbox": [ + 130, + 275, + 482, + 364 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 275, + 482, + 364 + ], + "type": "text", + "content": " to exclude distractors selectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 472 + ], + "type": "text", + "content": "and feeds them to another network. Note that above methods discard intensive noisy samples, which inevitably incurs information loss, limiting their performance especially when the noise ratio is high. To remedy this, DivideMix [24] employs Gaussian mixture models to split training dataset into labeled (clean) and unlabeled (noisy) sets, and then adopts MixMatch to make full use of those noisy samples. Besides, LongReMix [8] also uses MixMatch strategy and augments high-confidence samples to achieve superior robustness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 490, + 202, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 202, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 202, + 502 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": "In this section, we present a novel plug-and-play method (namely MemE), which is aimed at enhancing existing novel view synthesis methods when multi-view images contain discriminators. Existing novel view synthesis methods represent a 3D scene as a parameterized model " + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\theta}" + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": ", while facilitating a differentiable rendering technique " + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": " to render an image " + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": " as following:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 275, + 586, + 481, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 586, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 275, + 586, + 481, + 600 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {I}} = \\mathcal {R} (\\mathbf {S} _ {\\theta}, \\pi), \\tag {1}", + "image_path": "e5478969079421ac4643d6115f9fa80ecdf139df5d6655f64e7b6983cb6e212e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " is the image pose. Concretely, Neural Radiance Fields [36] leverages the Multilayer Perceptron (MLP) and volumetric rendering, whereas 3D Gaussian Splatting [20] is built upon a collection of learnable Gaussian points and splatting rasterization. Because multi-view images may contain distractors, we employ the proposed MemE to generate an attention weight " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " to indicate the distractor" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "text", + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "pixels. Therefore, existing methods with our MemE can achieve distractor-free novel view synthesis using the following loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 250, + 149, + 481, + 173 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 149, + 481, + 173 + ], + "spans": [ + { + "bbox": [ + 250, + 149, + 481, + 173 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) = \\sum_ {r} \\mathbf {M} _ {r} | | \\hat {\\mathbf {I}} _ {r} - \\mathbf {I} _ {r} | | _ {2} ^ {2}, \\tag {2}", + "image_path": "85b4cfd98b989412e87bb3e718919ecd06cc522749764469dbc132946b762bd1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 479, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 479, + 207 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 479, + 207 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 183, + 479, + 207 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_r" + }, + { + "bbox": [ + 130, + 183, + 479, + 207 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 183, + 479, + 207 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_r" + }, + { + "bbox": [ + 130, + 183, + 479, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 183, + 479, + 207 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_r" + }, + { + "bbox": [ + 130, + 183, + 479, + 207 + ], + "type": "text", + "content": " are the rendered color, ground-truth color, and attention weight at pixel " + }, + { + "bbox": [ + 130, + 183, + 479, + 207 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 183, + 479, + 207 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 208, + 482, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 208, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 482, + 340 + ], + "type": "text", + "content": "The framework of our MemE is depicted in Fig. 2. Our method consists of three components: hierarchical residual measurement (Sec. 3.1), purity-distractor mixture modeling (Sec. 3.2), and purity attention generation (Sec. 3.3). The hierarchical residual measurement module quantifies the residuals between noisy input " + }, + { + "bbox": [ + 130, + 208, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 208, + 482, + 340 + ], + "type": "text", + "content": " and rendered result " + }, + { + "bbox": [ + 130, + 208, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 208, + 482, + 340 + ], + "type": "text", + "content": ". The purity-distractor mixture modeling module fits the distribution of residuals. The purity attention generation module generates a noisy-excluded attention mask " + }, + { + "bbox": [ + 130, + 208, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 208, + 482, + 340 + ], + "type": "text", + "content": " for loss adjusting. By composing the three simple modules, MemE exploits the memorization effect existing in NeRF and 3DGS, and models the purity-distractor distribution of the scene in an unsupervised manner. The resulting mask " + }, + { + "bbox": [ + 130, + 208, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 208, + 482, + 340 + ], + "type": "text", + "content": " can effectively help NeRF and 3DGS focus only on purity and ignore distractors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 357, + 340, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 340, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 340, + 369 + ], + "type": "text", + "content": "3.1 Hierarchical Residual Measurement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 376, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 449 + ], + "type": "text", + "content": "MemE identifies which parts of the scene are prioritized for learning by calculating residuals of rendering results and noisy input. According to the observed memorization effect, the areas learned first more likely belong to purity, whereas distractors are more likely to be left out of training in the early stage. Further, we employ residual measurement on multiple scales to provide more comprehensive local information hierarchically." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": "Given a noisy input patch " + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": " and a rendering result patch " + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": ", the Hierarchical Residual Measurement module eventually outputs hierarchical residual vectors and residual maps to the other modules. Specifically, the residual map between " + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": " is first calculated by the L2 norms, noted as " + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": ". Then a residual pyramid " + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\{\\epsilon_n\\}" + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": " is generated by pixel-wise and tile-wise operations according to " + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\{\\epsilon_n\\}" + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": " are simply flattened to vectors " + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\{\\Gamma_n\\}" + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 280, + 531, + 480, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 531, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 280, + 531, + 480, + 544 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\epsilon} _ {n} = \\tau_ {n} \\odot \\boldsymbol {\\epsilon}, \\tag {3}", + "image_path": "a4fb84b841b956704fed04c04a9090bd3be8737c47d020087509cef2d1e45bfd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 277, + 564, + 480, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 564, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 277, + 564, + 480, + 577 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Gamma} _ {n} = \\nu \\odot \\boldsymbol {\\epsilon} _ {n}, \\tag {4}", + "image_path": "aefc0c9ec201c2df3bbf7ece47fba8af4ec27e3ba4cd70df41e7512f69647b6f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "content": " is the vectorization operation to flatten the patch, " + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\tau_{n}" + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "content": " represents the pixel-wise and tile-wise operations, and " + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 582, + 479, + 605 + ], + "type": "text", + "content": " represents the operation deployment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "The pixel-wise operations are the average convolutions with kernels of " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{i\\times i}" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "i\\in \\mathcal{I}" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": ", stride 1 and padding, where " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": " refers to the set of pixel-wise scales. As a result, the multi-level residuals " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon_{i}" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": " are with spatial size " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": ". The tile-wise operations are the average pooling with kernels of " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{j\\times j}" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "j\\in \\mathcal{J}" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": ", stride " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": " and no padding, where " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": " refers to the set of tile-wise scales. Thus, the multi-level" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": "residuals " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{j}" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": " are with spatial size " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H / j\\times W / j}" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": ". In conclusion, the outputs of Hierarchical Residual Measurement are multi-level residual maps " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\epsilon}_n\\}" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": " and residual vectors " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\Gamma}_n\\}" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "n\\in \\mathcal{I}\\cup \\mathcal{J}" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 168, + 344, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 168, + 344, + 182 + ], + "spans": [ + { + "bbox": [ + 131, + 168, + 344, + 182 + ], + "type": "text", + "content": "3.2 Purity-Distractor Mixture Modeling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 187, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 187, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 187, + 479, + 247 + ], + "type": "text", + "content": "Based on the residual vectors " + }, + { + "bbox": [ + 130, + 187, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\Gamma}_n\\}" + }, + { + "bbox": [ + 130, + 187, + 479, + 247 + ], + "type": "text", + "content": " from the Hierarchical Residual Measurement module, it is necessary to use a model to fit the distribution of purity and distractors in the NeRF and 3DGS parameter optimization process, so that MemE can use the model to identify purity and distractors in the next module, and calculate the mask to adjust training loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 247, + 480, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 480, + 307 + ], + "type": "text", + "content": "In the Purity-Distractor Mixture Modeling module, a one-dimensional and two-component GMM is utilized to fit the residual distributions of purity and distractor pixels because the residual distribution tends to be bi-modal in the early training stage. The probability density function (pdf) of a mixture model of " + }, + { + "bbox": [ + 130, + 247, + 480, + 307 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 247, + 480, + 307 + ], + "type": "text", + "content": " components, applied to a residual vector " + }, + { + "bbox": [ + 130, + 247, + 480, + 307 + ], + "type": "inline_equation", + "content": "\\pmb{\\Gamma}_{n}" + }, + { + "bbox": [ + 130, + 247, + 480, + 307 + ], + "type": "text", + "content": ", is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 236, + 316, + 480, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 316, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 236, + 316, + 480, + 348 + ], + "type": "interline_equation", + "content": "p \\left(\\boldsymbol {\\Gamma} _ {n}\\right) = \\sum_ {k = 1} ^ {K} \\lambda_ {k} \\cdot \\mathcal {N} \\left(\\boldsymbol {\\Gamma} _ {n} \\mid \\mu_ {k}, \\sigma_ {k} ^ {2}\\right), \\tag {5}", + "image_path": "195727643383e552daf99d75be5963f3a80643669c01c57e2b2d7623754c9ccf.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 354, + 479, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 479, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 354, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\lambda_{k}" + }, + { + "bbox": [ + 130, + 354, + 479, + 415 + ], + "type": "text", + "content": " represents the mixing coefficients for the convex combination of each pdf " + }, + { + "bbox": [ + 130, + 354, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\pmb{\\Gamma}_n\\mid \\mu_k,\\sigma_k^2)" + }, + { + "bbox": [ + 130, + 354, + 479, + 415 + ], + "type": "text", + "content": ". The fitting of GMM is conducted through the Expectation-Maximization (EM) algorithm [10]. As the residual vectors " + }, + { + "bbox": [ + 130, + 354, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\Gamma}_n\\}" + }, + { + "bbox": [ + 130, + 354, + 479, + 415 + ], + "type": "text", + "content": " are with multi-level, the GMM is implemented with each level respectively, and the output is a set of pdfs expressed as " + }, + { + "bbox": [ + 130, + 354, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\{p(\\pmb{\\Gamma}_n)\\}" + }, + { + "bbox": [ + 130, + 354, + 479, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "text", + "content": "Additionally, to achieve robust training, we update " + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "inline_equation", + "content": "\\{p(\\boldsymbol{\\Gamma}_n)\\}" + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "text", + "content": " every " + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "text", + "content": " steps. That is, we accumulate " + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "text", + "content": " residual vectors " + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "inline_equation", + "content": "\\{\\boldsymbol{\\Gamma}_n^t | t \\in \\{0,1,\\dots,T-1\\}\\}" + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "text", + "content": " for each level, and compute " + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "inline_equation", + "content": "\\{p(\\cup \\boldsymbol{\\Gamma}_n^t)\\}" + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "text", + "content": " all at once. For convenience, we continue to use the symbol " + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "inline_equation", + "content": "\\{p(\\boldsymbol{\\Gamma}_n)\\}" + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "inline_equation", + "content": "\\{p(\\cup \\boldsymbol{\\Gamma}_n^t)\\}" + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "text", + "content": " to denote the output pdfs in subsequent sections. And at the very beginning " + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 415, + 480, + 486 + ], + "type": "text", + "content": " steps, we do not employ our MemE for training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 502, + 305, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 305, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 305, + 514 + ], + "type": "text", + "content": "3.3 Purity Attention Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 479, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 569 + ], + "type": "text", + "content": "Utilized the residual maps " + }, + { + "bbox": [ + 130, + 521, + 479, + 569 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\epsilon}_n\\}" + }, + { + "bbox": [ + 130, + 521, + 479, + 569 + ], + "type": "text", + "content": " and pdf sets " + }, + { + "bbox": [ + 130, + 521, + 479, + 569 + ], + "type": "inline_equation", + "content": "\\{p(\\pmb{\\Gamma}_n)\\}" + }, + { + "bbox": [ + 130, + 521, + 479, + 569 + ], + "type": "text", + "content": " obtained from two former modules, Purity Attention Generation outputs an attention map " + }, + { + "bbox": [ + 130, + 521, + 479, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 521, + 479, + 569 + ], + "type": "text", + "content": " adjusting loss function to constraint view synthesis model only focuses on the purity and repel distractors out of optimization procedure." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "text", + "content": "Specifically, for the residual maps processed through tile-wise operations, we upsample their spatial size to the original size, i.e., " + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "text", + "content": ", using nearest interpolation. Then, for spatially aligned residual maps " + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "inline_equation", + "content": "\\{\\pmb {\\epsilon}_n\\}" + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "text", + "content": ", we have corresponding pdfs " + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "inline_equation", + "content": "\\{p(\\pmb {\\Gamma}_n)\\}" + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "text", + "content": ", and we compute adaptive thresholds " + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "inline_equation", + "content": "\\{\\rho_n\\}" + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "text", + "content": " according to " + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "inline_equation", + "content": "\\{p(\\pmb {\\Gamma}_n)\\}" + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "text", + "content": ". Next, the binary masks " + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_n|\\mathbf{M}_n\\triangleq \\pmb {\\epsilon}_n\\leq \\rho_n\\}" + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "text", + "content": " are generated according to " + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "inline_equation", + "content": "\\{\\rho_{n}\\}" + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_n\\}" + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "text", + "content": " are fused into a purity attention map " + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 570, + 479, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "Adaptive Threshold. We apply a threshold to the residual map to derive the mask at each level. Since our destination is to repel distractors from the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 448, + 103 + ], + "type": "text", + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "optimization process, we hope this threshold is: 1) smaller than the most of residuals of distractors in the total optimization process; 2) slightly larger than most of the residuals of purity at the later stage of optimization to preserve most purity and make sure to exclude the distractors as many as possible; and 3) starting from a moderate value and gradually decrease as the optimization progresses because purity pixels may also have large residual values in the early stages of optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "text", + "content": "Therefore, in each scale, we calculate a threshold from the position lying between the two means of " + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "inline_equation", + "content": "p(\\pmb{\\Gamma}_n)" + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "inline_equation", + "content": "\\mu_1(\\pmb{\\Gamma}_n)" + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "inline_equation", + "content": "\\mu_2(\\pmb{\\Gamma}_n)" + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "text", + "content": ". There are three reasons as follows: 1) the means indicate the most concentrated residual magnitude of the purity and distractors; 2) the " + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "inline_equation", + "content": "\\mu_1(\\pmb{\\Gamma}_n)" + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "text", + "content": " is the ideal value of our threshold at the end of optimization; and 3) the " + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "inline_equation", + "content": "\\mu_2(\\pmb{\\Gamma}_n)" + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "text", + "content": " provide a moderate starting point lying between the ideal value " + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "inline_equation", + "content": "\\mu_1(\\pmb{\\Gamma}_n)" + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "text", + "content": " and the unconstrained value, i.e., 1. Further, we design adaptive thresholds that exponentially move from " + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "inline_equation", + "content": "\\mu_2(\\pmb{\\Gamma}_n)" + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "inline_equation", + "content": "\\mu_1(\\pmb{\\Gamma}_n)" + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 188, + 300, + 481, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 300, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 188, + 300, + 481, + 314 + ], + "type": "interline_equation", + "content": "\\rho_ {n} = \\alpha \\cdot \\mu_ {1} (\\boldsymbol {\\Gamma} _ {n}) + (1 - \\alpha) \\cdot \\mu_ {2} (\\boldsymbol {\\Gamma} _ {n}), \\quad \\alpha = 1 - e ^ {\\left(- \\omega \\cdot \\frac {s}{N}\\right)} \\tag {6}", + "image_path": "b3568535ce5f86005d9ee65287aedf8ee31ce3c40a0ebc5c390b4831f349ac0a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "content": " controls the moving velocity, " + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "content": " are current training steps and total steps, and " + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\rho_{n}" + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "content": " is the threshold for generating the binary mask " + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_n" + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "content": ", where 0 refers to the prediction as distractors and 1 refers to the prediction as purity. Collecting the masks of different scales, we have a set of binary masks " + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_n\\}" + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": "Mask Fusion. After the binary masks " + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_n\\}" + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": " are generated, we further fuse the multi-level masks into a single purity attention map " + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": ". These masks contain information at multiple scales and are fused in this module. We split the " + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_n\\}" + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": " into pixel-wise operated subset " + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_i|i\\in \\mathcal{I}\\}" + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": " and tile-wise operated subset " + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_j|j\\in \\mathcal{J}\\}" + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": ". Generally, the masks in " + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_i\\}" + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": " are fine-grained and the masks in " + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_j\\}" + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": " are coarse grained. Within the " + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_i\\}" + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": ", the mask with smaller " + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": " is detailed but noisy; while the mask with larger " + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": " is noise-filtered but blurry. As such, we first fuse the " + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_i\\}" + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": " with OR operation, noted as " + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": ", to combine their advantages including detailed and noise-filtered, generating the fused pixel-wise mask " + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^p" + }, + { + "bbox": [ + 130, + 377, + 483, + 485 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 262, + 499, + 481, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 499, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 262, + 499, + 481, + 512 + ], + "type": "interline_equation", + "content": "\\mathbf {M} ^ {p} = \\oplus \\left\\{\\mathbf {M} _ {i} \\mid i \\in \\mathcal {I} \\right\\}. \\tag {7}", + "image_path": "0794b8f6aa505fb09186a2a114a99eb528bc7c9a2352c971baab4d0997b85fd3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 525, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 574 + ], + "type": "text", + "content": "The OR operation removes the noise of value 0 in binary masks while preserving the noise of value 1. As such we combine the hierarchical tile-wise masks " + }, + { + "bbox": [ + 130, + 525, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_j\\}" + }, + { + "bbox": [ + 130, + 525, + 482, + 574 + ], + "type": "text", + "content": " providing more comprehensive local information to generate the purity attention map " + }, + { + "bbox": [ + 130, + 525, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 525, + 482, + 574 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 245, + 580, + 481, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 580, + 481, + 618 + ], + "spans": [ + { + "bbox": [ + 245, + 580, + 481, + 618 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\frac {1}{\\mathcal {D}} \\left(\\mathbf {M} ^ {p} + \\sum_ {j} \\left(\\mathbf {M} _ {j}\\right)\\right), \\tag {8}", + "image_path": "ace3e7bba764cdcc36fc56326ea2134ec87b32f9a2257972d5aa0405a78e930b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " means the total number of masks, including " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_j\\}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^p" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{1,3,5,7\\}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{J} = \\{4,8,16\\}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " in our method. Therefore, the elements of purity attention map " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " are discrete values belongs to the set " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{0,0.25,0.5,0.75,1\\}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 229, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 229, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 229, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 204, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 204, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 204, + 150 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 159, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 479, + 217 + ], + "type": "text", + "content": "RobustNeRF Dataset. The RobustNeRF dataset [41] comprises four different indoor scenes. Within each scene, there exists a sub-training set containing distractors and a sub-testing set without distractors. Additionally, in the BabyYoda scene, there is an additional sub-validation set devoid of distractors, featuring views distinct from those in the training set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 219, + 479, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 479, + 265 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 479, + 265 + ], + "type": "text", + "content": "Kubric Dataset. The Kubric dataset introduced in [55], is a synthetic dataset comprising five scenes, generated using the Kubric dataset generator [13]. Each scene comprises an empty room with a texture-less background and several static objects. In each frame, there are distractors present with continuous motion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 267, + 479, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 479, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 479, + 324 + ], + "type": "text", + "content": "Phototourism Dataset. The Phototourism dataset [18] is a typical outdoor scene dataset, which comprises images collected from the internet, thus containing numerous distractors such as tourists and illumination changes. We randomly select six scenes from the Phototourism dataset to assess the performance of our MemE." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 345, + 343, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 345, + 343, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 345, + 343, + 357 + ], + "type": "text", + "content": "4.2 Metrics and Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 365, + 479, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 479, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 479, + 470 + ], + "type": "text", + "content": "We report PSNR, SSIM, and LPIPS to quantitatively evaluate the proposed MemE-NeRF and MemE-3DGS. MemE-NeRF is trained with 64 or 16 training patches per step for every scene. To ensure a fair comparison with RobustNeRF reported in [41], where training was conducted with 64 patches per step, we also train RobustNeRF with 16 patches (denoted as RobustNeRF*). MemE-NeRF is based on Mip-NeRF 360 [4] implemented using JAX, while MemE-3DGS is based on 3DGS [20] implemented using PyTorch. Experiments are conducted on 4 NVIDIA RTX-3090 GPUs for 64 patches and 1 for 16 patches. All models are trained with images containing distractors while tested on distractor-free images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 490, + 214, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 214, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 214, + 501 + ], + "type": "text", + "content": "4.3 Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 479, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 479, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 479, + 578 + ], + "type": "text", + "content": "RobustNeRF Dataset. On the RobustNeRF dataset, our MemE outperforms the counterparts, as shown in Table 1 and Fig. 3. MemE-NeRF surpasses RobustNeRF across most metrics, while MemE-3DGS significantly enhances the performance of 3DGS in scenes containing distractors. Notably, our MemE-NeRF achieves comparable or even higher PSNR than RobustNeRF, despite being trained with fewer patches." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 663 + ], + "type": "text", + "content": "As illustrated in Fig. 3, our MemE-NeRF accurately removes distractors while preserving more details. In Fig. 3 BabyYoda, MemE-NeRF effectively eliminates distractors and retains the detailed patterns of the table. In contrast, although RobustNeRF* removes distractors, it introduces incorrect floaters and misses details of the table. In Fig. 3 Statue, both MemE-NeRF and MemE-3DGS successfully remove distractors. However, RobustNeRF* is susceptible to contamination by distractors, leading to missing details and floaters. Compared" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 136, + 478, + 247 + ], + "blocks": [ + { + "bbox": [ + 175, + 114, + 437, + 126 + ], + "lines": [ + { + "bbox": [ + 175, + 114, + 437, + 126 + ], + "spans": [ + { + "bbox": [ + 175, + 114, + 437, + 126 + ], + "type": "text", + "content": "Table 1: Quantitative comparison on the RobustNeRF dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 136, + 478, + 247 + ], + "lines": [ + { + "bbox": [ + 136, + 136, + 478, + 247 + ], + "spans": [ + { + "bbox": [ + 136, + 136, + 478, + 247 + ], + "type": "table", + "html": "
Methodsp†StatueAndroidCrabBabyYoda
LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑
Mip-NeRF 360 (L2) [4]640.360.6619.090.400.6519.350.270.7725.730.310.7522.97
Mip-NeRF 360 (L1) [4]640.300.7219.550.400.6619.380.220.7926.690.220.8026.15
Mip-NeRF 360 (Ch.) [4]640.300.7319.640.400.6619.530.210.8027.720.220.8025.22
D2NeRF [55]-0.480.4919.090.430.5720.610.420.6821.180.440.6517.32
RobustNeRF [41]640.280.7520.890.310.6521.720.210.8130.750.200.8330.87
MemE-NeRF (Ours)640.230.7321.990.300.6821.740.200.8330.640.250.8331.36
RobustNeRF* [41]160.330.6621.900.310.6721.400.240.8230.580.310.7929.76
MemE-NeRF (Ours)160.310.6821.810.300.6821.540.230.8330.610.290.8131.08
3DGS [20]-0.200.7620.220.250.6821.540.190.8925.790.190.9128.12
MemE-3DGS (Ours)-0.140.8121.740.180.7322.720.190.9025.810.190.9128.40
", + "image_path": "68225af0af38e6ba430639b969488bf753fe35ee67af11515f395582c004967c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 136, + 247, + 271, + 255 + ], + "lines": [ + { + "bbox": [ + 136, + 247, + 271, + 255 + ], + "spans": [ + { + "bbox": [ + 136, + 247, + 271, + 255 + ], + "type": "text", + "content": "the number of patches in each training step." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 269, + 479, + 475 + ], + "blocks": [ + { + "bbox": [ + 138, + 269, + 479, + 475 + ], + "lines": [ + { + "bbox": [ + 138, + 269, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 479, + 475 + ], + "type": "image", + "image_path": "91f3b21e872f9dc506acbc673f69f00d709faa67f8d6530c561a500fe6eedd99.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 491, + 482, + 557 + ], + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 557 + ], + "type": "text", + "content": "Fig.3: Qualitative Comparison on the RobustNeRF Dataset. Compared to the baselines, i.e., Mip-NeRF 360 and 3DGS, our MemE-NeRF and MemE-3DGS not only efficiently eliminate distractors but also retain a higher level of detail. In comparison with RobustNeRF*, MemE models demonstrate superior performance in removing distractors, leading to a reduction in artifacts and an enhancement in detail preservation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "with 3DGS, MemE-3DGS exhibits significant improvements in both qualitative and quantitative comparisons. 3DGS is prone to overfitting distractors, resulting in incorrect view synthesis ( Fig. 3 Statue and Crab) or artifacts in the rendered images ( Fig. 3 Android and Baby Yoda). It is noteworthy that RobustNeRF leads to significant performance degradation in clean scene training, whereas the integration of our MemE is non-intrusive to existing NeRF and 3DGS frameworks, as demonstrated in [41] and Sec. 4.4." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 136, + 479, + 228 + ], + "blocks": [ + { + "bbox": [ + 187, + 114, + 425, + 126 + ], + "lines": [ + { + "bbox": [ + 187, + 114, + 425, + 126 + ], + "spans": [ + { + "bbox": [ + 187, + 114, + 425, + 126 + ], + "type": "text", + "content": "Table 2: Quantitative comparison on the Kubric dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 136, + 479, + 228 + ], + "lines": [ + { + "bbox": [ + 134, + 136, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 134, + 136, + 479, + 228 + ], + "type": "table", + "html": "
MethodsCarCarsBagChairsPillow
LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑LPIPS ↓SSIM ↑PSNR ↑
NeRF-W [34].218.81424.23.243.87324.51.139.79120.65.150.68123.77.088.93528.24
NSFF [25].200.80624.90.620.37610.29.108.89225.62.682.28412.82.782.3434.55
NeuralDiff [46].065.95231.89.098.92125.93.117.91029.02.112.72224.42.565.65220.09
D2NeRF [55].062.97534.27.090.95326.27.076.97934.14.095.70724.63.076.97936.58
RobustNeRF [41].013.98837.73.063.95726.31.006.99541.82.007.99241.23.018.99038.95
RobustNeRF* [41].021.97340.83.020.98643.70.027.98331.62.009.99346.46.026.98644.23
MemE-NeRF (Ours).021.97340.86.019.98744.34.003.99748.79.007.99447.87.016.98945.97
3DGS [20].037.96638.36.027.98341.85.005.99545.94.011.99144.76.034.98341.09
MemE-3DGS (Ours).032.96840.28.015.98945.40.002.99749.11.008.99447.82.016.98944.29
", + "image_path": "c22a64d089710108902eb54bbf5f3c0676c27a34ad5b49378efc2ea91f322045.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 247, + 477, + 533 + ], + "blocks": [ + { + "bbox": [ + 138, + 247, + 477, + 533 + ], + "lines": [ + { + "bbox": [ + 138, + 247, + 477, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 247, + 477, + 533 + ], + "type": "image", + "image_path": "3dcb2d4a3a60e5096771e39f238c7d03a2e1e2717f75daca3e44cee8f220723e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 547, + 482, + 581 + ], + "lines": [ + { + "bbox": [ + 130, + 547, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 482, + 581 + ], + "type": "text", + "content": "Fig. 4: Qualitative Comparison on the Kubric Dataset. Our MemE effectively manages scenes with distractors that exhibit continuous motion across successive frames, resulting in better distractor elimination and static object preserving." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Kubric Dataset. Compared with previous works, our MemE achieves consistent improvement in terms of all scenes and all metrics, as shown in Table 2 and Fig. 4. Note that the distractors in the Kubric dataset exhibit continuous motion between different frames, increasing the probability of overlap, particularly for objects with simple textures. Leveraging the inherent memorization" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 476, + 342 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 476, + 342 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 476, + 342 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 476, + 342 + ], + "type": "image", + "image_path": "c81c7c924afc35e6c52105fd2dd0c2072fbb89e60a259fb2e2f78feb1fb0b00e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 355, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 401 + ], + "type": "text", + "content": "Fig.5: Quality Results of 3DGS and MemE-3DGS(Ours) on the Phototourism Dataset. The pairs of tri-column figures illustrate the test view images and the rendering images of 3DGS and MemE-3DGS, respectively. Our MemE-3DGS effectively eliminates artifacts caused by distractors and significant illumination variations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 485 + ], + "type": "text", + "content": "effect in view synthesis models, our MemE utilizes GMM to adaptively perform customized fitting for each scene. Consequently, our MemE enhances the performance of NeRF and 3DGS to achieve SoTA results, regardless of whether the inputs are temporally continuous or discrete, and remains robust to inputs with varying distractor ratios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 487, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 583 + ], + "type": "text", + "content": "Phototourism Dataset. The Phototourism dataset remains challenging for existing novel view synthesis methods due to the diverse sources of images. This variability results in differences in distractors, illumination, and even camera equipment across images. Nevertheless, our MemE-3DGS achieves satisfactory view rendering results on this dataset, as shown in Fig. 5. Notably, severe lighting variations, as well as distractors, can significantly impact the quality of rendered images of 3DGS. However, our MemE is capable of effectively mitigating the influence of lighting variations and distractors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 605, + 239, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 239, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 239, + 617 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "We conduct ablation studies on the BabyYoda scene within the RobustNeRF dataset. This scene comprises subsets designated for training, testing, and validation. Specifically, the training subset contains distractors, while the testing and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 277, + 197 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 277, + 197 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 277, + 197 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 277, + 197 + ], + "type": "image", + "image_path": "f020e8c7a457d2652675ed60f0ad88e5eb0611ca046a1da214bb4495e601874a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 198, + 286, + 208 + ], + "lines": [ + { + "bbox": [ + 137, + 198, + 286, + 208 + ], + "spans": [ + { + "bbox": [ + 137, + 198, + 286, + 208 + ], + "type": "text", + "content": "(a) Adaptability analysis of MemE-NeRF." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 333, + 114, + 468, + 198 + ], + "blocks": [ + { + "bbox": [ + 333, + 114, + 468, + 198 + ], + "lines": [ + { + "bbox": [ + 333, + 114, + 468, + 198 + ], + "spans": [ + { + "bbox": [ + 333, + 114, + 468, + 198 + ], + "type": "image", + "image_path": "0ab7dc82ac893e40afde025aa2dadc98da1f9728abb0aa2164ff435b205fb31d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 198, + 477, + 209 + ], + "lines": [ + { + "bbox": [ + 327, + 198, + 477, + 209 + ], + "spans": [ + { + "bbox": [ + 327, + 198, + 477, + 209 + ], + "type": "text", + "content": "(b) Adaptability analysis of MemE-3DGS." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "text", + "content": "Fig. 6: Adaptability Analysis of MemE on NeRF and 3DGS in Distractor-Free Scenes. MemE demonstrates comparable performance to Mip-NeRF and 3DGS in scenes devoid of distractors. Its adaptability across various scenarios, regardless of the presence of distractors, enhances the versatility of NeRF/3DGS frameworks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 479, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 479, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 479, + 314 + ], + "type": "text", + "content": "validation subsets are devoid of distractors. Additionally, the validation subset encompasses different views from those in the training set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 316, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 495 + ], + "type": "text", + "content": "Adaptability of MemE. While MemE performs admirably in scenes with distractors, we sought to evaluate its performance in scenarios where the presence of distractors is unknown. We conducted a comparison between Mip-NeRF 360, RobustNeRF*, and 3DGS with our MemE on the distractor-free BabyYoda scene. The PSNR curve during training, as depicted in Fig. 6, demonstrates that our MemE-NeRF achieves performance very similar to Mip-NeRF 360 without experiencing significant performance loss. In contrast, RobustNeRF* suffers from considerable performance degradation in distractor-free scenarios. Similarly, MemE-3DGS exhibits a similar trend compared to 3DGS. Thus, MemE demonstrates adaptability in adjusting the focus of NeRF/3DGS in different types of inputs. That is, in scenes with distractors, MemE can effectively and adaptively eliminate them, while in distractor-free scenes, MemE can focus solely on purity without causing significant performance degradation for NeRF/3DGS. Consequently, MemE serves as a plug-and-play module capable of adaptively handling various scenarios with or without distractors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Mask Fusion. In our MemE, we utilize mask fusion to generate purity attention by aggregating binary masks " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_n\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " across multiple scales. These multi-scale masks consist of fine-grained pixel-wise masks " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_i\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " and coarse-grained tile-wise masks " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_j\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", with pixel-wise masks providing detailed information and tile-wise masks aggregating local information within each residual tile. We explore various strategies for integrating hierarchical masks " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_n\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ", as detailed in Table 3. Initially, using only the most fine-grained mask " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_1" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " resulted in a significant drop in PSNR metric due to its inability to aggregate local information, resulting in noise. Conversely, employing solely the pixel-wise masks " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_i|i\\in \\{1,3,5,7\\}\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " improved performance by leveraging the OR operation to aggregate their advantages: " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_1" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " providing detail and " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_i|i\\in \\{3,5,7\\}\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " filtering noise. Additionally, we tested the performance using the MEAN operation on pixel-wise masks, denoted as " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_i|i\\in \\{1,3,5,7\\}\\}^*" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " in Table 3. Similarly, solely utilizing tile-wise masks also improved results by leveraging hierarchical residual information. However, our" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "type": "text", + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 189, + 135, + 425, + 234 + ], + "blocks": [ + { + "bbox": [ + 204, + 114, + 408, + 126 + ], + "lines": [ + { + "bbox": [ + 204, + 114, + 408, + 126 + ], + "spans": [ + { + "bbox": [ + 204, + 114, + 408, + 126 + ], + "type": "text", + "content": "Table 3: Quantitative Results of Ablation Study." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 135, + 425, + 234 + ], + "lines": [ + { + "bbox": [ + 189, + 135, + 425, + 234 + ], + "spans": [ + { + "bbox": [ + 189, + 135, + 425, + 234 + ], + "type": "table", + "html": "
MethodsPSNR ↑SSIM ↑LPIPS ↓
w/o mask fusionM130.090.800.31
{Mi|i ∈ {1,3,5,7}}30.990.800.29
{Mi|i ∈ {1,3,5,7}}*30.760.810.29
{Mj|j ∈ {4,8,16}}30.930.800.30
w/o adaptive thresholds{μ1(Γn)}24.230.770.34
{μ2(Γn)}28.500.770.32
linear moving30.740.790.31
Ours31.080.810.29
", + "image_path": "5659d05990927762b1da5bde25c692fd89d083905555b3ac05f6d2eb8c4b6290.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 257, + 479, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 479, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 479, + 280 + ], + "type": "text", + "content": "method, combining both pixel-wise and tile-wise masks, achieved the best performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "text", + "content": "Adaptive Thresholds. In our MemE, masks are derived by applying adaptive thresholds to hierarchical residual maps " + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\{\\epsilon_n\\}" + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "text", + "content": ". These thresholds transition exponentially from the distractor means " + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\{\\mu_2(\\Gamma_n)\\}" + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "text", + "content": " to the purity means " + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\{\\mu_1(\\Gamma_n)\\}" + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "text", + "content": ". The means " + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\mu_1" + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\mu_2" + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "text", + "content": " of the GMM are determined from training images of a specific scene at specific scale. We evaluate various threshold-generating strategies, as detailed in Table 3. Initially, we utilize the purity means " + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\{\\mu_1(\\Gamma_n)\\}" + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "text", + "content": " as thresholds, aiming to attain purity and exclude distractors. However, this approach leads to a significant performance drop. This drop occurs because the purity means " + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\{\\mu_1(\\Gamma_n)\\}" + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "text", + "content": " decrease rapidly due to the memorization effect during the early stages. However, some areas of purity may exhibit large residuals early on, remaining unaccounted for in the training procedure. Conversely, employing the distractor mean " + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\{\\mu_2(\\Gamma_n)\\}" + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "text", + "content": " as thresholds introduces some distractors into the training procedure, resulting in a performance decrease. Additionally, we explore the linear threshold strategy, where the thresholds transition linearly from " + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\{\\mu_2(\\Gamma_n)\\}" + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\{\\mu_1(\\Gamma_n)\\}" + }, + { + "bbox": [ + 130, + 281, + 482, + 474 + ], + "type": "text", + "content": ". The ablation results demonstrate that our method achieves the best performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 494, + 220, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 220, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 220, + 506 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "In this study, we introduce MemE, an unsupervised plug-and-play module designed to facilitate high-quality novel view synthesis in noisy input scenarios. MemE capitalizes on the inherent property known as the memorization effect during the parameter optimization phase of view synthesis models, enabling effective distractor filtering. MemE seamlessly integrates with both NeRF and 3DGS frameworks. Notably, our work marks the first instance of leveraging the memorization effect in the training process of NeRF and 3DGS. Additionally, MemE proves to be effective in environments, irrespective of the presence of distractors, thereby substantially improving the versatility of NeRF and 3DGS across a variety of input scenarios. We evaluate MemE across three datasets to assess its effectiveness against other SoTA methods. Both quantitative and qualitative results demonstrate that our approach achieves SoTA performance." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "content": "We thank Yangdi Lu, Guangchi Fang, and Runqing Jiang for their insightful comments and valuable discussions. This work was partially supported by the National Natural Science Foundation of China (62301601), Guangdong Basic and Applied Basic Research Foundation (2022B1515020103, 2023B1515120087), the Shenzhen Science and Technology Program (No. RCYX20200714114641140)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 246, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 246, + 480, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 268 + ], + "type": "text", + "content": "1. Arazo, E., Ortego, D., Albert, P., O'Connor, N., McGuinness, K.: Unsupervised label noise modeling and loss correction. In: ICML. pp. 312-321. PMLR (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 269, + 480, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 301 + ], + "type": "text", + "content": "2. Arpit, D., Jastrzegski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M.S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., et al.: A closer look at memorization in deep networks. In: ICML. pp. 233-242 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 302, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 302, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 302, + 480, + 334 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV. pp. 5855-5864 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 335, + 480, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 366 + ], + "type": "text", + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: CVPR. pp. 5470-5479 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 368, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 480, + 399 + ], + "type": "text", + "content": "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-NeRF: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "type": "text", + "content": "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensorRF: Tensorial radiance fields. In: ECCV. pp. 333-350. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 423, + 480, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 455 + ], + "type": "text", + "content": "7. Chen, M., Wang, L., Lei, Y., Dong, Z., Guo, Y.: Learning spherical radiance field for efficient " + }, + { + "bbox": [ + 138, + 423, + 480, + 455 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 138, + 423, + 480, + 455 + ], + "type": "text", + "content": " unbounded novel view synthesis. IEEE Transactions on Image Processing (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "text", + "content": "8. Cordeiro, F.R., Sachdeva, R., Belagiannis, V., Reid, I., Carneiro, G.: LongReMix: Robust learning with high confidence samples in a noisy label environment. PR 133, 109013 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 489, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 480, + 510 + ], + "type": "text", + "content": "9. Dai, P., Zhang, Y., Yu, X., Lyu, X., Qi, X.: Hybrid neural rendering for large-scale scenes with motion blur. In: CVPR. pp. 154-164 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "10. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological) 39(1), 1-22 (1977)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "type": "text", + "content": "1. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR. pp. 5501-5510 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "text", + "content": "2. Goli, L., Reading, C., Sellán, S., Jacobson, A., Tagliasacchi, A.: Bayes' rays: Uncertainty quantification for neural radiance fields. arXiv preprint arXiv:2309.03185 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "3. Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., et al.: Kubric: A scalable dataset generator. In: CVPR. pp. 3749-3761 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "4. Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., Tsang, I., Sugiyama, M.: Co-teaching: Robust training of deep neural networks with extremely noisy labels. NeurIPS 31 (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "15. Huang, X., Zhang, Q., Feng, Y., Li, H., Wang, X., Wang, Q.: HDR-NeRF: High dynamic range neural radiance fields. In: CVPR. pp. 18398-18408 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "text", + "content": "16. Jiang, R., Yan, Y., Xue, J.H., Chen, S., Wang, N., Wang, H.: Knowledge distillation meets label noise learning: Ambiguity-guided mutual label refinery. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "17. Jiang, R., Yan, Y., Xue, J.H., Wang, B., Wang, H.: When sparse neural network meets label noise learning: A multistage learning framework. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "text", + "content": "18. Jin, Y., Mishkin, D., Mishchuk, A., Matas, J., Fua, P., Yi, K.M., Trulls, E.: Image matching across wide baselines: From paper to practice. IJCV 129(2), 517-547 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 134, + 238, + 480, + 270 + ], + "type": "text", + "content": "19. Karim, N., Rizve, M.N., Rahnavard, N., Mian, A., Shah, M.: Unicon: Combating label noise through uniform selection and contrastive learning. In: CVPR. pp. 9676-9686 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 271, + 480, + 293 + ], + "type": "text", + "content": "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG 42(4) (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "21. Lee, D., Lee, M., Shin, C., Lee, S.: DP-NeRF: Deblurred neural radiance field with physical scene priors. In: CVPR. pp. 12386-12396 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "22. Lee, D., Oh, J., Rim, J., Cho, S., Lee, K.M.: ExBluRF: Efficient radiance fields for extreme motion blurred images. In: ICCV. pp. 17639-17648 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 336, + 480, + 369 + ], + "type": "text", + "content": "23. Levy, D., Peleg, A., Pearl, N., Rosenbaum, D., Akkaynak, D., Korman, S., Treibitz, T.: SeaThru-NeRF: Neural radiance fields in scattering media. In: CVPR. pp. 56–65 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "24. Li, J., Socher, R., Hoi, S.C.: DivideMix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "25. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: CVPR. pp. 6498-6508 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "type": "text", + "content": "26. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: DynIBaR: Neural dynamic image-based rendering. In: CVPR. pp. 4273-4284 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "27. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: CVPR. pp. 13-23 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "28. Lu, Y., Zhang, Y., Han, B., Cheung, Y.m., Wang, H.: Label-noise learning with intrinsically long-tailed data. In: ICCV. pp. 1369-1378 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 480, + 511 + ], + "type": "text", + "content": "29. Lu, Y., Bo, Y., He, W.: An ensemble model for combating label noise. In: Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining. pp. 608-617 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "30. Lu, Y., Bo, Y., He, W.: Noise attention learning: Enhancing noise robustness by gradient scaling. NeurIPS 35, 23164-23177 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "31. Lu, Y., He, W.: SELC: Self-ensemble label correction improves learning with noisy labels. arXiv preprint arXiv:2205.01156 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "text", + "content": "32. Lu, Y., He, W.: Learning with noisy ground truth: From 2d classification to 3d reconstruction. arXiv preprint arXiv:2406.15982 (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "33. Ma, L., Li, X., Liao, J., Zhang, Q., Wang, X., Wang, J., Sander, P.V.: Deblur-NeRF: Neural radiance fields from blurry images. In: CVPR. pp. 12861–12870 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "34. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: NeRF in the wild: Neural radiance fields for unconstrained photo collections. In: CVPR. pp. 7210-7219 (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "35. Mildenhall, B., Hedman, P., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T.: NeRFs in the dark: High dynamic range view synthesis from noisy raw images. In: CVPR. pp. 16190–16199 (2022)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "36. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "37. Mirzaei, A., Aumentado-Armstrong, T., Derpanis, K.G., Kelly, J., Brubaker, M.A., Gilitschenski, I., Levinshtein, A.: SPIn-NeRF: Multiview segmentation and perceptual inpainting with neural radiance fields. In: CVPR. pp. 20669–20679 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "38. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "39. Pearl, N., Treibitz, T., Korman, S.: NAN: Noise-aware nerfs for burst-denoising. In: CVPR. pp. 12672-12681 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "40. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: CVPR. pp. 12932-12942 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "41. Sabour, S., Vora, S., Duckworth, D., Krasin, I., Fleet, D.J., Tagliasacchi, A.: RobustNeRF: Ignoring distractors with robust losses. In: CVPR. pp. 20626-20636 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "text", + "content": "42. Song, H., Kim, M., Park, D., Shin, Y., Lee, J.G.: Learning from noisy labels with deep neural networks: A survey. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "43. Tanay, T., Leonardis, A., Maggioni, M.: Efficient view synthesis and 3d-based multi-frame denoising with multiplane feature representations. In: CVPR. pp. 20898-20907 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "44. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-NeRF: Scalable large scene neural view synthesis. In: CVPR. pp. 8248-8258 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "45. Teigen, A.L., Yip, M., Hamran, V.P., Skui, V., Stahl, A., Mester, R.: Removing adverse volumetric effects from trained neural radiance fields. arXiv preprint arXiv:2311.10523 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "46. Tschernezki, V., Larlus, D., Vedaldi, A.: NeuralDiff: Segmenting 3D objects that move in egocentric videos. In: Proceedings of the IEEE International Conference on 3D Vision (3DV). pp. 910-919 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "47. Tu, Y., Zhang, B., Li, Y., Liu, L., Li, J., Wang, Y., Wang, C., Zhao, C.R.: Learning from noisy labels with decoupled meta label purifier. In: CVPR. pp. 19934-19943 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "48. Turki, H., Zhang, J.Y., Ferroni, F., Ramanan, D.: SUDS: Scalable urban dynamic scenes. In: CVPR. pp. 12375-12385 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "49. Wang, D., Zhang, T., Abboud, A., Süsstrunk, S.: InpaintNeRF360: Text-guided 3d inpainting on unbounded neural radiance fields. arXiv preprint arXiv:2305.15094 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "type": "text", + "content": "50. Wang, H., Xu, X., Xu, K., Lau, R.W.: Lighting up NeRF via unsupervised decomposition and enhancement. In: ICCV. pp. 12632-12641 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 577 + ], + "type": "text", + "content": "51. Wang, P., Zhao, L., Ma, R., Liu, P.: BAD-NeRF: Bundle adjusted deblur neural radiance fields. In: CVPR. pp. 4170-4179 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 577, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 610 + ], + "type": "text", + "content": "52. Warburg, F., Weber, E., Tancik, M., Holynski, A., Kanazawa, A.: NeRF-busters: Removing ghostly artifacts from casually captured NeRFs. arXiv preprint arXiv:2304.10532 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "53. Weder, S., Garcia-Hernando, G., Monszpart, A., Pollefeys, M., Brostow, G.J., Firman, M., Vicente, S.: Removing objects from neural radiance fields. In: CVPR. pp. 16528-16538 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "54. Wei, F., Funkhouser, T., Rusinkiewicz, S.: Clutter detection and removal in 3d scenes with view-consistent inpainting. In: ICCV. pp. 18131-18141 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "Distractor-Free NVS via Exploiting Memorization Effect in Optimization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 226 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "55. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\\~ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS 35, 32653-32666 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "text", + "content": "56. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: ICCV. pp. 5752-5761 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 482, + 194 + ], + "type": "text", + "content": "57. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: NeRF++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "text", + "content": "58. Zhang, M., Zhao, X., Yao, J., Yuan, C., Huang, W.: When noisy labels meet long tail dilemmas: A representation calibration method. In: ICCV. pp. 15890-15900 (2023)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Distractors-Immune Representation Learning with Cross-modal Contrastive Regularization for Change Captioning/01669dd0-b2b9-4e44-8fe5-812e59df2574_content_list.json b/2024/Distractors-Immune Representation Learning with Cross-modal Contrastive Regularization for Change Captioning/01669dd0-b2b9-4e44-8fe5-812e59df2574_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..269c75e384f3a89e2a8c049222e576ba274080cc --- /dev/null +++ b/2024/Distractors-Immune Representation Learning with Cross-modal Contrastive Regularization for Change Captioning/01669dd0-b2b9-4e44-8fe5-812e59df2574_content_list.json @@ -0,0 +1,1989 @@ +[ + { + "type": "text", + "text": "Distractors-Immune Representation Learning with Cross-modal Contrastive Regularization for Change Captioning", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunbin $\\mathrm{Tu}^{1}$ , Liang $\\mathrm{Li}^{2*}$ , Li $\\mathrm{Su}^{1*}$ , Chenggang $\\mathrm{Yan}^{3}$ , and Qingming $\\mathrm{Huang}^{1}$", + "bbox": [ + 233, + 233, + 767, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Chinese Academy of Sciences, Beijing, China", + "bbox": [ + 295, + 261, + 702, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tuyunbin22@mails.ucas.ac.cn, {suli,qmhuang}@ucas.ac.cn", + "bbox": [ + 294, + 277, + 709, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Key Laboratory of AI Safety of CAS, Institute of Computing Technology, Chinese", + "bbox": [ + 220, + 290, + 781, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Academy of Sciences (CAS), Beijing, China liang.li@ict.ac.cn", + "bbox": [ + 284, + 304, + 720, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Hangzhou Dianzi University (HDU) & Lishui Institute of HDU, China", + "bbox": [ + 259, + 318, + 741, + 332 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Change captioning aims to succinctly describe the semantic change between a pair of similar images, while being immune to distractors (illumination and viewpoint changes). Under these distractors, unchanged objects often appear pseudo changes about location and scale, and certain objects might overlap others, resulting in perturbational and discrimination-degraded features between two images. However, most existing methods directly capture the difference between them, which risk obtaining error-prone difference features. In this paper, we propose a distractors-immune representation learning network that correlates the corresponding channels of two image representations and decorrelates different ones in a self-supervised manner, thus attaining a pair of stable image representations under distractors. Then, the model can better interact them to capture the reliable difference features for caption generation. To yield words based on the most related difference features, we further design a cross-modal contrastive regularization, which regularizes the cross-modal alignment by maximizing the contrastive alignment between the attended difference features and generated words. Extensive experiments show that our method outperforms the state-of-the-art methods on four public datasets. The code is available at https://github.com/tuyunbin/DIRL.", + "bbox": [ + 259, + 366, + 740, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Change Captioning $\\cdot$ Distractors-Immune Representation Learning $\\cdot$ Cross-modal Contrastive Regularization", + "bbox": [ + 259, + 657, + 740, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 709, + 375, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, researchers have witnessed the great success toward vision-language understanding and generation [21,30,43,44]. As an emerging task, change captioning [27,36,46] is to describe what has semantically changed between two similar images in natural language (Fig. 1 (a)-(d)). This task has a wide range of practical applications, such as generating reports for surveillance area changes [10]", + "bbox": [ + 212, + 739, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors", + "bbox": [ + 217, + 824, + 387, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/780defeab1a9bb1cda6c3bbdf70f15c483cdc1273b25f76b6c773854a9c6f632.jpg", + "image_caption": [ + "Fig. 1: The examples of change captioning under different scenarios. The first and second cases show that object moving and dropping, respectively. The third one shows the color change under distractors (viewpoint and illumination changes), where the real change is overwhelmed by pseudo changes. The last one shows with only distractors. Changed objects are shown in red boxes." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 496, + 295 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9d7c3d1efb7740b04dacf79024b38cea503ffff60c367a686415151fcde44dd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 146, + 781, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and pathological changes between medical images [17], as well as providing visually-impaired users with the explanations of image editing effects [29].", + "bbox": [ + 212, + 412, + 784, + 443 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This task also poses a formidable challenge: models should be powerful enough to not only understand the contents of two images, but also describe the semantic change between them, while being immune to distractors (viewpoint and illumination changes). In a dynamic environment, two images of the same scene are usually obtained in the presence of distractors. In this situation, unchanged objects between two images often appear obviously pseudo changes about the scale and location (Fig. 1 (c) (d)), where the features of same objects within the image pair might be perturbational. Especially, as a viewpoint drastically changes, certain objects might partially overlap the others, e.g., the brown ball is partially occluded by the red block in the \"after\" image of Fig. 1 (d). Accordingly, the feature discrimination might be weakened under this viewpoint. In short, there are perturbational and discrimination-degraded features between two image representations under distractors.", + "bbox": [ + 212, + 444, + 787, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most previous methods directly subtract between such two image representations [9,24,28,38] or compute their feature similarity [25,26,45,47], which risk capturing error-prone difference features in between. The latest works SCORER [35] and SMART [34] overcome the distractors by implementing contrastive learning between similar/dissimilar image pairs. By maximizing the alignment of similar ones, both methods make the features of unchanged objects non-perturbational under distractors. However, both methods disregard the feature discrimination-degraded problem, leaving the influence of distractors upon a pair of image representations remained.", + "bbox": [ + 212, + 642, + 787, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a Distractors-Immune Representation Learning (DIRL) network to attain a pair of stable image representations that are non-perturbational and discriminative under distractors, for robust change captioning. Concretely, given two raw image representations, DIRL first computes a", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y.Tu et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "channel correlation matrix between them. Next, DIRL implements cross-channel decorrelation to optimize this matrix to be close to the identity matrix, i.e., making the corresponding channels of two image representations have similar semantics, while enforcing different channels to be independent. In this self-supervised fashion, DIRL not only alleviates feature perturbation between two images, but also enhances feature discrimination for each image. As such, the model can sufficiently interact the two stable image representations to infer their reliable difference features, which are then translated into a linguistic sentence via a transformer decoder.", + "bbox": [ + 212, + 146, + 782, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During sentence generation, as the changed object typically occurs in a local region with weak feature, traditional attention mechanisms used in mainstream methods [14,25,31] may lead to less satisfactory cross-modal alignment. To facilitate correct alignment, we further design a Cross-modal Contrastive Regularization (CCR), which is built upon the cross-attention module of transformer decoder. When the cross-attention module yields the attended difference features, CCR further regularizes them by maximizing the contrastive alignment between them and generated words, which helps the decoder generate the change caption based on the most related difference features.", + "bbox": [ + 212, + 282, + 784, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our key contributions are: (1) We propose the novel DIRL network to learn a pair of stable image representations under distractors by correlating their corresponding channels and decorrelating different channels. Based on the two distractors-immune representations, the model is able to learn the robust difference features between them for caption generation. (2) We design the CCR to regularize the cross-modal alignment by maximizing the contrastive alignment between the features of attended difference and generated words. This helps the decoder generate words based on the most related difference features. (3) Our method performs favourably against state-of-the-art methods on four public datasets with different change scenarios.", + "bbox": [ + 212, + 419, + 784, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 594, + 387, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Change captioning is a new task in the vision-and-language area [4, 6, 12, 16, 32, 39]. The pioneer work [13] releases the dataset about changes in surveillance scenarios with potential illumination changes. Afterwards, Tan et al. [29] collect a dataset about image editing scenes. Both works [14, 24] develop two datasets for simulating distractors (viewpoint and illumination changes), where the unchanged objects' scale and location would illustrate obviously pseudo changes.", + "bbox": [ + 212, + 628, + 782, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To learn the difference features under distractors, prior works [9, 20, 24] directly subtract between two image representations, which are hard to generalize to unaligned image pairs. SGCC [18] introduces 3D information of object depths to overcome viewpoint changes. Recent works [14, 26, 31, 47] first measure the feature similarity to summarize the shared features and then remove them to obtain difference features. On the other hand, the works [7, 25, 45] correlate the similar features of two images to implicitly deduce the difference features. However, the above methods do not well address the influence of distractors upon image rep", + "bbox": [ + 212, + 719, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "DIRL with CCR for Change Captioning", + "bbox": [ + 462, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cbb6b4300e8081854a7e786e33a17857cc2e9b848e0dd08743bd1ddc343aae34.jpg", + "image_caption": [ + "Fig. 2: The framework of our method, where the core blocks are distractors-immune representation learning network and cross-modal contrastive regularization. FC and Concat are short for the fully-connect layer and concatenation operation." + ], + "image_footnote": [], + "bbox": [ + 215, + 147, + 784, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "resentations. The latest works SCORER [35] and SMART [34] try contrastive learning between similar/dissimilar image pairs, which model the alignment between similar ones to handle distractors. There are two differences between both works and our DIRL. First, SCORER and SMART capture relations between all paired/unpaired images in a batch, whereas DIRL computes the correlation between the feature channels of each image pair. Second, SCORER and SMART only make the features between two images non-perturbational under distractors, while DIRL additionally considers enhancing feature discrimination for each image to better recognize semantic changes.", + "bbox": [ + 212, + 393, + 784, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During caption generation, since the changed object typically appears in a local region with weak feature, it is difficult to learn reliable alignment via current attention modules in prevalent methods. To address this issue, SMART [34] employs part-of-speech to guide the decoder to dynamically use visual information. These works [9, 14, 35] propose a cycle consistence module to enforce correct cross-modal alignment. The other works [3, 7, 45] design a pre-training and fine-tuning strategy to strengthen the alignment. In this paper, we regularize cross-modal alignment by proposing CCR, which maximizes the contrastive alignment between the generated words and attended difference features. We find that two works in single-image captioning [6, 44] tried the similar idea. Both works model the coarse-grained cross-modal contrastive regularization between the input image and generated caption. Instead, our CCR reuses the attended difference features; establishes the fine-grained contrastive regularization between them and generated words during decoding. In this way, the model can use the most relevant difference features to predict next words.", + "bbox": [ + 212, + 530, + 787, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 777, + 380, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the overall framework of our method includes following parts. First, a pre-trained CNN model extracts the raw representations of an", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y.Tu et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "image pair, which are then transformed into a low-dimensional space. Next, the two image representations are fed into the proposed DIRL to make them non-perturbational and discriminative under distractors. Subsequently, a cross-attention module interacts the two stable representations to mine the shared features, which are removed to learn the difference features for decoding into a sentence. During sentence generation, the designed CCR regularizes the cross-modal alignment, so as to help the decoder generate the sentence based on the most related difference features.", + "bbox": [ + 212, + 146, + 782, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Image Pair Feature Extraction", + "text_level": 1, + "bbox": [ + 214, + 287, + 514, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a pair of images \"before\" $I_{bef}$ and \"after\" $I_{aft}$ , we first use a pre-trained CNN to extract their feature representations. They are denoted as $F_{bef}$ and $F_{aft}$ , where $F_{o} \\in \\mathbb{R}^{C \\times H \\times W}$ . $C, H, W$ indicate the number of channels, height, and width. Then, we project them into a low-dimensional space of $\\mathbb{R}^D$ by a 2D-convolution:", + "bbox": [ + 214, + 310, + 782, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {F} _ {o} = \\operatorname {c o n v} _ {2} \\left(F _ {o}\\right) + \\operatorname {p o s} \\left(F _ {o}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 385, + 784, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $o \\in (bef, aft)$ . pos is the learnable position encoding function.", + "bbox": [ + 214, + 407, + 709, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Distractors-immune Representation Learning", + "text_level": 1, + "bbox": [ + 214, + 441, + 633, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the two position-embedded image representations, we propose a DIRL network to make them non-perturbational and discriminative under distractors in a self-supervised manner. Concretely, we first project $\\tilde{F}_{bef}$ and $\\tilde{F}_{aft}$ into a common embedding space with shared parameters:", + "bbox": [ + 214, + 465, + 782, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {Y} _ {o} = \\operatorname {M L P} \\left(\\tilde {F} _ {o}\\right), o \\in (b e f, a f t), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 537, + 784, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where MLP is a two-layer multi-layer perceptron with the ReLU activation function in between. Then, we compute a channel correlation matrix between $\\tilde{Y}_{bef}$ and $\\tilde{Y}_{aft}$ along the batch dimension:", + "bbox": [ + 214, + 569, + 782, + 616 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {i j} = \\frac {\\sum_ {b} \\tilde {y} _ {b , i} ^ {b e f} \\tilde {y} _ {b , j} ^ {a f t}}{\\sqrt {\\sum_ {b} \\left(\\tilde {y} _ {b , i} ^ {b e f}\\right) ^ {2}} \\sqrt {\\sum_ {b} \\left(\\tilde {y} _ {b , j} ^ {a f t}\\right) ^ {2}}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 627, + 784, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $b$ indexes batch samples and $i,j$ index the channel dimension of features. $\\mathcal{C} \\in \\mathbb{R}^{D \\times D}$ is a channel correlation matrix. Inspired by the recent self-supervised learning methods [42, 48], we perform cross-channel decorrelation to enforce this matrix to be close to the identity matrix, which is implemented by using the $\\ell_2$ -norm minimization:", + "bbox": [ + 214, + 685, + 782, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i r l}} = \\sum_ {i} (1 - \\mathcal {C} _ {i i}) ^ {2} + \\alpha \\sum_ {i} \\sum_ {j \\neq i} \\mathcal {C} _ {i j} ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 768, + 784, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first term equates the diagonal of $\\mathcal{C}$ to one, i.e., the corresponding channels of two image representations will be correlated and thus have similar semantics", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "DIRL with CCR for Change Captioning", + "bbox": [ + 462, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "under distractors. The second term equates the off diagonal of $\\mathcal{C}$ to zero, i.e., the different channels will be decorrelated. This enhances the discrimination of each image representation. $\\alpha$ is a trade-off parameter to balance the importance between two terms, which is discussed in the supplementary material.", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Difference Representation Learning", + "text_level": 1, + "bbox": [ + 215, + 227, + 555, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The difference features are modeled based on the two distractors-immune image representations. As shown in Fig. 1, most objects are identical between two images. Hence, it is natural to extract their shared features to infer the difference features. For a powerful model, it should capture all potential changes w.r.t. both images. To this end, we first learn the difference features for each image and then combine them to construct the difference features between two images.", + "bbox": [ + 212, + 250, + 782, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, we first compute the shared features on each image by the multi-head cross-attention (MHCA) mechanism [40]:", + "bbox": [ + 212, + 340, + 782, + 371 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {F} _ {b e f} ^ {s} = \\operatorname {M H C A} \\left(\\tilde {F} _ {b e f}, \\tilde {F} _ {a f t}, \\tilde {F} _ {a f t}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 380, + 784, + 412 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {F} _ {a f t} ^ {s} = \\mathrm {M H C A} \\left(\\tilde {F} _ {a f t}, \\tilde {F} _ {b e f}, \\tilde {F} _ {b e f}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 407, + 622, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsequently, we subtract each from the corresponding image representation to compute the difference features on each image, respectively:", + "bbox": [ + 212, + 440, + 782, + 470 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {F} _ {b e f} ^ {d} = \\tilde {F} _ {b e f} - \\tilde {F} _ {b e f} ^ {s}, \\\\ \\tilde {\\tilde {z}} _ {d} ^ {c} \\quad \\tilde {\\tilde {r}} _ {c} ^ {c} \\quad \\tilde {\\tilde {r}} _ {s} ^ {s} \\end{array} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 478, + 784, + 506 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {F} _ {a f t} ^ {d} = \\tilde {F} _ {a f t} - \\tilde {F} _ {a f t} ^ {s}.\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 500, + 570, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Both $\\tilde{F}_{bef}^{d}$ and $\\tilde{F}_{aft}^{d}$ are then concatenated as the omni-representation of difference features between two images, which is implemented by a fully-connected layer with the ReLU activation function:", + "bbox": [ + 212, + 527, + 782, + 574 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {F} _ {d} = \\operatorname {R e L U} \\left(\\left[ \\tilde {F} _ {b e f} ^ {d}; \\tilde {F} _ {a f t} ^ {d} \\right] W _ {c} + b _ {c}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 583, + 784, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $[;]$ is a concatenation operation.", + "bbox": [ + 214, + 614, + 491, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Caption Generation", + "text_level": 1, + "bbox": [ + 214, + 651, + 426, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After learning $\\tilde{F}_d\\in \\mathbb{R}^{HW\\times D}$ , we use a standard transformer decoder [40] to translate it into a sentence. First, we obtain the embedding features of $m$ words (ground-truth words during training, predicted words during inference). Then, we use the multi-head self-attention (MHSA) layer to model relationships among these word embedding features $E[W] = \\{E[w_1],\\dots,E[w_m]\\}$ :", + "bbox": [ + 212, + 672, + 782, + 750 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {E} [ W ] = \\mathrm {M H S A} (E [ W ], E [ W ], E [ W ]). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 758, + 784, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsequently, we use the relation-embedded word features $\\hat{E}[W]$ to attend to the related difference features from $\\tilde{F}_d$ based on the MHCA layer:", + "bbox": [ + 212, + 782, + 782, + 815 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {V} = \\text {M H C A} (E [ \\hat {W} ], \\tilde {F} _ {d}, \\tilde {F} _ {d}). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 823, + 784, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y.Tu et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, the $\\hat{V} \\in \\mathbb{R}^{HW \\times D}$ is passed to a feed-forward network (FFN) to obtain the enhanced difference features:", + "bbox": [ + 214, + 145, + 782, + 175 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {V} ^ {\\prime} = \\operatorname {L a y e r N o r m} (\\hat {V} + \\operatorname {F F N} (\\hat {V})). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 184, + 785, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, the probability distributions of target words are calculated via a single hidden layer:", + "bbox": [ + 214, + 207, + 782, + 236 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nW = \\operatorname {S o f t m a x} \\left(\\hat {V} ^ {\\prime} W _ {h} + b _ {h}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 237, + 785, + 262 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $W_{h} \\in \\mathbb{R}^{D \\times \\nu}$ and $b_{h} \\in \\mathbb{R}^{\\nu}$ are the learnable parameters. $\\nu$ is the dimension of vocabulary size.", + "bbox": [ + 214, + 265, + 782, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Cross-modal Contrastive Regularization", + "text_level": 1, + "bbox": [ + 215, + 314, + 589, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is often the case that the changed object appears in a local area with weak feature, so it is difficult for the cross-attention module to directly align the captured difference features with word features. To this end, we devise the CCR to regularize the cross-modal alignment, which reuses the attended difference features $\\hat{V}$ computed in Eq. (9) and maximizes the contrastive alignment between them and generated words $\\hat{E}[W]$ . Specifically, we first compute the global representation for $\\hat{E}[W]$ and $\\hat{V}$ via mean-pooling operation, respectively:", + "bbox": [ + 214, + 335, + 782, + 443 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {E} [ W ] = \\frac {1}{m} \\hat {E} [ W ], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 450, + 784, + 484 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {V} = \\frac {1}{H W} \\hat {V},\n$$\n", + "text_format": "latex", + "bbox": [ + 457, + 483, + 545, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $m$ and $HW$ are the length of $\\hat{E}[W]$ and $\\hat{V}$ . Then, given a training batch, we sample $B$ pairs of global representations of generated words and attended difference features. For $k$ -th global word representation $\\tilde{E}[W]_k$ , $k$ -th global difference representation $\\tilde{V}_k$ is its positive, while the others will be the negatives in this batch. The similarity between two global representations is computed by dot-product:", + "bbox": [ + 214, + 515, + 782, + 604 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s i m} \\left(\\tilde {E} [ W ], \\tilde {V}\\right) = \\tilde {E} [ W ] \\cdot \\tilde {V} ^ {\\top}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 606, + 784, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we leverage the InfoNCE loss [22] to pull semantically close pairs of $\\tilde{E}[W]_k$ and $\\tilde{V}_k$ together and push away non-related pairs:", + "bbox": [ + 214, + 635, + 782, + 666 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t 2 v} = - \\frac {1}{B} \\sum_ {k} ^ {B} \\log \\frac {e ^ {\\left(\\operatorname {s i m} \\left(\\tilde {E} [ W ] _ {k} , \\tilde {V} _ {k}\\right) / \\tau\\right)}}{\\sum_ {r} ^ {B} e ^ {\\left(\\operatorname {s i m} \\left(\\tilde {E} [ W ] _ {k} , \\tilde {V} _ {r}\\right) / \\tau\\right)}},\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 674, + 651, + 714 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v 2 t} = - \\frac {1}{B} \\sum_ {k} ^ {B} \\log \\frac {e ^ {\\left(\\operatorname* {s i m} \\left(\\tilde {V} _ {k} , \\tilde {E} [ W ] _ {k}\\right) / \\tau\\right)}}{\\sum_ {r} ^ {B} e ^ {\\left(\\operatorname* {s i m} \\left(\\tilde {V} _ {k} , \\tilde {E} [ W ] _ {r}\\right) / \\tau\\right)}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 717, + 782, + 758 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c c r} = \\frac {1}{2} (\\mathcal {L} _ {t 2 v} + \\mathcal {L} _ {v 2 t}),\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 762, + 583, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature hyper-parameter. This loss formulates a self-supervisory signal to regularize the cross-modal alignment during caption generation, so as to improve the quality of generated captions.", + "bbox": [ + 214, + 794, + 799, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "DIRL with CCR for Change Captioning", + "bbox": [ + 462, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.6 Joint Training", + "text_level": 1, + "bbox": [ + 215, + 146, + 382, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We train the overall architecture end-to-end by maximizing the likelihood of the observed word sequence. Given the ground-truth words $(w_1^*,\\ldots ,w_m^*)$ , we minimize the negative log-likelihood loss:", + "bbox": [ + 214, + 167, + 787, + 213 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c a p} (\\theta) = - \\sum_ {t = 1} ^ {m} \\log p _ {\\theta} \\left(w _ {t} ^ {*} \\mid w _ {< t} ^ {*}\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 219, + 785, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $p_{\\theta}\\left(w_t^* \\mid w_{