Add Batch 98c571ea-f557-4ddc-86f3-aa05ad2582a6 data

.gitattributes

@@ -3130,3 +3130,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/LaPE_[[:space:]]Layer-adaptive[[:space:]]Position[[:space:]]Embedding[[:space:]]for[[:space:]]Vision[[:space:]]Transformers[[:space:]]with[[:space:]]Independent[[:space:]]Layer[[:space:]]Normalization/881d69d4-8f69-41f3-95ce-4351c44e1c35_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/LaRS_[[:space:]]A[[:space:]]Diverse[[:space:]]Panoptic[[:space:]]Maritime[[:space:]]Obstacle[[:space:]]Detection[[:space:]]Dataset[[:space:]]and[[:space:]]Benchmark/7b55dd4e-abd8-4f11-aca5-6c10f323df34_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Label[[:space:]]Shift[[:space:]]Adapter[[:space:]]for[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]under[[:space:]]Covariate[[:space:]]and[[:space:]]Label[[:space:]]Shifts/8826026c-6da5-4749-8a23-22b76030d60b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Label-Efficient[[:space:]]Online[[:space:]]Continual[[:space:]]Object[[:space:]]Detection[[:space:]]in[[:space:]]Streaming[[:space:]]Video/bb408240-93f4-4420-9e47-a5facd88851a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Label-Free[[:space:]]Event-based[[:space:]]Object[[:space:]]Recognition[[:space:]]via[[:space:]]Joint[[:space:]]Learning[[:space:]]with[[:space:]]Image[[:space:]]Reconstruction[[:space:]]from[[:space:]]Events/cebb8d22-3e6f-4881-895d-c913e7109230_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Label-Guided[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]for[[:space:]]Continual[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]on[[:space:]]2D[[:space:]]Images[[:space:]]and[[:space:]]3D[[:space:]]Point[[:space:]]Clouds/3a801de3-50c8-469a-b26b-c00f633cc663_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Label-Noise[[:space:]]Learning[[:space:]]with[[:space:]]Intrinsically[[:space:]]Long-Tailed[[:space:]]Data/ca2911cd-1897-45f1-bc9c-ab89b815151c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Landscape[[:space:]]Learning[[:space:]]for[[:space:]]Neural[[:space:]]Network[[:space:]]Inversion/f1107abc-1d35-4283-b4db-d9fd38f81a61_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large[[:space:]]Selective[[:space:]]Kernel[[:space:]]Network[[:space:]]for[[:space:]]Remote[[:space:]]Sensing[[:space:]]Object[[:space:]]Detection/1c2f7154-2a88-4538-ac11-efbaf0ea97b7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large-Scale[[:space:]]Land[[:space:]]Cover[[:space:]]Mapping[[:space:]]with[[:space:]]Fine-Grained[[:space:]]Classes[[:space:]]via[[:space:]]Class-Aware[[:space:]]Semi-Supervised[[:space:]]Semantic[[:space:]]Segmentation/85af6521-8e24-45e3-ac82-fb3352f9480e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large-Scale[[:space:]]Person[[:space:]]Detection[[:space:]]and[[:space:]]Localization[[:space:]]Using[[:space:]]Overhead[[:space:]]Fisheye[[:space:]]Cameras/b691cb26-f3c1-4d93-ba53-9242b6258c4e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Late[[:space:]]Stopping_[[:space:]]Avoiding[[:space:]]Confidently[[:space:]]Learning[[:space:]]from[[:space:]]Mislabeled[[:space:]]Examples/c0b4cc8e-d6dd-4adb-8592-11f986f8c0fc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Latent-OFER_[[:space:]]Detect,[[:space:]]Mask,[[:space:]]and[[:space:]]Reconstruct[[:space:]]with[[:space:]]Latent[[:space:]]Vectors[[:space:]]for[[:space:]]Occluded[[:space:]]Facial[[:space:]]Expression[[:space:]]Recognition/b49e72cb-c8b1-4c66-830a-4cce4a977773_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LayoutDiffusion_[[:space:]]Improving[[:space:]]Graphic[[:space:]]Layout[[:space:]]Generation[[:space:]]by[[:space:]]Discrete[[:space:]]Diffusion[[:space:]]Probabilistic[[:space:]]Models/77145711-9475-40d3-bb1f-4deba9a37184_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LeaF_[[:space:]]Learning[[:space:]]Frames[[:space:]]for[[:space:]]4D[[:space:]]Point[[:space:]]Cloud[[:space:]]Sequence[[:space:]]Understanding/51c2f292-724b-40e6-8771-779ef4bbccc4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Leaping[[:space:]]Into[[:space:]]Memories_[[:space:]]Space-Time[[:space:]]Deep[[:space:]]Feature[[:space:]]Synthesis/5ad3f2c3-b163-4339-b33a-6f2219ed23a6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learn[[:space:]]TAROT[[:space:]]with[[:space:]]MENTOR_[[:space:]]A[[:space:]]Meta-Learned[[:space:]]Self-Supervised[[:space:]]Approach[[:space:]]for[[:space:]]Trajectory[[:space:]]Prediction/e84db066-834e-41d1-b5a0-8e0c6af0f9b7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learned[[:space:]]Compressive[[:space:]]Representations[[:space:]]for[[:space:]]Single-Photon[[:space:]]3D[[:space:]]Imaging/78c901d2-1e87-453e-b9f2-75e42afe6005_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learned[[:space:]]Image[[:space:]]Reasoning[[:space:]]Prior[[:space:]]Penetrates[[:space:]]Deep[[:space:]]Unfolding[[:space:]]Network[[:space:]]for[[:space:]]Panchromatic[[:space:]]and[[:space:]]Multi-spectral[[:space:]]Image[[:space:]]Fusion/e7814c90-67a9-4d35-afdb-629b41a8c850_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Adaptive[[:space:]]Neighborhoods[[:space:]]for[[:space:]]Graph[[:space:]]Neural[[:space:]]Networks/002465ad-e9c1-4d74-bc7a-fd7e74285b28_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Clothing[[:space:]]and[[:space:]]Pose[[:space:]]Invariant[[:space:]]3D[[:space:]]Shape[[:space:]]Representation[[:space:]]for[[:space:]]Long-Term[[:space:]]Person[[:space:]]Re-Identification/33efc350-232b-4a85-9272-b3026fb811bb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Concise[[:space:]]and[[:space:]]Descriptive[[:space:]]Attributes[[:space:]]for[[:space:]]Visual[[:space:]]Recognition/eb8f500b-e381-4a8d-9629-e0fa391759b6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Concordant[[:space:]]Attention[[:space:]]via[[:space:]]Target-aware[[:space:]]Alignment[[:space:]]for[[:space:]]Visible-Infrared[[:space:]]Person[[:space:]]Re-identification/f96bbbd7-1bf4-40cd-9545-e535cf6aa841_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Continuous[[:space:]]Exposure[[:space:]]Value[[:space:]]Representations[[:space:]]for[[:space:]]Single-Image[[:space:]]HDR[[:space:]]Reconstruction/4c1481b1-79c6-4632-b011-bcc21d585466_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Correction[[:space:]]Filter[[:space:]]via[[:space:]]Degradation-Adaptive[[:space:]]Regression[[:space:]]for[[:space:]]Blind[[:space:]]Single[[:space:]]Image[[:space:]]Super-Resolution/eaad2ba2-88cf-4d2a-a334-fd43747c9ca0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Cross-Modal[[:space:]]Affinity[[:space:]]for[[:space:]]Referring[[:space:]]Video[[:space:]]Object[[:space:]]Segmentation[[:space:]]Targeting[[:space:]]Limited[[:space:]]Samples/a0fe2dc9-eb10-4dc5-8bfa-8f5f03212fc3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Cross-Representation[[:space:]]Affinity[[:space:]]Consistency[[:space:]]for[[:space:]]Sparsely[[:space:]]Supervised[[:space:]]Biomedical[[:space:]]Instance[[:space:]]Segmentation/e6b0e920-39b7-4259-9938-916b268677a8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Data-Driven[[:space:]]Vector-Quantized[[:space:]]Degradation[[:space:]]Model[[:space:]]for[[:space:]]Animation[[:space:]]Video[[:space:]]Super-Resolution/8790a28a-4db4-4d06-856c-466efed197ec_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Depth[[:space:]]Estimation[[:space:]]for[[:space:]]Transparent[[:space:]]and[[:space:]]Mirror[[:space:]]Surfaces/4e49b9b1-7456-432b-aaae-0da1220f3a37_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Fine-Grained[[:space:]]Features[[:space:]]for[[:space:]]Pixel-Wise[[:space:]]Video[[:space:]]Correspondences/fd6b6adc-a24b-4ea2-b63f-ac1670224d69_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Foresightful[[:space:]]Dense[[:space:]]Visual[[:space:]]Affordance[[:space:]]for[[:space:]]Deformable[[:space:]]Object[[:space:]]Manipulation/b7d619c9-c5f2-438e-896d-a2c31b01ebe0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Gabor[[:space:]]Texture[[:space:]]Features[[:space:]]for[[:space:]]Fine-Grained[[:space:]]Recognition/6ca3f851-37f0-49f5-a456-09ea4b08c273_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Global-aware[[:space:]]Kernel[[:space:]]for[[:space:]]Image[[:space:]]Harmonization/edcde37a-d026-4b0d-9a7a-a8c99e117c7b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Hierarchical[[:space:]]Features[[:space:]]with[[:space:]]Joint[[:space:]]Latent[[:space:]]Space[[:space:]]Energy-Based[[:space:]]Prior/240d7938-e848-4274-a9c2-b84ff0fe9ab2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Human[[:space:]]Dynamics[[:space:]]in[[:space:]]Autonomous[[:space:]]Driving[[:space:]]Scenarios/cdd5a67c-1587-4c07-9231-ae20be13d8d8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Human-Human[[:space:]]Interactions[[:space:]]in[[:space:]]Images[[:space:]]from[[:space:]]Weak[[:space:]]Textual[[:space:]]Supervision/6803bdfe-c2d2-4fd6-a5a5-6f813d705eb4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Image[[:space:]]Harmonization[[:space:]]in[[:space:]]the[[:space:]]Linear[[:space:]]Color[[:space:]]Space/be202573-b803-48c5-8b87-29cda14139c4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Image-Adaptive[[:space:]]Codebooks[[:space:]]for[[:space:]]Class-Agnostic[[:space:]]Image[[:space:]]Restoration/4f85f017-28ce-47a6-a95c-6487947424f7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Long-Range[[:space:]]Information[[:space:]]with[[:space:]]Dual-Scale[[:space:]]Transformers[[:space:]]for[[:space:]]Indoor[[:space:]]Scene[[:space:]]Completion/4473583d-9b9f-42e0-9960-ac81cde7532b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Navigational[[:space:]]Visual[[:space:]]Representations[[:space:]]with[[:space:]]Semantic[[:space:]]Map[[:space:]]Supervision/94d1f5dd-be00-46b4-8d2e-2c6681cb4e80_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Neural[[:space:]]Eigenfunctions[[:space:]]for[[:space:]]Unsupervised[[:space:]]Semantic[[:space:]]Segmentation/4dcc0481-9970-4a10-9577-a5671c84635b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Neural[[:space:]]Implicit[[:space:]]Surfaces[[:space:]]with[[:space:]]Object-Aware[[:space:]]Radiance[[:space:]]Fields/3c0bb7d1-9040-4c42-8743-ea0aac8c43c4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Non-Local[[:space:]]Spatial-Angular[[:space:]]Correlation[[:space:]]for[[:space:]]Light[[:space:]]Field[[:space:]]Image[[:space:]]Super-Resolution/0facca26-8b6b-4a15-98c0-1ceb58a77d13_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Optical[[:space:]]Flow[[:space:]]from[[:space:]]Event[[:space:]]Camera[[:space:]]with[[:space:]]Rendered[[:space:]]Dataset/9da70e78-d6a8-4bda-bae6-2f6d222d7409_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Point[[:space:]]Cloud[[:space:]]Completion[[:space:]]without[[:space:]]Complete[[:space:]]Point[[:space:]]Clouds_[[:space:]]A[[:space:]]Pose-Aware[[:space:]]Approach/c5a386b7-062e-40bf-ab26-f63c862742fe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Pseudo-Relations[[:space:]]for[[:space:]]Cross-domain[[:space:]]Semantic[[:space:]]Segmentation/10b88ddb-f472-4de5-8cfb-db9e250c0176_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Rain[[:space:]]Location[[:space:]]Prior[[:space:]]for[[:space:]]Nighttime[[:space:]]Deraining/f13e3907-fb21-485a-916a-d1036549cfe5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Robust[[:space:]]Representations[[:space:]]with[[:space:]]Information[[:space:]]Bottleneck[[:space:]]and[[:space:]]Memory[[:space:]]Network[[:space:]]for[[:space:]]RGB-D-based[[:space:]]Gesture[[:space:]]Recognition/ffcfd550-d4a2-4cb4-a120-2a5976bb8508_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Semi-supervised[[:space:]]Gaussian[[:space:]]Mixture[[:space:]]Models[[:space:]]for[[:space:]]Generalized[[:space:]]Category[[:space:]]Discovery/c328493e-9994-4352-8a93-3b569e354069_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Shape[[:space:]]Primitives[[:space:]]via[[:space:]]Implicit[[:space:]]Convexity[[:space:]]Regularization/4387a009-eca0-4928-920d-1473c95da518_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Spatial-context-aware[[:space:]]Global[[:space:]]Visual[[:space:]]Feature[[:space:]]Representation[[:space:]]for[[:space:]]Instance[[:space:]]Image[[:space:]]Retrieval/4a7da155-d324-4f0d-9452-e7f76abd7205_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Support[[:space:]]and[[:space:]]Trivial[[:space:]]Prototypes[[:space:]]for[[:space:]]Interpretable[[:space:]]Image[[:space:]]Classification/1bf5aedf-f1e8-4793-8d34-430db9b0cd03_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Symmetry-Aware[[:space:]]Geometry[[:space:]]Correspondences[[:space:]]for[[:space:]]6D[[:space:]]Object[[:space:]]Pose[[:space:]]Estimation/c3bb6c4d-c63d-4614-92bb-10b441ef4c78_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Trajectory-Word[[:space:]]Alignments[[:space:]]for[[:space:]]Video-Language[[:space:]]Tasks/e49e204e-97bd-485d-8320-f15ff5bdec7b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Unified[[:space:]]Decompositional[[:space:]]and[[:space:]]Compositional[[:space:]]NeRF[[:space:]]for[[:space:]]Editable[[:space:]]Novel[[:space:]]View[[:space:]]Synthesis/d796471e-04ce-409b-90cf-4f80b508e362_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Versatile[[:space:]]3D[[:space:]]Shape[[:space:]]Generation[[:space:]]with[[:space:]]Improved[[:space:]]Auto-regressive[[:space:]]Models/4e583ffa-62ea-4806-8a7a-46ee129f849e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Vision-and-Language[[:space:]]Navigation[[:space:]]from[[:space:]]YouTube[[:space:]]Videos/05335e36-6140-429d-8434-dd018538220c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]a[[:space:]]More[[:space:]]Continuous[[:space:]]Zero[[:space:]]Level[[:space:]]Set[[:space:]]in[[:space:]]Unsigned[[:space:]]Distance[[:space:]]Fields[[:space:]]through[[:space:]]Level[[:space:]]Set[[:space:]]Projection/918f7816-6b1d-45a6-a6ec-23336f1dbd99_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]a[[:space:]]Room[[:space:]]with[[:space:]]the[[:space:]]Occ-SDF[[:space:]]Hybrid_[[:space:]]Signed[[:space:]]Distance[[:space:]]Function[[:space:]]Mingled[[:space:]]with[[:space:]]Occupancy[[:space:]]Aids[[:space:]]Scene[[:space:]]Representation/c8aa23d3-c76a-4daf-a8f1-9ca330b3d300_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]by[[:space:]]Sorting_[[:space:]]Self-supervised[[:space:]]Learning[[:space:]]with[[:space:]]Group[[:space:]]Ordering[[:space:]]Constraints/f7288260-030f-4ffe-bd8d-4ee76970c742_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]from[[:space:]]Noisy[[:space:]]Data[[:space:]]for[[:space:]]Semi-Supervised[[:space:]]3D[[:space:]]Object[[:space:]]Detection/a3c8a1dc-20e7-4f50-8e36-d4ddc2e6295d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]from[[:space:]]Noisy[[:space:]]Pseudo[[:space:]]Labels[[:space:]]for[[:space:]]Semi-Supervised[[:space:]]Temporal[[:space:]]Action[[:space:]]Localization/911172cc-132a-4651-a728-4f623936c716_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]from[[:space:]]Semantic[[:space:]]Alignment[[:space:]]between[[:space:]]Unpaired[[:space:]]Multiviews[[:space:]]for[[:space:]]Egocentric[[:space:]]Video[[:space:]]Recognition/3976a321-a1a8-401c-abba-91b9fbe78433_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]in[[:space:]]Imperfect[[:space:]]Environment_[[:space:]]Multi-Label[[:space:]]Classification[[:space:]]with[[:space:]]Long-Tailed[[:space:]]Distribution[[:space:]]and[[:space:]]Partial[[:space:]]Labels/0590a041-a804-4736-ab7e-fab59b5b776d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]to[[:space:]]Distill[[:space:]]Global[[:space:]]Representation[[:space:]]for[[:space:]]Sparse-View[[:space:]]CT/816ad801-b997-48a5-9faf-a09a225c6e9c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]to[[:space:]]Generate[[:space:]]Semantic[[:space:]]Layouts[[:space:]]for[[:space:]]Higher[[:space:]]Text-Image[[:space:]]Correspondence[[:space:]]in[[:space:]]Text-to-Image[[:space:]]Synthesis/580abc17-52d1-48f0-9975-3f39f0fd195f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]to[[:space:]]Ground[[:space:]]Instructional[[:space:]]Articles[[:space:]]in[[:space:]]Videos[[:space:]]through[[:space:]]Narrations/a5da633a-33ad-4510-84f7-55e37a75aa45_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/Label-Efficient Online Continual Object Detection in Streaming Video/bb408240-93f4-4420-9e47-a5facd88851a_content_list.json

@@ -0,0 +1,1526 @@
+[
+    {
+        "type": "text",
+        "text": "Label-Efficient Online Continual Object Detection in Streaming Video",
+        "text_level": 1,
+        "bbox": [
+            130,
+            130,
+            841,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jay Zhangjie Wu $^{1}$  David Junhao Zhang $^{1}$  Wynne Hsu $^{2}$  Mengmi Zhang $^{3,4}$  Mike Zheng Shou $^{1*}$",
+        "bbox": [
+            91,
+            179,
+            883,
+            200
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Show Lab,  $^{2}$ National University of Singapore  \n $^{3}$ School of Computer Science and Engineering, Nanyang Technological University, Singapore  \n $^{4}$ CFAR and I2R, Agency for Science, Technology and Research, Singapore",
+        "bbox": [
+            112,
+            210,
+            854,
+            265
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            297,
+            313,
+            316
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Humans can watch a continuous video stream and effortlessly perform continual acquisition and transfer of new knowledge with minimal supervision yet retaining previously learnt experiences. In contrast, existing continual learning (CL) methods require fully annotated labels to effectively learn from individual frames in a video stream. Here, we examine a more realistic and challenging problem—Label-Efficient Online Continual Object Detection (LEOCOD) in streaming video. We propose a plug-and-play module, Efficient-CLS, that can be easily inserted into and consistently improve existing CL algorithms for object detection in video streams with reduced data annotation costs and model retraining time. We show that our method has achieved significant improvement with minimal forgetting across all supervision levels on two challenging CL benchmarks for streaming real-world videos. Remarkably, with only  $25\\%$  annotated video frames, our proposed method still outperforms the state-of-the-art CL models trained with  $100\\%$  annotations on all video frames. The data and source code will be publicly available at https://github.com/showlab/Efficient-CLS.",
+        "bbox": [
+            73,
+            332,
+            473,
+            650
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            681,
+            209,
+            696
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Humans have the ability to continuously learn from an ever-changing environment, while retaining previously learnt experiences. In contrast to human learning, prior works [3, 2, 11, 34, 7] show that deep neural networks are prone to catastrophic forgetting. To address the forgetting problem, existing works in continual learning (CL) primarily focus on class-incremental image classification or object detection. Their experiment settings are often idealistic and simplified, where i.i.d. static images are usually grouped by class and incrementally presented to computational models in sequence. To learn a particular task containing specific",
+        "bbox": [
+            75,
+            707,
+            470,
+            875
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "classes, an agent can go through all the data of current task over multiple epochs. After that, the learned classes in current task become unavailable, i.e., no overlaps between the sets of learned classes and unseen classes.",
+        "bbox": [
+            500,
+            300,
+            890,
+            359
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, these experiment designs deviate from the online continual learning (OCL) setting in the real world, where an agent learns from temporally correlated non-i.i.d. video streams in one single pass. Given context regularities in natural environments, an agent is likely to encounter cases when objects of previously learnt classes co-occur with unknown objects from unseen classes, e.g., a computer mouse and a computer monitor often co-occur. Taking these considerations, [34] introduces OCL on object detection in real-world video streams. They evaluate existing CL approaches on this setting and report a huge performance gap compared with offline training.",
+        "bbox": [
+            496,
+            364,
+            892,
+            546
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Based on the setting in [34], we take a significant step further and introduce a novel problem setting called Label-Efficient Online Continual Object Detection (LEOCOD), which highlights two unique challenges. First, the setting in [34] is such that the CL algorithms are trained with every mini-batch over multiple passes. We tighten the training recipe in LEOCOD to strictly online, where data is allowed to have one single pass and models are trained on the entire video dataset for only one epoch. Second, existing CL models require fully supervised training where box-level ground truth labels of every object on every video frame have to be obtained from human annotators. Unlike static images, acquiring human annotations for object detection on videos can be expensive and daunting. Thus, in LEOCOD, the video frames per mini-batch are sparsely annotated to alleviate the burdens of extensive human labeling, making LEOCOD one step closer to real world application.",
+        "bbox": [
+            496,
+            549,
+            893,
+            806
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Cognitive science works [35, 19] show that humans are efficient at continuously learning from very few annotated data samples. We get inspirations from the theory of Complementary Learning Systems (CLS) in human brains [18], and propose a plug-and-play module for the LEOCOD task, dubbed as Efficient-CLS. In Efficient-CLS, we introduce",
+        "bbox": [
+            496,
+            809,
+            893,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding Author.",
+        "bbox": [
+            94,
+            886,
+            225,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "19246",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/a090acd2ce689d64a3637fba2655497866caff7450015ef2e8e9c44ef677bcb6.jpg",
+        "image_caption": [
+            "Figure 1. (a) Problem introduction: An agent continuously learns from a never-ending online video stream over time. In each training step, out of a mini-batch containing 16 consecutive video frames, only a fixed proportion of frames are labeled (green boundary), while the rest of the frames are unlabeled (orange boundary). Following [34], the video frame after every training mini-batch (transparent) is held out for testing. After every 100 training steps, the agent is evaluated on all the video frames from the test set for object detection. (b) Key results: Our proposed method (red) consistently outperforms the best competitive baseline (blue) by a margin of  $5\\%$ . Remarkably, our model, trained at  $25\\%$  annotation cost, surpasses the best baseline trained at  $100\\%$  (grey line). The orange cross denotes the performance of the state-of-the-art model, which is  $15\\%$  lower than our method."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            93,
+            529,
+            306
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/234e7cabfb9d3c0d9353c67dc8de8cfcc68f8333688ee4fc4fea263dba5918d3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            539,
+            93,
+            870,
+            305
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "two feed-forward neural networks as slow and fast learners. In the fast learner, memory is rapidly adapted to the current task. The weights of the slow learner change a little on each reinstatement, and are maintained by taking the exponential moving average (EMA) of the fast learner's weights over time. Though a few continual learning models in previous works [4, 24] also use a similar source of inspiration, they miss the effect of reciprocal connections from slow learners to fast learners, which we intend to address. Inspired by the bidirectional interaction in CLS [14], we reactivate the weights of the slow learners to predict meaningful pseudo labels from the unlabeled video frames and use these pseudo labels to guide the training of the fast learner, closing the loop between the two systems.",
+        "bbox": [
+            75,
+            422,
+            470,
+            633
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We demonstrate the versatility and effectiveness of our Efficient-CLS on two standard real-world video datasets, OAK [34] and EgoObjects [1]. Our proposed method can be easily integrated into existing CL models and consistently improve their performance by a large margin in LEOCOD. It is worth noting that, with only  $25\\%$  labeled data, our method surpasses the comparative baselines trained with full supervision (Figure 1(b)).",
+        "bbox": [
+            75,
+            633,
+            468,
+            755
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To summarize, we make the following key contributions:",
+        "bbox": [
+            96,
+            755,
+            467,
+            770
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We introduce a new, challenging and important problem of label-efficient online continual object detection (LEOCOD) in video streams. Solving this problem would greatly benefit real-world applications in reducing annotation cost and model retraining time.",
+            "- We propose Efficient-CLS, a plug-and-play module inspired from the Complementary Learning Systems (CLS) theory, which can be integrated into existing CL"
+        ],
+        "bbox": [
+            94,
+            775,
+            467,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "models and learn efficiently and effectively with less supervision and minimal forgetting.",
+        "bbox": [
+            529,
+            422,
+            890,
+            454
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- We benchmark existing CL methods on the task of LEOCOD and demonstrate the state-of-the-art performance of our method through extensive experiments.",
+        "bbox": [
+            517,
+            458,
+            893,
+            503
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            518,
+            640,
+            534
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Continual Learning",
+        "text_level": 1,
+        "bbox": [
+            500,
+            544,
+            689,
+            560
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To alleviate catastrophic forgetting, many continual learning methods exploit an external buffer where a limited number of old samples are stored and used for replay when adapting to a new task. iCaRL [26] stores the representative exemplars in past tasks for knowledge distillation and prototype rehearsal. Gradient Episodic Memory (GEM) [21] formulates optimization constraints on the exemplars in memory. Averaged GEM (A-GEM) [8] is an improved version of GEM that achieves faster training and less memory consumption. GDumb [25] greedily stores samples in memory as they come and trains a model using samples only in the memory. Dark Experience Replay++ (DER++) [6] combines replay with knowledge distillation and regularization, and samples logits along the entire optimization trajectory.",
+        "bbox": [
+            496,
+            568,
+            890,
+            779
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In contrast to classical continual learning where data are separated by task boundaries and models are trained with multiple iterations in every task, we examine a more realistic and challenging problem where data are provided in tiny batches and models are trained on these batches only once. Recently, online continual learning (OCL) has gained increasing interests in computer vision [3, 7, 29, 9, 34]. Many OCL methods rely on representative memory replays to pre",
+        "bbox": [
+            496,
+            779,
+            890,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "19247",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "vent forgetting. [3] utilizes gradients of network parameters to select replay samples of maximum diversity. Subsequent works [2, 29] propose to use losses and scoring functions as criteria for selecting the most representative samples for replay. However, these approaches tackle image classification problem in an artificial setting, where new classes appear in a specific order. Their performance in real-world vision tasks remains unclear.",
+        "bbox": [
+            75,
+            90,
+            470,
+            210
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Lately, [34] benchmarks CL methods in the real-world online setting with full supervision. As the video streams arrive endlessly in a real-time manner, assigning annotations to all the video frames for training computational models is laborious and time-consuming. It becomes even more daunting in object detection tasks where class labels and bounding boxes of all objects on a video frame have to be provided. Reducing burdensome costs of labeling remains an under-explored and challenging problem in online continual object detection. We propose a self-sustaining Efficient-CLS, which is capable of exploiting the unlabeled video frames by pseudo-labeling when the number of labeled frames is limited.",
+        "bbox": [
+            75,
+            212,
+            470,
+            407
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2. Complementary Learning Systems",
+        "text_level": 1,
+        "bbox": [
+            76,
+            419,
+            379,
+            435
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The essence of fast and slow learning in Complementary Learning Systems (CLS) has benefited several continual learning algorithms in image recognition [24, 23, 28, 4, 15]. However, these methods either require the task boundaries, which are not applicable in our online video setting, or they require to train fast and slow learning systems with replay samples from the same replay buffer, which could easily lead to overfitting problem when the replay buffer has limited capacity. To eliminate overfitting problem, [28] and [15] utilize generative replay models to couple sequential tasks in a latent embedding space. While generative approaches have succeeded in artificial and simple datasets, they often fail in complex vision tasks, e.g., object detection. DualNet [23] also employs a slow-fast learning architecture. However, its update process for the slow learner is more computationally demanding because it leans on both self-supervised and supervised learning objectives. The self-supervised phase demands additional training iterations on extensive batches to yield satisfactory results, thereby delaying the training of the fast learner. In contrast, our method updates the slow learner in real-time using the fast learner's weights, obviating the need for distinct slow learner training. Our approach is optimized for online CL scenarios where computational resources are constrained.",
+        "bbox": [
+            75,
+            443,
+            468,
+            806
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3. Semi-Supervised Learning",
+        "text_level": 1,
+        "bbox": [
+            76,
+            816,
+            316,
+            832
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The goal of semi-supervised learning (SSL) [33, 5, 37, 36] is to reduce the demand of labeled data and harness unlabeled data for performance improvement. [31] first introduce SSL to the context of CL, and propose a strategy",
+        "bbox": [
+            76,
+            839,
+            468,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "that combines pseudo-labeling, consistency regularization, out-of-distribution detection, and knowledge distillation to solve the problem of class-incremental image classification. However, to distill knowledge from previous tasks, their model relies heavily on the task boundary that identifies the change of training classes, which is not available in our LEOCOD setting. To reduce annotation costs in object detection, several methods [13, 32, 20] capitalize the teacher-student networks. In general, a teacher model predicts pseudo labels or enforces a consistency loss to guide the student networks. However, these previous works on semi-supervised object detection only consider the offline setting on static image datasets, while none of them has been extended to online continual learning on dynamic video streams. We discovered new insights that pseudo-labeling from slow learner to fast learner can not only reduce annotation overhead, but also alleviate forgetting, which is critical to the design of efficient continual learners in real world.",
+        "bbox": [
+            496,
+            90,
+            890,
+            363
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            500,
+            377,
+            589,
+            393
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Problem Setting",
+        "text_level": 1,
+        "bbox": [
+            500,
+            402,
+            660,
+            419
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We advance the online continual object detection in [34] to a label-efficient and computationally-efficient setting—Label-Efficient Online Continual Object Detection (LEOCOD). In contrast to the setting in [34] where video frames are extensively annotated and trained with multiple epochs, an agent in LEOCOD continuously learns from a sparsely annotated video stream in a single pass over time (see Figure 1(a)).",
+        "bbox": [
+            496,
+            426,
+            890,
+            547
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Formally, we consider the online continual object detection on a continuum of video streams  $\\mathcal{D} = \\{D_1,\\dots ,D_T\\}$  where at time step  $t$  , a learning agent receives a mini-batch of continuous video frames  $D_{t}$  from current environment for online training (one single pass). To perform label-efficient object detection, within the batch  $D_{t}$  , only a subset of video frames  $D_{t}^{s} = (X_{t}^{s},Y_{t}^{s})$  are labeled, while the remaining video frames  $D_{t}^{u} = (X_{t}^{u})$  are unlabeled. For each labeled data sample, its annotation contains the bounding box locations and their corresponding class labels.",
+        "bbox": [
+            496,
+            547,
+            890,
+            700
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Efficient Complementary Learning Systems",
+        "text_level": 1,
+        "bbox": [
+            500,
+            710,
+            870,
+            726
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We propose a plug-and-play module dubbed as EfficientCLS. Specifically, it consists of two feed-forward networks: (i) the fast learner is designed to quickly encode new knowledge from current data stream and then consolidate it to the slow learner; and (ii) the slow learner accumulates the acquired knowledge from fast learner over time and guides the fast learner with meaningful pseudo labels, when full supervision is not available. Same as [26, 8, 25, 6, 34], we maintain an external episodic memory, as a replay buffer, to store exemplars that can be retrieved for replays alongside ongoing video stream. As the fast and slow learners are",
+        "bbox": [
+            496,
+            734,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "19248",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/02a21dd849294a5f80d943914591152a18efa413983492843b7fea011e65dfed.jpg",
+        "image_caption": [
+            "Figure 2. The framework of Efficient-CLS. At each learning step, the system receives a batch of temporally continuous data  $D_{t}$ , including labeled (green) and unlabeled (orange) frames. The fast learner trains the labeled frames alongside a small subset of labeled exemplars retrieved from episodic memory with the supervised loss  $\\mathcal{L}_{sup}$ . Meanwhile, the fast learner leverages the pseudo labels generated by the slow learner to optimize a pseudo loss  $\\mathcal{L}_{pseudo}$ . To reinstate memory of the slow learner, the weights of the slow learner are updated by taking the Exponential Moving Average (EMA) of the fast learner's weights. The fast and slow learners are complementary to each other, forming a positive feedback loop."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            101,
+            92,
+            872,
+            329
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "model agnostic, our Efficient-CLS can be easily integrated into existing CL models, which leads to less supervision and minimal forgetting.",
+        "bbox": [
+            75,
+            431,
+            470,
+            479
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Learning with Labeled Frames. The fast learner and slow learner use the same standard Faster-RCNN [27] detector  $f$ . Despite the same architecture, the weights of the fast and slow learners are not shared. We use  $\\theta_{F}$  and  $\\theta_{S}$  to denote the network parameters for fast and slow learners respectively. As shown in Figure 2, at each training step  $t$ , we use the labeled video frames  $D_{t}^{s} = (X_{t}^{s}, Y_{t}^{s})$  to optimize the fast learner  $\\theta_{F}$  with the standard supervised loss  $\\mathcal{L}_{sup}$  in Faster-RCNN [27]. It consists of four losses: Region Proposal Network (RPN) classification loss  $\\mathcal{L}_{cls}^{rpn}$ , RPN regression loss  $\\mathcal{L}_{reg}^{rpn}$ , Region of Interest (ROI) classification loss  $\\mathcal{L}_{cls}^{roi}$ , and ROI regression loss  $\\mathcal{L}_{reg}^{roi}$ . We define  $\\mathcal{L}_{sup}$  as:",
+        "bbox": [
+            75,
+            493,
+            468,
+            679
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {s u p} = \\mathcal {L} _ {c l s} ^ {r p n} \\left(X _ {t} ^ {s}, Y _ {t} ^ {s}\\right) + \\mathcal {L} _ {r e g} ^ {r p n} \\left(X _ {t} ^ {s}, Y _ {t} ^ {s}\\right) + \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            686,
+            468,
+            713
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {c l s} ^ {r o i} \\left(X _ {t} ^ {s}, Y _ {t} ^ {s}\\right) + \\mathcal {L} _ {r e g} ^ {r o i} \\left(X _ {t} ^ {s}, Y _ {t} ^ {s}\\right).\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            210,
+            710,
+            421,
+            727
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Learning with Unlabeled Frames. We introduce a pseudo-labeling paradigm to capitalize the information from unlabeled video frames  $D_{t}^{u} = (X_{t}^{u})$  for training. In our early exploration, we intuitively use the fast learner for pseudo-labeling as it quickly adapts the knowledge of nearby frames. However, we observe that using the pseudo labels generated by the fast learner for self-replay exhibits biases towards recently seen objects, which is less effective in preventing forgetting. This has also been verified in our ablation study (Section 5.3). In contrast, the slow learner",
+        "bbox": [
+            75,
+            750,
+            470,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "preserves the semantic knowledge over a longer time span which generates pseudo labels with fewer biases. This encourages the fast learner to capture more generic scene representations, hence, in turn, contributing to reinstatement of memory in the slow learner (Section 3.2), resulting in a positive feedback loop.",
+        "bbox": [
+            496,
+            431,
+            890,
+            523
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Given all these design considerations, the slow learner takes the unlabeled video frames  $D_{t}^{u}$  as inputs to estimate the possible objects of interest and their corresponding bounding box locations. For brevity, we refer these \"pseudo bounding boxes and their corresponding class labels\" as \"pseudo labels\" in the paper. To get rid of false positives, we apply a threshold  $\\tau$  to filter out bounding boxes with predicted low confidence scores. Moreover, there also exist repetitive boxes which negatively impact the quality of pseudo-labeling. To address this issue, we use the technique of class-wise non-maximum suppression (NMS) [27] to remove the overlapped boxes and get the high-quality pseudo labels. Formally, the procedure of pseudo label generation is summarized below:",
+        "bbox": [
+            496,
+            523,
+            892,
+            733
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nY _ {t} ^ {u} = \\mathrm {N M S} ([ f (X _ {t} ^ {u}; \\theta_ {S}) ] _ {> \\tau}), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            594,
+            741,
+            890,
+            758
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $[\\cdot ]_{>\\tau}$  denotes the bounding box selection with confidence score larger than  $\\tau$",
+        "bbox": [
+            496,
+            763,
+            890,
+            795
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Given that the video streams are captured from the egocentric perspective in the real world, head and body motions may lead to undesired motion blur effects on some video frames. To enforce our module to learn invariant object representations from these video frames, same as the previous work [38], we apply data augmentation techniques on the pseudo-labeled frames, including 2D image",
+        "bbox": [
+            496,
+            795,
+            890,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "19249",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "crops, rotations, and flipping. Note that different from image classification, the predicted bounding box locations also need to be updated accordingly after image augmentations. We denote these pseudo-labeled video frames and their re-adjusted pseudo labels after data augmentations as  $(\\tilde{X}_t^u,\\tilde{Y}_t^u)$ . We can then use these pseudo pairs  $(\\tilde{X}_t^u,\\tilde{Y}_t^u)$  to train the fast learner by optimizing the pseudo loss  $\\mathcal{L}_{pseudo}:= \\mathcal{L}_{cls}^{roi}(\\tilde{X}_t^u,\\tilde{Y}_t^u) + \\mathcal{L}_{req}^{roi}(\\tilde{X}_t^u,\\tilde{Y}_t^u)$ .",
+        "bbox": [
+            75,
+            90,
+            468,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Overall, our Efficient-CLS is jointly trained with the following losses:  $\\mathcal{L}_{total} = \\mathcal{L}_{sup} + \\lambda_{pseudo}\\mathcal{L}_{pseudo}$ , where  $\\lambda_{pseudo}$  is the weight of  $\\mathcal{L}_{pseudo}$ .",
+        "bbox": [
+            75,
+            212,
+            468,
+            258
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Synapses Consolidation via EMA. To alleviate forgetting of obtained knowledge, we apply Exponential Moving Average (EMA) to gradually update the slow learner with the fast learner's weights. The evolving weight changes in the slow learner are functionally correlated with the memory consolidation mechanism in the hippocampus and the neocortex [4]. Formally, we define EMA process as:",
+        "bbox": [
+            75,
+            268,
+            468,
+            375
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\theta_ {S} = \\alpha \\theta_ {S} + (1 - \\alpha) \\theta_ {F}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            380,
+            468,
+            396
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where the  $\\alpha \\in [0,1]$  is EMA rate. According to the stability-plasticity dilemma, a smaller  $\\alpha$  means faster adaptation but less memorization. Empirically, we set  $\\alpha = 0.99$ , which leads to best performance.",
+        "bbox": [
+            75,
+            402,
+            468,
+            463
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experimental details",
+        "text_level": 1,
+        "bbox": [
+            76,
+            474,
+            274,
+            491
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Datasets",
+        "text_level": 1,
+        "bbox": [
+            76,
+            500,
+            179,
+            513
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We evaluate the continual learning methods on two realistic and challenging video datasets, OAK [34] and EgoObjects [1], for the task of online continual object detection.",
+        "bbox": [
+            75,
+            522,
+            468,
+            568
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "OAK [34] is a large egocentric video stream dataset spanning nine months of a graduate student's life, consisting of 7.6 million frames of 460 video clips with a total length of 70.2 hours [30]. The dataset contains 103 object categories. We follow [34] in the ordering of training and testing data splits. One frame every 16 consecutive video frames lasting for 30 seconds is held out to construct a test set and the remaining frames are used for training.",
+        "bbox": [
+            75,
+            568,
+            468,
+            688
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "EgoObjects [1] is one of the largest object-centric datasets focusing on object detection task. The dataset contains around 100k video frames with  $\\sim 250\\mathrm{k}$  annotations, which correspond to 277 categories and 1110 main objects [22]. Additionally, the dataset follows a long-tailed distribution which makes the task more challenging and real-world oriented. We adopt the same rule as OAK dataset to split the training and testing set.",
+        "bbox": [
+            75,
+            688,
+            468,
+            809
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Baselines",
+        "text_level": 1,
+        "bbox": [
+            76,
+            816,
+            184,
+            830
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We compare our model against the following baselines: 1) Vanilla training: Incremental Training is a naive baseline trained sequentially over the entire video stream without any measures to avoid catastrophic forgetting; Offline",
+        "bbox": [
+            76,
+            839,
+            468,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training is an upper bound that can access all the data throughout training; 2) Continual learning algorithms: EWC [16], iCaRL [26], A-GEM [8], GDumb [25], DER++ [6] and iOD [17].",
+        "bbox": [
+            498,
+            90,
+            890,
+            151
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The iCaRL model implemented by [34] stands as the SOTA method in online continual object detection. We reproduce their results using the released code<sup>1</sup>. When calculating RPN and ROI losses for replay samples, their iCaRL model neglects the losses of background proposals and penalizes the foreground losses according to the proportion of the current samples and replay samples. We empirically found that this trick hinders the model from effective episodic replay, thus resulting in severe forgetting. Therefore, we re-implement the iCaRL by discarding the reweighting trick and reverting back to the standard RPN and ROI losses. We name these two different implementations as iCaRL(Wang et al.) and iCaRL(our impl.), respectively.",
+        "bbox": [
+            496,
+            152,
+            890,
+            348
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3. Evaluation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            357,
+            620,
+            371
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Protocols. First, we define the annotation cost as the proportion of number of labeled frames versus the total 16 frames within a mini-batch  $D_{t}$ . For example, if 2 out of 16 consecutive frames within  $D_{t}$  get labeled, the annotation cost is  $2 / 16 = 12.5\\%$ . The frames to be labeled are randomly selected within each mini-batch  $D_{t}$ . Considering that different choices of labeled frames might influence the model performance, for fair comparisons between models, we fix the choice of randomly selected labeled frames and use the same labeled and unlabeled frames for training all models. Based on the various annotation costs, we introduce two training protocols: fully supervised protocol (100% annotation cost) and sparse annotation protocol (where the annotation cost is less than 100%). In sparse annotation protocol, we further split the training experiments based on 50%/25%/12.5%/6.25% annotation costs.",
+        "bbox": [
+            496,
+            380,
+            890,
+            621
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Metrics. We evaluate the models with three standard metrics: continual average precision (CAP), final average precision (FAP) and forgetfulness (F) [34]. We employ an Average Precision (AP) metric at an Intersection over Union (IoU) threshold of 0.5, commonly referred to as AP50.",
+        "bbox": [
+            496,
+            633,
+            890,
+            708
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "CAP shows the average performance over the entire video stream. That is,",
+        "bbox": [
+            500,
+            709,
+            890,
+            738
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathrm {C A P} = \\frac {1}{N} \\sum_ {i = 0} ^ {N} \\mathrm {C A P} _ {t _ {i}} = \\frac {1}{N C} \\sum_ {i = 0} ^ {N} \\sum_ {c = 0} ^ {C} \\mathrm {C A P} _ {t _ {i}} ^ {c}, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            535,
+            750,
+            890,
+            790
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathrm{CAP}_{t_i}^c$  is the AP of class  $c$  on test set at the  $i$ -th evaluation step,  $N$  is the total number of evaluation steps.",
+        "bbox": [
+            496,
+            801,
+            890,
+            832
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "FAP is the final performance of a model after seeing the entire video. That is,  $\\mathrm{FAP} = \\mathrm{CAP}_{t_N}$ , where  $t_N$  denotes the last evaluation step.",
+        "bbox": [
+            496,
+            832,
+            890,
+            877
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{1}$ https://github.com/oakdata/benchmark",
+        "bbox": [
+            517,
+            886,
+            728,
+            898
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "19250",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/aac219989275d564f8d39834abf219c67bce25adb0da56e6c1ae0309096c1c5d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td rowspan=\"2\">Annotation Cost</td><td colspan=\"3\">OAK</td><td colspan=\"3\">EgoObjects</td></tr><tr><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td></tr><tr><td>Incremental</td><td>100%</td><td>8.38</td><td>7.72</td><td>0.03</td><td>10.21</td><td>3.55</td><td>1.48</td></tr><tr><td>Offline Training</td><td>100%</td><td>48.28</td><td>35.23</td><td>-</td><td>86.18</td><td>59.81</td><td>-</td></tr><tr><td>EWC</td><td>100%</td><td>7.73</td><td>7.02</td><td>-0.12</td><td>5.15</td><td>1.60</td><td>0.57</td></tr><tr><td>iOD</td><td>100%</td><td>7.92</td><td>7.14</td><td>0.98</td><td>8.80</td><td>2.64</td><td>0.00</td></tr><tr><td>iCaRL(Wang et al.)</td><td>100%</td><td>22.89</td><td>16.60</td><td>-2.95</td><td>37.61</td><td>21.71</td><td>2.79</td></tr><tr><td>iCaRL(our impl.)</td><td>100%</td><td>36.14</td><td>26.26</td><td>-4.89</td><td>60.80</td><td>36.41</td><td>-0.60</td></tr><tr><td rowspan=\"2\">w/ Efficient-CLS</td><td>25%</td><td>38.36(+2.22)</td><td>26.64(+0.38)</td><td>-8.20(-3.31)</td><td>61.26(+0.46)</td><td>39.58(+3.17)</td><td>-3.48(-2.88)</td></tr><tr><td>100%</td><td>40.24(+4.10)</td><td>28.18(+1.92)</td><td>-8.10(-3.21)</td><td>67.05(+6.25)</td><td>40.36(+3.95)</td><td>-3.67(-3.07)</td></tr><tr><td>A-GEM</td><td>100%</td><td>36.94</td><td>26.19</td><td>-5.54</td><td>58.79</td><td>35.88</td><td>-8.38</td></tr><tr><td rowspan=\"2\">w/ Efficient-CLS</td><td>25%</td><td>37.06(+0.12)</td><td>26.36(+0.17)</td><td>-7.76(-2.22)</td><td>63.06(+4.27)</td><td>39.46(+3.58)</td><td>-7.49(+0.89)</td></tr><tr><td>100%</td><td>39.87(+2.93)</td><td>27.97(+1.78)</td><td>-7.17(-1.63)</td><td>66.94(+8.15)</td><td>39.57(+3.69)</td><td>-11.68(-3.30)</td></tr><tr><td>GDumb</td><td>100%</td><td>35.27</td><td>25.29</td><td>-6.59</td><td>58.85</td><td>36.38</td><td>-5.21</td></tr><tr><td rowspan=\"2\">w/ Efficient-CLS</td><td>25%</td><td>37.67(+2.40)</td><td>25.59(+0.30)</td><td>-9.30(-2.71)</td><td>62.70(+3.85)</td><td>38.78(+2.40)</td><td>-8.86(-3.65)</td></tr><tr><td>100%</td><td>38.61(+3.34)</td><td>26.04(+0.75)</td><td>-9.14(-2.55)</td><td>63.55(+4.70)</td><td>38.98(+2.60)</td><td>-7.50(-2.29)</td></tr><tr><td>DER++</td><td>100%</td><td>37.79</td><td>25.24</td><td>-2.87</td><td>55.82</td><td>30.84</td><td>-6.08</td></tr><tr><td rowspan=\"2\">w/ Efficient-CLS</td><td>25%</td><td>37.93(+0.14)</td><td>25.64(+0.4)</td><td>-8.90(-6.03)</td><td>59.70(+3.88)</td><td>34.15(+3.31)</td><td>-11.21(-5.13)</td></tr><tr><td>100%</td><td>39.61(+1.82)</td><td>26.73(+1.49)</td><td>-8.30(-5.43)</td><td>62.01(+6.19)</td><td>33.09(+2.25)</td><td>-11.05(-4.97)</td></tr></table>",
+        "bbox": [
+            84,
+            87,
+            885,
+            343
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Overall performance of existing algorithms and Efficient-CLS on OAK and EgoObjects. iCaRL(Wang et al.) denotes the SOTA model presented in [34], and iCaRL(our impl.) is the same method by our implementation.",
+        "bbox": [
+            75,
+            348,
+            893,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "$\\mathbf{F}$  estimates the forgetfulness of the model due to the sequential training. It takes into account the time interval between the presence of an object category and its subsequent presence. For a class  $c$ , we sort the  $\\mathrm{CAP}_{t_i}^c$  according to the time interval  $k$  between evaluation time  $t_i$  and the last time  $t_i - k$  the model is trained on  $c$ . After  $\\mathrm{CAP}_{t_i}^c$  is sorted, all  $\\mathrm{CAP}_{t_i}^c (i = 0,\\dots ,T)$  are divided into  $K$  bins  $B_{kmin},\\dots ,B_{kmax}$  according to the time interval  $k$ . The average CAP  $(\\mathrm{aCAP}_k)$  of each bin  $B_{k}$  is defined as the model's performance for detecting class  $c$  after the model has not been trained on  $c$  for  $k$  time steps. The forgetfulness (F) of the class  $c$  is defined as the weighted sum of the performance decrease at each time:",
+        "bbox": [
+            75,
+            395,
+            472,
+            592
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF ^ {c} = \\sum_ {k = k m i n} ^ {k m a x} \\frac {k - k m i n}{\\sum_ {k = k m i n} ^ {k m a x} k - k m i n} \\times \\left(\\mathrm {a C A P} _ {k m i n} - \\mathrm {a C A P} _ {k}\\right). \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            76,
+            603,
+            482,
+            655
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The overall forgetfulness is:  $\\mathrm{F} = \\frac{1}{C}\\sum_{c = 0}^{C}\\mathrm{F}^{c}$",
+        "bbox": [
+            76,
+            654,
+            383,
+            674
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4. Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            76,
+            681,
+            294,
+            696
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For a fair comparison, we followed the prior work [34] to use Faster-RCNN [27] with ResNet-50 backbone [12] as our object detection network, which is initialized by the weights pre-trained on PASCAL VOC [10]. We used Adam optimizer with a constant learning rate of 0.0001, and the batch size was set to 16 frames. Same as [34], we maintained a replay buffer with 5 samples per class. At each time step  $t$ , we first randomly retrieved 16 video frames from the replay buffer for joint training. We used confidence thresh  $\\tau = 0.7$  to generate pseudo-labels for unlabeled frames, and applied EMA rate  $\\alpha = 0.99$  to update the slow learner. The weight of pseudo loss  $\\lambda_{pseudo}$  was set to 1.0. All the experiments were conducted on 2 NVIDIA RTX 3090 GPUs.",
+        "bbox": [
+            75,
+            704,
+            470,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Results",
+        "text_level": 1,
+        "bbox": [
+            498,
+            393,
+            586,
+            409
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Fully Supervised Protocol",
+        "text_level": 1,
+        "bbox": [
+            500,
+            419,
+            733,
+            435
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As the previous work [34] focuses on online continual object detection (OCOD) in video streams, we first evaluated model performance in fully supervised setting (i.e.,  $100\\%$  annotation cost), where all video frames are paired with human labels. The reported results were measured by standard metrics (CAP, FAP, and F, Section 4.3) in Table 1.",
+        "bbox": [
+            496,
+            443,
+            890,
+            535
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Performance of CL baselines. [34] benchmarked Incremental, EWC, iCaRL(Wang et al.), and Offline Training on OAK dataset. They found the replay-based method (i.e. iCaRL(Wang et al.)) outperforms regularization-based method (i.e. EWC) by  $10\\%$  in FAP, while iCaRL(Wang et al.) has a huge gap of  $30\\%$  compared with Offline Training. Similar observations were made in Table 1, but the performance of Incremental and EWC were  $4\\%$  lower than that in [34], as in our setting we only trained each mini-batch of video frames once (they trained each mini-batch 10 times). As mentioned in Section 4.2, we introduce several variations to the original design of iCaRL(Wang et al.). Compared with iCaRL(Wang et al.), we observed a huge performance boost in FAP from  $22.89\\%$  to  $36.14\\%$  on OAK and from  $37.61\\%$  to  $60.80\\%$  on EgoObjects, abridging the gap between the baseline and Offline Training.",
+        "bbox": [
+            496,
+            551,
+            890,
+            792
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We further adapted other standard CL baselines, including iOD, A-GEM, GDumb, DER++, to the OCOD setting for comparisons. iOD is the SOTA method in offline class-incremental object detection. Though it performs well in prior setting, iOD collapses when adapted to online video streams. One explanation is that iOD requires explicit task boundary to trigger the reshape of model gradients that opti",
+        "bbox": [
+            496,
+            795,
+            890,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "19251",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/33fe0a64f37aecc71768e8c1c4c8ff86bb795b645e3f00e83a30203a3e49b417.jpg",
+        "image_caption": [
+            "Figure 3. The changes of  $\\mathrm{CAP}_{t_i}^c$  with sampled categories on OAK dataset. The x-axis denotes time step across the entire video stream. The y-axis denotes the AP50 of the category at specific time step (i.e.,  $\\mathrm{CAP}_{t_i}^c$ ). The grey line indicates the existence of the category."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            88,
+            92,
+            883,
+            256
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "mizes knowledge sharing between adjacent tasks. However, in online video streams, the task boundaries by classes are no longer available and the change of tasks is hard to identify, resulting in the failure to prevent forgetting. In contrast, replay-based methods (i.e., A-GEM, GDumb and DER++) generalize much better to the real-world streaming video.",
+        "bbox": [
+            75,
+            314,
+            470,
+            405
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Performance w/ Efficient-CLS. Inheriting from the benefit of fast and slow learning with EMA, our Efficient-CLS consistently improves all the state-of-the-art CL methods (i.e. iCaRL(our impl.), A-GEM, GDumb and DER++) by a significant margin. Taking iCaRL(our impl.) for example, with Efficient-CLS, we observed superior performance and minimal forgetting compared to baseline methods, even when categories appeared infrequently (e.g., sculpture in Figure 3). Since semantic contextual information is more important in indoor environments on EgoObjects compared to the outdoor environments in OAK, we noticed that the improvement brought by our method is even greater on EgoObjects with an increase of  $6.25\\%$  in FAP,  $3.95\\%$  in CAP  $\\%$  and  $3.07\\%$  in F. Consistent performance gains are also noticeable for A-GEM, GDumb and DER++, demonstrating the effectiveness and flexibility of our method.",
+        "bbox": [
+            75,
+            416,
+            468,
+            657
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2. Sparse Annotation Protocol",
+        "text_level": 1,
+        "bbox": [
+            76,
+            665,
+            326,
+            681
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Performance of CL baselines. The sparse annotation protocol is more challenging than the previous fully supervised protocol as shown by the performance differences when number of annotated video frames decreases (compare the performance of each colored bar along the x-axis within each subplot in Figure 4). We noted that GDumb is more resilient against the reduce of supervision. Specifically, from Table 2 at the lowest annotation cost of  $6.25\\%$  on EgoObjects dataset, GDumb achieves the highest performance of  $38.74\\%$ ,  $22.69\\%$  and  $-4.53\\%$  in FAP, CAP and F, which surpasses other baselines by a considerable margin. Same observations can be made on OAK dataset. One possible explanation is that GDumb only trains the data stored in the balanced replay buffer, which makes it less vulnerable",
+        "bbox": [
+            75,
+            688,
+            470,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/88aea2c26004d72c77b0f4cf9ad195f09c2423a021d54903470e4016ed0d6140.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"3\">OAK</td><td colspan=\"3\">EgoObjects</td></tr><tr><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td></tr><tr><td>iCaRL</td><td>23.04</td><td>17.75</td><td>-3.31</td><td>32.91</td><td>19.15</td><td>-1.36</td></tr><tr><td>w/ Efficient-CLS</td><td>29.72</td><td>20.31</td><td>-5.36</td><td>40.16</td><td>23.95</td><td>-2.01</td></tr><tr><td>A-GEM</td><td>23.59</td><td>16.15</td><td>-2.99</td><td>21.84</td><td>12.28</td><td>0.82</td></tr><tr><td>w/ Efficient-CLS</td><td>30.18</td><td>20.59</td><td>-5.44</td><td>38.96</td><td>23.79</td><td>-5.64</td></tr><tr><td>GDumb</td><td>27.37</td><td>19.64</td><td>-4.25</td><td>38.74</td><td>22.69</td><td>-4.53</td></tr><tr><td>w/ Efficient-CLS</td><td>29.07</td><td>19.99</td><td>-6.01</td><td>40.09</td><td>23.67</td><td>-5.16</td></tr><tr><td>DER++</td><td>24.21</td><td>15.93</td><td>-3.79</td><td>16.95</td><td>8.48</td><td>2.03</td></tr><tr><td>w/ Efficient-CLS</td><td>28.63</td><td>19.64</td><td>-4.60</td><td>35.78</td><td>20.74</td><td>-4.69</td></tr></table>",
+        "bbox": [
+            500,
+            310,
+            890,
+            422
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Performance of Efficient-CLS at low annotation cost of  $6.25\\%$ . The best results are bold-faced.",
+        "bbox": [
+            498,
+            429,
+            892,
+            458
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "able to class imbalance problem brought by the reduce of labeled samples in the training set.",
+        "bbox": [
+            498,
+            474,
+            892,
+            506
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Performance w/ Efficient-CLS. Our proposed Efficient-CLS is a plug-and-play module that can be easily inserted into and enhance existing continual learning algorithms with the ability to use unlabeled video frames effectively. In both OAK and EgoObjects dataset, Efficient-CLS consistently improves the comparative SOTAs in all three evaluation metrics regardless of various degrees of annotation cost. As shown in Table 2, at a lower annotation cost of  $6.25\\%$ , Efficient-CLS doubles the performance of DER++ and A-GEM in terms of FAP and CAP, and achieves an even larger improvement in preventing forgetting. Thanks to the useful information from pseudo labels predicted by the slow learner in Efficient-CLS, our method is more robust to various annotation costs, compared with SOTAs (compare the rate of change of blue bars vs. red bars over different degrees of annotation cost). Most remarkably, Efficient-CLS with  $25\\%$  annotation cost has already outperformed comparative SOTAs with  $100\\%$  annotation cost (see Table 1).",
+        "bbox": [
+            496,
+            518,
+            892,
+            790
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            500,
+            801,
+            653,
+            816
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We assessed the importance of our key design choices. The Complementary Learning Systems (CLS) in EfficientCLS is the key for rapidly adapting to learn new tasks, meanwhile, retaining previously learnt knowledge. It constitutes of two memory reinstatement mechanisms: one is",
+        "bbox": [
+            496,
+            825,
+            890,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "19252",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/dafc29ad3590cc3cfb2a7d358b163bca09c7a2127ab71373a07b5c75f8c90e61.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            88,
+            364,
+            218
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/c5dc86c1a627513b4e84ccece2bd45bd5df1c9042492552d7dca3c95edc54efc.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            364,
+            89,
+            607,
+            218
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/1ed6f3578d2d9af0843954a429ad7af6532263ab6fad0360d4c62a4e0a890394.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            609,
+            90,
+            848,
+            218
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/24693d7cad2f36899b31c8f33ce9bcb2974c9d1836f75eaf87ea557cde227c54.jpg",
+        "image_caption": [
+            "Figure 4. Evaluation of online continual object detection in video streams with three metrics (FAP, CAP and F, Section 4.3) on OAK dataset (first row) and EgoObjects dataset (second row). The higher the bars are, the better. The x-axis denotes the percentage of video frames that are labeled in the video stream. It ranges from  $6.25\\%$  to  $100\\%$  (full supervision). The y-axis indicates the performance using different evaluation metrics. Ours (iCaRL(our impl.) w/ Efficient-CLS, red) consistently beats the comparative SOTA (iCaRL(our impl.), blue) in all evaluation metrics."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            219,
+            362,
+            347
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/11853cb1106fcd715fd04101d300b2bdf81ad71ec85abd0a3a6aa42ea914aa0d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            362,
+            219,
+            604,
+            347
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/f73a9a17b60c67d99d85868822c5329faa2468187a6784f269ed11e0fc04b41f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            604,
+            219,
+            846,
+            347
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/3389289ced2f3b4692e364719e4348a28456981109f5af49c9f99a375f80b0f6.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">EMA</td><td rowspan=\"2\">PL</td><td colspan=\"3\">50%</td><td colspan=\"3\">25%</td><td colspan=\"3\">12.5%</td><td colspan=\"3\">6.25%</td></tr><tr><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td></tr><tr><td>X</td><td>X</td><td>34.68</td><td>25.78</td><td>-4.15</td><td>33.70</td><td>24.57</td><td>-4.30</td><td>28.76</td><td>19.80</td><td>-5.48</td><td>23.04</td><td>17.75</td><td>-3.31</td></tr><tr><td>✓</td><td>X</td><td>35.74</td><td>25.77</td><td>-4.82</td><td>34.79</td><td>25.62</td><td>-4.35</td><td>31.72</td><td>21.16</td><td>-7.24</td><td>27.84</td><td>20.03</td><td>-3.96</td></tr><tr><td>X</td><td>✓</td><td>35.61</td><td>25.56</td><td>-3.76</td><td>34.95</td><td>25.65</td><td>-3.65</td><td>31.60</td><td>22.44</td><td>-4.83</td><td>26.39</td><td>19.50</td><td>-1.99</td></tr><tr><td>✓</td><td>✓</td><td>38.61</td><td>26.90</td><td>-7.29</td><td>38.36</td><td>26.64</td><td>-8.20</td><td>33.92</td><td>23.04</td><td>-7.71</td><td>29.72</td><td>20.31</td><td>-5.36</td></tr></table>",
+        "bbox": [
+            91,
+            439,
+            874,
+            518
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Effectiveness of Exponential Moving Average (EMA) and Pseudo-labeling (PL) for iCaRL (our impl.) on OAK dataset at annotation cost  $50\\%$ ,  $25\\%$ ,  $12.5\\%$  and  $6.25\\%$ . The best results are bold-faced.",
+        "bbox": [
+            75,
+            522,
+            892,
+            551
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "synaptic weight transfer from fast to slow learner via exponential moving average (EMA); and the other is reciprocal replay from slow learner to fast learner with pseudolabeling (PL). Here we studied their effects individually.",
+        "bbox": [
+            75,
+            569,
+            470,
+            631
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Ablation: Exponential Moving Average (EMA). We removed EMA by setting the  $\\alpha$  in Equation 3 to 1, where the model weights of the fast learner and slow learner are now shared throughout the learning process. Note that in the fully supervised protocol, the pseudo-labeling is turned off and the Efficient-CLS equals to the EMA alone. From Table 1 at  $100\\%$  annotation cost, we observed that removing EMA leads to significant performance drops ranging from  $2\\%$  to  $8\\%$ , for all the comparative baselines on both OAK and EgoObjects. This shows that the slow learner can effectively consolidate the knowledge from the fast learner, and constructively alleviate catastrophic forgetting by synapses consolidation over time. Similar observations were made in Table 3 in the sparse annotation protocol (compare Row 2 vs. Row 1, Row 4 vs. Row 3). It is worth noting that, the performance difference between naive model (Row 1) and its variant with EMA (Row 2) is slightly larger in lower su",
+        "bbox": [
+            75,
+            643,
+            472,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "pervision (i.e.  $12.5\\%$ ,  $6.25\\%$ ) than higher supervision (i.e.  $50\\%$ ,  $25\\%$ ). One possible reason is that, compared with higher supervision, the fast learner suffers more forgetting in lower supervision; hence, the effect of removing EMA becomes stronger in lower supervision, again highlighting the importance of EMA.",
+        "bbox": [
+            496,
+            569,
+            892,
+            660
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Ablation: Pseudo-labeling (PL). We ablate our model by removing the PL of the slow learner across varying annotation costs and reported the results in Table 3. The removal of PL (Row 2) leads to a performance drop of around  $2\\%$  in FAP,  $1\\%$  in CAP and  $0.5 - 4\\%$  in F, compared with our full Efficient-CLS (Row 4). It implies that the slow learner captures useful semantic information from unlabeled video frames and these predicted pseudo labels are helpful in training the fast learner.",
+        "bbox": [
+            496,
+            672,
+            893,
+            808
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To investigate whether pseudo labels predicted by the fast learner itself could help stream learning, we conducted another ablation experiment where we performed PL without EMA (Row 3). Compared with the naive model (Row 1), we observed a performance increase from  $28.76\\%$  to  $31.60\\%$  in FAP and  $19.80\\%$  to  $22.44\\%$  in CAP. It indicates",
+        "bbox": [
+            496,
+            810,
+            893,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "19253",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "that, due to the temporal correlation in video stream, pseudo labels predicted by the fast learner can serve as an informative supervision for the training of the fast learner itself.",
+        "bbox": [
+            75,
+            90,
+            468,
+            136
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "However, replaying the self-predicted pseudo labels on the fast learner fails to prevent forgetting, as indicated by the drop from  $-5.48\\%$  to  $-4.83\\%$  in F. It is possible that the pseudo labels generated by the fast learner only bias towards the classes which have already been learnt very well and fail to reinforce the fast learner to improve on the poorly-learnt classes. Different from the fast learner, the slow learner integrates semantic information over time. The predicted pseudo labels carry more semantic information, which is useful for fast learner to capture more generic object representations during pseudo label replays. Again, this emphasizes that the reciprocal replay from the slow learner to the fast learner is critical for memory reinstatement, which has been missing in the computational modeling literature of CLS.",
+        "bbox": [
+            75,
+            137,
+            470,
+            363
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            76,
+            378,
+            194,
+            393
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "To imitate what humans see and learn in the real world, we introduce a more realistic and challenging problem of label-efficient online continual object detection (LEOCOD) in video streams. Addressing this problem would greatly benefit real-world applications by reducing model retraining time and data labeling costs. Inspired by the Complementary Learning Systems (CLS) theory, we propose a plug-and-play module, namely Efficient-CLS, that can be easily integrated into and improve existing continual leaning algorithms. We evaluate Efficient-CLS on two challenging real-world video datasets, where our method achieves the state-of-the-art performance in preventing catastrophic forgetting, all while requiring minimal annotation effort.",
+        "bbox": [
+            75,
+            404,
+            470,
+            602
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements This research is funded by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008 & NRF-NRFF15-2023-0001, AI Singapore Programme (AISG Award No: AISGC-2019-001-2A & AISG2-RP-2021-025), Mike Zheng Shou's Start-Up Grant from NUS, Mengmi Zhang's Startup Grant from Agency for Science, Technology, and Research (A*STAR), and Early Career Investigatorship from Center for Frontier AI Research (CFAR), A*STAR. The computational work for this article was partially performed on resources of the National Supercomputing Centre, Singapore.",
+        "bbox": [
+            75,
+            622,
+            468,
+            789
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            804,
+            173,
+            819
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Egoobjects dataset. https://ai.facebook.com/datasets/egoobjects-dataset/, May 2022. 2, 5",
+            "[2] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered"
+        ],
+        "bbox": [
+            84,
+            828,
+            468,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "retrieval. Advances in neural information processing systems, 32, 2019. 1, 3",
+            "[3] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Advances in neural information processing systems, 32, 2019. 1, 2, 3",
+            "[4] Elahe Arani, Fahad Sarfraz, and Bahram Zonooz. Learning fast, learning slow: A general continual learning method based on complementary learning system. arXiv preprint arXiv:2201.12604, 2022. 2, 3, 5",
+            "[5] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 3",
+            "[6] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 2, 3, 5",
+            "[7] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuytelaars, Joelle Pineau, and Eugene Belilovsky. New insights on reducing abrupt representation change in online continual learning. arXiv preprint arXiv:2203.03798, 2022. 1, 2",
+            "[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 2, 3, 5",
+            "[9] Hung-Jen Chen, An-Chieh Cheng, Da-Cheng Juan, Wei Wei, and Min Sun. Mitigating forgetting in online continual learning via instance-aware parameterization. Advances in Neural Information Processing Systems, 33:17466-17477, 2020. 2",
+            "[10] Mark Everingham, SM Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111(1):98-136, 2015. 6",
+            "[11] Enrico Fini, Stéphane Lathuiliere, Enver Sangineto, Moin Nabi, and Elisa Ricci. Online continual learning under extreme memory constraints. In European Conference on Computer Vision, pages 720-735. Springer, 2020. 1",
+            "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6",
+            "[13] Jisoo Jeong, Seungeui Lee, Jeesoo Kim, and Nojun Kwak. Consistency-based semi-supervised learning for object detection. Advances in neural information processing systems, 32, 2019. 3",
+            "[14] Daoyun Ji and Matthew A Wilson. Coordinated memory replay in the visual cortex and hippocampus during sleep. Nature neuroscience, 10(1):100-107, 2007. 2",
+            "[15] Nitin Kamra, Umang Gupta, and Yan Liu. Deep generative dual memory network for continual learning. arXiv preprint arXiv:1710.10368, 2017. 3",
+            "[16] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 5"
+        ],
+        "bbox": [
+            503,
+            92,
+            893,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "19254",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[17] Joseph Kj, Jathushan Rajasegaran, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Incremental object detection via meta-learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 5",
+            "[18] Dharshan Kumaran, Demis Hassabis, and James L McClelland. What learning systems do intelligent agents need? complementary learning systems theory updated. Trends in cognitive sciences, 20(7):512-534, 2016. 1",
+            "[19] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn and think like people. Behavioral and brain sciences, 40, 2017. 1",
+            "[20] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. arXiv preprint arXiv:2102.09480, 2021. 3",
+            "[21] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2",
+            "[22] Lorenzo Pellegrini, Chenchen Zhu, Fanyi Xiao, Zhicheng Yan, Antonio Carta, Matthias De Lange, Vincenzo Lomonaco, Roshan Sumbaly, Pau Rodriguez, and David Vazquez. 3rd continual learning workshop challenge on egocentric category and instance level object understanding. arXiv preprint arXiv:2212.06833, 2022. 5",
+            "[23] Quang Pham, Chenghao Liu, and Steven Hoi. Dualnet: Continual learning, fast and slow. Advances in Neural Information Processing Systems, 34, 2021. 3",
+            "[24] Quang Pham, Chenghao Liu, Doyen Sahoo, and HOI Steven. Contextual transformation networks for online continual learning. In International Conference on Learning Representations, 2020. 2, 3",
+            "[25] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In European conference on computer vision, pages 524-540. Springer, 2020. 2, 3, 5",
+            "[26] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 2, 3, 5",
+            "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 4, 6",
+            "[28] Mohammad Rostami, Soheil Kolouri, and Praveen K Pilly. Complementary learning for overcoming catastrophic forgetting using experience replay. arXiv preprint arXiv:1903.04566, 2019. 3",
+            "[29] Dongsub Shim, Zheda Mai, Jihwan Jeong, Scott Sanner, Hyunwoo Kim, and Jongseong Jang. Online class incremental continual learning with adversarial shapley value. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 9630-9638, 2021. 2, 3",
+            "[30] Krishna Kumar Singh, Kayvon Fatahalian, and Alexei A Efros. Krishnacam: Using a longitudinal, single-person,"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "egocentric dataset for scene understanding tasks. In 2016 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1-9. IEEE, 2016. 5",
+            "[31] James Smith, Jonathan Balloch, Yen-Chang Hsu, and Zsolt Kira. Memory-efficient semi-supervised continual learning: The world is its own replay buffer. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 3",
+            "[32] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 3",
+            "[33] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3",
+            "[34] Jianren Wang, Xin Wang, Yue Shang-Guan, and Abhinav Gupta. Wanderlust: Online continual object detection in the real world. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10829-10838, 2021. 1, 2, 3, 5, 6",
+            "[35] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020. 1",
+            "[36] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14421-14430, 2022. 3",
+            "[37] Xiangli Yang, Zixing Song, Irwin King, and Zenglin Xu. A survey on deep semi-supervised learning. IEEE Transactions on Knowledge and Data Engineering, 2022. 3",
+            "[38] Barret Zoph, Ekin D Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V Le. Learning data augmentation strategies for object detection. In European conference on computer vision, pages 566-583. Springer, 2020. 4"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            614
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "19255",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2023/Label-Efficient Online Continual Object Detection in Streaming Video/bb408240-93f4-4420-9e47-a5facd88851a_model.json

@@ -0,0 +1,1936 @@
+[
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.107,
+                0.003,
+                0.182,
+                0.043
+            ],
+            "angle": 0,
+            "content": "CVF"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.239,
+                0.001,
+                0.808,
+                0.047
+            ],
+            "angle": 0,
+            "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.131,
+                0.131,
+                0.842,
+                0.154
+            ],
+            "angle": 0,
+            "content": "Label-Efficient Online Continual Object Detection in Streaming Video"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.093,
+                0.18,
+                0.884,
+                0.201
+            ],
+            "angle": 0,
+            "content": "Jay Zhangjie Wu\\(^{1}\\) David Junhao Zhang\\(^{1}\\) Wynne Hsu\\(^{2}\\) Mengmi Zhang\\(^{3,4}\\) Mike Zheng Shou\\(^{1*}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.211,
+                0.856,
+                0.266
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)Show Lab, \\(^{2}\\)National University of Singapore  \n\\(^{3}\\)School of Computer Science and Engineering, Nanyang Technological University, Singapore  \n\\(^{4}\\)CFAR and I2R, Agency for Science, Technology and Research, Singapore"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.235,
+                0.299,
+                0.314,
+                0.317
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.075,
+                0.333,
+                0.474,
+                0.651
+            ],
+            "angle": 0,
+            "content": "Humans can watch a continuous video stream and effortlessly perform continual acquisition and transfer of new knowledge with minimal supervision yet retaining previously learnt experiences. In contrast, existing continual learning (CL) methods require fully annotated labels to effectively learn from individual frames in a video stream. Here, we examine a more realistic and challenging problem—Label-Efficient Online Continual Object Detection (LEOCOD) in streaming video. We propose a plug-and-play module, Efficient-CLS, that can be easily inserted into and consistently improve existing CL algorithms for object detection in video streams with reduced data annotation costs and model retraining time. We show that our method has achieved significant improvement with minimal forgetting across all supervision levels on two challenging CL benchmarks for streaming real-world videos. Remarkably, with only \\(25\\%\\) annotated video frames, our proposed method still outperforms the state-of-the-art CL models trained with \\(100\\%\\) annotations on all video frames. The data and source code will be publicly available at https://github.com/showlab/Efficient-CLS."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.682,
+                0.21,
+                0.698
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.708,
+                0.471,
+                0.875
+            ],
+            "angle": 0,
+            "content": "Humans have the ability to continuously learn from an ever-changing environment, while retaining previously learnt experiences. In contrast to human learning, prior works [3, 2, 11, 34, 7] show that deep neural networks are prone to catastrophic forgetting. To address the forgetting problem, existing works in continual learning (CL) primarily focus on class-incremental image classification or object detection. Their experiment settings are often idealistic and simplified, where i.i.d. static images are usually grouped by class and incrementally presented to computational models in sequence. To learn a particular task containing specific"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.301,
+                0.892,
+                0.361
+            ],
+            "angle": 0,
+            "content": "classes, an agent can go through all the data of current task over multiple epochs. After that, the learned classes in current task become unavailable, i.e., no overlaps between the sets of learned classes and unseen classes."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.365,
+                0.893,
+                0.547
+            ],
+            "angle": 0,
+            "content": "However, these experiment designs deviate from the online continual learning (OCL) setting in the real world, where an agent learns from temporally correlated non-i.i.d. video streams in one single pass. Given context regularities in natural environments, an agent is likely to encounter cases when objects of previously learnt classes co-occur with unknown objects from unseen classes, e.g., a computer mouse and a computer monitor often co-occur. Taking these considerations, [34] introduces OCL on object detection in real-world video streams. They evaluate existing CL approaches on this setting and report a huge performance gap compared with offline training."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.55,
+                0.895,
+                0.807
+            ],
+            "angle": 0,
+            "content": "Based on the setting in [34], we take a significant step further and introduce a novel problem setting called Label-Efficient Online Continual Object Detection (LEOCOD), which highlights two unique challenges. First, the setting in [34] is such that the CL algorithms are trained with every mini-batch over multiple passes. We tighten the training recipe in LEOCOD to strictly online, where data is allowed to have one single pass and models are trained on the entire video dataset for only one epoch. Second, existing CL models require fully supervised training where box-level ground truth labels of every object on every video frame have to be obtained from human annotators. Unlike static images, acquiring human annotations for object detection on videos can be expensive and daunting. Thus, in LEOCOD, the video frames per mini-batch are sparsely annotated to alleviate the burdens of extensive human labeling, making LEOCOD one step closer to real world application."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.81,
+                0.895,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Cognitive science works [35, 19] show that humans are efficient at continuously learning from very few annotated data samples. We get inspirations from the theory of Complementary Learning Systems (CLS) in human brains [18], and propose a plug-and-play module for the LEOCOD task, dubbed as Efficient-CLS. In Efficient-CLS, we introduce"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.096,
+                0.887,
+                0.227,
+                0.9
+            ],
+            "angle": 0,
+            "content": "*Corresponding Author."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "19246"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.099,
+                0.094,
+                0.531,
+                0.307
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.54,
+                0.094,
+                0.871,
+                0.306
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.312,
+                0.896,
+                0.41
+            ],
+            "angle": 0,
+            "content": "Figure 1. (a) Problem introduction: An agent continuously learns from a never-ending online video stream over time. In each training step, out of a mini-batch containing 16 consecutive video frames, only a fixed proportion of frames are labeled (green boundary), while the rest of the frames are unlabeled (orange boundary). Following [34], the video frame after every training mini-batch (transparent) is held out for testing. After every 100 training steps, the agent is evaluated on all the video frames from the test set for object detection. (b) Key results: Our proposed method (red) consistently outperforms the best competitive baseline (blue) by a margin of \\(5\\%\\). Remarkably, our model, trained at \\(25\\%\\) annotation cost, surpasses the best baseline trained at \\(100\\%\\) (grey line). The orange cross denotes the performance of the state-of-the-art model, which is \\(15\\%\\) lower than our method."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.424,
+                0.471,
+                0.635
+            ],
+            "angle": 0,
+            "content": "two feed-forward neural networks as slow and fast learners. In the fast learner, memory is rapidly adapted to the current task. The weights of the slow learner change a little on each reinstatement, and are maintained by taking the exponential moving average (EMA) of the fast learner's weights over time. Though a few continual learning models in previous works [4, 24] also use a similar source of inspiration, they miss the effect of reciprocal connections from slow learners to fast learners, which we intend to address. Inspired by the bidirectional interaction in CLS [14], we reactivate the weights of the slow learners to predict meaningful pseudo labels from the unlabeled video frames and use these pseudo labels to guide the training of the fast learner, closing the loop between the two systems."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.635,
+                0.47,
+                0.756
+            ],
+            "angle": 0,
+            "content": "We demonstrate the versatility and effectiveness of our Efficient-CLS on two standard real-world video datasets, OAK [34] and EgoObjects [1]. Our proposed method can be easily integrated into existing CL models and consistently improve their performance by a large margin in LEOCOD. It is worth noting that, with only \\(25\\%\\) labeled data, our method surpasses the comparative baselines trained with full supervision (Figure 1(b))."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.097,
+                0.756,
+                0.468,
+                0.771
+            ],
+            "angle": 0,
+            "content": "To summarize, we make the following key contributions:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.096,
+                0.776,
+                0.468,
+                0.852
+            ],
+            "angle": 0,
+            "content": "- We introduce a new, challenging and important problem of label-efficient online continual object detection (LEOCOD) in video streams. Solving this problem would greatly benefit real-world applications in reducing annotation cost and model retraining time."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.096,
+                0.856,
+                0.468,
+                0.902
+            ],
+            "angle": 0,
+            "content": "- We propose Efficient-CLS, a plug-and-play module inspired from the Complementary Learning Systems (CLS) theory, which can be integrated into existing CL"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.096,
+                0.776,
+                0.468,
+                0.902
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.531,
+                0.424,
+                0.892,
+                0.455
+            ],
+            "angle": 0,
+            "content": "models and learn efficiently and effectively with less supervision and minimal forgetting."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.518,
+                0.459,
+                0.894,
+                0.505
+            ],
+            "angle": 0,
+            "content": "- We benchmark existing CL methods on the task of LEOCOD and demonstrate the state-of-the-art performance of our method through extensive experiments."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.52,
+                0.642,
+                0.535
+            ],
+            "angle": 0,
+            "content": "2. Related Work"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.545,
+                0.691,
+                0.561
+            ],
+            "angle": 0,
+            "content": "2.1. Continual Learning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.569,
+                0.892,
+                0.78
+            ],
+            "angle": 0,
+            "content": "To alleviate catastrophic forgetting, many continual learning methods exploit an external buffer where a limited number of old samples are stored and used for replay when adapting to a new task. iCaRL [26] stores the representative exemplars in past tasks for knowledge distillation and prototype rehearsal. Gradient Episodic Memory (GEM) [21] formulates optimization constraints on the exemplars in memory. Averaged GEM (A-GEM) [8] is an improved version of GEM that achieves faster training and less memory consumption. GDumb [25] greedily stores samples in memory as they come and trains a model using samples only in the memory. Dark Experience Replay++ (DER++) [6] combines replay with knowledge distillation and regularization, and samples logits along the entire optimization trajectory."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.78,
+                0.892,
+                0.902
+            ],
+            "angle": 0,
+            "content": "In contrast to classical continual learning where data are separated by task boundaries and models are trained with multiple iterations in every task, we examine a more realistic and challenging problem where data are provided in tiny batches and models are trained on these batches only once. Recently, online continual learning (OCL) has gained increasing interests in computer vision [3, 7, 29, 9, 34]. Many OCL methods rely on representative memory replays to pre"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "19247"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.092,
+                0.471,
+                0.211
+            ],
+            "angle": 0,
+            "content": "vent forgetting. [3] utilizes gradients of network parameters to select replay samples of maximum diversity. Subsequent works [2, 29] propose to use losses and scoring functions as criteria for selecting the most representative samples for replay. However, these approaches tackle image classification problem in an artificial setting, where new classes appear in a specific order. Their performance in real-world vision tasks remains unclear."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.213,
+                0.471,
+                0.409
+            ],
+            "angle": 0,
+            "content": "Lately, [34] benchmarks CL methods in the real-world online setting with full supervision. As the video streams arrive endlessly in a real-time manner, assigning annotations to all the video frames for training computational models is laborious and time-consuming. It becomes even more daunting in object detection tasks where class labels and bounding boxes of all objects on a video frame have to be provided. Reducing burdensome costs of labeling remains an under-explored and challenging problem in online continual object detection. We propose a self-sustaining Efficient-CLS, which is capable of exploiting the unlabeled video frames by pseudo-labeling when the number of labeled frames is limited."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.42,
+                0.38,
+                0.436
+            ],
+            "angle": 0,
+            "content": "2.2. Complementary Learning Systems"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.444,
+                0.47,
+                0.807
+            ],
+            "angle": 0,
+            "content": "The essence of fast and slow learning in Complementary Learning Systems (CLS) has benefited several continual learning algorithms in image recognition [24, 23, 28, 4, 15]. However, these methods either require the task boundaries, which are not applicable in our online video setting, or they require to train fast and slow learning systems with replay samples from the same replay buffer, which could easily lead to overfitting problem when the replay buffer has limited capacity. To eliminate overfitting problem, [28] and [15] utilize generative replay models to couple sequential tasks in a latent embedding space. While generative approaches have succeeded in artificial and simple datasets, they often fail in complex vision tasks, e.g., object detection. DualNet [23] also employs a slow-fast learning architecture. However, its update process for the slow learner is more computationally demanding because it leans on both self-supervised and supervised learning objectives. The self-supervised phase demands additional training iterations on extensive batches to yield satisfactory results, thereby delaying the training of the fast learner. In contrast, our method updates the slow learner in real-time using the fast learner's weights, obviating the need for distinct slow learner training. Our approach is optimized for online CL scenarios where computational resources are constrained."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.817,
+                0.318,
+                0.833
+            ],
+            "angle": 0,
+            "content": "2.3. Semi-Supervised Learning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.84,
+                0.469,
+                0.902
+            ],
+            "angle": 0,
+            "content": "The goal of semi-supervised learning (SSL) [33, 5, 37, 36] is to reduce the demand of labeled data and harness unlabeled data for performance improvement. [31] first introduce SSL to the context of CL, and propose a strategy"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.092,
+                0.892,
+                0.364
+            ],
+            "angle": 0,
+            "content": "that combines pseudo-labeling, consistency regularization, out-of-distribution detection, and knowledge distillation to solve the problem of class-incremental image classification. However, to distill knowledge from previous tasks, their model relies heavily on the task boundary that identifies the change of training classes, which is not available in our LEOCOD setting. To reduce annotation costs in object detection, several methods [13, 32, 20] capitalize the teacher-student networks. In general, a teacher model predicts pseudo labels or enforces a consistency loss to guide the student networks. However, these previous works on semi-supervised object detection only consider the offline setting on static image datasets, while none of them has been extended to online continual learning on dynamic video streams. We discovered new insights that pseudo-labeling from slow learner to fast learner can not only reduce annotation overhead, but also alleviate forgetting, which is critical to the design of efficient continual learners in real world."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.378,
+                0.591,
+                0.394
+            ],
+            "angle": 0,
+            "content": "3. Method"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.404,
+                0.661,
+                0.42
+            ],
+            "angle": 0,
+            "content": "3.1. Problem Setting"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.428,
+                0.892,
+                0.548
+            ],
+            "angle": 0,
+            "content": "We advance the online continual object detection in [34] to a label-efficient and computationally-efficient setting—Label-Efficient Online Continual Object Detection (LEOCOD). In contrast to the setting in [34] where video frames are extensively annotated and trained with multiple epochs, an agent in LEOCOD continuously learns from a sparsely annotated video stream in a single pass over time (see Figure 1(a))."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.549,
+                0.892,
+                0.701
+            ],
+            "angle": 0,
+            "content": "Formally, we consider the online continual object detection on a continuum of video streams \\(\\mathcal{D} = \\{D_1,\\dots ,D_T\\}\\) where at time step \\(t\\) , a learning agent receives a mini-batch of continuous video frames \\(D_{t}\\) from current environment for online training (one single pass). To perform label-efficient object detection, within the batch \\(D_{t}\\) , only a subset of video frames \\(D_{t}^{s} = (X_{t}^{s},Y_{t}^{s})\\) are labeled, while the remaining video frames \\(D_{t}^{u} = (X_{t}^{u})\\) are unlabeled. For each labeled data sample, its annotation contains the bounding box locations and their corresponding class labels."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.711,
+                0.871,
+                0.727
+            ],
+            "angle": 0,
+            "content": "3.2. Efficient Complementary Learning Systems"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.735,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "We propose a plug-and-play module dubbed as EfficientCLS. Specifically, it consists of two feed-forward networks: (i) the fast learner is designed to quickly encode new knowledge from current data stream and then consolidate it to the slow learner; and (ii) the slow learner accumulates the acquired knowledge from fast learner over time and guides the fast learner with meaningful pseudo labels, when full supervision is not available. Same as [26, 8, 25, 6, 34], we maintain an external episodic memory, as a replay buffer, to store exemplars that can be retrieved for replays alongside ongoing video stream. As the fast and slow learners are"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "19248"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.102,
+                0.093,
+                0.873,
+                0.33
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.335,
+                0.895,
+                0.42
+            ],
+            "angle": 0,
+            "content": "Figure 2. The framework of Efficient-CLS. At each learning step, the system receives a batch of temporally continuous data \\( D_{t} \\), including labeled (green) and unlabeled (orange) frames. The fast learner trains the labeled frames alongside a small subset of labeled exemplars retrieved from episodic memory with the supervised loss \\( \\mathcal{L}_{sup} \\). Meanwhile, the fast learner leverages the pseudo labels generated by the slow learner to optimize a pseudo loss \\( \\mathcal{L}_{pseudo} \\). To reinstate memory of the slow learner, the weights of the slow learner are updated by taking the Exponential Moving Average (EMA) of the fast learner's weights. The fast and slow learners are complementary to each other, forming a positive feedback loop."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.433,
+                0.471,
+                0.48
+            ],
+            "angle": 0,
+            "content": "model agnostic, our Efficient-CLS can be easily integrated into existing CL models, which leads to less supervision and minimal forgetting."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.494,
+                0.47,
+                0.68
+            ],
+            "angle": 0,
+            "content": "Learning with Labeled Frames. The fast learner and slow learner use the same standard Faster-RCNN [27] detector \\( f \\). Despite the same architecture, the weights of the fast and slow learners are not shared. We use \\( \\theta_{F} \\) and \\( \\theta_{S} \\) to denote the network parameters for fast and slow learners respectively. As shown in Figure 2, at each training step \\( t \\), we use the labeled video frames \\( D_{t}^{s} = (X_{t}^{s}, Y_{t}^{s}) \\) to optimize the fast learner \\( \\theta_{F} \\) with the standard supervised loss \\( \\mathcal{L}_{sup} \\) in Faster-RCNN [27]. It consists of four losses: Region Proposal Network (RPN) classification loss \\( \\mathcal{L}_{cls}^{rpn} \\), RPN regression loss \\( \\mathcal{L}_{reg}^{rpn} \\), Region of Interest (ROI) classification loss \\( \\mathcal{L}_{cls}^{roi} \\), and ROI regression loss \\( \\mathcal{L}_{reg}^{roi} \\). We define \\( \\mathcal{L}_{sup} \\) as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.124,
+                0.688,
+                0.47,
+                0.714
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {s u p} = \\mathcal {L} _ {c l s} ^ {r p n} \\left(X _ {t} ^ {s}, Y _ {t} ^ {s}\\right) + \\mathcal {L} _ {r e g} ^ {r p n} \\left(X _ {t} ^ {s}, Y _ {t} ^ {s}\\right) + \\tag {1}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.211,
+                0.711,
+                0.422,
+                0.728
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {c l s} ^ {r o i} \\left(X _ {t} ^ {s}, Y _ {t} ^ {s}\\right) + \\mathcal {L} _ {r e g} ^ {r o i} \\left(X _ {t} ^ {s}, Y _ {t} ^ {s}\\right).\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.75,
+                0.471,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Learning with Unlabeled Frames. We introduce a pseudo-labeling paradigm to capitalize the information from unlabeled video frames \\( D_{t}^{u} = (X_{t}^{u}) \\) for training. In our early exploration, we intuitively use the fast learner for pseudo-labeling as it quickly adapts the knowledge of nearby frames. However, we observe that using the pseudo labels generated by the fast learner for self-replay exhibits biases towards recently seen objects, which is less effective in preventing forgetting. This has also been verified in our ablation study (Section 5.3). In contrast, the slow learner"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.433,
+                0.892,
+                0.524
+            ],
+            "angle": 0,
+            "content": "preserves the semantic knowledge over a longer time span which generates pseudo labels with fewer biases. This encourages the fast learner to capture more generic scene representations, hence, in turn, contributing to reinstatement of memory in the slow learner (Section 3.2), resulting in a positive feedback loop."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.524,
+                0.893,
+                0.734
+            ],
+            "angle": 0,
+            "content": "Given all these design considerations, the slow learner takes the unlabeled video frames \\( D_{t}^{u} \\) as inputs to estimate the possible objects of interest and their corresponding bounding box locations. For brevity, we refer these \"pseudo bounding boxes and their corresponding class labels\" as \"pseudo labels\" in the paper. To get rid of false positives, we apply a threshold \\( \\tau \\) to filter out bounding boxes with predicted low confidence scores. Moreover, there also exist repetitive boxes which negatively impact the quality of pseudo-labeling. To address this issue, we use the technique of class-wise non-maximum suppression (NMS) [27] to remove the overlapped boxes and get the high-quality pseudo labels. Formally, the procedure of pseudo label generation is summarized below:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.596,
+                0.742,
+                0.892,
+                0.76
+            ],
+            "angle": 0,
+            "content": "\\[\nY _ {t} ^ {u} = \\mathrm {N M S} ([ f (X _ {t} ^ {u}; \\theta_ {S}) ] _ {> \\tau}), \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.765,
+                0.891,
+                0.796
+            ],
+            "angle": 0,
+            "content": "where \\([\\cdot ]_{>\\tau}\\) denotes the bounding box selection with confidence score larger than \\(\\tau\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.796,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Given that the video streams are captured from the egocentric perspective in the real world, head and body motions may lead to undesired motion blur effects on some video frames. To enforce our module to learn invariant object representations from these video frames, same as the previous work [38], we apply data augmentation techniques on the pseudo-labeled frames, including 2D image"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "19249"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.092,
+                0.47,
+                0.213
+            ],
+            "angle": 0,
+            "content": "crops, rotations, and flipping. Note that different from image classification, the predicted bounding box locations also need to be updated accordingly after image augmentations. We denote these pseudo-labeled video frames and their re-adjusted pseudo labels after data augmentations as \\((\\tilde{X}_t^u,\\tilde{Y}_t^u)\\). We can then use these pseudo pairs \\((\\tilde{X}_t^u,\\tilde{Y}_t^u)\\) to train the fast learner by optimizing the pseudo loss \\(\\mathcal{L}_{pseudo}:= \\mathcal{L}_{cls}^{roi}(\\tilde{X}_t^u,\\tilde{Y}_t^u) + \\mathcal{L}_{req}^{roi}(\\tilde{X}_t^u,\\tilde{Y}_t^u)\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.213,
+                0.469,
+                0.259
+            ],
+            "angle": 0,
+            "content": "Overall, our Efficient-CLS is jointly trained with the following losses: \\(\\mathcal{L}_{total} = \\mathcal{L}_{sup} + \\lambda_{pseudo}\\mathcal{L}_{pseudo}\\), where \\(\\lambda_{pseudo}\\) is the weight of \\(\\mathcal{L}_{pseudo}\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.269,
+                0.469,
+                0.375
+            ],
+            "angle": 0,
+            "content": "Synapses Consolidation via EMA. To alleviate forgetting of obtained knowledge, we apply Exponential Moving Average (EMA) to gradually update the slow learner with the fast learner's weights. The evolving weight changes in the slow learner are functionally correlated with the memory consolidation mechanism in the hippocampus and the neocortex [4]. Formally, we define EMA process as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.19,
+                0.381,
+                0.469,
+                0.397
+            ],
+            "angle": 0,
+            "content": "\\[\n\\theta_ {S} = \\alpha \\theta_ {S} + (1 - \\alpha) \\theta_ {F}, \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.403,
+                0.469,
+                0.464
+            ],
+            "angle": 0,
+            "content": "where the \\(\\alpha \\in [0,1]\\) is EMA rate. According to the stability-plasticity dilemma, a smaller \\(\\alpha\\) means faster adaptation but less memorization. Empirically, we set \\(\\alpha = 0.99\\), which leads to best performance."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.475,
+                0.276,
+                0.492
+            ],
+            "angle": 0,
+            "content": "4. Experimental details"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.5,
+                0.18,
+                0.515
+            ],
+            "angle": 0,
+            "content": "4.1. Datasets"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.523,
+                0.469,
+                0.569
+            ],
+            "angle": 0,
+            "content": "We evaluate the continual learning methods on two realistic and challenging video datasets, OAK [34] and EgoObjects [1], for the task of online continual object detection."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.569,
+                0.469,
+                0.689
+            ],
+            "angle": 0,
+            "content": "OAK [34] is a large egocentric video stream dataset spanning nine months of a graduate student's life, consisting of 7.6 million frames of 460 video clips with a total length of 70.2 hours [30]. The dataset contains 103 object categories. We follow [34] in the ordering of training and testing data splits. One frame every 16 consecutive video frames lasting for 30 seconds is held out to construct a test set and the remaining frames are used for training."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.689,
+                0.469,
+                0.81
+            ],
+            "angle": 0,
+            "content": "EgoObjects [1] is one of the largest object-centric datasets focusing on object detection task. The dataset contains around 100k video frames with \\(\\sim 250\\mathrm{k}\\) annotations, which correspond to 277 categories and 1110 main objects [22]. Additionally, the dataset follows a long-tailed distribution which makes the task more challenging and real-world oriented. We adopt the same rule as OAK dataset to split the training and testing set."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.818,
+                0.185,
+                0.832
+            ],
+            "angle": 0,
+            "content": "4.2. Baselines"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.84,
+                0.469,
+                0.901
+            ],
+            "angle": 0,
+            "content": "We compare our model against the following baselines: 1) Vanilla training: Incremental Training is a naive baseline trained sequentially over the entire video stream without any measures to avoid catastrophic forgetting; Offline"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.092,
+                0.892,
+                0.152
+            ],
+            "angle": 0,
+            "content": "Training is an upper bound that can access all the data throughout training; 2) Continual learning algorithms: EWC [16], iCaRL [26], A-GEM [8], GDumb [25], DER++ [6] and iOD [17]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.153,
+                0.892,
+                0.349
+            ],
+            "angle": 0,
+            "content": "The iCaRL model implemented by [34] stands as the SOTA method in online continual object detection. We reproduce their results using the released code<sup>1</sup>. When calculating RPN and ROI losses for replay samples, their iCaRL model neglects the losses of background proposals and penalizes the foreground losses according to the proportion of the current samples and replay samples. We empirically found that this trick hinders the model from effective episodic replay, thus resulting in severe forgetting. Therefore, we re-implement the iCaRL by discarding the reweighting trick and reverting back to the standard RPN and ROI losses. We name these two different implementations as iCaRL(Wang et al.) and iCaRL(our impl.), respectively."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.358,
+                0.621,
+                0.372
+            ],
+            "angle": 0,
+            "content": "4.3. Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.381,
+                0.892,
+                0.622
+            ],
+            "angle": 0,
+            "content": "Protocols. First, we define the annotation cost as the proportion of number of labeled frames versus the total 16 frames within a mini-batch \\( D_{t} \\). For example, if 2 out of 16 consecutive frames within \\( D_{t} \\) get labeled, the annotation cost is \\( 2 / 16 = 12.5\\% \\). The frames to be labeled are randomly selected within each mini-batch \\( D_{t} \\). Considering that different choices of labeled frames might influence the model performance, for fair comparisons between models, we fix the choice of randomly selected labeled frames and use the same labeled and unlabeled frames for training all models. Based on the various annotation costs, we introduce two training protocols: fully supervised protocol (100% annotation cost) and sparse annotation protocol (where the annotation cost is less than 100%). In sparse annotation protocol, we further split the training experiments based on 50%/25%/12.5%/6.25% annotation costs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.634,
+                0.892,
+                0.709
+            ],
+            "angle": 0,
+            "content": "Metrics. We evaluate the models with three standard metrics: continual average precision (CAP), final average precision (FAP) and forgetfulness (F) [34]. We employ an Average Precision (AP) metric at an Intersection over Union (IoU) threshold of 0.5, commonly referred to as AP50."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.71,
+                0.892,
+                0.739
+            ],
+            "angle": 0,
+            "content": "CAP shows the average performance over the entire video stream. That is,"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.537,
+                0.75,
+                0.892,
+                0.791
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathrm {C A P} = \\frac {1}{N} \\sum_ {i = 0} ^ {N} \\mathrm {C A P} _ {t _ {i}} = \\frac {1}{N C} \\sum_ {i = 0} ^ {N} \\sum_ {c = 0} ^ {C} \\mathrm {C A P} _ {t _ {i}} ^ {c}, \\tag {4}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.802,
+                0.891,
+                0.833
+            ],
+            "angle": 0,
+            "content": "where \\(\\mathrm{CAP}_{t_i}^c\\) is the AP of class \\(c\\) on test set at the \\(i\\)-th evaluation step, \\(N\\) is the total number of evaluation steps."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.833,
+                0.892,
+                0.878
+            ],
+            "angle": 0,
+            "content": "FAP is the final performance of a model after seeing the entire video. That is, \\(\\mathrm{FAP} = \\mathrm{CAP}_{t_N}\\), where \\(t_N\\) denotes the last evaluation step."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.518,
+                0.887,
+                0.729,
+                0.9
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)https://github.com/oakdata/benchmark"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "19250"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.086,
+                0.088,
+                0.887,
+                0.344
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\"></td><td rowspan=\"2\">Annotation Cost</td><td colspan=\"3\">OAK</td><td colspan=\"3\">EgoObjects</td></tr><tr><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td></tr><tr><td>Incremental</td><td>100%</td><td>8.38</td><td>7.72</td><td>0.03</td><td>10.21</td><td>3.55</td><td>1.48</td></tr><tr><td>Offline Training</td><td>100%</td><td>48.28</td><td>35.23</td><td>-</td><td>86.18</td><td>59.81</td><td>-</td></tr><tr><td>EWC</td><td>100%</td><td>7.73</td><td>7.02</td><td>-0.12</td><td>5.15</td><td>1.60</td><td>0.57</td></tr><tr><td>iOD</td><td>100%</td><td>7.92</td><td>7.14</td><td>0.98</td><td>8.80</td><td>2.64</td><td>0.00</td></tr><tr><td>iCaRL(Wang et al.)</td><td>100%</td><td>22.89</td><td>16.60</td><td>-2.95</td><td>37.61</td><td>21.71</td><td>2.79</td></tr><tr><td>iCaRL(our impl.)</td><td>100%</td><td>36.14</td><td>26.26</td><td>-4.89</td><td>60.80</td><td>36.41</td><td>-0.60</td></tr><tr><td rowspan=\"2\">w/ Efficient-CLS</td><td>25%</td><td>38.36(+2.22)</td><td>26.64(+0.38)</td><td>-8.20(-3.31)</td><td>61.26(+0.46)</td><td>39.58(+3.17)</td><td>-3.48(-2.88)</td></tr><tr><td>100%</td><td>40.24(+4.10)</td><td>28.18(+1.92)</td><td>-8.10(-3.21)</td><td>67.05(+6.25)</td><td>40.36(+3.95)</td><td>-3.67(-3.07)</td></tr><tr><td>A-GEM</td><td>100%</td><td>36.94</td><td>26.19</td><td>-5.54</td><td>58.79</td><td>35.88</td><td>-8.38</td></tr><tr><td rowspan=\"2\">w/ Efficient-CLS</td><td>25%</td><td>37.06(+0.12)</td><td>26.36(+0.17)</td><td>-7.76(-2.22)</td><td>63.06(+4.27)</td><td>39.46(+3.58)</td><td>-7.49(+0.89)</td></tr><tr><td>100%</td><td>39.87(+2.93)</td><td>27.97(+1.78)</td><td>-7.17(-1.63)</td><td>66.94(+8.15)</td><td>39.57(+3.69)</td><td>-11.68(-3.30)</td></tr><tr><td>GDumb</td><td>100%</td><td>35.27</td><td>25.29</td><td>-6.59</td><td>58.85</td><td>36.38</td><td>-5.21</td></tr><tr><td rowspan=\"2\">w/ Efficient-CLS</td><td>25%</td><td>37.67(+2.40)</td><td>25.59(+0.30)</td><td>-9.30(-2.71)</td><td>62.70(+3.85)</td><td>38.78(+2.40)</td><td>-8.86(-3.65)</td></tr><tr><td>100%</td><td>38.61(+3.34)</td><td>26.04(+0.75)</td><td>-9.14(-2.55)</td><td>63.55(+4.70)</td><td>38.98(+2.60)</td><td>-7.50(-2.29)</td></tr><tr><td>DER++</td><td>100%</td><td>37.79</td><td>25.24</td><td>-2.87</td><td>55.82</td><td>30.84</td><td>-6.08</td></tr><tr><td rowspan=\"2\">w/ Efficient-CLS</td><td>25%</td><td>37.93(+0.14)</td><td>25.64(+0.4)</td><td>-8.90(-6.03)</td><td>59.70(+3.88)</td><td>34.15(+3.31)</td><td>-11.21(-5.13)</td></tr><tr><td>100%</td><td>39.61(+1.82)</td><td>26.73(+1.49)</td><td>-8.30(-5.43)</td><td>62.01(+6.19)</td><td>33.09(+2.25)</td><td>-11.05(-4.97)</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.076,
+                0.349,
+                0.895,
+                0.379
+            ],
+            "angle": 0,
+            "content": "Table 1. Overall performance of existing algorithms and Efficient-CLS on OAK and EgoObjects. iCaRL(Wang et al.) denotes the SOTA model presented in [34], and iCaRL(our impl.) is the same method by our implementation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.396,
+                0.473,
+                0.593
+            ],
+            "angle": 0,
+            "content": "\\(\\mathbf{F}\\) estimates the forgetfulness of the model due to the sequential training. It takes into account the time interval between the presence of an object category and its subsequent presence. For a class \\(c\\), we sort the \\(\\mathrm{CAP}_{t_i}^c\\) according to the time interval \\(k\\) between evaluation time \\(t_i\\) and the last time \\(t_i - k\\) the model is trained on \\(c\\). After \\(\\mathrm{CAP}_{t_i}^c\\) is sorted, all \\(\\mathrm{CAP}_{t_i}^c (i = 0,\\dots ,T)\\) are divided into \\(K\\) bins \\(B_{kmin},\\dots ,B_{kmax}\\) according to the time interval \\(k\\). The average CAP \\((\\mathrm{aCAP}_k)\\) of each bin \\(B_{k}\\) is defined as the model's performance for detecting class \\(c\\) after the model has not been trained on \\(c\\) for \\(k\\) time steps. The forgetfulness (F) of the class \\(c\\) is defined as the weighted sum of the performance decrease at each time:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.077,
+                0.604,
+                0.483,
+                0.656
+            ],
+            "angle": 0,
+            "content": "\\[\nF ^ {c} = \\sum_ {k = k m i n} ^ {k m a x} \\frac {k - k m i n}{\\sum_ {k = k m i n} ^ {k m a x} k - k m i n} \\times \\left(\\mathrm {a C A P} _ {k m i n} - \\mathrm {a C A P} _ {k}\\right). \\tag {5}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.655,
+                0.385,
+                0.675
+            ],
+            "angle": 0,
+            "content": "The overall forgetfulness is: \\(\\mathrm{F} = \\frac{1}{C}\\sum_{c = 0}^{C}\\mathrm{F}^{c}\\)"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.682,
+                0.295,
+                0.698
+            ],
+            "angle": 0,
+            "content": "4.4. Implementation Details"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.705,
+                0.471,
+                0.901
+            ],
+            "angle": 0,
+            "content": "For a fair comparison, we followed the prior work [34] to use Faster-RCNN [27] with ResNet-50 backbone [12] as our object detection network, which is initialized by the weights pre-trained on PASCAL VOC [10]. We used Adam optimizer with a constant learning rate of 0.0001, and the batch size was set to 16 frames. Same as [34], we maintained a replay buffer with 5 samples per class. At each time step \\( t \\), we first randomly retrieved 16 video frames from the replay buffer for joint training. We used confidence thresh \\( \\tau = 0.7 \\) to generate pseudo-labels for unlabeled frames, and applied EMA rate \\( \\alpha = 0.99 \\) to update the slow learner. The weight of pseudo loss \\( \\lambda_{pseudo} \\) was set to 1.0. All the experiments were conducted on 2 NVIDIA RTX 3090 GPUs."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.394,
+                0.587,
+                0.41
+            ],
+            "angle": 0,
+            "content": "5. Results"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.42,
+                0.735,
+                0.436
+            ],
+            "angle": 0,
+            "content": "5.1. Fully Supervised Protocol"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.444,
+                0.892,
+                0.536
+            ],
+            "angle": 0,
+            "content": "As the previous work [34] focuses on online continual object detection (OCOD) in video streams, we first evaluated model performance in fully supervised setting (i.e., \\(100\\%\\) annotation cost), where all video frames are paired with human labels. The reported results were measured by standard metrics (CAP, FAP, and F, Section 4.3) in Table 1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.553,
+                0.892,
+                0.794
+            ],
+            "angle": 0,
+            "content": "Performance of CL baselines. [34] benchmarked Incremental, EWC, iCaRL(Wang et al.), and Offline Training on OAK dataset. They found the replay-based method (i.e. iCaRL(Wang et al.)) outperforms regularization-based method (i.e. EWC) by \\(10\\%\\) in FAP, while iCaRL(Wang et al.) has a huge gap of \\(30\\%\\) compared with Offline Training. Similar observations were made in Table 1, but the performance of Incremental and EWC were \\(4\\%\\) lower than that in [34], as in our setting we only trained each mini-batch of video frames once (they trained each mini-batch 10 times). As mentioned in Section 4.2, we introduce several variations to the original design of iCaRL(Wang et al.). Compared with iCaRL(Wang et al.), we observed a huge performance boost in FAP from \\(22.89\\%\\) to \\(36.14\\%\\) on OAK and from \\(37.61\\%\\) to \\(60.80\\%\\) on EgoObjects, abridging the gap between the baseline and Offline Training."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.796,
+                0.892,
+                0.902
+            ],
+            "angle": 0,
+            "content": "We further adapted other standard CL baselines, including iOD, A-GEM, GDumb, DER++, to the OCOD setting for comparisons. iOD is the SOTA method in offline class-incremental object detection. Though it performs well in prior setting, iOD collapses when adapted to online video streams. One explanation is that iOD requires explicit task boundary to trigger the reshape of model gradients that opti"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "19251"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.089,
+                0.093,
+                0.885,
+                0.257
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.264,
+                0.893,
+                0.294
+            ],
+            "angle": 0,
+            "content": "Figure 3. The changes of \\(\\mathrm{CAP}_{t_i}^c\\) with sampled categories on OAK dataset. The x-axis denotes time step across the entire video stream. The y-axis denotes the AP50 of the category at specific time step (i.e., \\(\\mathrm{CAP}_{t_i}^c\\)). The grey line indicates the existence of the category."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.315,
+                0.471,
+                0.406
+            ],
+            "angle": 0,
+            "content": "mizes knowledge sharing between adjacent tasks. However, in online video streams, the task boundaries by classes are no longer available and the change of tasks is hard to identify, resulting in the failure to prevent forgetting. In contrast, replay-based methods (i.e., A-GEM, GDumb and DER++) generalize much better to the real-world streaming video."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.417,
+                0.47,
+                0.658
+            ],
+            "angle": 0,
+            "content": "Performance w/ Efficient-CLS. Inheriting from the benefit of fast and slow learning with EMA, our Efficient-CLS consistently improves all the state-of-the-art CL methods (i.e. iCaRL(our impl.), A-GEM, GDumb and DER++) by a significant margin. Taking iCaRL(our impl.) for example, with Efficient-CLS, we observed superior performance and minimal forgetting compared to baseline methods, even when categories appeared infrequently (e.g., sculpture in Figure 3). Since semantic contextual information is more important in indoor environments on EgoObjects compared to the outdoor environments in OAK, we noticed that the improvement brought by our method is even greater on EgoObjects with an increase of \\(6.25\\%\\) in FAP, \\(3.95\\%\\) in CAP \\(\\%\\) and \\(3.07\\%\\) in F. Consistent performance gains are also noticeable for A-GEM, GDumb and DER++, demonstrating the effectiveness and flexibility of our method."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.666,
+                0.327,
+                0.683
+            ],
+            "angle": 0,
+            "content": "5.2. Sparse Annotation Protocol"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.689,
+                0.471,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Performance of CL baselines. The sparse annotation protocol is more challenging than the previous fully supervised protocol as shown by the performance differences when number of annotated video frames decreases (compare the performance of each colored bar along the x-axis within each subplot in Figure 4). We noted that GDumb is more resilient against the reduce of supervision. Specifically, from Table 2 at the lowest annotation cost of \\(6.25\\%\\) on EgoObjects dataset, GDumb achieves the highest performance of \\(38.74\\%\\), \\(22.69\\%\\) and \\(-4.53\\%\\) in FAP, CAP and F, which surpasses other baselines by a considerable margin. Same observations can be made on OAK dataset. One possible explanation is that GDumb only trains the data stored in the balanced replay buffer, which makes it less vulnerable"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.501,
+                0.311,
+                0.892,
+                0.424
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\"></td><td colspan=\"3\">OAK</td><td colspan=\"3\">EgoObjects</td></tr><tr><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td></tr><tr><td>iCaRL</td><td>23.04</td><td>17.75</td><td>-3.31</td><td>32.91</td><td>19.15</td><td>-1.36</td></tr><tr><td>w/ Efficient-CLS</td><td>29.72</td><td>20.31</td><td>-5.36</td><td>40.16</td><td>23.95</td><td>-2.01</td></tr><tr><td>A-GEM</td><td>23.59</td><td>16.15</td><td>-2.99</td><td>21.84</td><td>12.28</td><td>0.82</td></tr><tr><td>w/ Efficient-CLS</td><td>30.18</td><td>20.59</td><td>-5.44</td><td>38.96</td><td>23.79</td><td>-5.64</td></tr><tr><td>GDumb</td><td>27.37</td><td>19.64</td><td>-4.25</td><td>38.74</td><td>22.69</td><td>-4.53</td></tr><tr><td>w/ Efficient-CLS</td><td>29.07</td><td>19.99</td><td>-6.01</td><td>40.09</td><td>23.67</td><td>-5.16</td></tr><tr><td>DER++</td><td>24.21</td><td>15.93</td><td>-3.79</td><td>16.95</td><td>8.48</td><td>2.03</td></tr><tr><td>w/ Efficient-CLS</td><td>28.63</td><td>19.64</td><td>-4.60</td><td>35.78</td><td>20.74</td><td>-4.69</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.499,
+                0.43,
+                0.893,
+                0.459
+            ],
+            "angle": 0,
+            "content": "Table 2. Performance of Efficient-CLS at low annotation cost of \\(6.25\\%\\). The best results are bold-faced."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.476,
+                0.893,
+                0.507
+            ],
+            "angle": 0,
+            "content": "able to class imbalance problem brought by the reduce of labeled samples in the training set."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.52,
+                0.893,
+                0.791
+            ],
+            "angle": 0,
+            "content": "Performance w/ Efficient-CLS. Our proposed Efficient-CLS is a plug-and-play module that can be easily inserted into and enhance existing continual learning algorithms with the ability to use unlabeled video frames effectively. In both OAK and EgoObjects dataset, Efficient-CLS consistently improves the comparative SOTAs in all three evaluation metrics regardless of various degrees of annotation cost. As shown in Table 2, at a lower annotation cost of \\(6.25\\%\\), Efficient-CLS doubles the performance of DER++ and A-GEM in terms of FAP and CAP, and achieves an even larger improvement in preventing forgetting. Thanks to the useful information from pseudo labels predicted by the slow learner in Efficient-CLS, our method is more robust to various annotation costs, compared with SOTAs (compare the rate of change of blue bars vs. red bars over different degrees of annotation cost). Most remarkably, Efficient-CLS with \\(25\\%\\) annotation cost has already outperformed comparative SOTAs with \\(100\\%\\) annotation cost (see Table 1)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.802,
+                0.655,
+                0.818
+            ],
+            "angle": 0,
+            "content": "5.3. Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.826,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "We assessed the importance of our key design choices. The Complementary Learning Systems (CLS) in EfficientCLS is the key for rapidly adapting to learn new tasks, meanwhile, retaining previously learnt knowledge. It constitutes of two memory reinstatement mechanisms: one is"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.52,
+                0.957
+            ],
+            "angle": 0,
+            "content": "19252"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.12,
+                0.089,
+                0.365,
+                0.219
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.365,
+                0.09,
+                0.609,
+                0.219
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.61,
+                0.091,
+                0.849,
+                0.219
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.12,
+                0.22,
+                0.363,
+                0.348
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.363,
+                0.22,
+                0.605,
+                0.348
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.605,
+                0.22,
+                0.848,
+                0.348
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.352,
+                0.895,
+                0.424
+            ],
+            "angle": 0,
+            "content": "Figure 4. Evaluation of online continual object detection in video streams with three metrics (FAP, CAP and F, Section 4.3) on OAK dataset (first row) and EgoObjects dataset (second row). The higher the bars are, the better. The x-axis denotes the percentage of video frames that are labeled in the video stream. It ranges from \\(6.25\\%\\) to \\(100\\%\\) (full supervision). The y-axis indicates the performance using different evaluation metrics. Ours (iCaRL(our impl.) w/ Efficient-CLS, red) consistently beats the comparative SOTA (iCaRL(our impl.), blue) in all evaluation metrics."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.093,
+                0.44,
+                0.875,
+                0.519
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">EMA</td><td rowspan=\"2\">PL</td><td colspan=\"3\">50%</td><td colspan=\"3\">25%</td><td colspan=\"3\">12.5%</td><td colspan=\"3\">6.25%</td></tr><tr><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td></tr><tr><td>X</td><td>X</td><td>34.68</td><td>25.78</td><td>-4.15</td><td>33.70</td><td>24.57</td><td>-4.30</td><td>28.76</td><td>19.80</td><td>-5.48</td><td>23.04</td><td>17.75</td><td>-3.31</td></tr><tr><td>✓</td><td>X</td><td>35.74</td><td>25.77</td><td>-4.82</td><td>34.79</td><td>25.62</td><td>-4.35</td><td>31.72</td><td>21.16</td><td>-7.24</td><td>27.84</td><td>20.03</td><td>-3.96</td></tr><tr><td>X</td><td>✓</td><td>35.61</td><td>25.56</td><td>-3.76</td><td>34.95</td><td>25.65</td><td>-3.65</td><td>31.60</td><td>22.44</td><td>-4.83</td><td>26.39</td><td>19.50</td><td>-1.99</td></tr><tr><td>✓</td><td>✓</td><td>38.61</td><td>26.90</td><td>-7.29</td><td>38.36</td><td>26.64</td><td>-8.20</td><td>33.92</td><td>23.04</td><td>-7.71</td><td>29.72</td><td>20.31</td><td>-5.36</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.076,
+                0.523,
+                0.893,
+                0.552
+            ],
+            "angle": 0,
+            "content": "Table 3. Effectiveness of Exponential Moving Average (EMA) and Pseudo-labeling (PL) for iCaRL (our impl.) on OAK dataset at annotation cost \\(50\\%\\), \\(25\\%\\), \\(12.5\\%\\) and \\(6.25\\%\\). The best results are bold-faced."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.57,
+                0.471,
+                0.632
+            ],
+            "angle": 0,
+            "content": "synaptic weight transfer from fast to slow learner via exponential moving average (EMA); and the other is reciprocal replay from slow learner to fast learner with pseudolabeling (PL). Here we studied their effects individually."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.644,
+                0.473,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Ablation: Exponential Moving Average (EMA). We removed EMA by setting the \\(\\alpha\\) in Equation 3 to 1, where the model weights of the fast learner and slow learner are now shared throughout the learning process. Note that in the fully supervised protocol, the pseudo-labeling is turned off and the Efficient-CLS equals to the EMA alone. From Table 1 at \\(100\\%\\) annotation cost, we observed that removing EMA leads to significant performance drops ranging from \\(2\\%\\) to \\(8\\%\\), for all the comparative baselines on both OAK and EgoObjects. This shows that the slow learner can effectively consolidate the knowledge from the fast learner, and constructively alleviate catastrophic forgetting by synapses consolidation over time. Similar observations were made in Table 3 in the sparse annotation protocol (compare Row 2 vs. Row 1, Row 4 vs. Row 3). It is worth noting that, the performance difference between naive model (Row 1) and its variant with EMA (Row 2) is slightly larger in lower su"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.57,
+                0.893,
+                0.661
+            ],
+            "angle": 0,
+            "content": "pervision (i.e. \\(12.5\\%\\), \\(6.25\\%\\)) than higher supervision (i.e. \\(50\\%\\), \\(25\\%\\)). One possible reason is that, compared with higher supervision, the fast learner suffers more forgetting in lower supervision; hence, the effect of removing EMA becomes stronger in lower supervision, again highlighting the importance of EMA."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.674,
+                0.895,
+                0.809
+            ],
+            "angle": 0,
+            "content": "Ablation: Pseudo-labeling (PL). We ablate our model by removing the PL of the slow learner across varying annotation costs and reported the results in Table 3. The removal of PL (Row 2) leads to a performance drop of around \\(2\\%\\) in FAP, \\(1\\%\\) in CAP and \\(0.5 - 4\\%\\) in F, compared with our full Efficient-CLS (Row 4). It implies that the slow learner captures useful semantic information from unlabeled video frames and these predicted pseudo labels are helpful in training the fast learner."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.811,
+                0.895,
+                0.9
+            ],
+            "angle": 0,
+            "content": "To investigate whether pseudo labels predicted by the fast learner itself could help stream learning, we conducted another ablation experiment where we performed PL without EMA (Row 3). Compared with the naive model (Row 1), we observed a performance increase from \\(28.76\\%\\) to \\(31.60\\%\\) in FAP and \\(19.80\\%\\) to \\(22.44\\%\\) in CAP. It indicates"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "19253"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.092,
+                0.47,
+                0.137
+            ],
+            "angle": 0,
+            "content": "that, due to the temporal correlation in video stream, pseudo labels predicted by the fast learner can serve as an informative supervision for the training of the fast learner itself."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.138,
+                0.471,
+                0.364
+            ],
+            "angle": 0,
+            "content": "However, replaying the self-predicted pseudo labels on the fast learner fails to prevent forgetting, as indicated by the drop from \\(-5.48\\%\\) to \\(-4.83\\%\\) in F. It is possible that the pseudo labels generated by the fast learner only bias towards the classes which have already been learnt very well and fail to reinforce the fast learner to improve on the poorly-learnt classes. Different from the fast learner, the slow learner integrates semantic information over time. The predicted pseudo labels carry more semantic information, which is useful for fast learner to capture more generic object representations during pseudo label replays. Again, this emphasizes that the reciprocal replay from the slow learner to the fast learner is critical for memory reinstatement, which has been missing in the computational modeling literature of CLS."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.379,
+                0.196,
+                0.394
+            ],
+            "angle": 0,
+            "content": "6. Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.405,
+                0.471,
+                0.603
+            ],
+            "angle": 0,
+            "content": "To imitate what humans see and learn in the real world, we introduce a more realistic and challenging problem of label-efficient online continual object detection (LEOCOD) in video streams. Addressing this problem would greatly benefit real-world applications by reducing model retraining time and data labeling costs. Inspired by the Complementary Learning Systems (CLS) theory, we propose a plug-and-play module, namely Efficient-CLS, that can be easily integrated into and improve existing continual leaning algorithms. We evaluate Efficient-CLS on two challenging real-world video datasets, where our method achieves the state-of-the-art performance in preventing catastrophic forgetting, all while requiring minimal annotation effort."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.623,
+                0.47,
+                0.79
+            ],
+            "angle": 0,
+            "content": "Acknowledgements This research is funded by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008 & NRF-NRFF15-2023-0001, AI Singapore Programme (AISG Award No: AISGC-2019-001-2A & AISG2-RP-2021-025), Mike Zheng Shou's Start-Up Grant from NUS, Mengmi Zhang's Startup Grant from Agency for Science, Technology, and Research (A*STAR), and Early Career Investigatorship from Center for Frontier AI Research (CFAR), A*STAR. The computational work for this article was partially performed on resources of the National Supercomputing Centre, Singapore."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.08,
+                0.805,
+                0.174,
+                0.82
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.085,
+                0.829,
+                0.468,
+                0.857
+            ],
+            "angle": 0,
+            "content": "[1] Egoobjects dataset. https://ai.facebook.com/datasets/egoobjects-dataset/, May 2022. 2, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.859,
+                0.469,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[2] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.085,
+                0.829,
+                0.469,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.533,
+                0.093,
+                0.892,
+                0.119
+            ],
+            "angle": 0,
+            "content": "retrieval. Advances in neural information processing systems, 32, 2019. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.121,
+                0.892,
+                0.176
+            ],
+            "angle": 0,
+            "content": "[3] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Advances in neural information processing systems, 32, 2019. 1, 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.178,
+                0.894,
+                0.232
+            ],
+            "angle": 0,
+            "content": "[4] Elahe Arani, Fahad Sarfraz, and Bahram Zonooz. Learning fast, learning slow: A general continual learning method based on complementary learning system. arXiv preprint arXiv:2201.12604, 2022. 2, 3, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.233,
+                0.894,
+                0.288
+            ],
+            "angle": 0,
+            "content": "[5] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.289,
+                0.892,
+                0.356
+            ],
+            "angle": 0,
+            "content": "[6] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 2, 3, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.358,
+                0.892,
+                0.413
+            ],
+            "angle": 0,
+            "content": "[7] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuytelaars, Joelle Pineau, and Eugene Belilovsky. New insights on reducing abrupt representation change in online continual learning. arXiv preprint arXiv:2203.03798, 2022. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.414,
+                0.892,
+                0.455
+            ],
+            "angle": 0,
+            "content": "[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 2, 3, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.456,
+                0.892,
+                0.512
+            ],
+            "angle": 0,
+            "content": "[9] Hung-Jen Chen, An-Chieh Cheng, Da-Cheng Juan, Wei Wei, and Min Sun. Mitigating forgetting in online continual learning via instance-aware parameterization. Advances in Neural Information Processing Systems, 33:17466-17477, 2020. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.512,
+                0.892,
+                0.567
+            ],
+            "angle": 0,
+            "content": "[10] Mark Everingham, SM Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111(1):98-136, 2015. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.568,
+                0.892,
+                0.622
+            ],
+            "angle": 0,
+            "content": "[11] Enrico Fini, Stéphane Lathuiliere, Enver Sangineto, Moin Nabi, and Elisa Ricci. Online continual learning under extreme memory constraints. In European Conference on Computer Vision, pages 720-735. Springer, 2020. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.623,
+                0.892,
+                0.678
+            ],
+            "angle": 0,
+            "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.679,
+                0.892,
+                0.734
+            ],
+            "angle": 0,
+            "content": "[13] Jisoo Jeong, Seungeui Lee, Jeesoo Kim, and Nojun Kwak. Consistency-based semi-supervised learning for object detection. Advances in neural information processing systems, 32, 2019. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.735,
+                0.892,
+                0.775
+            ],
+            "angle": 0,
+            "content": "[14] Daoyun Ji and Matthew A Wilson. Coordinated memory replay in the visual cortex and hippocampus during sleep. Nature neuroscience, 10(1):100-107, 2007. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.776,
+                0.892,
+                0.817
+            ],
+            "angle": 0,
+            "content": "[15] Nitin Kamra, Umang Gupta, and Yan Liu. Deep generative dual memory network for continual learning. arXiv preprint arXiv:1710.10368, 2017. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.818,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[16] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 5"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.504,
+                0.093,
+                0.894,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.52,
+                0.957
+            ],
+            "angle": 0,
+            "content": "19254"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.092,
+                0.47,
+                0.16
+            ],
+            "angle": 0,
+            "content": "[17] Joseph Kj, Jathushan Rajasegaran, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Incremental object detection via meta-learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.163,
+                0.47,
+                0.219
+            ],
+            "angle": 0,
+            "content": "[18] Dharshan Kumaran, Demis Hassabis, and James L McClelland. What learning systems do intelligent agents need? complementary learning systems theory updated. Trends in cognitive sciences, 20(7):512-534, 2016. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.221,
+                0.469,
+                0.275
+            ],
+            "angle": 0,
+            "content": "[19] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn and think like people. Behavioral and brain sciences, 40, 2017. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.277,
+                0.469,
+                0.332
+            ],
+            "angle": 0,
+            "content": "[20] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. arXiv preprint arXiv:2102.09480, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.334,
+                0.469,
+                0.375
+            ],
+            "angle": 0,
+            "content": "[21] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.377,
+                0.469,
+                0.459
+            ],
+            "angle": 0,
+            "content": "[22] Lorenzo Pellegrini, Chenchen Zhu, Fanyi Xiao, Zhicheng Yan, Antonio Carta, Matthias De Lange, Vincenzo Lomonaco, Roshan Sumbaly, Pau Rodriguez, and David Vazquez. 3rd continual learning workshop challenge on egocentric category and instance level object understanding. arXiv preprint arXiv:2212.06833, 2022. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.462,
+                0.469,
+                0.503
+            ],
+            "angle": 0,
+            "content": "[23] Quang Pham, Chenghao Liu, and Steven Hoi. Dualnet: Continual learning, fast and slow. Advances in Neural Information Processing Systems, 34, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.505,
+                0.469,
+                0.559
+            ],
+            "angle": 0,
+            "content": "[24] Quang Pham, Chenghao Liu, Doyen Sahoo, and HOI Steven. Contextual transformation networks for online continual learning. In International Conference on Learning Representations, 2020. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.562,
+                0.469,
+                0.617
+            ],
+            "angle": 0,
+            "content": "[25] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In European conference on computer vision, pages 524-540. Springer, 2020. 2, 3, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.619,
+                0.469,
+                0.686
+            ],
+            "angle": 0,
+            "content": "[26] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 2, 3, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.689,
+                0.469,
+                0.743
+            ],
+            "angle": 0,
+            "content": "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 4, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.746,
+                0.469,
+                0.8
+            ],
+            "angle": 0,
+            "content": "[28] Mohammad Rostami, Soheil Kolouri, and Praveen K Pilly. Complementary learning for overcoming catastrophic forgetting using experience replay. arXiv preprint arXiv:1903.04566, 2019. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.803,
+                0.469,
+                0.871
+            ],
+            "angle": 0,
+            "content": "[29] Dongsub Shim, Zheda Mai, Jihwan Jeong, Scott Sanner, Hyunwoo Kim, and Jongseong Jang. Online class incremental continual learning with adversarial shapley value. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 9630-9638, 2021. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.874,
+                0.469,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[30] Krishna Kumar Singh, Kayvon Fatahalian, and Alexei A Efros. Krishnacam: Using a longitudinal, single-person,"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.092,
+                0.47,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.535,
+                0.093,
+                0.892,
+                0.134
+            ],
+            "angle": 0,
+            "content": "egocentric dataset for scene understanding tasks. In 2016 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1-9. IEEE, 2016. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.136,
+                0.892,
+                0.204
+            ],
+            "angle": 0,
+            "content": "[31] James Smith, Jonathan Balloch, Yen-Chang Hsu, and Zsolt Kira. Memory-efficient semi-supervised continual learning: The world is its own replay buffer. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.206,
+                0.892,
+                0.261
+            ],
+            "angle": 0,
+            "content": "[32] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.263,
+                0.892,
+                0.318
+            ],
+            "angle": 0,
+            "content": "[33] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.32,
+                0.892,
+                0.387
+            ],
+            "angle": 0,
+            "content": "[34] Jianren Wang, Xin Wang, Yue Shang-Guan, and Abhinav Gupta. Wanderlust: Online continual object detection in the real world. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10829-10838, 2021. 1, 2, 3, 5, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.39,
+                0.892,
+                0.443
+            ],
+            "angle": 0,
+            "content": "[35] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.446,
+                0.892,
+                0.515
+            ],
+            "angle": 0,
+            "content": "[36] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14421-14430, 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.517,
+                0.892,
+                0.558
+            ],
+            "angle": 0,
+            "content": "[37] Xiangli Yang, Zixing Song, Irwin King, and Zenglin Xu. A survey on deep semi-supervised learning. IEEE Transactions on Knowledge and Data Engineering, 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.56,
+                0.892,
+                0.615
+            ],
+            "angle": 0,
+            "content": "[38] Barret Zoph, Ekin D Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V Le. Learning data augmentation strategies for object detection. In European conference on computer vision, pages 566-583. Springer, 2020. 4"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.503,
+                0.093,
+                0.892,
+                0.615
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "19255"
+        }
+    ]
+]

2023/Label-Efficient Online Continual Object Detection in Streaming Video/bb408240-93f4-4420-9e47-a5facd88851a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f0e80a36f4d370683922f406726e4161653dfb83dd142add2e8cb3e0374866b
+size 2745861

2023/Label-Efficient Online Continual Object Detection in Streaming Video/full.md

@@ -0,0 +1,282 @@
+# Label-Efficient Online Continual Object Detection in Streaming Video
+
+Jay Zhangjie Wu $^{1}$  David Junhao Zhang $^{1}$  Wynne Hsu $^{2}$  Mengmi Zhang $^{3,4}$  Mike Zheng Shou $^{1*}$
+
+$^{1}$ Show Lab,  $^{2}$ National University of Singapore  
+ $^{3}$ School of Computer Science and Engineering, Nanyang Technological University, Singapore  
+ $^{4}$ CFAR and I2R, Agency for Science, Technology and Research, Singapore
+
+# Abstract
+
+Humans can watch a continuous video stream and effortlessly perform continual acquisition and transfer of new knowledge with minimal supervision yet retaining previously learnt experiences. In contrast, existing continual learning (CL) methods require fully annotated labels to effectively learn from individual frames in a video stream. Here, we examine a more realistic and challenging problem—Label-Efficient Online Continual Object Detection (LEOCOD) in streaming video. We propose a plug-and-play module, Efficient-CLS, that can be easily inserted into and consistently improve existing CL algorithms for object detection in video streams with reduced data annotation costs and model retraining time. We show that our method has achieved significant improvement with minimal forgetting across all supervision levels on two challenging CL benchmarks for streaming real-world videos. Remarkably, with only  $25\%$  annotated video frames, our proposed method still outperforms the state-of-the-art CL models trained with  $100\%$  annotations on all video frames. The data and source code will be publicly available at https://github.com/showlab/Efficient-CLS.
+
+# 1. Introduction
+
+Humans have the ability to continuously learn from an ever-changing environment, while retaining previously learnt experiences. In contrast to human learning, prior works [3, 2, 11, 34, 7] show that deep neural networks are prone to catastrophic forgetting. To address the forgetting problem, existing works in continual learning (CL) primarily focus on class-incremental image classification or object detection. Their experiment settings are often idealistic and simplified, where i.i.d. static images are usually grouped by class and incrementally presented to computational models in sequence. To learn a particular task containing specific
+
+classes, an agent can go through all the data of current task over multiple epochs. After that, the learned classes in current task become unavailable, i.e., no overlaps between the sets of learned classes and unseen classes.
+
+However, these experiment designs deviate from the online continual learning (OCL) setting in the real world, where an agent learns from temporally correlated non-i.i.d. video streams in one single pass. Given context regularities in natural environments, an agent is likely to encounter cases when objects of previously learnt classes co-occur with unknown objects from unseen classes, e.g., a computer mouse and a computer monitor often co-occur. Taking these considerations, [34] introduces OCL on object detection in real-world video streams. They evaluate existing CL approaches on this setting and report a huge performance gap compared with offline training.
+
+Based on the setting in [34], we take a significant step further and introduce a novel problem setting called Label-Efficient Online Continual Object Detection (LEOCOD), which highlights two unique challenges. First, the setting in [34] is such that the CL algorithms are trained with every mini-batch over multiple passes. We tighten the training recipe in LEOCOD to strictly online, where data is allowed to have one single pass and models are trained on the entire video dataset for only one epoch. Second, existing CL models require fully supervised training where box-level ground truth labels of every object on every video frame have to be obtained from human annotators. Unlike static images, acquiring human annotations for object detection on videos can be expensive and daunting. Thus, in LEOCOD, the video frames per mini-batch are sparsely annotated to alleviate the burdens of extensive human labeling, making LEOCOD one step closer to real world application.
+
+Cognitive science works [35, 19] show that humans are efficient at continuously learning from very few annotated data samples. We get inspirations from the theory of Complementary Learning Systems (CLS) in human brains [18], and propose a plug-and-play module for the LEOCOD task, dubbed as Efficient-CLS. In Efficient-CLS, we introduce
+
+![](images/a090acd2ce689d64a3637fba2655497866caff7450015ef2e8e9c44ef677bcb6.jpg)  
+Figure 1. (a) Problem introduction: An agent continuously learns from a never-ending online video stream over time. In each training step, out of a mini-batch containing 16 consecutive video frames, only a fixed proportion of frames are labeled (green boundary), while the rest of the frames are unlabeled (orange boundary). Following [34], the video frame after every training mini-batch (transparent) is held out for testing. After every 100 training steps, the agent is evaluated on all the video frames from the test set for object detection. (b) Key results: Our proposed method (red) consistently outperforms the best competitive baseline (blue) by a margin of  $5\%$ . Remarkably, our model, trained at  $25\%$  annotation cost, surpasses the best baseline trained at  $100\%$  (grey line). The orange cross denotes the performance of the state-of-the-art model, which is  $15\%$  lower than our method.
+
+![](images/234e7cabfb9d3c0d9353c67dc8de8cfcc68f8333688ee4fc4fea263dba5918d3.jpg)
+
+two feed-forward neural networks as slow and fast learners. In the fast learner, memory is rapidly adapted to the current task. The weights of the slow learner change a little on each reinstatement, and are maintained by taking the exponential moving average (EMA) of the fast learner's weights over time. Though a few continual learning models in previous works [4, 24] also use a similar source of inspiration, they miss the effect of reciprocal connections from slow learners to fast learners, which we intend to address. Inspired by the bidirectional interaction in CLS [14], we reactivate the weights of the slow learners to predict meaningful pseudo labels from the unlabeled video frames and use these pseudo labels to guide the training of the fast learner, closing the loop between the two systems.
+
+We demonstrate the versatility and effectiveness of our Efficient-CLS on two standard real-world video datasets, OAK [34] and EgoObjects [1]. Our proposed method can be easily integrated into existing CL models and consistently improve their performance by a large margin in LEOCOD. It is worth noting that, with only  $25\%$  labeled data, our method surpasses the comparative baselines trained with full supervision (Figure 1(b)).
+
+To summarize, we make the following key contributions:
+
+- We introduce a new, challenging and important problem of label-efficient online continual object detection (LEOCOD) in video streams. Solving this problem would greatly benefit real-world applications in reducing annotation cost and model retraining time.  
+- We propose Efficient-CLS, a plug-and-play module inspired from the Complementary Learning Systems (CLS) theory, which can be integrated into existing CL
+
+models and learn efficiently and effectively with less supervision and minimal forgetting.
+
+- We benchmark existing CL methods on the task of LEOCOD and demonstrate the state-of-the-art performance of our method through extensive experiments.
+
+# 2. Related Work
+
+# 2.1. Continual Learning
+
+To alleviate catastrophic forgetting, many continual learning methods exploit an external buffer where a limited number of old samples are stored and used for replay when adapting to a new task. iCaRL [26] stores the representative exemplars in past tasks for knowledge distillation and prototype rehearsal. Gradient Episodic Memory (GEM) [21] formulates optimization constraints on the exemplars in memory. Averaged GEM (A-GEM) [8] is an improved version of GEM that achieves faster training and less memory consumption. GDumb [25] greedily stores samples in memory as they come and trains a model using samples only in the memory. Dark Experience Replay++ (DER++) [6] combines replay with knowledge distillation and regularization, and samples logits along the entire optimization trajectory.
+
+In contrast to classical continual learning where data are separated by task boundaries and models are trained with multiple iterations in every task, we examine a more realistic and challenging problem where data are provided in tiny batches and models are trained on these batches only once. Recently, online continual learning (OCL) has gained increasing interests in computer vision [3, 7, 29, 9, 34]. Many OCL methods rely on representative memory replays to pre
+
+vent forgetting. [3] utilizes gradients of network parameters to select replay samples of maximum diversity. Subsequent works [2, 29] propose to use losses and scoring functions as criteria for selecting the most representative samples for replay. However, these approaches tackle image classification problem in an artificial setting, where new classes appear in a specific order. Their performance in real-world vision tasks remains unclear.
+
+Lately, [34] benchmarks CL methods in the real-world online setting with full supervision. As the video streams arrive endlessly in a real-time manner, assigning annotations to all the video frames for training computational models is laborious and time-consuming. It becomes even more daunting in object detection tasks where class labels and bounding boxes of all objects on a video frame have to be provided. Reducing burdensome costs of labeling remains an under-explored and challenging problem in online continual object detection. We propose a self-sustaining Efficient-CLS, which is capable of exploiting the unlabeled video frames by pseudo-labeling when the number of labeled frames is limited.
+
+# 2.2. Complementary Learning Systems
+
+The essence of fast and slow learning in Complementary Learning Systems (CLS) has benefited several continual learning algorithms in image recognition [24, 23, 28, 4, 15]. However, these methods either require the task boundaries, which are not applicable in our online video setting, or they require to train fast and slow learning systems with replay samples from the same replay buffer, which could easily lead to overfitting problem when the replay buffer has limited capacity. To eliminate overfitting problem, [28] and [15] utilize generative replay models to couple sequential tasks in a latent embedding space. While generative approaches have succeeded in artificial and simple datasets, they often fail in complex vision tasks, e.g., object detection. DualNet [23] also employs a slow-fast learning architecture. However, its update process for the slow learner is more computationally demanding because it leans on both self-supervised and supervised learning objectives. The self-supervised phase demands additional training iterations on extensive batches to yield satisfactory results, thereby delaying the training of the fast learner. In contrast, our method updates the slow learner in real-time using the fast learner's weights, obviating the need for distinct slow learner training. Our approach is optimized for online CL scenarios where computational resources are constrained.
+
+# 2.3. Semi-Supervised Learning
+
+The goal of semi-supervised learning (SSL) [33, 5, 37, 36] is to reduce the demand of labeled data and harness unlabeled data for performance improvement. [31] first introduce SSL to the context of CL, and propose a strategy
+
+that combines pseudo-labeling, consistency regularization, out-of-distribution detection, and knowledge distillation to solve the problem of class-incremental image classification. However, to distill knowledge from previous tasks, their model relies heavily on the task boundary that identifies the change of training classes, which is not available in our LEOCOD setting. To reduce annotation costs in object detection, several methods [13, 32, 20] capitalize the teacher-student networks. In general, a teacher model predicts pseudo labels or enforces a consistency loss to guide the student networks. However, these previous works on semi-supervised object detection only consider the offline setting on static image datasets, while none of them has been extended to online continual learning on dynamic video streams. We discovered new insights that pseudo-labeling from slow learner to fast learner can not only reduce annotation overhead, but also alleviate forgetting, which is critical to the design of efficient continual learners in real world.
+
+# 3. Method
+
+# 3.1. Problem Setting
+
+We advance the online continual object detection in [34] to a label-efficient and computationally-efficient setting—Label-Efficient Online Continual Object Detection (LEOCOD). In contrast to the setting in [34] where video frames are extensively annotated and trained with multiple epochs, an agent in LEOCOD continuously learns from a sparsely annotated video stream in a single pass over time (see Figure 1(a)).
+
+Formally, we consider the online continual object detection on a continuum of video streams  $\mathcal{D} = \{D_1,\dots ,D_T\}$  where at time step  $t$  , a learning agent receives a mini-batch of continuous video frames  $D_{t}$  from current environment for online training (one single pass). To perform label-efficient object detection, within the batch  $D_{t}$  , only a subset of video frames  $D_{t}^{s} = (X_{t}^{s},Y_{t}^{s})$  are labeled, while the remaining video frames  $D_{t}^{u} = (X_{t}^{u})$  are unlabeled. For each labeled data sample, its annotation contains the bounding box locations and their corresponding class labels.
+
+# 3.2. Efficient Complementary Learning Systems
+
+We propose a plug-and-play module dubbed as EfficientCLS. Specifically, it consists of two feed-forward networks: (i) the fast learner is designed to quickly encode new knowledge from current data stream and then consolidate it to the slow learner; and (ii) the slow learner accumulates the acquired knowledge from fast learner over time and guides the fast learner with meaningful pseudo labels, when full supervision is not available. Same as [26, 8, 25, 6, 34], we maintain an external episodic memory, as a replay buffer, to store exemplars that can be retrieved for replays alongside ongoing video stream. As the fast and slow learners are
+
+![](images/02a21dd849294a5f80d943914591152a18efa413983492843b7fea011e65dfed.jpg)  
+Figure 2. The framework of Efficient-CLS. At each learning step, the system receives a batch of temporally continuous data  $D_{t}$ , including labeled (green) and unlabeled (orange) frames. The fast learner trains the labeled frames alongside a small subset of labeled exemplars retrieved from episodic memory with the supervised loss  $\mathcal{L}_{sup}$ . Meanwhile, the fast learner leverages the pseudo labels generated by the slow learner to optimize a pseudo loss  $\mathcal{L}_{pseudo}$ . To reinstate memory of the slow learner, the weights of the slow learner are updated by taking the Exponential Moving Average (EMA) of the fast learner's weights. The fast and slow learners are complementary to each other, forming a positive feedback loop.
+
+model agnostic, our Efficient-CLS can be easily integrated into existing CL models, which leads to less supervision and minimal forgetting.
+
+Learning with Labeled Frames. The fast learner and slow learner use the same standard Faster-RCNN [27] detector  $f$ . Despite the same architecture, the weights of the fast and slow learners are not shared. We use  $\theta_{F}$  and  $\theta_{S}$  to denote the network parameters for fast and slow learners respectively. As shown in Figure 2, at each training step  $t$ , we use the labeled video frames  $D_{t}^{s} = (X_{t}^{s}, Y_{t}^{s})$  to optimize the fast learner  $\theta_{F}$  with the standard supervised loss  $\mathcal{L}_{sup}$  in Faster-RCNN [27]. It consists of four losses: Region Proposal Network (RPN) classification loss  $\mathcal{L}_{cls}^{rpn}$ , RPN regression loss  $\mathcal{L}_{reg}^{rpn}$ , Region of Interest (ROI) classification loss  $\mathcal{L}_{cls}^{roi}$ , and ROI regression loss  $\mathcal{L}_{reg}^{roi}$ . We define  $\mathcal{L}_{sup}$  as:
+
+$$
+\mathcal {L} _ {s u p} = \mathcal {L} _ {c l s} ^ {r p n} \left(X _ {t} ^ {s}, Y _ {t} ^ {s}\right) + \mathcal {L} _ {r e g} ^ {r p n} \left(X _ {t} ^ {s}, Y _ {t} ^ {s}\right) + \tag {1}
+$$
+
+$$
+\mathcal {L} _ {c l s} ^ {r o i} \left(X _ {t} ^ {s}, Y _ {t} ^ {s}\right) + \mathcal {L} _ {r e g} ^ {r o i} \left(X _ {t} ^ {s}, Y _ {t} ^ {s}\right).
+$$
+
+Learning with Unlabeled Frames. We introduce a pseudo-labeling paradigm to capitalize the information from unlabeled video frames  $D_{t}^{u} = (X_{t}^{u})$  for training. In our early exploration, we intuitively use the fast learner for pseudo-labeling as it quickly adapts the knowledge of nearby frames. However, we observe that using the pseudo labels generated by the fast learner for self-replay exhibits biases towards recently seen objects, which is less effective in preventing forgetting. This has also been verified in our ablation study (Section 5.3). In contrast, the slow learner
+
+preserves the semantic knowledge over a longer time span which generates pseudo labels with fewer biases. This encourages the fast learner to capture more generic scene representations, hence, in turn, contributing to reinstatement of memory in the slow learner (Section 3.2), resulting in a positive feedback loop.
+
+Given all these design considerations, the slow learner takes the unlabeled video frames  $D_{t}^{u}$  as inputs to estimate the possible objects of interest and their corresponding bounding box locations. For brevity, we refer these "pseudo bounding boxes and their corresponding class labels" as "pseudo labels" in the paper. To get rid of false positives, we apply a threshold  $\tau$  to filter out bounding boxes with predicted low confidence scores. Moreover, there also exist repetitive boxes which negatively impact the quality of pseudo-labeling. To address this issue, we use the technique of class-wise non-maximum suppression (NMS) [27] to remove the overlapped boxes and get the high-quality pseudo labels. Formally, the procedure of pseudo label generation is summarized below:
+
+$$
+Y _ {t} ^ {u} = \mathrm {N M S} ([ f (X _ {t} ^ {u}; \theta_ {S}) ] _ {> \tau}), \tag {2}
+$$
+
+where  $[\cdot ]_{>\tau}$  denotes the bounding box selection with confidence score larger than  $\tau$
+
+Given that the video streams are captured from the egocentric perspective in the real world, head and body motions may lead to undesired motion blur effects on some video frames. To enforce our module to learn invariant object representations from these video frames, same as the previous work [38], we apply data augmentation techniques on the pseudo-labeled frames, including 2D image
+
+crops, rotations, and flipping. Note that different from image classification, the predicted bounding box locations also need to be updated accordingly after image augmentations. We denote these pseudo-labeled video frames and their re-adjusted pseudo labels after data augmentations as  $(\tilde{X}_t^u,\tilde{Y}_t^u)$ . We can then use these pseudo pairs  $(\tilde{X}_t^u,\tilde{Y}_t^u)$  to train the fast learner by optimizing the pseudo loss  $\mathcal{L}_{pseudo}:= \mathcal{L}_{cls}^{roi}(\tilde{X}_t^u,\tilde{Y}_t^u) + \mathcal{L}_{req}^{roi}(\tilde{X}_t^u,\tilde{Y}_t^u)$ .
+
+Overall, our Efficient-CLS is jointly trained with the following losses:  $\mathcal{L}_{total} = \mathcal{L}_{sup} + \lambda_{pseudo}\mathcal{L}_{pseudo}$ , where  $\lambda_{pseudo}$  is the weight of  $\mathcal{L}_{pseudo}$ .
+
+Synapses Consolidation via EMA. To alleviate forgetting of obtained knowledge, we apply Exponential Moving Average (EMA) to gradually update the slow learner with the fast learner's weights. The evolving weight changes in the slow learner are functionally correlated with the memory consolidation mechanism in the hippocampus and the neocortex [4]. Formally, we define EMA process as:
+
+$$
+\theta_ {S} = \alpha \theta_ {S} + (1 - \alpha) \theta_ {F}, \tag {3}
+$$
+
+where the  $\alpha \in [0,1]$  is EMA rate. According to the stability-plasticity dilemma, a smaller  $\alpha$  means faster adaptation but less memorization. Empirically, we set  $\alpha = 0.99$ , which leads to best performance.
+
+# 4. Experimental details
+
+# 4.1. Datasets
+
+We evaluate the continual learning methods on two realistic and challenging video datasets, OAK [34] and EgoObjects [1], for the task of online continual object detection.
+
+OAK [34] is a large egocentric video stream dataset spanning nine months of a graduate student's life, consisting of 7.6 million frames of 460 video clips with a total length of 70.2 hours [30]. The dataset contains 103 object categories. We follow [34] in the ordering of training and testing data splits. One frame every 16 consecutive video frames lasting for 30 seconds is held out to construct a test set and the remaining frames are used for training.
+
+EgoObjects [1] is one of the largest object-centric datasets focusing on object detection task. The dataset contains around 100k video frames with  $\sim 250\mathrm{k}$  annotations, which correspond to 277 categories and 1110 main objects [22]. Additionally, the dataset follows a long-tailed distribution which makes the task more challenging and real-world oriented. We adopt the same rule as OAK dataset to split the training and testing set.
+
+# 4.2. Baselines
+
+We compare our model against the following baselines: 1) Vanilla training: Incremental Training is a naive baseline trained sequentially over the entire video stream without any measures to avoid catastrophic forgetting; Offline
+
+Training is an upper bound that can access all the data throughout training; 2) Continual learning algorithms: EWC [16], iCaRL [26], A-GEM [8], GDumb [25], DER++ [6] and iOD [17].
+
+The iCaRL model implemented by [34] stands as the SOTA method in online continual object detection. We reproduce their results using the released code<sup>1</sup>. When calculating RPN and ROI losses for replay samples, their iCaRL model neglects the losses of background proposals and penalizes the foreground losses according to the proportion of the current samples and replay samples. We empirically found that this trick hinders the model from effective episodic replay, thus resulting in severe forgetting. Therefore, we re-implement the iCaRL by discarding the reweighting trick and reverting back to the standard RPN and ROI losses. We name these two different implementations as iCaRL(Wang et al.) and iCaRL(our impl.), respectively.
+
+# 4.3. Evaluation
+
+Protocols. First, we define the annotation cost as the proportion of number of labeled frames versus the total 16 frames within a mini-batch  $D_{t}$ . For example, if 2 out of 16 consecutive frames within  $D_{t}$  get labeled, the annotation cost is  $2 / 16 = 12.5\%$ . The frames to be labeled are randomly selected within each mini-batch  $D_{t}$ . Considering that different choices of labeled frames might influence the model performance, for fair comparisons between models, we fix the choice of randomly selected labeled frames and use the same labeled and unlabeled frames for training all models. Based on the various annotation costs, we introduce two training protocols: fully supervised protocol (100% annotation cost) and sparse annotation protocol (where the annotation cost is less than 100%). In sparse annotation protocol, we further split the training experiments based on 50%/25%/12.5%/6.25% annotation costs.
+
+Metrics. We evaluate the models with three standard metrics: continual average precision (CAP), final average precision (FAP) and forgetfulness (F) [34]. We employ an Average Precision (AP) metric at an Intersection over Union (IoU) threshold of 0.5, commonly referred to as AP50.
+
+CAP shows the average performance over the entire video stream. That is,
+
+$$
+\mathrm {C A P} = \frac {1}{N} \sum_ {i = 0} ^ {N} \mathrm {C A P} _ {t _ {i}} = \frac {1}{N C} \sum_ {i = 0} ^ {N} \sum_ {c = 0} ^ {C} \mathrm {C A P} _ {t _ {i}} ^ {c}, \tag {4}
+$$
+
+where  $\mathrm{CAP}_{t_i}^c$  is the AP of class  $c$  on test set at the  $i$ -th evaluation step,  $N$  is the total number of evaluation steps.
+
+FAP is the final performance of a model after seeing the entire video. That is,  $\mathrm{FAP} = \mathrm{CAP}_{t_N}$ , where  $t_N$  denotes the last evaluation step.
+
+<table><tr><td rowspan="2"></td><td rowspan="2">Annotation Cost</td><td colspan="3">OAK</td><td colspan="3">EgoObjects</td></tr><tr><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td></tr><tr><td>Incremental</td><td>100%</td><td>8.38</td><td>7.72</td><td>0.03</td><td>10.21</td><td>3.55</td><td>1.48</td></tr><tr><td>Offline Training</td><td>100%</td><td>48.28</td><td>35.23</td><td>-</td><td>86.18</td><td>59.81</td><td>-</td></tr><tr><td>EWC</td><td>100%</td><td>7.73</td><td>7.02</td><td>-0.12</td><td>5.15</td><td>1.60</td><td>0.57</td></tr><tr><td>iOD</td><td>100%</td><td>7.92</td><td>7.14</td><td>0.98</td><td>8.80</td><td>2.64</td><td>0.00</td></tr><tr><td>iCaRL(Wang et al.)</td><td>100%</td><td>22.89</td><td>16.60</td><td>-2.95</td><td>37.61</td><td>21.71</td><td>2.79</td></tr><tr><td>iCaRL(our impl.)</td><td>100%</td><td>36.14</td><td>26.26</td><td>-4.89</td><td>60.80</td><td>36.41</td><td>-0.60</td></tr><tr><td rowspan="2">w/ Efficient-CLS</td><td>25%</td><td>38.36(+2.22)</td><td>26.64(+0.38)</td><td>-8.20(-3.31)</td><td>61.26(+0.46)</td><td>39.58(+3.17)</td><td>-3.48(-2.88)</td></tr><tr><td>100%</td><td>40.24(+4.10)</td><td>28.18(+1.92)</td><td>-8.10(-3.21)</td><td>67.05(+6.25)</td><td>40.36(+3.95)</td><td>-3.67(-3.07)</td></tr><tr><td>A-GEM</td><td>100%</td><td>36.94</td><td>26.19</td><td>-5.54</td><td>58.79</td><td>35.88</td><td>-8.38</td></tr><tr><td rowspan="2">w/ Efficient-CLS</td><td>25%</td><td>37.06(+0.12)</td><td>26.36(+0.17)</td><td>-7.76(-2.22)</td><td>63.06(+4.27)</td><td>39.46(+3.58)</td><td>-7.49(+0.89)</td></tr><tr><td>100%</td><td>39.87(+2.93)</td><td>27.97(+1.78)</td><td>-7.17(-1.63)</td><td>66.94(+8.15)</td><td>39.57(+3.69)</td><td>-11.68(-3.30)</td></tr><tr><td>GDumb</td><td>100%</td><td>35.27</td><td>25.29</td><td>-6.59</td><td>58.85</td><td>36.38</td><td>-5.21</td></tr><tr><td rowspan="2">w/ Efficient-CLS</td><td>25%</td><td>37.67(+2.40)</td><td>25.59(+0.30)</td><td>-9.30(-2.71)</td><td>62.70(+3.85)</td><td>38.78(+2.40)</td><td>-8.86(-3.65)</td></tr><tr><td>100%</td><td>38.61(+3.34)</td><td>26.04(+0.75)</td><td>-9.14(-2.55)</td><td>63.55(+4.70)</td><td>38.98(+2.60)</td><td>-7.50(-2.29)</td></tr><tr><td>DER++</td><td>100%</td><td>37.79</td><td>25.24</td><td>-2.87</td><td>55.82</td><td>30.84</td><td>-6.08</td></tr><tr><td rowspan="2">w/ Efficient-CLS</td><td>25%</td><td>37.93(+0.14)</td><td>25.64(+0.4)</td><td>-8.90(-6.03)</td><td>59.70(+3.88)</td><td>34.15(+3.31)</td><td>-11.21(-5.13)</td></tr><tr><td>100%</td><td>39.61(+1.82)</td><td>26.73(+1.49)</td><td>-8.30(-5.43)</td><td>62.01(+6.19)</td><td>33.09(+2.25)</td><td>-11.05(-4.97)</td></tr></table>
+
+Table 1. Overall performance of existing algorithms and Efficient-CLS on OAK and EgoObjects. iCaRL(Wang et al.) denotes the SOTA model presented in [34], and iCaRL(our impl.) is the same method by our implementation.
+
+$\mathbf{F}$  estimates the forgetfulness of the model due to the sequential training. It takes into account the time interval between the presence of an object category and its subsequent presence. For a class  $c$ , we sort the  $\mathrm{CAP}_{t_i}^c$  according to the time interval  $k$  between evaluation time  $t_i$  and the last time  $t_i - k$  the model is trained on  $c$ . After  $\mathrm{CAP}_{t_i}^c$  is sorted, all  $\mathrm{CAP}_{t_i}^c (i = 0,\dots ,T)$  are divided into  $K$  bins  $B_{kmin},\dots ,B_{kmax}$  according to the time interval  $k$ . The average CAP  $(\mathrm{aCAP}_k)$  of each bin  $B_{k}$  is defined as the model's performance for detecting class  $c$  after the model has not been trained on  $c$  for  $k$  time steps. The forgetfulness (F) of the class  $c$  is defined as the weighted sum of the performance decrease at each time:
+
+$$
+F ^ {c} = \sum_ {k = k m i n} ^ {k m a x} \frac {k - k m i n}{\sum_ {k = k m i n} ^ {k m a x} k - k m i n} \times \left(\mathrm {a C A P} _ {k m i n} - \mathrm {a C A P} _ {k}\right). \tag {5}
+$$
+
+The overall forgetfulness is:  $\mathrm{F} = \frac{1}{C}\sum_{c = 0}^{C}\mathrm{F}^{c}$
+
+# 4.4. Implementation Details
+
+For a fair comparison, we followed the prior work [34] to use Faster-RCNN [27] with ResNet-50 backbone [12] as our object detection network, which is initialized by the weights pre-trained on PASCAL VOC [10]. We used Adam optimizer with a constant learning rate of 0.0001, and the batch size was set to 16 frames. Same as [34], we maintained a replay buffer with 5 samples per class. At each time step  $t$ , we first randomly retrieved 16 video frames from the replay buffer for joint training. We used confidence thresh  $\tau = 0.7$  to generate pseudo-labels for unlabeled frames, and applied EMA rate  $\alpha = 0.99$  to update the slow learner. The weight of pseudo loss  $\lambda_{pseudo}$  was set to 1.0. All the experiments were conducted on 2 NVIDIA RTX 3090 GPUs.
+
+# 5. Results
+
+# 5.1. Fully Supervised Protocol
+
+As the previous work [34] focuses on online continual object detection (OCOD) in video streams, we first evaluated model performance in fully supervised setting (i.e.,  $100\%$  annotation cost), where all video frames are paired with human labels. The reported results were measured by standard metrics (CAP, FAP, and F, Section 4.3) in Table 1.
+
+Performance of CL baselines. [34] benchmarked Incremental, EWC, iCaRL(Wang et al.), and Offline Training on OAK dataset. They found the replay-based method (i.e. iCaRL(Wang et al.)) outperforms regularization-based method (i.e. EWC) by  $10\%$  in FAP, while iCaRL(Wang et al.) has a huge gap of  $30\%$  compared with Offline Training. Similar observations were made in Table 1, but the performance of Incremental and EWC were  $4\%$  lower than that in [34], as in our setting we only trained each mini-batch of video frames once (they trained each mini-batch 10 times). As mentioned in Section 4.2, we introduce several variations to the original design of iCaRL(Wang et al.). Compared with iCaRL(Wang et al.), we observed a huge performance boost in FAP from  $22.89\%$  to  $36.14\%$  on OAK and from  $37.61\%$  to  $60.80\%$  on EgoObjects, abridging the gap between the baseline and Offline Training.
+
+We further adapted other standard CL baselines, including iOD, A-GEM, GDumb, DER++, to the OCOD setting for comparisons. iOD is the SOTA method in offline class-incremental object detection. Though it performs well in prior setting, iOD collapses when adapted to online video streams. One explanation is that iOD requires explicit task boundary to trigger the reshape of model gradients that opti
+
+![](images/33fe0a64f37aecc71768e8c1c4c8ff86bb795b645e3f00e83a30203a3e49b417.jpg)  
+Figure 3. The changes of  $\mathrm{CAP}_{t_i}^c$  with sampled categories on OAK dataset. The x-axis denotes time step across the entire video stream. The y-axis denotes the AP50 of the category at specific time step (i.e.,  $\mathrm{CAP}_{t_i}^c$ ). The grey line indicates the existence of the category.
+
+mizes knowledge sharing between adjacent tasks. However, in online video streams, the task boundaries by classes are no longer available and the change of tasks is hard to identify, resulting in the failure to prevent forgetting. In contrast, replay-based methods (i.e., A-GEM, GDumb and DER++) generalize much better to the real-world streaming video.
+
+Performance w/ Efficient-CLS. Inheriting from the benefit of fast and slow learning with EMA, our Efficient-CLS consistently improves all the state-of-the-art CL methods (i.e. iCaRL(our impl.), A-GEM, GDumb and DER++) by a significant margin. Taking iCaRL(our impl.) for example, with Efficient-CLS, we observed superior performance and minimal forgetting compared to baseline methods, even when categories appeared infrequently (e.g., sculpture in Figure 3). Since semantic contextual information is more important in indoor environments on EgoObjects compared to the outdoor environments in OAK, we noticed that the improvement brought by our method is even greater on EgoObjects with an increase of  $6.25\%$  in FAP,  $3.95\%$  in CAP  $\%$  and  $3.07\%$  in F. Consistent performance gains are also noticeable for A-GEM, GDumb and DER++, demonstrating the effectiveness and flexibility of our method.
+
+# 5.2. Sparse Annotation Protocol
+
+Performance of CL baselines. The sparse annotation protocol is more challenging than the previous fully supervised protocol as shown by the performance differences when number of annotated video frames decreases (compare the performance of each colored bar along the x-axis within each subplot in Figure 4). We noted that GDumb is more resilient against the reduce of supervision. Specifically, from Table 2 at the lowest annotation cost of  $6.25\%$  on EgoObjects dataset, GDumb achieves the highest performance of  $38.74\%$ ,  $22.69\%$  and  $-4.53\%$  in FAP, CAP and F, which surpasses other baselines by a considerable margin. Same observations can be made on OAK dataset. One possible explanation is that GDumb only trains the data stored in the balanced replay buffer, which makes it less vulnerable
+
+<table><tr><td rowspan="2"></td><td colspan="3">OAK</td><td colspan="3">EgoObjects</td></tr><tr><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td></tr><tr><td>iCaRL</td><td>23.04</td><td>17.75</td><td>-3.31</td><td>32.91</td><td>19.15</td><td>-1.36</td></tr><tr><td>w/ Efficient-CLS</td><td>29.72</td><td>20.31</td><td>-5.36</td><td>40.16</td><td>23.95</td><td>-2.01</td></tr><tr><td>A-GEM</td><td>23.59</td><td>16.15</td><td>-2.99</td><td>21.84</td><td>12.28</td><td>0.82</td></tr><tr><td>w/ Efficient-CLS</td><td>30.18</td><td>20.59</td><td>-5.44</td><td>38.96</td><td>23.79</td><td>-5.64</td></tr><tr><td>GDumb</td><td>27.37</td><td>19.64</td><td>-4.25</td><td>38.74</td><td>22.69</td><td>-4.53</td></tr><tr><td>w/ Efficient-CLS</td><td>29.07</td><td>19.99</td><td>-6.01</td><td>40.09</td><td>23.67</td><td>-5.16</td></tr><tr><td>DER++</td><td>24.21</td><td>15.93</td><td>-3.79</td><td>16.95</td><td>8.48</td><td>2.03</td></tr><tr><td>w/ Efficient-CLS</td><td>28.63</td><td>19.64</td><td>-4.60</td><td>35.78</td><td>20.74</td><td>-4.69</td></tr></table>
+
+Table 2. Performance of Efficient-CLS at low annotation cost of  $6.25\%$ . The best results are bold-faced.
+
+able to class imbalance problem brought by the reduce of labeled samples in the training set.
+
+Performance w/ Efficient-CLS. Our proposed Efficient-CLS is a plug-and-play module that can be easily inserted into and enhance existing continual learning algorithms with the ability to use unlabeled video frames effectively. In both OAK and EgoObjects dataset, Efficient-CLS consistently improves the comparative SOTAs in all three evaluation metrics regardless of various degrees of annotation cost. As shown in Table 2, at a lower annotation cost of  $6.25\%$ , Efficient-CLS doubles the performance of DER++ and A-GEM in terms of FAP and CAP, and achieves an even larger improvement in preventing forgetting. Thanks to the useful information from pseudo labels predicted by the slow learner in Efficient-CLS, our method is more robust to various annotation costs, compared with SOTAs (compare the rate of change of blue bars vs. red bars over different degrees of annotation cost). Most remarkably, Efficient-CLS with  $25\%$  annotation cost has already outperformed comparative SOTAs with  $100\%$  annotation cost (see Table 1).
+
+# 5.3. Ablation Study
+
+We assessed the importance of our key design choices. The Complementary Learning Systems (CLS) in EfficientCLS is the key for rapidly adapting to learn new tasks, meanwhile, retaining previously learnt knowledge. It constitutes of two memory reinstatement mechanisms: one is
+
+![](images/dafc29ad3590cc3cfb2a7d358b163bca09c7a2127ab71373a07b5c75f8c90e61.jpg)
+
+![](images/c5dc86c1a627513b4e84ccece2bd45bd5df1c9042492552d7dca3c95edc54efc.jpg)
+
+![](images/1ed6f3578d2d9af0843954a429ad7af6532263ab6fad0360d4c62a4e0a890394.jpg)
+
+![](images/24693d7cad2f36899b31c8f33ce9bcb2974c9d1836f75eaf87ea557cde227c54.jpg)  
+Figure 4. Evaluation of online continual object detection in video streams with three metrics (FAP, CAP and F, Section 4.3) on OAK dataset (first row) and EgoObjects dataset (second row). The higher the bars are, the better. The x-axis denotes the percentage of video frames that are labeled in the video stream. It ranges from  $6.25\%$  to  $100\%$  (full supervision). The y-axis indicates the performance using different evaluation metrics. Ours (iCaRL(our impl.) w/ Efficient-CLS, red) consistently beats the comparative SOTA (iCaRL(our impl.), blue) in all evaluation metrics.
+
+![](images/11853cb1106fcd715fd04101d300b2bdf81ad71ec85abd0a3a6aa42ea914aa0d.jpg)
+
+![](images/f73a9a17b60c67d99d85868822c5329faa2468187a6784f269ed11e0fc04b41f.jpg)
+
+<table><tr><td rowspan="2">EMA</td><td rowspan="2">PL</td><td colspan="3">50%</td><td colspan="3">25%</td><td colspan="3">12.5%</td><td colspan="3">6.25%</td></tr><tr><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td><td>FAP (↑)</td><td>CAP (↑)</td><td>F (↓)</td></tr><tr><td>X</td><td>X</td><td>34.68</td><td>25.78</td><td>-4.15</td><td>33.70</td><td>24.57</td><td>-4.30</td><td>28.76</td><td>19.80</td><td>-5.48</td><td>23.04</td><td>17.75</td><td>-3.31</td></tr><tr><td>✓</td><td>X</td><td>35.74</td><td>25.77</td><td>-4.82</td><td>34.79</td><td>25.62</td><td>-4.35</td><td>31.72</td><td>21.16</td><td>-7.24</td><td>27.84</td><td>20.03</td><td>-3.96</td></tr><tr><td>X</td><td>✓</td><td>35.61</td><td>25.56</td><td>-3.76</td><td>34.95</td><td>25.65</td><td>-3.65</td><td>31.60</td><td>22.44</td><td>-4.83</td><td>26.39</td><td>19.50</td><td>-1.99</td></tr><tr><td>✓</td><td>✓</td><td>38.61</td><td>26.90</td><td>-7.29</td><td>38.36</td><td>26.64</td><td>-8.20</td><td>33.92</td><td>23.04</td><td>-7.71</td><td>29.72</td><td>20.31</td><td>-5.36</td></tr></table>
+
+Table 3. Effectiveness of Exponential Moving Average (EMA) and Pseudo-labeling (PL) for iCaRL (our impl.) on OAK dataset at annotation cost  $50\%$ ,  $25\%$ ,  $12.5\%$  and  $6.25\%$ . The best results are bold-faced.
+
+synaptic weight transfer from fast to slow learner via exponential moving average (EMA); and the other is reciprocal replay from slow learner to fast learner with pseudolabeling (PL). Here we studied their effects individually.
+
+Ablation: Exponential Moving Average (EMA). We removed EMA by setting the  $\alpha$  in Equation 3 to 1, where the model weights of the fast learner and slow learner are now shared throughout the learning process. Note that in the fully supervised protocol, the pseudo-labeling is turned off and the Efficient-CLS equals to the EMA alone. From Table 1 at  $100\%$  annotation cost, we observed that removing EMA leads to significant performance drops ranging from  $2\%$  to  $8\%$ , for all the comparative baselines on both OAK and EgoObjects. This shows that the slow learner can effectively consolidate the knowledge from the fast learner, and constructively alleviate catastrophic forgetting by synapses consolidation over time. Similar observations were made in Table 3 in the sparse annotation protocol (compare Row 2 vs. Row 1, Row 4 vs. Row 3). It is worth noting that, the performance difference between naive model (Row 1) and its variant with EMA (Row 2) is slightly larger in lower su
+
+pervision (i.e.  $12.5\%$ ,  $6.25\%$ ) than higher supervision (i.e.  $50\%$ ,  $25\%$ ). One possible reason is that, compared with higher supervision, the fast learner suffers more forgetting in lower supervision; hence, the effect of removing EMA becomes stronger in lower supervision, again highlighting the importance of EMA.
+
+Ablation: Pseudo-labeling (PL). We ablate our model by removing the PL of the slow learner across varying annotation costs and reported the results in Table 3. The removal of PL (Row 2) leads to a performance drop of around  $2\%$  in FAP,  $1\%$  in CAP and  $0.5 - 4\%$  in F, compared with our full Efficient-CLS (Row 4). It implies that the slow learner captures useful semantic information from unlabeled video frames and these predicted pseudo labels are helpful in training the fast learner.
+
+To investigate whether pseudo labels predicted by the fast learner itself could help stream learning, we conducted another ablation experiment where we performed PL without EMA (Row 3). Compared with the naive model (Row 1), we observed a performance increase from  $28.76\%$  to  $31.60\%$  in FAP and  $19.80\%$  to  $22.44\%$  in CAP. It indicates
+
+that, due to the temporal correlation in video stream, pseudo labels predicted by the fast learner can serve as an informative supervision for the training of the fast learner itself.
+
+However, replaying the self-predicted pseudo labels on the fast learner fails to prevent forgetting, as indicated by the drop from  $-5.48\%$  to  $-4.83\%$  in F. It is possible that the pseudo labels generated by the fast learner only bias towards the classes which have already been learnt very well and fail to reinforce the fast learner to improve on the poorly-learnt classes. Different from the fast learner, the slow learner integrates semantic information over time. The predicted pseudo labels carry more semantic information, which is useful for fast learner to capture more generic object representations during pseudo label replays. Again, this emphasizes that the reciprocal replay from the slow learner to the fast learner is critical for memory reinstatement, which has been missing in the computational modeling literature of CLS.
+
+# 6. Conclusion
+
+To imitate what humans see and learn in the real world, we introduce a more realistic and challenging problem of label-efficient online continual object detection (LEOCOD) in video streams. Addressing this problem would greatly benefit real-world applications by reducing model retraining time and data labeling costs. Inspired by the Complementary Learning Systems (CLS) theory, we propose a plug-and-play module, namely Efficient-CLS, that can be easily integrated into and improve existing continual leaning algorithms. We evaluate Efficient-CLS on two challenging real-world video datasets, where our method achieves the state-of-the-art performance in preventing catastrophic forgetting, all while requiring minimal annotation effort.
+
+Acknowledgements This research is funded by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008 & NRF-NRFF15-2023-0001, AI Singapore Programme (AISG Award No: AISGC-2019-001-2A & AISG2-RP-2021-025), Mike Zheng Shou's Start-Up Grant from NUS, Mengmi Zhang's Startup Grant from Agency for Science, Technology, and Research (A*STAR), and Early Career Investigatorship from Center for Frontier AI Research (CFAR), A*STAR. The computational work for this article was partially performed on resources of the National Supercomputing Centre, Singapore.
+
+# References
+
+[1] Egoobjects dataset. https://ai.facebook.com/datasets/egoobjects-dataset/, May 2022. 2, 5  
+[2] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered
+
+retrieval. Advances in neural information processing systems, 32, 2019. 1, 3  
+[3] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Advances in neural information processing systems, 32, 2019. 1, 2, 3  
+[4] Elahe Arani, Fahad Sarfraz, and Bahram Zonooz. Learning fast, learning slow: A general continual learning method based on complementary learning system. arXiv preprint arXiv:2201.12604, 2022. 2, 3, 5  
+[5] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 3  
+[6] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 2, 3, 5  
+[7] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuytelaars, Joelle Pineau, and Eugene Belilovsky. New insights on reducing abrupt representation change in online continual learning. arXiv preprint arXiv:2203.03798, 2022. 1, 2  
+[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 2, 3, 5  
+[9] Hung-Jen Chen, An-Chieh Cheng, Da-Cheng Juan, Wei Wei, and Min Sun. Mitigating forgetting in online continual learning via instance-aware parameterization. Advances in Neural Information Processing Systems, 33:17466-17477, 2020. 2  
+[10] Mark Everingham, SM Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111(1):98-136, 2015. 6  
+[11] Enrico Fini, Stéphane Lathuiliere, Enver Sangineto, Moin Nabi, and Elisa Ricci. Online continual learning under extreme memory constraints. In European Conference on Computer Vision, pages 720-735. Springer, 2020. 1  
+[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6  
+[13] Jisoo Jeong, Seungeui Lee, Jeesoo Kim, and Nojun Kwak. Consistency-based semi-supervised learning for object detection. Advances in neural information processing systems, 32, 2019. 3  
+[14] Daoyun Ji and Matthew A Wilson. Coordinated memory replay in the visual cortex and hippocampus during sleep. Nature neuroscience, 10(1):100-107, 2007. 2  
+[15] Nitin Kamra, Umang Gupta, and Yan Liu. Deep generative dual memory network for continual learning. arXiv preprint arXiv:1710.10368, 2017. 3  
+[16] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 5
+
+[17] Joseph Kj, Jathushan Rajasegaran, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Incremental object detection via meta-learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 5  
+[18] Dharshan Kumaran, Demis Hassabis, and James L McClelland. What learning systems do intelligent agents need? complementary learning systems theory updated. Trends in cognitive sciences, 20(7):512-534, 2016. 1  
+[19] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn and think like people. Behavioral and brain sciences, 40, 2017. 1  
+[20] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. arXiv preprint arXiv:2102.09480, 2021. 3  
+[21] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2  
+[22] Lorenzo Pellegrini, Chenchen Zhu, Fanyi Xiao, Zhicheng Yan, Antonio Carta, Matthias De Lange, Vincenzo Lomonaco, Roshan Sumbaly, Pau Rodriguez, and David Vazquez. 3rd continual learning workshop challenge on egocentric category and instance level object understanding. arXiv preprint arXiv:2212.06833, 2022. 5  
+[23] Quang Pham, Chenghao Liu, and Steven Hoi. Dualnet: Continual learning, fast and slow. Advances in Neural Information Processing Systems, 34, 2021. 3  
+[24] Quang Pham, Chenghao Liu, Doyen Sahoo, and HOI Steven. Contextual transformation networks for online continual learning. In International Conference on Learning Representations, 2020. 2, 3  
+[25] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In European conference on computer vision, pages 524-540. Springer, 2020. 2, 3, 5  
+[26] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 2, 3, 5  
+[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 4, 6  
+[28] Mohammad Rostami, Soheil Kolouri, and Praveen K Pilly. Complementary learning for overcoming catastrophic forgetting using experience replay. arXiv preprint arXiv:1903.04566, 2019. 3  
+[29] Dongsub Shim, Zheda Mai, Jihwan Jeong, Scott Sanner, Hyunwoo Kim, and Jongseong Jang. Online class incremental continual learning with adversarial shapley value. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 9630-9638, 2021. 2, 3  
+[30] Krishna Kumar Singh, Kayvon Fatahalian, and Alexei A Efros. Krishnacam: Using a longitudinal, single-person,
+
+egocentric dataset for scene understanding tasks. In 2016 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1-9. IEEE, 2016. 5  
+[31] James Smith, Jonathan Balloch, Yen-Chang Hsu, and Zsolt Kira. Memory-efficient semi-supervised continual learning: The world is its own replay buffer. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 3  
+[32] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 3  
+[33] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3  
+[34] Jianren Wang, Xin Wang, Yue Shang-Guan, and Abhinav Gupta. Wanderlust: Online continual object detection in the real world. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10829-10838, 2021. 1, 2, 3, 5, 6  
+[35] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020. 1  
+[36] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14421-14430, 2022. 3  
+[37] Xiangli Yang, Zixing Song, Irwin King, and Zenglin Xu. A survey on deep semi-supervised learning. IEEE Transactions on Knowledge and Data Engineering, 2022. 3  
+[38] Barret Zoph, Ekin D Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V Le. Learning data augmentation strategies for object detection. In European conference on computer vision, pages 566-583. Springer, 2020. 4

2023/Label-Efficient Online Continual Object Detection in Streaming Video/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe6b37daa288956d62bd9d8dd3a263202303a3fba0ce74cee6dc13d1a84a6d31
+size 505436

2023/Label-Free Event-based Object Recognition via Joint Learning with Image Reconstruction from Events/cebb8d22-3e6f-4881-895d-c913e7109230_content_list.json

@@ -0,0 +1,1666 @@
+[
+    {
+        "type": "text",
+        "text": "Label-Free Event-based Object Recognition via Joint Learning with Image Reconstruction from Events",
+        "text_level": 1,
+        "bbox": [
+            142,
+            161,
+            826,
+            204
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Hoonhee Cho*, Hyeonseong Kim*, Yujeong Chae, and Kuk-Jin Yoon",
+        "bbox": [
+            208,
+            228,
+            759,
+            245
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Korea Advanced Institute of Science and Technology",
+        "bbox": [
+            273,
+            245,
+            695,
+            262
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{gnsgnsgml,brian617,yujeong,kjyoon}@kaist.ac.kr",
+        "bbox": [
+            275,
+            263,
+            692,
+            277
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            308,
+            312,
+            324
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recognizing objects from sparse and noisy events becomes extremely difficult when paired images and category labels do not exist. In this paper, we study label-free event-based object recognition where category labels and paired images are not available. To this end, we propose a joint formulation of object recognition and image reconstruction in a complementary manner. Our method first reconstructs images from events and performs object recognition through Contrastive Language-Image Pretraining (CLIP), enabling better recognition through a rich context of images. Since the category information is essential in reconstructing images, we propose category-guided attraction loss and category-agnostic repulsion loss to bridge the textual features of predicted categories and the visual features of reconstructed images using CLIP. Moreover, we introduce a reliable data sampling strategy and local-global reconstruction consistency to boost joint learning of two tasks. To enhance the accuracy of prediction and quality of reconstruction, we also propose a prototype-based approach using unpaired images. Extensive experiments demonstrate the superiority of our method and its extensibility for zero-shot object recognition. Our project code is available at https://github.com/Chohoonhee/Ev-LaFOR.",
+        "bbox": [
+            75,
+            338,
+            473,
+            669
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            77,
+            684,
+            208,
+            698
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Event cameras are neuromorphic vision sensors that asynchronously perceive per-pixel brightness changes with high temporal resolution. They show advantages over conventional frame-based cameras, such as high dynamic range (HDR) and low latency. However, despite the superiority of cameras, algorithms using events are still in their infancy. Existing event-based object recognition methods [1, 4, 22, 24, 25, 26, 30, 32, 34, 36, 40, 53] have been shown that object recognition is a crucial task for event feature learning [12, 29, 38] and endeavors have successfully",
+        "bbox": [
+            75,
+            707,
+            468,
+            845
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/0284fba41f1963b2940b0ca24aafc41c889a0aaf349f6e097983639d6b731466.jpg",
+        "image_caption": [
+            "Figure 1: We tackle label-free event-based object recognition where category labels and paired images are not available. Our approach can simultaneously perform object recognition and image reconstruction through proposed joint learning framework. Optionally, our method can leverage unpaired images to enhance object recognition performance and reconstructed image quality."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            507,
+            307,
+            884,
+            501
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "classified the object with events. However, they rely on supervision with category labels to achieve high performance. Furthermore, the sparsity of events poses a challenge for users to label them accurately when there are no paired images. Several unsupervised methods [25, 36] have been proposed for object recognition, however, their performances remain subpar. Hence, there is a natural need for robust and effective label-free event-based object recognition methods.",
+        "bbox": [
+            497,
+            640,
+            892,
+            751
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we study label-free event-based object recognition where category labels and paired images are not available as in Fig. 1. To tackle this task, we focus on two following facts. Firstly, as demonstrated in [45, 60], performing object recognition on reconstructed images with spatially dense information and rich context shows better performance than solely using events. Secondly, during reconstructing images from events, taking the categories into",
+        "bbox": [
+            497,
+            755,
+            892,
+            866
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            129,
+            49,
+            200,
+            85
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation.",
+        "bbox": [
+            253,
+            48,
+            789,
+            61
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "Except for this watermark, it is identical to the accepted version;",
+        "bbox": [
+            332,
+            62,
+            709,
+            74
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            304,
+            74,
+            739,
+            87
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*The first two authors contributed equally. In alphabetical order.",
+        "bbox": [
+            95,
+            853,
+            433,
+            866
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "19866",
+        "bbox": [
+            465,
+            887,
+            505,
+            897
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "account can improve the quality of the reconstructed images. From these observations, we propose a joint learning framework of two closely-linked tasks: event-based object recognition and event-to-image reconstruction.",
+        "bbox": [
+            75,
+            125,
+            468,
+            181
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In our joint learning framework, we first reconstruct images from events and perform object recognition on the reconstructed images using Contrastive Language-Image Pretraining (CLIP) [42]. CLIP is an image-text embedding model that has been trained on a large number of image and language pairs sourced from the web. We utilize the textual features encoded from CLIP since they are well-aligned with visual features, which is advantageous for our tasks that lack labels and images. To this end, we propose the category-guided attraction loss that aligns the visual features from reconstructed images close to the texture features of the predicted category from CLIP.",
+        "bbox": [
+            75,
+            184,
+            468,
+            349
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Although the above framework enables the joint learning of event-based object recognition and image reconstruction, using the predicted categories brings two issues. (1) Collapsing to the trivial solution, resulting in all outputs being assigned to a single category and (2) Interrupting the learning process due to unreliable (wrong) predictions. To alleviate these problems, we propose (1) category-agnostic repulsion loss to increase the distances between visual features for preventing collapse and (2) a reliable data sampling (RDS) method to reduce the effects of unreliable predictions. For RDS, we devise two reliable indicators named posterior probability indicator (PPI) and temporally reversed consistency indicator (TRCI). The PPI selects reliable samples based on the probability of prediction. For additional sampling, TRCI utilizes the characteristics of events that temporally reversed ones should be categorized into the same category as the original events. This reliable data sampling enables stable learning even with the predicted categories, leading to a substantial performance improvement in object recognition. Furthermore, to restore the local details, we introduce the local-global reconstruction consistency. Finally, to enhance the accuracy of recognition and visual quality of reconstructed images, we expand our method by employing the existing unpaired images. Extensive experiments demonstrate that our method shows outstanding results on the subset of the N-ImageNet dataset over unsupervised and image reconstruction-based approaches and even surpasses the supervised methods on the N-Caltech101 dataset without using labels and images. In addition, we show that our framework can be extended to event-based zero-shot object recognition.",
+        "bbox": [
+            75,
+            351,
+            468,
+            781
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our work presents four main contributions: (I) We propose a novel joint learning framework for event-based object recognition and image reconstruction, without requiring paired images and labels. (II) We introduce a reliable data sampling strategy and local-global reconstruction consistency that enhances label-free joint training. (III) We",
+        "bbox": [
+            75,
+            783,
+            468,
+            866
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "further introduce employing unpaired images in our framework for boosting performance. (IV) We conduct extensive experiments of our method along with zero-shot object recognition and superset categories learning, and demonstrate our method's superior performance and effectiveness.",
+        "bbox": [
+            497,
+            125,
+            890,
+            195
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            499,
+            207,
+            640,
+            220
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Event-based Object Recognition. Concurrent with the success of event-based low-level tasks [9, 10, 11, 23, 52, 54, 63], there has been a surge of active research focused on object recognition [3, 7, 19, 22, 27, 28, 31, 35, 43, 51, 53, 56], which is considered as a fundamental task. Considering the unique features of event data, numerous works [1, 4, 22, 24, 25, 26, 30, 32, 34, 36, 40, 53] have presented models and representations that can effectively extract significant patterns. However, these approaches require labels for each event data for supervised learning, and since event data cannot be crawled from the web like images, the labeling cost of event data is extremely high. On the other hand, our approach enables the learning of event-based object recognition and image reconstruction simultaneously, without requiring category labels and paired images.",
+        "bbox": [
+            497,
+            229,
+            890,
+            435
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Event-based Image Reconstruction is a well-known topic in the field of event-based vision [2, 13, 17, 20, 21, 46, 49, 57]. Recent researches have been utilizing deep learning to generate realistic outcomes [44, 45, 50, 62]. However, these supervised approaches require the paired images of each event stream to be accurately aligned and synchronized in terms of pixels and time. Another line of approach, such as Wang et al. [59] and Pini et al. [41], utilized generative adversarial networks (GANs) to reconstruct intensity with unpaired images that are not related to the events of the scene. However, GANs are prone to be sensitive to various parameters, and when images that are significantly irrelevant to objects are utilized, the quality of the reconstructed images from events is notably low. An alternative solution is a self-supervised method [37] that utilizes photometric consistency, but this approach requires a continuous sequence that spans an adequately long period. Differently, our framework can reconstruct the intensity of event data without images, very long event sequence and GANs. We only exploit the category prompts of a bundle of event data, not the class label of each event data.",
+        "bbox": [
+            497,
+            437,
+            890,
+            726
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Label-free Learning aims to solve the end-task only with unlabeled data, and has been actively explored [8, 58]. HOTS [25] proposed a category-free method for recognizing objects based on events and SSL-E2VID [37] proposed a self-supervised learning method for reconstructing images from only events. However, their performances fall short of supervised learning methods with significant performance gaps. Inspired by the success of using the pretrained model [5, 6, 39, 48, 55], we adopt CLIP [42] for label-free event-based object recognition. We propose a",
+        "bbox": [
+            497,
+            726,
+            890,
+            866
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "19867",
+        "bbox": [
+            465,
+            887,
+            504,
+            897
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/b9c0ce00c40f448ae65cf3529781931084efaa6040745f51c21d4fd1af0893e9.jpg",
+        "image_caption": [
+            "Figure 2: Overall joint learning framework of event-based object recognition with event-to-image reconstruction. Our approach involves CLIP's textual and visual encoders that derive textual and visual features, respectively, from given class sets and events. Textual features are obtained from category prompts, while visual features are obtained from reconstructed images. Then, logits and corresponding pseudo labels are generated by computing the cosine similarity between visual and textual features. Lastly, our framework employs pseudo labels to facilitate joint learning using category-agnostic repulsion loss and category-guided attraction loss. To mitigate the impact of unreliable pseudo labels and boost performance, we introduce a reliable data sampling strategy and local-global reconstruction consistency."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            119,
+            887,
+            391
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "novel joint learning framework that uses visual and textual knowledge to achieve superior performance in event-to-image reconstruction and event-based object recognition without category labels and paired images. To the best of our knowledge, we are the first to incorporate textual knowledge into event vision research, particularly in the context of label-free tasks.",
+        "bbox": [
+            75,
+            505,
+            470,
+            602
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            77,
+            612,
+            169,
+            625
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In Sec. 3.1, we first introduce our joint formulation of event-to-image reconstruction and object recognition for label-efficient learning. In Sec. 3.2, we describe our reliable data sampling (RDS) strategy for boosting joint learning performance by filtering the unreliable data samples during training. Sec. 3.3 introduces the local-global reconstruction consistency for spatially regularizing the reconstructed images, which ensures capturing the local details. In Sec. 3.4, we propose to further employ the unpaired images for better recognition performance and reconstructed image quality.",
+        "bbox": [
+            75,
+            633,
+            468,
+            771
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Joint Formulation",
+        "text_level": 1,
+        "bbox": [
+            77,
+            777,
+            253,
+            790
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We propose a joint formulation of event-to-image reconstruction and object recognition to perform event-based (zero-shot) object recognition without using labels and paired images of event data. Our joint formulation is based on the fact that 1) image reconstruction can be helpful in",
+        "bbox": [
+            75,
+            796,
+            470,
+            866
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "predicting categories from events, and 2) category information is important in reconstructing images from events. Therefore, we propose 1) the process of performing event-to-image reconstruction and CLIP-driven object recognition from the reconstructed images and 2) category-guided attraction loss and category-agnostic repulsion loss for learning this process without labels and paired images. The overall framework is shown in Fig. 2. Note that our joint learning framework can reconstruct intensity images from events without using any images.",
+        "bbox": [
+            497,
+            505,
+            890,
+            643
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Event-to-Image Reconstruction. For the joint formulation, we first reconstruct the intensity image  $\\mathcal{I}$  from the event stream  $\\mathcal{E} = (u_i,t_i,p_i)_i = 1$ , where  $u_{i},t_{i}$ , and  $p_i$  denote triggered pixel location, time, and polarity of each event, and  $N$  is the temporal length of the event stream. We convert the event stream into event spiking tensor (EST) [16], which is then fed into the image reconstruction network  $\\mathcal{G}$  to obtain the intensity image. The image reconstruction process is formulated as follows:",
+        "bbox": [
+            497,
+            643,
+            892,
+            767
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {I} = \\mathcal {G} (\\operatorname {E S T} (\\mathcal {E})). \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            633,
+            775,
+            890,
+            789
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "CLIP-driven Object Recognition. We perform object recognition on top of the reconstructed images instead of raw event data. The advantage of using reconstructed images rather than using event data directly for object recognition is that better performance can be achieved by using",
+        "bbox": [
+            497,
+            796,
+            892,
+            866
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "19868",
+        "bbox": [
+            465,
+            887,
+            505,
+            897
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "a network trained on rich image data [45, 60]. In particular, we use CLIP [42] for object recognition, which also enables zero-shot object recognition. For an event dataset of  $C$  classes, the category prompts are constructed by placing each category name into a pre-defined template, \"image of a [CLASS].\" Then, CLIP's textual encoder encodes category prompts to obtain the  $D$ -dimensional textual features  $F = (f_{i})_{i = 1}^{C}\\in \\mathbb{R}^{C\\times D}$ . Meantime, the reconstructed image  $\\mathcal{I}$  is encoded by CLIP's visual encoder to a visual feature  $v\\in \\mathbb{R}^{1\\times D}$  and the class probability  $p\\in \\mathbb{R}^{1\\times C}$  for the object recognition are obtained as,",
+        "bbox": [
+            75,
+            125,
+            472,
+            277
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {i} = \\operatorname {s o f t m a x} _ {i} \\left(v F ^ {\\top}\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            196,
+            285,
+            468,
+            302
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $p_i$  and  $\\mathrm{softmax}_i(\\cdot)$  denote the probability for category  $i$  and the softmax function. By taking the category of maximum probability, we can predict the category  $c$  for the event data  $\\mathcal{E}$  as,",
+        "bbox": [
+            75,
+            309,
+            468,
+            365
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nc = \\underset {i} {\\operatorname {a r g m a x}} \\left(p _ {i}\\right). \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            213,
+            375,
+            468,
+            396
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Category-guided Attraction Loss. The challenging part of our joint formulation is to train the reconstruction network  $\\mathcal{G}$  without using paired images of event data. Once we reconstruct the intensity images from the event stream (Eq. 1), we can perform object recognition using CLIP (Eq. 2). Motivated by the fact that category information is important in reconstructing images from events, we propose category-guided attraction loss to train the reconstruction network. The proposed category-guided attraction loss brings the visual feature  $v$  encoded from the reconstructed image  $\\mathcal{I}$  closer to the textual feature of the corresponding category, guiding the model to reconstruct the image corresponding to the category. Since category labels are not available for the event data, we use the predicted categories as pseudo labels. For event data with a batch size of  $B$ , the category-guided attraction loss  $\\mathcal{L}_{att}$  is formulated as InfoNCE loss [33]:",
+        "bbox": [
+            75,
+            403,
+            468,
+            625
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {a t t} = - \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(v _ {i} f _ {c _ {i}} ^ {\\top}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(v _ {i} f _ {c _ {j}} ^ {\\top}\\right)}, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            149,
+            627,
+            468,
+            666
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $v_{i}, c_{i}$ , and  $f_{j}$  denote the visual feature and predicted category of the  $i^{\\mathrm{th}}$  event data, and the textual feature of category  $j$ . Minimizing  $\\mathcal{L}_{att}$  enables the image reconstruction network to generate images of the corresponding category label, while object recognition benefits from the reconstructed images.",
+        "bbox": [
+            75,
+            672,
+            468,
+            755
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Category-agnostic Repulsion Loss. Although  $\\mathcal{L}_{att}$  enables joint learning, the use of pseudo labels may cause collapsing, resulting in all outputs being assigned to a single category. Therefore, we introduce a category-agnostic repulsion loss to prevent collapsing by increasing the distances between visual features. To be specific, the category-agnostic repulsion loss  $\\mathcal{L}_{rep}$  is applied to the set of visual features as follows:",
+        "bbox": [
+            75,
+            755,
+            468,
+            865
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {r e p} = \\sum_ {i = 1} ^ {B} \\log \\left(1 + \\sum_ {j = 1, j \\neq i} ^ {B} \\exp \\left(v _ {i} v _ {j} ^ {\\top}\\right)\\right). \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            559,
+            123,
+            890,
+            163
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "With  $\\mathcal{L}_{att}$  and  $\\mathcal{L}_{rep}$ , we can successfully formulate our joint learning framework, enabling simultaneous event-to-image reconstruction and (zero-shot) object recognition.",
+        "bbox": [
+            497,
+            168,
+            890,
+            209
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Reliable Data Sampling",
+        "text_level": 1,
+        "bbox": [
+            499,
+            214,
+            717,
+            229
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As  $\\mathcal{L}_{att}$  relies on pseudo labels, the reliability of predictions may affect the overall training procedure and performance. To reduce the effects of unreliable predictions, we devise two reliability indicators named posterior probability indicator (PPI) and temporally reversed consistency indicator (TRCI). Based on the resultant reliability of predictions, we sample data with reliable predictions for  $\\mathcal{L}_{att}$ . The remaining unreliable data is excluded from calculating  $\\mathcal{L}_{att}$  and used only for  $\\mathcal{L}_{rep}$ .",
+        "bbox": [
+            497,
+            232,
+            890,
+            356
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Posterior Probability Indicator. Given the reconstructed image  $\\mathcal{I}$  and the predicted category  $c$ , the posterior probability  $p_c = \\mathbf{p}(c|\\mathcal{I})$  indicates the model confidence for category  $c$ . Thus, we use the posterior probability as an indicator of the reliability of the prediction. In particular, we select  $K$  data samples corresponding to the top- $K$  posterior probabilities among  $B$  data samples in the mini-batch. The set of selected sample indices from PPI,  $S_{\\mathrm{PPI}}$  is defined as,",
+        "bbox": [
+            497,
+            357,
+            890,
+            467
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS _ {\\mathrm {P P I}} = \\underset {i} {\\operatorname {a r g s o r t}} (\\mathbf {p} (c _ {i} | \\mathcal {I} _ {i})) ] _ {K}, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            594,
+            473,
+            890,
+            495
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{I}_i$  and  $[\\cdot ]_K$  denote the reconstructed image of the  $i^{\\mathrm{th}}$  event data in the mini-batch and the set of the first  $K$  items. The  $\\operatorname {argsort}(\\cdot)$  function returns indices that would sort an input in descending order.",
+        "bbox": [
+            497,
+            497,
+            890,
+            552
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Temporally Reversed Consistency Indicator. Considering the temporal aspect of event data, we propose an additional reliability indicator based on the consistency between predictions from the original events and the temporally reversed events. From the fact that the temporally reversed events  $\\mathcal{E}^R = (u_i,\\max_j(t_j) - t_i,p_i)_{i = N}^1$  should be categorized into the same category as the original events  $\\mathcal{E}$ , we presume that the consistent model prediction would indicate the reliability of the data sample. To be specific, the set of selected sample indices from TRCI,  $S_{\\mathrm{TRCI}}$  is defined as follows:",
+        "bbox": [
+            497,
+            553,
+            890,
+            703
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS _ {\\mathrm {T R C I}} = \\{i \\mid c _ {i} = c _ {i} ^ {R}, i = 1, 2, \\dots , B \\}, \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            563,
+            704,
+            890,
+            720
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $c_{i}^{R}$  denotes a predicted category of  $i^{\\mathrm{th}}$  reversed event data  $\\mathcal{E}_i^R$  in the mini-batch.",
+        "bbox": [
+            497,
+            725,
+            890,
+            752
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our final set of selected sample indices  $S_{\\mathrm{RDS}}$  for reliable data sampling (RDS) is the intersection of  $S_{\\mathrm{PPI}}$  and  $S_{\\mathrm{TRCI}}$ :",
+        "bbox": [
+            497,
+            754,
+            890,
+            783
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS _ {\\mathrm {R D S}} = S _ {\\mathrm {P P I}} \\cap S _ {\\mathrm {T R C I}}. \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            613,
+            795,
+            890,
+            809
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Then, the modified attraction loss with RDS is as follows:",
+        "bbox": [
+            499,
+            813,
+            880,
+            827
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {a t t} = - \\sum_ {i \\in S _ {\\mathrm {R D S}}} \\log \\frac {\\exp \\left(v _ {i} f _ {c _ {i}} ^ {\\top}\\right)}{\\sum_ {j \\in S _ {\\mathrm {R D S}}} \\exp \\left(v _ {i} f _ {c _ {j}} ^ {\\top}\\right)}, \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            549,
+            832,
+            890,
+            869
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "19869",
+        "bbox": [
+            465,
+            887,
+            504,
+            897
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Local-global Reconstruction Consistency",
+        "text_level": 1,
+        "bbox": [
+            77,
+            125,
+            426,
+            140
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Along with  $\\mathcal{L}_{att}$  and  $\\mathcal{L}_{rep}$ , we further introduce the local-global reconstruction consistency for spatially regularizing the reconstructed images in a self-supervised manner. With random crop function  $\\mathcal{H}$ , we first crop the EST of original events  $\\mathrm{EST}(\\mathcal{E})$  to obtain the cropped tensor  $\\mathcal{H}(\\mathrm{EST}(\\mathcal{E}))$  for the local input. Then, the local-global reconstruction consistency loss  $L_{con}$  is defined as,",
+        "bbox": [
+            75,
+            144,
+            468,
+            243
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {c o n}} = \\| \\mathcal {G} (\\mathcal {H} (\\operatorname {E S T} (\\mathcal {E}))) - \\mathcal {H} (\\mathcal {G} (\\operatorname {E S T} (\\mathcal {E}))) \\| _ {1}. \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            100,
+            252,
+            468,
+            269
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Compared to  $\\mathcal{L}_{att}$  and  $\\mathcal{L}_{rep}$  that work on visual features,  $L_{con}$  is applied directly to the reconstructed images and considered as a spatial regularization for the reconstruction network  $\\mathcal{G}$ . In addition, matching the reconstructed images from local and global inputs enables the network to capture the local details, improving the reconstruction quality.",
+        "bbox": [
+            75,
+            279,
+            468,
+            363
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The total loss function for the joint learning framework is as follows:",
+        "bbox": [
+            75,
+            364,
+            468,
+            391
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {a t t}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e p}} + \\lambda_ {3} \\mathcal {L} _ {\\text {c o n}}, \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            144,
+            403,
+            468,
+            419
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda_1, \\lambda_2,$  and  $\\lambda_3$  are weights for  $\\mathcal{L}_{att},\\mathcal{L}_{rep}$ , and  $\\mathcal{L}_{con}$ .",
+        "bbox": [
+            75,
+            429,
+            467,
+            444
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4. Employing Unpaired Images",
+        "text_level": 1,
+        "bbox": [
+            77,
+            453,
+            332,
+            467
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we expand the use case of our method by introducing how to employ the existing unpaired images. Here, we assume that the category of each image is known, which is the common case for web crawling. Although our framework is capable of event-to-image reconstruction and object recognition without using any images and labels, utilizing existing images can possibly improve the image reconstruction quality and consequently improve object recognition performance. As shown in Fig. 3, we replace textual features in  $\\mathcal{L}_{att}$  with the prototype features generated from unpaired images. For each category  $c$ , we first encode all images of category  $c$  and apply a clustering algorithm to obtain  $L$  clusters in total. Then, we aggregate the features of each cluster by averaging them, resulting in  $L$  prototype features  $W^{c} = (w_{i}^{c})_{i=1}^{L} \\in \\mathbb{R}^{L \\times D}$ .",
+        "bbox": [
+            75,
+            473,
+            468,
+            681
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For data sample index  $i \\in S_{\\mathrm{RDS}}$  with corresponding visual feature  $v_{i}$  and the predicted category  $c_{i}$ , we replace the textual features  $f_{c_i}$  in Eq. 9 to the closest prototype feature  $w_{l_i}^{c_i}$ , where the cluster index  $l_{i}$  is calculated as,",
+        "bbox": [
+            75,
+            682,
+            468,
+            739
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nl _ {i} = \\underset {j} {\\operatorname {a r g m a x}} \\left(v _ {i} \\left(w _ {j} ^ {c _ {i}}\\right) ^ {\\top}\\right). \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            186,
+            749,
+            468,
+            775
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our modification of  $\\mathcal{L}_{att}$  to employ unpaired images is then defined as follows:",
+        "bbox": [
+            75,
+            783,
+            468,
+            811
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {a t t} = - \\sum_ {i \\in S _ {\\mathrm {R D S}}} \\log \\frac {\\exp \\left(v _ {i} \\left(w _ {l _ {i}} ^ {c _ {i}}\\right) ^ {\\top}\\right)}{\\sum_ {j \\in S _ {\\mathrm {R D S}}} \\exp \\left(v _ {i} \\left(w _ {l _ {j}} ^ {c _ {j}}\\right) ^ {\\top}\\right)}, \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            100,
+            819,
+            468,
+            858
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/1ad5b3a30d84555de23d6033ead8896ecaec074745b65999f993303a6652fe14.jpg",
+        "image_caption": [
+            "Figure 3: Proposed method employing unpaired images in the category-guided attraction loss. Unpaired images are further utilized to enhance object recognition performance and reconstructed image quality."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            122,
+            890,
+            346
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Datasets",
+        "text_level": 1,
+        "bbox": [
+            499,
+            429,
+            594,
+            444
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The experiments are conducted on two widely used event-based object recognition datasets, N-Caltech101 [35] and N-ImageNet [22], which are constructed by filming Caltech101 [15] and ImageNet [14] datasets on LCD monitor with event camera. Since N-Caltech101 and N-ImageNet datasets do not provide APS frames, the image in the original dataset and corresponding event sequence have different resolutions and are not aligned in pixel-level. To create a visual prototype in Sec. 3.4, we scrap images from the original dataset, mimicking the web crawling. To avoid including paired images of events in the train set, we randomly split the event train set in half for each class. One half is used to train the model, while the corresponding images from the other half are used to create the visual prototype.",
+        "bbox": [
+            497,
+            451,
+            890,
+            644
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "N-Caltech101 dataset consists of 8,246 event stream data belonging to 101 different object classes. Each object is captured with  $320 \\times 245$  resolution (average) for 300 milliseconds. Existing works [45, 60] utilized this dataset to evaluate the performance of image reconstruction and event-based object recognition.",
+        "bbox": [
+            497,
+            644,
+            890,
+            726
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "N-ImageNet is recently released large-scale dataset, containing 1,781,167 event stream data of 1,000 object classes. Compared to N-Caltech101, it has a larger resolution of  $640 \\times 480$  and a very short capturing time with  $5 \\times 10^{-2}$  milliseconds. Therefore, each event stream contains significantly sparse edge information. N-ImageNet is too challenging for unsupervised studies without ground truth, so we utilize a randomly sampled subset of 100 classes called N-ImageNet (Mini). Details about the split of N-ImageNet (Mini) are provided in the supplementary material.",
+        "bbox": [
+            497,
+            726,
+            890,
+            866
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "19870",
+        "bbox": [
+            465,
+            887,
+            505,
+            897
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/3a1715fb6849501fc70c9fa01ac0297af97b30bd9f8f4184985cfb16da0219bc.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td rowspan=\"2\">Image Pair</td><td rowspan=\"2\">Class Label</td><td colspan=\"3\">N-Caltech101</td><td colspan=\"3\">N-ImageNet (Mini)</td></tr><tr><td>Accuracy ↑</td><td>FID ↓</td><td>IS ↑</td><td>Accuracy ↑</td><td>FID ↓</td><td>IS ↑</td></tr><tr><td colspan=\"9\">Supervised Learning</td></tr><tr><td>Histogram [29]</td><td></td><td>✓</td><td>63.8</td><td>-</td><td>-</td><td>49.8</td><td>-</td><td>-</td></tr><tr><td>Ev-gait [61]</td><td></td><td>✓</td><td>61.9</td><td>-</td><td>-</td><td>51.3</td><td>-</td><td>-</td></tr><tr><td>EST [16]</td><td></td><td>✓</td><td>81.7*</td><td>-</td><td>-</td><td>52.5</td><td>-</td><td>-</td></tr><tr><td>DiST [22]</td><td></td><td>✓</td><td>73.8</td><td>-</td><td>-</td><td>57.9</td><td>-</td><td>-</td></tr><tr><td colspan=\"9\">Unsupervised Learning</td></tr><tr><td>HOTS [25]</td><td></td><td></td><td>21.0*</td><td>-</td><td>-</td><td>1.6</td><td>-</td><td>-</td></tr><tr><td colspan=\"9\">Image Reconstruction + CLIP</td></tr><tr><td>E2VID [44] (Pre-train)</td><td></td><td></td><td>64.0</td><td>83.90</td><td>12.81</td><td>13.3</td><td>147.14</td><td>7.82</td></tr><tr><td>E2VID [44] (Fine-tune)</td><td></td><td></td><td>59.8</td><td>85.31</td><td>12.91</td><td>13.4</td><td>144.57</td><td>7.48</td></tr><tr><td>E2VID [44] (Fine-tune)</td><td>✓</td><td></td><td>59.8</td><td>86.66</td><td>12.61</td><td>13.5</td><td>144.03</td><td>7.40</td></tr><tr><td>E2VID [44] (Fine-tune)</td><td></td><td>✓</td><td>60.1</td><td>85.50</td><td>12.85</td><td>12.7</td><td>145.79</td><td>7.33</td></tr><tr><td>E2VID [44] (Scratch)</td><td>✓</td><td></td><td>9.4</td><td>302.48</td><td>2.60</td><td>1.0</td><td>299.70</td><td>1.15</td></tr><tr><td>SSL-E2VID [37] (Pre-train)</td><td></td><td></td><td>28.2</td><td>281.09</td><td>2.36</td><td>1.6</td><td>324.43</td><td>2.08</td></tr><tr><td>SSL-E2VID [37] (Scratch)</td><td></td><td></td><td>30.5</td><td>228.30</td><td>4.37</td><td>1.7</td><td>250.55</td><td>3.00</td></tr><tr><td>Wang et al. [59] (Scratch)</td><td></td><td></td><td>42.7</td><td>160.34</td><td>6.83</td><td>10.1</td><td>231.94</td><td>5.00</td></tr><tr><td>Wang et al. [59] (Scratch)</td><td>✓</td><td></td><td>43.5</td><td>72.67</td><td>4.50</td><td>19.0</td><td>180.76</td><td>4.89</td></tr><tr><td>Wang et al. [59] (Scratch)</td><td></td><td>✓</td><td>39.7</td><td>179.79</td><td>4.18</td><td>13.2</td><td>216.37</td><td>4.70</td></tr><tr><td>Ours (Text Prompt)</td><td></td><td></td><td>82.46</td><td>62.29</td><td>14.81</td><td>30.16</td><td>150.58</td><td>7.92</td></tr><tr><td>Ours (Visual Prototype)</td><td></td><td></td><td>82.61</td><td>54.19</td><td>17.59</td><td>31.28</td><td>135.95</td><td>9.74</td></tr></table>",
+        "bbox": [
+            85,
+            123,
+            884,
+            444
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: The results of event-based object recognition and event-to-image reconstruction. The **bold** and **underline** denote the best and the second-best performance except for supervised methods, respectively. * denotes the values taken from the original paper and italic denotes the performance of supervised methods.",
+        "bbox": [
+            75,
+            450,
+            892,
+            492
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            75,
+            504,
+            208,
+            519
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We compare our methods utilizing text prompts or visual prototypes with other existing methods. We adopt four representations (Histogram [29], Ev-gait [61], EST [16], and DiST [22]) for supervised methods and an unsupervised method (HOTS [25]). Additionally, three image reconstruction networks, E2VID [44], SSL-E2VID [37] and Wang et al. [59], are utilized with CLIP. For a fair comparison with image reconstruction networks, we record the performance of three different models for each image reconstruction network: the official pre-trained model, fine-tuned model, and a model trained from scratch. We evaluate these models under three conditions: paired images, images belonging to the same class, and randomly selected images. Some cases are overlooked when the pre-trained model is unavailable or the training setup does not yield meaningful results. For implementation details, see the supplementary material.",
+        "bbox": [
+            75,
+            526,
+            470,
+            748
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Label-free Object Recognition",
+        "text_level": 1,
+        "bbox": [
+            75,
+            762,
+            346,
+            777
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Text Prompt. The quantitative results are presented in Table 1. The proposed method outperforms supervised object recognition methods on the N-Caltech101 dataset, despite not using category labels. Also, our method surpasses existing unsupervised methods and image reconstruction methods with CLIP by a large margin under all",
+        "bbox": [
+            75,
+            783,
+            468,
+            866
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "metrics. Specifically, ours surpasses the best image reconstruction with CLIP method, E2VID (Pre-train), by around  $28\\%$  increase in accuracy even though it utilizes the image while pre-training. On N-ImageNet, supervised methods record higher accuracy than ours since the dataset is challenging due to the sparse events captured within a short period. However, ours shows a significant improvement in all metrics compared to other unsupervised and image reconstruction-based methods. Our approach demonstrates the ability to perform object recognition and image reconstruction from event data, even when trained solely on textual features without any image data.",
+        "bbox": [
+            497,
+            504,
+            890,
+            671
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Visual Prototype. Besides text prompts, our framework can employ visual prototypes generated from scrapped images. By incorporating visual information, our visual prototype-based approach exceeds the performance of relying solely on textual prompts and achieves the best performance among unsupervised and image reconstruction-based methods under all metrics across both datasets. This demonstrates the superiority of the visual prototypes.",
+        "bbox": [
+            497,
+            671,
+            890,
+            783
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2. Image-free Image Reconstruction",
+        "text_level": 1,
+        "bbox": [
+            499,
+            790,
+            794,
+            806
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We also compare the reconstruction results of our methods with other event-to-image reconstruction networks in Table 1. Note that, as mentioned in Sec. 4, the resolution and pixel positions of events and corresponding orig-",
+        "bbox": [
+            497,
+            811,
+            892,
+            866
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "19871",
+        "bbox": [
+            465,
+            887,
+            504,
+            897
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/7812e7b24f7ef6d5ab6206f30e47da0bd3c45143d93840596bb805df70d09fcc.jpg",
+        "image_caption": [
+            "Events",
+            "Wang et. al.",
+            "E2VID",
+            "E2VID*",
+            "Ours (Prompt)",
+            "Ours (Proto.)",
+            "Ours (Proto.)*",
+            "Image",
+            "(a) N-Caltech101"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            138,
+            873,
+            346
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e58e7f9c7861d171cda9424194bfc75d07421857ace6eb51493299e456148679.jpg",
+        "image_caption": [
+            "(b) N-ImageNet",
+            "Figure 4: Visual results of event-to-image reconstruction network on (a) N-Caltech101 and (b) N-ImageNet datasets. Ours shows a higher quality of reconstructed images with fine-grained details than other methods (Wang et al. [59] and pre-trained E2VID [44]). * denotes the images with brightness and contrast adjustment."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            366,
+            873,
+            574
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "inal images are not aligned. Therefore, pixel-wise metrics (e.g., L1, PSNR) cannot be used for comparison. Instead, to assess the visual quality of the reconstructed images, we utilize two metrics: inception score (IS) [47] and Fréchet inception distance (FID) [18].",
+        "bbox": [
+            75,
+            657,
+            468,
+            726
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Despite not utilizing any images during training, our text prompt-based approach surpasses all other image reconstruction techniques in terms of FID and IS metrics on the N-Caltech101 dataset. However, in the case of N-ImageNet, our text prompt approach is the strongest by an IS indicator but falls behind pre-trained E2VID in FID. Since ours is trained from events of short periods without paired images, it becomes challenging to learn the distribution of real images. Instead, utilizing the visual prototype results in the highest FID and IS indicators in both datasets. This implies",
+        "bbox": [
+            75,
+            728,
+            468,
+            865
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "that it can acquire the ability to reconstruct images without requiring paired images that correspond to events.",
+        "bbox": [
+            497,
+            657,
+            890,
+            685
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We also present the qualitative results in Fig. 4. Wang et al. [59] utilize the paired image in training but the method lacks the ability to effectively display intensity contrast, resulting in the loss of many details. While pre-trained E2VID shows better image quality over [59], it still struggles to preserve fine details as it operates in a distinct training environment. This is particularly apparent in the N-ImageNet dataset, where E2VID generates results that deviate significantly from the original image. Conversely, our framework is able to recover intensity effectively and preserve details in reconstructed images, even without access to paired images. Our results closely resemble real images, even in the challenging N-ImageNet dataset.",
+        "bbox": [
+            497,
+            687,
+            892,
+            866
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "19872",
+        "bbox": [
+            465,
+            887,
+            504,
+            897
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/0cfd7f1f0fb8320a5f80648264f904e3a13da3d8a26c71fecd7dc2eda2a577c6.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td rowspan=\"2\">Image Pair</td><td rowspan=\"2\">Class Label</td><td colspan=\"3\">N-Caltech101</td><td colspan=\"3\">N-ImageNet (Mini)</td></tr><tr><td>Accuracy ↑</td><td>FID ↓</td><td>IS ↑</td><td>Accuracy ↑</td><td>FID ↓</td><td>IS ↑</td></tr><tr><td colspan=\"9\">Image Reconstruction + CLIP</td></tr><tr><td>E2VID [44] (Pre-train)</td><td></td><td></td><td>58.4</td><td>194.26</td><td>6.47</td><td>31.7</td><td>203.05</td><td>6.64</td></tr><tr><td>E2VID [44] (Fine-tune)</td><td></td><td></td><td>61.7</td><td>203.39</td><td>7.56</td><td>30.0</td><td>199.83</td><td>6.73</td></tr><tr><td>E2VID [44] (Fine-tune)</td><td>✓</td><td></td><td>61.1</td><td>203.40</td><td>7.45</td><td>30.9</td><td>199.44</td><td>6.80</td></tr><tr><td>E2VID [44] (Fine-tune)</td><td></td><td>✓</td><td>61.5</td><td>202.23</td><td>7.56</td><td>29.4</td><td>203.97</td><td>6.08</td></tr><tr><td>E2VID [44] (Scratch)</td><td>✓</td><td></td><td>21.3</td><td>362.08</td><td>2.43</td><td>4.9</td><td>333.13</td><td>1.36</td></tr><tr><td>SSL-E2VID [37] (Pre-train)</td><td></td><td></td><td>24.7</td><td>332.43</td><td>2.07</td><td>5.7</td><td>352.53</td><td>2.04</td></tr><tr><td>SSL-E2VID [37] (Scratch)</td><td></td><td></td><td>25.3</td><td>289.36</td><td>3.93</td><td>6.1</td><td>287.98</td><td>2.79</td></tr><tr><td>Wang et al. [59] (Scratch)</td><td></td><td></td><td>47.5</td><td>228.77</td><td>4.67</td><td>18.7</td><td>241.87</td><td>6.12</td></tr><tr><td>Wang et al. [59] (Scratch)</td><td>✓</td><td></td><td>48.5</td><td>230.74</td><td>4.91</td><td>20.5</td><td>236.78</td><td>3.90</td></tr><tr><td>Wang et al. [59] (Scratch)</td><td></td><td>✓</td><td>43.7</td><td>215.63</td><td>4.66</td><td>15.6</td><td>237.38</td><td>4.40</td></tr><tr><td>Ours (Text Prompt)</td><td></td><td></td><td>84.31</td><td>147.31</td><td>6.81</td><td>45.10</td><td>202.29</td><td>6.31</td></tr><tr><td>Ours (Visual Prototype)</td><td></td><td></td><td>85.56</td><td>107.58</td><td>9.35</td><td>51.50</td><td>174.24</td><td>7.32</td></tr></table>",
+        "bbox": [
+            85,
+            123,
+            884,
+            347
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3. Zero-shot Object Recognition",
+        "text_level": 1,
+        "bbox": [
+            75,
+            378,
+            339,
+            395
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We further demonstrate the advantages of our framework over supervised methods by enabling it to be used for zero-shot object recognition. The reconstructed images can be utilized for zero-shot object recognition thanks to CLIP's capacity to align images with any semantic concepts in an open vocabulary for zero-shot object recognition. To define \"unseen\" categories that are not seen during training, we use only 80 of the 100 categories for training and the remaining 20 for zero-shot testing.",
+        "bbox": [
+            75,
+            399,
+            468,
+            524
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In Table 2, we present the performance of zero-shot object recognition. Without any category labels, our methods with text prompt achieve a promising  $84.31\\%$  and  $45.10\\%$  on N-Caltech101 and N-ImageNet, respectively. This is  $25.91\\%$  and  $13.4\\%$  higher than E2VID pretrained with a large number of image pairs. These results validate that although our model learns through the process of being category-aware, generalization is possible for other categories, especially \"unseen\" in this case. Although FID and IS of our method using text prompts are relatively poor compared to pre-trained E2VID on N-ImageNet, this can also be overcome by using visual prototypes. Our method using the visual prototype is able to achieve a substantial improvement under all metrics. Specifically, in the N-Caltech101 dataset, our method increases the IS from 6.81 to 9.35, while in the N-ImageNet dataset, it increases the accuracy from 45.10 to 51.50.",
+        "bbox": [
+            75,
+            526,
+            470,
+            762
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Ablation Study and Discussion",
+        "text_level": 1,
+        "bbox": [
+            75,
+            777,
+            349,
+            793
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.1. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            75,
+            801,
+            231,
+            817
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We conduct an ablation study to understand how the components influence performance. Table 3 illustrates the effectiveness of each component on N-ImageNet dataset.",
+        "bbox": [
+            75,
+            824,
+            468,
+            866
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/3606d8eb0de6e8ed041a6a5e043c5349633a54dc568ac3ebe679df3232df95ca.jpg",
+        "table_caption": [
+            "Table 2: The results of event-based zero-shot object recognition using CLIP."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td rowspan=\"2\">Latt</td><td rowspan=\"2\">Lrep</td><td rowspan=\"2\">Lcon</td><td colspan=\"2\">RDS</td><td rowspan=\"2\">Accuracy ↑</td></tr><tr><td>PPI</td><td>TRCI</td></tr><tr><td>(1)</td><td>✓</td><td></td><td></td><td></td><td></td><td>9.50</td></tr><tr><td>(2)</td><td>✓</td><td>✓</td><td></td><td></td><td></td><td>14.00</td></tr><tr><td>(3)</td><td>✓</td><td>✓</td><td>✓</td><td></td><td></td><td>15.68</td></tr><tr><td>(4)</td><td>✓</td><td>✓</td><td></td><td>✓</td><td></td><td>28.42</td></tr><tr><td>(5)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>28.90</td></tr><tr><td>(6)</td><td>✓</td><td>✓</td><td></td><td>✓</td><td>✓</td><td>28.91</td></tr><tr><td>(7)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>30.16</td></tr></table>",
+        "bbox": [
+            505,
+            378,
+            885,
+            511
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Ablation study of the proposed method.",
+        "bbox": [
+            534,
+            513,
+            855,
+            527
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We start with a baseline, which denotes the removal of everything except attraction loss,  $\\mathcal{L}_{att}$ . During the experiment, we notice that if we only apply attraction loss towards the network without repulsion loss,  $\\mathcal{L}_{rep}$ , the collapse often takes place. Therefore,  $\\mathcal{L}_{rep}$  is essential for stable learning of  $\\mathcal{L}_{att}$ . It is worth mentioning the impact of the proposed PPI on performance. Upon analyzing (4), which displays the results when PPI is added to (2), it can be seen that the accuracy has increased by 14.42. Upon examining (5) and (7), it can be observed that the accuracy increases by 1.26, validating the impact of the sub-sampling process via TRCI. Furthermore, the ablation results between (6) and (7) show that the reconstruction consistency between global and local preserves the spatial details of the image, leading to performance improvement. Finally, the model (7) that has all of its components incorporated achieves the highest performance.",
+        "bbox": [
+            497,
+            549,
+            890,
+            774
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.2. Number of Samples in PPI",
+        "text_level": 1,
+        "bbox": [
+            499,
+            777,
+            741,
+            793
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The  $K$  used in PPI for selecting the sample with the highest probability impacts the performance of object recognition. Fig. 5 shows the evaluation of performance on N-ImageNet dataset based on different values of  $K$ . When  $K$  is set to 32, which includes all samples in the batch, the",
+        "bbox": [
+            497,
+            796,
+            890,
+            866
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "19873",
+        "bbox": [
+            465,
+            887,
+            505,
+            897
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/eb8c64942a8dac1e10eb30d1237c9becc4a3d28ea5a4f77e31027d0cad741418.jpg",
+        "image_caption": [
+            "Figure 5: Object recognition accuracy on N-ImageNet (Mini) according to  $K$  in PPI."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            84,
+            123,
+            460,
+            197
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/d94d636714e43c46ad2e957cfeb9cc1cfa7d442abd6bce46b365f99be7aa9a4a.jpg",
+        "image_caption": [
+            "Figure 6: Object recognition accuracy on N-Caltech101 according to  $K$  in PPI."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            245,
+            463,
+            317
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/90b415a0d004f3cb1321bdbf5ec937e872a4c9a47cf2ce06f9fdf56396414993.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td colspan=\"3\">N-Caltech101</td><td colspan=\"3\">N-ImageNet (Mini)</td></tr><tr><td>λ2\\λ3</td><td>0.5</td><td>1</td><td>2</td><td>0.5</td><td>1</td><td>2</td></tr><tr><td>0.001</td><td>82.71</td><td>82.93</td><td>81.26</td><td>29.62</td><td>26.34</td><td>30.08</td></tr><tr><td>0.005</td><td>83.52</td><td>82.02</td><td>81.66</td><td>29.38</td><td>26.92</td><td>28.72</td></tr><tr><td>0.01</td><td>82.24</td><td>82.46</td><td>81.99</td><td>29.34</td><td>30.16</td><td>29.80</td></tr><tr><td>0.05</td><td>81.99</td><td>82.39</td><td>82.61</td><td>27.50</td><td>27.92</td><td>29.00</td></tr><tr><td>0.1</td><td>81.37</td><td>82.46</td><td>82.17</td><td>25.76</td><td>26.28</td><td>26.32</td></tr></table>",
+        "bbox": [
+            80,
+            366,
+            463,
+            461
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 4: The accuracy according to hyper-parameter in Eq. 11 when  $\\lambda_1 = 1$ .",
+        "bbox": [
+            75,
+            465,
+            468,
+            492
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "attraction loss,  $\\mathcal{L}_{att}$ , is applied to even those samples with low reliability, resulting in the poorest performance. Due to the same reason, the performance is better when  $K$  is 16 than when it is 32, but since unreliable samples are also selected, the accuracy is relatively lower than in other favorable situations. As  $K$  decreases, the performance improves, but it starts to decline when  $K$  is less than 6, indicating that even with high reliability, it is challenging to achieve optimal training if  $\\mathcal{L}_{att}$  is applied with very few samples. Similarly, we report the accuracy according to the  $K$  value on the N-Caltech101 in Fig 6. Unless an extremely small number of samples are drawn, as the value of  $K$  increases, the performance also improves. However, once a certain threshold is reached, there is little significant change, and it starts to decline with further increases. The accuracy exceeds  $80\\%$  in all cases, showing robustness to  $K$  values. Although the level of sensitivity varies depending on the dataset, we do confirm that satisfactory results can be obtained between  $K = 4$  and  $K = 10$ .",
+        "bbox": [
+            75,
+            515,
+            468,
+            777
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6.3. Hyper-parameter Analysis",
+        "text_level": 1,
+        "bbox": [
+            75,
+            788,
+            317,
+            804
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In Table 4, we report the accuracy according to  $\\lambda_{2}$  and  $\\lambda_{3}$  in Eq. 11. If the weight of the repulsion loss,  $\\lambda_{2}$ , is set too high, the performance decreases. However, this can be mitigated by increasing the weight of the local-global re",
+        "bbox": [
+            75,
+            809,
+            468,
+            866
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/d578b2cc4dbd1da4c37ac78ccae19e553041cace94a5a3168bb1d7da94a52d11.jpg",
+        "image_caption": [
+            "Figure 7: Object recognition accuracy on N-Caltech101 according to the number of extra category prompts."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            124,
+            880,
+            205
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "construction consistency loss,  $\\lambda_3$ . The results verify the robustness of our method according to the  $\\lambda$  on both datasets.",
+        "bbox": [
+            497,
+            262,
+            890,
+            290
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6.4. Learning with Superset Categories",
+        "text_level": 1,
+        "bbox": [
+            499,
+            302,
+            801,
+            319
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In practical terms, it can be challenging for users to find corresponding categories of the event dataset. As a result, they often end up taking more categories as prompts. We tackle these issues by demonstrating that our joint learning framework works robustly even when employing more prompts than the categories of the event on the N-Caltech101 dataset. To avoid using additional prompts that are too irrelevant to the N-Caltech101 dataset, we randomly take additional prompts from N-ImageNet. We show the performance results according to the number of extra category prompts in Fig. 7. While the performance decreases with an increase in extra category prompts, it is worth noting that our algorithm remains reliable even when faced with over 100 additional prompts beyond the corresponding category. This suggests that our algorithm can function with superset categories, even in a situation when the number of extra categories is much larger than the actual categories.",
+        "bbox": [
+            497,
+            324,
+            890,
+            560
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            499,
+            577,
+            616,
+            592
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we propose a joint learning framework for event-based object recognition with image reconstruction from events without labels and paired images. With the proposed losses and the reliable data sampling strategy, our approach allows stable learning of two tasks simultaneously from pseudo labels. As a result, our approach achieves exceptional performance in event-based object recognition and event-to-image reconstruction, all without requiring category labels and paired images. Our approach successfully addresses the challenge of label dependency in event-based research, which has been a significant obstacle. We anticipate that our work can be further extended to other tasks such as event-based zero-shot object detection, zero-shot semantic segmentation, etc., that have not explored so far in the research field for event cameras.",
+        "bbox": [
+            497,
+            600,
+            890,
+            808
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements. This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF-2022R1A2B5B03002636).",
+        "bbox": [
+            497,
+            809,
+            890,
+            866
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "19874",
+        "bbox": [
+            465,
+            887,
+            505,
+            897
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            124,
+            171,
+            138
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7243-7252, 2017. 1, 2",
+            "[2] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 884-892, 2016. 2",
+            "[3] Yin Bi, Aaron Chadha, Alhabib Abbas, Eirina Bourtsoulatze, and Yiannis Andreopoulos. Graph-based object classification for neuromorphic vision sensing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 491-501, 2019. 2",
+            "[4] Marco Cannici, Marco Ciccone, Andrea Romanoni, and Matteo Matteucci. A differentiable recurrent surface for asynchronous event-based data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16, pages 136–152. Springer, 2020. 1, 2",
+            "[5] Hila Chefer, Sagie Benaim, Roni Paiss, and Lior Wolf. Image-based clip-guided essence transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 695-711. Springer, 2022. 2",
+            "[6] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. arXiv preprint arXiv:2301.04926, 2023. 2",
+            "[7] Wensheng Cheng, Hao Luo, Wen Yang, Lei Yu, Shoushun Chen, and Wei Li. Det: A high-resolution dvs dataset for lane extraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 2",
+            "[8] Hoonhee Cho, Jegyeong Cho, and Kuk-Jin Yoon. Learning adaptive dense event stereo from the image domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17797-17807, 2023. 2",
+            "[9] Hoonhee Cho, Jaeseok Jeong, and Kuk-Jin Yoon. Eomvs: Event-based omnidirectional multi-view stereo. IEEE Robotics and Automation Letters, 6(4):6709-6716, 2021. 2",
+            "[10] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 454-462, 2022. 2",
+            "[11] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In European Conference on Computer Vision, pages 470-486. Springer, 2022. 2",
+            "[12] Gregory Cohen, Saeed Afshar, Garrick Orchard, Jonathan Tapson, Ryad Benosman, and Andre van Schaik. Spatial and temporal downsampling in event-based visual classification. IEEE Transactions on Neural Networks and Learning Systems, 29(10):5030-5044, 2018. 1",
+            "[13] Matthew Cook, Luca Gugelmann, Florian Jug, Christoph Krautz, and Angelika Steger. Interacting maps for fast visual"
+        ],
+        "bbox": [
+            78,
+            147,
+            470,
+            855
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "interpretation. In The 2011 International Joint Conference on Neural Networks, pages 770-776. IEEE, 2011. 2",
+            "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5",
+            "[15] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 5",
+            "[16] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5633-5643, 2019. 3, 6",
+            "[17] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Asynchronous, photometric feature tracking using events and frames. In Proceedings of the European Conference on Computer Vision (ECCV), pages 750-765, 2018. 2",
+            "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7",
+            "[19] Yuhuang Hu, Hongjie Liu, Michael Pfeiffer, and Tobi Delbruck. Dvs benchmark datasets for object tracking, action recognition, and object recognition. Frontiers in neuroscience, 10:405, 2016. 2",
+            "[20] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew J Davison. Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ, 43:566-576, 2008. 2",
+            "[21] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 349-364. Springer, 2016. 2",
+            "[22] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2146-2156, 2021. 1, 2, 5, 6",
+            "[23] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 2",
+            "[24] Simon Klenk, David Bonello, Lukas Koestler, and Daniel Cremers. Masked event modeling: Self-supervised pretraining for event cameras. arXiv preprint arXiv:2212.10368, 2022.1,2",
+            "[25] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E Shi, and Ryad B Benosman. Hots: a hierarchy of event-based time-surfaces for pattern recognition. IEEE transactions on pattern analysis and machine intelligence, 39(7):1346-1359, 2016. 1, 2, 6"
+        ],
+        "bbox": [
+            502,
+            127,
+            890,
+            859
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "19875",
+        "bbox": [
+            465,
+            887,
+            504,
+            897
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[26] Jun Haeng Lee, Tobi Delbruck, and Michael Pfeiffer. Training deep spiking neural networks using backpropagation. Frontiers in neuroscience, 10:508, 2016. 1, 2",
+            "[27] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 2",
+            "[28] Iulia-Alexandra Lungu, Federico Corradi, and Tobi Delbrück. Live demonstration: Convolutional neural network driven by dynamic vision sensor playing roshambo. In 2017 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-1. IEEE, 2017. 2",
+            "[29] Ana I Maqueda, Antonio Loquercio, Guillermo Gallego, Narciso García, and Davide Scaramuzza. Event-based vision meets deep learning on steering prediction for self-driving cars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5419-5427, 2018. 1, 6",
+            "[30] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 415-431. Springer, 2020. 1, 2",
+            "[31] Diederik Paul Moeys, Daniel Neil, Federico Corradi, Emmett Kerr, Philip Vance, Gautham Das, Sonya A Coleman, Thomas M McGinnity, Dermot Kerr, and Tobi Delbruck. Pred18: Dataset and further experiments with davis event camera in predator-prey robot chasing. arXiv preprint arXiv:1807.03128, 2018. 2",
+            "[32] Daniel Neil, Michael Pfeiffer, and Shih-Chii Liu. Phased LSTM: Accelerating recurrent network training for long or event-based sequences. Advances in neural information processing systems, 29, 2016. 1, 2",
+            "[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4",
+            "[34] Garrick Orchard, Ryad Benosman, Ralph Etienne-Cummings, and Nitish V Thakor. A spiking neural network architecture for visual motion estimation. In 2013 IEEE Biomedical Circuits and Systems Conference (BioCAS), pages 298-301. IEEE, 2013. 1, 2",
+            "[35] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:437, 2015. 2, 5",
+            "[36] Garrick Orchard, Cedric Meyer, Ralph Etienne-Cummings, Christoph Posch, Nitish Thakor, and Ryad Benosman. Hfirst: A temporal approach to object recognition. IEEE transactions on pattern analysis and machine intelligence, 37(10):2028-2040, 2015. 1, 2",
+            "[37] Federico Paredes-Vallés and Guido CHE de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446-3455, 2021. 2, 6, 8",
+            "[38] Paul KJ Park, Baek Hwan Cho, Jin Man Park, Kyoobin Lee, Ha Young Kim, Hyo Ah Kang, Hyun Goo Lee, Jooyeon Woo, Yohan Roh, Won Jo Lee, et al. Performance improvement of deep learning based gesture recognition using spa"
+        ],
+        "bbox": [
+            78,
+            125,
+            468,
+            859
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "tiotemporal demosaicing technique. In 2016 IEEE International Conference on Image Processing (ICIP), pages 1624-1628. IEEE, 2016. 1",
+            "[39] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2",
+            "[40] José Antonio Pérez-Carrasco, Bo Zhao, Carmen Serrano, Begona Acha, Teresa Serrano-Gotarredona, Shouchun Chen, and Bernabé Linares-Barranco. Mapping from frame-driven to frame-free event-driven vision systems by low-rate rate coding and coincidence processing—application to feedforward convnets. IEEE transactions on pattern analysis and machine intelligence, 35(11):2706-2719, 2013. 1, 2",
+            "[41] Stefano Pini, Guido Borghi, and Roberto Vezzani. Learn to see by events: Color frame synthesis from event and rgb cameras. arXiv preprint arXiv:1812.02041, 2018. 2",
+            "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4",
+            "[43] Bharath Ramesh, Andrés Ussa, Luca Della Vedova, Hong Yang, and Garrick Orchard. Low-power dynamic object detection and classification with freely moving event cameras. Frontiers in neuroscience, 14:135, 2020. 2",
+            "[44] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3857-3866, 2019. 2, 6, 7, 8",
+            "[45] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 1, 2, 4, 5",
+            "[46] Christian Reinbacher, Gottfried Graber, and Thomas Pock. Real-time intensity-image reconstruction for event cameras using manifold regularisation. arXiv preprint arXiv:1607.06283, 2016. 2",
+            "[47] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 7",
+            "[48] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 2",
+            "[49] Cedric Scheerlinck, Nick Barnes, and Robert Mahony. Continuous-time intensity estimation using event cameras. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part V, pages 308–324. Springer, 2019. 2",
+            "[50] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast im"
+        ],
+        "bbox": [
+            502,
+            127,
+            890,
+            859
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "19876",
+        "bbox": [
+            465,
+            887,
+            504,
+            897
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "age reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 156-163, 2020. 2",
+            "[51] Teresa Serrano-Gotarredona and Bernabé Linares-Barranco. Poker-dvs and mnist-dvs. their history, how they were made, and other details. Frontiers in neuroscience, 9:481, 2015. 2",
+            "[52] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In European Conference on Computer Vision, pages 628-645. Springer, 2022. 2",
+            "[53] Amos Sironi, Manuele Brambilla, Nicolas Bourdis, Xavier Lagorce, and Ryad Benosman. Hats: Histograms of averaged time surfaces for robust event-based object classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1731-1740, 2018. 1, 2",
+            "[54] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision, pages 412-428. Springer, 2022. 2",
+            "[55] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918-17928, 2022. 2",
+            "[56] Ajay Vasudevan, Pablo Negri, Bernabe Linares-Barranco, and Teresa Serrano-Gotarredona. Introduction and analysis of an event-based sign language dataset. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 675-682. IEEE, 2020. 2",
+            "[57] Lin Wang, Yujeong Chae, and Kuk-Jin Yoon. Dual transfer learning for event-based end-task prediction via pluggable event to image translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2135-2145, 2021. 2",
+            "[58] Lin Wang, Yujeong Chae, Sung-Hoon Yoon, Tae-Kyun Kim, and Kuk-Jin Yoon. Evdistill: Asynchronous events to end-task learning via bidirectional reconstruction-guided cross-modal knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 608-619, 2021. 2",
+            "[59] Lin Wang, Yo-Sung Ho, Kuk-Jin Yoon, et al. Event-based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10081-10090, 2019. 2, 6, 7, 8",
+            "[60] Lin Wang, Tae-Kyun Kim, and Kuk-Jin Yoon. Joint framework for single image reconstruction and super-resolution with an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7657-7673, 2021. 1, 4, 5",
+            "[61] Yanxiang Wang, Bowen Du, Yiran Shen, Kai Wu, Guangrong Zhao, Jianguo Sun, and Hongkai Wen. Ev-gait: Event-based robust gait recognition using dynamic vision sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6358-6367, 2019. 6",
+            "[62] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Com"
+        ],
+        "bbox": [
+            78,
+            125,
+            468,
+            859
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "puter Vision, pages 2563-2572, 2021. 2",
+            "[63] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 2"
+        ],
+        "bbox": [
+            502,
+            127,
+            892,
+            201
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "19877",
+        "bbox": [
+            465,
+            887,
+            504,
+            897
+        ],
+        "page_idx": 11
+    }
+]

2023/Label-Free Event-based Object Recognition via Joint Learning with Image Reconstruction from Events/cebb8d22-3e6f-4881-895d-c913e7109230_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c74029b9914930cc74637dc334202b9450d6b575bde49a495fb91cb15b124c81
+size 1816931

2023/Label-Free Event-based Object Recognition via Joint Learning with Image Reconstruction from Events/full.md

@@ -0,0 +1,343 @@
+# Label-Free Event-based Object Recognition via Joint Learning with Image Reconstruction from Events
+
+Hoonhee Cho*, Hyeonseong Kim*, Yujeong Chae, and Kuk-Jin Yoon
+
+Korea Advanced Institute of Science and Technology
+
+{gnsgnsgml,brian617,yujeong,kjyoon}@kaist.ac.kr
+
+# Abstract
+
+Recognizing objects from sparse and noisy events becomes extremely difficult when paired images and category labels do not exist. In this paper, we study label-free event-based object recognition where category labels and paired images are not available. To this end, we propose a joint formulation of object recognition and image reconstruction in a complementary manner. Our method first reconstructs images from events and performs object recognition through Contrastive Language-Image Pretraining (CLIP), enabling better recognition through a rich context of images. Since the category information is essential in reconstructing images, we propose category-guided attraction loss and category-agnostic repulsion loss to bridge the textual features of predicted categories and the visual features of reconstructed images using CLIP. Moreover, we introduce a reliable data sampling strategy and local-global reconstruction consistency to boost joint learning of two tasks. To enhance the accuracy of prediction and quality of reconstruction, we also propose a prototype-based approach using unpaired images. Extensive experiments demonstrate the superiority of our method and its extensibility for zero-shot object recognition. Our project code is available at https://github.com/Chohoonhee/Ev-LaFOR.
+
+# 1. Introduction
+
+Event cameras are neuromorphic vision sensors that asynchronously perceive per-pixel brightness changes with high temporal resolution. They show advantages over conventional frame-based cameras, such as high dynamic range (HDR) and low latency. However, despite the superiority of cameras, algorithms using events are still in their infancy. Existing event-based object recognition methods [1, 4, 22, 24, 25, 26, 30, 32, 34, 36, 40, 53] have been shown that object recognition is a crucial task for event feature learning [12, 29, 38] and endeavors have successfully
+
+![](images/0284fba41f1963b2940b0ca24aafc41c889a0aaf349f6e097983639d6b731466.jpg)  
+Figure 1: We tackle label-free event-based object recognition where category labels and paired images are not available. Our approach can simultaneously perform object recognition and image reconstruction through proposed joint learning framework. Optionally, our method can leverage unpaired images to enhance object recognition performance and reconstructed image quality.
+
+classified the object with events. However, they rely on supervision with category labels to achieve high performance. Furthermore, the sparsity of events poses a challenge for users to label them accurately when there are no paired images. Several unsupervised methods [25, 36] have been proposed for object recognition, however, their performances remain subpar. Hence, there is a natural need for robust and effective label-free event-based object recognition methods.
+
+In this paper, we study label-free event-based object recognition where category labels and paired images are not available as in Fig. 1. To tackle this task, we focus on two following facts. Firstly, as demonstrated in [45, 60], performing object recognition on reconstructed images with spatially dense information and rich context shows better performance than solely using events. Secondly, during reconstructing images from events, taking the categories into
+
+account can improve the quality of the reconstructed images. From these observations, we propose a joint learning framework of two closely-linked tasks: event-based object recognition and event-to-image reconstruction.
+
+In our joint learning framework, we first reconstruct images from events and perform object recognition on the reconstructed images using Contrastive Language-Image Pretraining (CLIP) [42]. CLIP is an image-text embedding model that has been trained on a large number of image and language pairs sourced from the web. We utilize the textual features encoded from CLIP since they are well-aligned with visual features, which is advantageous for our tasks that lack labels and images. To this end, we propose the category-guided attraction loss that aligns the visual features from reconstructed images close to the texture features of the predicted category from CLIP.
+
+Although the above framework enables the joint learning of event-based object recognition and image reconstruction, using the predicted categories brings two issues. (1) Collapsing to the trivial solution, resulting in all outputs being assigned to a single category and (2) Interrupting the learning process due to unreliable (wrong) predictions. To alleviate these problems, we propose (1) category-agnostic repulsion loss to increase the distances between visual features for preventing collapse and (2) a reliable data sampling (RDS) method to reduce the effects of unreliable predictions. For RDS, we devise two reliable indicators named posterior probability indicator (PPI) and temporally reversed consistency indicator (TRCI). The PPI selects reliable samples based on the probability of prediction. For additional sampling, TRCI utilizes the characteristics of events that temporally reversed ones should be categorized into the same category as the original events. This reliable data sampling enables stable learning even with the predicted categories, leading to a substantial performance improvement in object recognition. Furthermore, to restore the local details, we introduce the local-global reconstruction consistency. Finally, to enhance the accuracy of recognition and visual quality of reconstructed images, we expand our method by employing the existing unpaired images. Extensive experiments demonstrate that our method shows outstanding results on the subset of the N-ImageNet dataset over unsupervised and image reconstruction-based approaches and even surpasses the supervised methods on the N-Caltech101 dataset without using labels and images. In addition, we show that our framework can be extended to event-based zero-shot object recognition.
+
+Our work presents four main contributions: (I) We propose a novel joint learning framework for event-based object recognition and image reconstruction, without requiring paired images and labels. (II) We introduce a reliable data sampling strategy and local-global reconstruction consistency that enhances label-free joint training. (III) We
+
+further introduce employing unpaired images in our framework for boosting performance. (IV) We conduct extensive experiments of our method along with zero-shot object recognition and superset categories learning, and demonstrate our method's superior performance and effectiveness.
+
+# 2. Related Work
+
+Event-based Object Recognition. Concurrent with the success of event-based low-level tasks [9, 10, 11, 23, 52, 54, 63], there has been a surge of active research focused on object recognition [3, 7, 19, 22, 27, 28, 31, 35, 43, 51, 53, 56], which is considered as a fundamental task. Considering the unique features of event data, numerous works [1, 4, 22, 24, 25, 26, 30, 32, 34, 36, 40, 53] have presented models and representations that can effectively extract significant patterns. However, these approaches require labels for each event data for supervised learning, and since event data cannot be crawled from the web like images, the labeling cost of event data is extremely high. On the other hand, our approach enables the learning of event-based object recognition and image reconstruction simultaneously, without requiring category labels and paired images.
+
+Event-based Image Reconstruction is a well-known topic in the field of event-based vision [2, 13, 17, 20, 21, 46, 49, 57]. Recent researches have been utilizing deep learning to generate realistic outcomes [44, 45, 50, 62]. However, these supervised approaches require the paired images of each event stream to be accurately aligned and synchronized in terms of pixels and time. Another line of approach, such as Wang et al. [59] and Pini et al. [41], utilized generative adversarial networks (GANs) to reconstruct intensity with unpaired images that are not related to the events of the scene. However, GANs are prone to be sensitive to various parameters, and when images that are significantly irrelevant to objects are utilized, the quality of the reconstructed images from events is notably low. An alternative solution is a self-supervised method [37] that utilizes photometric consistency, but this approach requires a continuous sequence that spans an adequately long period. Differently, our framework can reconstruct the intensity of event data without images, very long event sequence and GANs. We only exploit the category prompts of a bundle of event data, not the class label of each event data.
+
+Label-free Learning aims to solve the end-task only with unlabeled data, and has been actively explored [8, 58]. HOTS [25] proposed a category-free method for recognizing objects based on events and SSL-E2VID [37] proposed a self-supervised learning method for reconstructing images from only events. However, their performances fall short of supervised learning methods with significant performance gaps. Inspired by the success of using the pretrained model [5, 6, 39, 48, 55], we adopt CLIP [42] for label-free event-based object recognition. We propose a
+
+![](images/b9c0ce00c40f448ae65cf3529781931084efaa6040745f51c21d4fd1af0893e9.jpg)  
+Figure 2: Overall joint learning framework of event-based object recognition with event-to-image reconstruction. Our approach involves CLIP's textual and visual encoders that derive textual and visual features, respectively, from given class sets and events. Textual features are obtained from category prompts, while visual features are obtained from reconstructed images. Then, logits and corresponding pseudo labels are generated by computing the cosine similarity between visual and textual features. Lastly, our framework employs pseudo labels to facilitate joint learning using category-agnostic repulsion loss and category-guided attraction loss. To mitigate the impact of unreliable pseudo labels and boost performance, we introduce a reliable data sampling strategy and local-global reconstruction consistency.
+
+novel joint learning framework that uses visual and textual knowledge to achieve superior performance in event-to-image reconstruction and event-based object recognition without category labels and paired images. To the best of our knowledge, we are the first to incorporate textual knowledge into event vision research, particularly in the context of label-free tasks.
+
+# 3. Method
+
+In Sec. 3.1, we first introduce our joint formulation of event-to-image reconstruction and object recognition for label-efficient learning. In Sec. 3.2, we describe our reliable data sampling (RDS) strategy for boosting joint learning performance by filtering the unreliable data samples during training. Sec. 3.3 introduces the local-global reconstruction consistency for spatially regularizing the reconstructed images, which ensures capturing the local details. In Sec. 3.4, we propose to further employ the unpaired images for better recognition performance and reconstructed image quality.
+
+# 3.1. Joint Formulation
+
+We propose a joint formulation of event-to-image reconstruction and object recognition to perform event-based (zero-shot) object recognition without using labels and paired images of event data. Our joint formulation is based on the fact that 1) image reconstruction can be helpful in
+
+predicting categories from events, and 2) category information is important in reconstructing images from events. Therefore, we propose 1) the process of performing event-to-image reconstruction and CLIP-driven object recognition from the reconstructed images and 2) category-guided attraction loss and category-agnostic repulsion loss for learning this process without labels and paired images. The overall framework is shown in Fig. 2. Note that our joint learning framework can reconstruct intensity images from events without using any images.
+
+Event-to-Image Reconstruction. For the joint formulation, we first reconstruct the intensity image  $\mathcal{I}$  from the event stream  $\mathcal{E} = (u_i,t_i,p_i)_i = 1$ , where  $u_{i},t_{i}$ , and  $p_i$  denote triggered pixel location, time, and polarity of each event, and  $N$  is the temporal length of the event stream. We convert the event stream into event spiking tensor (EST) [16], which is then fed into the image reconstruction network  $\mathcal{G}$  to obtain the intensity image. The image reconstruction process is formulated as follows:
+
+$$
+\mathcal {I} = \mathcal {G} (\operatorname {E S T} (\mathcal {E})). \tag {1}
+$$
+
+CLIP-driven Object Recognition. We perform object recognition on top of the reconstructed images instead of raw event data. The advantage of using reconstructed images rather than using event data directly for object recognition is that better performance can be achieved by using
+
+a network trained on rich image data [45, 60]. In particular, we use CLIP [42] for object recognition, which also enables zero-shot object recognition. For an event dataset of  $C$  classes, the category prompts are constructed by placing each category name into a pre-defined template, "image of a [CLASS]." Then, CLIP's textual encoder encodes category prompts to obtain the  $D$ -dimensional textual features  $F = (f_{i})_{i = 1}^{C}\in \mathbb{R}^{C\times D}$ . Meantime, the reconstructed image  $\mathcal{I}$  is encoded by CLIP's visual encoder to a visual feature  $v\in \mathbb{R}^{1\times D}$  and the class probability  $p\in \mathbb{R}^{1\times C}$  for the object recognition are obtained as,
+
+$$
+p _ {i} = \operatorname {s o f t m a x} _ {i} \left(v F ^ {\top}\right), \tag {2}
+$$
+
+where  $p_i$  and  $\mathrm{softmax}_i(\cdot)$  denote the probability for category  $i$  and the softmax function. By taking the category of maximum probability, we can predict the category  $c$  for the event data  $\mathcal{E}$  as,
+
+$$
+c = \underset {i} {\operatorname {a r g m a x}} \left(p _ {i}\right). \tag {3}
+$$
+
+Category-guided Attraction Loss. The challenging part of our joint formulation is to train the reconstruction network  $\mathcal{G}$  without using paired images of event data. Once we reconstruct the intensity images from the event stream (Eq. 1), we can perform object recognition using CLIP (Eq. 2). Motivated by the fact that category information is important in reconstructing images from events, we propose category-guided attraction loss to train the reconstruction network. The proposed category-guided attraction loss brings the visual feature  $v$  encoded from the reconstructed image  $\mathcal{I}$  closer to the textual feature of the corresponding category, guiding the model to reconstruct the image corresponding to the category. Since category labels are not available for the event data, we use the predicted categories as pseudo labels. For event data with a batch size of  $B$ , the category-guided attraction loss  $\mathcal{L}_{att}$  is formulated as InfoNCE loss [33]:
+
+$$
+\mathcal {L} _ {a t t} = - \sum_ {i = 1} ^ {B} \log \frac {\exp \left(v _ {i} f _ {c _ {i}} ^ {\top}\right)}{\sum_ {j = 1} ^ {B} \exp \left(v _ {i} f _ {c _ {j}} ^ {\top}\right)}, \tag {4}
+$$
+
+where  $v_{i}, c_{i}$ , and  $f_{j}$  denote the visual feature and predicted category of the  $i^{\mathrm{th}}$  event data, and the textual feature of category  $j$ . Minimizing  $\mathcal{L}_{att}$  enables the image reconstruction network to generate images of the corresponding category label, while object recognition benefits from the reconstructed images.
+
+Category-agnostic Repulsion Loss. Although  $\mathcal{L}_{att}$  enables joint learning, the use of pseudo labels may cause collapsing, resulting in all outputs being assigned to a single category. Therefore, we introduce a category-agnostic repulsion loss to prevent collapsing by increasing the distances between visual features. To be specific, the category-agnostic repulsion loss  $\mathcal{L}_{rep}$  is applied to the set of visual features as follows:
+
+$$
+\mathcal {L} _ {r e p} = \sum_ {i = 1} ^ {B} \log \left(1 + \sum_ {j = 1, j \neq i} ^ {B} \exp \left(v _ {i} v _ {j} ^ {\top}\right)\right). \tag {5}
+$$
+
+With  $\mathcal{L}_{att}$  and  $\mathcal{L}_{rep}$ , we can successfully formulate our joint learning framework, enabling simultaneous event-to-image reconstruction and (zero-shot) object recognition.
+
+# 3.2. Reliable Data Sampling
+
+As  $\mathcal{L}_{att}$  relies on pseudo labels, the reliability of predictions may affect the overall training procedure and performance. To reduce the effects of unreliable predictions, we devise two reliability indicators named posterior probability indicator (PPI) and temporally reversed consistency indicator (TRCI). Based on the resultant reliability of predictions, we sample data with reliable predictions for  $\mathcal{L}_{att}$ . The remaining unreliable data is excluded from calculating  $\mathcal{L}_{att}$  and used only for  $\mathcal{L}_{rep}$ .
+
+Posterior Probability Indicator. Given the reconstructed image  $\mathcal{I}$  and the predicted category  $c$ , the posterior probability  $p_c = \mathbf{p}(c|\mathcal{I})$  indicates the model confidence for category  $c$ . Thus, we use the posterior probability as an indicator of the reliability of the prediction. In particular, we select  $K$  data samples corresponding to the top- $K$  posterior probabilities among  $B$  data samples in the mini-batch. The set of selected sample indices from PPI,  $S_{\mathrm{PPI}}$  is defined as,
+
+$$
+S _ {\mathrm {P P I}} = \underset {i} {\operatorname {a r g s o r t}} (\mathbf {p} (c _ {i} | \mathcal {I} _ {i})) ] _ {K}, \tag {6}
+$$
+
+where  $\mathcal{I}_i$  and  $[\cdot ]_K$  denote the reconstructed image of the  $i^{\mathrm{th}}$  event data in the mini-batch and the set of the first  $K$  items. The  $\operatorname {argsort}(\cdot)$  function returns indices that would sort an input in descending order.
+
+Temporally Reversed Consistency Indicator. Considering the temporal aspect of event data, we propose an additional reliability indicator based on the consistency between predictions from the original events and the temporally reversed events. From the fact that the temporally reversed events  $\mathcal{E}^R = (u_i,\max_j(t_j) - t_i,p_i)_{i = N}^1$  should be categorized into the same category as the original events  $\mathcal{E}$ , we presume that the consistent model prediction would indicate the reliability of the data sample. To be specific, the set of selected sample indices from TRCI,  $S_{\mathrm{TRCI}}$  is defined as follows:
+
+$$
+S _ {\mathrm {T R C I}} = \{i \mid c _ {i} = c _ {i} ^ {R}, i = 1, 2, \dots , B \}, \tag {7}
+$$
+
+where  $c_{i}^{R}$  denotes a predicted category of  $i^{\mathrm{th}}$  reversed event data  $\mathcal{E}_i^R$  in the mini-batch.
+
+Our final set of selected sample indices  $S_{\mathrm{RDS}}$  for reliable data sampling (RDS) is the intersection of  $S_{\mathrm{PPI}}$  and  $S_{\mathrm{TRCI}}$ :
+
+$$
+S _ {\mathrm {R D S}} = S _ {\mathrm {P P I}} \cap S _ {\mathrm {T R C I}}. \tag {8}
+$$
+
+Then, the modified attraction loss with RDS is as follows:
+
+$$
+\mathcal {L} _ {a t t} = - \sum_ {i \in S _ {\mathrm {R D S}}} \log \frac {\exp \left(v _ {i} f _ {c _ {i}} ^ {\top}\right)}{\sum_ {j \in S _ {\mathrm {R D S}}} \exp \left(v _ {i} f _ {c _ {j}} ^ {\top}\right)}, \tag {9}
+$$
+
+# 3.3. Local-global Reconstruction Consistency
+
+Along with  $\mathcal{L}_{att}$  and  $\mathcal{L}_{rep}$ , we further introduce the local-global reconstruction consistency for spatially regularizing the reconstructed images in a self-supervised manner. With random crop function  $\mathcal{H}$ , we first crop the EST of original events  $\mathrm{EST}(\mathcal{E})$  to obtain the cropped tensor  $\mathcal{H}(\mathrm{EST}(\mathcal{E}))$  for the local input. Then, the local-global reconstruction consistency loss  $L_{con}$  is defined as,
+
+$$
+\mathcal {L} _ {\text {c o n}} = \| \mathcal {G} (\mathcal {H} (\operatorname {E S T} (\mathcal {E}))) - \mathcal {H} (\mathcal {G} (\operatorname {E S T} (\mathcal {E}))) \| _ {1}. \tag {10}
+$$
+
+Compared to  $\mathcal{L}_{att}$  and  $\mathcal{L}_{rep}$  that work on visual features,  $L_{con}$  is applied directly to the reconstructed images and considered as a spatial regularization for the reconstruction network  $\mathcal{G}$ . In addition, matching the reconstructed images from local and global inputs enables the network to capture the local details, improving the reconstruction quality.
+
+The total loss function for the joint learning framework is as follows:
+
+$$
+\mathcal {L} _ {\text {t o t a l}} = \lambda_ {1} \mathcal {L} _ {\text {a t t}} + \lambda_ {2} \mathcal {L} _ {\text {r e p}} + \lambda_ {3} \mathcal {L} _ {\text {c o n}}, \tag {11}
+$$
+
+where  $\lambda_1, \lambda_2,$  and  $\lambda_3$  are weights for  $\mathcal{L}_{att},\mathcal{L}_{rep}$ , and  $\mathcal{L}_{con}$ .
+
+# 3.4. Employing Unpaired Images
+
+In this section, we expand the use case of our method by introducing how to employ the existing unpaired images. Here, we assume that the category of each image is known, which is the common case for web crawling. Although our framework is capable of event-to-image reconstruction and object recognition without using any images and labels, utilizing existing images can possibly improve the image reconstruction quality and consequently improve object recognition performance. As shown in Fig. 3, we replace textual features in  $\mathcal{L}_{att}$  with the prototype features generated from unpaired images. For each category  $c$ , we first encode all images of category  $c$  and apply a clustering algorithm to obtain  $L$  clusters in total. Then, we aggregate the features of each cluster by averaging them, resulting in  $L$  prototype features  $W^{c} = (w_{i}^{c})_{i=1}^{L} \in \mathbb{R}^{L \times D}$ .
+
+For data sample index  $i \in S_{\mathrm{RDS}}$  with corresponding visual feature  $v_{i}$  and the predicted category  $c_{i}$ , we replace the textual features  $f_{c_i}$  in Eq. 9 to the closest prototype feature  $w_{l_i}^{c_i}$ , where the cluster index  $l_{i}$  is calculated as,
+
+$$
+l _ {i} = \underset {j} {\operatorname {a r g m a x}} \left(v _ {i} \left(w _ {j} ^ {c _ {i}}\right) ^ {\top}\right). \tag {12}
+$$
+
+Our modification of  $\mathcal{L}_{att}$  to employ unpaired images is then defined as follows:
+
+$$
+\mathcal {L} _ {a t t} = - \sum_ {i \in S _ {\mathrm {R D S}}} \log \frac {\exp \left(v _ {i} \left(w _ {l _ {i}} ^ {c _ {i}}\right) ^ {\top}\right)}{\sum_ {j \in S _ {\mathrm {R D S}}} \exp \left(v _ {i} \left(w _ {l _ {j}} ^ {c _ {j}}\right) ^ {\top}\right)}, \tag {13}
+$$
+
+![](images/1ad5b3a30d84555de23d6033ead8896ecaec074745b65999f993303a6652fe14.jpg)  
+Figure 3: Proposed method employing unpaired images in the category-guided attraction loss. Unpaired images are further utilized to enhance object recognition performance and reconstructed image quality.
+
+# 4. Datasets
+
+The experiments are conducted on two widely used event-based object recognition datasets, N-Caltech101 [35] and N-ImageNet [22], which are constructed by filming Caltech101 [15] and ImageNet [14] datasets on LCD monitor with event camera. Since N-Caltech101 and N-ImageNet datasets do not provide APS frames, the image in the original dataset and corresponding event sequence have different resolutions and are not aligned in pixel-level. To create a visual prototype in Sec. 3.4, we scrap images from the original dataset, mimicking the web crawling. To avoid including paired images of events in the train set, we randomly split the event train set in half for each class. One half is used to train the model, while the corresponding images from the other half are used to create the visual prototype.
+
+N-Caltech101 dataset consists of 8,246 event stream data belonging to 101 different object classes. Each object is captured with  $320 \times 245$  resolution (average) for 300 milliseconds. Existing works [45, 60] utilized this dataset to evaluate the performance of image reconstruction and event-based object recognition.
+
+N-ImageNet is recently released large-scale dataset, containing 1,781,167 event stream data of 1,000 object classes. Compared to N-Caltech101, it has a larger resolution of  $640 \times 480$  and a very short capturing time with  $5 \times 10^{-2}$  milliseconds. Therefore, each event stream contains significantly sparse edge information. N-ImageNet is too challenging for unsupervised studies without ground truth, so we utilize a randomly sampled subset of 100 classes called N-ImageNet (Mini). Details about the split of N-ImageNet (Mini) are provided in the supplementary material.
+
+<table><tr><td rowspan="2">Methods</td><td rowspan="2">Image Pair</td><td rowspan="2">Class Label</td><td colspan="3">N-Caltech101</td><td colspan="3">N-ImageNet (Mini)</td></tr><tr><td>Accuracy ↑</td><td>FID ↓</td><td>IS ↑</td><td>Accuracy ↑</td><td>FID ↓</td><td>IS ↑</td></tr><tr><td colspan="9">Supervised Learning</td></tr><tr><td>Histogram [29]</td><td></td><td>✓</td><td>63.8</td><td>-</td><td>-</td><td>49.8</td><td>-</td><td>-</td></tr><tr><td>Ev-gait [61]</td><td></td><td>✓</td><td>61.9</td><td>-</td><td>-</td><td>51.3</td><td>-</td><td>-</td></tr><tr><td>EST [16]</td><td></td><td>✓</td><td>81.7*</td><td>-</td><td>-</td><td>52.5</td><td>-</td><td>-</td></tr><tr><td>DiST [22]</td><td></td><td>✓</td><td>73.8</td><td>-</td><td>-</td><td>57.9</td><td>-</td><td>-</td></tr><tr><td colspan="9">Unsupervised Learning</td></tr><tr><td>HOTS [25]</td><td></td><td></td><td>21.0*</td><td>-</td><td>-</td><td>1.6</td><td>-</td><td>-</td></tr><tr><td colspan="9">Image Reconstruction + CLIP</td></tr><tr><td>E2VID [44] (Pre-train)</td><td></td><td></td><td>64.0</td><td>83.90</td><td>12.81</td><td>13.3</td><td>147.14</td><td>7.82</td></tr><tr><td>E2VID [44] (Fine-tune)</td><td></td><td></td><td>59.8</td><td>85.31</td><td>12.91</td><td>13.4</td><td>144.57</td><td>7.48</td></tr><tr><td>E2VID [44] (Fine-tune)</td><td>✓</td><td></td><td>59.8</td><td>86.66</td><td>12.61</td><td>13.5</td><td>144.03</td><td>7.40</td></tr><tr><td>E2VID [44] (Fine-tune)</td><td></td><td>✓</td><td>60.1</td><td>85.50</td><td>12.85</td><td>12.7</td><td>145.79</td><td>7.33</td></tr><tr><td>E2VID [44] (Scratch)</td><td>✓</td><td></td><td>9.4</td><td>302.48</td><td>2.60</td><td>1.0</td><td>299.70</td><td>1.15</td></tr><tr><td>SSL-E2VID [37] (Pre-train)</td><td></td><td></td><td>28.2</td><td>281.09</td><td>2.36</td><td>1.6</td><td>324.43</td><td>2.08</td></tr><tr><td>SSL-E2VID [37] (Scratch)</td><td></td><td></td><td>30.5</td><td>228.30</td><td>4.37</td><td>1.7</td><td>250.55</td><td>3.00</td></tr><tr><td>Wang et al. [59] (Scratch)</td><td></td><td></td><td>42.7</td><td>160.34</td><td>6.83</td><td>10.1</td><td>231.94</td><td>5.00</td></tr><tr><td>Wang et al. [59] (Scratch)</td><td>✓</td><td></td><td>43.5</td><td>72.67</td><td>4.50</td><td>19.0</td><td>180.76</td><td>4.89</td></tr><tr><td>Wang et al. [59] (Scratch)</td><td></td><td>✓</td><td>39.7</td><td>179.79</td><td>4.18</td><td>13.2</td><td>216.37</td><td>4.70</td></tr><tr><td>Ours (Text Prompt)</td><td></td><td></td><td>82.46</td><td>62.29</td><td>14.81</td><td>30.16</td><td>150.58</td><td>7.92</td></tr><tr><td>Ours (Visual Prototype)</td><td></td><td></td><td>82.61</td><td>54.19</td><td>17.59</td><td>31.28</td><td>135.95</td><td>9.74</td></tr></table>
+
+Table 1: The results of event-based object recognition and event-to-image reconstruction. The **bold** and **underline** denote the best and the second-best performance except for supervised methods, respectively. * denotes the values taken from the original paper and italic denotes the performance of supervised methods.
+
+# 5. Experiments
+
+We compare our methods utilizing text prompts or visual prototypes with other existing methods. We adopt four representations (Histogram [29], Ev-gait [61], EST [16], and DiST [22]) for supervised methods and an unsupervised method (HOTS [25]). Additionally, three image reconstruction networks, E2VID [44], SSL-E2VID [37] and Wang et al. [59], are utilized with CLIP. For a fair comparison with image reconstruction networks, we record the performance of three different models for each image reconstruction network: the official pre-trained model, fine-tuned model, and a model trained from scratch. We evaluate these models under three conditions: paired images, images belonging to the same class, and randomly selected images. Some cases are overlooked when the pre-trained model is unavailable or the training setup does not yield meaningful results. For implementation details, see the supplementary material.
+
+# 5.1. Label-free Object Recognition
+
+Text Prompt. The quantitative results are presented in Table 1. The proposed method outperforms supervised object recognition methods on the N-Caltech101 dataset, despite not using category labels. Also, our method surpasses existing unsupervised methods and image reconstruction methods with CLIP by a large margin under all
+
+metrics. Specifically, ours surpasses the best image reconstruction with CLIP method, E2VID (Pre-train), by around  $28\%$  increase in accuracy even though it utilizes the image while pre-training. On N-ImageNet, supervised methods record higher accuracy than ours since the dataset is challenging due to the sparse events captured within a short period. However, ours shows a significant improvement in all metrics compared to other unsupervised and image reconstruction-based methods. Our approach demonstrates the ability to perform object recognition and image reconstruction from event data, even when trained solely on textual features without any image data.
+
+Visual Prototype. Besides text prompts, our framework can employ visual prototypes generated from scrapped images. By incorporating visual information, our visual prototype-based approach exceeds the performance of relying solely on textual prompts and achieves the best performance among unsupervised and image reconstruction-based methods under all metrics across both datasets. This demonstrates the superiority of the visual prototypes.
+
+# 5.2. Image-free Image Reconstruction
+
+We also compare the reconstruction results of our methods with other event-to-image reconstruction networks in Table 1. Note that, as mentioned in Sec. 4, the resolution and pixel positions of events and corresponding orig-
+
+![](images/7812e7b24f7ef6d5ab6206f30e47da0bd3c45143d93840596bb805df70d09fcc.jpg)  
+Events  
+Wang et. al.  
+E2VID  
+E2VID*  
+Ours (Prompt)  
+Ours (Proto.)  
+Ours (Proto.)*  
+Image  
+(a) N-Caltech101
+
+![](images/e58e7f9c7861d171cda9424194bfc75d07421857ace6eb51493299e456148679.jpg)  
+(b) N-ImageNet  
+Figure 4: Visual results of event-to-image reconstruction network on (a) N-Caltech101 and (b) N-ImageNet datasets. Ours shows a higher quality of reconstructed images with fine-grained details than other methods (Wang et al. [59] and pre-trained E2VID [44]). * denotes the images with brightness and contrast adjustment.
+
+inal images are not aligned. Therefore, pixel-wise metrics (e.g., L1, PSNR) cannot be used for comparison. Instead, to assess the visual quality of the reconstructed images, we utilize two metrics: inception score (IS) [47] and Fréchet inception distance (FID) [18].
+
+Despite not utilizing any images during training, our text prompt-based approach surpasses all other image reconstruction techniques in terms of FID and IS metrics on the N-Caltech101 dataset. However, in the case of N-ImageNet, our text prompt approach is the strongest by an IS indicator but falls behind pre-trained E2VID in FID. Since ours is trained from events of short periods without paired images, it becomes challenging to learn the distribution of real images. Instead, utilizing the visual prototype results in the highest FID and IS indicators in both datasets. This implies
+
+that it can acquire the ability to reconstruct images without requiring paired images that correspond to events.
+
+We also present the qualitative results in Fig. 4. Wang et al. [59] utilize the paired image in training but the method lacks the ability to effectively display intensity contrast, resulting in the loss of many details. While pre-trained E2VID shows better image quality over [59], it still struggles to preserve fine details as it operates in a distinct training environment. This is particularly apparent in the N-ImageNet dataset, where E2VID generates results that deviate significantly from the original image. Conversely, our framework is able to recover intensity effectively and preserve details in reconstructed images, even without access to paired images. Our results closely resemble real images, even in the challenging N-ImageNet dataset.
+
+<table><tr><td rowspan="2">Methods</td><td rowspan="2">Image Pair</td><td rowspan="2">Class Label</td><td colspan="3">N-Caltech101</td><td colspan="3">N-ImageNet (Mini)</td></tr><tr><td>Accuracy ↑</td><td>FID ↓</td><td>IS ↑</td><td>Accuracy ↑</td><td>FID ↓</td><td>IS ↑</td></tr><tr><td colspan="9">Image Reconstruction + CLIP</td></tr><tr><td>E2VID [44] (Pre-train)</td><td></td><td></td><td>58.4</td><td>194.26</td><td>6.47</td><td>31.7</td><td>203.05</td><td>6.64</td></tr><tr><td>E2VID [44] (Fine-tune)</td><td></td><td></td><td>61.7</td><td>203.39</td><td>7.56</td><td>30.0</td><td>199.83</td><td>6.73</td></tr><tr><td>E2VID [44] (Fine-tune)</td><td>✓</td><td></td><td>61.1</td><td>203.40</td><td>7.45</td><td>30.9</td><td>199.44</td><td>6.80</td></tr><tr><td>E2VID [44] (Fine-tune)</td><td></td><td>✓</td><td>61.5</td><td>202.23</td><td>7.56</td><td>29.4</td><td>203.97</td><td>6.08</td></tr><tr><td>E2VID [44] (Scratch)</td><td>✓</td><td></td><td>21.3</td><td>362.08</td><td>2.43</td><td>4.9</td><td>333.13</td><td>1.36</td></tr><tr><td>SSL-E2VID [37] (Pre-train)</td><td></td><td></td><td>24.7</td><td>332.43</td><td>2.07</td><td>5.7</td><td>352.53</td><td>2.04</td></tr><tr><td>SSL-E2VID [37] (Scratch)</td><td></td><td></td><td>25.3</td><td>289.36</td><td>3.93</td><td>6.1</td><td>287.98</td><td>2.79</td></tr><tr><td>Wang et al. [59] (Scratch)</td><td></td><td></td><td>47.5</td><td>228.77</td><td>4.67</td><td>18.7</td><td>241.87</td><td>6.12</td></tr><tr><td>Wang et al. [59] (Scratch)</td><td>✓</td><td></td><td>48.5</td><td>230.74</td><td>4.91</td><td>20.5</td><td>236.78</td><td>3.90</td></tr><tr><td>Wang et al. [59] (Scratch)</td><td></td><td>✓</td><td>43.7</td><td>215.63</td><td>4.66</td><td>15.6</td><td>237.38</td><td>4.40</td></tr><tr><td>Ours (Text Prompt)</td><td></td><td></td><td>84.31</td><td>147.31</td><td>6.81</td><td>45.10</td><td>202.29</td><td>6.31</td></tr><tr><td>Ours (Visual Prototype)</td><td></td><td></td><td>85.56</td><td>107.58</td><td>9.35</td><td>51.50</td><td>174.24</td><td>7.32</td></tr></table>
+
+# 5.3. Zero-shot Object Recognition
+
+We further demonstrate the advantages of our framework over supervised methods by enabling it to be used for zero-shot object recognition. The reconstructed images can be utilized for zero-shot object recognition thanks to CLIP's capacity to align images with any semantic concepts in an open vocabulary for zero-shot object recognition. To define "unseen" categories that are not seen during training, we use only 80 of the 100 categories for training and the remaining 20 for zero-shot testing.
+
+In Table 2, we present the performance of zero-shot object recognition. Without any category labels, our methods with text prompt achieve a promising  $84.31\%$  and  $45.10\%$  on N-Caltech101 and N-ImageNet, respectively. This is  $25.91\%$  and  $13.4\%$  higher than E2VID pretrained with a large number of image pairs. These results validate that although our model learns through the process of being category-aware, generalization is possible for other categories, especially "unseen" in this case. Although FID and IS of our method using text prompts are relatively poor compared to pre-trained E2VID on N-ImageNet, this can also be overcome by using visual prototypes. Our method using the visual prototype is able to achieve a substantial improvement under all metrics. Specifically, in the N-Caltech101 dataset, our method increases the IS from 6.81 to 9.35, while in the N-ImageNet dataset, it increases the accuracy from 45.10 to 51.50.
+
+# 6. Ablation Study and Discussion
+
+# 6.1. Ablation Study
+
+We conduct an ablation study to understand how the components influence performance. Table 3 illustrates the effectiveness of each component on N-ImageNet dataset.
+
+Table 2: The results of event-based zero-shot object recognition using CLIP.  
+
+<table><tr><td rowspan="2"></td><td rowspan="2">Latt</td><td rowspan="2">Lrep</td><td rowspan="2">Lcon</td><td colspan="2">RDS</td><td rowspan="2">Accuracy ↑</td></tr><tr><td>PPI</td><td>TRCI</td></tr><tr><td>(1)</td><td>✓</td><td></td><td></td><td></td><td></td><td>9.50</td></tr><tr><td>(2)</td><td>✓</td><td>✓</td><td></td><td></td><td></td><td>14.00</td></tr><tr><td>(3)</td><td>✓</td><td>✓</td><td>✓</td><td></td><td></td><td>15.68</td></tr><tr><td>(4)</td><td>✓</td><td>✓</td><td></td><td>✓</td><td></td><td>28.42</td></tr><tr><td>(5)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>28.90</td></tr><tr><td>(6)</td><td>✓</td><td>✓</td><td></td><td>✓</td><td>✓</td><td>28.91</td></tr><tr><td>(7)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>30.16</td></tr></table>
+
+Table 3: Ablation study of the proposed method.
+
+We start with a baseline, which denotes the removal of everything except attraction loss,  $\mathcal{L}_{att}$ . During the experiment, we notice that if we only apply attraction loss towards the network without repulsion loss,  $\mathcal{L}_{rep}$ , the collapse often takes place. Therefore,  $\mathcal{L}_{rep}$  is essential for stable learning of  $\mathcal{L}_{att}$ . It is worth mentioning the impact of the proposed PPI on performance. Upon analyzing (4), which displays the results when PPI is added to (2), it can be seen that the accuracy has increased by 14.42. Upon examining (5) and (7), it can be observed that the accuracy increases by 1.26, validating the impact of the sub-sampling process via TRCI. Furthermore, the ablation results between (6) and (7) show that the reconstruction consistency between global and local preserves the spatial details of the image, leading to performance improvement. Finally, the model (7) that has all of its components incorporated achieves the highest performance.
+
+# 6.2. Number of Samples in PPI
+
+The  $K$  used in PPI for selecting the sample with the highest probability impacts the performance of object recognition. Fig. 5 shows the evaluation of performance on N-ImageNet dataset based on different values of  $K$ . When  $K$  is set to 32, which includes all samples in the batch, the
+
+![](images/eb8c64942a8dac1e10eb30d1237c9becc4a3d28ea5a4f77e31027d0cad741418.jpg)  
+Figure 5: Object recognition accuracy on N-ImageNet (Mini) according to  $K$  in PPI.
+
+![](images/d94d636714e43c46ad2e957cfeb9cc1cfa7d442abd6bce46b365f99be7aa9a4a.jpg)  
+Figure 6: Object recognition accuracy on N-Caltech101 according to  $K$  in PPI.
+
+<table><tr><td></td><td colspan="3">N-Caltech101</td><td colspan="3">N-ImageNet (Mini)</td></tr><tr><td>λ2\λ3</td><td>0.5</td><td>1</td><td>2</td><td>0.5</td><td>1</td><td>2</td></tr><tr><td>0.001</td><td>82.71</td><td>82.93</td><td>81.26</td><td>29.62</td><td>26.34</td><td>30.08</td></tr><tr><td>0.005</td><td>83.52</td><td>82.02</td><td>81.66</td><td>29.38</td><td>26.92</td><td>28.72</td></tr><tr><td>0.01</td><td>82.24</td><td>82.46</td><td>81.99</td><td>29.34</td><td>30.16</td><td>29.80</td></tr><tr><td>0.05</td><td>81.99</td><td>82.39</td><td>82.61</td><td>27.50</td><td>27.92</td><td>29.00</td></tr><tr><td>0.1</td><td>81.37</td><td>82.46</td><td>82.17</td><td>25.76</td><td>26.28</td><td>26.32</td></tr></table>
+
+Table 4: The accuracy according to hyper-parameter in Eq. 11 when  $\lambda_1 = 1$ .
+
+attraction loss,  $\mathcal{L}_{att}$ , is applied to even those samples with low reliability, resulting in the poorest performance. Due to the same reason, the performance is better when  $K$  is 16 than when it is 32, but since unreliable samples are also selected, the accuracy is relatively lower than in other favorable situations. As  $K$  decreases, the performance improves, but it starts to decline when  $K$  is less than 6, indicating that even with high reliability, it is challenging to achieve optimal training if  $\mathcal{L}_{att}$  is applied with very few samples. Similarly, we report the accuracy according to the  $K$  value on the N-Caltech101 in Fig 6. Unless an extremely small number of samples are drawn, as the value of  $K$  increases, the performance also improves. However, once a certain threshold is reached, there is little significant change, and it starts to decline with further increases. The accuracy exceeds  $80\%$  in all cases, showing robustness to  $K$  values. Although the level of sensitivity varies depending on the dataset, we do confirm that satisfactory results can be obtained between  $K = 4$  and  $K = 10$ .
+
+# 6.3. Hyper-parameter Analysis
+
+In Table 4, we report the accuracy according to  $\lambda_{2}$  and  $\lambda_{3}$  in Eq. 11. If the weight of the repulsion loss,  $\lambda_{2}$ , is set too high, the performance decreases. However, this can be mitigated by increasing the weight of the local-global re
+
+![](images/d578b2cc4dbd1da4c37ac78ccae19e553041cace94a5a3168bb1d7da94a52d11.jpg)  
+Figure 7: Object recognition accuracy on N-Caltech101 according to the number of extra category prompts.
+
+construction consistency loss,  $\lambda_3$ . The results verify the robustness of our method according to the  $\lambda$  on both datasets.
+
+# 6.4. Learning with Superset Categories
+
+In practical terms, it can be challenging for users to find corresponding categories of the event dataset. As a result, they often end up taking more categories as prompts. We tackle these issues by demonstrating that our joint learning framework works robustly even when employing more prompts than the categories of the event on the N-Caltech101 dataset. To avoid using additional prompts that are too irrelevant to the N-Caltech101 dataset, we randomly take additional prompts from N-ImageNet. We show the performance results according to the number of extra category prompts in Fig. 7. While the performance decreases with an increase in extra category prompts, it is worth noting that our algorithm remains reliable even when faced with over 100 additional prompts beyond the corresponding category. This suggests that our algorithm can function with superset categories, even in a situation when the number of extra categories is much larger than the actual categories.
+
+# 7. Conclusion
+
+In this work, we propose a joint learning framework for event-based object recognition with image reconstruction from events without labels and paired images. With the proposed losses and the reliable data sampling strategy, our approach allows stable learning of two tasks simultaneously from pseudo labels. As a result, our approach achieves exceptional performance in event-based object recognition and event-to-image reconstruction, all without requiring category labels and paired images. Our approach successfully addresses the challenge of label dependency in event-based research, which has been a significant obstacle. We anticipate that our work can be further extended to other tasks such as event-based zero-shot object detection, zero-shot semantic segmentation, etc., that have not explored so far in the research field for event cameras.
+
+Acknowledgements. This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF-2022R1A2B5B03002636).
+
+# References
+
+[1] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7243-7252, 2017. 1, 2  
+[2] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 884-892, 2016. 2  
+[3] Yin Bi, Aaron Chadha, Alhabib Abbas, Eirina Bourtsoulatze, and Yiannis Andreopoulos. Graph-based object classification for neuromorphic vision sensing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 491-501, 2019. 2  
+[4] Marco Cannici, Marco Ciccone, Andrea Romanoni, and Matteo Matteucci. A differentiable recurrent surface for asynchronous event-based data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16, pages 136–152. Springer, 2020. 1, 2  
+[5] Hila Chefer, Sagie Benaim, Roni Paiss, and Lior Wolf. Image-based clip-guided essence transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 695-711. Springer, 2022. 2  
+[6] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. arXiv preprint arXiv:2301.04926, 2023. 2  
+[7] Wensheng Cheng, Hao Luo, Wen Yang, Lei Yu, Shoushun Chen, and Wei Li. Det: A high-resolution dvs dataset for lane extraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 2  
+[8] Hoonhee Cho, Jegyeong Cho, and Kuk-Jin Yoon. Learning adaptive dense event stereo from the image domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17797-17807, 2023. 2  
+[9] Hoonhee Cho, Jaeseok Jeong, and Kuk-Jin Yoon. Eomvs: Event-based omnidirectional multi-view stereo. IEEE Robotics and Automation Letters, 6(4):6709-6716, 2021. 2  
+[10] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 454-462, 2022. 2  
+[11] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In European Conference on Computer Vision, pages 470-486. Springer, 2022. 2  
+[12] Gregory Cohen, Saeed Afshar, Garrick Orchard, Jonathan Tapson, Ryad Benosman, and Andre van Schaik. Spatial and temporal downsampling in event-based visual classification. IEEE Transactions on Neural Networks and Learning Systems, 29(10):5030-5044, 2018. 1  
+[13] Matthew Cook, Luca Gugelmann, Florian Jug, Christoph Krautz, and Angelika Steger. Interacting maps for fast visual
+
+interpretation. In The 2011 International Joint Conference on Neural Networks, pages 770-776. IEEE, 2011. 2  
+[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5  
+[15] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 5  
+[16] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5633-5643, 2019. 3, 6  
+[17] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Asynchronous, photometric feature tracking using events and frames. In Proceedings of the European Conference on Computer Vision (ECCV), pages 750-765, 2018. 2  
+[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7  
+[19] Yuhuang Hu, Hongjie Liu, Michael Pfeiffer, and Tobi Delbruck. Dvs benchmark datasets for object tracking, action recognition, and object recognition. Frontiers in neuroscience, 10:405, 2016. 2  
+[20] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew J Davison. Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ, 43:566-576, 2008. 2  
+[21] Hanme Kim, Stefan Leutenegger, and Andrew J Davison. Real-time 3d reconstruction and 6-dof tracking with an event camera. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 349-364. Springer, 2016. 2  
+[22] Junho Kim, Jaehyeok Bae, Gangin Park, Dongsu Zhang, and Young Min Kim. N-imagenet: Towards robust, fine-grained object recognition with event cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2146-2156, 2021. 1, 2, 5, 6  
+[23] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 2  
+[24] Simon Klenk, David Bonello, Lukas Koestler, and Daniel Cremers. Masked event modeling: Self-supervised pretraining for event cameras. arXiv preprint arXiv:2212.10368, 2022.1,2  
+[25] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E Shi, and Ryad B Benosman. Hots: a hierarchy of event-based time-surfaces for pattern recognition. IEEE transactions on pattern analysis and machine intelligence, 39(7):1346-1359, 2016. 1, 2, 6
+
+[26] Jun Haeng Lee, Tobi Delbruck, and Michael Pfeiffer. Training deep spiking neural networks using backpropagation. Frontiers in neuroscience, 10:508, 2016. 1, 2  
+[27] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 2  
+[28] Iulia-Alexandra Lungu, Federico Corradi, and Tobi Delbrück. Live demonstration: Convolutional neural network driven by dynamic vision sensor playing roshambo. In 2017 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-1. IEEE, 2017. 2  
+[29] Ana I Maqueda, Antonio Loquercio, Guillermo Gallego, Narciso García, and Davide Scaramuzza. Event-based vision meets deep learning on steering prediction for self-driving cars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5419-5427, 2018. 1, 6  
+[30] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 415-431. Springer, 2020. 1, 2  
+[31] Diederik Paul Moeys, Daniel Neil, Federico Corradi, Emmett Kerr, Philip Vance, Gautham Das, Sonya A Coleman, Thomas M McGinnity, Dermot Kerr, and Tobi Delbruck. Pred18: Dataset and further experiments with davis event camera in predator-prey robot chasing. arXiv preprint arXiv:1807.03128, 2018. 2  
+[32] Daniel Neil, Michael Pfeiffer, and Shih-Chii Liu. Phased LSTM: Accelerating recurrent network training for long or event-based sequences. Advances in neural information processing systems, 29, 2016. 1, 2  
+[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4  
+[34] Garrick Orchard, Ryad Benosman, Ralph Etienne-Cummings, and Nitish V Thakor. A spiking neural network architecture for visual motion estimation. In 2013 IEEE Biomedical Circuits and Systems Conference (BioCAS), pages 298-301. IEEE, 2013. 1, 2  
+[35] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:437, 2015. 2, 5  
+[36] Garrick Orchard, Cedric Meyer, Ralph Etienne-Cummings, Christoph Posch, Nitish Thakor, and Ryad Benosman. Hfirst: A temporal approach to object recognition. IEEE transactions on pattern analysis and machine intelligence, 37(10):2028-2040, 2015. 1, 2  
+[37] Federico Paredes-Vallés and Guido CHE de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446-3455, 2021. 2, 6, 8  
+[38] Paul KJ Park, Baek Hwan Cho, Jin Man Park, Kyoobin Lee, Ha Young Kim, Hyo Ah Kang, Hyun Goo Lee, Jooyeon Woo, Yohan Roh, Won Jo Lee, et al. Performance improvement of deep learning based gesture recognition using spa
+
+tiotemporal demosaicing technique. In 2016 IEEE International Conference on Image Processing (ICIP), pages 1624-1628. IEEE, 2016. 1  
+[39] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2  
+[40] José Antonio Pérez-Carrasco, Bo Zhao, Carmen Serrano, Begona Acha, Teresa Serrano-Gotarredona, Shouchun Chen, and Bernabé Linares-Barranco. Mapping from frame-driven to frame-free event-driven vision systems by low-rate rate coding and coincidence processing—application to feedforward convnets. IEEE transactions on pattern analysis and machine intelligence, 35(11):2706-2719, 2013. 1, 2  
+[41] Stefano Pini, Guido Borghi, and Roberto Vezzani. Learn to see by events: Color frame synthesis from event and rgb cameras. arXiv preprint arXiv:1812.02041, 2018. 2  
+[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4  
+[43] Bharath Ramesh, Andrés Ussa, Luca Della Vedova, Hong Yang, and Garrick Orchard. Low-power dynamic object detection and classification with freely moving event cameras. Frontiers in neuroscience, 14:135, 2020. 2  
+[44] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3857-3866, 2019. 2, 6, 7, 8  
+[45] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 1, 2, 4, 5  
+[46] Christian Reinbacher, Gottfried Graber, and Thomas Pock. Real-time intensity-image reconstruction for event cameras using manifold regularisation. arXiv preprint arXiv:1607.06283, 2016. 2  
+[47] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 7  
+[48] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 2  
+[49] Cedric Scheerlinck, Nick Barnes, and Robert Mahony. Continuous-time intensity estimation using event cameras. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part V, pages 308–324. Springer, 2019. 2  
+[50] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast im
+
+age reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 156-163, 2020. 2  
+[51] Teresa Serrano-Gotarredona and Bernabé Linares-Barranco. Poker-dvs and mnist-dvs. their history, how they were made, and other details. Frontiers in neuroscience, 9:481, 2015. 2  
+[52] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In European Conference on Computer Vision, pages 628-645. Springer, 2022. 2  
+[53] Amos Sironi, Manuele Brambilla, Nicolas Bourdis, Xavier Lagorce, and Ryad Benosman. Hats: Histograms of averaged time surfaces for robust event-based object classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1731-1740, 2018. 1, 2  
+[54] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision, pages 412-428. Springer, 2022. 2  
+[55] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918-17928, 2022. 2  
+[56] Ajay Vasudevan, Pablo Negri, Bernabe Linares-Barranco, and Teresa Serrano-Gotarredona. Introduction and analysis of an event-based sign language dataset. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 675-682. IEEE, 2020. 2  
+[57] Lin Wang, Yujeong Chae, and Kuk-Jin Yoon. Dual transfer learning for event-based end-task prediction via pluggable event to image translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2135-2145, 2021. 2  
+[58] Lin Wang, Yujeong Chae, Sung-Hoon Yoon, Tae-Kyun Kim, and Kuk-Jin Yoon. Evdistill: Asynchronous events to end-task learning via bidirectional reconstruction-guided cross-modal knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 608-619, 2021. 2  
+[59] Lin Wang, Yo-Sung Ho, Kuk-Jin Yoon, et al. Event-based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10081-10090, 2019. 2, 6, 7, 8  
+[60] Lin Wang, Tae-Kyun Kim, and Kuk-Jin Yoon. Joint framework for single image reconstruction and super-resolution with an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7657-7673, 2021. 1, 4, 5  
+[61] Yanxiang Wang, Bowen Du, Yiran Shen, Kai Wu, Guangrong Zhao, Jianguo Sun, and Hongkai Wen. Ev-gait: Event-based robust gait recognition using dynamic vision sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6358-6367, 2019. 6  
+[62] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Com
+
+puter Vision, pages 2563-2572, 2021. 2  
+[63] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 2

2023/Label-Free Event-based Object Recognition via Joint Learning with Image Reconstruction from Events/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4c371f5e577a70d9acd57282dd90c666710f7d2f918111b3651f44409509b1e
+size 846694

2023/Label-Guided Knowledge Distillation for Continual Semantic Segmentation on 2D Images and 3D Point Clouds/3a801de3-50c8-469a-b26b-c00f633cc663_content_list.json

@@ -0,0 +1,1526 @@
+[
+    {
+        "type": "text",
+        "text": "Label-Guided Knowledge Distillation for Continual Semantic Segmentation on 2D Images and 3D Point Clouds",
+        "text_level": 1,
+        "bbox": [
+            104,
+            111,
+            867,
+            157
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ze Yang $^{1}$ , Ruibo Li $^{1}$ , Evan Ling $^{2}$ , Chi Zhang $^{1}$ , Yiming Wang $^{1}$ , Dezhao Huang $^{2}$ , Keng Teck Ma $^{2}$ , Minhoe Hur $^{3}$ , Guosheng Lin $^{1*}$",
+        "bbox": [
+            232,
+            184,
+            743,
+            220
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ School of Computer Science and Engineering, Nanyang Technological University, Singapore  \n $^{2}$ Hyundai Motor Group Innovation Center in Singapore (HMGICS)  \n $^{3}$ AIRS Company, Hyundai Motor Group",
+        "bbox": [
+            114,
+            220,
+            854,
+            273
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ze001@e.ntu.edu.sg, evan.ling@hmgics.com, minhoe.hur@hyundai.com, gslin@ntu.edu.sg",
+        "bbox": [
+            120,
+            277,
+            846,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            325,
+            313,
+            342
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Continual semantic segmentation (CSS) aims to extend an existing model to tackle unseen tasks while retaining its old knowledge. Naively fine-tuning the old model on new data leads to catastrophic forgetting. A common solution is knowledge distillation (KD), where the output distribution of the new model is regularized to be similar to that of the old model. However, in CSS, this is challenging because of the background shift issue. Existing KD-based CSS methods continue to suffer from confusion between the background and novel classes since they fail to establish a reliable class correspondence for distillation. To address this issue, we propose a new label-guided knowledge distillation (LGKD) loss, where the old model output is expanded and transplanted (with the guidance of the ground truth label) to form a semantically appropriate class correspondence with the new model output. Consequently, the useful knowledge from the old model can be effectively distilled into the new model without causing confusion. We conduct extensive experiments on two prevailing CSS benchmarks, Pascal-VOC and ADE20K, where our LGKD significantly boosts the performance of three competing methods, especially on novel mIoU by up to  $+76\\%$ , setting new state-of-the-art. Finally, to further demonstrate its generalization ability, we introduce the first CSS benchmark for 3D point cloud based on ScanNet, along with several re-implemented baselines for comparison. Experiments show that LGKD is versatile in both 2D and 3D modalities without requiring ad hoc design. Codes are available at https://github.com/Ze-Yang/LGKD.",
+        "bbox": [
+            75,
+            358,
+            473,
+            797
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            823,
+            209,
+            839
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Fully supervised semantic segmentation has witnessed tremendous success [33, 63, 27, 8, 16, 22, 47, 65, 9] in re",
+        "bbox": [
+            76,
+            849,
+            468,
+            880
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/5a842318a5f59736e908d670e8a3502e730702509059163ea18e851911ba5ad8.jpg",
+        "image_caption": [
+            "Figure 1. Illustration of knowledge distillation (KD) strategies in continual semantic segmentation. Standard KD and unbiased KD reduce the new class (bicycle) dimension(s) of the probabilities output by the new model via remove and combine respectively, which collapses the class correspondence across incremental steps and raises confusion between background and new classes (see Sec. 1). In contrast, our label-guided KD uses the ground truth label as guidance to expand the probabilities predicted by the old model. It builds a reliable class correspondence across different learning steps without discrepancy (ILT) or entanglement (MiB) (see Sec. 3.2). A  $\\xrightarrow{\\text{mimic}}$  B: encourage A to be similar to B."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            324,
+            890,
+            571
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "cent years. These algorithms generally assume a fixed number of classes to be learned. However, in real-world applications, it is often expected that a deployed model can be continuously generalized to handle new classes while not forgetting the old ones. A simple solution is to expand the original dataset with newly available samples and retrain a new model from scratch, dubbed as Joint Training. Obviously, this is computationally expensive and requires an increasing amount of space to store the old data over time.",
+        "bbox": [
+            496,
+            763,
+            893,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author: Guosheng Lin",
+        "bbox": [
+            94,
+            886,
+            302,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "18601",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Further, it may raise privacy issues in some circumstances, e.g., medical images and face data.",
+        "bbox": [
+            76,
+            90,
+            468,
+            121
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address this problem, continual semantic segmentation (CSS) has been proposed by [35] as an emerging research direction, where the training scheme is separated into several steps with each step tackling a set of unseen classes. Specifically, given the old model and new training data (only new classes are labeled while others are treated as background), the new model is supposed to recognize both old and new classes. Under this scenario, naively finetuning the old model on new data tends to suffer from catastrophic forgetting [17, 26], where the recognition capability of old classes is quickly lost.",
+        "bbox": [
+            75,
+            123,
+            468,
+            289
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Knowledge distillation (KD), first proposed in image classification [19], has recently been introduced to CSS [35, 4] to mitigate the forgetting issue. As Fig. 1 shows,  $ILT$  [35] removes new class probabilities (green bar) predicted by the new model, and simply distill old classes and background accordingly. However, they ignore the background shift problem [4], where the new class bicycle at current step  $t$  was labeled as background at last step  $t - 1$ . As a result, the old model, which regards bicycle as background, will output a high background score for the new class bicycle pixels. Via KD, it will mislead the new model to misclassify new classes as background. Obviously, this KD strategy hinders the learning of new classes because the class correspondence for distillation is corrupted, i.e., naively mapping the new background to the old background.",
+        "bbox": [
+            75,
+            291,
+            470,
+            518
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "This issue was highlighted in  $MiB$  [4] where they proposed a new class correspondence by combination. Concretely, they combined new classes with the new background via probability summation (Fig. 1) to form a pseudo class, which was treated as the counterpart to the old background for distillation. However, this strategy, though alleviating class mis-correspondence, entangles new classes with the new background and tends to misclassify background as new classes, as detailed in Sec. 3.2. In this paper, we term the error of mistaking background for new (novel) classes or vice-versa as novel-background confusion.",
+        "bbox": [
+            75,
+            520,
+            468,
+            686
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The key insight to overcome the novel-background confusion is to build a reliable class correspondence across different learning steps without corruption or entanglement. To this end, we devise a novel Label-Guided Knowledge Distillation (LGKD) loss, where the class probabilities predicted by the old model are expanded to have the same dimension as the output of the new model (see the blue block in Fig. 1). The background probability is then transplanted to the corresponding ground truth label (class) of the input pixels. In this way, the knowledge from the old background at the last step can be correctly distilled into its corresponding semantic class at the current step, i.e., either the new background or a novel class. Note that our LGKD is a generic regularization term with negligible computa",
+        "bbox": [
+            75,
+            688,
+            468,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "tional cost and can be easily incorporated into existing arts. We validate its effectiveness on two prevailing CSS benchmarks Pascal-VOC and ADE20K, where our LGKD consistently yields promising improvements upon three competing methods, especially on novel (new class) mIoU (up to  $+76\\%$ ), setting new state-of-the-art. To further demonstrate the generalization ability of our approach, we establish a challenging CSS benchmark based on ScanNet for 3D point cloud and re-implement multiple baselines for comparison. To our best knowledge, this is the first work to conduct CSS on 3D point cloud. Extensive experiments showcase that our LGKD loss is capable to handle both 2D and 3D modalities with no ad hoc design.",
+        "bbox": [
+            496,
+            90,
+            890,
+            287
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The main contributions of this paper can be summarized as:",
+        "bbox": [
+            500,
+            290,
+            890,
+            306
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose a new label-guided knowledge distillation (LGKD) loss for CSS, which builds a reliable class correspondence across incremental steps and alleviates novel-background confusion.",
+            "- LGKD is a generic regularization term with negligible computational cost, which can be readily combined with existing methods. Extensive experiments on two prevailing CSS benchmarks, Pascal-VOC and ADE20K, showcase that LGKD significantly improves three competitive methods, particularly on novel mIoU by up to  $+76\\%$ , setting new state-of-the-art.",
+            "- To further demonstrate its generalization capability, we establish the first CSS benchmark for 3D point cloud based on ScanNet and re-implement multiple baselines for comparison. Experiments illustrate that our LGKD is versatile in both 2D and 3D modalities without any ad hoc design."
+        ],
+        "bbox": [
+            514,
+            315,
+            890,
+            592
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Works",
+        "text_level": 1,
+        "bbox": [
+            500,
+            614,
+            648,
+            630
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Semantic Segmentation, aiming to perform pixel-level classification, has witnessed great advancement [63, 6, 7, 53, 22, 16] in recent years since the pioneering work Fully Convolutional Network (FCN) [33]. Among the above methods, the DeepLab series [6, 7] is well-known for their effective designs, e.g., atrous spatial pyramid pooling (ASPP) [6], atrous convolution in cascade [7]. DeepLabv3 [7] is commonly used as the segmentation framework in prior CSS works [4, 13, 61].",
+        "bbox": [
+            496,
+            643,
+            890,
+            779
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In 3D point cloud modality, PointNet [38] is a pioneering work to directly process unstructured point cloud. Subsequent work PointNet++ [39] learns to capture local structures and recognize fine-grained patterns via a hierarchical neural network that applies PointNet recursively. While more complicated techniques have been proposed in recent works [46, 45, 62, 68, 10], we employ PointNet++ as our 3D segmentation model for its simplicity and efficiency.",
+        "bbox": [
+            496,
+            779,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "18602",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/d20b9277ef6bef5cfa08df9dac0ba012aba9e3f4765a82868f524821621afb35.jpg",
+        "image_caption": [
+            "Figure 2. Illustration of our LGKD loss. At learning step  $t$ , an input sample  $x$  is fed into the old (top) and new (bottom) models to obtain the class probabilities for each element  $i \\in \\mathcal{I}$  (pixel or point). These elements are then separated into the novel and background sets according to their ground truth label available at the current step  $t$  (yellow and green as novel set). Finally, to build up a reliable class correspondence, we extend the old probabilities (output by the old model) to have the same dimension as the new ones by expand and transplant (for novel set only), such that useful prior knowledge from the old model can be effectively distilled into the new model to facilitate learning new classes. We highlight that our LGKD can effectively alleviate the novel-background confusion issue as occurs in [35, 4]."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            76,
+            88,
+            893,
+            354
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Continual Learning aims at training a deep neural network to learn multiple tasks in sequence continually. Distinct from one-step incremental learning [50, 51, 55, 56, 59, 58, 30, 60, 29, 57, 31, 44], continual learning generally confronts more severe forgetting issues during multiple-step learning. Numerous works tackle the catastrophic forgetting issue and they can be classified into three categories. The model growing category [2, 48, 52] tackles this problem by dynamically extending model capacity. The memory replay approaches retain the old-class knowledge by keeping a small amount of old-class data in current training. Those data can be raw data [41, 40, 3], features [18, 23] or generated data [24, 42, 32]. The regularization approaches constrain the model by regularizing the parameters [26, 1, 54], the gradients [21, 5], logits [40, 3, 4] or the features [20, 12, 14] of the model.",
+        "bbox": [
+            75,
+            465,
+            472,
+            705
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Continual Semantic Segmentation. Existing CSS methods can be divided into two categories: 1) replay-based. RECALL [34] rehearses old-class data obtained by web-crawling or generative models. Similarly, [49] utilizes an iterative relabeling strategy with rehearsal-based incremental learning. Following [4], replay-based methods are out of the scope of this paper since accessing old data violates the conventional CSS assumption. 2) regularization-based. ILT [35] is the first work to introduce KD into CSS. MiB [4] raises the background shift problem in CSS and proposes an unbiased KD to tackle it. PLOP [13] distills both short- and long-range spatial relations via local pooled output distilla",
+        "bbox": [
+            75,
+            719,
+            472,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "tion. SDR [36] leverages prototype matching, feature sparsification and contrastive learning to enforce feature consistency and discrimination. Most recently, class similarity KD, representation compensation, structure-preserving loss and feature projection, and the biased context are investigated by REMINDER [37], RCIL [61], SPPA [28] and RBC [64] respectively. However, the above approaches tend to suffer from the novel-background confusion problem since their distillation term lacks an appropriate class correspondence. Therefore, we propose a generic distillation term, dubbed LGKD, with a reliable class correspondence to alleviate this confusion issue.",
+        "bbox": [
+            496,
+            464,
+            893,
+            647
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            500,
+            666,
+            591,
+            681
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Problem Definition and Setups",
+        "text_level": 1,
+        "bbox": [
+            500,
+            694,
+            769,
+            710
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Before introducing the continual learning setting, we first present the setups of the standard semantic segmentation task. To ensure the generality, let  $\\mathcal{X} \\in \\mathbb{R}^{N \\times C_{\\mathrm{in}}}$  denote the input space (e.g., image, point cloud, etc.), where  $N$  is the number of input elements and  $C_{\\mathrm{in}}$  is the number of input channels (typically  $C_{\\mathrm{in}} = 3$ , i.e., RGB, for image and  $C_{\\mathrm{in}} >= 3$ , e.g., XYZ, RGB, normals, for point cloud). Let  $\\mathcal{Y} \\in \\mathcal{C}^N$  denote the output label space, where  $\\mathcal{C}$  is a label set that contains all classes including the background class  $\\mathsf{b} \\in \\mathcal{C}$ . Note that each sample  $x \\in \\mathcal{X}$  comprises a set of elements  $\\mathcal{I}$  (e.g., pixels, points, etc.) of corresponding label  $y \\in \\mathcal{Y}$  with constant cardinality  $|\\mathcal{I}| = N$ .",
+        "bbox": [
+            496,
+            719,
+            893,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "18603",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given a training set  $\\mathcal{T} \\subset \\mathcal{X} \\times \\mathcal{Y}$ , the target of semantic segmentation is to learn a model  $f_{\\theta}$  parametrized by  $\\theta$  that maps the input space  $\\mathcal{X}$  to a set of class probability vectors  $f_{\\theta}: \\mathcal{X} \\mapsto \\mathbb{R}^{N \\times |\\mathcal{C}|}$ . The output segmentation mask is then computed as  $\\hat{y} = \\{\\arg \\max_{c \\in \\mathcal{C}} f_{\\theta}(x)[i, c]\\}_{i=1}^{N}$ , where  $f_{\\theta}(x)[i, c]$  is the probability of element  $i$  belonging to category  $c$ .",
+        "bbox": [
+            75,
+            90,
+            468,
+            196
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As opposed to the standard training scheme where the model can access and learn from the full training set  $\\mathcal{T} \\subset \\mathcal{X} \\times \\mathcal{C}^N$  all at once, continual learning, instead, involves several learning steps  $\\{t\\}_{t=0}^{T}$  each with a subset of label  $\\mathcal{C}^t$ . More specifically, at training step  $t$ , the previous label set  $\\mathcal{C}^{0:t-1}$  will be expanded with a set of unseen classes  $\\mathcal{C}^t$  to form a new label set  $\\mathcal{C}^{0:t} = \\mathcal{C}^{0:t-1} \\cup \\mathcal{C}^t$ . Then a training set  $\\mathcal{T}^t \\subset \\mathcal{X} \\times (\\mathcal{C}^t)^N$  will be provided along with the last model  $f_{\\theta^{t-1}}: \\mathcal{X} \\mapsto \\mathbb{R}^{N \\times |\\mathcal{C}^{0:t-1}|}$  to obtain an updated model  $f_{\\theta^t}: \\mathcal{X} \\mapsto \\mathbb{R}^{N \\times |\\mathcal{C}^{0:t}|}$ . Note that the training set  $\\mathcal{T}^t$  may contain element  $i \\in \\mathcal{I}$  from either previous classes  $\\mathcal{C}^{0:t-1}$  or future classes  $\\mathcal{C}^{t+1:T}$ , though the annotations are collapsed into the background class  $b$  at the current step  $t$ . This is referred to as overlapped [4] setting. Following the standard incremental setup, we assume that the label sets of each step are disjoint from each other except for the special background class, i.e.,  $\\mathcal{C}^i \\cap \\mathcal{C}^j = b$  ( $i \\neq j$ ).",
+        "bbox": [
+            75,
+            198,
+            470,
+            457
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Revisiting Knowledge Distillation Loss",
+        "text_level": 1,
+        "bbox": [
+            76,
+            465,
+            408,
+            482
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "A naive way to tackle the continual learning problem is to initialize a new model  $f_{\\theta^t}$  with the weights of the last model  $f_{\\theta^{t-1}}$  and then optimize the network with the training set  $\\mathcal{T}^t$ . However, this will lead to the catastrophic forgetting of old classes  $\\mathcal{C}^{0:t-1}$ . Since no old class samples with annotation are available at the current step.",
+        "bbox": [
+            75,
+            489,
+            468,
+            580
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Standard knowledge distillation. To alleviate the forgetting problem, knowledge distillation [19] is a common solution to preserve old knowledge by regularizing the output class probability distribution of the new model  $f_{\\theta^t}$  to be close to that of the old model  $f_{\\theta^{t-1}}$  given the same input sample. Formally, the standard knowledge distillation loss  $\\ell_{\\mathrm{kd}}$  adopted in CSS by ILT [35] can be formulated as:",
+        "bbox": [
+            75,
+            584,
+            468,
+            690
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\ell_ {\\mathrm {k d}} ^ {\\theta^ {t}} (x, y) = - \\frac {1}{| \\mathcal {I} |} \\sum_ {i \\in \\mathcal {I}} \\sum_ {c \\in \\mathcal {C} ^ {0: t - 1}} q _ {x} ^ {t - 1} (i, c) \\log \\hat {q} _ {x} ^ {t} (i, c), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            89,
+            700,
+            468,
+            736
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\hat{q}_x^t (i,c)$  refers to the probability of class  $c$  for element  $i$  predicted by  $f_{\\theta^t}$  but re-normalized over all old classes:",
+        "bbox": [
+            76,
+            748,
+            468,
+            780
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {q} _ {x} ^ {t} (i, c) = q _ {x} ^ {t} (i, c) / \\sum_ {k \\in \\mathcal {C} ^ {0: t - 1}} q _ {x} ^ {t} (i, k). \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            137,
+            791,
+            468,
+            814
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "However, there exists a discrepancy issue in the standard distillation loss. Concretely, it fails to leverage prior knowledge to cope with the new classes, as the new class probabilities  $q_{x}^{t}(i,c)$  ( $c \\in \\mathcal{C}^t \\setminus b$ ) are not incorporated in distillation (Eq. 1). Moreover, it wrongly enhances the background",
+        "bbox": [
+            75,
+            825,
+            468,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "probability for novel class elements due to the background shift [4]. Hence, it tends to mistake the novel classes for the background (refer to supp. material for detailed discussion).",
+        "bbox": [
+            496,
+            90,
+            890,
+            137
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Unbiased knowledge distillation. To overcome the above issue, MiB [4] proposes an unbiased knowledge distillation loss for CSS by revising  $\\hat{q}_x^t (i,c)$  in Eq. 2 as:",
+        "bbox": [
+            496,
+            138,
+            890,
+            186
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {q} _ {x} ^ {t} (i, c) = \\left\\{ \\begin{array}{l l} q _ {x} ^ {t} (i, c) & \\text {i f} c \\neq \\mathfrak {b} \\\\ \\sum_ {k \\in \\mathcal {C} ^ {t}} q _ {x} ^ {t} (i, k) & \\text {i f} c = \\mathfrak {b}, \\end{array} \\right. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            558,
+            194,
+            890,
+            234
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where they compare the old background probability  $q_{x}^{t - 1}(i,\\mathsf{b})$  with the probability of being either a novel class or the new background, i.e.,  $\\sum_{k\\in \\mathcal{C}^t}q_x^t (i,k)$ , rather than directly with its counterpart  $q_{x}^{t}(i,\\mathsf{b})$ . Such class probability combination mechanism is meant to facilitate distilling the knowledge from the old background class defined at step  $t - 1$  to the novel classes at step  $t$ . Nevertheless, it entangles the novel classes with the new background due to the probability combination. As a consequence of entanglement, the probabilities of novel classes will be undesirably enhanced when the distillation loss only intends to improve the probabilities of the new background, and vice versa. As opposed to ILT [35], this leads to the other confusion — misclassifying background as novel classes.",
+        "bbox": [
+            496,
+            242,
+            892,
+            455
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Label-Guided Knowledge Distillation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            462,
+            821,
+            479
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Class separation. Unlike standard KD and unbiased KD, which apply the same KD strategy to all input elements, we propose to separate the input elements into two sets according to the ground truth label available at the current step, namely the background set  $S_{\\mathsf{b}} = \\{i|i\\in \\mathcal{I},y_i = \\mathsf{b}\\}$  and the novel set  $S_{n} = \\{i|i\\in \\mathcal{I},y_{i}\\in \\mathcal{C}^{t}\\setminus \\mathsf{b}\\}$ . In this way, we can customize the distillation strategy for each class set and establish a more accurate class correspondence, which facilitates more effective knowledge distillation.",
+        "bbox": [
+            496,
+            488,
+            890,
+            625
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Class correspondence. As opposed to revising the class probabilities of the new model  $q_{x}^{t}(i,c)$  as in [4], we instead resort to correcting the class probabilities of the old model  $q_{x}^{t - 1}(i,c)$  according to the ground truth label  $y_{i}$ . Our design is inspired by the wireless communication system. In order to improve the useful information captured by the receiver, one should ensure that the signal released by the transmitter is correct and that both transmitter and receiver share the same communication protocol. Likewise, in our case, the objective is to assure the class probabilities (signal) predicted by the old model  $f_{\\theta^{t - 1}}$  (transmitter) contain correct class semantics (after background shift) and are well-aligned with (protocol) the output class space of the new model  $f_{\\theta^t}$  (receiver). Protocol: For the elements from the novel set  $S_{n}$ , we expand the output of the old model  $q_{x}^{t - 1}(i,c)$  with zero probability assigned to the extra novel classes  $c \\in \\mathcal{C}^t \\setminus \\mathbf{b}$  such that it has the same output class space dimensions with the new model. Signal: Due to the",
+        "bbox": [
+            496,
+            628,
+            892,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "18604",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "background shift, we transplant the probability of the background to the corresponding ground truth novel class and set the background probability to zero while maintaining the others unchanged. The final corrected probability distribution to be distilled from can be formulated as:",
+        "bbox": [
+            75,
+            90,
+            468,
+            165
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\bar {q} _ {x} ^ {t - 1} (i, c) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} c \\in \\mathcal {C} ^ {t} \\text {a n d} c \\neq y _ {i} \\\\ q _ {x} ^ {t - 1} (i, \\mathrm {b}) & \\text {i f} c = y _ {i} \\\\ q _ {x} ^ {t - 1} (i, c) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            101,
+            174,
+            468,
+            231
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $i \\in S_n$ . We highlight that, this novel set distillation strategy can preserve the knowledge from  $f_{\\theta^{t-1}}$  for the old classes while also correcting the class probabilities from the old model for the new classes given the ground truth label  $y_i$  as guidance. As for the case of the background set  $i \\in S_b$ , it can be treated in a similar yet simpler way by merely expanding zero probability for the extra novel classes as:",
+        "bbox": [
+            75,
+            241,
+            468,
+            347
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\bar {q} _ {x} ^ {t - 1} (i, c) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} c \\in \\mathcal {C} ^ {t} \\backslash \\mathfrak {b} \\\\ q _ {x} ^ {t - 1} (i, c) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            137,
+            358,
+            468,
+            398
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In summary, our distillation loss can be given as:",
+        "bbox": [
+            76,
+            409,
+            400,
+            422
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\ell_ {\\mathrm {k d}} ^ {\\theta^ {t}} (x, y) = \\lambda_ {n} \\cdot \\bar {\\ell} _ {\\mathrm {k d}} ^ {\\theta^ {t}} (x, y, \\mathcal {S} _ {n}) + \\lambda_ {\\mathfrak {b}} \\cdot \\bar {\\ell} _ {\\mathrm {k d}} ^ {\\theta^ {t}} (x, y, \\mathcal {S} _ {\\mathfrak {b}}), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            96,
+            433,
+            468,
+            454
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda_{n}$  and  $\\lambda_{\\mathrm{b}}$  balances the contribution of the novel set  $S_{n}$  and background set  $S_{\\mathrm{b}}$ ;  $\\bar{\\ell}_{\\mathrm{kd}}^{\\theta^{t}}(x,y,\\mathcal{S})$  is the distillation loss defined within set  $S$  as:",
+        "bbox": [
+            76,
+            464,
+            468,
+            508
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\bar {\\ell} _ {\\mathrm {k d}} ^ {\\theta^ {t}} (x, y, \\mathcal {S}) = - \\frac {1}{N} \\sum_ {i \\in \\mathcal {S}} \\sum_ {c \\in \\mathcal {C} ^ {0: t}} \\bar {q} _ {x} ^ {t - 1} (i, c) \\log q _ {x} ^ {t} (i, c). \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            89,
+            517,
+            468,
+            554
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The overall training objective is then computed as:",
+        "bbox": [
+            76,
+            565,
+            411,
+            579
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} \\left(\\theta^ {t}\\right) = \\frac {1}{\\left| \\mathcal {T} ^ {t} \\right|} \\sum_ {(x, y) \\in \\mathcal {T} ^ {t}} \\left(\\lambda \\ell_ {c e} ^ {\\theta^ {t}} (x, y) + \\ell_ {\\mathrm {k d}} ^ {\\theta^ {t}} (x, y)\\right), \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            98,
+            590,
+            468,
+            628
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\ell_{ce}^{\\theta^t}(x,y)$  is the cross-entropy loss.  $\\lambda$  is the hyperparameter to balance the importance of the two terms, and is set as  $\\lambda = 1$  in experiments.",
+        "bbox": [
+            75,
+            641,
+            468,
+            688
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiment",
+        "text_level": 1,
+        "bbox": [
+            76,
+            700,
+            200,
+            718
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Benchmarking 3D Continual Segmentation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            726,
+            442,
+            742
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Existing CSS methods have mainly been evaluated on 2D CSS benchmarks, leaving a significant gap in the exploration of 3D CSS. In order to validate the versatility of LGKD across different modalities, we introduce the first 3D CSS benchmark based on ScanNet [11] and present several baseline methods for comparison.",
+        "bbox": [
+            75,
+            750,
+            468,
+            839
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "ScanNet contains 1201 and 312 indoor scenes in the training and validation set respectively with 20 classes in total, including a special class for \"other furniture\", i.e. background in the standard segmentation task. In Fig. 3, we",
+        "bbox": [
+            75,
+            839,
+            468,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/695c663d0d51fece92c00017d9f94956c57e4c16cfee9a6c6f6ae5e5d8454d2c.jpg",
+        "image_caption": [
+            "Figure 3. The scene-wise class frequency statistics of our proposed ScanNet benchmark for 3D continual semantic segmentation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            94,
+            880,
+            306
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "provide the scene-wise class frequency statistics of ScanNet [11], including both train and validation sets. The scene-wise class frequency indicates the number of scenes that a certain class appears in. Surprisingly, we find that ScanNet exhibits a challenging long-tail category distribution. Additionally, we provide the point-wise class frequency statistics and demonstrate the difficulty of our proposed benchmark in the supp. material. For a challenging CSS benchmark, we split the entire class set into multiple subsets according to the descending order of scene-wise class frequency as in Fig. 3, with the tail classes constituting the novel sets for continual learning. This means that the novel classes to be learned in each step are set to be those rare ones, a.k.a. tail classes in the long-tail field. Following the Pascal-VOC 2012 CSS benchmark [35, 4], we define three different incremental scenarios: adding one class (18-1), adding five classes simultaneously (14-5) and sequentially (14-1). Take 14-5 setting for instance, the last five rare classes, i.e., counter, refrigerator, bookshelf, bathtub, shower curtain, form the novel class set, and likewise for the other settings.",
+        "bbox": [
+            496,
+            369,
+            890,
+            686
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We form the training set by including all the scenes that contain at least one point from the novel classes at the current step, with the others (either previous or future classes) annotated as \"other furniture\". The validation set is established in a similar way except that the label of the previous classes  $\\mathcal{C}^{0:t-1}$  are maintained, as the model is required to predict all seen classes up to the current step. Finally, we report the old-, new- and all-class mIoU metrics as well as the background IoU at the end of all incremental steps.",
+        "bbox": [
+            496,
+            686,
+            890,
+            823
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Baselines and Setups",
+        "text_level": 1,
+        "bbox": [
+            500,
+            832,
+            694,
+            848
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To validate the effectiveness of our approach, we conduct thorough comparison with the latest state-of-the-art CSS methods as well as baselines. FT is a straightforward",
+        "bbox": [
+            498,
+            854,
+            890,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "18605",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/8e20b00d4cbaca6114da3fa18b73c1377263746772e3341beaaab61db876920c.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 1. Continual Semantic Segmentation performance (mIoU) on Pascal-VOC 2012 under different incremental scenarios. † indicates the results are excerpted from [4]. * suggests the results are reproduced with the official codes. 0 stands for the background class and all represents the mIoU over all classes including the background. Best results are highlighted in Red while runner-up in Blue."
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"4\">19-1 (2 steps)</td><td colspan=\"4\">15-5 (2 steps)</td><td colspan=\"4\">15-1 (6 steps)</td></tr><tr><td>0</td><td>1-19</td><td>20</td><td>all</td><td>0</td><td>1-15</td><td>16-20</td><td>all</td><td>0</td><td>1-15</td><td>16-20</td><td>all</td></tr><tr><td>FT†</td><td>-</td><td>6.8</td><td>12.9</td><td>-</td><td>-</td><td>2.1</td><td>33.1</td><td>-</td><td>-</td><td>0.2</td><td>1.8</td><td>-</td></tr><tr><td>EWC† [26]</td><td>-</td><td>26.9</td><td>14.0</td><td>-</td><td>-</td><td>24.3</td><td>35.5</td><td>-</td><td>-</td><td>0.3</td><td>4.3</td><td>-</td></tr><tr><td>LwF-MC† [40]</td><td>-</td><td>64.4</td><td>13.3</td><td>-</td><td>-</td><td>58.1</td><td>35.0</td><td>-</td><td>-</td><td>6.4</td><td>8.4</td><td>-</td></tr><tr><td>ILT† [35]</td><td>-</td><td>67.1</td><td>12.3</td><td>-</td><td>-</td><td>66.3</td><td>40.6</td><td>-</td><td>-</td><td>4.9</td><td>7.8</td><td>-</td></tr><tr><td>ILT* [35]</td><td>88.6</td><td>66.2</td><td>8.3</td><td>64.5</td><td>89.2</td><td>65.2</td><td>38.1</td><td>59.9</td><td>77.7</td><td>3.7</td><td>7.9</td><td>8.2</td></tr><tr><td>SDR* [36]</td><td>90.0</td><td>68.9</td><td>24.2</td><td>67.8</td><td>90.1</td><td>76.6</td><td>50.9</td><td>71.1</td><td>83.9</td><td>34.1</td><td>13.0</td><td>31.5</td></tr><tr><td>REIMINDER* [37]</td><td>91.9</td><td>75.6</td><td>33.9</td><td>74.4</td><td>90.2</td><td>75.1</td><td>49.1</td><td>69.6</td><td>83.3</td><td>65.2</td><td>27.0</td><td>57.0</td></tr><tr><td>SPPA [28]</td><td>-</td><td>-</td><td>36.2</td><td>74.6</td><td>-</td><td>-</td><td>52.9</td><td>72.1</td><td>-</td><td>-</td><td>23.3</td><td>56.0</td></tr><tr><td>MiB† [4]</td><td>-</td><td>70.2</td><td>22.1</td><td>-</td><td>-</td><td>75.5</td><td>49.4</td><td>-</td><td>-</td><td>35.1</td><td>13.5</td><td>-</td></tr><tr><td>MiB* [4]</td><td>90.3</td><td>71.5</td><td>23.2</td><td>70.1</td><td>89.6</td><td>74.2</td><td>46.3</td><td>68.3</td><td>83.2</td><td>38.0</td><td>13.9</td><td>34.5</td></tr><tr><td>LGKD+MiB (Ours)</td><td>91.4</td><td>72.1</td><td>40.4</td><td>71.5</td><td>91.6</td><td>75.2</td><td>54.8</td><td>71.1</td><td>79.7</td><td>39.3</td><td>17.0</td><td>35.9</td></tr><tr><td>PLOP* [13]</td><td>92.1</td><td>75.3</td><td>36.1</td><td>74.2</td><td>89.7</td><td>74.2</td><td>47.5</td><td>68.6</td><td>84.7</td><td>64.6</td><td>21.0</td><td>55.2</td></tr><tr><td>LGKD+PLOP (Ours)</td><td>92.9</td><td>76.5</td><td>42.9</td><td>75.7</td><td>91.4</td><td>78.7</td><td>56.1</td><td>73.9</td><td>89.3</td><td>69.3</td><td>30.9</td><td>61.1</td></tr><tr><td>RCIL* [61]</td><td>90.1</td><td>73.6</td><td>24.5</td><td>72.1</td><td>88.9</td><td>75.9</td><td>48.4</td><td>70.0</td><td>83.9</td><td>67.9</td><td>23.1</td><td>58.0</td></tr><tr><td>LGKD+RCIL (Ours)</td><td>92.8</td><td>76.5</td><td>37.5</td><td>75.5</td><td>91.4</td><td>77.6</td><td>54.3</td><td>72.7</td><td>89.4</td><td>69.0</td><td>29.1</td><td>60.5</td></tr><tr><td>Joint</td><td>93.7</td><td>77.6</td><td>78.1</td><td>78.4</td><td>93.7</td><td>79.0</td><td>73.6</td><td>78.4</td><td>93.7</td><td>79.0</td><td>73.6</td><td>78.4</td></tr></table>",
+        "bbox": [
+            101,
+            88,
+            870,
+            328
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "lower-bound baseline, which simply finetunes the model with the newly available data at each step. On the contrary, Joint, training a standard segmentation model on all classes within a single step, may serve as an upper bound for CSS approaches, though not always true if the CSS benchmark shows a long-tail nature with tail classes as new classes (e.g., ScanNet). For completeness, we include two general continual learning methods [26, 40] that are not specifically designed for semantic segmentation. Further, we mainly benchmark against the recent state-of-the-art methods taylored for segmentation [35, 4, 13, 36, 37, 61]. Following [4], we do not consider replay-based methods, e.g., [40]) in our comparison as the access to previous annotated data violates the standard class-incremental learning assumption. For the sake of generalization, we conduct experiments on both 2D image modality, i.e., Pascal-VOC 2012 [15] and ADE20K [66], and 3D point cloud modality, i.e. ScanNet [11]. Besides the old-, new- and all-class (including the background class) mIoU, we additionally report the background IoU because it is a special category under CSS with its semantic content varying across different learning steps.",
+        "bbox": [
+            75,
+            396,
+            472,
+            715
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "CSS protocols. As for the label masking policy for the incremental steps, [4] follows and adapts two different CSS protocols, namely disjoint [35] and overlapped [43]. Both protocols assume that only the novel classes at the current step are labeled while the others are masked out as background. However, in the disjoint protocol, input samples for step  $t$  should only contain elements (i.e., pixels or points) that belong to either current or previous classes, i.e.,  $\\mathcal{C}^{0:t}$ . On the contrary, the overlapped protocol allows the input samples to contain either previous, current or future classes, i.e.,  $\\mathcal{C}^{0:T}$ . The disjoint protocol makes a strong assumption that the current training data should not contain any samples",
+        "bbox": [
+            75,
+            719,
+            472,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "of the class we would like to learn in the future, which is impracticable. Therefore, we adopt the more realistic overlapped protocol throughout our experiments.",
+        "bbox": [
+            496,
+            396,
+            890,
+            443
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            498,
+            450,
+            718,
+            465
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In 2D image modality, we follow the framework DeepLabv3 [7] and training hyperparameters of the baselines when building our LGKD upon them, except that we adopt synchronized batch normalization with different batch size and learning rate settings. In 3D point cloud modality, we use PointNet++ [39] with multi-scale grouping as our base model. We use Adam [25] with an initial learning rate  $10^{-2}$  for the first step and  $5 \\times 10^{-3}$  for the subsequent steps and set the weight decay to 0. The learning rate is decayed by 0.7 every 100 epochs. We train our network for 500 epochs with a batch size of 32. We remove possible duplicated points before calculating the IoU metric. For all benchmarks, we report the mIoU results on the standard validation set. We run our experiments on two NVIDIA RTX 3090 GPUs for Pascal VOC 2012 and ADE20K while one for ScanNet. More details can be found in the supp. material.",
+        "bbox": [
+            496,
+            473,
+            890,
+            731
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4. Performance",
+        "text_level": 1,
+        "bbox": [
+            500,
+            738,
+            635,
+            753
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Pascal-VOC 2012 consists of 10,582 images for training and 1449 images for validation with 20 foreground categories. Following [4], we conduct three different continual learning scenarios, i.e., adding one class (19-1), five classes simultaneously (15-5) and sequentially (15-1). The split of the class set for each class-incremental step follows the alphabetical order. Table 1 shows comprehensive results of our method along with other baselines. Clearly, FT performs poorly across all the incremental settings as the old",
+        "bbox": [
+            496,
+            763,
+            893,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "18606",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/49553776dc6888e2ad3b0fc771d9019bd0f0290b9c3e9e8191daa66ebe224107.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"4\">100-50 (2 steps)</td><td colspan=\"4\">100-10 (6 steps)</td><td colspan=\"4\">50-50 (3 steps)</td></tr><tr><td>0</td><td>1-100</td><td>101-150</td><td>all</td><td>0</td><td>1-100</td><td>101-150</td><td>all</td><td>0</td><td>1-50</td><td>51-150</td><td>all</td></tr><tr><td>ILT† [35]</td><td>10.9</td><td>18.4</td><td>14.4</td><td>17.0</td><td>8.5</td><td>0.0</td><td>3.1</td><td>1.1</td><td>8.6</td><td>3.4</td><td>12.9</td><td>9.7</td></tr><tr><td>MiB† [4]</td><td>19.7</td><td>40.7</td><td>17.2</td><td>32.8</td><td>0.0</td><td>38.6</td><td>11.1</td><td>29.2</td><td>0.0</td><td>46.5</td><td>21.0</td><td>29.3</td></tr><tr><td>REMINDER* [37]</td><td>26.6</td><td>42.1</td><td>16.1</td><td>33.4</td><td>24.0</td><td>38.9</td><td>20.3</td><td>32.6</td><td>24.2</td><td>48.0</td><td>20.1</td><td>29.4</td></tr><tr><td>SPPA [28]</td><td>-</td><td>42.9</td><td>19.9</td><td>-</td><td>-</td><td>41.0</td><td>12.5</td><td>-</td><td>-</td><td>49.8</td><td>23.9</td><td>-</td></tr><tr><td>PLOP* [13]</td><td>26.7</td><td>42.1</td><td>14.6</td><td>32.9</td><td>23.9</td><td>39.4</td><td>14.5</td><td>31.1</td><td>24.5</td><td>48.2</td><td>20.7</td><td>29.8</td></tr><tr><td>LGKD+PLOP (Ours)</td><td>27.4</td><td>43.6</td><td>25.7</td><td>37.5</td><td>26.0</td><td>42.1</td><td>22.0</td><td>35.4</td><td>25.3</td><td>49.4</td><td>29.4</td><td>36.0</td></tr><tr><td>RCIL* [61]</td><td>26.3</td><td>41.7</td><td>17.3</td><td>33.5</td><td>0.0</td><td>37.2</td><td>14.9</td><td>29.5</td><td>21.2</td><td>47.8</td><td>23.0</td><td>31.2</td></tr><tr><td>LGKD+RCIL (Ours)</td><td>26.7</td><td>43.3</td><td>25.1</td><td>37.2</td><td>24.0</td><td>42.2</td><td>20.4</td><td>34.9</td><td>25.7</td><td>49.1</td><td>27.2</td><td>34.4</td></tr><tr><td>Joint</td><td>29.1</td><td>42.6</td><td>28.2</td><td>37.7</td><td>29.1</td><td>42.6</td><td>28.2</td><td>37.7</td><td>29.1</td><td>49.2</td><td>32.1</td><td>37.7</td></tr></table>",
+        "bbox": [
+            81,
+            88,
+            888,
+            250
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/dc850b93113b3d64538e2e5ad3dfb14299c416837475f7797d36c96aaffa76cc.jpg",
+        "table_caption": [
+            "Table 2. The mIoU(%) of the last step on the ADE20K dataset for different overlapped continual learning scenarios.  $\\dagger$  indicates the results are excerpted from [13]. * suggests the results are reproduced with the official codes. Best results are highlighted in Red while runner-up in Blue."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"4\">18-1 (2 steps)</td><td colspan=\"4\">14-5 (2 steps)</td><td colspan=\"4\">14-1 (6 steps)</td></tr><tr><td>0</td><td>1-18</td><td>19</td><td>all</td><td>0</td><td>1-14</td><td>15-19</td><td>all</td><td>0</td><td>1-14</td><td>15-19</td><td>all</td></tr><tr><td>FT</td><td>19.2</td><td>0.0</td><td>28.6</td><td>2.4</td><td>18.6</td><td>0.0</td><td>44.3</td><td>12.0</td><td>19.2</td><td>0.0</td><td>7.8</td><td>2.9</td></tr><tr><td>ILT [35]</td><td>23.8</td><td>14.9</td><td>25.8</td><td>15.9</td><td>26.7</td><td>22.2</td><td>43.4</td><td>27.7</td><td>19.3</td><td>0.0</td><td>18.5</td><td>5.6</td></tr><tr><td>MiB [4]</td><td>55.1</td><td>55.0</td><td>32.2</td><td>53.8</td><td>52.9</td><td>57.8</td><td>45.5</td><td>54.5</td><td>45.5</td><td>56.5</td><td>29.7</td><td>49.2</td></tr><tr><td>LGKD+MiB (Ours)</td><td>55.2</td><td>55.5</td><td>40.8</td><td>54.8</td><td>54.3</td><td>57.6</td><td>49.0</td><td>55.3</td><td>51.9</td><td>55.7</td><td>38.4</td><td>51.2</td></tr><tr><td>Joint</td><td>53.2</td><td>55.0</td><td>40.9</td><td>54.2</td><td>53.2</td><td>56.7</td><td>47.5</td><td>54.2</td><td>53.2</td><td>56.7</td><td>47.5</td><td>54.2</td></tr></table>",
+        "bbox": [
+            81,
+            301,
+            888,
+            405
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Continual Semantic Segmentation performance (mIoU) on ScanNet under different incremental scenarios.  $\\theta$  stands for the background class and all represents the mIoU over all classes including the background. All results are obtained with our implementation.",
+        "bbox": [
+            76,
+            406,
+            890,
+            434
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "class pixels are masked as background under CSS setup, and the model is explicitly taught to wrongly predict the old classes as background. EWC [26] yields similar results for novel classes while fairly outperforming the FT baseline in terms of base mIoU. Significant improvement on base classes across all settings can be observed when knowledge distillation is introduced in [40, 35]. MiB [4] obtains consistent improvement on both base and novel performance across different settings, due to addressing the background shift problem in CSS. Recent approaches [14, 61] obtain drastically higher performance against MiB when multiple incremental steps are involved (15-1). Despite their promising performance, our LGKD is able to further boost their performance by a large margin, especially on novel performance. For instance, LGKD+MiB achieves  $+74\\%$  (40.4 vs. 23.2) improvement against MiB on novel mIoU under 19-1 setting. These promising novel class results along with the improved background performance prove that our LGKD loss can effectively alleviate the novel-background confusion problem.",
+        "bbox": [
+            75,
+            458,
+            472,
+            760
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "ADE20K [67] is a challenging dataset consisting of 20,210 training and 2000 testing images of complex scenes. Similar to VOC, we conducted three different CSS scenarios, i.e., 100-50, 100-10 and 50-50, as shown in Table 2. Specifically, we build our LGKD upon two competitive methods, i.e., PLOP [13] and RCIL [61], where LGKD achieves considerable improvement, especially on novel mIoU (up to  $+76\\%$  from 14.6 to 25.7 with PLOP under 100-50 setting), setting new state-of-the-art across all incremental scenarios.",
+        "bbox": [
+            75,
+            763,
+            468,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "ScanNet. As shown in Table 3, we compare our method with FT baseline, ILT [35] and MiB [4] on our proposed ScanNet benchmark to illustrate the generality of LGKD in 3D modality. Similar to their performance in 2D modality, FT and ILT suffer from catastrophic forgetting when learning new classes. However, MiB significantly outperforms the former methods across all three settings and metrics, in particular, the base class performance, which once again demonstrates the effectiveness of tackling the background shift in CSS. Compared with MiB, our LGKD further improves the novel performance by a large margin:  $+27\\%$  (40.8 vs. 32.2) and  $+29\\%$  (38.4 vs. 29.7) relative improvement of novel mIoU against MiB are obtained under 18-1 and 14-1 settings respectively, which proves the considerable importance of overcoming the background shift with correct class correspondence. It is also noteworthy that our approach also outperforms joint training, which shows that LGKD is also effective in alleviating long-tail issues.",
+        "bbox": [
+            496,
+            458,
+            890,
+            731
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.5. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            500,
+            738,
+            653,
+            753
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Unless otherwise stated, all ablative studies are conducted upon VOC 15-5 setting with PLOP [13] as the baseline. Refer to the supp. material for the rationale.",
+        "bbox": [
+            496,
+            760,
+            890,
+            806
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Expand & Transplant. We compare three different designs for novel set distillation in Table 5. Firstly, One-hot is a hard-label distillation with all the class probabilities transplanted to the ground truth novel class, making the probability of the ground truth class equal to one. In this case, the distillation loss is equivalent to the standard one-hot cross",
+        "bbox": [
+            496,
+            809,
+            890,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "18607",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/8bf54184ea3c6ef69c3468d5ef88540764c7263aa98497c36dc727e5d8a6dc40.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td colspan=\"4\">VOC 15-5 (2 steps)</td><td colspan=\"4\">VOC 15-1 (6 steps)</td></tr><tr><td></td><td>0</td><td>1-15</td><td>16-20</td><td>all</td><td>0</td><td>1-15</td><td>16-20</td><td>all</td></tr><tr><td>PLOP [13]</td><td>89.7</td><td>74.2</td><td>47.5</td><td>68.6</td><td>84.7</td><td>64.6</td><td>21.0</td><td>55.2</td></tr><tr><td>+Standard KD [35]</td><td>90.7</td><td>74.4</td><td>49.2</td><td>69.2</td><td>88.4</td><td>66.4</td><td>13.4</td><td>54.8</td></tr><tr><td>+Unbiased KD [4]</td><td>90.2</td><td>77.1</td><td>51.8</td><td>71.7</td><td>86.3</td><td>67.1</td><td>22.6</td><td>57.4</td></tr><tr><td>+LGKD (Ours)</td><td>91.4</td><td>78.7</td><td>56.1</td><td>73.9</td><td>89.3</td><td>69.3</td><td>30.9</td><td>61.1</td></tr><tr><td></td><td colspan=\"4\">ADE20k 100-50 (2 steps)</td><td colspan=\"4\">ADE20k 100-10 (6 steps)</td></tr><tr><td></td><td>0</td><td>1-100</td><td>101-150</td><td>all</td><td>0</td><td>1-100</td><td>101-150</td><td>all</td></tr><tr><td>PLOP [13]</td><td>26.7</td><td>42.1</td><td>14.6</td><td>32.9</td><td>23.9</td><td>39.4</td><td>14.5</td><td>31.1</td></tr><tr><td>+Standard KD [35]</td><td>27.4</td><td>42.5</td><td>13.6</td><td>32.8</td><td>23.2</td><td>39.8</td><td>13.5</td><td>31.0</td></tr><tr><td>+Unbiased KD [4]</td><td>27.0</td><td>42.6</td><td>14.5</td><td>33.2</td><td>23.7</td><td>39.8</td><td>15.1</td><td>31.5</td></tr><tr><td>+LGKD (Ours)</td><td>27.4</td><td>43.6</td><td>25.7</td><td>37.5</td><td>26.0</td><td>42.1</td><td>22.0</td><td>35.4</td></tr></table>",
+        "bbox": [
+            163,
+            88,
+            810,
+            263
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/0fe51e8aff0ad7785689832e9c561f338211753f961dee1951db50c4c136b66c.jpg",
+        "table_caption": [
+            "Table 4. Effect of different knowledge distillation terms. Standard KD is adopted in ILT [35] while unbiased KD is adopted in MiB [4]."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>0</td><td>1-15</td><td>16-20</td><td>all</td></tr><tr><td>One-hot</td><td>91.4</td><td>76.7</td><td>51.9</td><td>71.5</td></tr><tr><td>Expand only</td><td>90.4</td><td>78.5</td><td>30.9</td><td>67.7</td></tr><tr><td>Expand &amp; Transplant</td><td>91.4</td><td>78.7</td><td>56.1</td><td>73.9</td></tr></table>",
+        "bbox": [
+            91,
+            300,
+            459,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/91f187122e7207ce24e5f6b8b7e3bbfa34290db23a8d2cbe8fef44321564ae96.jpg",
+        "table_caption": [
+            "Table 5. Effect of our novel set distillation design including expand and transplant operations."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>λb</td><td>λn</td><td>0</td><td>1-15</td><td>16-20</td><td>all</td></tr><tr><td>0</td><td>0</td><td>90.5</td><td>75.5</td><td>53.6</td><td>71.0</td></tr><tr><td>1</td><td>0</td><td>91.3</td><td>76.9</td><td>53.1</td><td>71.9</td></tr><tr><td>2</td><td>0</td><td>91.6</td><td>77.3</td><td>53.1</td><td>72.2</td></tr><tr><td>5</td><td>0</td><td>91.8</td><td>77.9</td><td>52.0</td><td>72.4</td></tr><tr><td>10</td><td>0</td><td>91.8</td><td>78.3</td><td>49.8</td><td>72.2</td></tr><tr><td>5</td><td>0.5</td><td>91.6</td><td>78.6</td><td>55.7</td><td>73.8</td></tr><tr><td>5</td><td>1</td><td>91.4</td><td>78.7</td><td>56.1</td><td>73.9</td></tr><tr><td>5</td><td>2</td><td>91.0</td><td>78.6</td><td>55.5</td><td>73.7</td></tr></table>",
+        "bbox": [
+            120,
+            410,
+            429,
+            547
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6. Effect of the weighted factors  ${\\lambda }_{n}$  and  ${\\lambda }_{\\mathrm{b}}$  .",
+        "bbox": [
+            124,
+            549,
+            419,
+            560
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "entropy loss, discarding all the old class probabilities predicted by the old model. As expected, the base performance is slightly compromised compared with Expand & Transplant since the old class knowledge is not distilled. Nevertheless, the novel performance also drops since using hard labels for the novel class pixels will develop a bias toward novel classes, resulting in false positive predictions. This illustrates the significance of retaining the old class probabilities during transplant. Secondly, Expand only will encourage the new model to misclassify a novel class into either background or an old class. Thus, a sharp drop in novel mIoU from 56.1 to 30.9 is observed. This uncovers the importance of correcting the old model output by transplant to handle the background shift. Finally, transplanting the background probability into the corresponding ground truth novel class while retaining the old class probabilities can establish an appropriate class correspondence and facilitate learning new classes, hence yielding the best novel performance.",
+        "bbox": [
+            75,
+            590,
+            468,
+            878
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Weighted factor. As table 6 shows, increasing  $\\lambda_{\\mathrm{b}}$  leads to",
+        "bbox": [
+            76,
+            885,
+            467,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "higher base performance, but it will compromise the novel performance when  $\\lambda_{\\mathrm{b}}$  becomes too large, e.g.,  $\\lambda_{\\mathrm{b}} = 10$ . It indicates that the background set distillation is mainly responsible for retaining old knowledge. We set  $\\lambda_{\\mathrm{b}} = 5$  as the optimal solution as it obtains promising base mIoU while not sacrificing the novel mIoU severely. With the novel set distillation kicking in, the novel mIoU is significantly improved from 52.0 to 56.1 with  $\\lambda_{n} = 1$ . Meanwhile, it also boosts the base mIoU from 77.9 to 78.7. This suggests that the novel set distillation facilitates learning new classes while maintaining the old class knowledge distilled from the old model.",
+        "bbox": [
+            496,
+            301,
+            890,
+            482
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of different KDs. To validate the superiority of LGKD, we compare our LGKD with other knowledge distillation (KD) terms on both VOC and ADE20k upon the same baseline PLOP [13] (Fig. 4), since PLOP does not adopt class probability-wise KD. Firstly, PLOP can hardly benefit from adding the standard KD [35], probably due to its tendency to misclassify new classes as the background (see Sec. 3.2). Secondly, though unbiased KD [4] yields a moderate boost on VOC, it makes almost no difference on the more challenging benchmark ADE20k. In contrast, LGKD consistently surpasses standard KD and unbiased KD by a significant margin, especially in terms of novel mIoU. Concretely, LGKD yields  $+8.3 / +11.2 / +6.9$  points improvement against unbiased KD on 15-1/100-50/100-10 setting.",
+        "bbox": [
+            496,
+            487,
+            890,
+            714
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.6. Error Analysis",
+        "text_level": 1,
+        "bbox": [
+            500,
+            722,
+            650,
+            738
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We carry out comprehensive error analysis on both Pascal-VOC 2012 and ScanNet datasets to validate the effectiveness of our LGKD in mitigating the novel-background confusion problem.",
+        "bbox": [
+            496,
+            746,
+            890,
+            806
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "2D image. As Fig. 4 shows, under the VOC 15-5 incremental scenario, naive finetuning completely forgets the old class bird while also mistaking the background for the novel class potted plant. Thanks to KD mechanism, ILT [35], REMINDER [37], MiB [4], PLOP [13] and RCIL [61] are capable to retain the old knowledge (bird and chair). Nev",
+        "bbox": [
+            496,
+            809,
+            890,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "18608",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5b46f10fce6c80ead65967ac3a82345c812a9e7b6fe052777680b2c8a4d37ab7.jpg",
+        "image_caption": [
+            "Old classes: bird chair",
+            "New classes: potted plant Sofa",
+            "Figure 4. Error analysis of baselines along with our LGKD under VOC 15-5 setting. MiB [4], PLOP [13] and RCIL [61] tend to mistake weed and suitcase (background—things of no interest) for potted plant and sofa (novel classes) respectively. Such confusion is effectively alleviated when equipped with our LGKD, except for the bird image with RCIL, where the confusion is slightly mitigated."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            84,
+            88,
+            767,
+            375
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/a0cad4b3110eedc47cfa79b7f07cf95012dbbc554e649b9e231d0e4f5a22cb39.jpg",
+        "image_caption": [
+            "Figure 5. Error analysis of MiB [4] vs. LGKD+MiB on step 4 (with bathtub depicted in pink as the novel class) of the ScanNet 14-1 incremental scenario. Our LGKD effectively alleviates the confusion made by MiB (see the black boxes) — mistaking the background (dark purple) for the novel class bathtub. Meanwhile, LGKD+MiB achieves decent accuracy on the novel class bathtub (see the red boxes)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            441,
+            468,
+            565
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "ertheless, these methods are consistently prone to misclassifying the background (weed and suitcase—things of no interest) as novel classes (potted plant and sofa). On the contrary, when incorporating our LGKD into MiB, PLOP and RCIL, the above novel-background confusion is effectively suppressed, except for the bird image with RCIL, where the confusion is slightly alleviated. The above examples demonstrate the effectiveness of LGKD in tackling the novel-background confusion issue. More visualization results can be found in the supp. material.",
+        "bbox": [
+            75,
+            696,
+            468,
+            849
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "3D point cloud. We show the error results in Fig. 5 for step 4 of the ScanNet 14-1 setting, where the current novel class is bathtub depicted in pink. Clearly, our method effic",
+        "bbox": [
+            75,
+            854,
+            470,
+            902
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "tively corrects the false prediction made by MiB [4] (see the black boxes), where the background points (highlighted in dark purple) are mistaken for the novel class bathtub. Meanwhile, our LGKD does not sacrifice the capability to recognize the true bathtub points (see red boxes).",
+        "bbox": [
+            496,
+            445,
+            893,
+            521
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            532,
+            619,
+            547
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we identify a key issue in continual semantic segmentation (CSS)—novel background confusion. This is generally caused by the corrupted class correspondence during knowledge distillation. To alleviate such confusion, we propose a novel solution termed Label-Guided Knowledge Distillation (LGKD), which can establish an appropriate class correspondence between the output of the old and the new model for distillation, guided by the ground truth label. We validate its effectiveness on both existing 2D benchmarks and our proposed 3D CSS benchmark, where LGKD can consistently outperform competing methods by a large margin, especially on novel mIoU, setting new state-of-the-art. Finally, we hope this work will inspire further in-depth exploration of 3D CSS in the future.",
+        "bbox": [
+            496,
+            558,
+            893,
+            770
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6. Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            500,
+            781,
+            679,
+            797
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is supported by the Hyundai research grant (04OIS000252C130).",
+        "bbox": [
+            500,
+            806,
+            893,
+            835
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            500,
+            847,
+            596,
+            863
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "ref_text",
+        "text": "[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware",
+        "bbox": [
+            506,
+            872,
+            892,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "18609",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "synapses: Learning what (not) to forget. In ECCV, 2018. 3",
+            "[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3366-3375, 2017. 3",
+            "[3] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In ECCV, 2018. 3",
+            "[4] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 2, 3, 4, 5, 6, 7, 8, 9",
+            "[5] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 3",
+            "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE T-PAMI, 40(4):834-848, 2017. 2",
+            "[7] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 2, 6",
+            "[8] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 1",
+            "[9] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 1",
+            "[10] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021. 2",
+            "[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5, 6",
+            "[12] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 3",
+            "[13] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4040-4050, 2021. 2, 3, 6, 7, 8, 9",
+            "[14] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distilla"
+        ],
+        "bbox": [
+            76,
+            90,
+            470,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "tion for small-tasks incremental learning. In European Conference on Computer Vision, pages 86-102. Springer, 2020. 3,7",
+            "[15] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html, 2012.6",
+            "[16] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3146-3154, 2019. 1, 2",
+            "[17] Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211, 2013. 2",
+            "[18] Tyler L Hayes, Kushal Kafle, Robik Shrestha, Manoj Acharya, and Christopher Kanan. Remind your neural network to prevent catastrophic forgetting. In European Conference on Computer Vision, pages 466-483. Springer, 2020. 3",
+            "[19] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 2, 4",
+            "[20] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 3",
+            "[21] Guannan Hu, Wu Zhang, Hu Ding, and Wenhao Zhu. Gradient episodic memory with a soft constraint for continual learning. arXiv preprint arXiv:2011.07801, 2020. 3",
+            "[22] Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu Liu. Ccnet: Criss-cross attention for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 603-612, 2019. 1, 2",
+            "[23] Ahmet Iscen, Jeffrey Zhang, Svetlana Lazebnik, and Cordelia Schmid. Memory-efficient incremental learning through feature adaptation. In European conference on computer vision, pages 699-715. Springer, 2020. 3",
+            "[24] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. arXiv preprint arXiv:1711.10563, 2017. 3",
+            "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6",
+            "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2, 3, 6, 7",
+            "[27] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In CVPR, 2017. 1",
+            "[28] Zihan Lin, Zilei Wang, and Yixin Zhang. Continual semantic segmentation via structure preserving and projected feature"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "18610",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "alignment. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 345-361. Springer, 2022. 3, 6, 7",
+            "[29] Weide Liu, Chi Zhang, Henghui Ding, Tzu-Yi Hung, and Guosheng Lin. Few-shot segmentation with optimal transport matching and message flow. arXiv preprint arXiv:2108.08518, 2021. 3",
+            "[30] Weide Liu, Chi Zhang, Guosheng Lin, and Fayao Liu. Crnet: Cross-reference networks for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4165-4173, 2020. 3",
+            "[31] Weide Liu, Chi Zhang, Guosheng Lin, and Fayao Liu. Crcnet: Few-shot segmentation with cross-reference and region-global conditional networks. International Journal of Computer Vision, 130(12):3140-3157, 2022. 3",
+            "[32] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12245-12254, 2020. 3",
+            "[33] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 1, 2",
+            "[34] Andrea Maracani, Umberto Michieli, Marco Toldo, and Pietro Zanuttigh. Recall: Replay-based continual learning in semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7026-7035, 2021. 3",
+            "[35] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In ICCV-WS, pages 0–0, 2019. 2, 3, 4, 5, 6, 7, 8",
+            "[36] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1114-1124, 2021. 3, 6",
+            "[37] Minh Hieu Phan, Son Lam Phung, Long Tran-Thanh, Abdesselam Bouzerdoum, et al. Class similarity weighted knowledge distillation for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16866-16875, 2022. 3, 6, 7, 8",
+            "[38] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2",
+            "[39] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 6",
+            "[40] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 3, 6, 7",
+            "[41] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 3"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[42] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. In NeurIPS, 2017. 3",
+            "[43] Konstantin Shmelkov, Cordelia Schmid, and Karteek Alahari. Incremental learning of object detectors without catastrophic forgetting. In ICCV, 2017. 6",
+            "[44] Nan Song, Chi Zhang, and Guosheng Lin. Few-shot open-set recognition using background as unknowns. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5970-5979, 2022. 3",
+            "[45] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European conference on computer vision, pages 685–702. Springer, 2020. 2",
+            "[46] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2",
+            "[47] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5463-5474, 2021. 1",
+            "[48] Ju Xu and Zhanxing Zhu. Reinforced continual learning. Advances in Neural Information Processing Systems, 31, 2018. 3",
+            "[49] Shipeng Yan, Jiale Zhou, Jiangwei Xie, Songyang Zhang, and Xuming He. An em framework for online incremental learning of semantic segmentation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3052-3060, 2021. 3",
+            "[50] Ze Yang, Yali Wang, Xianyu Chen, Jianzhuang Liu, and Yu Qiao. Context-transformer: Tackling object confusion for few-shot detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12653-12660, 2020. 3",
+            "[51] Ze Yang, Chi Zhang, Ruibo Li, Yi Xu, and Guosheng Lin. Efficient few-shot object detection via knowledge inheritance. IEEE Transactions on Image Processing, 32:321-334, 2022. 3",
+            "[52] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 3",
+            "[53] Yuhui Yuan, Lang Huang, Jianyuan Guo, Chao Zhang, Xilin Chen, and Jingdong Wang. Ocnet: Object context network for scene parsing. arXiv preprint arXiv:1809.00916, 2018. 2",
+            "[54] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, 2017. 3",
+            "[55] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12203-12213, 2020. 3",
+            "[56] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Differentiable earth mover's distance for few-shot"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "18611",
+        "bbox": [
+            480,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5632-5648, 2022. 3",
+            "[57] Chi Zhang, Henghui Ding, Guosheng Lin, Ruibo Li, Changhu Wang, and Chunhua Shen. Meta navigator: Search for a good adaptation policy for few-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9435-9444, 2021. 3",
+            "[58] Chi Zhang, Guosheng Lin, Fayao Liu, Jiushuang Guo, Qingyao Wu, and Rui Yao. Pyramid graph networks with connection attentions for region-based one-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9587-9595, 2019. 3",
+            "[59] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5217–5226, 2019. 3",
+            "[60] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12455-12464, 2021. 3",
+            "[61] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022. 2, 3, 6, 7, 8, 9",
+            "[62] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020. 2",
+            "[63] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 1, 2",
+            "[64] Hanbin Zhao, Fengyu Yang, Xinghe Fu, and Xi Li. Rbc: Rectifying the biased context in continual semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 55-72. Springer, 2022. 3",
+            "[65] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 1",
+            "[66] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 6",
+            "[67] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 7",
+            "[68] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical"
+        ],
+        "bbox": [
+            78,
+            90,
+            467,
+            898
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "and asymmetrical 3d convolution networks for lidar segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9939-9948, 2021. 2",
+        "bbox": [
+            535,
+            92,
+            890,
+            146
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "18612",
+        "bbox": [
+            480,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 11
+    }
+]

2023/Label-Guided Knowledge Distillation for Continual Semantic Segmentation on 2D Images and 3D Point Clouds/3a801de3-50c8-469a-b26b-c00f633cc663_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70afa47f926dc9f880b001f592a5034800c70f74da93d1a1c31b3612a4f992f2
+size 1275133

2023/Label-Guided Knowledge Distillation for Continual Semantic Segmentation on 2D Images and 3D Point Clouds/full.md

@@ -0,0 +1,318 @@
+# Label-Guided Knowledge Distillation for Continual Semantic Segmentation on 2D Images and 3D Point Clouds
+
+Ze Yang $^{1}$ , Ruibo Li $^{1}$ , Evan Ling $^{2}$ , Chi Zhang $^{1}$ , Yiming Wang $^{1}$ , Dezhao Huang $^{2}$ , Keng Teck Ma $^{2}$ , Minhoe Hur $^{3}$ , Guosheng Lin $^{1*}$
+
+$^{1}$ School of Computer Science and Engineering, Nanyang Technological University, Singapore  
+ $^{2}$ Hyundai Motor Group Innovation Center in Singapore (HMGICS)  
+ $^{3}$ AIRS Company, Hyundai Motor Group
+
+ze001@e.ntu.edu.sg, evan.ling@hmgics.com, minhoe.hur@hyundai.com, gslin@ntu.edu.sg
+
+# Abstract
+
+Continual semantic segmentation (CSS) aims to extend an existing model to tackle unseen tasks while retaining its old knowledge. Naively fine-tuning the old model on new data leads to catastrophic forgetting. A common solution is knowledge distillation (KD), where the output distribution of the new model is regularized to be similar to that of the old model. However, in CSS, this is challenging because of the background shift issue. Existing KD-based CSS methods continue to suffer from confusion between the background and novel classes since they fail to establish a reliable class correspondence for distillation. To address this issue, we propose a new label-guided knowledge distillation (LGKD) loss, where the old model output is expanded and transplanted (with the guidance of the ground truth label) to form a semantically appropriate class correspondence with the new model output. Consequently, the useful knowledge from the old model can be effectively distilled into the new model without causing confusion. We conduct extensive experiments on two prevailing CSS benchmarks, Pascal-VOC and ADE20K, where our LGKD significantly boosts the performance of three competing methods, especially on novel mIoU by up to  $+76\%$ , setting new state-of-the-art. Finally, to further demonstrate its generalization ability, we introduce the first CSS benchmark for 3D point cloud based on ScanNet, along with several re-implemented baselines for comparison. Experiments show that LGKD is versatile in both 2D and 3D modalities without requiring ad hoc design. Codes are available at https://github.com/Ze-Yang/LGKD.
+
+# 1. Introduction
+
+Fully supervised semantic segmentation has witnessed tremendous success [33, 63, 27, 8, 16, 22, 47, 65, 9] in re
+
+![](images/5a842318a5f59736e908d670e8a3502e730702509059163ea18e851911ba5ad8.jpg)  
+Figure 1. Illustration of knowledge distillation (KD) strategies in continual semantic segmentation. Standard KD and unbiased KD reduce the new class (bicycle) dimension(s) of the probabilities output by the new model via remove and combine respectively, which collapses the class correspondence across incremental steps and raises confusion between background and new classes (see Sec. 1). In contrast, our label-guided KD uses the ground truth label as guidance to expand the probabilities predicted by the old model. It builds a reliable class correspondence across different learning steps without discrepancy (ILT) or entanglement (MiB) (see Sec. 3.2). A  $\xrightarrow{\text{mimic}}$  B: encourage A to be similar to B.
+
+cent years. These algorithms generally assume a fixed number of classes to be learned. However, in real-world applications, it is often expected that a deployed model can be continuously generalized to handle new classes while not forgetting the old ones. A simple solution is to expand the original dataset with newly available samples and retrain a new model from scratch, dubbed as Joint Training. Obviously, this is computationally expensive and requires an increasing amount of space to store the old data over time.
+
+Further, it may raise privacy issues in some circumstances, e.g., medical images and face data.
+
+To address this problem, continual semantic segmentation (CSS) has been proposed by [35] as an emerging research direction, where the training scheme is separated into several steps with each step tackling a set of unseen classes. Specifically, given the old model and new training data (only new classes are labeled while others are treated as background), the new model is supposed to recognize both old and new classes. Under this scenario, naively finetuning the old model on new data tends to suffer from catastrophic forgetting [17, 26], where the recognition capability of old classes is quickly lost.
+
+Knowledge distillation (KD), first proposed in image classification [19], has recently been introduced to CSS [35, 4] to mitigate the forgetting issue. As Fig. 1 shows,  $ILT$  [35] removes new class probabilities (green bar) predicted by the new model, and simply distill old classes and background accordingly. However, they ignore the background shift problem [4], where the new class bicycle at current step  $t$  was labeled as background at last step  $t - 1$ . As a result, the old model, which regards bicycle as background, will output a high background score for the new class bicycle pixels. Via KD, it will mislead the new model to misclassify new classes as background. Obviously, this KD strategy hinders the learning of new classes because the class correspondence for distillation is corrupted, i.e., naively mapping the new background to the old background.
+
+This issue was highlighted in  $MiB$  [4] where they proposed a new class correspondence by combination. Concretely, they combined new classes with the new background via probability summation (Fig. 1) to form a pseudo class, which was treated as the counterpart to the old background for distillation. However, this strategy, though alleviating class mis-correspondence, entangles new classes with the new background and tends to misclassify background as new classes, as detailed in Sec. 3.2. In this paper, we term the error of mistaking background for new (novel) classes or vice-versa as novel-background confusion.
+
+The key insight to overcome the novel-background confusion is to build a reliable class correspondence across different learning steps without corruption or entanglement. To this end, we devise a novel Label-Guided Knowledge Distillation (LGKD) loss, where the class probabilities predicted by the old model are expanded to have the same dimension as the output of the new model (see the blue block in Fig. 1). The background probability is then transplanted to the corresponding ground truth label (class) of the input pixels. In this way, the knowledge from the old background at the last step can be correctly distilled into its corresponding semantic class at the current step, i.e., either the new background or a novel class. Note that our LGKD is a generic regularization term with negligible computa
+
+tional cost and can be easily incorporated into existing arts. We validate its effectiveness on two prevailing CSS benchmarks Pascal-VOC and ADE20K, where our LGKD consistently yields promising improvements upon three competing methods, especially on novel (new class) mIoU (up to  $+76\%$ ), setting new state-of-the-art. To further demonstrate the generalization ability of our approach, we establish a challenging CSS benchmark based on ScanNet for 3D point cloud and re-implement multiple baselines for comparison. To our best knowledge, this is the first work to conduct CSS on 3D point cloud. Extensive experiments showcase that our LGKD loss is capable to handle both 2D and 3D modalities with no ad hoc design.
+
+The main contributions of this paper can be summarized as:
+
+- We propose a new label-guided knowledge distillation (LGKD) loss for CSS, which builds a reliable class correspondence across incremental steps and alleviates novel-background confusion.  
+- LGKD is a generic regularization term with negligible computational cost, which can be readily combined with existing methods. Extensive experiments on two prevailing CSS benchmarks, Pascal-VOC and ADE20K, showcase that LGKD significantly improves three competitive methods, particularly on novel mIoU by up to  $+76\%$ , setting new state-of-the-art.  
+- To further demonstrate its generalization capability, we establish the first CSS benchmark for 3D point cloud based on ScanNet and re-implement multiple baselines for comparison. Experiments illustrate that our LGKD is versatile in both 2D and 3D modalities without any ad hoc design.
+
+# 2. Related Works
+
+Semantic Segmentation, aiming to perform pixel-level classification, has witnessed great advancement [63, 6, 7, 53, 22, 16] in recent years since the pioneering work Fully Convolutional Network (FCN) [33]. Among the above methods, the DeepLab series [6, 7] is well-known for their effective designs, e.g., atrous spatial pyramid pooling (ASPP) [6], atrous convolution in cascade [7]. DeepLabv3 [7] is commonly used as the segmentation framework in prior CSS works [4, 13, 61].
+
+In 3D point cloud modality, PointNet [38] is a pioneering work to directly process unstructured point cloud. Subsequent work PointNet++ [39] learns to capture local structures and recognize fine-grained patterns via a hierarchical neural network that applies PointNet recursively. While more complicated techniques have been proposed in recent works [46, 45, 62, 68, 10], we employ PointNet++ as our 3D segmentation model for its simplicity and efficiency.
+
+![](images/d20b9277ef6bef5cfa08df9dac0ba012aba9e3f4765a82868f524821621afb35.jpg)  
+Figure 2. Illustration of our LGKD loss. At learning step  $t$ , an input sample  $x$  is fed into the old (top) and new (bottom) models to obtain the class probabilities for each element  $i \in \mathcal{I}$  (pixel or point). These elements are then separated into the novel and background sets according to their ground truth label available at the current step  $t$  (yellow and green as novel set). Finally, to build up a reliable class correspondence, we extend the old probabilities (output by the old model) to have the same dimension as the new ones by expand and transplant (for novel set only), such that useful prior knowledge from the old model can be effectively distilled into the new model to facilitate learning new classes. We highlight that our LGKD can effectively alleviate the novel-background confusion issue as occurs in [35, 4].
+
+Continual Learning aims at training a deep neural network to learn multiple tasks in sequence continually. Distinct from one-step incremental learning [50, 51, 55, 56, 59, 58, 30, 60, 29, 57, 31, 44], continual learning generally confronts more severe forgetting issues during multiple-step learning. Numerous works tackle the catastrophic forgetting issue and they can be classified into three categories. The model growing category [2, 48, 52] tackles this problem by dynamically extending model capacity. The memory replay approaches retain the old-class knowledge by keeping a small amount of old-class data in current training. Those data can be raw data [41, 40, 3], features [18, 23] or generated data [24, 42, 32]. The regularization approaches constrain the model by regularizing the parameters [26, 1, 54], the gradients [21, 5], logits [40, 3, 4] or the features [20, 12, 14] of the model.
+
+Continual Semantic Segmentation. Existing CSS methods can be divided into two categories: 1) replay-based. RECALL [34] rehearses old-class data obtained by web-crawling or generative models. Similarly, [49] utilizes an iterative relabeling strategy with rehearsal-based incremental learning. Following [4], replay-based methods are out of the scope of this paper since accessing old data violates the conventional CSS assumption. 2) regularization-based. ILT [35] is the first work to introduce KD into CSS. MiB [4] raises the background shift problem in CSS and proposes an unbiased KD to tackle it. PLOP [13] distills both short- and long-range spatial relations via local pooled output distilla
+
+tion. SDR [36] leverages prototype matching, feature sparsification and contrastive learning to enforce feature consistency and discrimination. Most recently, class similarity KD, representation compensation, structure-preserving loss and feature projection, and the biased context are investigated by REMINDER [37], RCIL [61], SPPA [28] and RBC [64] respectively. However, the above approaches tend to suffer from the novel-background confusion problem since their distillation term lacks an appropriate class correspondence. Therefore, we propose a generic distillation term, dubbed LGKD, with a reliable class correspondence to alleviate this confusion issue.
+
+# 3. Method
+
+# 3.1. Problem Definition and Setups
+
+Before introducing the continual learning setting, we first present the setups of the standard semantic segmentation task. To ensure the generality, let  $\mathcal{X} \in \mathbb{R}^{N \times C_{\mathrm{in}}}$  denote the input space (e.g., image, point cloud, etc.), where  $N$  is the number of input elements and  $C_{\mathrm{in}}$  is the number of input channels (typically  $C_{\mathrm{in}} = 3$ , i.e., RGB, for image and  $C_{\mathrm{in}} >= 3$ , e.g., XYZ, RGB, normals, for point cloud). Let  $\mathcal{Y} \in \mathcal{C}^N$  denote the output label space, where  $\mathcal{C}$  is a label set that contains all classes including the background class  $\mathsf{b} \in \mathcal{C}$ . Note that each sample  $x \in \mathcal{X}$  comprises a set of elements  $\mathcal{I}$  (e.g., pixels, points, etc.) of corresponding label  $y \in \mathcal{Y}$  with constant cardinality  $|\mathcal{I}| = N$ .
+
+Given a training set  $\mathcal{T} \subset \mathcal{X} \times \mathcal{Y}$ , the target of semantic segmentation is to learn a model  $f_{\theta}$  parametrized by  $\theta$  that maps the input space  $\mathcal{X}$  to a set of class probability vectors  $f_{\theta}: \mathcal{X} \mapsto \mathbb{R}^{N \times |\mathcal{C}|}$ . The output segmentation mask is then computed as  $\hat{y} = \{\arg \max_{c \in \mathcal{C}} f_{\theta}(x)[i, c]\}_{i=1}^{N}$ , where  $f_{\theta}(x)[i, c]$  is the probability of element  $i$  belonging to category  $c$ .
+
+As opposed to the standard training scheme where the model can access and learn from the full training set  $\mathcal{T} \subset \mathcal{X} \times \mathcal{C}^N$  all at once, continual learning, instead, involves several learning steps  $\{t\}_{t=0}^{T}$  each with a subset of label  $\mathcal{C}^t$ . More specifically, at training step  $t$ , the previous label set  $\mathcal{C}^{0:t-1}$  will be expanded with a set of unseen classes  $\mathcal{C}^t$  to form a new label set  $\mathcal{C}^{0:t} = \mathcal{C}^{0:t-1} \cup \mathcal{C}^t$ . Then a training set  $\mathcal{T}^t \subset \mathcal{X} \times (\mathcal{C}^t)^N$  will be provided along with the last model  $f_{\theta^{t-1}}: \mathcal{X} \mapsto \mathbb{R}^{N \times |\mathcal{C}^{0:t-1}|}$  to obtain an updated model  $f_{\theta^t}: \mathcal{X} \mapsto \mathbb{R}^{N \times |\mathcal{C}^{0:t}|}$ . Note that the training set  $\mathcal{T}^t$  may contain element  $i \in \mathcal{I}$  from either previous classes  $\mathcal{C}^{0:t-1}$  or future classes  $\mathcal{C}^{t+1:T}$ , though the annotations are collapsed into the background class  $b$  at the current step  $t$ . This is referred to as overlapped [4] setting. Following the standard incremental setup, we assume that the label sets of each step are disjoint from each other except for the special background class, i.e.,  $\mathcal{C}^i \cap \mathcal{C}^j = b$  ( $i \neq j$ ).
+
+# 3.2. Revisiting Knowledge Distillation Loss
+
+A naive way to tackle the continual learning problem is to initialize a new model  $f_{\theta^t}$  with the weights of the last model  $f_{\theta^{t-1}}$  and then optimize the network with the training set  $\mathcal{T}^t$ . However, this will lead to the catastrophic forgetting of old classes  $\mathcal{C}^{0:t-1}$ . Since no old class samples with annotation are available at the current step.
+
+Standard knowledge distillation. To alleviate the forgetting problem, knowledge distillation [19] is a common solution to preserve old knowledge by regularizing the output class probability distribution of the new model  $f_{\theta^t}$  to be close to that of the old model  $f_{\theta^{t-1}}$  given the same input sample. Formally, the standard knowledge distillation loss  $\ell_{\mathrm{kd}}$  adopted in CSS by ILT [35] can be formulated as:
+
+$$
+\ell_ {\mathrm {k d}} ^ {\theta^ {t}} (x, y) = - \frac {1}{| \mathcal {I} |} \sum_ {i \in \mathcal {I}} \sum_ {c \in \mathcal {C} ^ {0: t - 1}} q _ {x} ^ {t - 1} (i, c) \log \hat {q} _ {x} ^ {t} (i, c), \tag {1}
+$$
+
+where  $\hat{q}_x^t (i,c)$  refers to the probability of class  $c$  for element  $i$  predicted by  $f_{\theta^t}$  but re-normalized over all old classes:
+
+$$
+\hat {q} _ {x} ^ {t} (i, c) = q _ {x} ^ {t} (i, c) / \sum_ {k \in \mathcal {C} ^ {0: t - 1}} q _ {x} ^ {t} (i, k). \tag {2}
+$$
+
+However, there exists a discrepancy issue in the standard distillation loss. Concretely, it fails to leverage prior knowledge to cope with the new classes, as the new class probabilities  $q_{x}^{t}(i,c)$  ( $c \in \mathcal{C}^t \setminus b$ ) are not incorporated in distillation (Eq. 1). Moreover, it wrongly enhances the background
+
+probability for novel class elements due to the background shift [4]. Hence, it tends to mistake the novel classes for the background (refer to supp. material for detailed discussion).
+
+Unbiased knowledge distillation. To overcome the above issue, MiB [4] proposes an unbiased knowledge distillation loss for CSS by revising  $\hat{q}_x^t (i,c)$  in Eq. 2 as:
+
+$$
+\hat {q} _ {x} ^ {t} (i, c) = \left\{ \begin{array}{l l} q _ {x} ^ {t} (i, c) & \text {i f} c \neq \mathfrak {b} \\ \sum_ {k \in \mathcal {C} ^ {t}} q _ {x} ^ {t} (i, k) & \text {i f} c = \mathfrak {b}, \end{array} \right. \tag {3}
+$$
+
+where they compare the old background probability  $q_{x}^{t - 1}(i,\mathsf{b})$  with the probability of being either a novel class or the new background, i.e.,  $\sum_{k\in \mathcal{C}^t}q_x^t (i,k)$ , rather than directly with its counterpart  $q_{x}^{t}(i,\mathsf{b})$ . Such class probability combination mechanism is meant to facilitate distilling the knowledge from the old background class defined at step  $t - 1$  to the novel classes at step  $t$ . Nevertheless, it entangles the novel classes with the new background due to the probability combination. As a consequence of entanglement, the probabilities of novel classes will be undesirably enhanced when the distillation loss only intends to improve the probabilities of the new background, and vice versa. As opposed to ILT [35], this leads to the other confusion — misclassifying background as novel classes.
+
+# 3.3. Label-Guided Knowledge Distillation
+
+Class separation. Unlike standard KD and unbiased KD, which apply the same KD strategy to all input elements, we propose to separate the input elements into two sets according to the ground truth label available at the current step, namely the background set  $S_{\mathsf{b}} = \{i|i\in \mathcal{I},y_i = \mathsf{b}\}$  and the novel set  $S_{n} = \{i|i\in \mathcal{I},y_{i}\in \mathcal{C}^{t}\setminus \mathsf{b}\}$ . In this way, we can customize the distillation strategy for each class set and establish a more accurate class correspondence, which facilitates more effective knowledge distillation.
+
+Class correspondence. As opposed to revising the class probabilities of the new model  $q_{x}^{t}(i,c)$  as in [4], we instead resort to correcting the class probabilities of the old model  $q_{x}^{t - 1}(i,c)$  according to the ground truth label  $y_{i}$ . Our design is inspired by the wireless communication system. In order to improve the useful information captured by the receiver, one should ensure that the signal released by the transmitter is correct and that both transmitter and receiver share the same communication protocol. Likewise, in our case, the objective is to assure the class probabilities (signal) predicted by the old model  $f_{\theta^{t - 1}}$  (transmitter) contain correct class semantics (after background shift) and are well-aligned with (protocol) the output class space of the new model  $f_{\theta^t}$  (receiver). Protocol: For the elements from the novel set  $S_{n}$ , we expand the output of the old model  $q_{x}^{t - 1}(i,c)$  with zero probability assigned to the extra novel classes  $c \in \mathcal{C}^t \setminus \mathbf{b}$  such that it has the same output class space dimensions with the new model. Signal: Due to the
+
+background shift, we transplant the probability of the background to the corresponding ground truth novel class and set the background probability to zero while maintaining the others unchanged. The final corrected probability distribution to be distilled from can be formulated as:
+
+$$
+\bar {q} _ {x} ^ {t - 1} (i, c) = \left\{ \begin{array}{l l} 0 & \text {i f} c \in \mathcal {C} ^ {t} \text {a n d} c \neq y _ {i} \\ q _ {x} ^ {t - 1} (i, \mathrm {b}) & \text {i f} c = y _ {i} \\ q _ {x} ^ {t - 1} (i, c) & \text {o t h e r w i s e ,} \end{array} \right. \tag {4}
+$$
+
+where  $i \in S_n$ . We highlight that, this novel set distillation strategy can preserve the knowledge from  $f_{\theta^{t-1}}$  for the old classes while also correcting the class probabilities from the old model for the new classes given the ground truth label  $y_i$  as guidance. As for the case of the background set  $i \in S_b$ , it can be treated in a similar yet simpler way by merely expanding zero probability for the extra novel classes as:
+
+$$
+\bar {q} _ {x} ^ {t - 1} (i, c) = \left\{ \begin{array}{l l} 0 & \text {i f} c \in \mathcal {C} ^ {t} \backslash \mathfrak {b} \\ q _ {x} ^ {t - 1} (i, c) & \text {o t h e r w i s e ,} \end{array} \right. \tag {5}
+$$
+
+In summary, our distillation loss can be given as:
+
+$$
+\ell_ {\mathrm {k d}} ^ {\theta^ {t}} (x, y) = \lambda_ {n} \cdot \bar {\ell} _ {\mathrm {k d}} ^ {\theta^ {t}} (x, y, \mathcal {S} _ {n}) + \lambda_ {\mathfrak {b}} \cdot \bar {\ell} _ {\mathrm {k d}} ^ {\theta^ {t}} (x, y, \mathcal {S} _ {\mathfrak {b}}), \tag {6}
+$$
+
+where  $\lambda_{n}$  and  $\lambda_{\mathrm{b}}$  balances the contribution of the novel set  $S_{n}$  and background set  $S_{\mathrm{b}}$ ;  $\bar{\ell}_{\mathrm{kd}}^{\theta^{t}}(x,y,\mathcal{S})$  is the distillation loss defined within set  $S$  as:
+
+$$
+\bar {\ell} _ {\mathrm {k d}} ^ {\theta^ {t}} (x, y, \mathcal {S}) = - \frac {1}{N} \sum_ {i \in \mathcal {S}} \sum_ {c \in \mathcal {C} ^ {0: t}} \bar {q} _ {x} ^ {t - 1} (i, c) \log q _ {x} ^ {t} (i, c). \tag {7}
+$$
+
+The overall training objective is then computed as:
+
+$$
+\mathcal {L} \left(\theta^ {t}\right) = \frac {1}{\left| \mathcal {T} ^ {t} \right|} \sum_ {(x, y) \in \mathcal {T} ^ {t}} \left(\lambda \ell_ {c e} ^ {\theta^ {t}} (x, y) + \ell_ {\mathrm {k d}} ^ {\theta^ {t}} (x, y)\right), \tag {8}
+$$
+
+where  $\ell_{ce}^{\theta^t}(x,y)$  is the cross-entropy loss.  $\lambda$  is the hyperparameter to balance the importance of the two terms, and is set as  $\lambda = 1$  in experiments.
+
+# 4. Experiment
+
+# 4.1. Benchmarking 3D Continual Segmentation
+
+Existing CSS methods have mainly been evaluated on 2D CSS benchmarks, leaving a significant gap in the exploration of 3D CSS. In order to validate the versatility of LGKD across different modalities, we introduce the first 3D CSS benchmark based on ScanNet [11] and present several baseline methods for comparison.
+
+ScanNet contains 1201 and 312 indoor scenes in the training and validation set respectively with 20 classes in total, including a special class for "other furniture", i.e. background in the standard segmentation task. In Fig. 3, we
+
+![](images/695c663d0d51fece92c00017d9f94956c57e4c16cfee9a6c6f6ae5e5d8454d2c.jpg)  
+Figure 3. The scene-wise class frequency statistics of our proposed ScanNet benchmark for 3D continual semantic segmentation.
+
+provide the scene-wise class frequency statistics of ScanNet [11], including both train and validation sets. The scene-wise class frequency indicates the number of scenes that a certain class appears in. Surprisingly, we find that ScanNet exhibits a challenging long-tail category distribution. Additionally, we provide the point-wise class frequency statistics and demonstrate the difficulty of our proposed benchmark in the supp. material. For a challenging CSS benchmark, we split the entire class set into multiple subsets according to the descending order of scene-wise class frequency as in Fig. 3, with the tail classes constituting the novel sets for continual learning. This means that the novel classes to be learned in each step are set to be those rare ones, a.k.a. tail classes in the long-tail field. Following the Pascal-VOC 2012 CSS benchmark [35, 4], we define three different incremental scenarios: adding one class (18-1), adding five classes simultaneously (14-5) and sequentially (14-1). Take 14-5 setting for instance, the last five rare classes, i.e., counter, refrigerator, bookshelf, bathtub, shower curtain, form the novel class set, and likewise for the other settings.
+
+We form the training set by including all the scenes that contain at least one point from the novel classes at the current step, with the others (either previous or future classes) annotated as "other furniture". The validation set is established in a similar way except that the label of the previous classes  $\mathcal{C}^{0:t-1}$  are maintained, as the model is required to predict all seen classes up to the current step. Finally, we report the old-, new- and all-class mIoU metrics as well as the background IoU at the end of all incremental steps.
+
+# 4.2. Baselines and Setups
+
+To validate the effectiveness of our approach, we conduct thorough comparison with the latest state-of-the-art CSS methods as well as baselines. FT is a straightforward
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">19-1 (2 steps)</td><td colspan="4">15-5 (2 steps)</td><td colspan="4">15-1 (6 steps)</td></tr><tr><td>0</td><td>1-19</td><td>20</td><td>all</td><td>0</td><td>1-15</td><td>16-20</td><td>all</td><td>0</td><td>1-15</td><td>16-20</td><td>all</td></tr><tr><td>FT†</td><td>-</td><td>6.8</td><td>12.9</td><td>-</td><td>-</td><td>2.1</td><td>33.1</td><td>-</td><td>-</td><td>0.2</td><td>1.8</td><td>-</td></tr><tr><td>EWC† [26]</td><td>-</td><td>26.9</td><td>14.0</td><td>-</td><td>-</td><td>24.3</td><td>35.5</td><td>-</td><td>-</td><td>0.3</td><td>4.3</td><td>-</td></tr><tr><td>LwF-MC† [40]</td><td>-</td><td>64.4</td><td>13.3</td><td>-</td><td>-</td><td>58.1</td><td>35.0</td><td>-</td><td>-</td><td>6.4</td><td>8.4</td><td>-</td></tr><tr><td>ILT† [35]</td><td>-</td><td>67.1</td><td>12.3</td><td>-</td><td>-</td><td>66.3</td><td>40.6</td><td>-</td><td>-</td><td>4.9</td><td>7.8</td><td>-</td></tr><tr><td>ILT* [35]</td><td>88.6</td><td>66.2</td><td>8.3</td><td>64.5</td><td>89.2</td><td>65.2</td><td>38.1</td><td>59.9</td><td>77.7</td><td>3.7</td><td>7.9</td><td>8.2</td></tr><tr><td>SDR* [36]</td><td>90.0</td><td>68.9</td><td>24.2</td><td>67.8</td><td>90.1</td><td>76.6</td><td>50.9</td><td>71.1</td><td>83.9</td><td>34.1</td><td>13.0</td><td>31.5</td></tr><tr><td>REIMINDER* [37]</td><td>91.9</td><td>75.6</td><td>33.9</td><td>74.4</td><td>90.2</td><td>75.1</td><td>49.1</td><td>69.6</td><td>83.3</td><td>65.2</td><td>27.0</td><td>57.0</td></tr><tr><td>SPPA [28]</td><td>-</td><td>-</td><td>36.2</td><td>74.6</td><td>-</td><td>-</td><td>52.9</td><td>72.1</td><td>-</td><td>-</td><td>23.3</td><td>56.0</td></tr><tr><td>MiB† [4]</td><td>-</td><td>70.2</td><td>22.1</td><td>-</td><td>-</td><td>75.5</td><td>49.4</td><td>-</td><td>-</td><td>35.1</td><td>13.5</td><td>-</td></tr><tr><td>MiB* [4]</td><td>90.3</td><td>71.5</td><td>23.2</td><td>70.1</td><td>89.6</td><td>74.2</td><td>46.3</td><td>68.3</td><td>83.2</td><td>38.0</td><td>13.9</td><td>34.5</td></tr><tr><td>LGKD+MiB (Ours)</td><td>91.4</td><td>72.1</td><td>40.4</td><td>71.5</td><td>91.6</td><td>75.2</td><td>54.8</td><td>71.1</td><td>79.7</td><td>39.3</td><td>17.0</td><td>35.9</td></tr><tr><td>PLOP* [13]</td><td>92.1</td><td>75.3</td><td>36.1</td><td>74.2</td><td>89.7</td><td>74.2</td><td>47.5</td><td>68.6</td><td>84.7</td><td>64.6</td><td>21.0</td><td>55.2</td></tr><tr><td>LGKD+PLOP (Ours)</td><td>92.9</td><td>76.5</td><td>42.9</td><td>75.7</td><td>91.4</td><td>78.7</td><td>56.1</td><td>73.9</td><td>89.3</td><td>69.3</td><td>30.9</td><td>61.1</td></tr><tr><td>RCIL* [61]</td><td>90.1</td><td>73.6</td><td>24.5</td><td>72.1</td><td>88.9</td><td>75.9</td><td>48.4</td><td>70.0</td><td>83.9</td><td>67.9</td><td>23.1</td><td>58.0</td></tr><tr><td>LGKD+RCIL (Ours)</td><td>92.8</td><td>76.5</td><td>37.5</td><td>75.5</td><td>91.4</td><td>77.6</td><td>54.3</td><td>72.7</td><td>89.4</td><td>69.0</td><td>29.1</td><td>60.5</td></tr><tr><td>Joint</td><td>93.7</td><td>77.6</td><td>78.1</td><td>78.4</td><td>93.7</td><td>79.0</td><td>73.6</td><td>78.4</td><td>93.7</td><td>79.0</td><td>73.6</td><td>78.4</td></tr></table>
+
+Table 1. Continual Semantic Segmentation performance (mIoU) on Pascal-VOC 2012 under different incremental scenarios. † indicates the results are excerpted from [4]. * suggests the results are reproduced with the official codes. 0 stands for the background class and all represents the mIoU over all classes including the background. Best results are highlighted in Red while runner-up in Blue.
+
+lower-bound baseline, which simply finetunes the model with the newly available data at each step. On the contrary, Joint, training a standard segmentation model on all classes within a single step, may serve as an upper bound for CSS approaches, though not always true if the CSS benchmark shows a long-tail nature with tail classes as new classes (e.g., ScanNet). For completeness, we include two general continual learning methods [26, 40] that are not specifically designed for semantic segmentation. Further, we mainly benchmark against the recent state-of-the-art methods taylored for segmentation [35, 4, 13, 36, 37, 61]. Following [4], we do not consider replay-based methods, e.g., [40]) in our comparison as the access to previous annotated data violates the standard class-incremental learning assumption. For the sake of generalization, we conduct experiments on both 2D image modality, i.e., Pascal-VOC 2012 [15] and ADE20K [66], and 3D point cloud modality, i.e. ScanNet [11]. Besides the old-, new- and all-class (including the background class) mIoU, we additionally report the background IoU because it is a special category under CSS with its semantic content varying across different learning steps.
+
+CSS protocols. As for the label masking policy for the incremental steps, [4] follows and adapts two different CSS protocols, namely disjoint [35] and overlapped [43]. Both protocols assume that only the novel classes at the current step are labeled while the others are masked out as background. However, in the disjoint protocol, input samples for step  $t$  should only contain elements (i.e., pixels or points) that belong to either current or previous classes, i.e.,  $\mathcal{C}^{0:t}$ . On the contrary, the overlapped protocol allows the input samples to contain either previous, current or future classes, i.e.,  $\mathcal{C}^{0:T}$ . The disjoint protocol makes a strong assumption that the current training data should not contain any samples
+
+of the class we would like to learn in the future, which is impracticable. Therefore, we adopt the more realistic overlapped protocol throughout our experiments.
+
+# 4.3. Implementation Details
+
+In 2D image modality, we follow the framework DeepLabv3 [7] and training hyperparameters of the baselines when building our LGKD upon them, except that we adopt synchronized batch normalization with different batch size and learning rate settings. In 3D point cloud modality, we use PointNet++ [39] with multi-scale grouping as our base model. We use Adam [25] with an initial learning rate  $10^{-2}$  for the first step and  $5 \times 10^{-3}$  for the subsequent steps and set the weight decay to 0. The learning rate is decayed by 0.7 every 100 epochs. We train our network for 500 epochs with a batch size of 32. We remove possible duplicated points before calculating the IoU metric. For all benchmarks, we report the mIoU results on the standard validation set. We run our experiments on two NVIDIA RTX 3090 GPUs for Pascal VOC 2012 and ADE20K while one for ScanNet. More details can be found in the supp. material.
+
+# 4.4. Performance
+
+Pascal-VOC 2012 consists of 10,582 images for training and 1449 images for validation with 20 foreground categories. Following [4], we conduct three different continual learning scenarios, i.e., adding one class (19-1), five classes simultaneously (15-5) and sequentially (15-1). The split of the class set for each class-incremental step follows the alphabetical order. Table 1 shows comprehensive results of our method along with other baselines. Clearly, FT performs poorly across all the incremental settings as the old
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">100-50 (2 steps)</td><td colspan="4">100-10 (6 steps)</td><td colspan="4">50-50 (3 steps)</td></tr><tr><td>0</td><td>1-100</td><td>101-150</td><td>all</td><td>0</td><td>1-100</td><td>101-150</td><td>all</td><td>0</td><td>1-50</td><td>51-150</td><td>all</td></tr><tr><td>ILT† [35]</td><td>10.9</td><td>18.4</td><td>14.4</td><td>17.0</td><td>8.5</td><td>0.0</td><td>3.1</td><td>1.1</td><td>8.6</td><td>3.4</td><td>12.9</td><td>9.7</td></tr><tr><td>MiB† [4]</td><td>19.7</td><td>40.7</td><td>17.2</td><td>32.8</td><td>0.0</td><td>38.6</td><td>11.1</td><td>29.2</td><td>0.0</td><td>46.5</td><td>21.0</td><td>29.3</td></tr><tr><td>REMINDER* [37]</td><td>26.6</td><td>42.1</td><td>16.1</td><td>33.4</td><td>24.0</td><td>38.9</td><td>20.3</td><td>32.6</td><td>24.2</td><td>48.0</td><td>20.1</td><td>29.4</td></tr><tr><td>SPPA [28]</td><td>-</td><td>42.9</td><td>19.9</td><td>-</td><td>-</td><td>41.0</td><td>12.5</td><td>-</td><td>-</td><td>49.8</td><td>23.9</td><td>-</td></tr><tr><td>PLOP* [13]</td><td>26.7</td><td>42.1</td><td>14.6</td><td>32.9</td><td>23.9</td><td>39.4</td><td>14.5</td><td>31.1</td><td>24.5</td><td>48.2</td><td>20.7</td><td>29.8</td></tr><tr><td>LGKD+PLOP (Ours)</td><td>27.4</td><td>43.6</td><td>25.7</td><td>37.5</td><td>26.0</td><td>42.1</td><td>22.0</td><td>35.4</td><td>25.3</td><td>49.4</td><td>29.4</td><td>36.0</td></tr><tr><td>RCIL* [61]</td><td>26.3</td><td>41.7</td><td>17.3</td><td>33.5</td><td>0.0</td><td>37.2</td><td>14.9</td><td>29.5</td><td>21.2</td><td>47.8</td><td>23.0</td><td>31.2</td></tr><tr><td>LGKD+RCIL (Ours)</td><td>26.7</td><td>43.3</td><td>25.1</td><td>37.2</td><td>24.0</td><td>42.2</td><td>20.4</td><td>34.9</td><td>25.7</td><td>49.1</td><td>27.2</td><td>34.4</td></tr><tr><td>Joint</td><td>29.1</td><td>42.6</td><td>28.2</td><td>37.7</td><td>29.1</td><td>42.6</td><td>28.2</td><td>37.7</td><td>29.1</td><td>49.2</td><td>32.1</td><td>37.7</td></tr></table>
+
+Table 2. The mIoU(%) of the last step on the ADE20K dataset for different overlapped continual learning scenarios.  $\dagger$  indicates the results are excerpted from [13]. * suggests the results are reproduced with the official codes. Best results are highlighted in Red while runner-up in Blue.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">18-1 (2 steps)</td><td colspan="4">14-5 (2 steps)</td><td colspan="4">14-1 (6 steps)</td></tr><tr><td>0</td><td>1-18</td><td>19</td><td>all</td><td>0</td><td>1-14</td><td>15-19</td><td>all</td><td>0</td><td>1-14</td><td>15-19</td><td>all</td></tr><tr><td>FT</td><td>19.2</td><td>0.0</td><td>28.6</td><td>2.4</td><td>18.6</td><td>0.0</td><td>44.3</td><td>12.0</td><td>19.2</td><td>0.0</td><td>7.8</td><td>2.9</td></tr><tr><td>ILT [35]</td><td>23.8</td><td>14.9</td><td>25.8</td><td>15.9</td><td>26.7</td><td>22.2</td><td>43.4</td><td>27.7</td><td>19.3</td><td>0.0</td><td>18.5</td><td>5.6</td></tr><tr><td>MiB [4]</td><td>55.1</td><td>55.0</td><td>32.2</td><td>53.8</td><td>52.9</td><td>57.8</td><td>45.5</td><td>54.5</td><td>45.5</td><td>56.5</td><td>29.7</td><td>49.2</td></tr><tr><td>LGKD+MiB (Ours)</td><td>55.2</td><td>55.5</td><td>40.8</td><td>54.8</td><td>54.3</td><td>57.6</td><td>49.0</td><td>55.3</td><td>51.9</td><td>55.7</td><td>38.4</td><td>51.2</td></tr><tr><td>Joint</td><td>53.2</td><td>55.0</td><td>40.9</td><td>54.2</td><td>53.2</td><td>56.7</td><td>47.5</td><td>54.2</td><td>53.2</td><td>56.7</td><td>47.5</td><td>54.2</td></tr></table>
+
+Table 3. Continual Semantic Segmentation performance (mIoU) on ScanNet under different incremental scenarios.  $\theta$  stands for the background class and all represents the mIoU over all classes including the background. All results are obtained with our implementation.
+
+class pixels are masked as background under CSS setup, and the model is explicitly taught to wrongly predict the old classes as background. EWC [26] yields similar results for novel classes while fairly outperforming the FT baseline in terms of base mIoU. Significant improvement on base classes across all settings can be observed when knowledge distillation is introduced in [40, 35]. MiB [4] obtains consistent improvement on both base and novel performance across different settings, due to addressing the background shift problem in CSS. Recent approaches [14, 61] obtain drastically higher performance against MiB when multiple incremental steps are involved (15-1). Despite their promising performance, our LGKD is able to further boost their performance by a large margin, especially on novel performance. For instance, LGKD+MiB achieves  $+74\%$  (40.4 vs. 23.2) improvement against MiB on novel mIoU under 19-1 setting. These promising novel class results along with the improved background performance prove that our LGKD loss can effectively alleviate the novel-background confusion problem.
+
+ADE20K [67] is a challenging dataset consisting of 20,210 training and 2000 testing images of complex scenes. Similar to VOC, we conducted three different CSS scenarios, i.e., 100-50, 100-10 and 50-50, as shown in Table 2. Specifically, we build our LGKD upon two competitive methods, i.e., PLOP [13] and RCIL [61], where LGKD achieves considerable improvement, especially on novel mIoU (up to  $+76\%$  from 14.6 to 25.7 with PLOP under 100-50 setting), setting new state-of-the-art across all incremental scenarios.
+
+ScanNet. As shown in Table 3, we compare our method with FT baseline, ILT [35] and MiB [4] on our proposed ScanNet benchmark to illustrate the generality of LGKD in 3D modality. Similar to their performance in 2D modality, FT and ILT suffer from catastrophic forgetting when learning new classes. However, MiB significantly outperforms the former methods across all three settings and metrics, in particular, the base class performance, which once again demonstrates the effectiveness of tackling the background shift in CSS. Compared with MiB, our LGKD further improves the novel performance by a large margin:  $+27\%$  (40.8 vs. 32.2) and  $+29\%$  (38.4 vs. 29.7) relative improvement of novel mIoU against MiB are obtained under 18-1 and 14-1 settings respectively, which proves the considerable importance of overcoming the background shift with correct class correspondence. It is also noteworthy that our approach also outperforms joint training, which shows that LGKD is also effective in alleviating long-tail issues.
+
+# 4.5. Ablation Study
+
+Unless otherwise stated, all ablative studies are conducted upon VOC 15-5 setting with PLOP [13] as the baseline. Refer to the supp. material for the rationale.
+
+Expand & Transplant. We compare three different designs for novel set distillation in Table 5. Firstly, One-hot is a hard-label distillation with all the class probabilities transplanted to the ground truth novel class, making the probability of the ground truth class equal to one. In this case, the distillation loss is equivalent to the standard one-hot cross
+
+<table><tr><td></td><td colspan="4">VOC 15-5 (2 steps)</td><td colspan="4">VOC 15-1 (6 steps)</td></tr><tr><td></td><td>0</td><td>1-15</td><td>16-20</td><td>all</td><td>0</td><td>1-15</td><td>16-20</td><td>all</td></tr><tr><td>PLOP [13]</td><td>89.7</td><td>74.2</td><td>47.5</td><td>68.6</td><td>84.7</td><td>64.6</td><td>21.0</td><td>55.2</td></tr><tr><td>+Standard KD [35]</td><td>90.7</td><td>74.4</td><td>49.2</td><td>69.2</td><td>88.4</td><td>66.4</td><td>13.4</td><td>54.8</td></tr><tr><td>+Unbiased KD [4]</td><td>90.2</td><td>77.1</td><td>51.8</td><td>71.7</td><td>86.3</td><td>67.1</td><td>22.6</td><td>57.4</td></tr><tr><td>+LGKD (Ours)</td><td>91.4</td><td>78.7</td><td>56.1</td><td>73.9</td><td>89.3</td><td>69.3</td><td>30.9</td><td>61.1</td></tr><tr><td></td><td colspan="4">ADE20k 100-50 (2 steps)</td><td colspan="4">ADE20k 100-10 (6 steps)</td></tr><tr><td></td><td>0</td><td>1-100</td><td>101-150</td><td>all</td><td>0</td><td>1-100</td><td>101-150</td><td>all</td></tr><tr><td>PLOP [13]</td><td>26.7</td><td>42.1</td><td>14.6</td><td>32.9</td><td>23.9</td><td>39.4</td><td>14.5</td><td>31.1</td></tr><tr><td>+Standard KD [35]</td><td>27.4</td><td>42.5</td><td>13.6</td><td>32.8</td><td>23.2</td><td>39.8</td><td>13.5</td><td>31.0</td></tr><tr><td>+Unbiased KD [4]</td><td>27.0</td><td>42.6</td><td>14.5</td><td>33.2</td><td>23.7</td><td>39.8</td><td>15.1</td><td>31.5</td></tr><tr><td>+LGKD (Ours)</td><td>27.4</td><td>43.6</td><td>25.7</td><td>37.5</td><td>26.0</td><td>42.1</td><td>22.0</td><td>35.4</td></tr></table>
+
+Table 4. Effect of different knowledge distillation terms. Standard KD is adopted in ILT [35] while unbiased KD is adopted in MiB [4].  
+
+<table><tr><td></td><td>0</td><td>1-15</td><td>16-20</td><td>all</td></tr><tr><td>One-hot</td><td>91.4</td><td>76.7</td><td>51.9</td><td>71.5</td></tr><tr><td>Expand only</td><td>90.4</td><td>78.5</td><td>30.9</td><td>67.7</td></tr><tr><td>Expand &amp; Transplant</td><td>91.4</td><td>78.7</td><td>56.1</td><td>73.9</td></tr></table>
+
+Table 5. Effect of our novel set distillation design including expand and transplant operations.  
+
+<table><tr><td>λb</td><td>λn</td><td>0</td><td>1-15</td><td>16-20</td><td>all</td></tr><tr><td>0</td><td>0</td><td>90.5</td><td>75.5</td><td>53.6</td><td>71.0</td></tr><tr><td>1</td><td>0</td><td>91.3</td><td>76.9</td><td>53.1</td><td>71.9</td></tr><tr><td>2</td><td>0</td><td>91.6</td><td>77.3</td><td>53.1</td><td>72.2</td></tr><tr><td>5</td><td>0</td><td>91.8</td><td>77.9</td><td>52.0</td><td>72.4</td></tr><tr><td>10</td><td>0</td><td>91.8</td><td>78.3</td><td>49.8</td><td>72.2</td></tr><tr><td>5</td><td>0.5</td><td>91.6</td><td>78.6</td><td>55.7</td><td>73.8</td></tr><tr><td>5</td><td>1</td><td>91.4</td><td>78.7</td><td>56.1</td><td>73.9</td></tr><tr><td>5</td><td>2</td><td>91.0</td><td>78.6</td><td>55.5</td><td>73.7</td></tr></table>
+
+Table 6. Effect of the weighted factors  ${\lambda }_{n}$  and  ${\lambda }_{\mathrm{b}}$  .
+
+entropy loss, discarding all the old class probabilities predicted by the old model. As expected, the base performance is slightly compromised compared with Expand & Transplant since the old class knowledge is not distilled. Nevertheless, the novel performance also drops since using hard labels for the novel class pixels will develop a bias toward novel classes, resulting in false positive predictions. This illustrates the significance of retaining the old class probabilities during transplant. Secondly, Expand only will encourage the new model to misclassify a novel class into either background or an old class. Thus, a sharp drop in novel mIoU from 56.1 to 30.9 is observed. This uncovers the importance of correcting the old model output by transplant to handle the background shift. Finally, transplanting the background probability into the corresponding ground truth novel class while retaining the old class probabilities can establish an appropriate class correspondence and facilitate learning new classes, hence yielding the best novel performance.
+
+Weighted factor. As table 6 shows, increasing  $\lambda_{\mathrm{b}}$  leads to
+
+higher base performance, but it will compromise the novel performance when  $\lambda_{\mathrm{b}}$  becomes too large, e.g.,  $\lambda_{\mathrm{b}} = 10$ . It indicates that the background set distillation is mainly responsible for retaining old knowledge. We set  $\lambda_{\mathrm{b}} = 5$  as the optimal solution as it obtains promising base mIoU while not sacrificing the novel mIoU severely. With the novel set distillation kicking in, the novel mIoU is significantly improved from 52.0 to 56.1 with  $\lambda_{n} = 1$ . Meanwhile, it also boosts the base mIoU from 77.9 to 78.7. This suggests that the novel set distillation facilitates learning new classes while maintaining the old class knowledge distilled from the old model.
+
+Effect of different KDs. To validate the superiority of LGKD, we compare our LGKD with other knowledge distillation (KD) terms on both VOC and ADE20k upon the same baseline PLOP [13] (Fig. 4), since PLOP does not adopt class probability-wise KD. Firstly, PLOP can hardly benefit from adding the standard KD [35], probably due to its tendency to misclassify new classes as the background (see Sec. 3.2). Secondly, though unbiased KD [4] yields a moderate boost on VOC, it makes almost no difference on the more challenging benchmark ADE20k. In contrast, LGKD consistently surpasses standard KD and unbiased KD by a significant margin, especially in terms of novel mIoU. Concretely, LGKD yields  $+8.3 / +11.2 / +6.9$  points improvement against unbiased KD on 15-1/100-50/100-10 setting.
+
+# 4.6. Error Analysis
+
+We carry out comprehensive error analysis on both Pascal-VOC 2012 and ScanNet datasets to validate the effectiveness of our LGKD in mitigating the novel-background confusion problem.
+
+2D image. As Fig. 4 shows, under the VOC 15-5 incremental scenario, naive finetuning completely forgets the old class bird while also mistaking the background for the novel class potted plant. Thanks to KD mechanism, ILT [35], REMINDER [37], MiB [4], PLOP [13] and RCIL [61] are capable to retain the old knowledge (bird and chair). Nev
+
+![](images/5b46f10fce6c80ead65967ac3a82345c812a9e7b6fe052777680b2c8a4d37ab7.jpg)  
+Old classes: bird chair  
+New classes: potted plant Sofa  
+Figure 4. Error analysis of baselines along with our LGKD under VOC 15-5 setting. MiB [4], PLOP [13] and RCIL [61] tend to mistake weed and suitcase (background—things of no interest) for potted plant and sofa (novel classes) respectively. Such confusion is effectively alleviated when equipped with our LGKD, except for the bird image with RCIL, where the confusion is slightly mitigated.
+
+![](images/a0cad4b3110eedc47cfa79b7f07cf95012dbbc554e649b9e231d0e4f5a22cb39.jpg)  
+Figure 5. Error analysis of MiB [4] vs. LGKD+MiB on step 4 (with bathtub depicted in pink as the novel class) of the ScanNet 14-1 incremental scenario. Our LGKD effectively alleviates the confusion made by MiB (see the black boxes) — mistaking the background (dark purple) for the novel class bathtub. Meanwhile, LGKD+MiB achieves decent accuracy on the novel class bathtub (see the red boxes).
+
+ertheless, these methods are consistently prone to misclassifying the background (weed and suitcase—things of no interest) as novel classes (potted plant and sofa). On the contrary, when incorporating our LGKD into MiB, PLOP and RCIL, the above novel-background confusion is effectively suppressed, except for the bird image with RCIL, where the confusion is slightly alleviated. The above examples demonstrate the effectiveness of LGKD in tackling the novel-background confusion issue. More visualization results can be found in the supp. material.
+
+3D point cloud. We show the error results in Fig. 5 for step 4 of the ScanNet 14-1 setting, where the current novel class is bathtub depicted in pink. Clearly, our method effic
+
+tively corrects the false prediction made by MiB [4] (see the black boxes), where the background points (highlighted in dark purple) are mistaken for the novel class bathtub. Meanwhile, our LGKD does not sacrifice the capability to recognize the true bathtub points (see red boxes).
+
+# 5. Conclusion
+
+In this work, we identify a key issue in continual semantic segmentation (CSS)—novel background confusion. This is generally caused by the corrupted class correspondence during knowledge distillation. To alleviate such confusion, we propose a novel solution termed Label-Guided Knowledge Distillation (LGKD), which can establish an appropriate class correspondence between the output of the old and the new model for distillation, guided by the ground truth label. We validate its effectiveness on both existing 2D benchmarks and our proposed 3D CSS benchmark, where LGKD can consistently outperform competing methods by a large margin, especially on novel mIoU, setting new state-of-the-art. Finally, we hope this work will inspire further in-depth exploration of 3D CSS in the future.
+
+# 6. Acknowledgement
+
+This work is supported by the Hyundai research grant (04OIS000252C130).
+
+# References
+
+[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware
+
+synapses: Learning what (not) to forget. In ECCV, 2018. 3  
+[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3366-3375, 2017. 3  
+[3] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In ECCV, 2018. 3  
+[4] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 2, 3, 4, 5, 6, 7, 8, 9  
+[5] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 3  
+[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE T-PAMI, 40(4):834-848, 2017. 2  
+[7] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 2, 6  
+[8] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 1  
+[9] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 1  
+[10] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021. 2  
+[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5, 6  
+[12] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 3  
+[13] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4040-4050, 2021. 2, 3, 6, 7, 8, 9  
+[14] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distilla
+
+tion for small-tasks incremental learning. In European Conference on Computer Vision, pages 86-102. Springer, 2020. 3,7  
+[15] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html, 2012.6  
+[16] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3146-3154, 2019. 1, 2  
+[17] Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211, 2013. 2  
+[18] Tyler L Hayes, Kushal Kafle, Robik Shrestha, Manoj Acharya, and Christopher Kanan. Remind your neural network to prevent catastrophic forgetting. In European Conference on Computer Vision, pages 466-483. Springer, 2020. 3  
+[19] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 2, 4  
+[20] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 3  
+[21] Guannan Hu, Wu Zhang, Hu Ding, and Wenhao Zhu. Gradient episodic memory with a soft constraint for continual learning. arXiv preprint arXiv:2011.07801, 2020. 3  
+[22] Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu Liu. Ccnet: Criss-cross attention for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 603-612, 2019. 1, 2  
+[23] Ahmet Iscen, Jeffrey Zhang, Svetlana Lazebnik, and Cordelia Schmid. Memory-efficient incremental learning through feature adaptation. In European conference on computer vision, pages 699-715. Springer, 2020. 3  
+[24] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. arXiv preprint arXiv:1711.10563, 2017. 3  
+[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6  
+[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2, 3, 6, 7  
+[27] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In CVPR, 2017. 1  
+[28] Zihan Lin, Zilei Wang, and Yixin Zhang. Continual semantic segmentation via structure preserving and projected feature
+
+alignment. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 345-361. Springer, 2022. 3, 6, 7  
+[29] Weide Liu, Chi Zhang, Henghui Ding, Tzu-Yi Hung, and Guosheng Lin. Few-shot segmentation with optimal transport matching and message flow. arXiv preprint arXiv:2108.08518, 2021. 3  
+[30] Weide Liu, Chi Zhang, Guosheng Lin, and Fayao Liu. Crnet: Cross-reference networks for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4165-4173, 2020. 3  
+[31] Weide Liu, Chi Zhang, Guosheng Lin, and Fayao Liu. Crcnet: Few-shot segmentation with cross-reference and region-global conditional networks. International Journal of Computer Vision, 130(12):3140-3157, 2022. 3  
+[32] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12245-12254, 2020. 3  
+[33] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 1, 2  
+[34] Andrea Maracani, Umberto Michieli, Marco Toldo, and Pietro Zanuttigh. Recall: Replay-based continual learning in semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7026-7035, 2021. 3  
+[35] Umberto Michieli and Pietro Zanuttigh. Incremental learning techniques for semantic segmentation. In ICCV-WS, pages 0–0, 2019. 2, 3, 4, 5, 6, 7, 8  
+[36] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1114-1124, 2021. 3, 6  
+[37] Minh Hieu Phan, Son Lam Phung, Long Tran-Thanh, Abdesselam Bouzerdoum, et al. Class similarity weighted knowledge distillation for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16866-16875, 2022. 3, 6, 7, 8  
+[38] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2  
+[39] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 6  
+[40] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 3, 6, 7  
+[41] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 3
+
+[42] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. In NeurIPS, 2017. 3  
+[43] Konstantin Shmelkov, Cordelia Schmid, and Karteek Alahari. Incremental learning of object detectors without catastrophic forgetting. In ICCV, 2017. 6  
+[44] Nan Song, Chi Zhang, and Guosheng Lin. Few-shot open-set recognition using background as unknowns. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5970-5979, 2022. 3  
+[45] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European conference on computer vision, pages 685–702. Springer, 2020. 2  
+[46] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2  
+[47] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5463-5474, 2021. 1  
+[48] Ju Xu and Zhanxing Zhu. Reinforced continual learning. Advances in Neural Information Processing Systems, 31, 2018. 3  
+[49] Shipeng Yan, Jiale Zhou, Jiangwei Xie, Songyang Zhang, and Xuming He. An em framework for online incremental learning of semantic segmentation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3052-3060, 2021. 3  
+[50] Ze Yang, Yali Wang, Xianyu Chen, Jianzhuang Liu, and Yu Qiao. Context-transformer: Tackling object confusion for few-shot detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12653-12660, 2020. 3  
+[51] Ze Yang, Chi Zhang, Ruibo Li, Yi Xu, and Guosheng Lin. Efficient few-shot object detection via knowledge inheritance. IEEE Transactions on Image Processing, 32:321-334, 2022. 3  
+[52] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 3  
+[53] Yuhui Yuan, Lang Huang, Jianyuan Guo, Chao Zhang, Xilin Chen, and Jingdong Wang. Ocnet: Object context network for scene parsing. arXiv preprint arXiv:1809.00916, 2018. 2  
+[54] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, 2017. 3  
+[55] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12203-12213, 2020. 3  
+[56] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Differentiable earth mover's distance for few-shot
+
+learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5632-5648, 2022. 3  
+[57] Chi Zhang, Henghui Ding, Guosheng Lin, Ruibo Li, Changhu Wang, and Chunhua Shen. Meta navigator: Search for a good adaptation policy for few-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9435-9444, 2021. 3  
+[58] Chi Zhang, Guosheng Lin, Fayao Liu, Jiushuang Guo, Qingyao Wu, and Rui Yao. Pyramid graph networks with connection attentions for region-based one-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9587-9595, 2019. 3  
+[59] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5217–5226, 2019. 3  
+[60] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12455-12464, 2021. 3  
+[61] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053-7064, 2022. 2, 3, 6, 7, 8, 9  
+[62] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020. 2  
+[63] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 1, 2  
+[64] Hanbin Zhao, Fengyu Yang, Xinghe Fu, and Xi Li. Rbc: Rectifying the biased context in continual semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 55-72. Springer, 2022. 3  
+[65] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 1  
+[66] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 6  
+[67] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 7  
+[68] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical
+
+and asymmetrical 3d convolution networks for lidar segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9939-9948, 2021. 2

2023/Label-Guided Knowledge Distillation for Continual Semantic Segmentation on 2D Images and 3D Point Clouds/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:acf9ebd154d68e86347b37c030baeb0d92620a4862f4910817fe1e122552fd07
+size 757427

2023/Label-Noise Learning with Intrinsically Long-Tailed Data/ca2911cd-1897-45f1-bc9c-ab89b815151c_content_list.json

@@ -0,0 +1,1569 @@
+[
+    {
+        "type": "text",
+        "text": "Label-Noise Learning with Intrinsically Long-Tailed Data",
+        "text_level": 1,
+        "bbox": [
+            192,
+            130,
+            777,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yang Lu $^{1,2*}$  Yiliang Zhang $^{1,2}$  Bo Han $^{3}$  Yiu-ming Cheung $^{3}$  Hanzi Wang $^{1,2}$",
+        "bbox": [
+            158,
+            178,
+            807,
+            199
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Fujian Key Laboratory of Sensing and Computing for Smart City, School of Informatics, Xiamen University, Xiamen, China",
+        "bbox": [
+            114,
+            200,
+            854,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing,",
+        "bbox": [
+            254,
+            217,
+            712,
+            233
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ministry of Education of China, Xiamen University, Xiamen, China",
+        "bbox": [
+            281,
+            234,
+            684,
+            250
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{3}$ Department of Computer Science, Hong Kong Baptist University, Hong Kong, China",
+        "bbox": [
+            228,
+            253,
+            738,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "luyang@xmu.edu.cn ylzhangcs@hotmail.com {bhanml, ymc}@comp.hkbu.edu.hk hanzi.wang@xmu.edu.cn",
+        "bbox": [
+            122,
+            272,
+            844,
+            286
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            321,
+            313,
+            338
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Label noise is one of the key factors that lead to the poor generalization of deep learning models. Existing label-noise learning methods usually assume that the ground-truth classes of the training data are balanced. However, the real-world data is often imbalanced, leading to the inconsistency between observed and intrinsic class distribution with label noises. In this case, it is hard to distinguish clean samples from noisy samples on the intrinsic tail classes with the unknown intrinsic class distribution. In this paper, we propose a learning framework for label-noise learning with intrinsically long-tailed data. Specifically, we propose two-stage bi-dimensional sample selection (TABASCO) to better separate clean samples from noisy samples, especially for the tail classes. TABASCO consists of two new separation metrics that complement each other to compensate for the limitation of using a single metric in sample separation. Extensive experiments on benchmarks demonstrate the effectiveness of our method. Our code is available at https://github.com/Wakings/TABASCO.",
+        "bbox": [
+            75,
+            353,
+            473,
+            655
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            670,
+            209,
+            686
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Under the support of a large amount of high-quality labeled data, deep neural networks have achieved great success in various fields [27, 40, 7]. However, it is expensive and difficult to obtain a large amount of high-quality labeled data in many practical applications. Instead, the commonly used large-scale training data is usually obtained from the Internet or crowdsourcing platforms like Amazon Mechanical Turk, which is unreliable and may be mislabeled [52, 43]. The models trained on this kind of unreliable data, called noisy-labeled data, often produce poor generalization performance because deep neural networks tend to overfit noisy samples due to their large model capacity. In",
+        "bbox": [
+            75,
+            695,
+            468,
+            878
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/22e148d9b056a7487ba405eda466053370f0d945b7ff477a2b438a49a21ec20e.jpg",
+        "image_caption": [
+            "(a) Class distribution"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            321,
+            687,
+            425
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/348c53ee3757a4e47bad2047be5b399853813846bb127ad12b77972385732830.jpg",
+        "image_caption": [
+            "(b) Loss distribution",
+            "Figure 1: (a) An example of observed class distribution with noisy labels with long-tailed intrinsic class distribution. (b) The training loss of each sample under noisy-labeled and long-tailed data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            694,
+            321,
+            870,
+            425
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "the literature, there are some works to obtain a robust model trained on noisy-labeled data [11, 44]. Among them, the most straightforward and effective way is to differentiate between clean and noisy samples based on their differences in specific metrics, such as training loss [12, 29, 23].",
+        "bbox": [
+            496,
+            522,
+            892,
+            598
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "These label-noise learning methods generally assume that the intrinsic class distribution of the training data is balanced, where each class has almost the same number of samples in terms of their unknown ground-truth labels. However, the data in real-world applications are often imbalanced, e.g., the LVIS dataset [9] and the iNaturalist dataset [17]. The class imbalance usually exhibits in the form of a long-tail distribution, where a small portion of classes possess a large number of samples, and the other classes possess a small number of samples only [38, 18]. In this case, the model training tends to the head classes and ignores the tail classes [60, 62]. When both noisy labels and long-tail distribution exist, training a robust model is even more challenging. There are two key challenges in this scenario. (1) Distribution inconsistency: The observed and intrinsic distributions are likely inconsistent due to noise labels, making the model more difficult to discover and focus on the intrinsic tail classes. As illustrated in Fig. 1(a), the intrinsic class distribution of the dataset is long-tailed, while the existence of noisy labels makes the distribution more",
+        "bbox": [
+            496,
+            599,
+            893,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Yang Lu is the corresponding author: luyang@xmu.edu.cn",
+        "bbox": [
+            94,
+            886,
+            413,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "1369",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "balanced. The intrinsic tail classes, e.g., classes 6 and 9, are occupied by a large number of noisy data, making them no longer tail classes by observation. (2) Tail inseparability: Even if the tail class is identified, it is more difficult than ever to distinguish between clean and noisy samples in the tail class because clean samples are overwhelmed by noises that make their values of the separation metric highly similar. As illustrated in Fig. 1(b), the training loss of clean and noisy samples in the tail class are generally inseparable compared with the ones in the other classes. Several preliminary works have studied the joint problem of label noise and long-tail distribution [21, 49, 24, 2]. These methods implicitly reduce the complexity of the problem by assuming a similar noise rate for each class. However, this assumption is too strong to apply because noisy samples from the head classes may be the majority, resulting in a higher noise rate in the tail class than in other classes.",
+        "bbox": [
+            75,
+            90,
+            472,
+            347
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a Two-stAge Bi-dimensionAl Sample seleCtiOn (TABASCO) strategy to address the problem of label-noise learning with intrinsically long-tailed data. In the first stage, we propose to use two new separation metrics for sample separation, i.e., weighted Jensen-Shannon divergence (WJSD) and adaptive centroid distance (ACD), which work corporately to separate clean samples from noisy samples in the tail classes. The proposed metrics are complementary, where WJSD separates the samples from the output perspective while ACD does that from the feature perspective. In the second stage, we determine the separation dimension with better separability for each class and perform sample selection. In order to evaluate the method uniformly and effectively, we introduce two benchmarks with real-world noise and intrinsically long-tailed distribution. The main contributions of our work can be summarized as follows:",
+        "bbox": [
+            75,
+            348,
+            472,
+            604
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We present a more general problem of label-noise learning with intrinsically long-tailed data. The key challenges in this problem are distribution inconsistency and tail inseparability.",
+            "- We propose an effective solution called TABASCO. With the help of two new separation metrics, it is able to effectively identify and select clean samples of the intrinsic tail class.",
+            "- We introduce two benchmarks with real-world noise and intrinsically long-tailed distribution. Extensive experiments on them show the effectiveness of our method and the limitations of existing methods."
+        ],
+        "bbox": [
+            94,
+            614,
+            468,
+            806
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            821,
+            218,
+            835
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Label-Noise Learning",
+        "text_level": 1,
+        "bbox": [
+            76,
+            845,
+            281,
+            863
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "A straightforward strategy to deal with noisy data is to reduce the proportion of noise in training samples by sepa",
+        "bbox": [
+            76,
+            869,
+            468,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "rating noisy samples from clean samples. Methods such as co-teaching [12] and DivideMix [29] adopt the small loss trick, while Jo-SRC [55] and UNICON [23] use Jensen-Shannon divergence instead for sample selection. In contrast to methods that separate at the label level, some methods [30, 34] attempt to separate samples at the feature space. There are also some methods to avoid over-fitting the model to noisy data by imposing regularization constraints on model parameters [50, 10] or labels [37, 58, 32]. Other methods mitigate the influence of noisy data by adjusting the loss functions, such as backward and forward loss correction [36], gold loss correction [16], MW-Net [41] and Dual-T [54].",
+        "bbox": [
+            496,
+            90,
+            890,
+            287
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Long-tail Learning",
+        "text_level": 1,
+        "bbox": [
+            500,
+            295,
+            683,
+            311
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Re-balancing the data for long-tail distributions is a classical strategy to solve the problem of long-tail learning, such as re-sampling [4, 31, 8, 13] and data augmentation [61, 57, 5]. In addition, there are methods to improve the model generalization by introducing long-tail robust loss functions [6, 3, 45, 39]. Methods such as FTL [56], RIDE [46] and DiVE [15] try to solve the problem by using the idea of transfer learning. Recently, approaches based on decoupling [22, 61, 59] split end-to-end learning into feature learning and classifier retraining such that the obtained feature extractor is less affected by the long-tail distribution. In contrast to the above methods with the supervised learning paradigm, CReST [47], ABC [28] and DARP [25] attempt to solve the long-tail problem in the manner of semi-supervised learning.",
+        "bbox": [
+            496,
+            318,
+            890,
+            546
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3. Label-Noise Learning on Long-tailed Data",
+        "text_level": 1,
+        "bbox": [
+            500,
+            553,
+            861,
+            569
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Research on this joint problem has just been explored. CNLCU [51] relaxes the constraint of the small loss trick by regarding a portion of large loss samples as clean samples to reduce the probability of misclassifying clean samples to noisy samples in the tail class. RoLT [49] uses the distance from the samples to the centroid of the current class instead of the training loss for sample selection. HAR [2] uses an adaptive approach to regularize noise and tail class samples. Karthik et al. [24] uses the idea of decoupling to fine-tune the loss function for better robustness after feature learning by the self-supervised method. ULC [19] introduces uncertainty to enhance the separation ability of noise samples. H2E [53] reduces hard noises to easy ones by learning a classifier as noise identifier invariant to the class and context distributional changes.",
+        "bbox": [
+            496,
+            575,
+            890,
+            801
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3. Problem Definition",
+        "text_level": 1,
+        "bbox": [
+            500,
+            815,
+            684,
+            830
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Given a training set  $\\mathcal{D} = \\{x_i, \\hat{y}_i\\}_{i=1}^N$ , where  $\\hat{y}_i \\in [M]$  is the observed label of the sample  $x_i$ .  $N$  is the number of training samples, and  $M$  is the number of classes. In our problem,  $\\mathcal{D}$  has the following properties:",
+        "bbox": [
+            496,
+            838,
+            890,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "1370",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/e2fb5fbcd490fc645823e0de49b016a33494bce676d8b24cf9f2e12b8ee9f4bb.jpg",
+        "image_caption": [
+            "Figure 2: The proposed framework for label-noise learning with intrinsically long-tailed data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            83,
+            90,
+            890,
+            272
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Noisy-labeled. There is a subset of samples  $\\hat{\\mathcal{D}} \\in \\mathcal{D}$ , where a sample  $\\{x, \\hat{y}\\} \\in \\hat{\\mathcal{D}}$  has an unknown ground-truth label  $y$  different from its observed label  $\\hat{y}$ .",
+            "- Long-tailed. The ground-truth class distribution is long-tailed. Supposing  $n_c$  is the number of the samples in class  $c$ , we have  $n_1 > n_2 > \\dots > n_M$ ."
+        ],
+        "bbox": [
+            94,
+            308,
+            468,
+            405
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We call this kind of training data intrinsically long-tailed because the observed class distribution may not be long-tailed due to the existence of noisy labels. Therefore, our goal is to learn from noisy-labeled data with the intrinsically long-tailed class distribution. The problem is challenging from two aspects. On the one hand, directly applying long-tail learning methods [6, 3, 45] is infeasible because the observed class distribution may be inconsistent with the intrinsic class distribution. On the other hand, label-noise learning methods perform poorly in the tail classes because they usually contain more noisy samples due to insufficient data, as shown in Fig. 1(a). It results in high noise rates in the tail classes, which brings great challenges to separating clean samples and noisy samples for further training. Under this circumstance, the existing sample selection methods [29, 23] often fail to select clean samples in the tail class. The main reason is that it is difficult to distinguish the tail class data from the noisy label data by existing metrics like cross-entropy loss, as shown in Fig. 1(b).",
+        "bbox": [
+            75,
+            410,
+            472,
+            698
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4. Proposed Method",
+        "text_level": 1,
+        "bbox": [
+            76,
+            709,
+            250,
+            726
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a two-stage bi-dimensional sample selection (TABASCO) to address the problem of label-noise learning with intrinsically long-tailed data. TABASCO decouples the sample selection process into two stages: (1) sample separation, and (2) dimension and cluster selection. First, given an initial model  $\\theta$  trained on the original training data  $\\mathcal{D}$ , we propose to calculate bi-dimension metrics for each sample in each observed class based on the outputs and features of the model. Second, we determine the separation dimension with better separability for each class, and we adopt the corresponding selection strategy to",
+        "bbox": [
+            75,
+            734,
+            472,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "select the cluster with more clean samples based on the selected metric. Last, we adopt semi-supervised learning to update the model by regarding the selected clean cluster as labeled data. The overall framework is shown in Fig. 2.",
+        "bbox": [
+            496,
+            309,
+            890,
+            369
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1. Bi-Dimensional Separation Metrics",
+        "text_level": 1,
+        "bbox": [
+            498,
+            378,
+            807,
+            395
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Due to the deficiency of a single separation metric to distinguish clean samples from noisy samples in the complex situation of noisy labels with intrinsic long-tail distribution, we propose to jointly use two metrics from different perspectives: weighted Jensen-Shannon divergence (WJSD) and adaptive centroid distance (ACD). Both metrics are specifically designed for the joint problem and are complementary to cover the case when only one of them cannot separate clean samples from noisy samples well. WJSD fully utilizes the information of prediction confidence, while ACD relies on the distance in the feature space. Thus, using clustering on samples according to the values of their bi-dimensional metrics has the flexibility to separate samples with different imbalance ratios, noise ratios, and noise types.",
+        "bbox": [
+            496,
+            402,
+            890,
+            630
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We first reduce the separation granularity to alleviate the interference from the head class to the tail class. Specifically, the separation is performed within each observed class according to the proposed bi-dimensional metrics. We first divide training set  $\\mathcal{D}$  into subsets according to the observed labels  $\\mathcal{D}_c = \\{(x,\\hat{y})\\mid \\hat{y} = c\\}$ . Given a sample  $(x_{i},\\hat{y}_{i})\\in \\mathcal{D}_{c}$ , the model predictions obtained from the model  $\\theta$  is denoted as  $\\mathbf{p}_i = [p_i^1,p_i^2,\\dots,p_i^M ]$ , where  $p_i^j$  is the  $j$ 's dimension of vector  $\\mathbf{p}_i$ . The one-hot representation of the observed label  $\\hat{y}_i$  is denoted as  $\\hat{\\mathbf{y}}_i$ .",
+        "bbox": [
+            496,
+            631,
+            890,
+            782
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Weighted Jensen-Shannon Divergence The Jensen-Shannon Divergence (JSD) is a commonly used metric to separate samples by assessing the variability of model predictions [55, 23]. It is defined as:",
+        "bbox": [
+            496,
+            787,
+            890,
+            848
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nJ S D \\left(x _ {i}\\right) = \\frac {1}{2} K L \\left(\\mathbf {p} _ {i} \\left\\| \\frac {\\mathbf {p} _ {i} + \\hat {\\mathbf {y}} _ {i}}{2}\\right) + \\frac {1}{2} K L \\left(\\hat {\\mathbf {y}} _ {i} \\left\\| \\frac {\\mathbf {p} _ {i} + \\hat {\\mathbf {y}} _ {i}}{2}\\right), \\right. \\right. \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            500,
+            852,
+            890,
+            896
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "1371",
+        "bbox": [
+            482,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $KL(\\cdot \\|\\cdot)$  is the Kullback-Leibler divergence. When the intrinsic distribution is balanced, the JSD values of clean samples are generally lower than the noisy samples, so the separation can be easily modeled. However, when the distribution is intrinsically long-tailed, the noisy and clean samples in a tail class tend to obtain similar prediction confidence on the observed class  $c$ , because their predictions are both towards the head classes. In this case, using JSD as a separation metric fails to separate them. The reason is given by Theorem 1, which shows the upper bound of the absolute value of the JSD difference between two samples.",
+        "bbox": [
+            75,
+            90,
+            472,
+            257
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Theorem 1 (Upper bound of JSD difference) Suppose  $x_{i}$  and  $x_{j}$  are two samples in class  $c$ ,  $p_{i}^{c}$  and  $q_{i}^{c}$  are the  $c$ 's dimension of their prediction confidence  $\\mathbf{p}_{i} = [p_{i}^{1}, p_{i}^{2}, \\dots, p_{i}^{M}]$  and  $\\mathbf{p}_{j} = [p_{j}^{1}, p_{j}^{2}, \\dots, p_{j}^{M}]$ , respectively. The upper bound of the absolute value of the difference between their JSD values is given by:",
+        "bbox": [
+            76,
+            257,
+            472,
+            349
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n| J S D (x _ {i}) - J S D (x _ {j}) | \\leq \\frac {1}{2} \\log \\left(\\frac {p _ {i} ^ {c} + 1}{p _ {i} ^ {c}}\\right) | p _ {i} ^ {c} - p _ {j} ^ {c} |.\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            93,
+            352,
+            454,
+            386
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Theorem 1 shows that the difference between values of JSD is only determined by the prediction confidence of the observed class, i.e.,  $p_i^c$  and  $p_j^c$ . In addition, as  $|p_i^c - p_j^c|$  gets smaller with a fixed value of  $p_i^c$ ,  $|JSD(x_i) - JSD(x_j)|$  is also smaller. This indicates that when two samples are in the same class  $c$  with close values of  $p_i^c$ , their JSD values are also close, which makes it difficult to separate them.",
+        "bbox": [
+            75,
+            388,
+            468,
+            494
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Nevertheless, we can utilize the prediction confidence on other classes, i.e.,  $p_i^d$  for  $d \\neq c$ , to distinguish two samples even if their values of  $p_i^c$  are close. Specifically, we propose WJSD by imposing an additional weight on JSD to further distinguish samples by inspecting their prediction confidence on other classes rather than the observed class. First, we take the maximum prediction confidence  $\\max(\\mathbf{p}_i)$  into account because it may reflect the confidence of a noisy sample's ground-truth class  $y_i$ . Then, we calculate the ratio of the maximum prediction confidence to the prediction confidence of the observed class  $\\max(\\mathbf{p}_i) / p_i^c$ . In this manner, the additional weight is greater than one only if the class with the maximum prediction confidence is not the observed class. The larger gap between  $\\max(\\mathbf{p}_i)$  and  $p_i^c$  makes the weight higher. To avoid exceptionally large weights by the division during normalization over all samples in class  $c$  for later clustering, we set the upper bound according to the averaged prediction confidence over all samples in class  $c$ . Finally, the additional weight can be calculated as follows:",
+        "bbox": [
+            75,
+            494,
+            472,
+            796
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nW \\left(x _ {i}\\right) = \\min  \\left(\\max  \\left(\\mathbf {p} _ {i}\\right) / p _ {i} ^ {c}, \\max  \\left(\\bar {\\mathbf {p}} _ {c}\\right) / \\bar {p} _ {c} ^ {c}\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            127,
+            801,
+            468,
+            820
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\bar{\\mathbf{p}}_c = [\\bar{p}_c^1,\\bar{p}_c^2,\\dots,\\bar{p}_c^M] = \\frac{1}{|\\mathcal{D}_c|}\\sum_{i = 1}^{|\\mathcal{D}_c|}\\mathbf{p}_i$  is the average prediction confidence of class  $c$ . Thus, WJSD can be calculated by:",
+        "bbox": [
+            76,
+            825,
+            468,
+            875
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nW J S D \\left(x _ {i}\\right) = W \\left(x _ {i}\\right) \\times J S D \\left(x _ {i}\\right). \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            151,
+            880,
+            468,
+            897
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Remarks: Compared with the value of JSD that is only related to the prediction confidence on the observed class, adopting the maximum prediction confidence in WJSD can better separate clean samples from noisy samples.",
+        "bbox": [
+            496,
+            90,
+            890,
+            152
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Adaptive Centroid Distance. Although the sample separability of WJSD is greatly improved compared with JSD, the separation metric still relies on model prediction. When all the noisy samples labeled in a class are from another class, e.g., asymmetric noise, the model prediction for both noisy and clean samples will be highly similar because it is easy to learn a classifier that maps the features from two different classes into one. It results in low discrimination between the model predictions of clean and noisy samples in this case. Thus, solely using WJSD may not be enough to separate the clean and noisy samples when model predictions (in terms of all classes, not only about  $p^c$ ) are close.",
+        "bbox": [
+            496,
+            156,
+            890,
+            338
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Except for separating the samples in the output space, we propose another metric calculated in the feature space to eliminate the bias from the classifier because the learned features are more robust to noise labels. Specifically, for a given sample, we can calculate the distance between its feature and the class feature centroid to evaluate how the feature of a sample deviates from its class centroid. This approach is only practical when the quality of the centroid is high. Directly calculating the centroid according to the observed label may not be accurate because it involves a certain number of features of the noisy samples. Therefore, we define purity in Definition 1 to assess the quality of the centroid of class  $c$ .",
+        "bbox": [
+            496,
+            338,
+            890,
+            532
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Definition 1 (Purity) Suppose a noisy sample set of class  $c$  be  $\\mathcal{D}_c = \\{(x,\\hat{y})\\mid \\hat{y} = c\\}$ , where the intrinsic label corresponding to the sample is  $y$ . The purity  $P_{\\mathcal{D}_c}$  of set  $\\mathcal{D}_c$  is calculated by:",
+        "bbox": [
+            496,
+            534,
+            890,
+            594
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP _ {\\mathcal {D} _ {c}} = \\max  _ {k = 1, \\dots , M} \\left\\{\\frac {\\sum_ {i = 1} ^ {N} I \\left(y _ {i} = k\\right)}{N} \\right\\}, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            563,
+            598,
+            890,
+            640
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $I(\\cdot)$  is an indicator function and  $N$  is the number of samples in the set  $\\mathcal{D}_c$ .",
+        "bbox": [
+            496,
+            643,
+            893,
+            672
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Purity indicates the proportion of an intrinsic class that takes the majority in the observed class  $c$ . The higher the purity of a class set used to calculate the centroid, the better the centroid helps to distinguish between noise and clean samples. RoLT [49] adopts a similar idea to use class feature centroid for noisy sample detection. However it directly calculates the feature centroid on the observed class, such that it inevitably suffers from the existence of noise features when the purity is low, especially in our problem.",
+        "bbox": [
+            496,
+            674,
+            890,
+            810
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In order to improve the purity of the class feature centroid for distance calculation between features, we propose ACD as the second separation metric to separate samples jointly with WJSD. The feature centroid for each class is adaptively updated by involving samples with high confidence from the observed class. The class centroid  $\\mathbf{o}_c$  based",
+        "bbox": [
+            496,
+            810,
+            890,
+            898
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "1372",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "on a high-confidence sample set  $\\mathcal{D}_c^H$  is calculated by:",
+        "bbox": [
+            76,
+            90,
+            434,
+            107
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {o} _ {c} = \\frac {1}{| \\mathcal {D} _ {c} ^ {H} |} \\sum_ {i = 1} ^ {| \\mathcal {D} _ {c} ^ {H} |} \\mathbf {f} _ {i}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            169,
+            111,
+            468,
+            155
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {D} _ {c} ^ {H} = \\left\\{x _ {i} \\mid x _ {i} \\in \\mathcal {D} _ {c}, p _ {i} ^ {t _ {c}} > H _ {c} \\right\\}, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            163,
+            157,
+            468,
+            176
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{f}_i$  is the feature of  $x_{i}$ , and  $t_c = \\mathrm{argmax}_c\\{\\bar{p}_c^j\\}$  is the class index with the largest average prediction confidence of class  $c$ . Thus, we can use the prediction confidence of class  $t_c$  for sample  $x_{i}$ , e.g.,  $p_i^{t_c}$ , as the selection criteria compare with the threshold  $H_{c}$ .  $\\mathcal{D}_c^H$  is constructed by the samples in  $D_{c}$  whose corresponding prediction confidence of class  $t_c$  is higher than  $H_{c}$ . The high-confidence threshold  $H_{c}$  is defined as:",
+        "bbox": [
+            76,
+            181,
+            472,
+            303
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nH _ {c} = \\frac {1}{D _ {c}} \\sum_ {i = 1} ^ {| D _ {c} |} w _ {i} \\times p _ {i} ^ {t _ {c}}, \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            305,
+            468,
+            347
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nw _ {i} = \\max  \\left(1, p _ {i} ^ {t _ {c}} / \\bar {p} _ {c} ^ {t _ {c}}\\right). \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            194,
+            349,
+            468,
+            369
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "$H_{c}$  is calculated by the weighted sum of the prediction confidence of class  $t_c$  for all samples  $x_{i}$ . The weight  $w_{i}$  increases when a sample's prediction confidence of class  $t_c$  is higher than its class average. Thus, the threshold is high enough to ensure the purity of the selected samples. High confidence samples are more representative and more likely to be in the same class. It is worth noting that we choose more representative samples rather than clean samples because the clean samples in the tail classes are easily overwhelmed by the noise samples in our problem. In this case, it is difficult to select the clean samples directly because the noise samples may be more representative in the tail class. As for asymmetric noise, the choice of clean or noisy samples to obtain centroid for sample separation is equivalent.",
+        "bbox": [
+            75,
+            373,
+            468,
+            585
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In each round, the centroids are adaptively adjusted because the feature extractor in the model is updated, as well as the set  $\\mathcal{D}_c^H$ . Therefore, the proposed metric ACD can be calculated by:",
+        "bbox": [
+            75,
+            585,
+            470,
+            646
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nA C D \\left(x _ {i}\\right) = \\cos \\left(\\mathbf {f} _ {i}, \\mathbf {o} _ {c}\\right). \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            651,
+            468,
+            667
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Remarks: The proposed two metrics, WJSD and ACD, are complementary for sample separation in the tail classes. When the ground-truth classes of the noisy samples are diverse, e.g. symmetric noise, WJSD plays the dominant role because the predictions can hardly be unified due to the class diversity of noisy samples. They are more likely to be predicted towards their ground-truth class, which can be captured by  $\\max (\\mathbf{p}_i)$ . In this case, ACD may not be effective in separating clean samples from noisy samples because the messy noisy samples may affect the purity of the calculated centroid. On the other hand, when the ground-truth classes of the noisy samples belong to only one class, e.g., asymmetric noise, WJSD tends to produce similar predictions for noisy and clean samples, while the adaptive centroid in ACD can help distinguish them in the feature space.",
+        "bbox": [
+            75,
+            672,
+            472,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [
+            "Algorithm 1: The Dimension Selection Strategy"
+        ],
+        "code_body": "Input: Noise sample set  $\\mathcal{D}_c$  in class  $c$ , threshold  $\\eta$   \nOutput: Cluster  $\\mathcal{G}_c^1, \\mathcal{G}_c^2$   \n1 Obtain  $\\mathcal{G}_{wjsd}^1, \\mathcal{G}_{wjsd}^2$  and separation threshold  $d$  by applying GMM with values of WJSD to all samples in  $\\mathcal{D}_c$   \n2 Obtain  $\\mathcal{G}_{acd}^1, \\mathcal{G}_{acd}^2$  by applying GMM with values of ACD to all samples in  $\\mathcal{D}_c$   \n3 Calculate the mean  $\\mu_1, \\mu_2$  and variance  $\\sigma_1, \\sigma_2$  of WJSD for  $\\mathcal{G}_{acd}^1, \\mathcal{G}_{acd}^2$   \n4 if  $\\mu_1 < d < \\mu_2$  and  $\\sigma_2 / \\sigma_1 < \\eta$  then  \n5 return  $\\mathcal{G}_{wjsd}^1, \\mathcal{G}_{wjsd}^2$   \n6 else if  $\\mu_1 > d$  and  $\\mu_2 > d$  then  \n7 return  $\\mathcal{G}_{wjsd}^1, \\mathcal{G}_{wjsd}^2$   \n8 else  \n9 return  $\\mathcal{G}_{acd}^1, \\mathcal{G}_{acd}^2$   \n10 end",
+        "bbox": [
+            506,
+            112,
+            862,
+            356
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Bi-Dimensional Sample Selection",
+        "text_level": 1,
+        "bbox": [
+            500,
+            383,
+            790,
+            401
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Once the values of the bi-dimensional metrics for all samples in class  $c$  are calculated, each sample can be represented by a point in a 2D space. For each dimension, Gaussian mixture model (GMM) can be adopted to separate the samples in class  $c$  into two clusters  $\\mathcal{G}_c^1$ ,  $\\mathcal{G}_c^2$ .",
+        "bbox": [
+            498,
+            410,
+            892,
+            488
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Dimension Selection. Because each metric shows different separability for different noise types, we first propose a dimension selection strategy to select a proper dimension to separate clean and noisy samples. We consider three cases to select the proper dimension: (a) both of them are acceptable; (b) the separability of WJSD is better; (c) the separability of ACD is better. In short, we only need to figure out if WJSD is suitable for sample separation. If it is certain that WJSD does not show good separability, ACD will be adopted. Fig. 3 shows three cases for dimension selection.",
+        "bbox": [
+            496,
+            494,
+            892,
+            646
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Therefore, we use the statistics of WJSD as the criterion of dimension selection. We measure the means  $(\\mu_1,\\mu_2)$  and standard deviations  $(\\sigma_{1},\\sigma_{2})$  of WJSD for two clusters separated by GMM in terms of ACD. We have two cases that can determine when WJSD is better. We mainly compare",
+        "bbox": [
+            496,
+            647,
+            893,
+            724
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/2de398f5880d8ca6a5cb6c69c9c5ba5bff5667799a17761e1e0b4df02a61fdfb.jpg",
+        "image_caption": [
+            "Figure 3: Examples of three cases of the optimal separation dimension correspond to sample distribution in bidimensional metrics."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            734,
+            877,
+            851
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "1373",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "$\\mu_{1}$  and  $\\mu_{2}$  with the threshold  $d$  of GMM in WJSD. First, as shown in Fig. 3 (b), if one cluster  $\\mathcal{G}_{acd}^{1}$  of ACD has a  $\\mu_{1}$  less than  $d$  and a  $\\sigma_{1}$  much larger than  $\\sigma_{2}$  in another cluster with a  $\\mu_{2}$  larger than  $d$ , it means that cluster  $\\mathcal{G}_{acd}^{1}$  has more clean samples but also many noise samples. So in this case, ACD is improper and WJSD is proper. Second, if both  $\\mu_{1}, \\mu_{2}$  are greater than  $d$ , it means that it is a representation of the previous case in the tail class because there are fewer clean samples in the cluster. So, in this case, WJSD is proper. The strategy is summarized in Algorithm 1.",
+        "bbox": [
+            75,
+            90,
+            472,
+            243
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Cluster Selection. Once we have determined the separation dimension, we can obtain two sample clusters  $\\mathcal{G}_c^1$  and  $\\mathcal{G}_c^2$ . However, it is not clear which cluster contains more clean samples because the noisy samples may overwhelm the clean samples in some cases. Therefore, we need to further select a cluster between  $\\mathcal{G}_c^1$  and  $\\mathcal{G}_c^2$  as the clean one. For the cluster separated based on WJSD, we simply choose the one with a smaller average WJSD value as the cluster with more clean samples, denoted as  $\\mathcal{D}_c^{clean}$  [29, 23]. For the cluster separated based on ACD, we cannot directly select the cluster closest to the centroid as the cluster with more clean samples since we take more representative samples, which may compose of noisy samples, rather than clean samples as the centroid. Therefore, we need to determine the selection criteria based on whether the current class centroid is obtained by clean samples. Especially, the criterion is based on the similarity between centroids of the current class and other classes. Assume that the sample in  $\\mathcal{G}_c^1$  is near the centroid and the sample in  $\\mathcal{G}_c^2$  is the opposite. We determined the choice of a more clean cluster  $\\mathcal{D}_c^{clean}$  based on the following criteria:",
+        "bbox": [
+            75,
+            247,
+            472,
+            565
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {D} _ {c} ^ {\\text {c l e a n}} = \\left\\{ \\begin{array}{l} \\mathcal {G} _ {c} ^ {2}, \\text {i f} | \\cos (o _ {c}, o _ {k}) - 1 | <   \\varepsilon \\text {a n d} \\left| \\mathcal {D} _ {c} ^ {H} \\right| <   \\left| \\mathcal {D} _ {k} ^ {H} \\right| \\\\ \\mathcal {G} _ {c} ^ {1}, \\text {o t h e r w i s e} \\end{array} \\right. \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            76,
+            570,
+            477,
+            625
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "It means that if the centroid of another class is similar to the centroid of the current class, whichever class of high-confidence sample set has fewer samples, then the high-confidence samples in this class are the noise samples. It is because the asymmetric noise ratio cannot exceed  $50\\%$  in the past assumption [29, 23]. Therefore, if the centroid is obtained from noise samples, we choose the cluster far away from the centroid as the cluster with more clean samples; otherwise, the cluster close to the centroid is selected.",
+        "bbox": [
+            75,
+            630,
+            468,
+            767
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Overall Training Process",
+        "text_level": 1,
+        "bbox": [
+            76,
+            773,
+            305,
+            789
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "After sample selection is conducted for each class, we adopt semi-supervised learning [1, 42] to train with all clean samples  $\\mathcal{D}^{clean}$  as labeled data and all noisy samples  $\\mathcal{D}^{noisy}$  as unlabeled data. The model is thus updated for the next round of training. One may also use long-tailed semi-supervised learning methods instead of normal semi-supervised learning in this stage.",
+        "bbox": [
+            75,
+            799,
+            470,
+            905
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            90,
+            633,
+            107
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            500,
+            114,
+            689,
+            131
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Datasets. In order to comprehensively evaluate the effectiveness of method, we perform experiments respectively in the scenarios with synthetic noise and realistic noise in intrinsically long-tailed distribution. We construct synthetic scenarios based on CIFAR-10 and CIFAR-100 [26]. We construct two benchmarks to imitate realistic scenarios, which is built on real-world noise datasets including Red Mini-ImageNet [20], CIFAR-10N and CIFAR-100N [48]. For all datasets, we adopt the standard paradigm of constructing a long-tailed distribution first and injecting noise later. We imitate realistic long-tailed distribution using the same setting in previous long-tailed learning works [3, 35] that long-tailed imbalance follows an exponential decay in sample sizes across different classes. The imbalance factor is denoted by the ratio between the size of the largest class and that of the smallest class.",
+        "bbox": [
+            496,
+            137,
+            890,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The details for dataset construction are as follows. (1) As for CIFAR-10/100 [26], we adopt symmetric and asymmetric noise to inject synthetic noise, which is commonly used in the area of label-noise learning [41]. The noise ratio is denoted by the ratio between the number of noisy samples and the total number of samples. It should be noted that different from the previous methods [21, 49] to deal with the joint problem of noisy labeled and long-tailed data, the noise transition matrix is randomly generated only related to the noise ratio. (2) As for CIFAR-10N/100N (10N/100N) [48], we replace sample labels with human-annotated noisy labels after constructing a long-tailed distribution based on real labels. (3) As for Red mini-[ImageNet (Red) [20], we inject web label noise samples after constructing a long-tailed distribution based on the original mini-[ImageNet. Therefore, the observed and intrinsic distribution are likely inconsistent, especially in long-tail distribution.",
+        "bbox": [
+            496,
+            378,
+            892,
+            636
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Compared Methods. We compare our method with the following three types of approaches: (1) Long-tail learning methods (LT). They are LA [33], LDAM [3] and IB [35]. (2) Label-noise learning methods (NL). They are DivideMix [29] and UNICON [23]; (3) Methods aiming at dealing with noisy label and long-tail distribution (NL-LT). They are MW-Net [41], RoLT [49], HAR [2] and ULC [19].",
+        "bbox": [
+            496,
+            638,
+            893,
+            744
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. We use PreAct ResNet18 [14] as backbone for CIFAR datasets and ResNet18 as backbone for Red mini-ImageNet dataset. Both of backbones adopt SGD with an initial learning rate of 0.02, a momentum of 0.9, and a weight decay of  $5 \\times 10^{-4}$  and a batch size of 64. For fair comparison, we train all methods for 100 epochs.",
+        "bbox": [
+            496,
+            747,
+            893,
+            838
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2. Comparative Results",
+        "text_level": 1,
+        "bbox": [
+            500,
+            847,
+            697,
+            863
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "CIFAR-10/100. Tab. 1 and 2 report the accuracy of different methods for intrinsically long-tailed CIFAR-10/100",
+        "bbox": [
+            500,
+            869,
+            892,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "1374",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/ea9d963df6985beed275a024776a2540336124c053b00ec4ff65e620cbe32b89.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Dataset</td><td colspan=\"2\">CIFAR-10</td><td colspan=\"2\">CIFAR-100</td></tr><tr><td colspan=\"2\">Imbalance Factor</td><td colspan=\"4\">0.1</td></tr><tr><td colspan=\"2\">Noise Ratio (Sym.)</td><td>0.4</td><td>0.6</td><td>0.4</td><td>0.6</td></tr><tr><td>Baseline</td><td>CE</td><td>71.67</td><td>61.16</td><td>34.53</td><td>23.63</td></tr><tr><td rowspan=\"3\">LT</td><td>LA</td><td>70.56</td><td>54.92</td><td>29.07</td><td>23.21</td></tr><tr><td>LDAM</td><td>70.53</td><td>61.97</td><td>31.30</td><td>23.13</td></tr><tr><td>IB</td><td>73.24</td><td>62.62</td><td>32.40</td><td>25.84</td></tr><tr><td rowspan=\"2\">NL</td><td>DivideMix</td><td>82.67</td><td>80.17</td><td>54.71</td><td>44.98</td></tr><tr><td>UNICON</td><td>84.25</td><td>82.29</td><td>52.34</td><td>45.87</td></tr><tr><td rowspan=\"4\">NL-LT</td><td>MW-Net</td><td>70.90</td><td>59.85</td><td>32.03</td><td>21.71</td></tr><tr><td>RoLT</td><td>81.62</td><td>76.58</td><td>42.95</td><td>32.59</td></tr><tr><td>HAR</td><td>77.44</td><td>63.75</td><td>38.17</td><td>26.09</td></tr><tr><td>ULC</td><td>84.46</td><td>83.25</td><td>54.91</td><td>44.66</td></tr><tr><td>Our</td><td>TABASCO</td><td>85.53</td><td>84.83</td><td>56.52</td><td>45.98</td></tr></table>",
+        "bbox": [
+            81,
+            88,
+            465,
+            332
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "with symmetric and asymmetric noise, respectively. It can be observed three phenomena as follows: (1) The performance of existing long-tail methods is lower than baseline in most situations. It is because such methods do not have the ability to distinguish the noise samples, and the wrong attention to the noise samples leads to the further decline of the model generalization ability. (2) The advantages of existing methods aiming at both long-tailed distribution and label noise are not obvious compared with single methods. It is because existing methods do not take into account the distribution inconsistency caused by noise labels. (3) Our method significantly improved over other methods in all cases. Moreover, as the proportion of noise increases, the accuracy of our method decreases less compared to other methods. It validates that our method can effectively identify the noise in long-tail distribution.",
+        "bbox": [
+            75,
+            382,
+            468,
+            625
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Red mini-ImageNet and CIFAR-10N/100N. Tab. 4 reports the accuracy of different methods on real-world noise datasets with intrinsically long-tailed distribution. According to the results of label-noise learning methods on CIFAR-10N/100N, it can be observed that real-world noise is more difficult to separate than synthetic noise to some extent. It further increases the difficulty of dealing with such problem. Nevertheless, our method still performs well, which again confirms its effectiveness.",
+        "bbox": [
+            75,
+            627,
+            468,
+            763
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3. Ablation Study and Discussions",
+        "text_level": 1,
+        "bbox": [
+            76,
+            768,
+            356,
+            785
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In the following experiments, we analyze each component of the proposed TABASCO on CIFAR-10/100 to verify its effectiveness. We use an imbalance factor of 0.1 and a noise ratio of 0.4 to construct the synthetic noise dataset.",
+        "bbox": [
+            75,
+            792,
+            467,
+            852
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Effectiveness of WJSD. To validate the effectiveness of the proposed WJSD, we compare the accuracy of models trained by different separation metrics. Specifically, we",
+        "bbox": [
+            75,
+            854,
+            468,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/b215880e41e8da7af173718bfa3d28b46dd8ef80e9389bc9e30c3a9930ac3088.jpg",
+        "table_caption": [
+            "Table 1: Performance comparison under symmetric noise. The best results are shown in bold."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Dataset</td><td colspan=\"2\">CIFAR-10</td><td colspan=\"2\">CIFAR-100</td></tr><tr><td colspan=\"2\">Imbalance Factor</td><td colspan=\"4\">0.1</td></tr><tr><td colspan=\"2\">Noise Ratio (Asym.)</td><td>0.2</td><td>0.4</td><td>0.2</td><td>0.4</td></tr><tr><td>Baseline</td><td>CE</td><td>79.90</td><td>62.88</td><td>44.45</td><td>32.05</td></tr><tr><td rowspan=\"3\">LT</td><td>LA</td><td>71.49</td><td>59.88</td><td>39.34</td><td>28.49</td></tr><tr><td>LDAM</td><td>74.58</td><td>62.29</td><td>40.06</td><td>33.26</td></tr><tr><td>IB</td><td>73.49</td><td>58.36</td><td>45.02</td><td>35.25</td></tr><tr><td rowspan=\"2\">NL</td><td>DivideMix</td><td>80.92</td><td>69.35</td><td>58.09</td><td>41.99</td></tr><tr><td>UNICON</td><td>72.81</td><td>69.04</td><td>55.99</td><td>44.70</td></tr><tr><td rowspan=\"4\">NL-LT</td><td>MW-Net</td><td>79.34</td><td>65.49</td><td>42.52</td><td>30.42</td></tr><tr><td>RoLT</td><td>73.30</td><td>58.29</td><td>48.19</td><td>39.32</td></tr><tr><td>HAR</td><td>82.85</td><td>69.19</td><td>48.50</td><td>33.20</td></tr><tr><td>ULC</td><td>74.07</td><td>73.19</td><td>54.45</td><td>43.20</td></tr><tr><td>Our</td><td>TABASCO</td><td>82.10</td><td>80.57</td><td>59.39</td><td>50.51</td></tr></table>",
+        "bbox": [
+            504,
+            88,
+            887,
+            332
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/8c09be3167e27219cb9986d82e33b13e0a10393aac6ad71bc39d04361880e1ce.jpg",
+        "table_caption": [
+            "Table 2: Performance comparison under asymmetric noise. The best results are shown in bold."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Separation metric</td><td>Accuracy</td></tr><tr><td rowspan=\"2\">CIFAR-10</td><td>JSD</td><td>83.97</td></tr><tr><td>WJSD</td><td>85.57</td></tr><tr><td rowspan=\"2\">CIFAR-100</td><td>JSD</td><td>55.80</td></tr><tr><td>WJSD</td><td>56.67</td></tr></table>",
+        "bbox": [
+            553,
+            388,
+            836,
+            481
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Performance comparison between JSD and WJSD.",
+        "bbox": [
+            500,
+            491,
+            890,
+            507
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "train the model by using JSD and WJSD for sample separation with a symmetric noise dataset, respectively. As for each separation metric, we simply use the small value strategy [29, 23] for sample selection. As Tab. 3 shows, the model trained by WJSD achieves better results on both symmetric noise datasets compared with JSD owing to the stronger separability of WJSD.",
+        "bbox": [
+            496,
+            513,
+            890,
+            621
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Effectiveness of ACD. To validate the advantage of the proposed ACD, we compare it with the method which calculates the feature centroid on the observed class (hereinafter referred to as centroid distance or CD) to show their separability in CIFAR-10. Specifically, based on the same backbone, we calculate the distance from the sample to the centroid using ACD and CD in CIFAR-10 with asymmetric noise, respectively. Fig. 4 shows the distributions of samples in the tail class by CD and ACD. The purity of the tail class is 0.53. It can be observed that the CD values of noise and clean samples are highly overlapped, which leads to failure sample separation. It is because the purity of the tail class may be quite low due to the negative effects of both label-noise and long-tailed distribution. The centroid obtained by ACD has a higher purity, so there is a significant difference in ACD between them. As shown in Fig. 4(b), most clean samples are concentrated in the interval of [0.8,1], which is easily clustered by GMM.",
+        "bbox": [
+            496,
+            628,
+            890,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "1375",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/9ee4f8e0776b34ab050d7b71fb326a4b046e28174c07d9c52c5c21b18b98e4da.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Dataset</td><td colspan=\"2\">Red</td><td>10N</td><td>100N</td></tr><tr><td colspan=\"2\">Imbalance Factor</td><td colspan=\"2\">≈ 0.1</td><td colspan=\"2\">0.1</td></tr><tr><td colspan=\"2\">Noise Ratio</td><td>0.2</td><td>0.4</td><td colspan=\"2\">≈ 0.4</td></tr><tr><td>Baseline</td><td>CE</td><td>40.42</td><td>31.46</td><td>60.44</td><td>38.10</td></tr><tr><td rowspan=\"3\">LT</td><td>LA</td><td>26.82</td><td>25.88</td><td>65.74</td><td>36.50</td></tr><tr><td>LDAM</td><td>26.64</td><td>23.46</td><td>62.50</td><td>38.48</td></tr><tr><td>IB</td><td>23.80</td><td>22.08</td><td>65.91</td><td>42.48</td></tr><tr><td rowspan=\"2\">NL</td><td>DivideMix</td><td>48.76</td><td>48.96</td><td>67.85</td><td>44.25</td></tr><tr><td>UNICON</td><td>40.18</td><td>41.64</td><td>69.54</td><td>51.93</td></tr><tr><td rowspan=\"4\">NL-LT</td><td>MW-Net</td><td>42.66</td><td>40.26</td><td>69.73</td><td>44.20</td></tr><tr><td>RoLT</td><td>22.56</td><td>24.22</td><td>75.24</td><td>46.61</td></tr><tr><td>HAR</td><td>46.61</td><td>38.71</td><td>74.97</td><td>44.54</td></tr><tr><td>ULC</td><td>48.12</td><td>47.06</td><td>75.71</td><td>51.72</td></tr><tr><td>Our</td><td>TABASCO</td><td>50.20</td><td>49.68</td><td>80.61</td><td>53.83</td></tr></table>",
+        "bbox": [
+            81,
+            88,
+            467,
+            333
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Performance comparison with real-world noise and long-tail distribution. The best results are shown in bold.",
+        "bbox": [
+            76,
+            343,
+            468,
+            375
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effectiveness of Bi-dimensional Sample Separation. To validate how the proposed bi-dimensional metrics complement each other, we plot the values of bi-dimensional metrics of both clean and noisy samples under different noise types in Fig. 5. For the case of symmetric noise shown in Fig. 5(a), it can be observed that ACD cannot effectively distinguish clean samples from noisy samples in the tail class. In this case, WJSD shows its advantage because the values of WJSD for most noisy samples are clustered in the top region, and the clean samples are scattered throughout the rest region. For the case of asymmetric noise shown in Fig. 5(b), it can be observed that WJSD cannot distinguish clean samples from noisy samples well for the tail class, while ACD can distinguish them well although the clean and noisy samples are closer in the tail class. The experimental observation is consistent with the previous discussion, which validates the complementarity of bi-dimensional metrics.",
+        "bbox": [
+            75,
+            383,
+            468,
+            655
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effectiveness of Sample Selection. In order to validate the effectiveness of the proposed sample selection in dimension selection and cluster selection, we compare the optimal clean ratio of clusters obtained by WJSD, ACD and our selection method in CIFAR-10 with different noise types. The clean ratio of a cluster is calculated by the proportion of clean samples in the cluster to the total number of cluster samples. As shown in Fig. 6, on the tail class, the clean ratio of the cluster obtained by our selection method is aligned with the higher one between WJSD and ACD for both noise cases. It shows that we can choose the most appropriate separation dimension in different situations. Besides, the clean ratio of the cluster we selected is consistent with that of the optimal cluster in the corresponding dimension, which is sufficient to show that the method we proposed can select clean samples.",
+        "bbox": [
+            75,
+            657,
+            470,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/8978744f5c52019e23a6df9ddbcefd4b3d31544d821d855b90bb486837f20e8a.jpg",
+        "image_caption": [
+            "Figure 4: Comparison of the sample distributions between CD and ACD in the tail class."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            89,
+            684,
+            196
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/2ab50acdf51b02f9a848c4538c99922c871944a12dfcdf0eedd93ee4d564c139.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            691,
+            90,
+            870,
+            196
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6a8bf68f9db9b8598d1392047cd47c220a1be42a8cd501b7dec94e1c1f3ce57f.jpg",
+        "image_caption": [
+            "(a) Symmetric noise"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            511,
+            236,
+            687,
+            337
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/87cf4887cb473a40cbc410da6221f0e73f402470ce88fd0ef33b08ab0eaab4d3.jpg",
+        "image_caption": [
+            "(b) Asymmetric noise"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            696,
+            237,
+            867,
+            337
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/df444e84913d8f40b83c37295349c5d42fb46363247f2eca2f479539d74cc1f3.jpg",
+        "image_caption": [
+            "Figure 5: Scatter plot of the values of the proposed bidimensional metrics with (a) symmetric noise and (b) asymmetric noise.",
+            "(a) Symmetric noise",
+            "Figure 6: The performance of the proposed sample selection with (a) symmetric noise and (b) asymmetric noise."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            412,
+            684,
+            517
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/34026ab00885b70d2d07597d1d8b93c9ba16dae610525ee4a0436ad8f17d4e94.jpg",
+        "image_caption": [
+            "(b) Asymmetric noise"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            691,
+            412,
+            870,
+            517
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            583,
+            617,
+            598
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This paper studies a more general and realistic problem of label-noise learning with intrinsically long-tailed data. The major challenge in this problem is that it is hard to distinguish clean samples from noisy samples on intrinsic tail classes. Accordingly, we propose a learning framework TABASCO for this problem. In TABASCO, two new metrics are explicitly proposed to address the problem of sample selection in tail classes. Extensive experiments on noisy-labeled datasets with long-tail distribution demonstrate its effectiveness.",
+        "bbox": [
+            496,
+            607,
+            890,
+            758
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements. We thank Dr. Shiquan Ruan for his discussion on the theory proof. This study was supported in part by the National Natural Science Foundation of China under Grants 62002302 and U21A20514; in part by the FuXiaQuan National Independent Innovation Demonstration Zone Collaborative Innovation Platform under Grant 3502ZCQXT2022008; in part by NSFC/RGC Joint Research Scheme under Grant N_HKBU214/21; and in part by the General Research Fund of RGC under Grants 12201321 and 12202622.",
+        "bbox": [
+            496,
+            760,
+            890,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "1376",
+        "bbox": [
+            483,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            173,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, 2019. 6",
+            "[2] Kaidi Cao, Yining Chen, Junwei Lu, Nikos Aréchiga, Adrien Gaidon, and Tengyu Ma. Heteroskedastic and imbalanced deep learning with adaptive regularization. In ICLR, 2021. 2, 6",
+            "[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. In NeurIPS, 2019. 2, 3, 6",
+            "[4] Nitesh V. Chawla, Kevin W. Bowyer, Lawrence O. Hall, and W. Philip Kegelmeyer. SMOTE: synthetic minority oversampling technique. J. Artif. Intell. Res., 2002. 2",
+            "[5] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: Rebalanced mixup. In ECCV, 2020. 2",
+            "[6] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge J. Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 2, 3",
+            "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 1",
+            "[8] Andrew Estabrooks, Taeho Jo, and Nathalie Japkowicz. A multiple resampling method for learning from imbalanced data sets. Comput. Intell., 2004. 2",
+            "[9] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 1",
+            "[10] Bo Han, Gang Niu, Xingrui Yu, Quanming Yao, Miao Xu, Ivor W. Tsang, and Masashi Sugiyama. SIGUA: forgetting may make learning with noisy labels more robust. In ICML, 2020. 2",
+            "[11] Bo Han, Quanming Yao, Tongliang Liu, Gang Niu, Ivor W. Tsang, James T. Kwok, and Masashi Sugiyama. A survey of label-noise representation learning: Past, present and future. CoRR, 2020. 1",
+            "[12] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor W. Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, 2018. 1, 2",
+            "[13] Hui Han, Wenyuan Wang, and Binghamuan Mao. Borderlinesmote: A new over-sampling method in imbalanced data sets learning. In ICIC 2005, 2005. 2",
+            "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6",
+            "[15] Yin-Yin He, Jianxin Wu, and Xiu-Shen Wei. Distilling virtual examples for long-tailed recognition. In ICCV, 2021. 2",
+            "[16] Dan Hendrycks, Mantas Mazeika, Duncan Wilson, and Kevin Gimpel. Using trusted data to train deep networks on labels corrupted by severe noise. In NeurIPS, 2018. 2",
+            "[17] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Per"
+        ],
+        "bbox": [
+            78,
+            114,
+            468,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "ona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 1",
+            "[18] Grant Van Horn and Pietro Perona. The devil is in the tails: Fine-grained classification in the wild. CoRR, 2017. 1",
+            "[19] Yingsong Huang, Bing Bai, Shengwei Zhao, Kun Bai, and Fei Wang. Uncertainty-aware learning against label noise on imbalanced datasets. In AAAI, 2022. 2, 6",
+            "[20] Lu Jiang, Di Huang, Mason Liu, and Weilong Yang. Beyond synthetic noise: Deep learning on controlled noisy labels. In ICML, 2020. 6",
+            "[21] Shenwang Jiang, Jianan Li, Ying Wang, Bo Huang, Zhang Zhang, and Tingfa Xu. Delving into sample loss curve to embrace noisy and imbalanced data. In AAAI, 2022. 2, 6",
+            "[22] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In ICLR, 2020. 2",
+            "[23] Nazmul Karim, Mamshad Nayeem Rizve, Nazanin Rahnavard, Ajmal Mian, and Mubarak Shah. Unicon: Combating label noise through uniform selection and contrastive learning. In CVPR, 2022. 1, 2, 3, 6, 7",
+            "[24] Shyamgopal Karthik, Jérôme Revaud, and Chidlovskii Boris. Learning from long-tailed data with noisy labels. CoRR, 2021. 2",
+            "[25] Jaehyung Kim, Youngbum Hur, Sejun Park, Eunho Yang, Sung Ju Hwang, and Jinwoo Shin. Distribution aligning refinery of pseudo-label for imbalanced semi-supervised learning. In NeurIPS, 2020. 2",
+            "[26] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 6",
+            "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 1",
+            "[28] Hyuck Lee, Seungjae Shin, and Heeyoung Kim. ABC: auxiliary balanced classifier for class-imbalanced semi-supervised learning. In NeurIPS, 2021. 2",
+            "[29] Junnan Li, Richard Socher, and Steven C. H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In ICLR, 2020. 1, 2, 3, 6, 7",
+            "[30] Shikun Li, Xiaobo Xia, Shiming Ge, and Tongliang Liu. Selective-supervised contrastive learning with noisy labels. In CVPR, 2022. 2",
+            "[31] Xu-Ying Liu, Jianxin Wu, and Zhi-Hua Zhou. Exploratory undersampling for class-imbalance learning. IEEE Trans. Syst. Man Cybern. Part B, 2009. 2",
+            "[32] Michal Lukasik, Srinadh Bhojanapalli, Aditya Krishna Menon, and Sanjiv Kumar. Does label smoothing mitigate label noise? In ICML, 2020. 2",
+            "[33] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In ICLR, 2021. 6",
+            "[34] Diego Ortego, Eric Arazo, Paul Albert, Noel E. O'Connor, and Kevin McGuinness. Multi-objective interpolation training for robustness to label noise. In CVPR, 2021. 2",
+            "[35] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In ICCV, 2021. 6"
+        ],
+        "bbox": [
+            503,
+            93,
+            888,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "1377",
+        "bbox": [
+            483,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[36] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In CVPR, 2017. 2",
+            "[37] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey E. Hinton. Regularizing neural networks by penalizing confident output distributions. In ICLR, 2017. 2",
+            "[38] William J Reed. The pareto, zipf and other power laws. Economics Letters, 74(1):15-19, 2001. 1",
+            "[39] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In NeurIPS, 2020. 2",
+            "[40] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1",
+            "[41] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In NeurIPS, 2019. 2, 6",
+            "[42] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 6",
+            "[43] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. SELFIE: refurbishing unclean samples for robust deep learning. In ICML, 2019. 1",
+            "[44] Hwanjun Song, Minseok Kim, Dongmin Park, and Jae-Gil Lee. Learning from noisy labels with deep neural networks: A survey. CoRR, 2020. 1",
+            "[45] Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In CVPR, 2020. 2, 3",
+            "[46] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X. Yu. Long-tailed recognition by routing diverse distribution-aware experts. In ICLR, 2021. 2",
+            "[47] Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan L. Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In CVPR, 2021. 2",
+            "[48] Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. Learning with noisy labels revisited: A study using real-world human annotations. In ICLR, 2022. 6",
+            "[49] Tong Wei, Jiang-Xin Shi, Wei-Wei Tu, and Yu-Feng Li. Robust long-tailed learning under label noise. CoRR, 2021. 2, 4, 6, 14",
+            "[50] Xiaobo Xia, Tongliang Liu, Bo Han, Chen Gong, Nannan Wang, Zongyuan Ge, and Yi Chang. Robust early-learning: Hinding the memorization of noisy labels. In ICLR, 2021. 2",
+            "[51] Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Jun Yu, Gang Niu, and Masashi Sugiyama. Sample selection with uncertainty of losses for learning with noisy labels. In ICLR, 2022. 2"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[52] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In CVPR, 2015. 1",
+            "[53] Yi Xuanyu, Tang Kaihua, Hua Xian-Sheng, Lim Joo-Hwee, and Zhang Hanwang. Identifying hard noise in long-tailed sample distribution. In ECCV, 2022. 2",
+            "[54] Yu Yao, Tongliang Liu, Bo Han, Mingming Gong, Jiankang Deng, Gang Niu, and Masashi Sugiyama. Dual T: reducing estimation error for transition matrix in label-noise learning. In NeurIPS, 2020. 2",
+            "[55] Yazhou Yao, Zeren Sun, Chuanyi Zhang, Fumin Shen, Qi Wu, Jian Zhang, and Zhenmin Tang. Jo-src: A contrastive approach for combating noisy labels. In CVPR, 2021. 2, 3",
+            "[56] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In CVPR, 2019. 2",
+            "[57] Yuhang Zang, Chen Huang, and Chen Change Loy. FASA: feature augmentation and sampling adaptation for long-tailed instance segmentation. In ICCV, 2021. 2",
+            "[58] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 2",
+            "[59] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In CVPR, 2021. 2",
+            "[60] Yongshun Zhang, Xiu-Shen Wei, Boyan Zhou, and Jianxin Wu. Bag of tricks for long-tailed visual recognition with deep convolutional neural networks. In AAAI, 2021. 1",
+            "[61] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2",
+            "[62] Beier Zhu, Yulei Niu, Xian-Sheng Hua, and Hanwang Zhang. Cross-domain empirical risk minimization for unbiased long-tailed classification. In AAAI, 2022. 1"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            575
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "1378",
+        "bbox": [
+            483,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2023/Label-Noise Learning with Intrinsically Long-Tailed Data/ca2911cd-1897-45f1-bc9c-ab89b815151c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3ef382d1e47fe886472c0473e75814315fc5bb063e272c2f13775b89a87382b
+size 550494

2023/Label-Noise Learning with Intrinsically Long-Tailed Data/full.md

@@ -0,0 +1,339 @@
+# Label-Noise Learning with Intrinsically Long-Tailed Data
+
+Yang Lu $^{1,2*}$  Yiliang Zhang $^{1,2}$  Bo Han $^{3}$  Yiu-ming Cheung $^{3}$  Hanzi Wang $^{1,2}$
+
+$^{1}$ Fujian Key Laboratory of Sensing and Computing for Smart City, School of Informatics, Xiamen University, Xiamen, China
+
+$^{2}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing,
+
+Ministry of Education of China, Xiamen University, Xiamen, China
+
+$^{3}$ Department of Computer Science, Hong Kong Baptist University, Hong Kong, China
+
+luyang@xmu.edu.cn ylzhangcs@hotmail.com {bhanml, ymc}@comp.hkbu.edu.hk hanzi.wang@xmu.edu.cn
+
+# Abstract
+
+Label noise is one of the key factors that lead to the poor generalization of deep learning models. Existing label-noise learning methods usually assume that the ground-truth classes of the training data are balanced. However, the real-world data is often imbalanced, leading to the inconsistency between observed and intrinsic class distribution with label noises. In this case, it is hard to distinguish clean samples from noisy samples on the intrinsic tail classes with the unknown intrinsic class distribution. In this paper, we propose a learning framework for label-noise learning with intrinsically long-tailed data. Specifically, we propose two-stage bi-dimensional sample selection (TABASCO) to better separate clean samples from noisy samples, especially for the tail classes. TABASCO consists of two new separation metrics that complement each other to compensate for the limitation of using a single metric in sample separation. Extensive experiments on benchmarks demonstrate the effectiveness of our method. Our code is available at https://github.com/Wakings/TABASCO.
+
+# 1. Introduction
+
+Under the support of a large amount of high-quality labeled data, deep neural networks have achieved great success in various fields [27, 40, 7]. However, it is expensive and difficult to obtain a large amount of high-quality labeled data in many practical applications. Instead, the commonly used large-scale training data is usually obtained from the Internet or crowdsourcing platforms like Amazon Mechanical Turk, which is unreliable and may be mislabeled [52, 43]. The models trained on this kind of unreliable data, called noisy-labeled data, often produce poor generalization performance because deep neural networks tend to overfit noisy samples due to their large model capacity. In
+
+![](images/22e148d9b056a7487ba405eda466053370f0d945b7ff477a2b438a49a21ec20e.jpg)  
+(a) Class distribution
+
+![](images/348c53ee3757a4e47bad2047be5b399853813846bb127ad12b77972385732830.jpg)  
+(b) Loss distribution  
+Figure 1: (a) An example of observed class distribution with noisy labels with long-tailed intrinsic class distribution. (b) The training loss of each sample under noisy-labeled and long-tailed data.
+
+the literature, there are some works to obtain a robust model trained on noisy-labeled data [11, 44]. Among them, the most straightforward and effective way is to differentiate between clean and noisy samples based on their differences in specific metrics, such as training loss [12, 29, 23].
+
+These label-noise learning methods generally assume that the intrinsic class distribution of the training data is balanced, where each class has almost the same number of samples in terms of their unknown ground-truth labels. However, the data in real-world applications are often imbalanced, e.g., the LVIS dataset [9] and the iNaturalist dataset [17]. The class imbalance usually exhibits in the form of a long-tail distribution, where a small portion of classes possess a large number of samples, and the other classes possess a small number of samples only [38, 18]. In this case, the model training tends to the head classes and ignores the tail classes [60, 62]. When both noisy labels and long-tail distribution exist, training a robust model is even more challenging. There are two key challenges in this scenario. (1) Distribution inconsistency: The observed and intrinsic distributions are likely inconsistent due to noise labels, making the model more difficult to discover and focus on the intrinsic tail classes. As illustrated in Fig. 1(a), the intrinsic class distribution of the dataset is long-tailed, while the existence of noisy labels makes the distribution more
+
+balanced. The intrinsic tail classes, e.g., classes 6 and 9, are occupied by a large number of noisy data, making them no longer tail classes by observation. (2) Tail inseparability: Even if the tail class is identified, it is more difficult than ever to distinguish between clean and noisy samples in the tail class because clean samples are overwhelmed by noises that make their values of the separation metric highly similar. As illustrated in Fig. 1(b), the training loss of clean and noisy samples in the tail class are generally inseparable compared with the ones in the other classes. Several preliminary works have studied the joint problem of label noise and long-tail distribution [21, 49, 24, 2]. These methods implicitly reduce the complexity of the problem by assuming a similar noise rate for each class. However, this assumption is too strong to apply because noisy samples from the head classes may be the majority, resulting in a higher noise rate in the tail class than in other classes.
+
+In this paper, we propose a Two-stAge Bi-dimensionAl Sample seleCtiOn (TABASCO) strategy to address the problem of label-noise learning with intrinsically long-tailed data. In the first stage, we propose to use two new separation metrics for sample separation, i.e., weighted Jensen-Shannon divergence (WJSD) and adaptive centroid distance (ACD), which work corporately to separate clean samples from noisy samples in the tail classes. The proposed metrics are complementary, where WJSD separates the samples from the output perspective while ACD does that from the feature perspective. In the second stage, we determine the separation dimension with better separability for each class and perform sample selection. In order to evaluate the method uniformly and effectively, we introduce two benchmarks with real-world noise and intrinsically long-tailed distribution. The main contributions of our work can be summarized as follows:
+
+- We present a more general problem of label-noise learning with intrinsically long-tailed data. The key challenges in this problem are distribution inconsistency and tail inseparability.  
+- We propose an effective solution called TABASCO. With the help of two new separation metrics, it is able to effectively identify and select clean samples of the intrinsic tail class.  
+- We introduce two benchmarks with real-world noise and intrinsically long-tailed distribution. Extensive experiments on them show the effectiveness of our method and the limitations of existing methods.
+
+# 2. Related Work
+
+# 2.1. Label-Noise Learning
+
+A straightforward strategy to deal with noisy data is to reduce the proportion of noise in training samples by sepa
+
+rating noisy samples from clean samples. Methods such as co-teaching [12] and DivideMix [29] adopt the small loss trick, while Jo-SRC [55] and UNICON [23] use Jensen-Shannon divergence instead for sample selection. In contrast to methods that separate at the label level, some methods [30, 34] attempt to separate samples at the feature space. There are also some methods to avoid over-fitting the model to noisy data by imposing regularization constraints on model parameters [50, 10] or labels [37, 58, 32]. Other methods mitigate the influence of noisy data by adjusting the loss functions, such as backward and forward loss correction [36], gold loss correction [16], MW-Net [41] and Dual-T [54].
+
+# 2.2. Long-tail Learning
+
+Re-balancing the data for long-tail distributions is a classical strategy to solve the problem of long-tail learning, such as re-sampling [4, 31, 8, 13] and data augmentation [61, 57, 5]. In addition, there are methods to improve the model generalization by introducing long-tail robust loss functions [6, 3, 45, 39]. Methods such as FTL [56], RIDE [46] and DiVE [15] try to solve the problem by using the idea of transfer learning. Recently, approaches based on decoupling [22, 61, 59] split end-to-end learning into feature learning and classifier retraining such that the obtained feature extractor is less affected by the long-tail distribution. In contrast to the above methods with the supervised learning paradigm, CReST [47], ABC [28] and DARP [25] attempt to solve the long-tail problem in the manner of semi-supervised learning.
+
+# 2.3. Label-Noise Learning on Long-tailed Data
+
+Research on this joint problem has just been explored. CNLCU [51] relaxes the constraint of the small loss trick by regarding a portion of large loss samples as clean samples to reduce the probability of misclassifying clean samples to noisy samples in the tail class. RoLT [49] uses the distance from the samples to the centroid of the current class instead of the training loss for sample selection. HAR [2] uses an adaptive approach to regularize noise and tail class samples. Karthik et al. [24] uses the idea of decoupling to fine-tune the loss function for better robustness after feature learning by the self-supervised method. ULC [19] introduces uncertainty to enhance the separation ability of noise samples. H2E [53] reduces hard noises to easy ones by learning a classifier as noise identifier invariant to the class and context distributional changes.
+
+# 3. Problem Definition
+
+Given a training set  $\mathcal{D} = \{x_i, \hat{y}_i\}_{i=1}^N$ , where  $\hat{y}_i \in [M]$  is the observed label of the sample  $x_i$ .  $N$  is the number of training samples, and  $M$  is the number of classes. In our problem,  $\mathcal{D}$  has the following properties:
+
+![](images/e2fb5fbcd490fc645823e0de49b016a33494bce676d8b24cf9f2e12b8ee9f4bb.jpg)  
+Figure 2: The proposed framework for label-noise learning with intrinsically long-tailed data.
+
+- Noisy-labeled. There is a subset of samples  $\hat{\mathcal{D}} \in \mathcal{D}$ , where a sample  $\{x, \hat{y}\} \in \hat{\mathcal{D}}$  has an unknown ground-truth label  $y$  different from its observed label  $\hat{y}$ .  
+- Long-tailed. The ground-truth class distribution is long-tailed. Supposing  $n_c$  is the number of the samples in class  $c$ , we have  $n_1 > n_2 > \dots > n_M$ .
+
+We call this kind of training data intrinsically long-tailed because the observed class distribution may not be long-tailed due to the existence of noisy labels. Therefore, our goal is to learn from noisy-labeled data with the intrinsically long-tailed class distribution. The problem is challenging from two aspects. On the one hand, directly applying long-tail learning methods [6, 3, 45] is infeasible because the observed class distribution may be inconsistent with the intrinsic class distribution. On the other hand, label-noise learning methods perform poorly in the tail classes because they usually contain more noisy samples due to insufficient data, as shown in Fig. 1(a). It results in high noise rates in the tail classes, which brings great challenges to separating clean samples and noisy samples for further training. Under this circumstance, the existing sample selection methods [29, 23] often fail to select clean samples in the tail class. The main reason is that it is difficult to distinguish the tail class data from the noisy label data by existing metrics like cross-entropy loss, as shown in Fig. 1(b).
+
+# 4. Proposed Method
+
+In this paper, we propose a two-stage bi-dimensional sample selection (TABASCO) to address the problem of label-noise learning with intrinsically long-tailed data. TABASCO decouples the sample selection process into two stages: (1) sample separation, and (2) dimension and cluster selection. First, given an initial model  $\theta$  trained on the original training data  $\mathcal{D}$ , we propose to calculate bi-dimension metrics for each sample in each observed class based on the outputs and features of the model. Second, we determine the separation dimension with better separability for each class, and we adopt the corresponding selection strategy to
+
+select the cluster with more clean samples based on the selected metric. Last, we adopt semi-supervised learning to update the model by regarding the selected clean cluster as labeled data. The overall framework is shown in Fig. 2.
+
+# 4.1. Bi-Dimensional Separation Metrics
+
+Due to the deficiency of a single separation metric to distinguish clean samples from noisy samples in the complex situation of noisy labels with intrinsic long-tail distribution, we propose to jointly use two metrics from different perspectives: weighted Jensen-Shannon divergence (WJSD) and adaptive centroid distance (ACD). Both metrics are specifically designed for the joint problem and are complementary to cover the case when only one of them cannot separate clean samples from noisy samples well. WJSD fully utilizes the information of prediction confidence, while ACD relies on the distance in the feature space. Thus, using clustering on samples according to the values of their bi-dimensional metrics has the flexibility to separate samples with different imbalance ratios, noise ratios, and noise types.
+
+We first reduce the separation granularity to alleviate the interference from the head class to the tail class. Specifically, the separation is performed within each observed class according to the proposed bi-dimensional metrics. We first divide training set  $\mathcal{D}$  into subsets according to the observed labels  $\mathcal{D}_c = \{(x,\hat{y})\mid \hat{y} = c\}$ . Given a sample  $(x_{i},\hat{y}_{i})\in \mathcal{D}_{c}$ , the model predictions obtained from the model  $\theta$  is denoted as  $\mathbf{p}_i = [p_i^1,p_i^2,\dots,p_i^M ]$ , where  $p_i^j$  is the  $j$ 's dimension of vector  $\mathbf{p}_i$ . The one-hot representation of the observed label  $\hat{y}_i$  is denoted as  $\hat{\mathbf{y}}_i$ .
+
+Weighted Jensen-Shannon Divergence The Jensen-Shannon Divergence (JSD) is a commonly used metric to separate samples by assessing the variability of model predictions [55, 23]. It is defined as:
+
+$$
+J S D \left(x _ {i}\right) = \frac {1}{2} K L \left(\mathbf {p} _ {i} \left\| \frac {\mathbf {p} _ {i} + \hat {\mathbf {y}} _ {i}}{2}\right) + \frac {1}{2} K L \left(\hat {\mathbf {y}} _ {i} \left\| \frac {\mathbf {p} _ {i} + \hat {\mathbf {y}} _ {i}}{2}\right), \right. \right. \tag {1}
+$$
+
+where  $KL(\cdot \|\cdot)$  is the Kullback-Leibler divergence. When the intrinsic distribution is balanced, the JSD values of clean samples are generally lower than the noisy samples, so the separation can be easily modeled. However, when the distribution is intrinsically long-tailed, the noisy and clean samples in a tail class tend to obtain similar prediction confidence on the observed class  $c$ , because their predictions are both towards the head classes. In this case, using JSD as a separation metric fails to separate them. The reason is given by Theorem 1, which shows the upper bound of the absolute value of the JSD difference between two samples.
+
+Theorem 1 (Upper bound of JSD difference) Suppose  $x_{i}$  and  $x_{j}$  are two samples in class  $c$ ,  $p_{i}^{c}$  and  $q_{i}^{c}$  are the  $c$ 's dimension of their prediction confidence  $\mathbf{p}_{i} = [p_{i}^{1}, p_{i}^{2}, \dots, p_{i}^{M}]$  and  $\mathbf{p}_{j} = [p_{j}^{1}, p_{j}^{2}, \dots, p_{j}^{M}]$ , respectively. The upper bound of the absolute value of the difference between their JSD values is given by:
+
+$$
+| J S D (x _ {i}) - J S D (x _ {j}) | \leq \frac {1}{2} \log \left(\frac {p _ {i} ^ {c} + 1}{p _ {i} ^ {c}}\right) | p _ {i} ^ {c} - p _ {j} ^ {c} |.
+$$
+
+Theorem 1 shows that the difference between values of JSD is only determined by the prediction confidence of the observed class, i.e.,  $p_i^c$  and  $p_j^c$ . In addition, as  $|p_i^c - p_j^c|$  gets smaller with a fixed value of  $p_i^c$ ,  $|JSD(x_i) - JSD(x_j)|$  is also smaller. This indicates that when two samples are in the same class  $c$  with close values of  $p_i^c$ , their JSD values are also close, which makes it difficult to separate them.
+
+Nevertheless, we can utilize the prediction confidence on other classes, i.e.,  $p_i^d$  for  $d \neq c$ , to distinguish two samples even if their values of  $p_i^c$  are close. Specifically, we propose WJSD by imposing an additional weight on JSD to further distinguish samples by inspecting their prediction confidence on other classes rather than the observed class. First, we take the maximum prediction confidence  $\max(\mathbf{p}_i)$  into account because it may reflect the confidence of a noisy sample's ground-truth class  $y_i$ . Then, we calculate the ratio of the maximum prediction confidence to the prediction confidence of the observed class  $\max(\mathbf{p}_i) / p_i^c$ . In this manner, the additional weight is greater than one only if the class with the maximum prediction confidence is not the observed class. The larger gap between  $\max(\mathbf{p}_i)$  and  $p_i^c$  makes the weight higher. To avoid exceptionally large weights by the division during normalization over all samples in class  $c$  for later clustering, we set the upper bound according to the averaged prediction confidence over all samples in class  $c$ . Finally, the additional weight can be calculated as follows:
+
+$$
+W \left(x _ {i}\right) = \min  \left(\max  \left(\mathbf {p} _ {i}\right) / p _ {i} ^ {c}, \max  \left(\bar {\mathbf {p}} _ {c}\right) / \bar {p} _ {c} ^ {c}\right), \tag {2}
+$$
+
+where  $\bar{\mathbf{p}}_c = [\bar{p}_c^1,\bar{p}_c^2,\dots,\bar{p}_c^M] = \frac{1}{|\mathcal{D}_c|}\sum_{i = 1}^{|\mathcal{D}_c|}\mathbf{p}_i$  is the average prediction confidence of class  $c$ . Thus, WJSD can be calculated by:
+
+$$
+W J S D \left(x _ {i}\right) = W \left(x _ {i}\right) \times J S D \left(x _ {i}\right). \tag {3}
+$$
+
+Remarks: Compared with the value of JSD that is only related to the prediction confidence on the observed class, adopting the maximum prediction confidence in WJSD can better separate clean samples from noisy samples.
+
+Adaptive Centroid Distance. Although the sample separability of WJSD is greatly improved compared with JSD, the separation metric still relies on model prediction. When all the noisy samples labeled in a class are from another class, e.g., asymmetric noise, the model prediction for both noisy and clean samples will be highly similar because it is easy to learn a classifier that maps the features from two different classes into one. It results in low discrimination between the model predictions of clean and noisy samples in this case. Thus, solely using WJSD may not be enough to separate the clean and noisy samples when model predictions (in terms of all classes, not only about  $p^c$ ) are close.
+
+Except for separating the samples in the output space, we propose another metric calculated in the feature space to eliminate the bias from the classifier because the learned features are more robust to noise labels. Specifically, for a given sample, we can calculate the distance between its feature and the class feature centroid to evaluate how the feature of a sample deviates from its class centroid. This approach is only practical when the quality of the centroid is high. Directly calculating the centroid according to the observed label may not be accurate because it involves a certain number of features of the noisy samples. Therefore, we define purity in Definition 1 to assess the quality of the centroid of class  $c$ .
+
+Definition 1 (Purity) Suppose a noisy sample set of class  $c$  be  $\mathcal{D}_c = \{(x,\hat{y})\mid \hat{y} = c\}$ , where the intrinsic label corresponding to the sample is  $y$ . The purity  $P_{\mathcal{D}_c}$  of set  $\mathcal{D}_c$  is calculated by:
+
+$$
+P _ {\mathcal {D} _ {c}} = \max  _ {k = 1, \dots , M} \left\{\frac {\sum_ {i = 1} ^ {N} I \left(y _ {i} = k\right)}{N} \right\}, \tag {4}
+$$
+
+where  $I(\cdot)$  is an indicator function and  $N$  is the number of samples in the set  $\mathcal{D}_c$ .
+
+Purity indicates the proportion of an intrinsic class that takes the majority in the observed class  $c$ . The higher the purity of a class set used to calculate the centroid, the better the centroid helps to distinguish between noise and clean samples. RoLT [49] adopts a similar idea to use class feature centroid for noisy sample detection. However it directly calculates the feature centroid on the observed class, such that it inevitably suffers from the existence of noise features when the purity is low, especially in our problem.
+
+In order to improve the purity of the class feature centroid for distance calculation between features, we propose ACD as the second separation metric to separate samples jointly with WJSD. The feature centroid for each class is adaptively updated by involving samples with high confidence from the observed class. The class centroid  $\mathbf{o}_c$  based
+
+on a high-confidence sample set  $\mathcal{D}_c^H$  is calculated by:
+
+$$
+\mathbf {o} _ {c} = \frac {1}{| \mathcal {D} _ {c} ^ {H} |} \sum_ {i = 1} ^ {| \mathcal {D} _ {c} ^ {H} |} \mathbf {f} _ {i}, \tag {5}
+$$
+
+$$
+\mathcal {D} _ {c} ^ {H} = \left\{x _ {i} \mid x _ {i} \in \mathcal {D} _ {c}, p _ {i} ^ {t _ {c}} > H _ {c} \right\}, \tag {6}
+$$
+
+where  $\mathbf{f}_i$  is the feature of  $x_{i}$ , and  $t_c = \mathrm{argmax}_c\{\bar{p}_c^j\}$  is the class index with the largest average prediction confidence of class  $c$ . Thus, we can use the prediction confidence of class  $t_c$  for sample  $x_{i}$ , e.g.,  $p_i^{t_c}$ , as the selection criteria compare with the threshold  $H_{c}$ .  $\mathcal{D}_c^H$  is constructed by the samples in  $D_{c}$  whose corresponding prediction confidence of class  $t_c$  is higher than  $H_{c}$ . The high-confidence threshold  $H_{c}$  is defined as:
+
+$$
+H _ {c} = \frac {1}{D _ {c}} \sum_ {i = 1} ^ {| D _ {c} |} w _ {i} \times p _ {i} ^ {t _ {c}}, \tag {7}
+$$
+
+$$
+w _ {i} = \max  \left(1, p _ {i} ^ {t _ {c}} / \bar {p} _ {c} ^ {t _ {c}}\right). \tag {8}
+$$
+
+$H_{c}$  is calculated by the weighted sum of the prediction confidence of class  $t_c$  for all samples  $x_{i}$ . The weight  $w_{i}$  increases when a sample's prediction confidence of class  $t_c$  is higher than its class average. Thus, the threshold is high enough to ensure the purity of the selected samples. High confidence samples are more representative and more likely to be in the same class. It is worth noting that we choose more representative samples rather than clean samples because the clean samples in the tail classes are easily overwhelmed by the noise samples in our problem. In this case, it is difficult to select the clean samples directly because the noise samples may be more representative in the tail class. As for asymmetric noise, the choice of clean or noisy samples to obtain centroid for sample separation is equivalent.
+
+In each round, the centroids are adaptively adjusted because the feature extractor in the model is updated, as well as the set  $\mathcal{D}_c^H$ . Therefore, the proposed metric ACD can be calculated by:
+
+$$
+A C D \left(x _ {i}\right) = \cos \left(\mathbf {f} _ {i}, \mathbf {o} _ {c}\right). \tag {9}
+$$
+
+Remarks: The proposed two metrics, WJSD and ACD, are complementary for sample separation in the tail classes. When the ground-truth classes of the noisy samples are diverse, e.g. symmetric noise, WJSD plays the dominant role because the predictions can hardly be unified due to the class diversity of noisy samples. They are more likely to be predicted towards their ground-truth class, which can be captured by  $\max (\mathbf{p}_i)$ . In this case, ACD may not be effective in separating clean samples from noisy samples because the messy noisy samples may affect the purity of the calculated centroid. On the other hand, when the ground-truth classes of the noisy samples belong to only one class, e.g., asymmetric noise, WJSD tends to produce similar predictions for noisy and clean samples, while the adaptive centroid in ACD can help distinguish them in the feature space.
+
+Algorithm 1: The Dimension Selection Strategy  
+Input: Noise sample set  $\mathcal{D}_c$  in class  $c$ , threshold  $\eta$   
+Output: Cluster  $\mathcal{G}_c^1, \mathcal{G}_c^2$   
+1 Obtain  $\mathcal{G}_{wjsd}^1, \mathcal{G}_{wjsd}^2$  and separation threshold  $d$  by applying GMM with values of WJSD to all samples in  $\mathcal{D}_c$   
+2 Obtain  $\mathcal{G}_{acd}^1, \mathcal{G}_{acd}^2$  by applying GMM with values of ACD to all samples in  $\mathcal{D}_c$   
+3 Calculate the mean  $\mu_1, \mu_2$  and variance  $\sigma_1, \sigma_2$  of WJSD for  $\mathcal{G}_{acd}^1, \mathcal{G}_{acd}^2$   
+4 if  $\mu_1 < d < \mu_2$  and  $\sigma_2 / \sigma_1 < \eta$  then  
+5 return  $\mathcal{G}_{wjsd}^1, \mathcal{G}_{wjsd}^2$   
+6 else if  $\mu_1 > d$  and  $\mu_2 > d$  then  
+7 return  $\mathcal{G}_{wjsd}^1, \mathcal{G}_{wjsd}^2$   
+8 else  
+9 return  $\mathcal{G}_{acd}^1, \mathcal{G}_{acd}^2$   
+10 end
+
+# 4.2. Bi-Dimensional Sample Selection
+
+Once the values of the bi-dimensional metrics for all samples in class  $c$  are calculated, each sample can be represented by a point in a 2D space. For each dimension, Gaussian mixture model (GMM) can be adopted to separate the samples in class  $c$  into two clusters  $\mathcal{G}_c^1$ ,  $\mathcal{G}_c^2$ .
+
+Dimension Selection. Because each metric shows different separability for different noise types, we first propose a dimension selection strategy to select a proper dimension to separate clean and noisy samples. We consider three cases to select the proper dimension: (a) both of them are acceptable; (b) the separability of WJSD is better; (c) the separability of ACD is better. In short, we only need to figure out if WJSD is suitable for sample separation. If it is certain that WJSD does not show good separability, ACD will be adopted. Fig. 3 shows three cases for dimension selection.
+
+Therefore, we use the statistics of WJSD as the criterion of dimension selection. We measure the means  $(\mu_1,\mu_2)$  and standard deviations  $(\sigma_{1},\sigma_{2})$  of WJSD for two clusters separated by GMM in terms of ACD. We have two cases that can determine when WJSD is better. We mainly compare
+
+![](images/2de398f5880d8ca6a5cb6c69c9c5ba5bff5667799a17761e1e0b4df02a61fdfb.jpg)  
+Figure 3: Examples of three cases of the optimal separation dimension correspond to sample distribution in bidimensional metrics.
+
+$\mu_{1}$  and  $\mu_{2}$  with the threshold  $d$  of GMM in WJSD. First, as shown in Fig. 3 (b), if one cluster  $\mathcal{G}_{acd}^{1}$  of ACD has a  $\mu_{1}$  less than  $d$  and a  $\sigma_{1}$  much larger than  $\sigma_{2}$  in another cluster with a  $\mu_{2}$  larger than  $d$ , it means that cluster  $\mathcal{G}_{acd}^{1}$  has more clean samples but also many noise samples. So in this case, ACD is improper and WJSD is proper. Second, if both  $\mu_{1}, \mu_{2}$  are greater than  $d$ , it means that it is a representation of the previous case in the tail class because there are fewer clean samples in the cluster. So, in this case, WJSD is proper. The strategy is summarized in Algorithm 1.
+
+Cluster Selection. Once we have determined the separation dimension, we can obtain two sample clusters  $\mathcal{G}_c^1$  and  $\mathcal{G}_c^2$ . However, it is not clear which cluster contains more clean samples because the noisy samples may overwhelm the clean samples in some cases. Therefore, we need to further select a cluster between  $\mathcal{G}_c^1$  and  $\mathcal{G}_c^2$  as the clean one. For the cluster separated based on WJSD, we simply choose the one with a smaller average WJSD value as the cluster with more clean samples, denoted as  $\mathcal{D}_c^{clean}$  [29, 23]. For the cluster separated based on ACD, we cannot directly select the cluster closest to the centroid as the cluster with more clean samples since we take more representative samples, which may compose of noisy samples, rather than clean samples as the centroid. Therefore, we need to determine the selection criteria based on whether the current class centroid is obtained by clean samples. Especially, the criterion is based on the similarity between centroids of the current class and other classes. Assume that the sample in  $\mathcal{G}_c^1$  is near the centroid and the sample in  $\mathcal{G}_c^2$  is the opposite. We determined the choice of a more clean cluster  $\mathcal{D}_c^{clean}$  based on the following criteria:
+
+$$
+\mathcal {D} _ {c} ^ {\text {c l e a n}} = \left\{ \begin{array}{l} \mathcal {G} _ {c} ^ {2}, \text {i f} | \cos (o _ {c}, o _ {k}) - 1 | <   \varepsilon \text {a n d} \left| \mathcal {D} _ {c} ^ {H} \right| <   \left| \mathcal {D} _ {k} ^ {H} \right| \\ \mathcal {G} _ {c} ^ {1}, \text {o t h e r w i s e} \end{array} \right. \tag {10}
+$$
+
+It means that if the centroid of another class is similar to the centroid of the current class, whichever class of high-confidence sample set has fewer samples, then the high-confidence samples in this class are the noise samples. It is because the asymmetric noise ratio cannot exceed  $50\%$  in the past assumption [29, 23]. Therefore, if the centroid is obtained from noise samples, we choose the cluster far away from the centroid as the cluster with more clean samples; otherwise, the cluster close to the centroid is selected.
+
+# 4.3. Overall Training Process
+
+After sample selection is conducted for each class, we adopt semi-supervised learning [1, 42] to train with all clean samples  $\mathcal{D}^{clean}$  as labeled data and all noisy samples  $\mathcal{D}^{noisy}$  as unlabeled data. The model is thus updated for the next round of training. One may also use long-tailed semi-supervised learning methods instead of normal semi-supervised learning in this stage.
+
+# 5. Experiments
+
+# 5.1. Experimental Setup
+
+Datasets. In order to comprehensively evaluate the effectiveness of method, we perform experiments respectively in the scenarios with synthetic noise and realistic noise in intrinsically long-tailed distribution. We construct synthetic scenarios based on CIFAR-10 and CIFAR-100 [26]. We construct two benchmarks to imitate realistic scenarios, which is built on real-world noise datasets including Red Mini-ImageNet [20], CIFAR-10N and CIFAR-100N [48]. For all datasets, we adopt the standard paradigm of constructing a long-tailed distribution first and injecting noise later. We imitate realistic long-tailed distribution using the same setting in previous long-tailed learning works [3, 35] that long-tailed imbalance follows an exponential decay in sample sizes across different classes. The imbalance factor is denoted by the ratio between the size of the largest class and that of the smallest class.
+
+The details for dataset construction are as follows. (1) As for CIFAR-10/100 [26], we adopt symmetric and asymmetric noise to inject synthetic noise, which is commonly used in the area of label-noise learning [41]. The noise ratio is denoted by the ratio between the number of noisy samples and the total number of samples. It should be noted that different from the previous methods [21, 49] to deal with the joint problem of noisy labeled and long-tailed data, the noise transition matrix is randomly generated only related to the noise ratio. (2) As for CIFAR-10N/100N (10N/100N) [48], we replace sample labels with human-annotated noisy labels after constructing a long-tailed distribution based on real labels. (3) As for Red mini-[ImageNet (Red) [20], we inject web label noise samples after constructing a long-tailed distribution based on the original mini-[ImageNet. Therefore, the observed and intrinsic distribution are likely inconsistent, especially in long-tail distribution.
+
+Compared Methods. We compare our method with the following three types of approaches: (1) Long-tail learning methods (LT). They are LA [33], LDAM [3] and IB [35]. (2) Label-noise learning methods (NL). They are DivideMix [29] and UNICON [23]; (3) Methods aiming at dealing with noisy label and long-tail distribution (NL-LT). They are MW-Net [41], RoLT [49], HAR [2] and ULC [19].
+
+Implementation Details. We use PreAct ResNet18 [14] as backbone for CIFAR datasets and ResNet18 as backbone for Red mini-ImageNet dataset. Both of backbones adopt SGD with an initial learning rate of 0.02, a momentum of 0.9, and a weight decay of  $5 \times 10^{-4}$  and a batch size of 64. For fair comparison, we train all methods for 100 epochs.
+
+# 5.2. Comparative Results
+
+CIFAR-10/100. Tab. 1 and 2 report the accuracy of different methods for intrinsically long-tailed CIFAR-10/100
+
+<table><tr><td colspan="2">Dataset</td><td colspan="2">CIFAR-10</td><td colspan="2">CIFAR-100</td></tr><tr><td colspan="2">Imbalance Factor</td><td colspan="4">0.1</td></tr><tr><td colspan="2">Noise Ratio (Sym.)</td><td>0.4</td><td>0.6</td><td>0.4</td><td>0.6</td></tr><tr><td>Baseline</td><td>CE</td><td>71.67</td><td>61.16</td><td>34.53</td><td>23.63</td></tr><tr><td rowspan="3">LT</td><td>LA</td><td>70.56</td><td>54.92</td><td>29.07</td><td>23.21</td></tr><tr><td>LDAM</td><td>70.53</td><td>61.97</td><td>31.30</td><td>23.13</td></tr><tr><td>IB</td><td>73.24</td><td>62.62</td><td>32.40</td><td>25.84</td></tr><tr><td rowspan="2">NL</td><td>DivideMix</td><td>82.67</td><td>80.17</td><td>54.71</td><td>44.98</td></tr><tr><td>UNICON</td><td>84.25</td><td>82.29</td><td>52.34</td><td>45.87</td></tr><tr><td rowspan="4">NL-LT</td><td>MW-Net</td><td>70.90</td><td>59.85</td><td>32.03</td><td>21.71</td></tr><tr><td>RoLT</td><td>81.62</td><td>76.58</td><td>42.95</td><td>32.59</td></tr><tr><td>HAR</td><td>77.44</td><td>63.75</td><td>38.17</td><td>26.09</td></tr><tr><td>ULC</td><td>84.46</td><td>83.25</td><td>54.91</td><td>44.66</td></tr><tr><td>Our</td><td>TABASCO</td><td>85.53</td><td>84.83</td><td>56.52</td><td>45.98</td></tr></table>
+
+with symmetric and asymmetric noise, respectively. It can be observed three phenomena as follows: (1) The performance of existing long-tail methods is lower than baseline in most situations. It is because such methods do not have the ability to distinguish the noise samples, and the wrong attention to the noise samples leads to the further decline of the model generalization ability. (2) The advantages of existing methods aiming at both long-tailed distribution and label noise are not obvious compared with single methods. It is because existing methods do not take into account the distribution inconsistency caused by noise labels. (3) Our method significantly improved over other methods in all cases. Moreover, as the proportion of noise increases, the accuracy of our method decreases less compared to other methods. It validates that our method can effectively identify the noise in long-tail distribution.
+
+Red mini-ImageNet and CIFAR-10N/100N. Tab. 4 reports the accuracy of different methods on real-world noise datasets with intrinsically long-tailed distribution. According to the results of label-noise learning methods on CIFAR-10N/100N, it can be observed that real-world noise is more difficult to separate than synthetic noise to some extent. It further increases the difficulty of dealing with such problem. Nevertheless, our method still performs well, which again confirms its effectiveness.
+
+# 5.3. Ablation Study and Discussions
+
+In the following experiments, we analyze each component of the proposed TABASCO on CIFAR-10/100 to verify its effectiveness. We use an imbalance factor of 0.1 and a noise ratio of 0.4 to construct the synthetic noise dataset.
+
+Effectiveness of WJSD. To validate the effectiveness of the proposed WJSD, we compare the accuracy of models trained by different separation metrics. Specifically, we
+
+Table 1: Performance comparison under symmetric noise. The best results are shown in bold.  
+
+<table><tr><td colspan="2">Dataset</td><td colspan="2">CIFAR-10</td><td colspan="2">CIFAR-100</td></tr><tr><td colspan="2">Imbalance Factor</td><td colspan="4">0.1</td></tr><tr><td colspan="2">Noise Ratio (Asym.)</td><td>0.2</td><td>0.4</td><td>0.2</td><td>0.4</td></tr><tr><td>Baseline</td><td>CE</td><td>79.90</td><td>62.88</td><td>44.45</td><td>32.05</td></tr><tr><td rowspan="3">LT</td><td>LA</td><td>71.49</td><td>59.88</td><td>39.34</td><td>28.49</td></tr><tr><td>LDAM</td><td>74.58</td><td>62.29</td><td>40.06</td><td>33.26</td></tr><tr><td>IB</td><td>73.49</td><td>58.36</td><td>45.02</td><td>35.25</td></tr><tr><td rowspan="2">NL</td><td>DivideMix</td><td>80.92</td><td>69.35</td><td>58.09</td><td>41.99</td></tr><tr><td>UNICON</td><td>72.81</td><td>69.04</td><td>55.99</td><td>44.70</td></tr><tr><td rowspan="4">NL-LT</td><td>MW-Net</td><td>79.34</td><td>65.49</td><td>42.52</td><td>30.42</td></tr><tr><td>RoLT</td><td>73.30</td><td>58.29</td><td>48.19</td><td>39.32</td></tr><tr><td>HAR</td><td>82.85</td><td>69.19</td><td>48.50</td><td>33.20</td></tr><tr><td>ULC</td><td>74.07</td><td>73.19</td><td>54.45</td><td>43.20</td></tr><tr><td>Our</td><td>TABASCO</td><td>82.10</td><td>80.57</td><td>59.39</td><td>50.51</td></tr></table>
+
+Table 2: Performance comparison under asymmetric noise. The best results are shown in bold.  
+
+<table><tr><td>Dataset</td><td>Separation metric</td><td>Accuracy</td></tr><tr><td rowspan="2">CIFAR-10</td><td>JSD</td><td>83.97</td></tr><tr><td>WJSD</td><td>85.57</td></tr><tr><td rowspan="2">CIFAR-100</td><td>JSD</td><td>55.80</td></tr><tr><td>WJSD</td><td>56.67</td></tr></table>
+
+Table 3: Performance comparison between JSD and WJSD.
+
+train the model by using JSD and WJSD for sample separation with a symmetric noise dataset, respectively. As for each separation metric, we simply use the small value strategy [29, 23] for sample selection. As Tab. 3 shows, the model trained by WJSD achieves better results on both symmetric noise datasets compared with JSD owing to the stronger separability of WJSD.
+
+Effectiveness of ACD. To validate the advantage of the proposed ACD, we compare it with the method which calculates the feature centroid on the observed class (hereinafter referred to as centroid distance or CD) to show their separability in CIFAR-10. Specifically, based on the same backbone, we calculate the distance from the sample to the centroid using ACD and CD in CIFAR-10 with asymmetric noise, respectively. Fig. 4 shows the distributions of samples in the tail class by CD and ACD. The purity of the tail class is 0.53. It can be observed that the CD values of noise and clean samples are highly overlapped, which leads to failure sample separation. It is because the purity of the tail class may be quite low due to the negative effects of both label-noise and long-tailed distribution. The centroid obtained by ACD has a higher purity, so there is a significant difference in ACD between them. As shown in Fig. 4(b), most clean samples are concentrated in the interval of [0.8,1], which is easily clustered by GMM.
+
+<table><tr><td colspan="2">Dataset</td><td colspan="2">Red</td><td>10N</td><td>100N</td></tr><tr><td colspan="2">Imbalance Factor</td><td colspan="2">≈ 0.1</td><td colspan="2">0.1</td></tr><tr><td colspan="2">Noise Ratio</td><td>0.2</td><td>0.4</td><td colspan="2">≈ 0.4</td></tr><tr><td>Baseline</td><td>CE</td><td>40.42</td><td>31.46</td><td>60.44</td><td>38.10</td></tr><tr><td rowspan="3">LT</td><td>LA</td><td>26.82</td><td>25.88</td><td>65.74</td><td>36.50</td></tr><tr><td>LDAM</td><td>26.64</td><td>23.46</td><td>62.50</td><td>38.48</td></tr><tr><td>IB</td><td>23.80</td><td>22.08</td><td>65.91</td><td>42.48</td></tr><tr><td rowspan="2">NL</td><td>DivideMix</td><td>48.76</td><td>48.96</td><td>67.85</td><td>44.25</td></tr><tr><td>UNICON</td><td>40.18</td><td>41.64</td><td>69.54</td><td>51.93</td></tr><tr><td rowspan="4">NL-LT</td><td>MW-Net</td><td>42.66</td><td>40.26</td><td>69.73</td><td>44.20</td></tr><tr><td>RoLT</td><td>22.56</td><td>24.22</td><td>75.24</td><td>46.61</td></tr><tr><td>HAR</td><td>46.61</td><td>38.71</td><td>74.97</td><td>44.54</td></tr><tr><td>ULC</td><td>48.12</td><td>47.06</td><td>75.71</td><td>51.72</td></tr><tr><td>Our</td><td>TABASCO</td><td>50.20</td><td>49.68</td><td>80.61</td><td>53.83</td></tr></table>
+
+Table 4: Performance comparison with real-world noise and long-tail distribution. The best results are shown in bold.
+
+Effectiveness of Bi-dimensional Sample Separation. To validate how the proposed bi-dimensional metrics complement each other, we plot the values of bi-dimensional metrics of both clean and noisy samples under different noise types in Fig. 5. For the case of symmetric noise shown in Fig. 5(a), it can be observed that ACD cannot effectively distinguish clean samples from noisy samples in the tail class. In this case, WJSD shows its advantage because the values of WJSD for most noisy samples are clustered in the top region, and the clean samples are scattered throughout the rest region. For the case of asymmetric noise shown in Fig. 5(b), it can be observed that WJSD cannot distinguish clean samples from noisy samples well for the tail class, while ACD can distinguish them well although the clean and noisy samples are closer in the tail class. The experimental observation is consistent with the previous discussion, which validates the complementarity of bi-dimensional metrics.
+
+Effectiveness of Sample Selection. In order to validate the effectiveness of the proposed sample selection in dimension selection and cluster selection, we compare the optimal clean ratio of clusters obtained by WJSD, ACD and our selection method in CIFAR-10 with different noise types. The clean ratio of a cluster is calculated by the proportion of clean samples in the cluster to the total number of cluster samples. As shown in Fig. 6, on the tail class, the clean ratio of the cluster obtained by our selection method is aligned with the higher one between WJSD and ACD for both noise cases. It shows that we can choose the most appropriate separation dimension in different situations. Besides, the clean ratio of the cluster we selected is consistent with that of the optimal cluster in the corresponding dimension, which is sufficient to show that the method we proposed can select clean samples.
+
+![](images/8978744f5c52019e23a6df9ddbcefd4b3d31544d821d855b90bb486837f20e8a.jpg)  
+Figure 4: Comparison of the sample distributions between CD and ACD in the tail class.
+
+![](images/2ab50acdf51b02f9a848c4538c99922c871944a12dfcdf0eedd93ee4d564c139.jpg)
+
+![](images/6a8bf68f9db9b8598d1392047cd47c220a1be42a8cd501b7dec94e1c1f3ce57f.jpg)  
+(a) Symmetric noise
+
+![](images/87cf4887cb473a40cbc410da6221f0e73f402470ce88fd0ef33b08ab0eaab4d3.jpg)  
+(b) Asymmetric noise
+
+![](images/df444e84913d8f40b83c37295349c5d42fb46363247f2eca2f479539d74cc1f3.jpg)  
+Figure 5: Scatter plot of the values of the proposed bidimensional metrics with (a) symmetric noise and (b) asymmetric noise.  
+(a) Symmetric noise  
+Figure 6: The performance of the proposed sample selection with (a) symmetric noise and (b) asymmetric noise.
+
+![](images/34026ab00885b70d2d07597d1d8b93c9ba16dae610525ee4a0436ad8f17d4e94.jpg)  
+(b) Asymmetric noise
+
+# 6. Conclusion
+
+This paper studies a more general and realistic problem of label-noise learning with intrinsically long-tailed data. The major challenge in this problem is that it is hard to distinguish clean samples from noisy samples on intrinsic tail classes. Accordingly, we propose a learning framework TABASCO for this problem. In TABASCO, two new metrics are explicitly proposed to address the problem of sample selection in tail classes. Extensive experiments on noisy-labeled datasets with long-tail distribution demonstrate its effectiveness.
+
+Acknowledgements. We thank Dr. Shiquan Ruan for his discussion on the theory proof. This study was supported in part by the National Natural Science Foundation of China under Grants 62002302 and U21A20514; in part by the FuXiaQuan National Independent Innovation Demonstration Zone Collaborative Innovation Platform under Grant 3502ZCQXT2022008; in part by NSFC/RGC Joint Research Scheme under Grant N_HKBU214/21; and in part by the General Research Fund of RGC under Grants 12201321 and 12202622.
+
+# References
+
+[1] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, 2019. 6  
+[2] Kaidi Cao, Yining Chen, Junwei Lu, Nikos Aréchiga, Adrien Gaidon, and Tengyu Ma. Heteroskedastic and imbalanced deep learning with adaptive regularization. In ICLR, 2021. 2, 6  
+[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. In NeurIPS, 2019. 2, 3, 6  
+[4] Nitesh V. Chawla, Kevin W. Bowyer, Lawrence O. Hall, and W. Philip Kegelmeyer. SMOTE: synthetic minority oversampling technique. J. Artif. Intell. Res., 2002. 2  
+[5] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: Rebalanced mixup. In ECCV, 2020. 2  
+[6] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge J. Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 2, 3  
+[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 1  
+[8] Andrew Estabrooks, Taeho Jo, and Nathalie Japkowicz. A multiple resampling method for learning from imbalanced data sets. Comput. Intell., 2004. 2  
+[9] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 1  
+[10] Bo Han, Gang Niu, Xingrui Yu, Quanming Yao, Miao Xu, Ivor W. Tsang, and Masashi Sugiyama. SIGUA: forgetting may make learning with noisy labels more robust. In ICML, 2020. 2  
+[11] Bo Han, Quanming Yao, Tongliang Liu, Gang Niu, Ivor W. Tsang, James T. Kwok, and Masashi Sugiyama. A survey of label-noise representation learning: Past, present and future. CoRR, 2020. 1  
+[12] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor W. Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, 2018. 1, 2  
+[13] Hui Han, Wenyuan Wang, and Binghamuan Mao. Borderlinesmote: A new over-sampling method in imbalanced data sets learning. In ICIC 2005, 2005. 2  
+[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6  
+[15] Yin-Yin He, Jianxin Wu, and Xiu-Shen Wei. Distilling virtual examples for long-tailed recognition. In ICCV, 2021. 2  
+[16] Dan Hendrycks, Mantas Mazeika, Duncan Wilson, and Kevin Gimpel. Using trusted data to train deep networks on labels corrupted by severe noise. In NeurIPS, 2018. 2  
+[17] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexander Shepard, Hartwig Adam, Pietro Per
+
+ona, and Serge J. Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 1  
+[18] Grant Van Horn and Pietro Perona. The devil is in the tails: Fine-grained classification in the wild. CoRR, 2017. 1  
+[19] Yingsong Huang, Bing Bai, Shengwei Zhao, Kun Bai, and Fei Wang. Uncertainty-aware learning against label noise on imbalanced datasets. In AAAI, 2022. 2, 6  
+[20] Lu Jiang, Di Huang, Mason Liu, and Weilong Yang. Beyond synthetic noise: Deep learning on controlled noisy labels. In ICML, 2020. 6  
+[21] Shenwang Jiang, Jianan Li, Ying Wang, Bo Huang, Zhang Zhang, and Tingfa Xu. Delving into sample loss curve to embrace noisy and imbalanced data. In AAAI, 2022. 2, 6  
+[22] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In ICLR, 2020. 2  
+[23] Nazmul Karim, Mamshad Nayeem Rizve, Nazanin Rahnavard, Ajmal Mian, and Mubarak Shah. Unicon: Combating label noise through uniform selection and contrastive learning. In CVPR, 2022. 1, 2, 3, 6, 7  
+[24] Shyamgopal Karthik, Jérôme Revaud, and Chidlovskii Boris. Learning from long-tailed data with noisy labels. CoRR, 2021. 2  
+[25] Jaehyung Kim, Youngbum Hur, Sejun Park, Eunho Yang, Sung Ju Hwang, and Jinwoo Shin. Distribution aligning refinery of pseudo-label for imbalanced semi-supervised learning. In NeurIPS, 2020. 2  
+[26] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 6  
+[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 1  
+[28] Hyuck Lee, Seungjae Shin, and Heeyoung Kim. ABC: auxiliary balanced classifier for class-imbalanced semi-supervised learning. In NeurIPS, 2021. 2  
+[29] Junnan Li, Richard Socher, and Steven C. H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In ICLR, 2020. 1, 2, 3, 6, 7  
+[30] Shikun Li, Xiaobo Xia, Shiming Ge, and Tongliang Liu. Selective-supervised contrastive learning with noisy labels. In CVPR, 2022. 2  
+[31] Xu-Ying Liu, Jianxin Wu, and Zhi-Hua Zhou. Exploratory undersampling for class-imbalance learning. IEEE Trans. Syst. Man Cybern. Part B, 2009. 2  
+[32] Michal Lukasik, Srinadh Bhojanapalli, Aditya Krishna Menon, and Sanjiv Kumar. Does label smoothing mitigate label noise? In ICML, 2020. 2  
+[33] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In ICLR, 2021. 6  
+[34] Diego Ortego, Eric Arazo, Paul Albert, Noel E. O'Connor, and Kevin McGuinness. Multi-objective interpolation training for robustness to label noise. In CVPR, 2021. 2  
+[35] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In ICCV, 2021. 6
+
+[36] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In CVPR, 2017. 2  
+[37] Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey E. Hinton. Regularizing neural networks by penalizing confident output distributions. In ICLR, 2017. 2  
+[38] William J Reed. The pareto, zipf and other power laws. Economics Letters, 74(1):15-19, 2001. 1  
+[39] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In NeurIPS, 2020. 2  
+[40] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1  
+[41] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In NeurIPS, 2019. 2, 6  
+[42] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 6  
+[43] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. SELFIE: refurbishing unclean samples for robust deep learning. In ICML, 2019. 1  
+[44] Hwanjun Song, Minseok Kim, Dongmin Park, and Jae-Gil Lee. Learning from noisy labels with deep neural networks: A survey. CoRR, 2020. 1  
+[45] Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In CVPR, 2020. 2, 3  
+[46] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X. Yu. Long-tailed recognition by routing diverse distribution-aware experts. In ICLR, 2021. 2  
+[47] Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan L. Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In CVPR, 2021. 2  
+[48] Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. Learning with noisy labels revisited: A study using real-world human annotations. In ICLR, 2022. 6  
+[49] Tong Wei, Jiang-Xin Shi, Wei-Wei Tu, and Yu-Feng Li. Robust long-tailed learning under label noise. CoRR, 2021. 2, 4, 6, 14  
+[50] Xiaobo Xia, Tongliang Liu, Bo Han, Chen Gong, Nannan Wang, Zongyuan Ge, and Yi Chang. Robust early-learning: Hinding the memorization of noisy labels. In ICLR, 2021. 2  
+[51] Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Jun Yu, Gang Niu, and Masashi Sugiyama. Sample selection with uncertainty of losses for learning with noisy labels. In ICLR, 2022. 2
+
+[52] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In CVPR, 2015. 1  
+[53] Yi Xuanyu, Tang Kaihua, Hua Xian-Sheng, Lim Joo-Hwee, and Zhang Hanwang. Identifying hard noise in long-tailed sample distribution. In ECCV, 2022. 2  
+[54] Yu Yao, Tongliang Liu, Bo Han, Mingming Gong, Jiankang Deng, Gang Niu, and Masashi Sugiyama. Dual T: reducing estimation error for transition matrix in label-noise learning. In NeurIPS, 2020. 2  
+[55] Yazhou Yao, Zeren Sun, Chuanyi Zhang, Fumin Shen, Qi Wu, Jian Zhang, and Zhenmin Tang. Jo-src: A contrastive approach for combating noisy labels. In CVPR, 2021. 2, 3  
+[56] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In CVPR, 2019. 2  
+[57] Yuhang Zang, Chen Huang, and Chen Change Loy. FASA: feature augmentation and sampling adaptation for long-tailed instance segmentation. In ICCV, 2021. 2  
+[58] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 2  
+[59] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In CVPR, 2021. 2  
+[60] Yongshun Zhang, Xiu-Shen Wei, Boyan Zhou, and Jianxin Wu. Bag of tricks for long-tailed visual recognition with deep convolutional neural networks. In AAAI, 2021. 1  
+[61] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2  
+[62] Beier Zhu, Yulei Niu, Xian-Sheng Hua, and Hanwang Zhang. Cross-domain empirical risk minimization for unbiased long-tailed classification. In AAAI, 2022. 1

2023/Label-Noise Learning with Intrinsically Long-Tailed Data/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c45add9f26c9b1264021224109326c5671f686b0042b28d64a9eb3f68df49ae4
+size 390335

2023/Landscape Learning for Neural Network Inversion/f1107abc-1d35-4283-b4db-d9fd38f81a61_content_list.json

@@ -0,0 +1,1836 @@
+[
+    {
+        "type": "text",
+        "text": "Landscape Learning for Neural Network Inversion",
+        "text_level": 1,
+        "bbox": [
+            227,
+            130,
+            743,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ruoshi Liu<sup>1</sup> Chengzhi Mao<sup>1</sup> Purva Tendulkar<sup>1</sup> Hao Wang<sup>2</sup> Carl Vondrick<sup>1</sup>  \n<sup>1</sup> Columbia University <sup>2</sup> Rutgers University",
+        "bbox": [
+            158,
+            179,
+            803,
+            220
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            257,
+            313,
+            273
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Many machine learning methods operate by inverting a neural network at inference time, which has become a popular technique for solving inverse problems in computer vision, robotics, and graphics. However, these methods often involve gradient descent through a highly non-convex loss landscape, causing the optimization process to be unstable and slow. We introduce a method that learns a loss landscape where gradient descent is efficient, bringing massive improvement and acceleration to the inversion process. We demonstrate this advantage on a number of methods for both generative and discriminative tasks, including GAN inversion, adversarial defense, and 3D human pose reconstruction.",
+        "bbox": [
+            75,
+            277,
+            473,
+            460
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            493,
+            209,
+            508
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Many inference problems in machine learning are formulated as inverting a forward model  $F(x)$  by optimizing an objective over the input space  $x$ . This approach, which we term optimization-based inference (OBI), has traditionally been used to solve a range of inverse problems in vision, graphics, robotics, recommendation systems, and security [28, 38, 22, 10, 62, 17]. Recently, neural networks have emerged as the paramterization of choice for forward models [43, 53, 1, 49, 76, 68, 13, 78], which can be pretrained on large collections of data, and inverted at testing time in order to solve inference queries.",
+        "bbox": [
+            75,
+            518,
+            470,
+            686
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Optimization-based inference (OBI) has many advantages over feed-forward or encoder-based inference (EBI). Without encoder, OBI provides flexibility to adapt to new tasks, allowing one to define new constraints into the objective during inference. When observations are partially missing, OBI can adapt without additional training. Moreover, OBI naturally supports generating multiple and diverse hypotheses when there is uncertainty. Finally, OBI has intrinsic advantages for robustness, both adapting to new data distributions as well as defending against adversarial examples.",
+        "bbox": [
+            75,
+            686,
+            470,
+            838
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, the key bottleneck for OBI in practice is the computational efficiency and the speed of inference. Feedforward models are fast because they only require a single forward pass of a neural network, but OBI requires many",
+        "bbox": [
+            75,
+            839,
+            470,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/5886baaa4e06470f25ae06068a39dd9120e390fe68b37291340f97a2996cbbf0.jpg",
+        "image_caption": [
+            "Figure 1: Loss Landscapes Comparison. The loss landscape of optimization-based inference (OBI) is often highly non-convex. We propose to learn a smoother loss landscape through a mapping network to accelerate the optimization. Plotted from real data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            258,
+            694,
+            386
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/ea580b0c066fa2958e2c0629e07c010bdcb00c7fb1d73626bf901eaa0425efd6.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            696,
+            258,
+            880,
+            386
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "(often hundreds) steps of optimization in order to obtain strong results for one example. Forward models in OBI are often trained with generative or discriminative tasks, but they are not trained for the purpose of performing gradient descent in the input space. Fig. 9 visualizes the loss landscape for uncurated examples. The loss landscape is not guaranteed to be an efficient path from the initialization to the solution, causing the instability and inefficiency.",
+        "bbox": [
+            496,
+            476,
+            893,
+            595
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a framework to accelerate and stabilize the inversion of forward neural networks. Instead of optimizing over the original input space, we learn a new input space such that gradient descent converges quickly. Our approach uses an alternating algorithm to learn the mapping between these spaces. The first step collects optimization trajectories in the new space until convergence. The second step updates the mapping parameters to reduce the distance between the convergence point and each point along the trajectory. By repeating these steps, our approach will learn a mapping between the spaces that allows gradient descent over the input to converge in significantly fewer steps.",
+        "bbox": [
+            496,
+            597,
+            895,
+            777
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Empirical experiments and visualizations on both generative and discriminative models show that our method can largely improve the convergence speed for optimization. We validate our approach on a diverse set of vision tasks, including GAN inversion [1], adversarial defense [45], and 3D human pose reconstruction [53]. Our experiments show that our method converges an order of magnitude faster without loss in absolute performance after convergence. As our ap",
+        "bbox": [
+            496,
+            777,
+            895,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "2239",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/6a230d29c2b6a3599d5ea8047c1a3131a7bc79ba84216272c6c01daec7e57e56.jpg",
+        "image_caption": [
+            "Figure 2: Method. The left and middle figure show the loss landscape for our latent space and the original latent space, respectively. While walking to the optimal solution in a few steps is hard in  $X$  space, it can be done in our learned loss landscapes."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            140,
+            74,
+            834,
+            200
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "proach does not require retraining of the forward model, it is compatible to all existing OBI methods with a differentiable forward model and objective function.",
+        "bbox": [
+            75,
+            244,
+            468,
+            289
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The primary contribution of this paper is an efficient optimization-based inference framework. In Sec. 2, we survey the related literature to provide an overview of forward model inversion problem. In Sec. 3, we formally define OBI (3.1); our method to learn an efficient loss landscape for OBI (3.2); a training algorithm for better generalization and robustness (3.3). In Sec. 4, we experimentally study and analyze the effectiveness of the mapping network for OBI.",
+        "bbox": [
+            75,
+            289,
+            470,
+            411
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            422,
+            218,
+            436
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The different approaches for inference with a neural network can be partitioned into either encoder-based inference, which is feedforward, or optimization-based inference, which is iterative. We briefly review these two approaches in the context of our work. In addition, we also review representative work in meta-learning and discuss the similarities and differences with our work.",
+        "bbox": [
+            75,
+            446,
+            472,
+            551
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Encoder-based Inference",
+        "text_level": 1,
+        "bbox": [
+            76,
+            560,
+            305,
+            574
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Encoder-based inference trains a neural network  $F$  to directly map from the output space to the input space. Autoencoder based approach [55] learns an encoder that maps the input data to the latent space. [59, 65, 70, 54] learn an encoder from the image to the latent space in GAN. Encoder based inference requires training the encoder on the anticipated distribution in advance, which is often less effective and can fail on unexpected samples [21, 33]. In addition, encoder-based method can only produce one reconstruction for under-constrained inverse problems [75], even if multiple outcomes can all be true given the partial input.",
+        "bbox": [
+            75,
+            583,
+            470,
+            750
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Optimization-based Inference",
+        "text_level": 1,
+        "bbox": [
+            76,
+            756,
+            344,
+            772
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "OBI methods perform inference by solving an optimization problem with gradient-based methods such as Stochastic Gradient Descent (SGD) [60] and Projected Gradient Descent (PGD) [44]. In these cases, the objective function specifies the inference task. Besides these methods which use a point estimate for the latent variable, one can estimate the posterior distribution of the latent variable through Bayesian optimization, such as SGLD [71].",
+        "bbox": [
+            75,
+            780,
+            470,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Gradient based optimization methods have been used to infer the latent code of query samples in deep generative models like GANs [26] via GAN inversion [36, 31, 61, 81, 1, 2, 8, 29, 52]. Style transfer relies on gradient based optimization to change the style of the input images [32]. It can also create adversarial attacks that fool the classifier [18, 11, 48, 63]. Recently, back-propagation-based optimization has shown to be effective in defending adversarial examples [45, 47] and compressed sensing [9]. [74] uses MAML [23] to accelerate the optimization process in compressed sensing.",
+        "bbox": [
+            496,
+            244,
+            893,
+            410
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recently, constrained optimization was popularized for text-to-image synthesis by [19, 41]. They search in the latent space to produce an image that has the highest similarity with the given text as measured by a multi-modal similarity model like CLIP [56]. Test-time constrained optimization is also related to the idea of 'prompt-tuning' for large language models and vision-language models. [39] learns \"soft prompts\" to condition frozen language models to perform specific downstream tasks. Soft prompts are learned through backpropagation to incorporate signal from just a few labeled examples (few-shot learning). Visual prompting at both training time [7, 46] and testing time [67] also relies on effective latent variable optimization.",
+        "bbox": [
+            496,
+            410,
+            893,
+            606
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "A major challenge for optimization-based inference is how to perform efficient optimization in a highly non-convex space. To address this, input convex model [4] was proposed so that gradient descent can be performed in a convex space. [66] introduced a method to retrain the generative model such as it learns a latent manifold that is easy to optimize. When the model cannot be changed and retrained, bidirectional inference [68] and hybrid inference [81, 80] uses an encoder to provide a good initialization for the optimization-based inference in a non-convex space. Our method does not retrain the generative model, but instead maps a new latent space into the original input space, allowing more efficient optimization.",
+        "bbox": [
+            496,
+            607,
+            895,
+            803
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3. Meta-Learning",
+        "text_level": 1,
+        "bbox": [
+            500,
+            811,
+            653,
+            828
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Given a distribution of tasks, meta-learning aims to adapt quickly when presented with a previously unseen task. MAML [23] and related methods [25, 24] propose a method to learn a parameter initialization of a neural net",
+        "bbox": [
+            496,
+            835,
+            893,
+            897
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "2240",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "work by differentiating through the fine-tuning process. To reduce the computational cost of MAML due to the 2nd degree gradient, [51] proposes a first-order meta-learning algorithm. Unlike MAML which include modifying the forward model (e.g. in [74]), our approach is able to keep the forward model fixed in order to maintain its learned rich priors. Instead of meta-learning of model initialization, a line of work [5, 57, 14, 72] proposed to learn an optimizer, often in the form of an RNN, in order to accelerate gradient descent. Unlike learned optimizers that try to create better optimization algorithms, our approach instead learns to remap the loss landscape which is compatible with any choice of optimizer, including standard SGD or learned ones.",
+        "bbox": [
+            75,
+            90,
+            472,
+            287
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Learning Landscapes for Efficient Inference",
+        "text_level": 1,
+        "bbox": [
+            76,
+            301,
+            470,
+            319
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We present our framework to learn an efficient loss landscape for optimization-based inference (OBI) methods. In Sec. 3.1, we will define OBI. In Sec. 3.2, we will introduce our framework as well as the training objective. In Sec. 3.3, we will describe how to train our model with an alternating optimization algorithm and an experience-replay buffer.",
+        "bbox": [
+            75,
+            328,
+            472,
+            420
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Optimization-based Inference",
+        "text_level": 1,
+        "bbox": [
+            76,
+            429,
+            344,
+            445
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Let  $F(\\mathbf{x}) = \\hat{y}$  be a differentiable forward model that generates an output  $\\hat{y}$  given an input variable  $\\mathbf{x} \\in X$ . For example,  $\\hat{y}$  might be an image, and  $\\mathbf{x}$  might be the latent variables for a generative model. Given an observation  $y$ , the goal of OBI is to find the input  $\\hat{\\mathbf{x}} \\in X$  such that an objective function  $L(\\hat{y}, y)$  is minimized. Formally, we write this procedure as:",
+        "bbox": [
+            75,
+            452,
+            472,
+            559
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {x}} = \\underset {\\mathbf {x} \\in X} {\\operatorname {a r g m i n}} L (F (\\mathbf {x}), y) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            571,
+            468,
+            595
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "When the objective function  $L$  and the model  $F$  are both differentiable, we can perform the optimization over input space  $X$  with gradient descent.",
+        "bbox": [
+            75,
+            607,
+            468,
+            654
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Remapping the Input Space",
+        "text_level": 1,
+        "bbox": [
+            76,
+            662,
+            328,
+            680
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Instead of operating in the original input space  $X$ , we will create a new space  $Z$  where gradient descent is efficient and converges in a small number of iterations. We will use a neural network  $\\Theta: Z \\to X$  that maps from the new space  $Z$  to the original space  $X$ . The parameters of the mapping  $\\Theta$  is the vector  $\\theta$ . The learning problem we need solve is to estimate  $\\theta$  so that there is a short gradient descent path in  $Z$  from the initialization to the solution. Fig. 2 shows an overview of this setup.",
+        "bbox": [
+            75,
+            686,
+            468,
+            823
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We formulate an objective by rolling out the gradient updates on  $\\mathbf{z}$ , where we write  $\\mathbf{z}_t \\gets \\mathbf{z}_{t-1} + \\lambda \\frac{\\partial L}{\\partial \\mathbf{z}_{t-1}}$  as the  $t^{\\text{th}}$  update with a step size of  $\\lambda$ . For gradient descent in space  $Z$  to be efficient, the goal of our introduced  $\\Theta$  is to move every step  $\\mathbf{z}_t$  as close as possible to the global minima:",
+        "bbox": [
+            75,
+            823,
+            470,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/69201484fa2b47fa9b0c4acbde6be430b7c869aedea1d4120233a92138ff2b2f.jpg",
+        "image_caption": [
+            "Figure 3: Landscape Learning. An optimization trajectory  $\\{\\mathbf{z}_t\\}_{t=0}^5$  collected is used to train  $\\Theta$ .  $\\mathbf{z}_i$  that corresponds to a higher  $L_i$  will yield a higher gradient when training  $\\Theta$ . Optimization over multiple steps along the trajectory causes  $\\Theta$  to learn patterns of trajectories and create a more efficient loss landscape."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            88,
+            893,
+            233
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\theta} = \\underset {\\theta} {\\operatorname {a r g m i n}} \\mathbb {E} _ {(\\mathbf {z}, y)} \\left[ \\sum_ {t = 0} ^ {T} L (F (\\Theta (\\mathbf {z} _ {t})), y) \\right] \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            542,
+            314,
+            890,
+            363
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\text {w h e r e} \\quad \\mathbf {z} _ {t} = \\left\\{ \\begin{array}{l l} 0, & t = 0 \\\\ \\mathbf {z} _ {t - 1} + \\lambda \\frac {\\partial L}{\\partial \\mathbf {z} _ {t - 1}}, & t > 0 \\end{array} \\right.\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            545,
+            359,
+            812,
+            400
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We visualize this process with a toy example in Fig. 3. Gradient updates on  $\\Theta$  w.r.t multiple steps  $\\mathbf{z}_t$  along a trajectory will cause the loss value on each step to be lowered. By learning the parameters of  $\\Theta$  across many examples,  $\\Theta$  can learn the patterns of optimization trajectories in  $X$ . For example,  $\\Theta$  can learn to estimate the step size in  $X$  and dynamically adjust it to each example. Moreover,  $\\Theta$  can learn to reduce non-convexity of the landscape.",
+        "bbox": [
+            496,
+            409,
+            893,
+            530
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Once we obtain  $\\hat{\\Theta}$  with parameters  $\\hat{\\theta}$ , we do inference on a new example  $y$  through the standard optimization-based inference procedure, except in  $Z$  space now. Given the observation  $y$ , we find the corresponding  $\\hat{\\mathbf{x}}$  through the optimization:",
+        "bbox": [
+            496,
+            530,
+            893,
+            604
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {x}} = \\hat {\\Theta} (\\hat {\\mathbf {z}}) \\quad \\text {w h e r e} \\quad \\hat {\\mathbf {z}} = \\underset {\\mathbf {z} \\in Z} {\\operatorname {a r g m i n}} L (F (\\hat {\\Theta} (\\mathbf {z})), y) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            521,
+            616,
+            893,
+            643
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "When the inverse problem is under-constrained, one can infer multiple hypotheses for  $\\hat{\\mathbf{x}}$  by repeating the above optimization with a different random initialization for  $\\mathbf{z}_0$ .",
+        "bbox": [
+            496,
+            652,
+            893,
+            698
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3. Training",
+        "text_level": 1,
+        "bbox": [
+            500,
+            707,
+            604,
+            724
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We use alternating optimization (AO) in order to train  $\\Theta$  jointly with estimating  $\\mathbf{z}$  for each example in the training set. Specifically, we first fix parameters of  $\\Theta$  and collect  $N$  optimization trajectories of  $\\mathbf{z}$ , each with length  $T$ . Adopting the same terminology from the robotics community for learning on continuous states [50], we term this an experience replay buffer. Subsequently, we randomly sample data from this buffer and train  $\\Theta$  to optimize the loss function. We alternate between these two steps for a fixed number of iterations with gradient descent for both. Depending on the application, the training time for  $\\Theta$  varied from one hour to",
+        "bbox": [
+            496,
+            731,
+            893,
+            898
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "2241",
+        "bbox": [
+            482,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "one day using a four GPU server. Please see the appendix for more implementation details.",
+        "bbox": [
+            76,
+            90,
+            468,
+            119
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We also experimented with an online version of the training algorithm, where we update parameters of  $\\Theta$  immediately after one update to  $\\mathbf{z}$ . However, in our experiments, we found this resulted in a slower convergence rate. We show these comparisons in the ablation experiments.",
+        "bbox": [
+            76,
+            121,
+            468,
+            196
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [
+            "Algorithm 1 Learning Mapping Network  $\\Theta$"
+        ],
+        "code_body": "1: Input: Ground truth  $y$ , step size  $\\lambda_z$  and  $\\lambda_{\\theta}$ , number of buffers  $B$ , number of data samples in a buffer  $N$ , number of optimization steps per sample  $T$ , loss function  $L$ , and forward model  $F$ .  \n2: Output: Learned mapping network  $\\Theta$   \n3: Randomly initialize a mapping network  $\\Theta$  with parameters  $\\theta$   \n4: for  $b = 1, \\dots, B$  do  \n5: Initialize Experience Replay Buffer  $[\\{\\mathbf{z}_{t,i}\\}_{t=1}^{T}]_{i=1}^{N}$   \n6: for  $i = 1, \\dots, N$  do  \n7:  $\\mathbf{z}_0 \\gets \\mathbf{0}$   \n8: for  $t = 1, \\dots, T$  do  \n9:  $\\mathbf{z}_t \\gets \\mathbf{z}_{t-1} + \\lambda_z \\frac{\\partial}{\\partial \\mathbf{z}_{t-1}} L(F(\\Theta(\\mathbf{z})), y)$   \n10:  $\\mathbf{z}_{t,i} \\gets z_t$   \n11: end for  \n12: end for  \n13: for  $j = 1, \\dots, T \\cdot N$  do  \n14: Randomly sample  $\\mathbf{z}$  from previously collected buffer  \n15:  $\\theta_j \\gets \\theta_{j-1} + \\lambda_\\theta \\frac{\\partial}{\\partial \\theta_{j-1}} L(F(\\Theta(\\mathbf{z})), y)$   \n16: end for  \n17: end for  \n18: Return  $\\theta$",
+        "bbox": [
+            84,
+            229,
+            472,
+            579
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            609,
+            207,
+            626
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The goal of our experiments is to analyze how well our proposed method can be applied to various existing OBI methods to improve the optimization efficiency. We demonstrate application of our method to three diverse OBI methods in computer vision, including both generative models and discriminative models. For each OBI method, the inference-time optimization objective of the baseline and ours can be written as:",
+        "bbox": [
+            75,
+            635,
+            468,
+            755
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\text {B a s i l e :} \\hat {\\mathbf {x}} = \\min  _ {\\mathbf {x} \\in X} L (F (\\mathbf {x}), y) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            766,
+            468,
+            787
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\text {O u r s :} \\hat {\\mathbf {z}} = \\min  _ {\\mathbf {z} \\in Z} L (F (\\Theta (\\mathbf {z})), y) \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            169,
+            797,
+            468,
+            820
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Next, we provide the specific implementation of the loss term  $L$  for each application, along with quantitative and qualitative results. We also perform experiments to understand the loss landscape in Sec. 4.5 and perform ablations on different parts of our approach in Sec. 4.4.",
+        "bbox": [
+            75,
+            824,
+            468,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1. GAN Inversion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            90,
+            653,
+            104
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We first validate our method on StyleGAN inversion [1]. We take a pretrained generator of StyleGAN [35] denoted as  $F$ . Let  $y$  be an observed image whose input variable we are recovering, we optimize the objective of Eq. 5 where the loss can be written as,",
+        "bbox": [
+            496,
+            113,
+            893,
+            188
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL (\\hat {y}, y) = L _ {l p i p s} (\\hat {y}, y) + \\left\\| \\hat {y} - y \\right\\| _ {2} ^ {2} \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            200,
+            892,
+            218
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $L_{lpips}$  is a perceptual similarity loss introduced in [79],  $\\hat{y} = F(\\hat{x})$  for baseline and  $\\hat{y} = F(\\Theta(\\hat{z}))$  for ours. We train  $\\Theta$  on the train split of CelebA-HQ [34] dataset and evaluate on CelebA-HQ validation split for in-distribution experiments and LSUN-Cat[77] for distribution shifting (OOD) experiments. We compare the results from our method against the state-of-the-art encoder-based GAN inversion model [59].",
+        "bbox": [
+            496,
+            229,
+            890,
+            349
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Quantitative Results. From Fig. 4a, we see that in all experiments, optimization in our space  $Z$  consistently outperforms the baseline from the first optimization step to after convergence. This gap in performance is even larger when evaluated on OOD data. This suggests that the improvement in performance is not caused by memorizing the training data. Note that our image reconstruction performance after only 2 steps of optimization is able to outperform or be on-par with the baseline at 20 steps of optimization, resulting in a 10-fold improvement in efficiency. Even after convergence (after 2000 optimization steps), our reconstruction performance improves over the baseline by  $15\\%$  for in-distribution evaluation and  $10\\%$  for OOD evaluation. When compared with encoder-based GAN inversion [59], our method achieves better reconstruction after 11 steps of optimization for in-distribution data and 3 steps for OOD data.",
+        "bbox": [
+            496,
+            349,
+            893,
+            590
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Qualitative Results. From Fig. 8, we can see that our method already shows improvements on in-distribution data - it can almost perfectly reconstruct details like fine hair strands, the cap on the person's head, the object in the person's mouth as well-as the text on it. Interestingly, our method is able to discover and reconstruct latents for cats while the encoder-based model fails miserably as shown in Fig 8. The performance on OOD data truly highlights the benefits of our method. We also visualize how the face generations evolve over the process of optimization in Fig. 5. We can see that in just 4 steps, our method is already able to reconstruct coarse features such as face orientation, skin tone and hair color, while the baseline has hardly deviated from the initialization in regard to any of these features. Further, in Fig. 7 we visualize reconstructions from partial observations where only the center of the face (row 1) or everything other than the mouth (row 2) is visible. We can see a variety of feasible possibilities for the hidden regions (e.g., different hairstyles, lip colors, expressions, etc) showcasing the diversity of the new latent space.",
+        "bbox": [
+            496,
+            592,
+            893,
+            893
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "2242",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/2daf20a2d6fe761dfa56132ec9e29b47d5dde8a8a850ca1615201afad50035ad.jpg",
+        "image_caption": [
+            "(a) GAN Inversion"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            160,
+            89,
+            475,
+            276
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/82a7d3c0e94e4dd8570ef6f006ec5b47b3e61d5cfec59d603b9e10f242c75ac2.jpg",
+        "image_caption": [
+            "(b) 3D Human Pose Reconstruction"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            483,
+            90,
+            805,
+            275
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/eb62ba9e9a69d0902c6842a6553b58d24853ec7e130a36463ff2845add34e314.jpg",
+        "image_caption": [
+            "Figure 4: Optimization Performance. We visualize the trends of optimization performance compared with the baseline. In GAN Inversion (Left), we evaluate all models on test splits of CelebA-HQ [34] and LSUN-cat [77] (OOD) with loss defined in Eq. 6. Since encoder-based inference doesn't involve optimization, we use a flat line to represent it. We perform 2000 steps of gradient descent for all models except encoder-based models. In 3D Human Pose Reconstruction (Right), we evaluate all models on test splits of GRAB [64] and PROX [27] (OOD) with loss defined in Eq. 7. We perform 200 steps of gradient descent for all models. For each step, we plot the average loss value of test splits.",
+            "Figure 5: Optimization Process for GAN Inversion. Comparing optimization process of our method and the baseline in order to reconstruct the ground truth image. Left shows the results from the baseline where optimization is done in the original input space  $X$ . Middle shows the results from our method where optimization is done in our space  $Z$ . Right column contains the ground truth image to each example. Each row corresponds to the same example."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            142,
+            367,
+            831,
+            612
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2.3D Human Pose Reconstruction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            683,
+            359,
+            698
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In addition to image generation, our framework also works for 3D reconstruction. For this, we use VPoser [53] – a variational autoencoder [37] that has learnt a pose prior over body pose. VPoser was trained on SMPL-X body pose parameters  $y \\in \\mathbb{R}^{63}$  obtained by applying MoSh [42] on three publicly available human motion capture datasets: CMU Mocap [16], training set of Human3.6M [30], and the PosePrior dataset [3].",
+        "bbox": [
+            73,
+            705,
+            470,
+            825
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We take a pretrained VPoser decoder denoted as  $F$ . Our trained mapping network  $\\Theta$  projects a vector from the new input space  $\\mathbf{z} \\in Z$  to a vector in the original VPoser decoder's input space  $\\mathbf{x} \\in X$ . Similar to GAN Inversion, we optimize the objective of Eq. 5, where the loss function",
+        "bbox": [
+            75,
+            827,
+            472,
+            905
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "between predicted and ground truth pose parameters is,",
+        "bbox": [
+            498,
+            684,
+            864,
+            700
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL (\\hat {y}, y) = \\left\\| \\hat {y} - y \\right\\| _ {2} ^ {2} \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            625,
+            710,
+            893,
+            729
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\hat{y} = F(\\hat{x})$  for the baseline and  $\\hat{y} = F(\\Theta(\\hat{z}))$  for ours. For training  $\\Theta$ , we use the GRAB dataset [64] which contains poses of humans interacting with everyday objects. We construct splits for novel video sequences – thus the test split will contain a seen human subject but a potentially unseen pose / demonstration by that subject. We evaluate on this test split for in-distribution experiments and on the PROX dataset [27] for OOD experiments, which contains poses of humans interacting in 3D scenes (e.g., living room, office, etc).",
+        "bbox": [
+            496,
+            738,
+            893,
+            888
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Quantitative Results. For SMPL-X human pose recon",
+        "bbox": [
+            517,
+            890,
+            893,
+            906
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "2243",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/583c9ad7156d0677d7f63e45a5447bf6c025fb747bbcfdadd1c56eac1fd1e3a5.jpg",
+        "image_caption": [
+            "Figure 6: Optimization Process for 3D Human Pose Reconstruction. Results shown are for out-of-distribution PROX dataset for sitting (Top) and standing (Bottom) poses."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            138,
+            74,
+            830,
+            332
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "struction experiment, the results follow a similar trend as GAN inversion, with massive improvement in both convergence speed and final loss values after convergence (see Fig. 4b). Our method outperforms the baseline by  $19\\%$  for in-distribution evaluation and  $11\\%$  for OOD evaluation.",
+        "bbox": [
+            75,
+            375,
+            470,
+            450
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Qualitative Results. In Fig. 6 we visualize how the human pose reconstructions evolve over the process of optimization. Here, we observe that the reconstructions from steps 0 to 6 of the baseline are similar for both examples. On the other hand, our method caters to fast convergence for the varying examples, highlighting the general, yet efficient properties of our search space. Further, in Fig. 7 we visualize reconstructions from partial observations where the only joints visible are that of the upper body (row 3) or lower body (row 4). We obtain a wide range of feasible possibilities for the hidden joints demonstrating the diversity of the latent space.",
+        "bbox": [
+            75,
+            450,
+            473,
+            632
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Defending Adversarial Attacks",
+        "text_level": 1,
+        "bbox": [
+            76,
+            640,
+            352,
+            656
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Our method is also applicable to discriminative models. A state-of-the-art defense [45] for adversarial attack optimizes the self-supervision task at inference time, such that the model can dynamically restore the corrupted structure from the input image for robust inference. Following the existing algorithm implementation, we optimize the input image via our method by minimizing the following discriminative loss function:",
+        "bbox": [
+            75,
+            662,
+            468,
+            785
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} L (F (\\mathbf {r} + \\mathbf {a}), y) = L (F (\\Theta (\\mathbf {z}) + \\mathbf {a}), y) \\\\ = \\mathbb {E} _ {i, j} \\left[ - y _ {i j} ^ {(s)} \\log \\frac {\\exp (\\cos (\\mathbf {f} _ {i} , \\mathbf {f} _ {j}))}{\\sum_ {k} \\exp (\\cos (\\mathbf {f} _ {i} , \\mathbf {f} _ {k}))} \\right] + \\lambda | | \\Theta (\\mathbf {z}) | | _ {2} ^ {2}, \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            89,
+            792,
+            455,
+            849
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{a}$  is the adversarial attacks that we aim to defend against,  $\\mathbf{r} = \\Theta (\\mathbf{z})$  is our additive defense vector to optimize,  $\\mathbf{f}_i$  are the contrastive features produced by the neural network",
+        "bbox": [
+            76,
+            854,
+            470,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "$F$  from the  $i^{\\mathrm{th}}$  image instance, and  $\\mathbf{y}_{ij}^{(s)}$  is the indicator for the positive pairs and negative pairs.",
+        "bbox": [
+            498,
+            375,
+            890,
+            406
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "After obtaining the mapping network  $\\Theta$  and the input variable  $\\hat{\\mathbf{z}}$ , the prediction is  $\\hat{\\mathbf{y}} = F'(\\mathbf{a} + \\Theta (\\mathbf{z}))$ , where  $F^{\\prime}$  is the classification model. Note that the self-supervision loss is optimized as a proxy for increasing the robust classification accuracy. In addition, we add a  $L_{2}$  norm decay term for the generated noise  $z$  to avoid generating reversal vector that is too large.",
+        "bbox": [
+            496,
+            406,
+            892,
+            512
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/4558638513fc0c2eb65fbd49d8071a116d781a8ff2e97de897e12d7dd675c482.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"3\">Model</td><td colspan=\"7\">Optimization Steps</td></tr><tr><td>None</td><td colspan=\"2\">1 step</td><td colspan=\"2\">3 steps</td><td colspan=\"2\">5 steps</td></tr><tr><td></td><td>BL</td><td>Ours</td><td>BL</td><td>Ours</td><td>BL</td><td>Ours</td></tr><tr><td>RO</td><td>31.99</td><td>34.62</td><td>44.65</td><td>36.77</td><td>44.23</td><td>38.38</td><td>43.43</td></tr><tr><td>AWP</td><td>35.61</td><td>39.54</td><td>51.39</td><td>42.81</td><td>51.67</td><td>44.96</td><td>51.05</td></tr><tr><td>MART</td><td>35.66</td><td>39.77</td><td>51.77</td><td>42.50</td><td>51.77</td><td>45.42</td><td>50.96</td></tr><tr><td>SemiSL</td><td>29.78</td><td>34.53</td><td>52.11</td><td>37.27</td><td>51.23</td><td>40.93</td><td>49.83</td></tr></table>",
+        "bbox": [
+            504,
+            523,
+            888,
+            609
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Quantitative Results. We evaluate our method on four popular pretrained robust models [58, 69, 73, 12] on CIFAR-10 dataset. The results in [45] require many steps to optimize the objective to improve the adversarial robustness, which slows down the inference by hundreds of times than the original forward pass. Ideally, we desire test-time optimization that can adapt to the attacked images in just one step,",
+        "bbox": [
+            496,
+            685,
+            893,
+            792
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/624b755753205044a76e21533495571167f9dc117854b1eb23495d16f5b5a1c6.jpg",
+        "table_caption": [
+            "Table 1: Experiment on improving adversarial robust accuracy. Our goal is to defend 200 steps of  $L_{2} = 256 / 255$  norm bounded attack [44], where the attack's step size is 64/255. Our baseline (BL) is the SOTA test-time optimization-based defense [45], which minimizes the loss of self-supervision task."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Defense Steps</td><td>0</td><td>1</td><td>3</td><td>5</td></tr><tr><td>Reversal (Baseline)</td><td>29.53</td><td>31.36</td><td>34.07</td><td>36.81</td></tr><tr><td>OBI (Ours)</td><td>29.53</td><td>44.48</td><td>46.84</td><td>50.42</td></tr></table>",
+        "bbox": [
+            514,
+            813,
+            877,
+            875
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Adaptive attack experiments with BPDA (SemiSL). When the attacker adapts to the defense, we can still improve robustness.",
+        "bbox": [
+            498,
+            876,
+            890,
+            905
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "2244",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/dcb86b669b6740b4ba9ea94d12806f2e3854d659d434a086ad906bc8f905d78a.jpg",
+        "image_caption": [
+            "Figure 7: Diversity of Masked Reconstructions. We visualize reconstructions for partially observable inputs from random initializations. The masked regions are not considered for loss computation, i.e., the gradient is set to be zero. By optimizing only on the partial observation, we obtain diverse, feasible solutions for the hidden regions. 5 or 4 out of 20 most representative samples are presented."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            151,
+            74,
+            821,
+            436
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/62729aff825b347aded27aa88e9ba7a231605d53bf0db4a15ac72f061549356f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Number of Steps</td><td>Full Model</td><td>Without AO and Buffer</td><td>Random Θ</td><td>Baseline</td></tr><tr><td>In-distribution</td><td>20</td><td>3.082</td><td>3.583</td><td>4.417</td><td>4.456</td></tr><tr><td>OOD</td><td>20</td><td>4.932</td><td>5.127</td><td>5.135</td><td>5.292</td></tr><tr><td>In-distribution</td><td>200</td><td>1.964</td><td>3.034</td><td>2.723</td><td>2.569</td></tr><tr><td>OOD</td><td>200</td><td>3.498</td><td>4.756</td><td>3.823</td><td>4.617</td></tr></table>",
+        "bbox": [
+            76,
+            492,
+            470,
+            594
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Ablation Study on mapping network. Number of Steps indicates the number of optimization steps performed during inference.",
+        "bbox": [
+            75,
+            604,
+            470,
+            647
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "causing the minimal delay at inference time. In Table 1, our method outperforms the gradient descent method in [45] by up to  $18\\%$  robust accuracy using a single step, providing a real-time robust inference method. In Table 2, even under the adaptive attack BPDA [6], our method also improves accuracy by up to  $14\\%$ . Note that our method converges after 1 step of optimization, demonstrating the effectiveness of our approach.",
+        "bbox": [
+            75,
+            661,
+            468,
+            782
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            76,
+            787,
+            230,
+            804
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present an ablation study by removing the proposed alternating optimization (AO) scheme and the experience replay buffer. We also compare against a  $\\Theta$  that is randomly initialized. From Table 3 in appendix, we discovered that for in-distribution, AO and training of Theta improves the optimization performance by  $14\\%$  and  $30\\%$",
+        "bbox": [
+            75,
+            805,
+            470,
+            896
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "respectively. Such gap becomes  $35\\%$  and  $28\\%$  for evaluation on 200 steps. For OOD data, the advantage is furthered enlarged as shown in Table 3.",
+        "bbox": [
+            496,
+            494,
+            892,
+            541
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "One surprising result we discovered experimentally is that OBI under a randomly initialized mapping network  $\\Theta$  consistently outperforms the baseline. We believe this is due to the fact that adding a Gaussian distribution to an underlying latent distribution of StyleGAN is beneficial in smoothing out loss landscape, making it easier to perform OBI. Similar random projection can also be found in [74].",
+        "bbox": [
+            496,
+            542,
+            893,
+            648
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.5. Loss Landscape",
+        "text_level": 1,
+        "bbox": [
+            500,
+            662,
+            660,
+            679
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To understand the underlying cause of the significant improvement in optimization brought by our mapping network  $\\theta$ , we visualize in Fig. 9 the loss landscape for performing optimization in the original input space  $X$ , and our projected space  $Z$ . To generate this visualization, we first perform 20 steps of optimization on the validation dataset to collect a set of recovered latents. We then perform principle component analysis (PCA) on these recovered latents to obtain two principle directions. Finally, for individual examples, we evaluate the loss for vertices on a meshgrid spanned by the two principle directions where the center is the last step of optimization.",
+        "bbox": [
+            496,
+            686,
+            893,
+            868
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "From the visualization, we can see that the loss landscapes of the baseline are highly non-convex and contain",
+        "bbox": [
+            500,
+            869,
+            893,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "2245",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4989d1af3ecdd68a877d72ca94a3441b4a17d62a06c6ce8e7b5f208a9c0b1cab.jpg",
+        "image_caption": [
+            "In-Distribution",
+            "Encoder",
+            "Ours",
+            "Ground Truth"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            161,
+            107,
+            480,
+            354
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            375,
+            191,
+            480,
+            354
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/d1a31fd19b5cdb469c4da96c60e6ee20716775786cb911d3761d9fd6c4f496dd.jpg",
+        "image_caption": [
+            "Out-of-Distribution",
+            "Encoder",
+            "Ours",
+            "Ground Truth"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            488,
+            107,
+            807,
+            354
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/53b4913a7839fced8f224bb719df4235681cc332cf2ddb564415a29b685dd227.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            158,
+            416,
+            320,
+            512
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b3b1463c7a5a2c1e7915da3c1b95e5ebf1a627e742c368d6dcd1d0123675edec.jpg",
+        "image_caption": [
+            "Figure 8: Comparison Against Encoder-Based Inference. Left shows the results on the test split of CelebA-HQ; Right shows the results on the LSUN-cat dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            321,
+            415,
+            480,
+            512
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/69afcfa8539fd65e8bb4e8be3a416b9ca34ddff77f40cf1a285b5eed6aa4f637.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            482,
+            415,
+            640,
+            512
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/0b15995df62b7fd6efe12425d4f1aaa5f0bf943796309e5ac13111bc299506ca.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            642,
+            415,
+            803,
+            512
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/37f9e7d3fab8c35a691240b0efe20708bd672bca55d45784606a0450aca620da.jpg",
+        "image_caption": [
+            "Figure 9: Visualizing Loss Landscape (Uncurated). Visualizing the loss landscape of StyleGAN inversion spanned by two principle directions. Top row shows 4 examples of the loss landscapes corresponding to our space  $Z$ . Bottom row shows the loss landscapes corresponding to the original input space  $X$  for the same 4 examples. Note that the top row and the bottom row have different minimum loss values on the landscape because the minimas are obtained by performing independent optimization runs in space  $Z$  and  $X$  respectively. Given a fixed number of optimization steps, optimization in  $Z$  reaches lower loss values than optimization in  $X$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            156,
+            520,
+            321,
+            617
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/d97445046b79e18a36a67a3ceb47b8f90ec9093be9b1c95045d4d5e3968e4ab6.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            321,
+            521,
+            480,
+            616
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/83720abffbc6a5c377b720bd81e988339fce701fd224bb63921ab9f17f736a39.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            480,
+            521,
+            638,
+            616
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/51c5ab68c7753665d5fea9a93da9d269916673ee362f3194915de8f7ee83c30b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            638,
+            521,
+            803,
+            617
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "points whose loss are significantly higher than its neighboring regions, while our loss landscapes are significantly smoother, with the \"spikes\" removed. Besides, our loss landscapes also tend to be steeper than the baseline ones. These two phenomena directly cause our method to perform gradient descent faster and stabler.",
+        "bbox": [
+            75,
+            712,
+            470,
+            803
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Analysis",
+        "text_level": 1,
+        "bbox": [
+            76,
+            820,
+            171,
+            835
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.1. L-BFGS",
+        "text_level": 1,
+        "bbox": [
+            76,
+            845,
+            179,
+            859
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "After learning the mapping network  $\\theta$ , in addition to the main results presented in Fig. 4a, we also run the same ex",
+        "bbox": [
+            76,
+            869,
+            470,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "periments but with L-BFGS as optimizer [40], which tends to perform better empirically with a smoother loss landscape. As shown in table 5, after 10 steps of optimization, L-BFGS achieves much better results than Adam in our learned landscape, which is already significantly better than the baseline. Notably, applying L-BFGS in the original landscape does not achieve significant speedup, and for Celeb-HQ, around  $18\\%$  of the examples failed to converge due to instability. These results further verify that our method successfully learns a smoother landscape for efficient optimization.",
+        "bbox": [
+            496,
+            712,
+            893,
+            864
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "2246",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/1f056855aaebce7f361d28a23f0c3b2589590e8539d156dc7f46056f9a42af9c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>level of Corruption</td><td>N(0,0.02)</td><td>N(0,0.12)</td><td>N(0,0.22)</td><td>N(0,0.32)</td><td>N(0,0.42)</td></tr><tr><td>Improvement</td><td>29.53%</td><td>18.95%</td><td>12.57%</td><td>6.30%</td><td>6.72%</td></tr></table>",
+        "bbox": [
+            179,
+            88,
+            787,
+            135
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/0b37228c72ca54287adb99aa5e9103054a9715204e9a921f47fefc2b2d4fd48a.jpg",
+        "table_caption": [
+            "Table 4: Evaluation of a mapping network trained with Celeba-HQ trainset and tested on the Celeba-HQ testset corrupted with different levels of Gaussian noise. The improvement is calculated by the percentage improvement of loss defined by Eq. 6 from baseline (no mapping network) to ours (with mapping network) after 200 steps of optimization."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>dataset</td><td>CelebA-HQ</td><td>AFHQ-Cat</td><td>LSUN-Cat</td><td>Container-Ship</td><td>N(0.5, 0.52)</td></tr><tr><td>Improvement</td><td>29.53%</td><td>35.99%</td><td>24.36%</td><td>16.87%</td><td>6.60%</td></tr></table>",
+        "bbox": [
+            176,
+            204,
+            794,
+            250
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/aacf76d00fd80c3b90c304c11695ad7ea8418a709a9e04b0a59aa5bbb4febaec.jpg",
+        "table_caption": [
+            "Table 6: Evaluation of a mapping network trained with Celeba-HQ trainset and tested on a spectrum of datasets from the original CelebA-HQ (in-distribution) to Gaussian noise (OOD). AFHQ-Cat [15] is a dataset with aligned cat faces. LSUN-Cat [77] is a dataset with unaligned cat images. Container-Ship is 200 images sampled from \"Container Ship\" class of ImageNet [20]. The improvement is calculated by the percentage improvement of loss defined by Eq. 6 from baseline (without  $\\theta$ ) to ours (with  $\\theta$ ) after 200 steps of optimization."
+        ],
+        "table_footnote": [
+            "Table 5: L-BFGS vs. Adam in Leanred Landscape (Ours)"
+        ],
+        "table_body": "<table><tr><td></td><td>ours\nL-BFGS</td><td>ours\nAdam</td><td>baseline\nL-BFGS</td><td>baseline\nAdam</td></tr><tr><td>Human</td><td>2.239</td><td>3.504</td><td>N/A</td><td>4.875</td></tr><tr><td>Cat (OOD)</td><td>4.066</td><td>5.087</td><td>5.227</td><td>5.475</td></tr></table>",
+        "bbox": [
+            91,
+            318,
+            454,
+            395
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.2. More Evaluation on OOD Generalization",
+        "text_level": 1,
+        "bbox": [
+            76,
+            416,
+            431,
+            431
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Since our proposed mapping network  $\\Theta$  is parameterized by a neural network, there's no guarantee that the learned mapping function is surjective. Therefore, we empirically study the generalization performance by testing a mapping network trained on CelebA-HQ against a spectrum of datasets from very similar ones (in-distribution) to completely different ones (OOD).",
+        "bbox": [
+            75,
+            439,
+            470,
+            545
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Synthetic Spectrum We first created a synthetic version of this spectrum of datasets by varying the level of Gaussian noise injected into the images of CelebA-HQ. With a higher level of Gaussian noise injected into the original images, more of the original image content is corrupted, creating a distribution of images further away from the original images.",
+        "bbox": [
+            75,
+            545,
+            468,
+            636
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Natural Image Spectrum We also created a natural image version of this spectrum of datasets containing: CelebA-HQ (original in-distribution testset), AFHQ-Cat (center-aligned cat faces), LSUN-Cat (unaligned cat images), Container-Ship (\"Container Ship\" class from ImageNet), and Gaussian noise with individual pixel sampled from  $\\mathcal{N}(0.5, 0.5^2)$ . From left to right, images vary from very similar to human faces to very different.",
+        "bbox": [
+            75,
+            636,
+            470,
+            756
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "From results evaluated on both synthetic spectrum and real spectrum, we observed consistent improvements of our method over the baseline, which shows the generalization performance of our method when applied to OOD data. From table 6, we see the improvements drop as the evaluation data is less and less in-distribution with the training data, which indicates that the mapping network does learn a prior from the training data.",
+        "bbox": [
+            75,
+            757,
+            470,
+            878
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/6e592a18fdce047f8fc40b97a05228ac9702873e8f2b302df7c6efe2f390af9a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Baseline</td><td>Ours</td><td>Improvement</td></tr><tr><td>CelebA-HQ</td><td>2.569</td><td>2.396</td><td>6.72%</td></tr><tr><td>LSUN-Cat</td><td>4.617</td><td>3.717</td><td>19.49%</td></tr></table>",
+        "bbox": [
+            524,
+            329,
+            867,
+            392
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Evaluation of a mapping network trained with images of Gaussian noise sampled from  $\\mathcal{N}(0.5,0.5^2)$  clipped to [0, 1]. Numbers show the average loss defined in Eq. 6 evaluated on testset of CelebA-HQ and LSUN-Cat.",
+        "bbox": [
+            498,
+            392,
+            893,
+            449
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.3. Dependence on Training Dataset",
+        "text_level": 1,
+        "bbox": [
+            500,
+            460,
+            785,
+            477
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "From previous section, we know that the learned mapping network contains priors learned from the training data. To exclude such influence, we train our mapping network using randomly generated images made of Gaussian noise and evaluate on testset of CelebA-HQ. From table 7, we see that even trained with the task of reconstructing images of Gaussian noise, our mapping network still bring some improvement over the baseline model, though less significant.",
+        "bbox": [
+            496,
+            484,
+            893,
+            606
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            618,
+            617,
+            633
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This paper presents a method to accelerate optimization-based inference to invert a forward model. We propose an approach that learns a new space that is easier than the original input space to optimize with gradient descent at testing time. Our experiments and analysis on three different applications in computer vision have shown that by learning this mapping function, optimization becomes more efficient and generalizes better to out-of-distribution data. Through quantitative and qualitative analysis, we found that such improvement in optimization performance comes from a more efficient loss landscape. Since optimization-based inference has many advantages over encoder-based inference, we believe methods to accelerate them will have many impacts in a variety of applications.",
+        "bbox": [
+            496,
+            643,
+            893,
+            853
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement This research is based on work partially supported by the NSF NRI Award 1925157, NSF STC LEAP, DARPA MCS program, and DARPA CCU program.",
+        "bbox": [
+            498,
+            854,
+            893,
+            902
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "2247",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            174,
+            104
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 1, 2, 4",
+            "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2",
+            "[3] Ijaz Akhter and Michael Black. Pose-conditioned joint angle limits for 3d human pose reconstruction. 06 2015. 5",
+            "[4] Brandon Amos, Lei Xu, and J Zico Kolter. Input convex neural networks. In International Conference on Machine Learning, pages 146-155. PMLR, 2017. 2",
+            "[5] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas. Learning to learn by gradient descent by gradient descent. Advances in neural information processing systems, 29, 2016. 3",
+            "[6] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 7",
+            "[7] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting largescale models. arXiv preprint arXiv:2203.17274, 1(3):4, 2022. 2",
+            "[8] David Bau, Jun-Yan Zhu, Jonas Wulff, William Peebles, Hendrik Strobelt, Bolei Zhou, and Antonio Torralba. Seeing what a gan cannot generate. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4502-4511, 2019. 2",
+            "[9] Ashish Bora, Ajil Jalal, Eric Price, and Alexandros G Dimakis. Compressed sensing using generative models. In International Conference on Machine Learning, pages 537-546. PMLR, 2017. 2",
+            "[10] Philémon Brakel, Dirk Stroobandt, and Benjamin Schrauwen. Training energy-based models for time-series imputation. The Journal of Machine Learning Research, 14(1):2771-2797, 2013. 1",
+            "[11] Nicholas Carlini and David A. Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy, pages 39-57, 2017. 2",
+            "[12] Yair Carmon, Aditi Raghunathan, Ludwig Schmidt, John C Duchi, and Percy S Liang. Unlabeled data improves adversarial robustness. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 6",
+            "[13] Boyuan Chen, Robert Kwiatkowski, Carl Vondrick, and Hod Lipson. Full-body visual self-modeling of robot morphologies. arXiv preprint arXiv:2111.06389, 2021. 1",
+            "[14] Tianlong Chen, Xiaohan Chen, Wuyang Chen, Howard Heaton, Jialin Liu, Zhangyang Wang, and Wotao Yin. Learning to optimize: A primer and a benchmark. arXiv preprint arXiv:2103.12828, 2021. 3"
+        ],
+        "bbox": [
+            78,
+            114,
+            472,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[15] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 9",
+            "[16] CMU. Cmu mocap dataset. 5",
+            "[17] Jochen L Cremer, Ioannis Konstantelos, and Goran Strbac. From optimization-based machine learning to interpretable security rules for operation. IEEE Transactions on Power Systems, 34(5):3826-3836, 2019. 1",
+            "[18] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 2",
+            "[19] Katherine Crowson, Stella Rose Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. ArXiv, abs/2204.08583, 2022. 2",
+            "[20] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 9",
+            "[21] Tan M Dinh, Anh Tuan Tran, Rang Nguyen, and Binh-Son Hua. Hyperinverter: Improving stylegan inversion via hypernetwork. arXiv preprint arXiv:2112.00719, 2021. 2",
+            "[22] Justin Domke. Generic methods for optimization-based modeling. In Artificial Intelligence and Statistics, pages 318-326. PMLR, 2012. 1",
+            "[23] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2",
+            "[24] Chelsea Finn, Aravind Rajeswaran, Sham Kakade, and Sergey Levine. Online meta-learning. In International Conference on Machine Learning, pages 1920-1930. PMLR, 2019. 2",
+            "[25] Chelsea Finn, Kelvin Xu, and Sergey Levine. Probabilistic model-agnostic meta-learning. Advances in neural information processing systems, 31, 2018. 2",
+            "[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2",
+            "[27] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 5",
+            "[28] Carlos Hernandez, George Vogiatzis, and Roberto Cipolla. Multiview photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(3):548-554, 2008. 1",
+            "[29] Minyoung Huh, Richard Zhang, Jun-Yan Zhu, Sylvain Paris, and Aaron Hertzmann. Transforming and projecting images into class-conditional generative networks. In European Conference on Computer Vision, pages 17–34. Springer, 2020. 2",
+            "[30] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments."
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            912
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "2248",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 5",
+            "[31] Ali Jahanian, Lucy Chai, and Phillip Isola. On the\"steerability\" of generative adversarial networks. arXiv preprint arXiv:1907.07171, 2019. 2",
+            "[32] Yongcheng Jing, Yezhou Yang, Zunlei Feng, Jingwen Ye, Yizhou Yu, and Mingli Song. Neural style transfer: A review. IEEE transactions on visualization and computer graphics, 26(11):3365-3385, 2019. 2",
+            "[33] Kyoungkook Kang, Seongtae Kim, and Sunghyun Cho. Gan inversion for out-of-range images with geometric transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13941-13949, 2021. 2",
+            "[34] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 4, 5",
+            "[35] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 4",
+            "[36] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2",
+            "[37] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5",
+            "[38] Kyoung Mu Lee and C-CJ Kuo. Shape from shading with a linear triangular element surface model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 15(8):815-822, 1993. 1",
+            "[39] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. ArXiv, abs/2104.08691, 2021. 2",
+            "[40] Dong C Liu and Jorge Nocedal. On the limited memory bfgs method for large scale optimization. Mathematical programming, 45(1-3):503-528, 1989. 8",
+            "[41] Xingchao Liu, Chengyue Gong, Lemeng Wu, Shujian Zhang, Haoran Su, and Qiang Liu. Fusedream: Training-free text-to-image generation with improved clip+gan space optimization. ArXiv, abs/2112.01573, 2021. 2",
+            "[42] Matthew Loper, Naureen Mahmood, and Michael J. Black. Mosh: Motion and shape capture from sparse markers. ACM Trans. Graph., 33(6), nov 2014. 5",
+            "[43] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 1",
+            "[44] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 2, 6",
+            "[45] Chengzhi Mao, Mia Chiquier, Hao Wang, Junfeng Yang, and Carl Vondrick. Adversarial attacks are reversible with natural supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 661-671, 2021. 1, 2, 6, 7"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[46] Chengzhi Mao, Revant Teotia, Amrutha Sundar, Sachit Menon, Junfeng Yang, Xin Wang, and Carl Vondrick. Doubly right object recognition: A why prompt for visual rationales. arXiv preprint arXiv:2212.06202, 2022. 2",
+            "[47] Chengzhi Mao, Lingyu Zhang, Abhishek Joshi, Junfeng Yang, Hao Wang, and Carl Vondrick. Robust perception through equivariance. arXiv preprint arXiv:2212.06079, 2022. 2",
+            "[48] Chengzhi Mao, Ziyuan Zhong, Junfeng Yang, Carl Vondrick, and Baishakhi Ray. Metric learning for adversarial robustness. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2",
+            "[49] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 2437-2445, 2020. 1",
+            "[50] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Ried-miller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. 3",
+            "[51] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 3",
+            "[52] Xingang Pan, Xiaohang Zhan, Bo Dai, Dahua Lin, Chen Change Loy, and Ping Luo. Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2",
+            "[53] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 1, 5",
+            "[54] Guim Perarnau, Joost Van De Weijer, Bogdan Raducanu, and Jose M Álvarez. Invertible conditional gans for image editing. arXiv preprint arXiv:1611.06355, 2016. 2",
+            "[55] Stanislav Pidhorskyi, Donald A Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14104-14113, 2020. 2",
+            "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2",
+            "[57] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. 2016. 3",
+            "[58] Leslie Rice, Eric Wong, and J. Zico Kolter. Overfitting in adversarially robust deep learning, 2020. 6",
+            "[59] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 2, 4",
+            "[60] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 2"
+        ],
+        "bbox": [
+            501,
+            92,
+            892,
+            911
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "2249",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[61] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243–9252, 2020. 2",
+            "[62] Veselin Stoyanov, Alexander Ropson, and Jason Eisner. Empirical risk minimization of graphical model parameters given approximate inference, decoding, and model structure. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 725-733. JMLR Workshop and Conference Proceedings, 2011. 1",
+            "[63] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv:1312.6199, 2013. 2",
+            "[64] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In European conference on computer vision, pages 581-600. Springer, 2020. 5",
+            "[65] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 2",
+            "[66] Austin Tripp, Erik Daxberger, and José Miguel Hernández-Lobato. Sample-efficient optimization in the latent space of deep generative models via weighted retraining. Advances in Neural Information Processing Systems, 33:11259-11272, 2020. 2",
+            "[67] Yun-Yun Tsai, Chengzhi Mao, Yow-Kuan Lin, and Junfeng Yang. Self-supervised convolutional visual prompts. arXiv preprint arXiv:2303.00198, 2023. 2",
+            "[68] Hao Wang, Chengzhi Mao, Hao He, Mingmin Zhao, Tommi S Jaakkola, and Dina Katabi. Bidirectional inference networks: A class of deep bayesian networks for health profiling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 766-773, 2019. 1, 2",
+            "[69] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 6",
+            "[70] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Weiming Zhang, Lu Yuan, Gang Hua, and Nenghai Yu. A simple baseline for stylegan inversion. ArXiv, abs/2104.07661, 2021. 2",
+            "[71] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688. CiteSeer, 2011. 2",
+            "[72] Olga Wichrowska, Niru Maheswaranathan, Matthew W Hoffman, Sergio Gomez Colmenarejo, Misha Denil, Nando Freitas, and Jascha Sohl-Dickstein. Learned optimizers that scale and generalize. In International Conference on Machine Learning, pages 3751-3760. PMLR, 2017. 3",
+            "[73] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. In NeurIPS, 2020. 6",
+            "[74] Yan Wu, Mihaela Rosca, and Timothy Lillicrap. Deep compressed sensing. In International Conference on Machine Learning, pages 6850-6860. PMLR, 2019. 2, 3, 7"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            900
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[75] Raymond A Yeh, Chen Chen, Teck Yian Lim, Alexander G Schwing, Mark Hasegawa-Johnson, and Minh N Do. Semantic image inpainting with deep generative models. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5485-5493, 2017. 2",
+            "[76] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 1",
+            "[77] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 4, 5, 9",
+            "[78] Junzhe Zhang, Xinyi Chen, Zhongang Cai, Liang Pan, Haiyu Zhao, Shuai Yi, Chai Kiat Yeo, Bo Dai, and Chen Change Loy. Unsupervised 3d shape completion through gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1768-1777, 2021. 1",
+            "[79] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4",
+            "[80] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2",
+            "[81] Jun-Yan Zhu, Philipp Kähenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In European conference on computer vision, pages 597-613. Springer, 2016. 2"
+        ],
+        "bbox": [
+            501,
+            90,
+            893,
+            529
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "2250",
+        "bbox": [
+            482,
+            944,
+            516,
+            955
+        ],
+        "page_idx": 11
+    }
+]

2023/Landscape Learning for Neural Network Inversion/f1107abc-1d35-4283-b4db-d9fd38f81a61_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee99da27cdc693a7248ce42861519c67152e6be157461c040f4713054ea53690
+size 4728297

2023/Landscape Learning for Neural Network Inversion/full.md

@@ -0,0 +1,405 @@
+# Landscape Learning for Neural Network Inversion
+
+Ruoshi Liu<sup>1</sup> Chengzhi Mao<sup>1</sup> Purva Tendulkar<sup>1</sup> Hao Wang<sup>2</sup> Carl Vondrick<sup>1</sup>  
+<sup>1</sup> Columbia University <sup>2</sup> Rutgers University
+
+# Abstract
+
+Many machine learning methods operate by inverting a neural network at inference time, which has become a popular technique for solving inverse problems in computer vision, robotics, and graphics. However, these methods often involve gradient descent through a highly non-convex loss landscape, causing the optimization process to be unstable and slow. We introduce a method that learns a loss landscape where gradient descent is efficient, bringing massive improvement and acceleration to the inversion process. We demonstrate this advantage on a number of methods for both generative and discriminative tasks, including GAN inversion, adversarial defense, and 3D human pose reconstruction.
+
+# 1. Introduction
+
+Many inference problems in machine learning are formulated as inverting a forward model  $F(x)$  by optimizing an objective over the input space  $x$ . This approach, which we term optimization-based inference (OBI), has traditionally been used to solve a range of inverse problems in vision, graphics, robotics, recommendation systems, and security [28, 38, 22, 10, 62, 17]. Recently, neural networks have emerged as the paramterization of choice for forward models [43, 53, 1, 49, 76, 68, 13, 78], which can be pretrained on large collections of data, and inverted at testing time in order to solve inference queries.
+
+Optimization-based inference (OBI) has many advantages over feed-forward or encoder-based inference (EBI). Without encoder, OBI provides flexibility to adapt to new tasks, allowing one to define new constraints into the objective during inference. When observations are partially missing, OBI can adapt without additional training. Moreover, OBI naturally supports generating multiple and diverse hypotheses when there is uncertainty. Finally, OBI has intrinsic advantages for robustness, both adapting to new data distributions as well as defending against adversarial examples.
+
+However, the key bottleneck for OBI in practice is the computational efficiency and the speed of inference. Feedforward models are fast because they only require a single forward pass of a neural network, but OBI requires many
+
+![](images/5886baaa4e06470f25ae06068a39dd9120e390fe68b37291340f97a2996cbbf0.jpg)  
+Figure 1: Loss Landscapes Comparison. The loss landscape of optimization-based inference (OBI) is often highly non-convex. We propose to learn a smoother loss landscape through a mapping network to accelerate the optimization. Plotted from real data.
+
+![](images/ea580b0c066fa2958e2c0629e07c010bdcb00c7fb1d73626bf901eaa0425efd6.jpg)
+
+(often hundreds) steps of optimization in order to obtain strong results for one example. Forward models in OBI are often trained with generative or discriminative tasks, but they are not trained for the purpose of performing gradient descent in the input space. Fig. 9 visualizes the loss landscape for uncurated examples. The loss landscape is not guaranteed to be an efficient path from the initialization to the solution, causing the instability and inefficiency.
+
+In this paper, we propose a framework to accelerate and stabilize the inversion of forward neural networks. Instead of optimizing over the original input space, we learn a new input space such that gradient descent converges quickly. Our approach uses an alternating algorithm to learn the mapping between these spaces. The first step collects optimization trajectories in the new space until convergence. The second step updates the mapping parameters to reduce the distance between the convergence point and each point along the trajectory. By repeating these steps, our approach will learn a mapping between the spaces that allows gradient descent over the input to converge in significantly fewer steps.
+
+Empirical experiments and visualizations on both generative and discriminative models show that our method can largely improve the convergence speed for optimization. We validate our approach on a diverse set of vision tasks, including GAN inversion [1], adversarial defense [45], and 3D human pose reconstruction [53]. Our experiments show that our method converges an order of magnitude faster without loss in absolute performance after convergence. As our ap
+
+![](images/6a230d29c2b6a3599d5ea8047c1a3131a7bc79ba84216272c6c01daec7e57e56.jpg)  
+Figure 2: Method. The left and middle figure show the loss landscape for our latent space and the original latent space, respectively. While walking to the optimal solution in a few steps is hard in  $X$  space, it can be done in our learned loss landscapes.
+
+proach does not require retraining of the forward model, it is compatible to all existing OBI methods with a differentiable forward model and objective function.
+
+The primary contribution of this paper is an efficient optimization-based inference framework. In Sec. 2, we survey the related literature to provide an overview of forward model inversion problem. In Sec. 3, we formally define OBI (3.1); our method to learn an efficient loss landscape for OBI (3.2); a training algorithm for better generalization and robustness (3.3). In Sec. 4, we experimentally study and analyze the effectiveness of the mapping network for OBI.
+
+# 2. Related Work
+
+The different approaches for inference with a neural network can be partitioned into either encoder-based inference, which is feedforward, or optimization-based inference, which is iterative. We briefly review these two approaches in the context of our work. In addition, we also review representative work in meta-learning and discuss the similarities and differences with our work.
+
+# 2.1. Encoder-based Inference
+
+Encoder-based inference trains a neural network  $F$  to directly map from the output space to the input space. Autoencoder based approach [55] learns an encoder that maps the input data to the latent space. [59, 65, 70, 54] learn an encoder from the image to the latent space in GAN. Encoder based inference requires training the encoder on the anticipated distribution in advance, which is often less effective and can fail on unexpected samples [21, 33]. In addition, encoder-based method can only produce one reconstruction for under-constrained inverse problems [75], even if multiple outcomes can all be true given the partial input.
+
+# 2.2. Optimization-based Inference
+
+OBI methods perform inference by solving an optimization problem with gradient-based methods such as Stochastic Gradient Descent (SGD) [60] and Projected Gradient Descent (PGD) [44]. In these cases, the objective function specifies the inference task. Besides these methods which use a point estimate for the latent variable, one can estimate the posterior distribution of the latent variable through Bayesian optimization, such as SGLD [71].
+
+Gradient based optimization methods have been used to infer the latent code of query samples in deep generative models like GANs [26] via GAN inversion [36, 31, 61, 81, 1, 2, 8, 29, 52]. Style transfer relies on gradient based optimization to change the style of the input images [32]. It can also create adversarial attacks that fool the classifier [18, 11, 48, 63]. Recently, back-propagation-based optimization has shown to be effective in defending adversarial examples [45, 47] and compressed sensing [9]. [74] uses MAML [23] to accelerate the optimization process in compressed sensing.
+
+Recently, constrained optimization was popularized for text-to-image synthesis by [19, 41]. They search in the latent space to produce an image that has the highest similarity with the given text as measured by a multi-modal similarity model like CLIP [56]. Test-time constrained optimization is also related to the idea of 'prompt-tuning' for large language models and vision-language models. [39] learns "soft prompts" to condition frozen language models to perform specific downstream tasks. Soft prompts are learned through backpropagation to incorporate signal from just a few labeled examples (few-shot learning). Visual prompting at both training time [7, 46] and testing time [67] also relies on effective latent variable optimization.
+
+A major challenge for optimization-based inference is how to perform efficient optimization in a highly non-convex space. To address this, input convex model [4] was proposed so that gradient descent can be performed in a convex space. [66] introduced a method to retrain the generative model such as it learns a latent manifold that is easy to optimize. When the model cannot be changed and retrained, bidirectional inference [68] and hybrid inference [81, 80] uses an encoder to provide a good initialization for the optimization-based inference in a non-convex space. Our method does not retrain the generative model, but instead maps a new latent space into the original input space, allowing more efficient optimization.
+
+# 2.3. Meta-Learning
+
+Given a distribution of tasks, meta-learning aims to adapt quickly when presented with a previously unseen task. MAML [23] and related methods [25, 24] propose a method to learn a parameter initialization of a neural net
+
+work by differentiating through the fine-tuning process. To reduce the computational cost of MAML due to the 2nd degree gradient, [51] proposes a first-order meta-learning algorithm. Unlike MAML which include modifying the forward model (e.g. in [74]), our approach is able to keep the forward model fixed in order to maintain its learned rich priors. Instead of meta-learning of model initialization, a line of work [5, 57, 14, 72] proposed to learn an optimizer, often in the form of an RNN, in order to accelerate gradient descent. Unlike learned optimizers that try to create better optimization algorithms, our approach instead learns to remap the loss landscape which is compatible with any choice of optimizer, including standard SGD or learned ones.
+
+# 3. Learning Landscapes for Efficient Inference
+
+We present our framework to learn an efficient loss landscape for optimization-based inference (OBI) methods. In Sec. 3.1, we will define OBI. In Sec. 3.2, we will introduce our framework as well as the training objective. In Sec. 3.3, we will describe how to train our model with an alternating optimization algorithm and an experience-replay buffer.
+
+# 3.1. Optimization-based Inference
+
+Let  $F(\mathbf{x}) = \hat{y}$  be a differentiable forward model that generates an output  $\hat{y}$  given an input variable  $\mathbf{x} \in X$ . For example,  $\hat{y}$  might be an image, and  $\mathbf{x}$  might be the latent variables for a generative model. Given an observation  $y$ , the goal of OBI is to find the input  $\hat{\mathbf{x}} \in X$  such that an objective function  $L(\hat{y}, y)$  is minimized. Formally, we write this procedure as:
+
+$$
+\hat {\mathbf {x}} = \underset {\mathbf {x} \in X} {\operatorname {a r g m i n}} L (F (\mathbf {x}), y) \tag {1}
+$$
+
+When the objective function  $L$  and the model  $F$  are both differentiable, we can perform the optimization over input space  $X$  with gradient descent.
+
+# 3.2. Remapping the Input Space
+
+Instead of operating in the original input space  $X$ , we will create a new space  $Z$  where gradient descent is efficient and converges in a small number of iterations. We will use a neural network  $\Theta: Z \to X$  that maps from the new space  $Z$  to the original space  $X$ . The parameters of the mapping  $\Theta$  is the vector  $\theta$ . The learning problem we need solve is to estimate  $\theta$  so that there is a short gradient descent path in  $Z$  from the initialization to the solution. Fig. 2 shows an overview of this setup.
+
+We formulate an objective by rolling out the gradient updates on  $\mathbf{z}$ , where we write  $\mathbf{z}_t \gets \mathbf{z}_{t-1} + \lambda \frac{\partial L}{\partial \mathbf{z}_{t-1}}$  as the  $t^{\text{th}}$  update with a step size of  $\lambda$ . For gradient descent in space  $Z$  to be efficient, the goal of our introduced  $\Theta$  is to move every step  $\mathbf{z}_t$  as close as possible to the global minima:
+
+![](images/69201484fa2b47fa9b0c4acbde6be430b7c869aedea1d4120233a92138ff2b2f.jpg)  
+Figure 3: Landscape Learning. An optimization trajectory  $\{\mathbf{z}_t\}_{t=0}^5$  collected is used to train  $\Theta$ .  $\mathbf{z}_i$  that corresponds to a higher  $L_i$  will yield a higher gradient when training  $\Theta$ . Optimization over multiple steps along the trajectory causes  $\Theta$  to learn patterns of trajectories and create a more efficient loss landscape.
+
+$$
+\hat {\theta} = \underset {\theta} {\operatorname {a r g m i n}} \mathbb {E} _ {(\mathbf {z}, y)} \left[ \sum_ {t = 0} ^ {T} L (F (\Theta (\mathbf {z} _ {t})), y) \right] \tag {2}
+$$
+
+$$
+\text {w h e r e} \quad \mathbf {z} _ {t} = \left\{ \begin{array}{l l} 0, & t = 0 \\ \mathbf {z} _ {t - 1} + \lambda \frac {\partial L}{\partial \mathbf {z} _ {t - 1}}, & t > 0 \end{array} \right.
+$$
+
+We visualize this process with a toy example in Fig. 3. Gradient updates on  $\Theta$  w.r.t multiple steps  $\mathbf{z}_t$  along a trajectory will cause the loss value on each step to be lowered. By learning the parameters of  $\Theta$  across many examples,  $\Theta$  can learn the patterns of optimization trajectories in  $X$ . For example,  $\Theta$  can learn to estimate the step size in  $X$  and dynamically adjust it to each example. Moreover,  $\Theta$  can learn to reduce non-convexity of the landscape.
+
+Once we obtain  $\hat{\Theta}$  with parameters  $\hat{\theta}$ , we do inference on a new example  $y$  through the standard optimization-based inference procedure, except in  $Z$  space now. Given the observation  $y$ , we find the corresponding  $\hat{\mathbf{x}}$  through the optimization:
+
+$$
+\hat {\mathbf {x}} = \hat {\Theta} (\hat {\mathbf {z}}) \quad \text {w h e r e} \quad \hat {\mathbf {z}} = \underset {\mathbf {z} \in Z} {\operatorname {a r g m i n}} L (F (\hat {\Theta} (\mathbf {z})), y) \tag {3}
+$$
+
+When the inverse problem is under-constrained, one can infer multiple hypotheses for  $\hat{\mathbf{x}}$  by repeating the above optimization with a different random initialization for  $\mathbf{z}_0$ .
+
+# 3.3. Training
+
+We use alternating optimization (AO) in order to train  $\Theta$  jointly with estimating  $\mathbf{z}$  for each example in the training set. Specifically, we first fix parameters of  $\Theta$  and collect  $N$  optimization trajectories of  $\mathbf{z}$ , each with length  $T$ . Adopting the same terminology from the robotics community for learning on continuous states [50], we term this an experience replay buffer. Subsequently, we randomly sample data from this buffer and train  $\Theta$  to optimize the loss function. We alternate between these two steps for a fixed number of iterations with gradient descent for both. Depending on the application, the training time for  $\Theta$  varied from one hour to
+
+one day using a four GPU server. Please see the appendix for more implementation details.
+
+We also experimented with an online version of the training algorithm, where we update parameters of  $\Theta$  immediately after one update to  $\mathbf{z}$ . However, in our experiments, we found this resulted in a slower convergence rate. We show these comparisons in the ablation experiments.
+
+Algorithm 1 Learning Mapping Network  $\Theta$  
+1: Input: Ground truth  $y$ , step size  $\lambda_z$  and  $\lambda_{\theta}$ , number of buffers  $B$ , number of data samples in a buffer  $N$ , number of optimization steps per sample  $T$ , loss function  $L$ , and forward model  $F$ .  
+2: Output: Learned mapping network  $\Theta$   
+3: Randomly initialize a mapping network  $\Theta$  with parameters  $\theta$   
+4: for  $b = 1, \dots, B$  do  
+5: Initialize Experience Replay Buffer  $[\{\mathbf{z}_{t,i}\}_{t=1}^{T}]_{i=1}^{N}$   
+6: for  $i = 1, \dots, N$  do  
+7:  $\mathbf{z}_0 \gets \mathbf{0}$   
+8: for  $t = 1, \dots, T$  do  
+9:  $\mathbf{z}_t \gets \mathbf{z}_{t-1} + \lambda_z \frac{\partial}{\partial \mathbf{z}_{t-1}} L(F(\Theta(\mathbf{z})), y)$   
+10:  $\mathbf{z}_{t,i} \gets z_t$   
+11: end for  
+12: end for  
+13: for  $j = 1, \dots, T \cdot N$  do  
+14: Randomly sample  $\mathbf{z}$  from previously collected buffer  
+15:  $\theta_j \gets \theta_{j-1} + \lambda_\theta \frac{\partial}{\partial \theta_{j-1}} L(F(\Theta(\mathbf{z})), y)$   
+16: end for  
+17: end for  
+18: Return  $\theta$
+
+# 4. Experiments
+
+The goal of our experiments is to analyze how well our proposed method can be applied to various existing OBI methods to improve the optimization efficiency. We demonstrate application of our method to three diverse OBI methods in computer vision, including both generative models and discriminative models. For each OBI method, the inference-time optimization objective of the baseline and ours can be written as:
+
+$$
+\text {B a s i l e :} \hat {\mathbf {x}} = \min  _ {\mathbf {x} \in X} L (F (\mathbf {x}), y) \tag {4}
+$$
+
+$$
+\text {O u r s :} \hat {\mathbf {z}} = \min  _ {\mathbf {z} \in Z} L (F (\Theta (\mathbf {z})), y) \tag {5}
+$$
+
+Next, we provide the specific implementation of the loss term  $L$  for each application, along with quantitative and qualitative results. We also perform experiments to understand the loss landscape in Sec. 4.5 and perform ablations on different parts of our approach in Sec. 4.4.
+
+# 4.1. GAN Inversion
+
+We first validate our method on StyleGAN inversion [1]. We take a pretrained generator of StyleGAN [35] denoted as  $F$ . Let  $y$  be an observed image whose input variable we are recovering, we optimize the objective of Eq. 5 where the loss can be written as,
+
+$$
+L (\hat {y}, y) = L _ {l p i p s} (\hat {y}, y) + \left\| \hat {y} - y \right\| _ {2} ^ {2} \tag {6}
+$$
+
+where  $L_{lpips}$  is a perceptual similarity loss introduced in [79],  $\hat{y} = F(\hat{x})$  for baseline and  $\hat{y} = F(\Theta(\hat{z}))$  for ours. We train  $\Theta$  on the train split of CelebA-HQ [34] dataset and evaluate on CelebA-HQ validation split for in-distribution experiments and LSUN-Cat[77] for distribution shifting (OOD) experiments. We compare the results from our method against the state-of-the-art encoder-based GAN inversion model [59].
+
+Quantitative Results. From Fig. 4a, we see that in all experiments, optimization in our space  $Z$  consistently outperforms the baseline from the first optimization step to after convergence. This gap in performance is even larger when evaluated on OOD data. This suggests that the improvement in performance is not caused by memorizing the training data. Note that our image reconstruction performance after only 2 steps of optimization is able to outperform or be on-par with the baseline at 20 steps of optimization, resulting in a 10-fold improvement in efficiency. Even after convergence (after 2000 optimization steps), our reconstruction performance improves over the baseline by  $15\%$  for in-distribution evaluation and  $10\%$  for OOD evaluation. When compared with encoder-based GAN inversion [59], our method achieves better reconstruction after 11 steps of optimization for in-distribution data and 3 steps for OOD data.
+
+Qualitative Results. From Fig. 8, we can see that our method already shows improvements on in-distribution data - it can almost perfectly reconstruct details like fine hair strands, the cap on the person's head, the object in the person's mouth as well-as the text on it. Interestingly, our method is able to discover and reconstruct latents for cats while the encoder-based model fails miserably as shown in Fig 8. The performance on OOD data truly highlights the benefits of our method. We also visualize how the face generations evolve over the process of optimization in Fig. 5. We can see that in just 4 steps, our method is already able to reconstruct coarse features such as face orientation, skin tone and hair color, while the baseline has hardly deviated from the initialization in regard to any of these features. Further, in Fig. 7 we visualize reconstructions from partial observations where only the center of the face (row 1) or everything other than the mouth (row 2) is visible. We can see a variety of feasible possibilities for the hidden regions (e.g., different hairstyles, lip colors, expressions, etc) showcasing the diversity of the new latent space.
+
+![](images/2daf20a2d6fe761dfa56132ec9e29b47d5dde8a8a850ca1615201afad50035ad.jpg)  
+(a) GAN Inversion
+
+![](images/82a7d3c0e94e4dd8570ef6f006ec5b47b3e61d5cfec59d603b9e10f242c75ac2.jpg)  
+(b) 3D Human Pose Reconstruction
+
+![](images/eb62ba9e9a69d0902c6842a6553b58d24853ec7e130a36463ff2845add34e314.jpg)  
+Figure 4: Optimization Performance. We visualize the trends of optimization performance compared with the baseline. In GAN Inversion (Left), we evaluate all models on test splits of CelebA-HQ [34] and LSUN-cat [77] (OOD) with loss defined in Eq. 6. Since encoder-based inference doesn't involve optimization, we use a flat line to represent it. We perform 2000 steps of gradient descent for all models except encoder-based models. In 3D Human Pose Reconstruction (Right), we evaluate all models on test splits of GRAB [64] and PROX [27] (OOD) with loss defined in Eq. 7. We perform 200 steps of gradient descent for all models. For each step, we plot the average loss value of test splits.  
+Figure 5: Optimization Process for GAN Inversion. Comparing optimization process of our method and the baseline in order to reconstruct the ground truth image. Left shows the results from the baseline where optimization is done in the original input space  $X$ . Middle shows the results from our method where optimization is done in our space  $Z$ . Right column contains the ground truth image to each example. Each row corresponds to the same example.
+
+# 4.2.3D Human Pose Reconstruction
+
+In addition to image generation, our framework also works for 3D reconstruction. For this, we use VPoser [53] – a variational autoencoder [37] that has learnt a pose prior over body pose. VPoser was trained on SMPL-X body pose parameters  $y \in \mathbb{R}^{63}$  obtained by applying MoSh [42] on three publicly available human motion capture datasets: CMU Mocap [16], training set of Human3.6M [30], and the PosePrior dataset [3].
+
+We take a pretrained VPoser decoder denoted as  $F$ . Our trained mapping network  $\Theta$  projects a vector from the new input space  $\mathbf{z} \in Z$  to a vector in the original VPoser decoder's input space  $\mathbf{x} \in X$ . Similar to GAN Inversion, we optimize the objective of Eq. 5, where the loss function
+
+between predicted and ground truth pose parameters is,
+
+$$
+L (\hat {y}, y) = \left\| \hat {y} - y \right\| _ {2} ^ {2} \tag {7}
+$$
+
+where  $\hat{y} = F(\hat{x})$  for the baseline and  $\hat{y} = F(\Theta(\hat{z}))$  for ours. For training  $\Theta$ , we use the GRAB dataset [64] which contains poses of humans interacting with everyday objects. We construct splits for novel video sequences – thus the test split will contain a seen human subject but a potentially unseen pose / demonstration by that subject. We evaluate on this test split for in-distribution experiments and on the PROX dataset [27] for OOD experiments, which contains poses of humans interacting in 3D scenes (e.g., living room, office, etc).
+
+Quantitative Results. For SMPL-X human pose recon
+
+![](images/583c9ad7156d0677d7f63e45a5447bf6c025fb747bbcfdadd1c56eac1fd1e3a5.jpg)  
+Figure 6: Optimization Process for 3D Human Pose Reconstruction. Results shown are for out-of-distribution PROX dataset for sitting (Top) and standing (Bottom) poses.
+
+struction experiment, the results follow a similar trend as GAN inversion, with massive improvement in both convergence speed and final loss values after convergence (see Fig. 4b). Our method outperforms the baseline by  $19\%$  for in-distribution evaluation and  $11\%$  for OOD evaluation.
+
+Qualitative Results. In Fig. 6 we visualize how the human pose reconstructions evolve over the process of optimization. Here, we observe that the reconstructions from steps 0 to 6 of the baseline are similar for both examples. On the other hand, our method caters to fast convergence for the varying examples, highlighting the general, yet efficient properties of our search space. Further, in Fig. 7 we visualize reconstructions from partial observations where the only joints visible are that of the upper body (row 3) or lower body (row 4). We obtain a wide range of feasible possibilities for the hidden joints demonstrating the diversity of the latent space.
+
+# 4.3. Defending Adversarial Attacks
+
+Our method is also applicable to discriminative models. A state-of-the-art defense [45] for adversarial attack optimizes the self-supervision task at inference time, such that the model can dynamically restore the corrupted structure from the input image for robust inference. Following the existing algorithm implementation, we optimize the input image via our method by minimizing the following discriminative loss function:
+
+$$
+\begin{array}{l} L (F (\mathbf {r} + \mathbf {a}), y) = L (F (\Theta (\mathbf {z}) + \mathbf {a}), y) \\ = \mathbb {E} _ {i, j} \left[ - y _ {i j} ^ {(s)} \log \frac {\exp (\cos (\mathbf {f} _ {i} , \mathbf {f} _ {j}))}{\sum_ {k} \exp (\cos (\mathbf {f} _ {i} , \mathbf {f} _ {k}))} \right] + \lambda | | \Theta (\mathbf {z}) | | _ {2} ^ {2}, \\ \end{array}
+$$
+
+where  $\mathbf{a}$  is the adversarial attacks that we aim to defend against,  $\mathbf{r} = \Theta (\mathbf{z})$  is our additive defense vector to optimize,  $\mathbf{f}_i$  are the contrastive features produced by the neural network
+
+$F$  from the  $i^{\mathrm{th}}$  image instance, and  $\mathbf{y}_{ij}^{(s)}$  is the indicator for the positive pairs and negative pairs.
+
+After obtaining the mapping network  $\Theta$  and the input variable  $\hat{\mathbf{z}}$ , the prediction is  $\hat{\mathbf{y}} = F'(\mathbf{a} + \Theta (\mathbf{z}))$ , where  $F^{\prime}$  is the classification model. Note that the self-supervision loss is optimized as a proxy for increasing the robust classification accuracy. In addition, we add a  $L_{2}$  norm decay term for the generated noise  $z$  to avoid generating reversal vector that is too large.
+
+<table><tr><td rowspan="3">Model</td><td colspan="7">Optimization Steps</td></tr><tr><td>None</td><td colspan="2">1 step</td><td colspan="2">3 steps</td><td colspan="2">5 steps</td></tr><tr><td></td><td>BL</td><td>Ours</td><td>BL</td><td>Ours</td><td>BL</td><td>Ours</td></tr><tr><td>RO</td><td>31.99</td><td>34.62</td><td>44.65</td><td>36.77</td><td>44.23</td><td>38.38</td><td>43.43</td></tr><tr><td>AWP</td><td>35.61</td><td>39.54</td><td>51.39</td><td>42.81</td><td>51.67</td><td>44.96</td><td>51.05</td></tr><tr><td>MART</td><td>35.66</td><td>39.77</td><td>51.77</td><td>42.50</td><td>51.77</td><td>45.42</td><td>50.96</td></tr><tr><td>SemiSL</td><td>29.78</td><td>34.53</td><td>52.11</td><td>37.27</td><td>51.23</td><td>40.93</td><td>49.83</td></tr></table>
+
+Quantitative Results. We evaluate our method on four popular pretrained robust models [58, 69, 73, 12] on CIFAR-10 dataset. The results in [45] require many steps to optimize the objective to improve the adversarial robustness, which slows down the inference by hundreds of times than the original forward pass. Ideally, we desire test-time optimization that can adapt to the attacked images in just one step,
+
+Table 1: Experiment on improving adversarial robust accuracy. Our goal is to defend 200 steps of  $L_{2} = 256 / 255$  norm bounded attack [44], where the attack's step size is 64/255. Our baseline (BL) is the SOTA test-time optimization-based defense [45], which minimizes the loss of self-supervision task.  
+
+<table><tr><td>Defense Steps</td><td>0</td><td>1</td><td>3</td><td>5</td></tr><tr><td>Reversal (Baseline)</td><td>29.53</td><td>31.36</td><td>34.07</td><td>36.81</td></tr><tr><td>OBI (Ours)</td><td>29.53</td><td>44.48</td><td>46.84</td><td>50.42</td></tr></table>
+
+Table 2: Adaptive attack experiments with BPDA (SemiSL). When the attacker adapts to the defense, we can still improve robustness.
+
+![](images/dcb86b669b6740b4ba9ea94d12806f2e3854d659d434a086ad906bc8f905d78a.jpg)  
+Figure 7: Diversity of Masked Reconstructions. We visualize reconstructions for partially observable inputs from random initializations. The masked regions are not considered for loss computation, i.e., the gradient is set to be zero. By optimizing only on the partial observation, we obtain diverse, feasible solutions for the hidden regions. 5 or 4 out of 20 most representative samples are presented.
+
+<table><tr><td></td><td>Number of Steps</td><td>Full Model</td><td>Without AO and Buffer</td><td>Random Θ</td><td>Baseline</td></tr><tr><td>In-distribution</td><td>20</td><td>3.082</td><td>3.583</td><td>4.417</td><td>4.456</td></tr><tr><td>OOD</td><td>20</td><td>4.932</td><td>5.127</td><td>5.135</td><td>5.292</td></tr><tr><td>In-distribution</td><td>200</td><td>1.964</td><td>3.034</td><td>2.723</td><td>2.569</td></tr><tr><td>OOD</td><td>200</td><td>3.498</td><td>4.756</td><td>3.823</td><td>4.617</td></tr></table>
+
+Table 3: Ablation Study on mapping network. Number of Steps indicates the number of optimization steps performed during inference.
+
+causing the minimal delay at inference time. In Table 1, our method outperforms the gradient descent method in [45] by up to  $18\%$  robust accuracy using a single step, providing a real-time robust inference method. In Table 2, even under the adaptive attack BPDA [6], our method also improves accuracy by up to  $14\%$ . Note that our method converges after 1 step of optimization, demonstrating the effectiveness of our approach.
+
+# 4.4. Ablation Study
+
+In this section, we present an ablation study by removing the proposed alternating optimization (AO) scheme and the experience replay buffer. We also compare against a  $\Theta$  that is randomly initialized. From Table 3 in appendix, we discovered that for in-distribution, AO and training of Theta improves the optimization performance by  $14\%$  and  $30\%$
+
+respectively. Such gap becomes  $35\%$  and  $28\%$  for evaluation on 200 steps. For OOD data, the advantage is furthered enlarged as shown in Table 3.
+
+One surprising result we discovered experimentally is that OBI under a randomly initialized mapping network  $\Theta$  consistently outperforms the baseline. We believe this is due to the fact that adding a Gaussian distribution to an underlying latent distribution of StyleGAN is beneficial in smoothing out loss landscape, making it easier to perform OBI. Similar random projection can also be found in [74].
+
+# 4.5. Loss Landscape
+
+To understand the underlying cause of the significant improvement in optimization brought by our mapping network  $\theta$ , we visualize in Fig. 9 the loss landscape for performing optimization in the original input space  $X$ , and our projected space  $Z$ . To generate this visualization, we first perform 20 steps of optimization on the validation dataset to collect a set of recovered latents. We then perform principle component analysis (PCA) on these recovered latents to obtain two principle directions. Finally, for individual examples, we evaluate the loss for vertices on a meshgrid spanned by the two principle directions where the center is the last step of optimization.
+
+From the visualization, we can see that the loss landscapes of the baseline are highly non-convex and contain
+
+![](images/4989d1af3ecdd68a877d72ca94a3441b4a17d62a06c6ce8e7b5f208a9c0b1cab.jpg)  
+In-Distribution  
+Encoder  
+Ours  
+Ground Truth
+
+![](images/d1a31fd19b5cdb469c4da96c60e6ee20716775786cb911d3761d9fd6c4f496dd.jpg)  
+Out-of-Distribution  
+Encoder  
+Ours  
+Ground Truth
+
+![](images/53b4913a7839fced8f224bb719df4235681cc332cf2ddb564415a29b685dd227.jpg)
+
+![](images/b3b1463c7a5a2c1e7915da3c1b95e5ebf1a627e742c368d6dcd1d0123675edec.jpg)  
+Figure 8: Comparison Against Encoder-Based Inference. Left shows the results on the test split of CelebA-HQ; Right shows the results on the LSUN-cat dataset.
+
+![](images/69afcfa8539fd65e8bb4e8be3a416b9ca34ddff77f40cf1a285b5eed6aa4f637.jpg)
+
+![](images/0b15995df62b7fd6efe12425d4f1aaa5f0bf943796309e5ac13111bc299506ca.jpg)
+
+![](images/37f9e7d3fab8c35a691240b0efe20708bd672bca55d45784606a0450aca620da.jpg)  
+Figure 9: Visualizing Loss Landscape (Uncurated). Visualizing the loss landscape of StyleGAN inversion spanned by two principle directions. Top row shows 4 examples of the loss landscapes corresponding to our space  $Z$ . Bottom row shows the loss landscapes corresponding to the original input space  $X$  for the same 4 examples. Note that the top row and the bottom row have different minimum loss values on the landscape because the minimas are obtained by performing independent optimization runs in space  $Z$  and  $X$  respectively. Given a fixed number of optimization steps, optimization in  $Z$  reaches lower loss values than optimization in  $X$ .
+
+![](images/d97445046b79e18a36a67a3ceb47b8f90ec9093be9b1c95045d4d5e3968e4ab6.jpg)
+
+![](images/83720abffbc6a5c377b720bd81e988339fce701fd224bb63921ab9f17f736a39.jpg)
+
+![](images/51c5ab68c7753665d5fea9a93da9d269916673ee362f3194915de8f7ee83c30b.jpg)
+
+points whose loss are significantly higher than its neighboring regions, while our loss landscapes are significantly smoother, with the "spikes" removed. Besides, our loss landscapes also tend to be steeper than the baseline ones. These two phenomena directly cause our method to perform gradient descent faster and stabler.
+
+# 5. Analysis
+
+# 5.1. L-BFGS
+
+After learning the mapping network  $\theta$ , in addition to the main results presented in Fig. 4a, we also run the same ex
+
+periments but with L-BFGS as optimizer [40], which tends to perform better empirically with a smoother loss landscape. As shown in table 5, after 10 steps of optimization, L-BFGS achieves much better results than Adam in our learned landscape, which is already significantly better than the baseline. Notably, applying L-BFGS in the original landscape does not achieve significant speedup, and for Celeb-HQ, around  $18\%$  of the examples failed to converge due to instability. These results further verify that our method successfully learns a smoother landscape for efficient optimization.
+
+<table><tr><td>level of Corruption</td><td>N(0,0.02)</td><td>N(0,0.12)</td><td>N(0,0.22)</td><td>N(0,0.32)</td><td>N(0,0.42)</td></tr><tr><td>Improvement</td><td>29.53%</td><td>18.95%</td><td>12.57%</td><td>6.30%</td><td>6.72%</td></tr></table>
+
+Table 4: Evaluation of a mapping network trained with Celeba-HQ trainset and tested on the Celeba-HQ testset corrupted with different levels of Gaussian noise. The improvement is calculated by the percentage improvement of loss defined by Eq. 6 from baseline (no mapping network) to ours (with mapping network) after 200 steps of optimization.  
+
+<table><tr><td>dataset</td><td>CelebA-HQ</td><td>AFHQ-Cat</td><td>LSUN-Cat</td><td>Container-Ship</td><td>N(0.5, 0.52)</td></tr><tr><td>Improvement</td><td>29.53%</td><td>35.99%</td><td>24.36%</td><td>16.87%</td><td>6.60%</td></tr></table>
+
+Table 6: Evaluation of a mapping network trained with Celeba-HQ trainset and tested on a spectrum of datasets from the original CelebA-HQ (in-distribution) to Gaussian noise (OOD). AFHQ-Cat [15] is a dataset with aligned cat faces. LSUN-Cat [77] is a dataset with unaligned cat images. Container-Ship is 200 images sampled from "Container Ship" class of ImageNet [20]. The improvement is calculated by the percentage improvement of loss defined by Eq. 6 from baseline (without  $\theta$ ) to ours (with  $\theta$ ) after 200 steps of optimization.  
+
+<table><tr><td></td><td>ours
+L-BFGS</td><td>ours
+Adam</td><td>baseline
+L-BFGS</td><td>baseline
+Adam</td></tr><tr><td>Human</td><td>2.239</td><td>3.504</td><td>N/A</td><td>4.875</td></tr><tr><td>Cat (OOD)</td><td>4.066</td><td>5.087</td><td>5.227</td><td>5.475</td></tr></table>
+
+Table 5: L-BFGS vs. Adam in Leanred Landscape (Ours)
+
+# 5.2. More Evaluation on OOD Generalization
+
+Since our proposed mapping network  $\Theta$  is parameterized by a neural network, there's no guarantee that the learned mapping function is surjective. Therefore, we empirically study the generalization performance by testing a mapping network trained on CelebA-HQ against a spectrum of datasets from very similar ones (in-distribution) to completely different ones (OOD).
+
+Synthetic Spectrum We first created a synthetic version of this spectrum of datasets by varying the level of Gaussian noise injected into the images of CelebA-HQ. With a higher level of Gaussian noise injected into the original images, more of the original image content is corrupted, creating a distribution of images further away from the original images.
+
+Natural Image Spectrum We also created a natural image version of this spectrum of datasets containing: CelebA-HQ (original in-distribution testset), AFHQ-Cat (center-aligned cat faces), LSUN-Cat (unaligned cat images), Container-Ship ("Container Ship" class from ImageNet), and Gaussian noise with individual pixel sampled from  $\mathcal{N}(0.5, 0.5^2)$ . From left to right, images vary from very similar to human faces to very different.
+
+From results evaluated on both synthetic spectrum and real spectrum, we observed consistent improvements of our method over the baseline, which shows the generalization performance of our method when applied to OOD data. From table 6, we see the improvements drop as the evaluation data is less and less in-distribution with the training data, which indicates that the mapping network does learn a prior from the training data.
+
+<table><tr><td></td><td>Baseline</td><td>Ours</td><td>Improvement</td></tr><tr><td>CelebA-HQ</td><td>2.569</td><td>2.396</td><td>6.72%</td></tr><tr><td>LSUN-Cat</td><td>4.617</td><td>3.717</td><td>19.49%</td></tr></table>
+
+Table 7: Evaluation of a mapping network trained with images of Gaussian noise sampled from  $\mathcal{N}(0.5,0.5^2)$  clipped to [0, 1]. Numbers show the average loss defined in Eq. 6 evaluated on testset of CelebA-HQ and LSUN-Cat.
+
+# 5.3. Dependence on Training Dataset
+
+From previous section, we know that the learned mapping network contains priors learned from the training data. To exclude such influence, we train our mapping network using randomly generated images made of Gaussian noise and evaluate on testset of CelebA-HQ. From table 7, we see that even trained with the task of reconstructing images of Gaussian noise, our mapping network still bring some improvement over the baseline model, though less significant.
+
+# 6. Conclusion
+
+This paper presents a method to accelerate optimization-based inference to invert a forward model. We propose an approach that learns a new space that is easier than the original input space to optimize with gradient descent at testing time. Our experiments and analysis on three different applications in computer vision have shown that by learning this mapping function, optimization becomes more efficient and generalizes better to out-of-distribution data. Through quantitative and qualitative analysis, we found that such improvement in optimization performance comes from a more efficient loss landscape. Since optimization-based inference has many advantages over encoder-based inference, we believe methods to accelerate them will have many impacts in a variety of applications.
+
+Acknowledgement This research is based on work partially supported by the NSF NRI Award 1925157, NSF STC LEAP, DARPA MCS program, and DARPA CCU program.
+
+# References
+
+[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 1, 2, 4  
+[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2  
+[3] Ijaz Akhter and Michael Black. Pose-conditioned joint angle limits for 3d human pose reconstruction. 06 2015. 5  
+[4] Brandon Amos, Lei Xu, and J Zico Kolter. Input convex neural networks. In International Conference on Machine Learning, pages 146-155. PMLR, 2017. 2  
+[5] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas. Learning to learn by gradient descent by gradient descent. Advances in neural information processing systems, 29, 2016. 3  
+[6] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 7  
+[7] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting largescale models. arXiv preprint arXiv:2203.17274, 1(3):4, 2022. 2  
+[8] David Bau, Jun-Yan Zhu, Jonas Wulff, William Peebles, Hendrik Strobelt, Bolei Zhou, and Antonio Torralba. Seeing what a gan cannot generate. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4502-4511, 2019. 2  
+[9] Ashish Bora, Ajil Jalal, Eric Price, and Alexandros G Dimakis. Compressed sensing using generative models. In International Conference on Machine Learning, pages 537-546. PMLR, 2017. 2  
+[10] Philémon Brakel, Dirk Stroobandt, and Benjamin Schrauwen. Training energy-based models for time-series imputation. The Journal of Machine Learning Research, 14(1):2771-2797, 2013. 1  
+[11] Nicholas Carlini and David A. Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy, pages 39-57, 2017. 2  
+[12] Yair Carmon, Aditi Raghunathan, Ludwig Schmidt, John C Duchi, and Percy S Liang. Unlabeled data improves adversarial robustness. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 6  
+[13] Boyuan Chen, Robert Kwiatkowski, Carl Vondrick, and Hod Lipson. Full-body visual self-modeling of robot morphologies. arXiv preprint arXiv:2111.06389, 2021. 1  
+[14] Tianlong Chen, Xiaohan Chen, Wuyang Chen, Howard Heaton, Jialin Liu, Zhangyang Wang, and Wotao Yin. Learning to optimize: A primer and a benchmark. arXiv preprint arXiv:2103.12828, 2021. 3
+
+[15] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 9  
+[16] CMU. Cmu mocap dataset. 5  
+[17] Jochen L Cremer, Ioannis Konstantelos, and Goran Strbac. From optimization-based machine learning to interpretable security rules for operation. IEEE Transactions on Power Systems, 34(5):3826-3836, 2019. 1  
+[18] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 2  
+[19] Katherine Crowson, Stella Rose Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. ArXiv, abs/2204.08583, 2022. 2  
+[20] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 9  
+[21] Tan M Dinh, Anh Tuan Tran, Rang Nguyen, and Binh-Son Hua. Hyperinverter: Improving stylegan inversion via hypernetwork. arXiv preprint arXiv:2112.00719, 2021. 2  
+[22] Justin Domke. Generic methods for optimization-based modeling. In Artificial Intelligence and Statistics, pages 318-326. PMLR, 2012. 1  
+[23] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2  
+[24] Chelsea Finn, Aravind Rajeswaran, Sham Kakade, and Sergey Levine. Online meta-learning. In International Conference on Machine Learning, pages 1920-1930. PMLR, 2019. 2  
+[25] Chelsea Finn, Kelvin Xu, and Sergey Levine. Probabilistic model-agnostic meta-learning. Advances in neural information processing systems, 31, 2018. 2  
+[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2  
+[27] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 5  
+[28] Carlos Hernandez, George Vogiatzis, and Roberto Cipolla. Multiview photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(3):548-554, 2008. 1  
+[29] Minyoung Huh, Richard Zhang, Jun-Yan Zhu, Sylvain Paris, and Aaron Hertzmann. Transforming and projecting images into class-conditional generative networks. In European Conference on Computer Vision, pages 17–34. Springer, 2020. 2  
+[30] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments.
+
+IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 5  
+[31] Ali Jahanian, Lucy Chai, and Phillip Isola. On the"steerability" of generative adversarial networks. arXiv preprint arXiv:1907.07171, 2019. 2  
+[32] Yongcheng Jing, Yezhou Yang, Zunlei Feng, Jingwen Ye, Yizhou Yu, and Mingli Song. Neural style transfer: A review. IEEE transactions on visualization and computer graphics, 26(11):3365-3385, 2019. 2  
+[33] Kyoungkook Kang, Seongtae Kim, and Sunghyun Cho. Gan inversion for out-of-range images with geometric transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13941-13949, 2021. 2  
+[34] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 4, 5  
+[35] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 4  
+[36] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2  
+[37] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5  
+[38] Kyoung Mu Lee and C-CJ Kuo. Shape from shading with a linear triangular element surface model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 15(8):815-822, 1993. 1  
+[39] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. ArXiv, abs/2104.08691, 2021. 2  
+[40] Dong C Liu and Jorge Nocedal. On the limited memory bfgs method for large scale optimization. Mathematical programming, 45(1-3):503-528, 1989. 8  
+[41] Xingchao Liu, Chengyue Gong, Lemeng Wu, Shujian Zhang, Haoran Su, and Qiang Liu. Fusedream: Training-free text-to-image generation with improved clip+gan space optimization. ArXiv, abs/2112.01573, 2021. 2  
+[42] Matthew Loper, Naureen Mahmood, and Michael J. Black. Mosh: Motion and shape capture from sparse markers. ACM Trans. Graph., 33(6), nov 2014. 5  
+[43] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 1  
+[44] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 2, 6  
+[45] Chengzhi Mao, Mia Chiquier, Hao Wang, Junfeng Yang, and Carl Vondrick. Adversarial attacks are reversible with natural supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 661-671, 2021. 1, 2, 6, 7
+
+[46] Chengzhi Mao, Revant Teotia, Amrutha Sundar, Sachit Menon, Junfeng Yang, Xin Wang, and Carl Vondrick. Doubly right object recognition: A why prompt for visual rationales. arXiv preprint arXiv:2212.06202, 2022. 2  
+[47] Chengzhi Mao, Lingyu Zhang, Abhishek Joshi, Junfeng Yang, Hao Wang, and Carl Vondrick. Robust perception through equivariance. arXiv preprint arXiv:2212.06079, 2022. 2  
+[48] Chengzhi Mao, Ziyuan Zhong, Junfeng Yang, Carl Vondrick, and Baishakhi Ray. Metric learning for adversarial robustness. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2  
+[49] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 2437-2445, 2020. 1  
+[50] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Ried-miller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. 3  
+[51] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 3  
+[52] Xingang Pan, Xiaohang Zhan, Bo Dai, Dahua Lin, Chen Change Loy, and Ping Luo. Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2  
+[53] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 1, 5  
+[54] Guim Perarnau, Joost Van De Weijer, Bogdan Raducanu, and Jose M Álvarez. Invertible conditional gans for image editing. arXiv preprint arXiv:1611.06355, 2016. 2  
+[55] Stanislav Pidhorskyi, Donald A Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14104-14113, 2020. 2  
+[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2  
+[57] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. 2016. 3  
+[58] Leslie Rice, Eric Wong, and J. Zico Kolter. Overfitting in adversarially robust deep learning, 2020. 6  
+[59] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 2, 4  
+[60] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 2
+
+[61] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243–9252, 2020. 2  
+[62] Veselin Stoyanov, Alexander Ropson, and Jason Eisner. Empirical risk minimization of graphical model parameters given approximate inference, decoding, and model structure. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 725-733. JMLR Workshop and Conference Proceedings, 2011. 1  
+[63] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv:1312.6199, 2013. 2  
+[64] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In European conference on computer vision, pages 581-600. Springer, 2020. 5  
+[65] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 2  
+[66] Austin Tripp, Erik Daxberger, and José Miguel Hernández-Lobato. Sample-efficient optimization in the latent space of deep generative models via weighted retraining. Advances in Neural Information Processing Systems, 33:11259-11272, 2020. 2  
+[67] Yun-Yun Tsai, Chengzhi Mao, Yow-Kuan Lin, and Junfeng Yang. Self-supervised convolutional visual prompts. arXiv preprint arXiv:2303.00198, 2023. 2  
+[68] Hao Wang, Chengzhi Mao, Hao He, Mingmin Zhao, Tommi S Jaakkola, and Dina Katabi. Bidirectional inference networks: A class of deep bayesian networks for health profiling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 766-773, 2019. 1, 2  
+[69] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 6  
+[70] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Weiming Zhang, Lu Yuan, Gang Hua, and Nenghai Yu. A simple baseline for stylegan inversion. ArXiv, abs/2104.07661, 2021. 2  
+[71] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688. CiteSeer, 2011. 2  
+[72] Olga Wichrowska, Niru Maheswaranathan, Matthew W Hoffman, Sergio Gomez Colmenarejo, Misha Denil, Nando Freitas, and Jascha Sohl-Dickstein. Learned optimizers that scale and generalize. In International Conference on Machine Learning, pages 3751-3760. PMLR, 2017. 3  
+[73] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. In NeurIPS, 2020. 6  
+[74] Yan Wu, Mihaela Rosca, and Timothy Lillicrap. Deep compressed sensing. In International Conference on Machine Learning, pages 6850-6860. PMLR, 2019. 2, 3, 7
+
+[75] Raymond A Yeh, Chen Chen, Teck Yian Lim, Alexander G Schwing, Mark Hasegawa-Johnson, and Minh N Do. Semantic image inpainting with deep generative models. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5485-5493, 2017. 2  
+[76] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 1  
+[77] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 4, 5, 9  
+[78] Junzhe Zhang, Xinyi Chen, Zhongang Cai, Liang Pan, Haiyu Zhao, Shuai Yi, Chai Kiat Yeo, Bo Dai, and Chen Change Loy. Unsupervised 3d shape completion through gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1768-1777, 2021. 1  
+[79] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4  
+[80] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2  
+[81] Jun-Yan Zhu, Philipp Kähenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In European conference on computer vision, pages 597-613. Springer, 2016. 2

2023/Landscape Learning for Neural Network Inversion/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9434e349eeb07fe67d9cecf9b2fddf7a3b9404602f31b69367e49520ba278560
+size 768865

2023/Large Selective Kernel Network for Remote Sensing Object Detection/1c2f7154-2a88-4538-ac11-efbaf0ea97b7_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc141ba277e97ab6af94fe4a451a2675c1f391aa837736812f84fa1d50558e85
+size 1302487

2023/Large Selective Kernel Network for Remote Sensing Object Detection/full.md

@@ -0,0 +1,448 @@
+# Large Selective Kernel Network for Remote Sensing Object Detection
+
+Yuxuan Li†, Qibin Hou, Zhaohui Zheng, Ming-Ming Cheng, Jian Yang* and Xiang Li*  
+VCIP, CS, Nankai University
+
+yuxuan.li.17@ucl.ac.uk†, {csjyang, xiang.li.implus}@nankai.edu.cn*
+
+# Abstract
+
+Recent research on remote sensing object detection has largely focused on improving the representation of oriented bounding boxes but has overlooked the unique prior knowledge presented in remote sensing scenarios. Such prior knowledge can be useful because tiny remote sensing objects may be mistakenly detected without referencing a sufficiently long-range context, which can vary for different objects. This paper considers these priors and proposes the lightweight Large Selective Kernel Network (LSKNet). LSKNet can dynamically adjust its large spatial receptive field to better model the ranging context of various objects in remote sensing scenarios. To our knowledge, large and selective kernel mechanisms have not been previously explored in remote sensing object detection. Without bells and whistles, our lightweight LSKNet sets new state-of-the-art scores on standard benchmarks, i.e., HRSC2016  $(98.46\% \text{mAP})$ , DOTA-v1.0  $(81.85\% \text{mAP})$ , and FAIR1M-v1.0  $(47.87\% \text{mAP})$ .
+
+# 1. Introduction
+
+Remote sensing object detection [47, 57, 82] focuses on identifying and locating objects of interest in aerial images, such as vehicles, ships or aircraft. In recent years, one mainstream trend has been to generate bounding boxes that accurately fit the orientation of the objects being detected rather than simply drawing horizontal boxes around them. Consequently, much research has focused on improving the representation of oriented bounding boxes for remote sensing object detection. This has largely been achieved through the development of specialized detection frameworks (i.e., RoI Transformer [12], Oriented R-CNN [69], and R3Det [75]) and oriented box encoding (i.e., gliding vertex [71] and midpoint offset box encoding [69]). Additionally, several loss functions, including GWD [77], KLD [79], and Modulated Loss [54], have been further proposed to enhance the
+
+![](images/c3bb16fa2ce72726663a286ca45a641a4d1fbd556a43bec0cac9b91882fe8b49.jpg)  
+Ship or Vehicle?
+
+![](images/6db7f7b631b5099d1fc4667c304046d7449724738601f65a7fa3b3c761dc2f65.jpg)  
+Figure 1. Successfully detecting remote sensing objects requires using a wide range of contextual information. Detectors with a limited receptive field may easily lead to incorrect results.
+
+![](images/910dff187a27be78cab0ab7c96a028b932c451fc85de589cb9fbfa24ca5b559a.jpg)
+
+performance of these approaches.
+
+Despite these advances, relatively few works have considered the strong prior knowledge of remote sensing images. Aerial images are typically captured at high resolutions from a bird's eye view. In particular, most objects in aerial images may be small and difficult to identify based on their appearance alone. Instead, recognizing these objects relies on their context, as the surrounding environment can provide valuable clues about their shape, orientation, and other characteristics. According to an analysis of the remote sensing data, we identify two important priors:
+
+- Accurate detection often requires a wide range of contextual information. As illustrated in Fig. 1, the limited context used by object detectors in remote sensing images can often lead to incorrect classifications. Rather than their appearance, the context distinguishes the ship from the vehicle.  
+- The contextual information required for different objects is very different. As shown in Fig. 2, the soccer field requires relatively less contextual information because of the unique distinguishable court borderlines. In contrast, the roundabout may require more context information to distinguish between gardens and ring-like buildings. Intersections, especially those partially covered by trees, require an extremely large receptive field due to the long-range dependencies between the intersecting roads.
+
+To address the challenge of accurately detecting objects
+
+![](images/84fbfd0eeb92108394a7ef9b0a7e672251aa231fbdfdf86c42802e44b6f03bb8.jpg)  
+Figure 2. The wide range of contextual information required for different object types is very different by human criteria. The objects with red boxes are the exact ground-truth annotations.
+
+in remote sensing images, which often require a wide and dynamic range of contextual information, we propose a novel lightweight detection backbone called Large Selective Kernel Network (LSKNet). Our approach involves dynamically adjusting the receptive field of the feature extraction backbone to more effectively process the varying wide context of the objects being detected. This is achieved through a spatial selective mechanism, which efficiently weights the features processed by a sequence of large depth-wise kernels and then spatially merge them. The weights of these kernels are determined dynamically based on the input, allowing the model to adaptively use different large kernels and adjust the receptive field for each target in space as needed.
+
+To our knowledge, the proposed LSKNet is the first to investigate using large and selective kernels for remote sensing object detection. Despite its simplicity and lightweight nature, our model achieves state-of-the-art performance on three popular datasets: HRSC2016 (98.46% mAP), DOTAv1.0 (81.85% mAP), and FAIR1M-v1.0 (47.87% mAP), surpassing previously published results. Furthermore, we demonstrate that our model's behaviour aligns with the two priors above, which in turn verifies the effectiveness of the proposed mechanism.
+
+# 2. Related Work
+
+# 2.1. Remote Sensing Object Detection Framework
+
+High-performance remote sensing object detectors often rely on the RCNN [56] framework, which consists of a region proposal network and regional CNN detection heads. The RPN proposes high-quality regions of interest (RoIs) from the backbone feature maps, while the regional CNN detection heads are responsible for object classification and bounding box regression. Several variations on the RCNN framework have been proposed in recent years. The two-
+
+stage RoI transformer [12] uses fully-connected layers to rotate candidate horizontal anchor boxes in the first stage, and then features within the boxes are extracted for further regression and classification. SCRDet [78] uses an attention mechanism to reduce background noise and improve the modelling of crowded and small objects. Oriented RCNN [69] and Gliding Vertex [71] introduce new box encoding systems to address the instability of training losses caused by rotation angle periodicity. Some approaches [32, 61, 87] treat remote sensing detection as a point detection task [74], providing an alternative way of addressing remote sensing detection problems.
+
+Rather than relying on the proposed anchors, one-stage detection frameworks classify and regress oriented bounding boxes directly from grid densely sampled anchors. The one-stage S²A network [23] extracts robust object features via oriented feature alignment and orientation-invariant feature extraction. DRN [50], on the other hand, leverages attention mechanisms to dynamically refined the backbone's extracted features for more accurate predictions. In contrast with Oriented RCNN and Gliding Vertex, RSDet [54] addresses the discontinuity of regression loss by introducing a modulated loss. LD [86] enhances the localization quality of oriented bounding boxes by distillation. AOPG [6] and R3Det [75] adopt a progressive regression approach, refining bounding boxes from coarse to fine granularity. In addition to CNN, AO2-DETR [9] introduces a transformer-based detection framework, DETR [4], into remote sensing detection tasks, which brings more research diversity.
+
+While these approaches have achieved promising results in addressing the issue of rotation variance, they do not consider the strong and valuable prior information presented in aerial images. Instead, our approach uses the large kernel and spatial selective mechanism to better model these priors without modifying the current detection framework.
+
+# 2.2. Large Kernel Networks
+
+Transformer-based [59] models, such as the Vision Transformer (ViT) [1, 11, 14, 53, 60], Swin transformer [25, 39, 51, 64, 70, 83], and pyramid transformer [62, 67], have gained popularity in computer vision. Research [17, 45, 55, 72, 85] has demonstrated that the large receptive field is a key factor in their success. Recent work has shown that well-designed convolutional networks with large receptive fields can also be highly competitive with transformer-based models. For example, ConvNeXt [40] uses  $7 \times 7$  depth-wise convolutions in its backbone, resulting in significant performance improvements in downstream tasks. In addition, RepLKNet [13] even uses a  $31 \times 31$  convolutional kernel via re-parameterization, achieving compelling performance. A subsequent work SLaK [38], further expands the kernel size to  $51 \times 51$  through kernel decomposition and sparse group techniques. RF-Next [18] automatically searches for a fixed
+
+![](images/c5a076b661b5d458804786abcc0a7b8380a683b072777292ba4b9ea3993b0fdd.jpg)  
+Figure 3. Architectural comparison between our proposed LSK module and other representative selective mechanism modules. K: Kernel.
+
+![](images/2cebb7e29b6ec97ac377dd67b9c69e0c2d7594fc42c85702b004e2241f3172b8.jpg)
+
+![](images/8ce330c0dfca81cc999baa38a4d05a2bc2de3eed51e13836546f3c7d641b9c8f.jpg)
+
+![](images/a52af307ee375073e004c1f68353bc71e84e097d84aadeed78c7291c2e00b1e7.jpg)
+
+large kernel for various tasks. VAN [19] introduces an efficient decomposition of large kernels as convolutional attention. Similarly, SegNeXt [20] and Conv2Former [28] demonstrate that large kernel convolution plays an important role in modulating the convolutional features with a richer context.
+
+Although large kernel convolutions have received attention in general object recognition, there has been a lack of research examining their significance in remote sensing detection. As previously noted in Sec. 1, aerial images possess unique characteristics that make large kernels particularly well-suited for remote sensing. As far as we know, our work represents the first attempt to introduce large kernel convolutions for remote sensing and to examine their importance in this field.
+
+# 2.3. Attention/Selective Mechanism
+
+The attention mechanism [21] is a simple but effective way to enhance neural representations for various tasks. The channel attention SE block [30] uses global average information to reweight feature channels, while spatial attention modules like GENet [29], GCNet [3], and SGE [34] enhance a network's ability to model context information via spatial masks. CBAM [66] and BAM [52] combine both channel and spatial attention.
+
+In addition to channel/spatial attention mechanisms, kernel selection is a self-adaptive and effective technique for dynamic context modelling. CondConv [73] and Dynamic convolution [5] use parallel kernels to adaptively aggregate features from multiple convolution kernels. SKNet [33] introduces multiple branches with different convolutional kernels and selectively combines them along the channel dimension. ResNeSt [84] extends the idea of SKNet by partitioning the input feature map into several groups. Similarly to the SKNet, SCNet [37] uses branch attention to capturing richer information and spatial attention to improve localization ability. Deformable Convnets [8, 89] introduce a flexible kernel shape for convolution units.
+
+Our approach bears the most similarity to SKNet [33]. However, there are two key distinctions between the two
+
+![](images/4d4de59c7dcfa5bf07dbaba2286cf630d5a38a4e92187d05be2691e3bf643343.jpg)  
+Figure 4. A block of LSKNet.
+
+methods. Firstly, our proposed selective mechanism relies explicitly on a sequence of large kernels via decomposition, a departure from most existing attention-based approaches. Secondly, our method adaptively aggregates information across large kernels in the spatial dimension rather than the channel dimension utilized by SKNet. This design is more intuitive and effective for remote sensing tasks because channel-wise selection fails to model the spatial variance for different targets across the image space. The detailed structural comparisons are listed in Fig. 3.
+
+# 3. Methods
+
+# 3.1. LSKNet Architecture
+
+The overall architecture of the LSKNet backbone is simply built upon repeated LSKNet Blocks. Fig 4 illustrates an LSKNet Block, which is inspired by ConvNeXt [41], MetaFormer [81], PVT-v2 [63], Conv2Former [28] and VAN [19]. Each LSKNet block consists of two residual sub-blocks: the Large Kernel Selection (LK Selection) subblock and the Feed-forward Network (FFN) sub-block.
+
+The LK Selection sub-block dynamically adjusts the network's receptive field as needed. The core LSK module
+
+![](images/4fc7ef940ab39965d1a30be538f026fe14d1fd03f7677f06bb30cbb8a0c2c15d.jpg)  
+Figure 5. A conceptual illustration of LSK module.
+
+<table><tr><td>Model</td><td>{C1,C2,C3,C4}</td><td>{D1,D2,D3,D4}</td><td>#P</td></tr><tr><td>★LSKNet-T</td><td>{32,64,160,256}</td><td>{3,3,5,2}</td><td>4.3M</td></tr><tr><td>★LSKNet-S</td><td>{64,128,320,512}</td><td>{2,2,4,2}</td><td>14.4M</td></tr></table>
+
+Table 1. Variants of LSKNet used in this paper.  $C_i$ : feature channel number;  $D_i$ : number of LSKNet blocks of each stage  $i$ .  
+
+<table><tr><td>RF</td><td>(k,d) sequence</td><td>#P</td><td>FLOPs</td></tr><tr><td rowspan="2">23</td><td>(23,1)</td><td>40.4K</td><td>42.4G</td></tr><tr><td>(5,1) → (7,3)</td><td>11.3K</td><td>11.9G</td></tr><tr><td rowspan="2">29</td><td>(29,1)</td><td>60.4K</td><td>63.3G</td></tr><tr><td>(3,1) → (5,2) → (7,3)</td><td>11.3K</td><td>13.6G</td></tr></table>
+
+Table 2. Theoretical efficiency comparisons of two representative examples by expanding a single large depth-wise kernel into a sequence, given channels being 64.  $k$ : kernel size;  $d$ : dilation.
+
+(Fig. 5) is embedded in the LK Selection sub-block. It consists of a sequence of large kernel convolutions and a spatial kernel selection mechanism, which will be elaborated on later. The FFN sub-block is used for channel mixing and feature refinement, which consists of a sequence of a fully connected layer, a depth-wise convolution, a GELU [26] activation, and a second fully connected layer.
+
+The detailed configuration of different variants of LSKNet used in this paper is listed in Tab. 1.
+
+# 3.2. Large Kernel Convolutions
+
+According to the prior (2), as stated in Sec. 1, it is suggested to model a series of multiple long-range contexts for adaptive selection. Therefore, we propose constructing a larger kernel convolution by explicitly decomposing it into a sequence of depth-wise convolutions with a large growing kernel and increasing dilation. Specifically, for the  $i$ -th depth-wise convolution, the expansion of the kernel size  $k$ , dilation rate  $d$ , and the receptive field  $RF$  are defined as follows:
+
+$$
+k _ {i - 1} \leq k _ {i}; d _ {1} = 1, d _ {i - 1} <   d _ {i} \leq R F _ {i - 1}, \tag {1}
+$$
+
+$$
+R F _ {1} = k _ {1}, R F _ {i} = d _ {i} \left(k _ {i} - 1\right) + R F _ {i - 1}. \tag {2}
+$$
+
+The increasing kernel size and dilation rate ensure that the receptive field expands quickly enough. We set an upper
+
+bound on the dilation rate to guarantee that the dilation convolution does not introduce gaps between feature maps. For instance, we can decompose a large kernel into 2 or 3 depth-wise convolutions as in Tab. 2, which have a theoretical receptive field of 23 and 29, respectively.
+
+There are two advantages of the proposed designs. First, it explicitly yields multiple features with various large receptive fields, which makes it easier for the later kernel selection. Second, sequential decomposition is more efficient than simply applying a single larger kernel. As shown in Tab. 2, under the same resulted theoretical receptive field, our decomposition greatly reduces the number of parameters compared to the standard large convolution kernels. To obtain features with rich contextual information from different ranges for input  $\mathbf{X}$ , a series of decomposed depth-wise convolutions with different receptive fields are applied:
+
+$$
+\mathbf {U} _ {0} = \mathbf {X}, \quad \mathbf {U} _ {i + 1} = \mathcal {F} _ {i} ^ {d w} (\mathbf {U} _ {i}), \tag {3}
+$$
+
+where  $\mathcal{F}_i^{dw}(\cdot)$  are depth-wise convolutions with kernel  $k_{i}$  and dilation  $d_{i}$ . Assuming there are  $N$  decomposed kernels, each of which is further processed by a  $1\times 1$  convolution layer  $\mathcal{F}^{1\times 1}(\cdot)$ :
+
+$$
+\widetilde {\mathbf {U}} _ {i} = \mathcal {F} _ {i} ^ {1 \times 1} \left(\mathbf {U} _ {i}\right), \text {f o r} i \text {i n} [ 1, N ], \tag {4}
+$$
+
+allowing channel mixing for each spatial feature vector. Then, a selection mechanism is proposed to dynamically select kernels for various objects based on the multi-scale features obtained, which would be introduced next.
+
+# 3.3. Spatial Kernel Selection
+
+To enhance the network's ability to focus on the most relevant spatial context regions for detecting targets, we use a spatial selection mechanism to spatially select the feature maps from large convolution kernels at different scales. Firstly, we concatenate the features obtained from different kernels with different ranges of receptive field:
+
+$$
+\widetilde {\mathbf {U}} = \left[ \widetilde {\mathbf {U}} _ {1}; \dots ; \widetilde {\mathbf {U}} _ {i} \right], \tag {5}
+$$
+
+and then efficiently extract the spatial relationship by applying channel-based average and maximum pooling (denoted
+
+as  $\mathcal{P}_{avg}(\cdot)$  and  $\mathcal{P}_{max}(\cdot)$  to  $\tilde{\mathbf{U}}$ :
+
+$$
+\mathbf {S A} _ {a v g} = \mathcal {P} _ {a v g} (\widetilde {\mathbf {U}}), \mathbf {S A} _ {m a x} = \mathcal {P} _ {m a x} (\widetilde {\mathbf {U}}), \qquad (6)
+$$
+
+where  $\mathbf{SA}_{avg}$  and  $\mathbf{SA}_{max}$  are the average and maximum pooled spatial feature descriptors. To allow information interaction among different spatial descriptors, we concatenate the spatially pooled features and use a convolution layer  $\mathcal{F}^{2\rightarrow N}(\cdot)$  to transform the pooled features (with 2 channels) into  $N$  spatial attention maps:
+
+$$
+\widehat {\mathbf {S A}} = \mathcal {F} ^ {2 \rightarrow N} \left(\left[ \mathbf {S A} _ {a v g}; \mathbf {S A} _ {m a x} \right]\right). \tag {7}
+$$
+
+For each of the spatial attention maps,  $\widehat{\mathbf{SA}_i}$ , a sigmoid activation function is applied to obtain the individual spatial selection mask for each of the decomposed large kernels:
+
+$$
+\widetilde {\mathbf {S A} _ {i}} = \sigma (\widehat {\mathbf {S A} _ {i}}), \tag {8}
+$$
+
+where  $\sigma(\cdot)$  denotes the sigmoid function. The feature maps from the sequence of decomposed large kernels are weighted by their corresponding spatial selection masks and then fused by a convolution layer  $\mathcal{F}(\cdot)$  to obtain the attention feature  $\mathbf{S}$ :
+
+$$
+\mathbf {S} = \mathcal {F} \left(\sum_ {i = 1} ^ {N} \left(\widetilde {\mathbf {S A}} _ {i} \cdot \widetilde {\mathbf {U}} _ {i}\right)\right). \tag {9}
+$$
+
+The final output of the LSK module is the element-wise product between the input feature  $\mathbf{X}$  and  $\mathbf{S}$ , similarly in [19, 20, 28]:
+
+$$
+\mathbf {Y} = \mathbf {X} \cdot \mathbf {S}. \tag {10}
+$$
+
+Fig. 5 shows a detailed conceptual illustration of an LSK module where we intuitively demonstrate how the large selective kernel works by adaptively collecting the corresponding large receptive field for different objects.
+
+# 4. Experiments
+
+# 4.1. Datasets
+
+HRSC2016 [42] is a high-resolution remote sensing dataset which is collected for ship detection. It consists of 1,061 images which contain 2,976 instances of ships.
+
+DOTA-v1.0 [68] consists of 2,806 remote sensing images. It contains 188,282 instances of 15 categories: Plane (PL), Baseball diamond (BD), Bridge (BR), Ground track field (GTF), Small vehicle (SV), Large vehicle (LV), Ship (SH), Tennis court (TC), Basketball court (BC), Storage tank (ST), Soccer-ball field (SBF), Roundabout (RA), Harbor (HA), Swimming pool (SP), and Helicopter (HC).
+
+FAIR1M-v1.0 [58] is a recently published remote sensing dataset that consists of 15,266 high-resolution images and more than 1 million instances. It contains 5 categories and 37 sub-categories objects.
+
+# 4.2. Implementation Details
+
+In our experiment, we report the results of the oriented object detection models on HRSC2016, DOTA-v1.0 and FAIR1M-v1.0 datasets. To ensure fairness, we follow the same dataset processing approach as other mainstream methods [23, 24, 69]. For DOTA-v1.0 and FAIR1M-v1.0 datasets, we adopt multi-scale training and testing strategy by first rescaling the images into three scales (0.5, 1.0, 1.5), and then cropping each scaled image into  $1024 \times 1024$  sub-images with a patch overlap of 500 pixels. For the HRSC2016 dataset, we rescale the images by setting the longer side of the image to 800 pixels, without changing their aspect ratios.
+
+During our experiments, the backbones are first pretrained on the ImageNet-1K [10] dataset and then finetuned on the target remote sensing benchmarks. In ablation studies, we adopt the 100-epoch backbone pretraining schedule for experimental efficiency (Tab. 3, 5, 4, 6, 7). We adopt a 300-epoch backbone pretraining strategy to pursue higher accuracy in main results (Tab. 8, 9, 10), similarly to [6,23,69,75]. In main results (Tab. 8, 9), the "Pre." column stands for the dataset on which the networks/backbones are pretrained (IN: Imagenet [10] dataset; CO: Microsoft COCO [36] dataset; MA: Million-AID [43] dataset). Unless otherwise stated, LSKNet is defaulting to be built within the framework of Oriented RCNN [69] due to its compelling performance and efficiency. All the models are trained on the training and validation sets and tested on the testing set. Following [69], we train the models for 36 epochs on the HRSC2016 datasets and 12 epochs on the DOTA-v1.0 and FAIR1M-v1.0 datasets, with the AdamW [44] optimizer. The initial learning rate is set to 0.0004 for HRSC2016, and 0.0002 for the other two datasets, with a weight decay of 0.05. The models are implemented under MMRotate [88] framework. We use 8 RTX3090 GPUs with a batch size of 8 for model training and use a single RTX3090 GPU for testing. All the FLOPs we report in this paper are calculated with a  $1024 \times 1024$  image input.
+
+# 4.3. Ablation Study
+
+In this section, we report ablation study results on the DOTA-v1.0 test set to investigate its effectiveness.
+
+Large Kernel Decomposition. Deciding on the number of kernels to decompose is a critical choice for the LSK module. We follow Eq. (1) to configure the decomposed kernels. The results of the ablation study on the number of large kernel decompositions, when the theoretical receptive field is fixed at 29, are shown in Tab. 3. It suggests that decomposing the large kernel into two depth-wise large kernels results in a good trade-off between speed and accuracy, achieving the best performance in terms of both FPS (frames per second) and mAP (mean average precision).
+
+Receptive Field Size and Selection Type. Based on our
+
+<table><tr><td>(k,d) sequence</td><td>RF</td><td>Num.</td><td>FPS</td><td>mAP (%)</td></tr><tr><td>(29,1)</td><td>29</td><td>1</td><td>18.6</td><td>80.66</td></tr><tr><td>(5,1)→(7,4)</td><td>29</td><td>2</td><td>20.5</td><td>80.91</td></tr><tr><td>(3,1)→(5,2)→(7,3)</td><td>29</td><td>3</td><td>19.2</td><td>80.77</td></tr></table>
+
+Table 3. The effects of the number of decomposed large kernels on the inference FPS and mAP, given the theoretical receptive field being 29. We adopt LSKNet-T backbones pretrained on ImageNet for 100 epochs. Decomposing the large kernel into two depth-wise kernels achieves the best performance of speed and accuracy.  
+
+<table><tr><td>(k1, d1)</td><td>(k2, d2)</td><td>CS</td><td>SS</td><td>RF</td><td>FPS</td><td>mAP (%)</td></tr><tr><td>(3, 1)</td><td>(5, 2)</td><td>-</td><td>-</td><td>11</td><td>22.1</td><td>80.80</td></tr><tr><td>(5, 1)</td><td>(7, 3)</td><td>-</td><td>-</td><td>23</td><td>21.7</td><td>80.94</td></tr><tr><td>(5, 1)</td><td>(7, 4)</td><td>-</td><td>-</td><td>29</td><td>20.5</td><td>80.91</td></tr><tr><td>(7, 1)</td><td>(9, 4)</td><td>-</td><td>-</td><td>39</td><td>21.3</td><td>80.84</td></tr><tr><td>(5, 1)</td><td>(7, 3)</td><td>✓</td><td>-</td><td>23</td><td>19.6</td><td>80.57</td></tr><tr><td>(5, 1)</td><td>(7, 3)</td><td>-</td><td>✓</td><td>23</td><td>20.7</td><td>81.31</td></tr></table>
+
+Table 4. The effectiveness of the key design components of the LSKNet when the large kernel is decomposed into a sequence of two depth-wise kernels. CS: channel selection (likewise in SKNet [33]); SS: spatial selection (ours). We adopt LSKNet-T backbones pretrained on ImageNet for 100 epochs. The LSKNet achieves the best performance when using a reasonably large receptive field with spatial selection.  
+
+<table><tr><td colspan="2">Pooling</td><td rowspan="2">FPS</td><td rowspan="2">mAP (%)</td></tr><tr><td>Max.</td><td>Avg.</td></tr><tr><td>✓</td><td></td><td>20.7</td><td>81.23</td></tr><tr><td></td><td>✓</td><td>20.7</td><td>81.12</td></tr><tr><td>✓</td><td>✓</td><td>20.7</td><td>81.31</td></tr></table>
+
+evaluations presented in Tab. 3, we find that the optimal solution for our proposed LSKNet is to decompose the large kernel into two depth-wise kernels in series. Furthermore, Tab. 4 shows that excessively small or large receptive fields can hinder the performance of the LSKNet, and a receptive field size of approximately 23 is determined to be the most effective. In addition, our experiments indicate that the proposed spatial selection approach is more effective than channel attention (similarly in SKNet [33]) for remote sensing object detection tasks. It suggests that in detection tasks, spatial information is more critical.
+
+Pooling Layers in Spatial Selection. We conduct experiments to determine the optimal pooling layers for spatial selection, as reported in Tab. 5. The results suggest that using both max and average pooling in the spatial selection component of our LSK module provides the best performance without sacrificing inference speed.
+
+Performance of LSKNet backbone under different detection frameworks. To validate the generality and ef
+
+Table 5. Ablation study on the effectiveness of the maximum and average pooling in spatial selection of our proposed LSK module. We adopt LSKNet-T backbones pretrained on ImageNet for 100 epochs. The best result is obtained when using both.  
+
+<table><tr><td>Frameworks</td><td>ResNet-18</td><td>★ LSKNet-T</td></tr><tr><td>ORCNN [69]</td><td>79.27</td><td>81.31 (+2.04)</td></tr><tr><td>RoI Trans. [12]</td><td>78.32</td><td>80.89 (+2.57)</td></tr><tr><td>S²A-Net [23]</td><td>76.82</td><td>80.15 (+3.33)</td></tr><tr><td>R3Det [75]</td><td>74.16</td><td>78.39 (+4.23)</td></tr><tr><td>#P (backbone only)</td><td>11.2M</td><td>4.3M (-62%)</td></tr><tr><td>FLOPs (backbone only)</td><td>38.1G</td><td>19.1G (-50%)</td></tr></table>
+
+Table 6. Comparison of LSKNet-T and ResNet-18 as backbones with different detection frameworks on DOTA-v1.0. The LSKNet-T backbone is pretrained on ImageNet for 100 epochs. The lightweight LSKNet-T achieves significantly higher mAP in various frameworks than ResNet-18.  
+
+<table><tr><td>Group</td><td>Model (backbone only)</td><td>#P</td><td>FLOPs</td><td>mAP (%)</td></tr><tr><td>Baseline</td><td>ResNet-18</td><td>11.2M</td><td>38.1G</td><td>79.27</td></tr><tr><td rowspan="3">Large Kernel</td><td>VAN-B1 [19]</td><td>13.4M</td><td>52.7G</td><td>81.15</td></tr><tr><td>ConvNeXt V2-N [65]</td><td>15.0M</td><td>51.2G</td><td>80.81</td></tr><tr><td>MSCAN-S [20]</td><td>13.1M</td><td>45.0G</td><td>81.12</td></tr><tr><td rowspan="3">Selective Attention</td><td>SKNet-26 [33]</td><td>14.5M</td><td>58.5G</td><td>80.67</td></tr><tr><td>ResNeSt-14 [84]</td><td>8.6M</td><td>57.9G</td><td>79.51</td></tr><tr><td>SCNet-18 [37]</td><td>14.0M</td><td>50.7G</td><td>79.69</td></tr><tr><td>Ours</td><td>★ LSKNet-S</td><td>14.4M</td><td>54.4G</td><td>81.48</td></tr><tr><td>Prev Best</td><td>CSPNeXt [46]</td><td>26.1M</td><td>87.6G</td><td>81.33</td></tr></table>
+
+Table 7. Comparison on LSKNet-S and other (large kernel/selective attention) backbones under O-RCNN [69] framework on DOTA-v1.0, except that the Prev Best is under RT-MDet [46] framework. All backbones are pretrained on ImageNet for 100 epochs. Our LSKNet achieves the best mAP under similar complexity budgets.  
+
+<table><tr><td>Method</td><td>Pre.</td><td>mAP(07)↑</td><td>mAP(12)↑</td><td>#P↓</td><td>FLOPs↓</td></tr><tr><td>DRN [50]</td><td>IN</td><td>-</td><td>92.70</td><td>-</td><td>-</td></tr><tr><td>CenterMap [61]</td><td>IN</td><td>-</td><td>92.80</td><td>41.1M</td><td>198G</td></tr><tr><td>Rol Trans. [12]</td><td>IN</td><td>86.20</td><td>-</td><td>55.1M</td><td>200G</td></tr><tr><td>G. V. [71]</td><td>IN</td><td>88.20</td><td>-</td><td>41.1M</td><td>198G</td></tr><tr><td>R3Det [75]</td><td>IN</td><td>89.26</td><td>96.01</td><td>41.9M</td><td>336G</td></tr><tr><td>DAL [48]</td><td>IN</td><td>89.77</td><td>-</td><td>36.4M</td><td>216G</td></tr><tr><td>GWD [77]</td><td>IN</td><td>89.85</td><td>97.37</td><td>47.4M</td><td>456G</td></tr><tr><td>S2ANet [23]</td><td>IN</td><td>90.17</td><td>95.01</td><td>38.6M</td><td>198G</td></tr><tr><td>AOPG [6]</td><td>IN</td><td>90.34</td><td>96.22</td><td>-</td><td>-</td></tr><tr><td>ReDet [24]</td><td>IN</td><td>90.46</td><td>97.63</td><td>31.6M</td><td>-</td></tr><tr><td>O-RCNN [69]</td><td>IN</td><td>90.50</td><td>97.60</td><td>41.1M</td><td>199G</td></tr><tr><td>RTMDet [46]</td><td>CO</td><td>90.60</td><td>97.10</td><td>52.3M</td><td>205G</td></tr><tr><td>★LSKNet-T</td><td>IN</td><td>90.54</td><td>98.13</td><td>21.0M</td><td>124G</td></tr><tr><td>★LSKNet-S</td><td>IN</td><td>90.65</td><td>98.46</td><td>31.0M</td><td>161G</td></tr></table>
+
+Table 8. Comparison with state-of-the-art methods on the HRSC2016 dataset. The LSKNet-S backbone is pretrained on ImageNet for 300 epochs, the same with most compared methods [23,69,75]. mAP (07/12): VOC 2007 [15]/2012 [16] metrics.
+
+fectiveness of our proposed LSKNet backbone, we evaluate its performance under various remote sensing detection frameworks, including two-stage frameworks O-RCNN [69] and RoI Transformer [12] as well as one-stage frameworks  $\mathrm{S}^2\mathrm{A}$  -Net [23] and R3Det [75]. The re
+
+Method |Pre.  $\left|\mathbf{mAP}\uparrow\right|\# \mathbf{P}\downarrow$  FLOPs↓ PL BD BR GTF SV LV SH TC BC ST SBF RA HA SP HC  
+
+<table><tr><td colspan="17">One-stage</td><td></td><td></td><td></td></tr><tr><td>R3Det [75]</td><td>IN</td><td>76.47</td><td>41.9M</td><td>336G</td><td>89.80</td><td>83.77</td><td>48.11</td><td>66.77</td><td>78.76</td><td>83.27</td><td>87.84</td><td>90.82</td><td>85.38</td><td>85.51</td><td>65.57</td><td>62.68</td><td>67.53</td><td>78.56</td><td>72.62</td></tr><tr><td>CFA [22]</td><td>IN</td><td>76.67</td><td>-</td><td>-</td><td>89.08</td><td>83.20</td><td>54.37</td><td>66.87</td><td>81.23</td><td>80.96</td><td>87.17</td><td>90.21</td><td>84.32</td><td>86.09</td><td>52.34</td><td>69.94</td><td>75.52</td><td>80.76</td><td>67.96</td></tr><tr><td>DAFNe [31]</td><td>IN</td><td>76.95</td><td>-</td><td>-</td><td>89.40</td><td>86.27</td><td>53.70</td><td>60.51</td><td>82.04</td><td>81.17</td><td>88.66</td><td>90.37</td><td>83.81</td><td>87.27</td><td>53.93</td><td>69.38</td><td>75.61</td><td>81.26</td><td>70.86</td></tr><tr><td>SASM [27]</td><td>IN</td><td>79.17</td><td>-</td><td>-</td><td>89.54</td><td>85.94</td><td>57.73</td><td>78.41</td><td>79.78</td><td>84.19</td><td>89.25</td><td>90.87</td><td>58.80</td><td>87.27</td><td>63.82</td><td>67.81</td><td>78.67</td><td>79.35</td><td>69.37</td></tr><tr><td>AO2-DETR [9]</td><td>IN</td><td>79.22</td><td>74.3M</td><td>304G</td><td>89.95</td><td>84.52</td><td>56.90</td><td>74.83</td><td>80.86</td><td>83.47</td><td>88.47</td><td>90.87</td><td>86.12</td><td>88.55</td><td>63.21</td><td>65.09</td><td>79.09</td><td>82.88</td><td>73.46</td></tr><tr><td>\( S^2ANet \) [23]</td><td>IN</td><td>79.42</td><td>-</td><td>-</td><td>88.89</td><td>83.60</td><td>57.74</td><td>81.95</td><td>79.94</td><td>83.19</td><td>89.11</td><td>90.78</td><td>84.87</td><td>87.81</td><td>70.30</td><td>68.25</td><td>78.30</td><td>77.01</td><td>69.58</td></tr><tr><td>R3Det-GWD [77]</td><td>IN</td><td>80.23</td><td>41.9M</td><td>336G</td><td>89.66</td><td>84.99</td><td>59.26</td><td>82.19</td><td>78.97</td><td>84.83</td><td>87.70</td><td>90.21</td><td>86.54</td><td>86.85</td><td>73.47</td><td>67.77</td><td>76.92</td><td>79.22</td><td>74.92</td></tr><tr><td>RTMDet-R [46]</td><td>IN</td><td>80.54</td><td>52.3M</td><td>205G</td><td>88.36</td><td>84.96</td><td>57.33</td><td>80.46</td><td>80.58</td><td>84.88</td><td>88.08</td><td>90.90</td><td>86.32</td><td>87.57</td><td>69.29</td><td>70.61</td><td>78.63</td><td>80.97</td><td>79.24</td></tr><tr><td>R3Det-KLD [79]</td><td>IN</td><td>80.63</td><td>41.9M</td><td>336G</td><td>89.92</td><td>85.13</td><td>59.19</td><td>81.33</td><td>78.82</td><td>84.38</td><td>87.50</td><td>89.80</td><td>87.33</td><td>87.00</td><td>72.57</td><td>71.35</td><td>77.12</td><td>79.34</td><td>78.68</td></tr><tr><td>RTMDet-R [46]</td><td>CO</td><td>81.33</td><td>52.3M</td><td>205G</td><td>88.01</td><td>86.17</td><td>58.54</td><td>82.44</td><td>81.30</td><td>84.82</td><td>88.71</td><td>90.89</td><td>88.77</td><td>87.37</td><td>71.96</td><td>71.18</td><td>81.23</td><td>81.40</td><td>77.13</td></tr><tr><td>Two-stage</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>SCRDet [78]</td><td>IN</td><td>72.61</td><td>-</td><td>-</td><td>89.98</td><td>80.65</td><td>52.09</td><td>68.36</td><td>68.36</td><td>60.32</td><td>72.41</td><td>90.85</td><td>87.94</td><td>86.86</td><td>65.02</td><td>66.68</td><td>66.25</td><td>68.24</td><td>65.21</td></tr><tr><td>Rol Trans. [12]</td><td>IN</td><td>74.61</td><td>55.1M</td><td>200G</td><td>88.65</td><td>82.60</td><td>52.53</td><td>70.87</td><td>77.93</td><td>76.67</td><td>86.87</td><td>90.71</td><td>83.83</td><td>82.51</td><td>53.95</td><td>67.61</td><td>74.67</td><td>68.75</td><td>61.03</td></tr><tr><td>G.V. [71]</td><td>IN</td><td>75.02</td><td>41.1M</td><td>198G</td><td>89.64</td><td>85.00</td><td>52.26</td><td>77.34</td><td>73.01</td><td>73.14</td><td>86.82</td><td>90.74</td><td>79.02</td><td>86.81</td><td>59.55</td><td>70.91</td><td>72.94</td><td>70.86</td><td>57.32</td></tr><tr><td>CenterMap [61]</td><td>IN</td><td>76.03</td><td>41.1M</td><td>198G</td><td>89.83</td><td>84.41</td><td>54.60</td><td>70.25</td><td>77.66</td><td>78.32</td><td>87.19</td><td>90.66</td><td>84.89</td><td>85.27</td><td>56.46</td><td>69.23</td><td>74.13</td><td>71.56</td><td>66.06</td></tr><tr><td>CSL [76]</td><td>IN</td><td>76.17</td><td>37.4M</td><td>236G</td><td>90.25</td><td>85.53</td><td>54.64</td><td>75.31</td><td>70.44</td><td>73.51</td><td>77.62</td><td>90.84</td><td>86.15</td><td>86.69</td><td>69.60</td><td>68.04</td><td>73.83</td><td>71.10</td><td>68.93</td></tr><tr><td>ReDet [24]</td><td>IN</td><td>80.10</td><td>-</td><td>-</td><td>88.81</td><td>82.48</td><td>60.83</td><td>80.82</td><td>78.34</td><td>86.06</td><td>88.31</td><td>90.87</td><td>88.77</td><td>87.03</td><td>68.65</td><td>66.90</td><td>79.26</td><td>79.71</td><td>74.67</td></tr><tr><td>DODet [7]</td><td>IN</td><td>80.62</td><td>-</td><td>-</td><td>89.96</td><td>85.52</td><td>58.01</td><td>81.22</td><td>78.71</td><td>85.46</td><td>88.59</td><td>90.89</td><td>87.12</td><td>87.80</td><td>70.50</td><td>71.54</td><td>82.06</td><td>77.43</td><td>74.47</td></tr><tr><td>AOPG [6]</td><td>IN</td><td>80.66</td><td>-</td><td>-</td><td>89.88</td><td>85.57</td><td>60.90</td><td>81.51</td><td>78.70</td><td>85.29</td><td>88.85</td><td>90.89</td><td>87.60</td><td>87.65</td><td>71.66</td><td>68.69</td><td>82.31</td><td>77.32</td><td>73.10</td></tr><tr><td>O-RCNN [69]</td><td>IN</td><td>80.87</td><td>41.1M</td><td>199G</td><td>89.84</td><td>85.43</td><td>61.09</td><td>79.82</td><td>79.71</td><td>85.35</td><td>88.82</td><td>90.88</td><td>86.68</td><td>87.73</td><td>72.21</td><td>70.80</td><td>82.42</td><td>78.18</td><td>74.11</td></tr><tr><td>KFloU [80]</td><td>IN</td><td>80.93</td><td>58.8M</td><td>206G</td><td>89.44</td><td>84.41</td><td>62.22</td><td>82.51</td><td>80.10</td><td>86.07</td><td>88.68</td><td>90.90</td><td>87.32</td><td>88.38</td><td>72.80</td><td>71.95</td><td>78.96</td><td>74.95</td><td>75.27</td></tr><tr><td>RVSA [60]</td><td>MA</td><td>81.24</td><td>114.4M</td><td>414G</td><td>88.97</td><td>85.76</td><td>61.46</td><td>81.27</td><td>79.98</td><td>85.31</td><td>88.30</td><td>90.84</td><td>85.06</td><td>87.50</td><td>66.77</td><td>73.11</td><td>84.75</td><td>81.88</td><td>77.58</td></tr><tr><td>★LSKNet-T (ours)</td><td>IN</td><td>81.37</td><td>21.0M</td><td>124G</td><td>89.14</td><td>84.90</td><td>61.78</td><td>83.50</td><td>81.54</td><td>85.87</td><td>88.64</td><td>90.89</td><td>88.02</td><td>87.31</td><td>71.55</td><td>70.74</td><td>78.66</td><td>79.81</td><td>78.16</td></tr><tr><td>★LSKNet-S (ours)</td><td>IN</td><td>81.64</td><td>31.0M</td><td>161G</td><td>89.57</td><td>86.34</td><td>63.13</td><td>83.67</td><td>82.20</td><td>86.10</td><td>88.66</td><td>90.89</td><td>88.41</td><td>87.42</td><td>71.72</td><td>69.58</td><td>78.88</td><td>81.77</td><td>76.52</td></tr><tr><td>★LSKNet-S* (ours)</td><td>IN</td><td>81.85</td><td>31.0M</td><td>161G</td><td>89.69</td><td>85.70</td><td>61.47</td><td>83.23</td><td>81.37</td><td>86.05</td><td>88.64</td><td>90.88</td><td>88.49</td><td>87.40</td><td>71.67</td><td>71.35</td><td>79.19</td><td>81.77</td><td>80.86</td></tr></table>
+
+Table 9. Comparison with state-of-the-art methods on the DOTA-v1.0 dataset with multi-scale training and testing. The LSKNet backbones are pretrained on ImageNet for 300 epochs, similarly to [23, 69, 75]. *: With EMA finetune similarly to the compared methods [46].  
+
+<table><tr><td>Model</td><td>G. V.* [71]</td><td>RetinaNet* [35]</td><td>C-RCNN* [2]</td><td>F-RCNN* [56]</td><td>RoI Trans.* [12]</td><td>O-RCNN [69]</td><td>★ LSKNet-T</td><td>★ LSKNet-S</td></tr><tr><td>mAP(%)</td><td>29.92</td><td>30.67</td><td>31.18</td><td>32.12</td><td>35.29</td><td>45.60</td><td>46.93</td><td>47.87</td></tr></table>
+
+Table 10. Comparison with state-of-the-art methods on the FAIR1M-v1.0 dataset. The LSKNet backbones are pretrained on ImageNet for 300 epochs, similarly to [23,69,75].*: Results are referenced from FAIR1M paper [58].
+
+sults in Tab. 6 show that our proposed LSKNet-T backbone significantly improves detection performance compared to ResNet-18, while using only  $38\%$  of its parameters and with  $50\%$  fewer FLOPs.
+
+Comparison with Other Large Kernel/Selective Attention Backbones. We also compare our LSKNet with 6 popular high-performance backbone models with large kernels or selective attention. As shown in Tab. 7, under similar model sizes and complexity budgets, our LSKNet outperforms all other models on the DOTA-v1.0 dataset.
+
+# 4.4. Main Results
+
+Results on HRSC2016. We evaluated the performance of our LSKNet against 12 state-of-the-art methods on the HRSC2016 dataset. The results presented in Tab. 8 demonstrate that our LSKNet-S outperforms all other methods with an mAP of  $90.65\%$  and  $98.46\%$  under the PASCAL VOC 2007 [15] and VOC 2012 [16] metrics, respectively.
+
+Results on DOTA-v1.0. We compare our LSKNet with
+
+20 state-of-the-art methods on the DOTA-v1.0 dataset, as reported in Tab. 9. Our LSKNet-T, LSKNet-S and LSKNet-S* achieve state-of-the-art with mAP of  $81.37\%$ ,  $81.64\%$  and  $81.85\%$  respectively. Notably, our high-performing LSKNet-S reaches an inference speed of 18.1 FPS on 1024x1024 images with a single RTX3090 GPU.
+
+Results on FAIR1M-v1.0. We compare our LSKNet against 6 other models on the FAIR1M-v1.0 dataset, as shown in Tab. 10. The results reveal that our LSKNet-T and LSKNet-S perform exceptionally well, achieving state-of-the-art mAP scores of  $46.93\%$  and  $47.87\%$  respectively, surpassing all other models by a significant margin.
+
+# 4.5. Analysis
+
+Detection Results Visualization. Visualization examples of detection results and Eigen-CAM [49] are shown in Fig. 6. LSKNet can capture much more context information relevant to the detected targets, leading to better performance in various hard cases, which justifies our prior (1).
+
+![](images/d7a8ed27195bfe79d5b31025d48d78e26b0cf161cd1c956ec42cbe49a8b53159.jpg)  
+(a) Robustness to obstacles
+
+![](images/f1cf7274765a362675051a08069ec03e4afad2981538c85f047aa29184ccefb7.jpg)  
+(b) Accurate Object Classification
+
+![](images/20c4bb773aecb4dfaa342f1407399b8017616b0288915928d51eca7471115d0d.jpg)  
+(c) Accurate Object Localization  
+Figure 6. Eigen-CAM visualization of Oriented RCNN detection framework with ResNet-50 and LSKNet-S. Our proposed LSKNet can model a much long range of context information, leading to better performance in various hard cases.
+
+![](images/7ebedb060824893151b42938746c503367c232b389eb40f079979e13898ef94d.jpg)  
+Figure 7. Normalised Ratio  $R_{c}$  of Expected Selective RF Area and GT Bounding Box Area for object categories in DOTA-v1.0. The relative range of context required for different object categories varies a lot. Examples of Bridge and Soccer-ball-field are given, where the visualized receptive field is obtained from Eq. (8) (i.e., the spatial activation) of our well-trained LSKNet model.
+
+Relative Context Range for Different Objects. To investigate the relative range of receptive field for each object category, we define  $R_{c}$  as the Ratio of Expected Selective RF Area and GT Bounding Box Area for category  $c$ :
+
+$$
+R _ {c} = \frac {\sum_ {i = 1} ^ {I _ {c}} A _ {i} / B _ {i}}{I _ {c}}, \tag {11}
+$$
+
+$$
+A _ {i} = \sum_ {d = 1} ^ {D} \sum_ {n = 1} ^ {N} | \widetilde {\mathbf {S A}} _ {n} ^ {d} \cdot R F _ {n} |, B _ {i} = \sum_ {j = 1} ^ {J _ {i}} \operatorname {A r e a} \left(\mathrm {G T} _ {j}\right), \tag {12}
+$$
+
+where  $I_{c}$  is the number of images that contain the object
+
+![](images/2cea3cd55e85b011419d017dc50c7bf5a4d9ed7c8006248b9021ae97b8a82202.jpg)  
+Figure 8. Normalised Kernel Selection Difference in the LSKNet-T blocks for Bridge, Roundabout and Soccer-ball-field. B.i.j represents the j-th LSK block in stage i. A greater value is indicative of a dependence on a broader context.
+
+category  $c$  only. The  $A_{i}$  is the sum of spatial selection activation in all LSK blocks for input image  $i$ , where  $D$  is the number of blocks in an LSKNet, and  $N$  is the number of decomposed large kernels in an LSK module.  $B_{i}$  is the total pixel area of all  $J_{i}$  annotated oriented object bounding boxes (GT). We plot the normalized  $R_{c}$  in Fig. 7 which represents the relative range of context required for different object categories for a better view.
+
+The results suggest that the Bridge category stands out as requiring a greater amount of additional contextual information compared to other categories, primarily due to its similarity in features with roads and the necessity of contextual clues to ascertain whether it is enveloped by water. Conversely, the Court categories, such as Soccer-ball-field, necessitate minimal contextual information owing to their distinctive textural attributes, specifically the court bound-
+
+![](images/512746ec2faceb1f4d3f46b3c748134cc1ed9cd1b4b997a852ac398188fdc385.jpg)  
+Bridge
+
+![](images/774d319db09b06d7db67c5605096e6b694516cfcfa790b7b542c00b5d59ea0f0.jpg)  
+Roundabout
+
+![](images/9beac37f773233e704842ffe7f81a23da7b961f5aed405c42fa90713a3b3e76d.jpg)  
+Small Vehicle
+
+![](images/7b6c39586a57fb0f698de55b9729fdff4f3165450b308e73fb43cf1a8c48e0b9.jpg)  
+Swimming Pool
+
+![](images/14125c3a6618e4fe0f91e7083dfe795d4cd53808d21b584b272c5953042ea552.jpg)  
+Ship
+
+![](images/3338637b3ebcde3702afc0b4ad60f0e84901eb63040f50d49a4f244714c25a21.jpg)  
+Storage tank
+
+![](images/c00d4df859a995651d0ab7c8e189bd3f6ead78cbb8164e8fa4ec9be90673080a.jpg)  
+Baseball diamond
+
+![](images/00181f35da6ab9d12fb110f9895f6c76c877a742f654a3f8d066eee074c16625.jpg)  
+Large Vehicle
+
+![](images/a2363b95673378b8e57f9ac7a60a2158b40c320a68867e0cbe9d13e7234cbc09.jpg)  
+Harbor
+
+![](images/0a790f8e132bbd87c9e1c168efde2a477c333e87deed8d4e8136587cb3c4ec06.jpg)  
+Tennis court
+
+![](images/e496d46bd283e0c36260a9d4452442e62c87f0a15a93235d83e1178fe29dcc57.jpg)  
+Plane  
+Figure 9. Receptive field activation for more object categories in DOTA-v1.0, where the activation map is obtained from the Eq. (8) (i.e., the spatial activation) of our well-trained LSKNet model.
+
+![](images/0604412e8dcab7a2f7eacbee42261b7d60ceda6a452313709ee4c2b3385f7d7f.jpg)  
+Soccer-ball field
+
+ary lines. It aligns with our knowledge and further supports prior (2) that the relative range of contextual information required for different object categories varies greatly.
+
+Kernel Selection Behaviour. We further investigate the kernel selection behaviour in our LSKNet. For object category  $c$ , the Kernel Selection Difference  $\Delta A_{c}$  (i.e., larger kernel selection - smaller kernel selection) of an LSKNet-T block is defined as:
+
+$$
+\Delta A _ {c} = \left| \widetilde {\mathbf {S A}} _ {\text {l a r g e r}} - \widetilde {\mathbf {S A}} _ {\text {s m a l l e r}} \right|. \tag {13}
+$$
+
+We demonstrate the normalised  $\Delta A_{c}$  over all images for three typical categories: Bridge, Roundabout and Soccerball-field and for each LSKNet-T block in Fig. 8. As expected, the participation of larger kernels of all blocks for Bridge is higher than that of Roundabout, and Roundabout is higher than Soccer-ball-field. This aligns with the common sense that Soccer-ball-field indeed does not require a large amount of context, since its own texture characteristics are already sufficiently distinct and discriminatory.
+
+We also surprisingly discover another selection pattern of LSKNet across network depth: LSKNet usually utilizes larger kernels in its shallow layers and smaller kernels in higher levels. This indicates that networks tend to quickly focus on capturing information from large receptive fields in low-level layers so that higher-level semantics can contain sufficient receptive fields for better discrimination.
+
+Spatial Activation Visualisations. Spatial activation map examples for more object categories in DOTA-v1.0
+
+are shown in Fig. 9, where the activation map is obtained from Eq. (8) (i.e., the spatial activation) of our well-trained LSKNet model. The object categories are arranged in decreasing order from top left to bottom right based on the Ratio of Expected Selective RF Area and GT Bounding Box Area as illustrated in Fig. 7. The spatial activation visualisation results also demonstrate that the model's behaviour aligns with our proposed two priors and the above analysis, which in turn verifies the effectiveness of the proposed mechanism.
+
+# 5. Conclusion
+
+In this paper, we propose the Large Selective Kernel Network (LSKNet) for remote sensing object detection tasks, which is designed to utilize the inherent characteristics in remote sensing images: the need for a wider and adaptable contextual understanding. By adapting its large spatial receptive field, LSKNet can effectively model the varying contextual nuances of different object types. Extensive experiments demonstrate that our proposed lightweight model achieves state-of-the-art performance on competitive remote sensing benchmarks.
+
+Acknowledgement. This research was supported by the NSFC (NO. 62176130, 62206134, 62225604) and the Fundamental Research Funds for the Central Universities (Nankai University, 070-63233084, 070-63233089). Computation is supported by the Supercomputing Center of Nankai University.
+
+# References
+
+[1] Yakoub Bazi, Laila Bashmal, Mohamad M. Al Rahhal, Reham Al Dayil, and Naif Al Ajan. Vision transformers for remote sensing image classification. Remote Sensing, 2021. 2  
+[2] Zhaowei Cai and Nuno Vasconcelos. Cascade R-CNN: Delving into high quality object detection. In CVPR, 2018. 7  
+[3] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. GCNet: Non-local networks meet squeeze-excitation networks and beyond. In ICCV Workshops, 2019. 3  
+[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2  
+[5] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In CVPR, 2020. 3  
+[6] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. TGARS, 2022. 2, 5, 6, 7  
+[7] Gong Cheng, Yanqing Yao, Shengyang Li, Ke Li, Xingxing Xie, Jiabao Wang, Xiwen Yao, and Junwei Han. Dual-aligned oriented detector. TGARS, 2022. 7  
+[8] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, 2017. 3  
+[9] Linhui Dai, Hong Liu, Hao Tang, Zhiwei Wu, and Pinhao Song. AO2-DETR: Arbitrary-oriented object detection transformer. TCSVT, 2022. 2, 7  
+[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5  
+[11] Peifang Deng, Kejie Xu, and Hong Huang. When cnns meet vision transformer: A joint framework for remote sensing scene classification. IEEE Geoscience and Remote Sensing Letters, 2022. 2  
+[12] Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning RoI transformer for oriented object detection in aerial images. In CVPR, 2019. 1, 2, 6, 7  
+[13] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 2  
+[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR 2021, 2021. 2  
+[15] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2007 (VOC2007) Results. 6, 7  
+[16] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. 6, 7  
+[17] Deng-Ping Fan, Ge-Peng Ji, Peng Xu, Ming-Ming Cheng, Christos Sakaridis, and Luc Van Gool. Advances in deep concealed scene understanding. Visual Intelligence, 2023. 2
+
+[18] Shanghua Gao, Zhong-Yu Li, Qi Han, Ming-Ming Cheng, and Liang Wang. Rf- next: Efficient receptive field search for convolutional neural networks. IEEE TPAMI, 45(3):2984- 3002, 2023. 2  
+[19] Meng-Hao Guo, Chengrou Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shiyong Hu. Visual attention network. ArXiv, 2022. 3, 5, 6  
+[20] Meng-Hao Guo, Cheng-Ze Lu, Qibin Hou, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Segnext: Rethinking convolutional attention design for semantic segmentation. In NeurIPS, 2022. 3, 5, 6  
+[21] Meng-Hao Guo, Tianxing Xu, Jiang-Jiang Liu, Zheng-Ning Liu, Peng-Tao Jiang, Tai-Jiang Mu, Song-Hai Zhang, Ralph Martin, Ming-Ming Cheng, and Shi-Min Hu. Attention mechanisms in computer vision: A survey. Computational Visual Media, 11 2021. 3  
+[22] Zonghao Guo, Chang Liu, Xiaosong Zhang, Jianbin Jiao, Xi-angyang Ji, and Qixiang Ye. Beyond bounding-box: Convex-hull feature adaptation for oriented and densely packed object detection. In CVPR, 2021. 7  
+[23] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. TGARS, 2020. 2, 5, 6, 7  
+[24] Jiaming Han, Jian Ding, Nan Xue, and Gui-Song Xia. Redet: A rotation-equivariant detector for aerial object detection. In CVPR, 2021. 5, 6, 7  
+[25] Siyuan Hao, Bin Wu, Kun Zhao, Yuanxin Ye, and Wei Wang. Two-stream swim transformer with differentiable sobel operator for remote sensing image classification. Remote Sensing, 2022. 2  
+[26] Dan Hendrycks and Kevin Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. CoRR, 2016. 4  
+[27] Liping Hou, Ke Lu, Jian Xue, and Yuqiu Li. Shape-adaptive selection and measurement for oriented object detection. In AAAI, 2022. 7  
+[28] Qibin Hou, Cheng-Ze Lu, Ming-Ming Cheng, and Jiashi Feng. Conv2former: A simple transformer-style ConvNet for visual recognition. ArXiv, 2022. 3, 5  
+[29] Jie Hu, Li Shen, Samuel Albanie, Gang Sun, and Andrea Vedaldi. Gather-excite: Exploiting feature context in convolutional neural networks. In NeurPIS, 2018. 3  
+[30] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 3  
+[31] Steven Lang, Fabrizio Ventola, and Kristian Kersting. Dafne: A one-stage anchor-free deep model for oriented object detection. CoRR, 2021. 7  
+[32] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In ECCV, 2018. 2  
+[33] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In CVPR, 2019. 3, 6  
+[34] Yuxuan Li, Xiang Li, and Jian Yang. Spatial group-wise enhance: Enhancing semantic feature learning in cnn. In ACCV, 2022. 3  
+[35] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In ICCV, 2017. 7
+
+[36] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5  
+[37] Jiang-Jiang Liu, Qibin Hou, Ming-Ming Cheng, Changhu Wang, and Jiashi Feng. Improving convolutional networks with self-calibrated convolutions. In CVPR, 2020. 3, 6  
+[38] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity. ArXiv, 2022. 2  
+[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2  
+[40] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 2  
+[41] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 3  
+[42] Zikun Liu, Hongzhen Wang, Lubin Weng, and Yiping Yang. Ship rotated bounding box space for ship extraction from high-resolution optical satellite images with complex backgrounds. IEEE Geoscience and Remote Sensing Letters, 2016. 5  
+[43] Yang Long, Gui-Song Xia, Shengyang Li, Wen Yang, Michael Ying Yang, Xiao Xiang Zhu, Liangpei Zhang, and Deren Li. On creating benchmark dataset for aerial image interpretation: Reviews, guidances, and million-aid. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021. 5  
+[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. ArXiv, 2017. 5  
+[45] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. In NeurIPS, 2016. 2  
+[46] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. CoRR, 2022. 6, 7  
+[47] Jie Mei, Yi-Bo Zheng, and Ming-Ming Cheng. D2anet: Difference-aware attention network for multi-level change detection from satellite imagery. Computational Visual Media, 2023. 1  
+[48] Qi Ming, Zhiqiang Zhou, Lingjuan Miao, Hongwei Zhang, and Linhao Li. Dynamic anchor learning for arbitrary-oriented object detection. CoRR, 2020. 6  
+[49] Mohammed Bany Muhammad and Mohammed Yeasin. Eigen-cam: Class activation map using principal components. CoRR, 2020. 7  
+[50] Xingjia Pan, Yuqiang Ren, Kekai Sheng, Weiming Dong, Haolei Yuan, Xiaowei Guo, Chongyang Ma, and Changsheng Xu. Dynamic refinement network for oriented and densely packed object detection. In CVPR, 2020. 2, 6  
+[51] Teerapong Panboonyuen, Kulsawasd Jitkajornwanich, Siam Lawawirojwong, Panu Srestasathiern, and Peerapon Vateekul. Transformer-based decoder designs for semantic
+
+segmentation on remotely sensed images. Remote Sensing, 2021. 2  
+[52] Jongchan Park, Sanghyun Woo, Joon-Young Lee, and In-So Kweon. Bam: Bottleneck attention module. In British Machine Vision Conference, 2018. 3  
+[53] Malsha V. Perera, Wele Gedara Chaminda Bandara, Jeya Maria Jose Valanarasu, and Vishal M. Patel. Transformer-based SAR image despeckling. CoRR, 2022. 2  
+[54] Wen Qian, Xue Yang, Silong Peng, Junchi Yan, and Yue Guo. Learning modulated loss for rotated object detection. In AAAI, 2021. 1, 2  
+[55] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 2  
+[56] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2, 7  
+[57] Xian Sun, Yu Tian, Wanxuan Lu, Peijin Wang, Ruigang Niu, Hongfeng Yu, and Kun Fu. From single- to multi-modal remote sensing imagery interpretation: a survey and taxonomy. Science China Information Sciences, 2023. 1  
+[58] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, Martin Weinmann, Stefan Hinz, Cheng Wang, and Kun Fu. FAIR1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. IS-PRS Journal of Photogrammetry and Remote Sensing, 2022. 5, 7  
+[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2  
+[60] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer towards remote sensing foundation model. TGARS, 2022. 2, 7  
+[61] Jinwang Wang, Wen Yang, Heng-Chao Li, Haijian Zhang, and Gui-Song Xia. Learning center probability map for detecting objects in aerial images. TGARS, 2021. 2, 6, 7  
+[62] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. 2  
+[63] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. CVM, 2022. 3  
+[64] Xiao Wang, Guangyao Chen, Guangwu Qian, Pengcheng Gao, Xiao-Yong Wei, Yaowei Wang, Yonghong Tian, and Wen Gao. Large-scale multi-modal pre-trained models: A comprehensive survey. Machine Intelligence Research, 20(4):447-482, 2023. 2  
+[65] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In-So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. Arxiv, 2023. 6  
+[66] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. CBAM: Convolutional block attention module. In ECCV, 2018. 3
+
+[67] Yu-Huan Wu, Yun Liu, Xin Zhan, and Ming-Ming Cheng. P2T: Pyramid pooling transformer for scene understanding. IEEE TPAMI, pages 1-12, 2022. 2  
+[68] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In CVPR, 2018. 5  
+[69] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented R-CNN for object detection. In ICCV, 2021. 1, 2, 5, 6, 7  
+[70] Xiangkai Xu, Zhejun Feng, Changqing Cao, Mengyuan Li, Jin Wu, Zengyan Wu, Yajie Shang, and Shubing Ye. An improved swim transformer-based model for remote sensing object detection and instance segmentation. Remote Sensing, 2021. 2  
+[71] Yongchao Xu, Mingtao Fu, Qimeng Wang, Yukang Wang, Kai Chen, Gui-Song Xia, and Xiang Bai. Gliding vertex on the horizontal bounding box for multi-oriented object detection. IEEE TPAMI, 2021. 1, 2, 6, 7  
+[72] Haotian Yan, Zhe Li, Weijian Li, Changhu Wang, Ming Wu, and Chuang Zhang. Contnet: Why not use convolution and transformer at the same time? CoRR, 2021. 2  
+[73] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. NeurIPS, 2019. 3  
+[74] Jing Yang, Qingshan Liu, and Kaihua Zhang. Stacked hourglass network for robust facial landmark localisation. In CVPR Workshops, 2017. 2  
+[75] Xue Yang, Qingqing Liu, Junchi Yan, and Ang Li. R3det: Refined single-stage detector with feature refinement for rotating object. CoRR, 2019. 1, 2, 5, 6, 7  
+[76] Xue Yang and Junchi Yan. Arbitrary-oriented object detection with circular smooth label. In ECCV, 2020. 7  
+[77] Xue Yang, Junchi Yan, Qi Ming, Wentao Wang, Xiaopeng Zhang, and Qi Tian. Rethinking rotated object detection with gaussian Wasserstein distance loss. In ICML, 2021. 1, 6, 7  
+[78] Xue Yang, Jirui Yang, Junchi Yan, Yue Zhang, Tengfei Zhang, Zhi Guo, Xian Sun, and Kun Fu. SCRDet: Towards more robust detection for small, cluttered and rotated objects. In ICCV, 2019. 2, 7  
+[79] Xue Yang, Xiaojiang Yang, Jirui Yang, Qi Ming, Wentao Wang, Qi Tian, and Junchi Yan. Learning high-precision bounding box for rotated object detection via kullback-leibler divergence. In NeurIPS, 2021. 1, 7  
+[80] Xue Yang, Yue Zhou, Gefan Zhang, Jirui Yang, Wentao Wang, Junchi Yan, Xiaopeng Zhang, and Qi Tian. The KFloU loss for rotated object detection. In ICLR, 2022. 7  
+[81] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 3  
+[82] Syed Sahil Abbas Zaidi, Mohammad Samar Ansari, Asra Aslam, Nadia Kanwal, Mamoona Asghar, and Brian Lee. A survey of modern deep learning based object detection models. Digital Signal Processing, 2022. 1  
+[83] Cui Zhang, Liejun Wang, Shuli Cheng, and Yongming Li. Swinsunet: Pure transformer network for remote sensing image change detection. IEEE Transactions on Geoscience and Remote Sensing, 2022. 2
+
+[84] Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R. Manmatha, Mu Li, and Alexander Smola. Resnest: Split-attention networks. In CVPR Workshops, 2022. 3, 6  
+[85] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H.S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, 2021. 2  
+[86] Zhaohui Zheng, Rongguang Ye, Qibin Hou, Dongwei Ren, Ping Wang, Wangmeng Zuo, and Ming-Ming Cheng. Localization distillation for object detection. IEEE TPAMI, 2023. 2  
+[87] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. ArXiv, 2019. 2  
+[88] Yue Zhou, Xue Yang, Gefan Zhang, Jiabao Wang, Yanyi Liu, Liping Hou, Xue Jiang, Xingzhao Liu, Junchi Yan, Chengqi Lyu, et al. Mmrotate: A rotated object detection benchmark using pytorch. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7331-7334, 2022. 5  
+[89] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In CVPR, 2019. 3

2023/Large Selective Kernel Network for Remote Sensing Object Detection/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7634af88f64c1b3a8403c5bae396450fbb88f1285d97079603a97f910d8cdf92
+size 963598

2023/Large-Scale Land Cover Mapping with Fine-Grained Classes via Class-Aware Semi-Supervised Semantic Segmentation/85af6521-8e24-45e3-ac82-fb3352f9480e_content_list.json

@@ -0,0 +1,1475 @@
+[
+    {
+        "type": "text",
+        "text": "Large-Scale Land Cover Mapping with Fine-Grained Classes via Class-Aware Semi-Supervised Semantic Segmentation",
+        "text_level": 1,
+        "bbox": [
+            91,
+            128,
+            879,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Runmin Dong $^{1,4}$ , Lichao Mou $^{2}$ , Mengxuan Chen $^{1,4}$ , Weijia Li $^{3}$ , Xin-Yi Tong $^{2}$ , Shuai Yuan $^{1,4}$ , Lixian Zhang $^{1,4}$ , Juepeng Zheng $^{3,4}$ , Xiao Xiang Zhu $^{2}$ , Haohuan Fu $^{1,4,*}$ , Tsinghua University  ${}^{2}$ Technical University of Munich  ${}^{3}$ Sun Yat-Sen University  ${}^{4}$ Tsinghua University - Xi'an Institute of Surveying and Mapping Joint Research Center",
+        "bbox": [
+            138,
+            202,
+            833,
+            276
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "drm@mail.tsinghua.edu.cn, haohuan@tsinghua.edu.cn",
+        "bbox": [
+            264,
+            277,
+            702,
+            291
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            325,
+            313,
+            343
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Semi-supervised learning has attracted increasing attention in the large-scale land cover mapping task. However, existing methods overlook the potential to alleviate the class imbalance problem by selecting a suitable set of unlabeled data. Besides, in class-imbalanced scenarios, existing pseudo-labeling methods mostly only pick confident samples, failing to exploit the hard samples during training. To tackle these issues, we propose a unified Class-Aware Semi-Supervised Semantic Segmentation framework. The proposed framework consists of three key components. To construct a better semi-supervised learning dataset, we propose a class-aware unlabeled data selection method that is more balanced towards the minority classes. Based on the built dataset with improved class balance, we propose a Class-Balanced Cross Entropy loss, jointly considering the annotation bias and the class bias to re-weight the loss in both sample and class levels to alleviate the class imbalance problem. Moreover, we propose the Class Center Contrast method to jointly utilize the labeled and unlabeled data. Specifically, we decompose the feature embedding space using the ground truth and pseudo-labels, and employ the embedding centers for hard and easy samples of each class per image in the contrast loss to exploit the hard samples during training. Compared with state-of-the-art class-balanced pseudo-labeling methods, the proposed method improves the mean accuracy and mIoU by  $4.28\\%$  and  $1.70\\%$ , respectively, on the large-scale Sentinel-2 dataset with 24 land cover classes.",
+        "bbox": [
+            76,
+            358,
+            473,
+            781
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            808,
+            209,
+            824
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Land cover mapping provides pixel-level information for urban management, climate change research, ecosystem protection, and other sustainability-related applica",
+        "bbox": [
+            76,
+            834,
+            468,
+            878
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/7b5448a33ca75d31ea5097cb4e3ad58bf9f15c5950fbe43a104bff7189823f84.jpg",
+        "image_caption": [
+            "Figure 1. Examples of the labeled data. The top row shows Sentinel-2 images containing 13 bands. The bottom row shows the corresponding labels. We use a fine-grained classification system with 24 land cover classes in this work."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            328,
+            893,
+            554
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tions [21, 27]. With the rapid progress of computer vision technology, it is desired to automatically gain large-scale and fine-grained land cover information from remote sensing images to support the demand of existing earth system science studies and bring new opportunities for intelligent city study [31]. However, fine-grained annotations for land cover mapping are high-cost and time-consuming, limiting labeled data collection for large-scale applications. Semi-supervised learning (SSL) is a potential solution to this problem, as it utilizes both existing labeled data and a large number of unlabeled data [37].",
+        "bbox": [
+            496,
+            655,
+            893,
+            821
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large-scale land cover mapping with fine-grained classes usually suffers from class imbalance, as shown in Figure 1 and 3. The labeled data tends to be biased due to the difficulty of data collection and label annotation varying by class. The class imbalance leads to suboptimal per",
+        "bbox": [
+            496,
+            824,
+            893,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author",
+        "bbox": [
+            94,
+            887,
+            220,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "16783",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "formance in the minority classes. Most SSL methods for large-scale land cover mapping are typically based on the assumption that labeled and unlabeled data have the same class distribution [41, 37]. However, this assumption does not hold in most scenarios, resulting in the bias of class estimation. As such, it is non-trivial to jointly utilize the labeled and unlabeled data to train the network for semi-supervised semantic segmentation. Although many pseudo-labeling methods aim to assign pseudo-labels more steadily and correctly for different classes on the unlabeled data, e.g., dynamically setting threshold or proportion [17, 44], they still rely on setting hyperparameters and empirical principles, resulting in a lack of flexibility and generalization ability in applications. Besides, pseudo-labeling methods mostly rely on easy samples with high confidence scores, thus failing to exploit the hard samples on the unlabeled data.",
+        "bbox": [
+            80,
+            90,
+            467,
+            330
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To alleviate the above issues, we propose a unified Class-Aware Semi-Supervised Semantic Segmentation framework. We exploit the potential of the unlabeled data and alleviate the class imbalance problem in large-scale land cover mapping with fine-grained classes. The proposed class-aware unlabeled data selection method constructs an SSL dataset and compensates for the class imbalance on the labeled data. Based on the built SSL dataset, we propose a class-balanced learning method to remove the annotation bias and class bias on the SSL dataset. We dynamically estimate the class prior information on the entire SSL dataset during training to re-weight both samples and classes of the loss function for the labeled data. Finally, we propose the Class Center Contrast method to jointly use the labeled and unlabeled data for training. The ground truth and pseudolabels are leveraged as the guide to decompose the feature embedding space. We estimate the hard, easy, and overall embedding centers for each class per image and apply the contrastive loss to optimize the distances between class centers to utilize both easy and hard samples. The class imbalance issue is alleviated by utilizing the proposed method, and the performance of large-scale land cover mapping with fine-grained classes is effectively improved.",
+        "bbox": [
+            80,
+            333,
+            467,
+            679
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            80,
+            694,
+            215,
+            709
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Large-Scale Land Cover Mapping. Land cover mapping classifies each pixel of images into a unique class, which can be regarded as a semantic segmentation problem [6, 11]. Large-scale land cover mapping with fine-grained classes can provide necessary information for earth system science studies and land spatial layout optimization. However, manual or semi-manual labeling is high-cost and time-consuming [34]. Besides, the classification systems of most large-scale studies are coarse. Thus the provided land cover information is limited. For example, the Chesapeake dataset covering  $160,000\\mathrm{km}^2$  costs $1.3 million but only contains four classes [33]. DynamicEarthNet",
+        "bbox": [
+            80,
+            720,
+            467,
+            898
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "dataset covers  $17,000\\mathrm{km}^2$  and includes 7 basic land cover classes [36]. To overcome the limited area of the labeled data and improve the model generalization, some works utilize openly available produced data (e.g., OpenStreetMap data and land cover products), which contains many noisy labels but can provide prior information in a large-scale area [19, 10]. These works still use coarse classification systems of land cover mapping. Five-Billion-Pixels [37] is a large-scale land cover dataset with 24 categories, covering about  $60,000\\mathrm{km}^2$ , which opens the door to large-scale land cover mapping with fine-grained classes. However, this dataset suffers from a severe class-imbalance problem (the common categories cover hundreds of times more pixels than the rare ones), due to the difficulty of data collection and label annotation varying by class.",
+        "bbox": [
+            503,
+            90,
+            888,
+            316
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Class-Imbalanced Learning. To solve the class-imbalance problem, related works include data resampling [3, 4], loss re-weighting [22, 32, 18], margin modification [5, 35], and decoupled learning [20, 47]. In self-supervised learning and SSL, recent works devote to rebalancing the class distribution of SSL data by assigning pseudo-labels for the unlabeled data to address the class imbalance issue [41, 15]. For example, Gui et al. [14] propose a class-aware pseudo-labeling method for SSL, dynamically adjusting the threshold for selecting pseudo-labels to obtain better performance on minority classes. Similarly, Hu et al. [17] propose a label bias removal method and dynamically determine the threshold for each class to assign more pseudo-labels for minority classes. However, the hard samples with relatively low confidence scores are ignored for optimizing the model, leading to sub-optimal performance.",
+        "bbox": [
+            503,
+            321,
+            888,
+            563
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Semi-Supervised Semantic Segmentation. SSL methods leverage the relationships between the labeled and unlabeled data to improve the model accuracy and generalization, alleviating the limited area of the labeled data [8]. Semi-supervised semantic segmentation methods include entropy minimization [38], consistency regularization [30], and pseudo-labeling [42]. Pseudo-labeling is widely used in semi-supervised semantic segmentation, and many works contribute to selecting pseudo-labels more steadily and correctly [29]. For example, Tong et al. [37] propose a dynamic pseudo-label assignment method, in which the number of pseudo-labels is dynamically increased with training iterations. However, these methods still rely on the manual setting of hyperparameters and suffer from the uncertainty of pseudo-labels. Alternatively, contrastive learning methods in supervised learning [39, 16] and SSL [46, 40] constrain the representations of positive samples against these negative samples in feature space, leading to promising results. Different from the manner of data augmentation-based positive sample construction [43, 23] and memory bank-based negative sample construction [1, 40], we leverage the ground-truth and pseudo-labels as a guide to decom",
+        "bbox": [
+            503,
+            569,
+            888,
+            898
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "16784",
+        "bbox": [
+            482,
+            946,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "pose the feature embedding space, and estimate the embedding centers for each class per image to stably select easy and hard samples for the proposed Class Center Contrast method.",
+        "bbox": [
+            75,
+            90,
+            470,
+            151
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Methodology",
+        "text_level": 1,
+        "bbox": [
+            76,
+            166,
+            212,
+            184
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. The Overall Workflow of Our Approach",
+        "text_level": 1,
+        "bbox": [
+            76,
+            191,
+            421,
+            209
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our approach consists of four steps. First, a baseline model is trained on the labeled data. Second, we propose a class-aware unlabeled data selection method to build an SSL dataset that is more balanced towards minority classes. The baseline model is used to predict land cover classes on all collected data from a large unlabeled data pool and calculate the statistics of the class distribution for each image patch. The image patches containing pixels of minority classes are selected as unlabeled data. Then the SSL dataset is constructed by combining the labeled data and the selected unlabeled data. Third, we propose a Class-Balanced Cross Entropy loss and a Class Center Contrast method. To alleviate the class imbalance problem, we reweight the cross entropy loss on labeled data in both sample and class level by jointly considering the annotation bias and the class bias on the built SSL dataset. Besides, we decompose the feature embedding space using the ground truth and pseudo-labels, and propose an embedding center selection and contrast method for class-balanced and annotation-guided learning on unlabeled data. The pipeline of the proposed class-aware SSL model is shown in Figure 2. Finally, we perform model inference for large-scale land cover mapping.",
+        "bbox": [
+            75,
+            215,
+            472,
+            564
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Class-Aware Unlabeled Data Selection",
+        "text_level": 1,
+        "bbox": [
+            76,
+            574,
+            406,
+            589
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "A suitable set selection of unlabeled data is important for SSL model development. Inspired by the observation in [41] that existing SSL algorithms can produce pseudolabels on minority classes with high precision, we propose an unlabeled data selection strategy that tends to select unlabeled image patches containing pixels with minority classes.",
+        "bbox": [
+            75,
+            598,
+            468,
+            703
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "First, we train a semantic segmentation model using labeled data. We apply the off-the-shelf class balance strategy [4] designed for supervised learning to guarantee the model ability of minority class estimation. The definition of minority classes depends on the class proportion and the class accuracy on the labeled data. Then, we perform model inference on all candidate unlabeled data to predict the land cover class of each pixel, thereby obtaining the class distribution of each candidate image patch. We denote the proportion of pixels of all minority classes in an image patch as  $P_{m}$ . Finally, we formulate a principle to automatically select unlabeled data according to the class distribution of each candidate image patch. Specifically, the pro",
+        "bbox": [
+            75,
+            704,
+            470,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "portions of the 24 classes on the labeled dataset are shown in Figure 3. The classes with a proportion lower than  $1\\%$  generally achieve an accuracy of less than  $30\\%$  within our dataset. We collect all the images with  $P_{m} \\geq 1\\%$ , which leads to about 25,000 images over the about 500,000 candidates (see Section 4.3). Then we randomly select another 25,000 images from those with  $0 < P_{m} < 1\\%$ , leading to a total of 50,000 unlabeled images, which is five times the number of the labeled training data. The authors suggest selecting unlabeled patches with relatively more pixels of minority classes (i.e., larger  $P_{m}$ ), and those encompassing rare classes characterized by fewer pixels within the minority classes. Note that although the prediction on the unlabeled data is not absolutely accurate, it can preserve image patches that are most likely to contain pixels with minority classes, which contributes to building a more balanced SSL dataset.",
+        "bbox": [
+            496,
+            90,
+            893,
+            347
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3. Class-Balanced Learning",
+        "text_level": 1,
+        "bbox": [
+            498,
+            357,
+            730,
+            373
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In our built SSL dataset, the class distribution of unlabeled data differs from that of the labeled data. It can be regarded as an annotation bias problem on the entire dataset. Besides, the SSL dataset still suffers from a class imbalance issue. Consequently, using the class prior information estimated on the labeled data to the re-weight loss function is inappropriate. To solve the above issues, we propose a class-balanced learning method to jointly remove the annotation bias and class bias on the SSL dataset.",
+        "bbox": [
+            496,
+            380,
+            890,
+            515
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "First, we re-weight samples of labeled data to remove the annotation bias. Owing to our unlabeled data selection method, the entire SSL dataset is more balanced towards minority classes than the original labeled dataset. Therefore, we can estimate the class prior on the entire SSL dataset to partially remove the annotation bias. Specifically, we denote  $L$  as a label indicator, where  $L = 0$  and  $L = 1$  represent the unlabeled and labeled data, respectively. The model parameters  $\\theta$  of labeled data in traditional SSL methods are obtained by maximizing likelihood estimation in the formula (1).",
+        "bbox": [
+            496,
+            516,
+            892,
+            681
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\hat {\\theta} = \\arg \\max  _ {\\theta} \\log P (Y | X, L = 1; \\theta) \\\\ = \\arg \\max  _ {\\theta} \\sum_ {(x, y) \\in D _ {L = 1}} \\log P (y | x; \\theta), \\tag {1} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            563,
+            704,
+            890,
+            766
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\hat{\\theta}$  is the estimated model parameter,  $X$  represents samples,  $Y$  represents classes, and  $D_{L = 1}$  denotes the set of labeled data. Due to the annotation bias in the SSL dataset, we have  $P(y|x;L = 1)\\neq P(y|x)$ . According to causal inference [13, 17], in the entire SSL dataset,  $X\\rightarrow Y\\rightarrow L$  can be regarded as a one-way chain, where the class is dependent on its pixel, and whether the pixel has a label or not depends on its class. According to the",
+        "bbox": [
+            496,
+            777,
+            893,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "16785",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/4374f975a37edd04ac6a43821175eb63e023e918e341f7644a90e03609a8add0.jpg",
+        "image_caption": [
+            "Figure 2. The pipeline of the proposed Class-Aware Semi-Supervised Semantic Segmentation framework with class-balanced learning, embedding center estimation, and class center contrast. The class-aware unlabeled data selection process is not shown in the figure."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            92,
+            851,
+            349
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "conditional independence rule, we obtain the expression,  $P(X|Y,L = 1) = P(X|Y)$  when conditioning on Y. Therefore, estimating the model parameters of the entire SSL dataset can be formalized as in the formula (2) [17], where  $s(x,y) = \\frac{\\log P(y|x;\\theta) - \\log P(y;\\theta)}{\\log P(y|x;\\theta)}$  is the loss weight of each labeled sample  $(x,y)$ . Thus, it can be interpreted as modifying the class prior  $P(y;\\theta)$  to remove the annotation bias of the labeled data.",
+        "bbox": [
+            75,
+            424,
+            473,
+            547
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\hat {\\theta} = \\arg \\max  _ {\\theta} \\log P (X | Y; \\theta) \\\\ = \\arg \\max  _ {\\theta} \\sum_ {(x, y) \\in D _ {L = 1}} \\log P (x | y; \\theta) \\\\ = \\arg \\max  _ {\\theta} \\sum_ {(x, y) \\in D _ {L = 1}} \\log \\frac {P (y | x ; \\theta) P (x ; \\theta)}{P (y ; \\theta)} \\tag {2} \\\\ = \\arg \\max  _ {\\theta} \\sum_ {(x, y) \\in D _ {L = 1}} \\log \\frac {P (y | x ; \\theta)}{P (y ; \\theta)} \\\\ = \\arg \\max  _ {\\theta} \\sum_ {(x, y) \\in D _ {L = 1}} s (x, y) \\cdot \\log P (y | x; \\theta), \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            107,
+            569,
+            468,
+            751
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Second, although the SSL dataset is more balanced than the labeled data, it still suffers from the class imbalance problem. Therefore, we further estimate the class bias on the SSL dataset. A widely used class re-weighting loss function on supervised learning can be formalized as follows:",
+        "bbox": [
+            75,
+            760,
+            468,
+            849
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {w c e} = \\sum_ {k = 1} ^ {K} W _ {k} \\sum_ {i = 1} ^ {n _ {k}} y _ {i k} \\log p _ {i k}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            862,
+            468,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $W_{k} = \\frac{1}{\\log(1 + r_{k})}$  and  $r_k$  is the ratio of the class  $k$  on the labeled data.  $K$  is the number of classes,  $n_k$  is the number of samples of class  $k$ ,  $y_{ik}$  and  $p_{ik}$  is the ground-truth and predicted probability of the  $i$ -th sample for class  $k$ , respectively. In SSL, we approximate  $r_k$  with  $P(y; \\theta)$ , the class distribution on the entire dataset.",
+        "bbox": [
+            496,
+            422,
+            890,
+            516
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Finally, we perform an online estimation strategy for  $P(y; \\theta)$  to simplify the training process. We compute the average of class probability distributions over all pixels in a mini-batch, which is denoted as  $P(Y; B_t, \\theta_t)$ .  $B_t$  and  $\\theta_t$  are the samples and model parameters of the current batch, respectively. Then the approximate class distribution, denoted as  $\\tilde{P}(Y)$ , is updated by the weighted sum of the obtained class distribution of this mini-batch and the approximate class distribution in the previous iteration, which is formalized as follows:",
+        "bbox": [
+            496,
+            520,
+            892,
+            671
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\tilde {P} (Y) \\leftarrow \\mu \\tilde {P} (Y) + (1 - \\mu) P (Y; B _ {t}, \\theta_ {t}), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            557,
+            700,
+            890,
+            719
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mu \\in [0,1]$  is a weighting coefficient. We set  $\\mu = \\frac{B}{N}$ , in which  $B$  is the batch size and  $N$  is the total number of samples in the dataset. We initialize  $\\tilde{P}(Y)$  with the class distribution calculated on the labeled data. As such, the proposed Class-Balanced Cross Entropy loss on the labeled data is formulated as follows:",
+        "bbox": [
+            496,
+            742,
+            892,
+            833
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {c b c e} = \\sum_ {k = 1} ^ {K} W _ {k} \\sum_ {i = 1} ^ {n _ {k}} s \\left(x _ {i k}, y _ {i k}\\right) y _ {i k} \\log p _ {i k}. \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            553,
+            862,
+            890,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "16786",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4. Class Center Contrast",
+        "text_level": 1,
+        "bbox": [
+            76,
+            90,
+            284,
+            104
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training via pseudo-labeling is commonly used to utilize unlabeled data in SSL. In fine-grained classes, assigning reliable labels for minority classes is challenging because the segmentation model tends to be biased towards the majority classes [26]. Besides, existing label assignment methods usually use only high-confidence predictions, which may offer limited information to model learning. In this work, our method selects reliable and hard samples. Moreover, instead of directly using the pseudo-labels for supervised learning on unlabeled data, we leverage the pseudo-labels as the guide to decompose the feature embedding space. In general, the distance of feature embeddings from the same class should be minimized, while the distance of those from different classes should be maximized. Therefore, we propose an online embedding center selection method that divides each class into hard and easy sample regions on each image. Specifically, we use the prediction probabilities of ground truth or pseudo-labels to separate easy and hard samples. The predicted probability of each pixel is associated with its feature embedding, and we then calculate the average prediction probability of each class for each image. Feature embeddings with above-average probabilities for the corresponding class are considered easy samples and vice versa. Thus we can obtain three feature embedding centers, including easy, hard, and all feature embedding centers, as contrastive samples. With this method, on the one hand, the obtained feature embeddings for each class are with relatively high confidence, as the samples are close to feature local centers for each class in the embedding space. On the other hand, these feature embeddings include hard samples of each class.",
+        "bbox": [
+            76,
+            113,
+            472,
+            583
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Inspired by the existing contrastive learning method [39], we perform cross-image contrastive learning in each mini-batch. Since both labeled and unlabeled data are trained in a mini-batch, the embeddings of the labeled data could help the feature embedding learning of the unlabeled data. Formally, denote  $P_{m}$  and  $N_{m}$  as the feature embedding collections of the positive and negative samples in a mini-batch for the  $m$ -th feature embedding center  $c_{m}$ . The proposed Class Center Contrastive loss for the  $m$ -th feature embedding center is formulated as follows:",
+        "bbox": [
+            75,
+            583,
+            470,
+            734
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {c c t} ^ {m} = - \\frac {1}{| P _ {m} |} \\sum_ {\\boldsymbol {c} _ {m} ^ {+} \\in P _ {m}} \\log \\frac {e ^ {\\left(\\boldsymbol {c} _ {m} \\cdot \\boldsymbol {c} _ {m} ^ {+} / \\tau\\right)}}{e ^ {\\left(\\boldsymbol {c} _ {m} \\cdot \\boldsymbol {c} _ {m} ^ {+} / \\tau\\right)} + \\sum_ {\\boldsymbol {c} _ {m} ^ {-} \\in N _ {m}} e ^ {\\left(\\boldsymbol {c} _ {m} \\cdot \\boldsymbol {c} _ {m} ^ {-} / \\tau\\right)}}, \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            76,
+            758,
+            485,
+            821
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "in which  $\\tau$  is the temperature parameter,  $c_{m}^{+}$  and  $c_{m}^{-}$  are the positive and negative samples, respectively.",
+        "bbox": [
+            76,
+            823,
+            467,
+            853
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The overall loss function can be formulated as follows:",
+        "bbox": [
+            96,
+            854,
+            460,
+            868
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c b c e} + \\lambda \\mathcal {L} _ {c c t}, \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            204,
+            885,
+            468,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda$  is the loss weight.",
+        "bbox": [
+            500,
+            90,
+            679,
+            106
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Data",
+        "text_level": 1,
+        "bbox": [
+            500,
+            125,
+            566,
+            138
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Image Collection and Preparation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            148,
+            797,
+            165
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Considering free access, multispectral bands, and relatively high spatial resolution, all satellite images used in this work are acquired from Sentinel-2 satellites. Sentinel-2 imagery has 13 spectral bands at  $10\\mathrm{m}$  (R, G, B, and near-infrared bands),  $20\\mathrm{m}$  (six red edge and shortwave infrared bands), and  $60\\mathrm{m}$  (three atmospheric correction bands) resolutions [12]. The abundant multispectral information and up to  $10\\mathrm{m}$  spatial resolution benefit land cover mapping with a fine-grained classification system.",
+        "bbox": [
+            496,
+            174,
+            890,
+            309
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We collect Sentinel-2 images with radiometric correction, geometric corrections, and atmospheric correction, covering more than 60 cities in China. We apply an off-the-shelf Sentinel-2 sharpening method, DSen2Net [24], to reconstruct low-resolution bands and uniform the resolution of all bands of Sentinel-2 images to  $10\\mathrm{m}$ .",
+        "bbox": [
+            496,
+            311,
+            890,
+            401
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Label Collection and Classification System",
+        "text_level": 1,
+        "bbox": [
+            500,
+            415,
+            862,
+            431
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The land cover labels and classification system are from the Five-Billion-Pixels dataset [37], labeled with  $4\\mathrm{m}$  resolution. It includes 24 land cover classes. We resize the  $4\\mathrm{m}$  labels to  $10\\mathrm{m}$  to match Sentinel-2 images. The labeled area covers about  $60,000\\mathrm{km}^2$  in China. The classification system includes the industrial area (C1), paddy field (C2), irrigated field (C3), dry cropland (C4), garden land (C5), arbor forest (C6), shrub forest (C7), park (C8), natural meadow (C9), artificial meadow (C10), river (C11), urban residential (C12), lake (C13), pond (C14), fish pond (C15), snow (C16), bare land (C17), rural residential (C18), stadium (C19), square (C20), road (C21), overpass (C22), railway station (C23), and airport (C24).",
+        "bbox": [
+            496,
+            440,
+            890,
+            636
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3. Dataset",
+        "text_level": 1,
+        "bbox": [
+            500,
+            648,
+            594,
+            664
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Examples of the labeled data are shown in Figure 1. The training, validation, and test datasets include 9,995, 2,000, and 2,503 images with a size of  $256 \\times 256$ , respectively, according to the geographical division. The unlabeled data is from more than 60 dispersed administrative districts in China and has no overlap with the labeled data. The number of candidate unlabeled images is about 500,000, each with a size of  $256 \\times 256$ . According to our minority class-biased unlabeled data selection strategy, we filter 50,000 images to build the unlabeled set, which covers more scenes with minority classes. Figure 3 shows the proportion of each land cover class on labeled training data, unlabeled training data, and all training data. The majority classes cover hundreds of times more pixels than the minority classes on the labeled training data.",
+        "bbox": [
+            496,
+            674,
+            890,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "16787",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/66af3d5eaff49cc0c9ef9bd8aca6e0f9f856cf57c02cf9dde286574d88af7d03.jpg",
+        "image_caption": [
+            "Figure 3. Percentage of each land cover class on the labeled training data, unlabeled training data, and all training data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            92,
+            890,
+            390
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/1e6d46ac96b704cc9c89296589fcec31ff0587896a4d10eea3a4cdc437201d23.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>OA</td><td>mAcc</td><td>mIoU</td><td>C1</td><td>C2</td><td>C3</td><td>C4</td><td>C5</td><td>C6</td><td>C7</td><td>C8</td><td>C9</td><td>C10</td></tr><tr><td>Baseline</td><td>73.24</td><td>45.77</td><td>35.61</td><td>71.24</td><td>66.25</td><td>88.66</td><td>17.90</td><td>16.93</td><td>62.26</td><td>27.01</td><td>17.26</td><td>44.65</td><td>11.33</td></tr><tr><td>SimCLR [7] +finetune</td><td>74.07</td><td>45.81</td><td>35.17</td><td>74.49</td><td>63.33</td><td>90.88</td><td>14.10</td><td>17.80</td><td>62.10</td><td>23.57</td><td>18.39</td><td>48.42</td><td>21.02</td></tr><tr><td>Pseudo -labeling [25]</td><td>74.61</td><td>48.83</td><td>38.03</td><td>76.85</td><td>69.60</td><td>91.55</td><td>13.47</td><td>20.52</td><td>59.19</td><td>31.63</td><td>37.53</td><td>44.90</td><td>10.68</td></tr><tr><td>Advent [38]</td><td>74.81</td><td>49.25</td><td>38.98</td><td>75.88</td><td>68.38</td><td>91.98</td><td>12.20</td><td>15.01</td><td>62.46</td><td>25.91</td><td>44.77</td><td>54.75</td><td>8.10</td></tr><tr><td>CADR [17]</td><td>75.49</td><td>47.83</td><td>36.63</td><td>76.32</td><td>52.43</td><td>91.48</td><td>21.54</td><td>24.35</td><td>64.74</td><td>31.85</td><td>22.59</td><td>59.88</td><td>17.25</td></tr><tr><td>DPA [37]</td><td>75.36</td><td>48.57</td><td>37.77</td><td>76.05</td><td>69.40</td><td>92.33</td><td>15.51</td><td>19.79</td><td>66.30</td><td>27.05</td><td>23.25</td><td>52.41</td><td>12.48</td></tr><tr><td>Ours</td><td>75.97</td><td>53.53</td><td>40.68</td><td>81.71</td><td>80.07</td><td>88.89</td><td>22.27</td><td>37.32</td><td>67.06</td><td>34.87</td><td>26.44</td><td>56.73</td><td>17.16</td></tr><tr><td>C11</td><td>C12</td><td>C13</td><td>C14</td><td>C15</td><td>C16</td><td>C17</td><td>C18</td><td>C19</td><td>C20</td><td>C21</td><td>C22</td><td>C23</td><td>C24</td></tr><tr><td>70.86</td><td>82.76</td><td>78.93</td><td>24.49</td><td>57.71</td><td>3.04</td><td>48.67</td><td>77.23</td><td>10.80</td><td>7.90</td><td>71.42</td><td>45.83</td><td>34.95</td><td>60.34</td></tr><tr><td>71.47</td><td>83.27</td><td>79.41</td><td>30.54</td><td>64.73</td><td>0.75</td><td>48.77</td><td>68.82</td><td>2.00</td><td>0</td><td>73.59</td><td>44.53</td><td>33.87</td><td>63.72</td></tr><tr><td>72.40</td><td>82.90</td><td>83.56</td><td>16.94</td><td>52.44</td><td>13.66</td><td>55.06</td><td>72.58</td><td>21.05</td><td>13.03</td><td>70.27</td><td>60.60</td><td>38.78</td><td>62.83</td></tr><tr><td>69.98</td><td>83.19</td><td>83.38</td><td>19.72</td><td>49.34</td><td>15.50</td><td>46.56</td><td>75.16</td><td>22.40</td><td>17.55</td><td>70.99</td><td>58.04</td><td>41.62</td><td>69.23</td></tr><tr><td>65.90</td><td>86.27</td><td>84.53</td><td>26.09</td><td>72.57</td><td>0</td><td>51.78</td><td>74.96</td><td>0.95</td><td>0</td><td>65.23</td><td>48.51</td><td>33.64</td><td>75.02</td></tr><tr><td>71.97</td><td>83.33</td><td>81.80</td><td>16.77</td><td>50.66</td><td>0.88</td><td>52.37</td><td>72.61</td><td>21.42</td><td>6.64</td><td>72.15</td><td>59.10</td><td>49.15</td><td>72.37</td></tr><tr><td>70.61</td><td>82.86</td><td>81.69</td><td>17.13</td><td>78.53</td><td>26.28</td><td>55.80</td><td>80.69</td><td>31.83</td><td>26.49</td><td>66.23</td><td>64.65</td><td>46.39</td><td>42.97</td></tr></table>",
+        "bbox": [
+            86,
+            435,
+            883,
+            693
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Quantitative comparison with different methods. We also show the precision of each class. Bold indicates the best results  $(\\%)$",
+        "bbox": [
+            76,
+            696,
+            875,
+            712
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            734,
+            209,
+            752
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Implementation Details and Metrics",
+        "text_level": 1,
+        "bbox": [
+            76,
+            763,
+            390,
+            781
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We choose ResUNet [9] in our experiments, as it is stable and lightweight for land cover mapping. We train for 100 epochs for all experiments using stochastic gradient descent (SGD) with a momentum of 0.9 and a weight decay of  $10^{-4}$ . The initial learning rate is set to 0.05. We adopt the batch size as 12, including 6 labeled images and 6 unlabeled images. The loss weight  $\\lambda$  is set to 0.005. The temperature",
+        "bbox": [
+            75,
+            794,
+            470,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "$\\tau$  of contrastive loss is set to 0.07.",
+        "bbox": [
+            500,
+            736,
+            723,
+            750
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The performance of the proposed approach and other competing methods are assessed with overall accuracy (OA), mean accuracy (mAcc), mean intersection over union (mIoU), and precision of each class.",
+        "bbox": [
+            498,
+            753,
+            890,
+            814
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2. Comparison Results",
+        "text_level": 1,
+        "bbox": [
+            500,
+            829,
+            692,
+            845
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1 shows the comparison results with typical and state-of-the-art class-balanced pseudo-labeling methods. For a fair comparison, all competing methods and our",
+        "bbox": [
+            498,
+            854,
+            892,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "16788",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/a6759b955130e942fd9658c3df227bdcea20218473c6fb420ad98b36f2536fd3.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>Random selection</td><td>75.50</td><td>51.35</td><td>39.09</td></tr><tr><td>Majority class -biased selection</td><td>75.75</td><td>49.98</td><td>39.33</td></tr><tr><td>Minority class -biased selection</td><td>75.97</td><td>53.53</td><td>40.68</td></tr></table>",
+        "bbox": [
+            130,
+            88,
+            416,
+            189
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/53c82a41da47d48dd81f00bdb2d8097ccbb54a351ff4aa10a5e1221579f63e06.jpg",
+        "table_caption": [
+            "Table 2. Ablation study on the proposed class-aware unlabeled data selection method. Bold indicates the best results  $(\\%)$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>Without class balance</td><td>75.41</td><td>48.16</td><td>37.27</td></tr><tr><td>Class-balanced method in [38]</td><td>73.61</td><td>51.33</td><td>38.61</td></tr><tr><td>Class-balanced method in [37]</td><td>75.51</td><td>49.66</td><td>35.51</td></tr><tr><td>Ours</td><td>75.97</td><td>53.53</td><td>40.68</td></tr></table>",
+        "bbox": [
+            78,
+            239,
+            475,
+            321
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Comparison results on different class-balanced learning methods. Bold indicates the best results  $(\\%)$ .",
+        "bbox": [
+            76,
+            330,
+            467,
+            357
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "method use the same model architecture. All experiments are measured on the same computing platform. The baseline applies the supervised learning method to the labeled data. The SimCLR+finetune method uses a self-supervised learning method [7] on the unlabeled data to obtain a pretrain model and performs fine-tuning on the labeled data. Pseudo-labeling method [25] first generates pseudo-labels of the unlabeled data by using the baseline model, then retransmits the model with ground truth and pseudo-labels. Advent [38], DPA [37], and CADR [17] are three state-of-the-art SSL methods.",
+        "bbox": [
+            75,
+            383,
+            467,
+            547
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The results demonstrate that our approach significantly improved over  $4.28\\%$  in mAcc,  $1.70\\%$  in mIoU, and  $0.58\\%$  in OA. Compared to the baseline for each class, our method can significantly improve the accuracy of the minority classes and maintain the accuracy of the majority classes simultaneously. In contrast, the competing methods often tend to predict the pixels to majority classes and suffer from low precision on minority classes. Regarding the computational cost, our method incurs  $53\\%$  additional training time than the basic SSL algorithm (i.e., Advent) due to Class-Balanced Learning and Class Center Contrast. The inference times are the same for all SSL methods, as we use the same backbone. Please find the visual comparison in the supplementary.",
+        "bbox": [
+            75,
+            549,
+            467,
+            762
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            76,
+            770,
+            228,
+            787
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Class-Aware Unlabeled Data Selection. We compare our method with the opposite strategy (i.e., majority class-biased data selection) and widely opted random selection method. Table 2 shows that our method achieves the best results. On the contrary, the mAcc value of the majority class-biased data selection method is unsatisfactory because the SSL model is biased toward predicting the majority classes.",
+        "bbox": [
+            75,
+            794,
+            467,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/10d83bb2ceae84f2e5bf7fb6a2db8ac3888294dc074b804f6a1d10733d745232.jpg",
+        "image_caption": [
+            "Figure 4. Results of using the different number of labeled images (i.e., 500, 1,000, and 3,000) and the different ratios of unlabeled to labeled images in SSL."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            89,
+            890,
+            263
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Besides, to verify the effectiveness of unlabeled data selection, we show the statistics of class distribution in Figure 3. Estimating the class distribution of unlabeled data is based on the pseudo-labels produced by the baseline model. Owing to the proposed class-aware unlabeled data selection method, the class distribution of the unlabeled data is more balanced than that of the labeled data. As a result, the entire training dataset is more balanced towards minority classes than the original labeled dataset. Therefore, we can estimate the class prior information on the entire dataset to partially remove the annotation bias.",
+        "bbox": [
+            496,
+            351,
+            890,
+            517
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Class-Balanced Learning. In general, class weighting sacrifices the accuracy of majority classes to improve the accuracy of minority classes, which decreases OA and mIoU. From Table 3, two comparison methods suffer from the above problem, while our method improves the mAcc, OA, and mIoU compared to the baseline (i.e., without loss weights). The reason is that the two comparison methods use class prior information of labeled data and ignore the discrepancy of class distributions between labeled and unlabeled data in SSL. Besides, Table 4 shows the ablation study on Class-Balanced Cross Entropy loss. Utilizing sample re-weighting improves performance by removing the annotation bias on the semi-supervised learning dataset. Using class re-weighting for cross-entropy loss removes class bias and improves the performance of mAcc and mIoU by a large margin. Furthermore, jointly utilizing the sample and class weights achieves the best results in Table 4.",
+        "bbox": [
+            496,
+            520,
+            890,
+            790
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Class Center Contrast Method. Table 5 shows the ablation study on the proposed class center contrast method. Contrastive loss utilizes inter-class and intra-class information and achieves better results than center loss used in [28]. Compared with only training contrastive loss with pseudolabels on unlabeled data, using pseudo-labels on both labeled and unlabeled data achieves better results. Further-",
+        "bbox": [
+            496,
+            794,
+            890,
+            898
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "16789",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/a32f4d62a56ee63264f1724813b1316e98741cd4edf88c634ee78c788460f4d3.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Cross entropy loss</td><td>Sample re-weighting</td><td>Class re-weighting</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>✓</td><td></td><td></td><td>75.41</td><td>48.16</td><td>37.27</td></tr><tr><td>✓</td><td>✓</td><td></td><td>75.74</td><td>48.96</td><td>38.40</td></tr><tr><td>✓</td><td></td><td>✓</td><td>75.43</td><td>52.06</td><td>40.07</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>75.97</td><td>53.53</td><td>40.68</td></tr></table>",
+        "bbox": [
+            166,
+            87,
+            803,
+            167
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/3a3ef5b9cf4e13b3adf267b42001abaab364006405379d2ac79606baa42796f0.jpg",
+        "table_caption": [
+            "Table 4. Ablation study on class-balanced cross entropy loss. Bold indicates the best results  $(\\%)$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Center loss</td><td>Contrastive loss</td><td>Pseudo label</td><td>Ground truth</td><td>On labeled data</td><td>On unlabeled data</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>✓</td><td></td><td>✓</td><td></td><td></td><td>✓</td><td>74.07</td><td>48.13</td><td>37.34</td></tr><tr><td></td><td>✓</td><td>✓</td><td></td><td></td><td>✓</td><td>75.23</td><td>50.06</td><td>39.29</td></tr><tr><td></td><td>✓</td><td>✓</td><td></td><td>✓</td><td>✓</td><td>75.26</td><td>51.42</td><td>39.43</td></tr><tr><td></td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>75.97</td><td>53.53</td><td>40.68</td></tr></table>",
+        "bbox": [
+            135,
+            203,
+            834,
+            299
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/45efef6b56a47580d8a4a354b5af92acbf71321f28fc5d16b255426c7de47944.jpg",
+        "table_caption": [
+            "Table 5. Ablation study on the class center contrast method. Bold indicates the best results  $(\\%)$"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>Random sample selection</td><td>74.68</td><td>49.31</td><td>38.80</td></tr><tr><td>Simple sample selection</td><td>74.74</td><td>51.57</td><td>40.42</td></tr><tr><td>Hard sample selection</td><td>74.57</td><td>51.14</td><td>40.31</td></tr><tr><td>Feature embedding centers</td><td>75.97</td><td>53.53</td><td>40.68</td></tr></table>",
+        "bbox": [
+            86,
+            345,
+            460,
+            425
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/e259d892fbe004a71ca134cd182697c5d7c50116bbad327954eaca1f14937489.jpg",
+        "table_caption": [
+            "Table 6. Results of different contrastive sample selection manners. Bold indicates the best results  $(\\%)$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>Within-image method</td><td>73.36</td><td>53.48</td><td>40.03</td></tr><tr><td>Cross-image method</td><td>75.97</td><td>53.53</td><td>40.68</td></tr></table>",
+        "bbox": [
+            102,
+            479,
+            444,
+            529
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "more, introducing the ground truth of labeled data to the training of contrastive loss can significantly improve performance. We also compare different sample selection strategies for contrastive learning. We select three contrastive samples for each class per image for a fair comparison. As shown in Table 6, leveraging high-confidence predicted samples from the unlabeled data for semi-supervised learning yields better results than employing uncertain samples. Nonetheless, samples with high-confidence predictions provide limited information for model learning, leading to only marginal enhancements. The proposed Class Center Contrast method involves both high-confidence and hard samples, achieving the best results. Besides, we compare within-image and cross-image contrastive learning results in Table 7. Cross-image method uses labeled and unlabeled images in a mini-batch and achieves better results, while the within-image method only uses local information of labeled or unlabeled images.",
+        "bbox": [
+            75,
+            595,
+            468,
+            867
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Hyperparameter Tuning of Loss Weight.  $\\lambda$  is to balance two loss terms in the formula (7). Table 8 shows the",
+        "bbox": [
+            76,
+            869,
+            468,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/11eb0834942d54c4f2a712e19739ea1fba553cf61a5ca055604034830108765f.jpg",
+        "table_caption": [
+            "Table 7. Results of within-image and cross-image contrastive learning. Bold indicates the best results  $(\\%)$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Loss weight</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>λ = 0.05</td><td>75.07</td><td>51.62</td><td>40.15</td></tr><tr><td>λ = 0.01</td><td>75.26</td><td>51.42</td><td>39.43</td></tr><tr><td>λ = 0.005</td><td>75.97</td><td>53.53</td><td>40.68</td></tr><tr><td>λ = 0.001</td><td>75.27</td><td>53.84</td><td>40.26</td></tr></table>",
+        "bbox": [
+            555,
+            345,
+            836,
+            425
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 8. Hyperparameter tuning of loss weight. Bold indicates the best results  $(\\%)$ .",
+        "bbox": [
+            500,
+            435,
+            890,
+            462
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "results using different loss weights  $\\lambda$ . We conclude that a large loss weight for  $\\mathcal{L}_{cct}$  reduces the model performance. In this work, we set  $\\lambda = 0.005$ .",
+        "bbox": [
+            498,
+            488,
+            890,
+            532
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The Number and Ratio of Labeled Images in SSL. We conduct a comparative analysis of the enhancement in accuracy resulting from the increase of unlabeled data. As shown in Figure 4, the utilization of a larger quantity of unlabeled data, specifically ten times the amount of labeled training data containing only 500 images, has been observed to result in a substantial improvement in accuracy. Meanwhile, the performance becomes unstable as the ratio between unlabeled and labeled images increases further because of the limited number of labeled images. As the amount of labeled data increased, using the unlabeled data, which is five times the amount of labeled data, can yield superior results compared to using just one time the amount of labeled data. Subsequently, the performance is relatively stable as the number of unlabeled data increases. Therefore, to balance the computational cost and improvement in accuracy, we use a ratio of 5:1 for unlabeled and labeled data.",
+        "bbox": [
+            496,
+            534,
+            890,
+            790
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4. Examples of Land Cover Mapping in China",
+        "text_level": 1,
+        "bbox": [
+            500,
+            801,
+            870,
+            816
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As an application of the proposed method, we produce large-scale land cover maps with 24 classes over 1 million  $\\mathrm{km}^2$ . As shown in Figure 5, we compare our land cover mapping results with two existing public products produced by ESA [45] and Google [2]. The example in the top row",
+        "bbox": [
+            496,
+            825,
+            890,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "16790",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/99ab48134b1f93faa14350de580fb1f22554bef78b70c53c9974b5cf369dd558.jpg",
+        "image_caption": [
+            "Tree cover",
+            "Shrubland",
+            "Grassland",
+            "Cropland",
+            "Built-up",
+            "Bare/sparse.",
+            "vegetation",
+            "Snow and ice",
+            "Perm",
+            "water bodies",
+            "Herbaceous wetland",
+            "Mangroves",
+            "Moss and lichen",
+            "(a) ESA"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            83,
+            98,
+            269,
+            239
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/f7f357d0a572cca03a1ba1ff2bd5083e6e4a81c05ae6d022c9b05c5dd55adfd2.jpg",
+        "image_caption": [
+            "Water",
+            "Trees",
+            "Grass",
+            "Flooded",
+            "vegetation",
+            "Crops",
+            "Shrub &",
+            "Scrub",
+            "Built",
+            "Bare ground",
+            "Snow & ice"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            349,
+            98,
+            531,
+            238
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/0aa2424f575f40c7123070bad49ed39ca9a28a3c17339dbbfcbb6f86bc0e8fae.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            614,
+            98,
+            883,
+            239
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/e7b8435da0c4cbe62f2e8d8cc0317ea4202a3afe0372c0b324bbe76068c871f6.jpg",
+        "image_caption": [
+            "(b) Google",
+            "(c) Ours",
+            "Figure 5. Examples of Land Cover Mapping in China. The land cover products of ESA and Google are also generated from Sentinel-2 images but use different classification systems. In the black rectangles, our land cover results with fine-grained classes can provide more land cover information. In the red rectangles, our method recognizes the paddy field correctly, while the other two products classify it as cropland and water, respectively, due to the coarse classification system and the misleading spectral information of the paddy field."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            614,
+            243,
+            879,
+            383
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "shows the results of an urban area located in the east of China, and that in the bottom row gives the results of a rural area located in the center of China. The land cover products of ESA and Google contain 11 and 9 classes, while our classification system includes 24 classes. From the black rectangles, the two comparison products mainly distinguish built areas, water, trees, and crops. Owing to the fine-grained classification system and the proposed class-aware SSL methods, our results can recognize more land cover categories, including industrial area, urban residential, rural residential, road, overpass, railway station, arbor forest, artificial meadow, irrigated field, bare land, lake, river, pond, and fish pond. Our land cover maps with fine-grained classes can provide more land cover information. Besides, from the red rectangles, our method recognizes the paddy field correctly, while the other two products classify it as cropland and water, respectively, due to the coarse classification system and the misleading spectral information of the paddy field. From the above qualitative comparison, our method provides land cover maps with fine-grained classes, significantly improving the land cover information.",
+        "bbox": [
+            75,
+            505,
+            472,
+            824
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            76,
+            843,
+            194,
+            858
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we propose a unified Class-Aware Semi-Supervised Semantic Segmentation approach for large-",
+        "bbox": [
+            76,
+            869,
+            468,
+            902
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "scale land cover mapping with fine-grained classes. To address the class-imbalance issue, we propose a class-aware unlabeled data selection method to build an SSL dataset more balanced towards minority classes. Then, we propose a Class-Balanced Cross Entropy loss, considering the annotation and class biases on the built SSL dataset. Moreover, we propose the Class Center Contrast method to jointly use the labeled and unlabeled data for training. Experimental results validate the effectiveness of our method in large-scale land cover mapping with fine-grained classes.",
+        "bbox": [
+            496,
+            505,
+            893,
+            657
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            500,
+            710,
+            666,
+            728
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This research was supported in part by the National Natural Science Foundation of China (Grant No. T2125006), Jiangsu Innovation Capacity Building Program (Project No. BM2022028), China Postdoctoral Science Foundation (Grant No. 2023M731871), the German Federal Ministry of Education and Research (BMBF) in the framework of the international future AI lab \"AI4EO - Artificial Intelligence for Earth Observation: Reasoning, Uncertainties, Ethics and Beyond\" (Grant No. 01DD20001), and Shuimu Tsinghua Scholar Project.",
+        "bbox": [
+            496,
+            750,
+            893,
+            902
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "16791",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            173,
+            104
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Inigo Alonso, Alberto Sabater, David Ferstl, Luis Montesano, and Ana C Murillo. Semi-supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8219-8228, 2021. 2",
+            "[2] Christopher F Brown, Steven P Brumby, Brooke Guzder-Williams, Tanya Birch, Samantha Brooks Hyde, Joseph Mazzariello, Wanda Czerwinski, Valerie J Pasquarella, Robert Haertel, Simon Ilyushchenko, et al. Dynamic world, near real-time global  $10\\mathrm{m}$  land use land cover mapping. Scientific Data, 9(1):251, 2022. 8",
+            "[3] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural networks, 106:249-259, 2018. 2",
+            "[4] Jonathon Byrd and Zachary Lipton. What is the effect of importance weighting in deep learning? In International Conference on Machine Learning, pages 872-881, 2019. 2, 3",
+            "[5] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 2",
+            "[6] Bingyu Chen, Min Xia, and Junqing Huang. Mfanet: A multi-level feature aggregation network for semantic segmentation of land cover. Remote Sensing, 13(4):731, 2021. 2",
+            "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 6, 7",
+            "[8] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 2",
+            "[9] Foivos I Diakogiannis, François Waldner, Peter Caccetta, and Chen Wu. Resunet-a: A deep learning framework for semantic segmentation of remotely sensed data. ISPRS Journal of Photogrammetry and Remote Sensing, 162:94-114, 2020. 6",
+            "[10] Runmin Dong, Weizhen Fang, Haohuan Fu, Lin Gan, Jie Wang, and Peng Gong. High-resolution land cover mapping through learning with noise correction. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2021. 2",
+            "[11] Runmin Dong, Cong Li, Haohuan Fu, Jie Wang, Weijia Li, Yi Yao, Lin Gan, Le Yu, and Peng Gong. Improving 3-m resolution land cover mapping through efficient learning from an imperfect 10-m resolution map. Remote Sensing, 12(9):1418, 2020. 2",
+            "[12] Runmin Dong, Lixian Zhang, Weijia Li, Shuai Yuan, Lin Gan, Juepeng Zheng, Haohuan Fu, Lichao Mou, and Xiao Xiang Zhu. An adaptive image fusion method for sentinel-2 images and high-resolution images with long-time"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "intervals. International Journal of Applied Earth Observation and Geoinformation, 121:103381, 2023. 5",
+            "[13] Madelyn Glymour, Judea Pearl, and Nicholas P Jewell. Causal inference in statistics: A primer. John Wiley & Sons, 2016. 3",
+            "[14] Qian Gui, Xinting Wu, and Baoning Niu. Class-aware pseudo labeling for non-random missing labels in semi-supervised learning. In 2022 IEEE Eighth International Conference on Multimedia Big Data, pages 138-143. IEEE, 2022. 2",
+            "[15] Lanzhe Guo and Yufeng Li. Class-imbalanced semi-supervised learning with adaptive thresholding. In International Conference on Machine Learning, pages 8082-8094, 2022. 2",
+            "[16] Hanzhe Hu, Jinshi Cui, and Liwei Wang. Region-aware contrastive learning for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16291-16301, 2021. 2",
+            "[17] Xinting Hu, Yulei Niu, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. On non-random missing labels in semi-supervised learning. In International Conference on Learning Representations, 2022. 2, 3, 4, 6, 7",
+            "[18] Muhammad Abdullah Jamal, Matthew Brown, Ming-Hsuan Yang, Liqiang Wang, and Boqing Gong. Rethinking class-balanced methods for long-tailed visual recognition from a domain adaptation perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7610-7619, 2020. 2",
+            "[19] Noah Johnson, Wayne Treible, and Daniel Crispell. Opensentinelmap: A large-scale land use dataset using openstreetmap and sentinel-2 imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1333–1341, 2022. 2",
+            "[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. 2019. 2",
+            "[21] Krishna Karra, Caitlin Kontgis, Zoe Statman-Weil, Joseph C Mazzariello, Mark Mathis, and Steven P Brumby. Global land use/land cover with sentinel 2 and deep learning. In 2021 IEEE international Geoscience and Remote Sensing Symposium IGARSS, pages 4704–4707, 2021. 1",
+            "[22] Salman H Khan, Munawar Hayat, Mohammed Bennamoun, Ferdous A Sohel, and Roberto Togneri. Cost-sensitive learning of deep feature representations from imbalanced data. IEEE Transactions on Neural Networks and Learning Systems, 29(8):3573-3587, 2017. 2",
+            "[23] Donghyeon Kwon and Suha Kwak. Semi-supervised semantic segmentation with error localization network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9957-9967, 2022. 2",
+            "[24] Charis Lanaras, José Bioucas-Dias, Silvano Galliani, Emmanuel Baltsavias, and Konrad Schindler. Super-resolution of sentinel-2 images: Learning a globally applicable deep neural network. ISPRS Journal of Photogrammetry and Remote Sensing, 146:305–319, 2018. 5",
+            "[25] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks."
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "16792",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "In Workshop on challenges in representation learning, International Conference on Machine Learning, volume 3, page 896, 2013. 6, 7",
+            "[26] Ruihuang Li, Shuai Li, Chenhang He, Yabin Zhang, Xu Jia, and Lei Zhang. Class-balanced pixel-level self-labeling for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11593–11603, 2022. 5",
+            "[27] Weijia Li, Runmin Dong, Haohuan Fu, Jie Wang, Le Yu, and Peng Gong. Integrating google earth imagery with landsat data to improve 30-m resolution land cover mapping. Remote Sensing of Environment, 237:111563, 2020. 1",
+            "[28] Zhuohong Li, Hongyan Zhang, Fangxiao Lu, Ruoyao Xue, Guangyi Yang, and Liangpei Zhang. Breaking the resolution barrier: A low-to-high network for large-scale high-resolution land-cover mapping using low-resolution labels. ISPRS Journal of Photogrammetry and Remote Sensing, 192:244-267, 2022. 7",
+            "[29] Xiaoqiang Lu, Licheng Jiao, Fang Liu, Shuyuan Yang, Xu Liu, Zhixi Feng, Lingling Li, and Puhua Chen. Simple and efficient: A semisupervised learning framework for remote sensing image semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 2",
+            "[30] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12674-12684, 2020. 2",
+            "[31] Prem Chandra Pandey, Nikos Koutsias, George P Petropoulos, Prashant K Srivastava, and Eyal Ben Dor. Land use/land cover in view of earth observation: data sources, input dimensions, and classifiers—a review of the state of the art. Geocarto International, 36(9):957-988, 2021. 1",
+            "[32] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In International Conference on Machine Learning, pages 4334-4343, 2018. 2",
+            "[33] Caleb Robinson, Le Hou, Kolya Malkin, Rachel Soobitsky, Jacob Czawlytko, Bistra Dilkina, and Nebojsa Jojic. Large scale high-resolution land cover mapping with multi-resolution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12726-12735, 2019. 2",
+            "[34] Vladan Stojnic and Vladimir Risojevic. Self-supervised learning of remote sensing scene representations using contrastive multiview coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1182-1191, 2021. 2",
+            "[35] Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11662-11671, 2020. 2",
+            "[36] Aysim Token, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Senaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite"
+        ],
+        "bbox": [
+            78,
+            92,
+            470,
+            901
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "dataset for semantic change segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21158-21167, 2022. 2",
+            "[37] Xinyi Tong, Guisong Xia, and Xiao Xiang Zhu. Enabling country-scale land cover mapping with meter-resolution satellite imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 196:178-196, 2023. 1, 2, 5, 6, 7",
+            "[38] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2517-2526, 2019. 2, 6, 7",
+            "[39] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021. 2, 5",
+            "[40] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo-labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4248-4257, 2022. 2",
+            "[41] Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10857-10866, 2021. 2, 3",
+            "[42] Linshan Wu, Leyuan Fang, Xingxin He, Min He, Jiayi Ma, and Zhun Zhong. Querying labeled for unlabeled: Cross-image semantic consistency guided semi-supervised semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2",
+            "[43] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14421-14430, 2022. 2",
+            "[44] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4268-4277, 2022. 2",
+            "[45] D Zanaga, R Van De Kerchove, W De Keersmaecker, N Souverijns, C Brockmann, R Quast, J Wevers, A Grosu, A Paccini, S Vergnaud, et al. Esa worldwide 10 m 2020 v100. zenodo, 2021. 8",
+            "[46] Yuanyi Zhong, Bodi Yuan, Hong Wu, Zhiqiang Yuan, Jian Peng, and Yu-Xiong Wang. Pixel contrastive-consistent semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7273-7282, 2021. 2",
+            "[47] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, 2020. 2"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            897
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "16793",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2023/Large-Scale Land Cover Mapping with Fine-Grained Classes via Class-Aware Semi-Supervised Semantic Segmentation/85af6521-8e24-45e3-ac82-fb3352f9480e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1814308282930a153d77fd16ab314ddb8e5b10d019eda4ccba093ca30f6b6121
+size 5430027

2023/Large-Scale Land Cover Mapping with Fine-Grained Classes via Class-Aware Semi-Supervised Semantic Segmentation/full.md

@@ -0,0 +1,309 @@
+# Large-Scale Land Cover Mapping with Fine-Grained Classes via Class-Aware Semi-Supervised Semantic Segmentation
+
+Runmin Dong $^{1,4}$ , Lichao Mou $^{2}$ , Mengxuan Chen $^{1,4}$ , Weijia Li $^{3}$ , Xin-Yi Tong $^{2}$ , Shuai Yuan $^{1,4}$ , Lixian Zhang $^{1,4}$ , Juepeng Zheng $^{3,4}$ , Xiao Xiang Zhu $^{2}$ , Haohuan Fu $^{1,4,*}$ , Tsinghua University  ${}^{2}$ Technical University of Munich  ${}^{3}$ Sun Yat-Sen University  ${}^{4}$ Tsinghua University - Xi'an Institute of Surveying and Mapping Joint Research Center
+
+drm@mail.tsinghua.edu.cn, haohuan@tsinghua.edu.cn
+
+# Abstract
+
+Semi-supervised learning has attracted increasing attention in the large-scale land cover mapping task. However, existing methods overlook the potential to alleviate the class imbalance problem by selecting a suitable set of unlabeled data. Besides, in class-imbalanced scenarios, existing pseudo-labeling methods mostly only pick confident samples, failing to exploit the hard samples during training. To tackle these issues, we propose a unified Class-Aware Semi-Supervised Semantic Segmentation framework. The proposed framework consists of three key components. To construct a better semi-supervised learning dataset, we propose a class-aware unlabeled data selection method that is more balanced towards the minority classes. Based on the built dataset with improved class balance, we propose a Class-Balanced Cross Entropy loss, jointly considering the annotation bias and the class bias to re-weight the loss in both sample and class levels to alleviate the class imbalance problem. Moreover, we propose the Class Center Contrast method to jointly utilize the labeled and unlabeled data. Specifically, we decompose the feature embedding space using the ground truth and pseudo-labels, and employ the embedding centers for hard and easy samples of each class per image in the contrast loss to exploit the hard samples during training. Compared with state-of-the-art class-balanced pseudo-labeling methods, the proposed method improves the mean accuracy and mIoU by  $4.28\%$  and  $1.70\%$ , respectively, on the large-scale Sentinel-2 dataset with 24 land cover classes.
+
+# 1. Introduction
+
+Land cover mapping provides pixel-level information for urban management, climate change research, ecosystem protection, and other sustainability-related applica
+
+![](images/7b5448a33ca75d31ea5097cb4e3ad58bf9f15c5950fbe43a104bff7189823f84.jpg)  
+Figure 1. Examples of the labeled data. The top row shows Sentinel-2 images containing 13 bands. The bottom row shows the corresponding labels. We use a fine-grained classification system with 24 land cover classes in this work.
+
+tions [21, 27]. With the rapid progress of computer vision technology, it is desired to automatically gain large-scale and fine-grained land cover information from remote sensing images to support the demand of existing earth system science studies and bring new opportunities for intelligent city study [31]. However, fine-grained annotations for land cover mapping are high-cost and time-consuming, limiting labeled data collection for large-scale applications. Semi-supervised learning (SSL) is a potential solution to this problem, as it utilizes both existing labeled data and a large number of unlabeled data [37].
+
+Large-scale land cover mapping with fine-grained classes usually suffers from class imbalance, as shown in Figure 1 and 3. The labeled data tends to be biased due to the difficulty of data collection and label annotation varying by class. The class imbalance leads to suboptimal per
+
+formance in the minority classes. Most SSL methods for large-scale land cover mapping are typically based on the assumption that labeled and unlabeled data have the same class distribution [41, 37]. However, this assumption does not hold in most scenarios, resulting in the bias of class estimation. As such, it is non-trivial to jointly utilize the labeled and unlabeled data to train the network for semi-supervised semantic segmentation. Although many pseudo-labeling methods aim to assign pseudo-labels more steadily and correctly for different classes on the unlabeled data, e.g., dynamically setting threshold or proportion [17, 44], they still rely on setting hyperparameters and empirical principles, resulting in a lack of flexibility and generalization ability in applications. Besides, pseudo-labeling methods mostly rely on easy samples with high confidence scores, thus failing to exploit the hard samples on the unlabeled data.
+
+To alleviate the above issues, we propose a unified Class-Aware Semi-Supervised Semantic Segmentation framework. We exploit the potential of the unlabeled data and alleviate the class imbalance problem in large-scale land cover mapping with fine-grained classes. The proposed class-aware unlabeled data selection method constructs an SSL dataset and compensates for the class imbalance on the labeled data. Based on the built SSL dataset, we propose a class-balanced learning method to remove the annotation bias and class bias on the SSL dataset. We dynamically estimate the class prior information on the entire SSL dataset during training to re-weight both samples and classes of the loss function for the labeled data. Finally, we propose the Class Center Contrast method to jointly use the labeled and unlabeled data for training. The ground truth and pseudolabels are leveraged as the guide to decompose the feature embedding space. We estimate the hard, easy, and overall embedding centers for each class per image and apply the contrastive loss to optimize the distances between class centers to utilize both easy and hard samples. The class imbalance issue is alleviated by utilizing the proposed method, and the performance of large-scale land cover mapping with fine-grained classes is effectively improved.
+
+# 2. Related Work
+
+Large-Scale Land Cover Mapping. Land cover mapping classifies each pixel of images into a unique class, which can be regarded as a semantic segmentation problem [6, 11]. Large-scale land cover mapping with fine-grained classes can provide necessary information for earth system science studies and land spatial layout optimization. However, manual or semi-manual labeling is high-cost and time-consuming [34]. Besides, the classification systems of most large-scale studies are coarse. Thus the provided land cover information is limited. For example, the Chesapeake dataset covering  $160,000\mathrm{km}^2$  costs $1.3 million but only contains four classes [33]. DynamicEarthNet
+
+dataset covers  $17,000\mathrm{km}^2$  and includes 7 basic land cover classes [36]. To overcome the limited area of the labeled data and improve the model generalization, some works utilize openly available produced data (e.g., OpenStreetMap data and land cover products), which contains many noisy labels but can provide prior information in a large-scale area [19, 10]. These works still use coarse classification systems of land cover mapping. Five-Billion-Pixels [37] is a large-scale land cover dataset with 24 categories, covering about  $60,000\mathrm{km}^2$ , which opens the door to large-scale land cover mapping with fine-grained classes. However, this dataset suffers from a severe class-imbalance problem (the common categories cover hundreds of times more pixels than the rare ones), due to the difficulty of data collection and label annotation varying by class.
+
+Class-Imbalanced Learning. To solve the class-imbalance problem, related works include data resampling [3, 4], loss re-weighting [22, 32, 18], margin modification [5, 35], and decoupled learning [20, 47]. In self-supervised learning and SSL, recent works devote to rebalancing the class distribution of SSL data by assigning pseudo-labels for the unlabeled data to address the class imbalance issue [41, 15]. For example, Gui et al. [14] propose a class-aware pseudo-labeling method for SSL, dynamically adjusting the threshold for selecting pseudo-labels to obtain better performance on minority classes. Similarly, Hu et al. [17] propose a label bias removal method and dynamically determine the threshold for each class to assign more pseudo-labels for minority classes. However, the hard samples with relatively low confidence scores are ignored for optimizing the model, leading to sub-optimal performance.
+
+Semi-Supervised Semantic Segmentation. SSL methods leverage the relationships between the labeled and unlabeled data to improve the model accuracy and generalization, alleviating the limited area of the labeled data [8]. Semi-supervised semantic segmentation methods include entropy minimization [38], consistency regularization [30], and pseudo-labeling [42]. Pseudo-labeling is widely used in semi-supervised semantic segmentation, and many works contribute to selecting pseudo-labels more steadily and correctly [29]. For example, Tong et al. [37] propose a dynamic pseudo-label assignment method, in which the number of pseudo-labels is dynamically increased with training iterations. However, these methods still rely on the manual setting of hyperparameters and suffer from the uncertainty of pseudo-labels. Alternatively, contrastive learning methods in supervised learning [39, 16] and SSL [46, 40] constrain the representations of positive samples against these negative samples in feature space, leading to promising results. Different from the manner of data augmentation-based positive sample construction [43, 23] and memory bank-based negative sample construction [1, 40], we leverage the ground-truth and pseudo-labels as a guide to decom
+
+pose the feature embedding space, and estimate the embedding centers for each class per image to stably select easy and hard samples for the proposed Class Center Contrast method.
+
+# 3. Methodology
+
+# 3.1. The Overall Workflow of Our Approach
+
+Our approach consists of four steps. First, a baseline model is trained on the labeled data. Second, we propose a class-aware unlabeled data selection method to build an SSL dataset that is more balanced towards minority classes. The baseline model is used to predict land cover classes on all collected data from a large unlabeled data pool and calculate the statistics of the class distribution for each image patch. The image patches containing pixels of minority classes are selected as unlabeled data. Then the SSL dataset is constructed by combining the labeled data and the selected unlabeled data. Third, we propose a Class-Balanced Cross Entropy loss and a Class Center Contrast method. To alleviate the class imbalance problem, we reweight the cross entropy loss on labeled data in both sample and class level by jointly considering the annotation bias and the class bias on the built SSL dataset. Besides, we decompose the feature embedding space using the ground truth and pseudo-labels, and propose an embedding center selection and contrast method for class-balanced and annotation-guided learning on unlabeled data. The pipeline of the proposed class-aware SSL model is shown in Figure 2. Finally, we perform model inference for large-scale land cover mapping.
+
+# 3.2. Class-Aware Unlabeled Data Selection
+
+A suitable set selection of unlabeled data is important for SSL model development. Inspired by the observation in [41] that existing SSL algorithms can produce pseudolabels on minority classes with high precision, we propose an unlabeled data selection strategy that tends to select unlabeled image patches containing pixels with minority classes.
+
+First, we train a semantic segmentation model using labeled data. We apply the off-the-shelf class balance strategy [4] designed for supervised learning to guarantee the model ability of minority class estimation. The definition of minority classes depends on the class proportion and the class accuracy on the labeled data. Then, we perform model inference on all candidate unlabeled data to predict the land cover class of each pixel, thereby obtaining the class distribution of each candidate image patch. We denote the proportion of pixels of all minority classes in an image patch as  $P_{m}$ . Finally, we formulate a principle to automatically select unlabeled data according to the class distribution of each candidate image patch. Specifically, the pro
+
+portions of the 24 classes on the labeled dataset are shown in Figure 3. The classes with a proportion lower than  $1\%$  generally achieve an accuracy of less than  $30\%$  within our dataset. We collect all the images with  $P_{m} \geq 1\%$ , which leads to about 25,000 images over the about 500,000 candidates (see Section 4.3). Then we randomly select another 25,000 images from those with  $0 < P_{m} < 1\%$ , leading to a total of 50,000 unlabeled images, which is five times the number of the labeled training data. The authors suggest selecting unlabeled patches with relatively more pixels of minority classes (i.e., larger  $P_{m}$ ), and those encompassing rare classes characterized by fewer pixels within the minority classes. Note that although the prediction on the unlabeled data is not absolutely accurate, it can preserve image patches that are most likely to contain pixels with minority classes, which contributes to building a more balanced SSL dataset.
+
+# 3.3. Class-Balanced Learning
+
+In our built SSL dataset, the class distribution of unlabeled data differs from that of the labeled data. It can be regarded as an annotation bias problem on the entire dataset. Besides, the SSL dataset still suffers from a class imbalance issue. Consequently, using the class prior information estimated on the labeled data to the re-weight loss function is inappropriate. To solve the above issues, we propose a class-balanced learning method to jointly remove the annotation bias and class bias on the SSL dataset.
+
+First, we re-weight samples of labeled data to remove the annotation bias. Owing to our unlabeled data selection method, the entire SSL dataset is more balanced towards minority classes than the original labeled dataset. Therefore, we can estimate the class prior on the entire SSL dataset to partially remove the annotation bias. Specifically, we denote  $L$  as a label indicator, where  $L = 0$  and  $L = 1$  represent the unlabeled and labeled data, respectively. The model parameters  $\theta$  of labeled data in traditional SSL methods are obtained by maximizing likelihood estimation in the formula (1).
+
+$$
+\begin{array}{l} \hat {\theta} = \arg \max  _ {\theta} \log P (Y | X, L = 1; \theta) \\ = \arg \max  _ {\theta} \sum_ {(x, y) \in D _ {L = 1}} \log P (y | x; \theta), \tag {1} \\ \end{array}
+$$
+
+where  $\hat{\theta}$  is the estimated model parameter,  $X$  represents samples,  $Y$  represents classes, and  $D_{L = 1}$  denotes the set of labeled data. Due to the annotation bias in the SSL dataset, we have  $P(y|x;L = 1)\neq P(y|x)$ . According to causal inference [13, 17], in the entire SSL dataset,  $X\rightarrow Y\rightarrow L$  can be regarded as a one-way chain, where the class is dependent on its pixel, and whether the pixel has a label or not depends on its class. According to the
+
+![](images/4374f975a37edd04ac6a43821175eb63e023e918e341f7644a90e03609a8add0.jpg)  
+Figure 2. The pipeline of the proposed Class-Aware Semi-Supervised Semantic Segmentation framework with class-balanced learning, embedding center estimation, and class center contrast. The class-aware unlabeled data selection process is not shown in the figure.
+
+conditional independence rule, we obtain the expression,  $P(X|Y,L = 1) = P(X|Y)$  when conditioning on Y. Therefore, estimating the model parameters of the entire SSL dataset can be formalized as in the formula (2) [17], where  $s(x,y) = \frac{\log P(y|x;\theta) - \log P(y;\theta)}{\log P(y|x;\theta)}$  is the loss weight of each labeled sample  $(x,y)$ . Thus, it can be interpreted as modifying the class prior  $P(y;\theta)$  to remove the annotation bias of the labeled data.
+
+$$
+\begin{array}{l} \hat {\theta} = \arg \max  _ {\theta} \log P (X | Y; \theta) \\ = \arg \max  _ {\theta} \sum_ {(x, y) \in D _ {L = 1}} \log P (x | y; \theta) \\ = \arg \max  _ {\theta} \sum_ {(x, y) \in D _ {L = 1}} \log \frac {P (y | x ; \theta) P (x ; \theta)}{P (y ; \theta)} \tag {2} \\ = \arg \max  _ {\theta} \sum_ {(x, y) \in D _ {L = 1}} \log \frac {P (y | x ; \theta)}{P (y ; \theta)} \\ = \arg \max  _ {\theta} \sum_ {(x, y) \in D _ {L = 1}} s (x, y) \cdot \log P (y | x; \theta), \\ \end{array}
+$$
+
+Second, although the SSL dataset is more balanced than the labeled data, it still suffers from the class imbalance problem. Therefore, we further estimate the class bias on the SSL dataset. A widely used class re-weighting loss function on supervised learning can be formalized as follows:
+
+$$
+\mathcal {L} _ {w c e} = \sum_ {k = 1} ^ {K} W _ {k} \sum_ {i = 1} ^ {n _ {k}} y _ {i k} \log p _ {i k}, \tag {3}
+$$
+
+where  $W_{k} = \frac{1}{\log(1 + r_{k})}$  and  $r_k$  is the ratio of the class  $k$  on the labeled data.  $K$  is the number of classes,  $n_k$  is the number of samples of class  $k$ ,  $y_{ik}$  and  $p_{ik}$  is the ground-truth and predicted probability of the  $i$ -th sample for class  $k$ , respectively. In SSL, we approximate  $r_k$  with  $P(y; \theta)$ , the class distribution on the entire dataset.
+
+Finally, we perform an online estimation strategy for  $P(y; \theta)$  to simplify the training process. We compute the average of class probability distributions over all pixels in a mini-batch, which is denoted as  $P(Y; B_t, \theta_t)$ .  $B_t$  and  $\theta_t$  are the samples and model parameters of the current batch, respectively. Then the approximate class distribution, denoted as  $\tilde{P}(Y)$ , is updated by the weighted sum of the obtained class distribution of this mini-batch and the approximate class distribution in the previous iteration, which is formalized as follows:
+
+$$
+\tilde {P} (Y) \leftarrow \mu \tilde {P} (Y) + (1 - \mu) P (Y; B _ {t}, \theta_ {t}), \tag {4}
+$$
+
+where  $\mu \in [0,1]$  is a weighting coefficient. We set  $\mu = \frac{B}{N}$ , in which  $B$  is the batch size and  $N$  is the total number of samples in the dataset. We initialize  $\tilde{P}(Y)$  with the class distribution calculated on the labeled data. As such, the proposed Class-Balanced Cross Entropy loss on the labeled data is formulated as follows:
+
+$$
+\mathcal {L} _ {c b c e} = \sum_ {k = 1} ^ {K} W _ {k} \sum_ {i = 1} ^ {n _ {k}} s \left(x _ {i k}, y _ {i k}\right) y _ {i k} \log p _ {i k}. \tag {5}
+$$
+
+# 3.4. Class Center Contrast
+
+Training via pseudo-labeling is commonly used to utilize unlabeled data in SSL. In fine-grained classes, assigning reliable labels for minority classes is challenging because the segmentation model tends to be biased towards the majority classes [26]. Besides, existing label assignment methods usually use only high-confidence predictions, which may offer limited information to model learning. In this work, our method selects reliable and hard samples. Moreover, instead of directly using the pseudo-labels for supervised learning on unlabeled data, we leverage the pseudo-labels as the guide to decompose the feature embedding space. In general, the distance of feature embeddings from the same class should be minimized, while the distance of those from different classes should be maximized. Therefore, we propose an online embedding center selection method that divides each class into hard and easy sample regions on each image. Specifically, we use the prediction probabilities of ground truth or pseudo-labels to separate easy and hard samples. The predicted probability of each pixel is associated with its feature embedding, and we then calculate the average prediction probability of each class for each image. Feature embeddings with above-average probabilities for the corresponding class are considered easy samples and vice versa. Thus we can obtain three feature embedding centers, including easy, hard, and all feature embedding centers, as contrastive samples. With this method, on the one hand, the obtained feature embeddings for each class are with relatively high confidence, as the samples are close to feature local centers for each class in the embedding space. On the other hand, these feature embeddings include hard samples of each class.
+
+Inspired by the existing contrastive learning method [39], we perform cross-image contrastive learning in each mini-batch. Since both labeled and unlabeled data are trained in a mini-batch, the embeddings of the labeled data could help the feature embedding learning of the unlabeled data. Formally, denote  $P_{m}$  and  $N_{m}$  as the feature embedding collections of the positive and negative samples in a mini-batch for the  $m$ -th feature embedding center  $c_{m}$ . The proposed Class Center Contrastive loss for the  $m$ -th feature embedding center is formulated as follows:
+
+$$
+\mathcal {L} _ {c c t} ^ {m} = - \frac {1}{| P _ {m} |} \sum_ {\boldsymbol {c} _ {m} ^ {+} \in P _ {m}} \log \frac {e ^ {\left(\boldsymbol {c} _ {m} \cdot \boldsymbol {c} _ {m} ^ {+} / \tau\right)}}{e ^ {\left(\boldsymbol {c} _ {m} \cdot \boldsymbol {c} _ {m} ^ {+} / \tau\right)} + \sum_ {\boldsymbol {c} _ {m} ^ {-} \in N _ {m}} e ^ {\left(\boldsymbol {c} _ {m} \cdot \boldsymbol {c} _ {m} ^ {-} / \tau\right)}}, \tag {6}
+$$
+
+in which  $\tau$  is the temperature parameter,  $c_{m}^{+}$  and  $c_{m}^{-}$  are the positive and negative samples, respectively.
+
+The overall loss function can be formulated as follows:
+
+$$
+\mathcal {L} = \mathcal {L} _ {c b c e} + \lambda \mathcal {L} _ {c c t}, \tag {7}
+$$
+
+where  $\lambda$  is the loss weight.
+
+# 4. Data
+
+# 4.1. Image Collection and Preparation
+
+Considering free access, multispectral bands, and relatively high spatial resolution, all satellite images used in this work are acquired from Sentinel-2 satellites. Sentinel-2 imagery has 13 spectral bands at  $10\mathrm{m}$  (R, G, B, and near-infrared bands),  $20\mathrm{m}$  (six red edge and shortwave infrared bands), and  $60\mathrm{m}$  (three atmospheric correction bands) resolutions [12]. The abundant multispectral information and up to  $10\mathrm{m}$  spatial resolution benefit land cover mapping with a fine-grained classification system.
+
+We collect Sentinel-2 images with radiometric correction, geometric corrections, and atmospheric correction, covering more than 60 cities in China. We apply an off-the-shelf Sentinel-2 sharpening method, DSen2Net [24], to reconstruct low-resolution bands and uniform the resolution of all bands of Sentinel-2 images to  $10\mathrm{m}$ .
+
+# 4.2. Label Collection and Classification System
+
+The land cover labels and classification system are from the Five-Billion-Pixels dataset [37], labeled with  $4\mathrm{m}$  resolution. It includes 24 land cover classes. We resize the  $4\mathrm{m}$  labels to  $10\mathrm{m}$  to match Sentinel-2 images. The labeled area covers about  $60,000\mathrm{km}^2$  in China. The classification system includes the industrial area (C1), paddy field (C2), irrigated field (C3), dry cropland (C4), garden land (C5), arbor forest (C6), shrub forest (C7), park (C8), natural meadow (C9), artificial meadow (C10), river (C11), urban residential (C12), lake (C13), pond (C14), fish pond (C15), snow (C16), bare land (C17), rural residential (C18), stadium (C19), square (C20), road (C21), overpass (C22), railway station (C23), and airport (C24).
+
+# 4.3. Dataset
+
+Examples of the labeled data are shown in Figure 1. The training, validation, and test datasets include 9,995, 2,000, and 2,503 images with a size of  $256 \times 256$ , respectively, according to the geographical division. The unlabeled data is from more than 60 dispersed administrative districts in China and has no overlap with the labeled data. The number of candidate unlabeled images is about 500,000, each with a size of  $256 \times 256$ . According to our minority class-biased unlabeled data selection strategy, we filter 50,000 images to build the unlabeled set, which covers more scenes with minority classes. Figure 3 shows the proportion of each land cover class on labeled training data, unlabeled training data, and all training data. The majority classes cover hundreds of times more pixels than the minority classes on the labeled training data.
+
+![](images/66af3d5eaff49cc0c9ef9bd8aca6e0f9f856cf57c02cf9dde286574d88af7d03.jpg)  
+Figure 3. Percentage of each land cover class on the labeled training data, unlabeled training data, and all training data.
+
+<table><tr><td>Method</td><td>OA</td><td>mAcc</td><td>mIoU</td><td>C1</td><td>C2</td><td>C3</td><td>C4</td><td>C5</td><td>C6</td><td>C7</td><td>C8</td><td>C9</td><td>C10</td></tr><tr><td>Baseline</td><td>73.24</td><td>45.77</td><td>35.61</td><td>71.24</td><td>66.25</td><td>88.66</td><td>17.90</td><td>16.93</td><td>62.26</td><td>27.01</td><td>17.26</td><td>44.65</td><td>11.33</td></tr><tr><td>SimCLR [7] +finetune</td><td>74.07</td><td>45.81</td><td>35.17</td><td>74.49</td><td>63.33</td><td>90.88</td><td>14.10</td><td>17.80</td><td>62.10</td><td>23.57</td><td>18.39</td><td>48.42</td><td>21.02</td></tr><tr><td>Pseudo -labeling [25]</td><td>74.61</td><td>48.83</td><td>38.03</td><td>76.85</td><td>69.60</td><td>91.55</td><td>13.47</td><td>20.52</td><td>59.19</td><td>31.63</td><td>37.53</td><td>44.90</td><td>10.68</td></tr><tr><td>Advent [38]</td><td>74.81</td><td>49.25</td><td>38.98</td><td>75.88</td><td>68.38</td><td>91.98</td><td>12.20</td><td>15.01</td><td>62.46</td><td>25.91</td><td>44.77</td><td>54.75</td><td>8.10</td></tr><tr><td>CADR [17]</td><td>75.49</td><td>47.83</td><td>36.63</td><td>76.32</td><td>52.43</td><td>91.48</td><td>21.54</td><td>24.35</td><td>64.74</td><td>31.85</td><td>22.59</td><td>59.88</td><td>17.25</td></tr><tr><td>DPA [37]</td><td>75.36</td><td>48.57</td><td>37.77</td><td>76.05</td><td>69.40</td><td>92.33</td><td>15.51</td><td>19.79</td><td>66.30</td><td>27.05</td><td>23.25</td><td>52.41</td><td>12.48</td></tr><tr><td>Ours</td><td>75.97</td><td>53.53</td><td>40.68</td><td>81.71</td><td>80.07</td><td>88.89</td><td>22.27</td><td>37.32</td><td>67.06</td><td>34.87</td><td>26.44</td><td>56.73</td><td>17.16</td></tr><tr><td>C11</td><td>C12</td><td>C13</td><td>C14</td><td>C15</td><td>C16</td><td>C17</td><td>C18</td><td>C19</td><td>C20</td><td>C21</td><td>C22</td><td>C23</td><td>C24</td></tr><tr><td>70.86</td><td>82.76</td><td>78.93</td><td>24.49</td><td>57.71</td><td>3.04</td><td>48.67</td><td>77.23</td><td>10.80</td><td>7.90</td><td>71.42</td><td>45.83</td><td>34.95</td><td>60.34</td></tr><tr><td>71.47</td><td>83.27</td><td>79.41</td><td>30.54</td><td>64.73</td><td>0.75</td><td>48.77</td><td>68.82</td><td>2.00</td><td>0</td><td>73.59</td><td>44.53</td><td>33.87</td><td>63.72</td></tr><tr><td>72.40</td><td>82.90</td><td>83.56</td><td>16.94</td><td>52.44</td><td>13.66</td><td>55.06</td><td>72.58</td><td>21.05</td><td>13.03</td><td>70.27</td><td>60.60</td><td>38.78</td><td>62.83</td></tr><tr><td>69.98</td><td>83.19</td><td>83.38</td><td>19.72</td><td>49.34</td><td>15.50</td><td>46.56</td><td>75.16</td><td>22.40</td><td>17.55</td><td>70.99</td><td>58.04</td><td>41.62</td><td>69.23</td></tr><tr><td>65.90</td><td>86.27</td><td>84.53</td><td>26.09</td><td>72.57</td><td>0</td><td>51.78</td><td>74.96</td><td>0.95</td><td>0</td><td>65.23</td><td>48.51</td><td>33.64</td><td>75.02</td></tr><tr><td>71.97</td><td>83.33</td><td>81.80</td><td>16.77</td><td>50.66</td><td>0.88</td><td>52.37</td><td>72.61</td><td>21.42</td><td>6.64</td><td>72.15</td><td>59.10</td><td>49.15</td><td>72.37</td></tr><tr><td>70.61</td><td>82.86</td><td>81.69</td><td>17.13</td><td>78.53</td><td>26.28</td><td>55.80</td><td>80.69</td><td>31.83</td><td>26.49</td><td>66.23</td><td>64.65</td><td>46.39</td><td>42.97</td></tr></table>
+
+Table 1. Quantitative comparison with different methods. We also show the precision of each class. Bold indicates the best results  $(\%)$
+
+# 5. Experiments
+
+# 5.1. Implementation Details and Metrics
+
+We choose ResUNet [9] in our experiments, as it is stable and lightweight for land cover mapping. We train for 100 epochs for all experiments using stochastic gradient descent (SGD) with a momentum of 0.9 and a weight decay of  $10^{-4}$ . The initial learning rate is set to 0.05. We adopt the batch size as 12, including 6 labeled images and 6 unlabeled images. The loss weight  $\lambda$  is set to 0.005. The temperature
+
+$\tau$  of contrastive loss is set to 0.07.
+
+The performance of the proposed approach and other competing methods are assessed with overall accuracy (OA), mean accuracy (mAcc), mean intersection over union (mIoU), and precision of each class.
+
+# 5.2. Comparison Results
+
+Table 1 shows the comparison results with typical and state-of-the-art class-balanced pseudo-labeling methods. For a fair comparison, all competing methods and our
+
+<table><tr><td>Method</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>Random selection</td><td>75.50</td><td>51.35</td><td>39.09</td></tr><tr><td>Majority class -biased selection</td><td>75.75</td><td>49.98</td><td>39.33</td></tr><tr><td>Minority class -biased selection</td><td>75.97</td><td>53.53</td><td>40.68</td></tr></table>
+
+Table 2. Ablation study on the proposed class-aware unlabeled data selection method. Bold indicates the best results  $(\%)$ .  
+
+<table><tr><td>Method</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>Without class balance</td><td>75.41</td><td>48.16</td><td>37.27</td></tr><tr><td>Class-balanced method in [38]</td><td>73.61</td><td>51.33</td><td>38.61</td></tr><tr><td>Class-balanced method in [37]</td><td>75.51</td><td>49.66</td><td>35.51</td></tr><tr><td>Ours</td><td>75.97</td><td>53.53</td><td>40.68</td></tr></table>
+
+Table 3. Comparison results on different class-balanced learning methods. Bold indicates the best results  $(\%)$ .
+
+method use the same model architecture. All experiments are measured on the same computing platform. The baseline applies the supervised learning method to the labeled data. The SimCLR+finetune method uses a self-supervised learning method [7] on the unlabeled data to obtain a pretrain model and performs fine-tuning on the labeled data. Pseudo-labeling method [25] first generates pseudo-labels of the unlabeled data by using the baseline model, then retransmits the model with ground truth and pseudo-labels. Advent [38], DPA [37], and CADR [17] are three state-of-the-art SSL methods.
+
+The results demonstrate that our approach significantly improved over  $4.28\%$  in mAcc,  $1.70\%$  in mIoU, and  $0.58\%$  in OA. Compared to the baseline for each class, our method can significantly improve the accuracy of the minority classes and maintain the accuracy of the majority classes simultaneously. In contrast, the competing methods often tend to predict the pixels to majority classes and suffer from low precision on minority classes. Regarding the computational cost, our method incurs  $53\%$  additional training time than the basic SSL algorithm (i.e., Advent) due to Class-Balanced Learning and Class Center Contrast. The inference times are the same for all SSL methods, as we use the same backbone. Please find the visual comparison in the supplementary.
+
+# 5.3. Ablation Study
+
+Class-Aware Unlabeled Data Selection. We compare our method with the opposite strategy (i.e., majority class-biased data selection) and widely opted random selection method. Table 2 shows that our method achieves the best results. On the contrary, the mAcc value of the majority class-biased data selection method is unsatisfactory because the SSL model is biased toward predicting the majority classes.
+
+![](images/10d83bb2ceae84f2e5bf7fb6a2db8ac3888294dc074b804f6a1d10733d745232.jpg)  
+Figure 4. Results of using the different number of labeled images (i.e., 500, 1,000, and 3,000) and the different ratios of unlabeled to labeled images in SSL.
+
+Besides, to verify the effectiveness of unlabeled data selection, we show the statistics of class distribution in Figure 3. Estimating the class distribution of unlabeled data is based on the pseudo-labels produced by the baseline model. Owing to the proposed class-aware unlabeled data selection method, the class distribution of the unlabeled data is more balanced than that of the labeled data. As a result, the entire training dataset is more balanced towards minority classes than the original labeled dataset. Therefore, we can estimate the class prior information on the entire dataset to partially remove the annotation bias.
+
+Class-Balanced Learning. In general, class weighting sacrifices the accuracy of majority classes to improve the accuracy of minority classes, which decreases OA and mIoU. From Table 3, two comparison methods suffer from the above problem, while our method improves the mAcc, OA, and mIoU compared to the baseline (i.e., without loss weights). The reason is that the two comparison methods use class prior information of labeled data and ignore the discrepancy of class distributions between labeled and unlabeled data in SSL. Besides, Table 4 shows the ablation study on Class-Balanced Cross Entropy loss. Utilizing sample re-weighting improves performance by removing the annotation bias on the semi-supervised learning dataset. Using class re-weighting for cross-entropy loss removes class bias and improves the performance of mAcc and mIoU by a large margin. Furthermore, jointly utilizing the sample and class weights achieves the best results in Table 4.
+
+Class Center Contrast Method. Table 5 shows the ablation study on the proposed class center contrast method. Contrastive loss utilizes inter-class and intra-class information and achieves better results than center loss used in [28]. Compared with only training contrastive loss with pseudolabels on unlabeled data, using pseudo-labels on both labeled and unlabeled data achieves better results. Further-
+
+<table><tr><td>Cross entropy loss</td><td>Sample re-weighting</td><td>Class re-weighting</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>✓</td><td></td><td></td><td>75.41</td><td>48.16</td><td>37.27</td></tr><tr><td>✓</td><td>✓</td><td></td><td>75.74</td><td>48.96</td><td>38.40</td></tr><tr><td>✓</td><td></td><td>✓</td><td>75.43</td><td>52.06</td><td>40.07</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>75.97</td><td>53.53</td><td>40.68</td></tr></table>
+
+Table 4. Ablation study on class-balanced cross entropy loss. Bold indicates the best results  $(\%)$ .  
+
+<table><tr><td>Center loss</td><td>Contrastive loss</td><td>Pseudo label</td><td>Ground truth</td><td>On labeled data</td><td>On unlabeled data</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>✓</td><td></td><td>✓</td><td></td><td></td><td>✓</td><td>74.07</td><td>48.13</td><td>37.34</td></tr><tr><td></td><td>✓</td><td>✓</td><td></td><td></td><td>✓</td><td>75.23</td><td>50.06</td><td>39.29</td></tr><tr><td></td><td>✓</td><td>✓</td><td></td><td>✓</td><td>✓</td><td>75.26</td><td>51.42</td><td>39.43</td></tr><tr><td></td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>75.97</td><td>53.53</td><td>40.68</td></tr></table>
+
+Table 5. Ablation study on the class center contrast method. Bold indicates the best results  $(\%)$  
+
+<table><tr><td>Method</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>Random sample selection</td><td>74.68</td><td>49.31</td><td>38.80</td></tr><tr><td>Simple sample selection</td><td>74.74</td><td>51.57</td><td>40.42</td></tr><tr><td>Hard sample selection</td><td>74.57</td><td>51.14</td><td>40.31</td></tr><tr><td>Feature embedding centers</td><td>75.97</td><td>53.53</td><td>40.68</td></tr></table>
+
+Table 6. Results of different contrastive sample selection manners. Bold indicates the best results  $(\%)$ .  
+
+<table><tr><td>Method</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>Within-image method</td><td>73.36</td><td>53.48</td><td>40.03</td></tr><tr><td>Cross-image method</td><td>75.97</td><td>53.53</td><td>40.68</td></tr></table>
+
+more, introducing the ground truth of labeled data to the training of contrastive loss can significantly improve performance. We also compare different sample selection strategies for contrastive learning. We select three contrastive samples for each class per image for a fair comparison. As shown in Table 6, leveraging high-confidence predicted samples from the unlabeled data for semi-supervised learning yields better results than employing uncertain samples. Nonetheless, samples with high-confidence predictions provide limited information for model learning, leading to only marginal enhancements. The proposed Class Center Contrast method involves both high-confidence and hard samples, achieving the best results. Besides, we compare within-image and cross-image contrastive learning results in Table 7. Cross-image method uses labeled and unlabeled images in a mini-batch and achieves better results, while the within-image method only uses local information of labeled or unlabeled images.
+
+Hyperparameter Tuning of Loss Weight.  $\lambda$  is to balance two loss terms in the formula (7). Table 8 shows the
+
+Table 7. Results of within-image and cross-image contrastive learning. Bold indicates the best results  $(\%)$ .  
+
+<table><tr><td>Loss weight</td><td>OA</td><td>mAcc</td><td>mIoU</td></tr><tr><td>λ = 0.05</td><td>75.07</td><td>51.62</td><td>40.15</td></tr><tr><td>λ = 0.01</td><td>75.26</td><td>51.42</td><td>39.43</td></tr><tr><td>λ = 0.005</td><td>75.97</td><td>53.53</td><td>40.68</td></tr><tr><td>λ = 0.001</td><td>75.27</td><td>53.84</td><td>40.26</td></tr></table>
+
+Table 8. Hyperparameter tuning of loss weight. Bold indicates the best results  $(\%)$ .
+
+results using different loss weights  $\lambda$ . We conclude that a large loss weight for  $\mathcal{L}_{cct}$  reduces the model performance. In this work, we set  $\lambda = 0.005$ .
+
+The Number and Ratio of Labeled Images in SSL. We conduct a comparative analysis of the enhancement in accuracy resulting from the increase of unlabeled data. As shown in Figure 4, the utilization of a larger quantity of unlabeled data, specifically ten times the amount of labeled training data containing only 500 images, has been observed to result in a substantial improvement in accuracy. Meanwhile, the performance becomes unstable as the ratio between unlabeled and labeled images increases further because of the limited number of labeled images. As the amount of labeled data increased, using the unlabeled data, which is five times the amount of labeled data, can yield superior results compared to using just one time the amount of labeled data. Subsequently, the performance is relatively stable as the number of unlabeled data increases. Therefore, to balance the computational cost and improvement in accuracy, we use a ratio of 5:1 for unlabeled and labeled data.
+
+# 5.4. Examples of Land Cover Mapping in China
+
+As an application of the proposed method, we produce large-scale land cover maps with 24 classes over 1 million  $\mathrm{km}^2$ . As shown in Figure 5, we compare our land cover mapping results with two existing public products produced by ESA [45] and Google [2]. The example in the top row
+
+![](images/99ab48134b1f93faa14350de580fb1f22554bef78b70c53c9974b5cf369dd558.jpg)  
+Tree cover  
+Shrubland  
+Grassland  
+Cropland  
+Built-up  
+Bare/sparse.  
+vegetation  
+Snow and ice  
+Perm  
+water bodies  
+Herbaceous wetland  
+Mangroves  
+Moss and lichen  
+(a) ESA
+
+![](images/f7f357d0a572cca03a1ba1ff2bd5083e6e4a81c05ae6d022c9b05c5dd55adfd2.jpg)  
+Water  
+Trees  
+Grass  
+Flooded  
+vegetation  
+Crops  
+Shrub &  
+Scrub  
+Built  
+Bare ground  
+Snow & ice
+
+![](images/0aa2424f575f40c7123070bad49ed39ca9a28a3c17339dbbfcbb6f86bc0e8fae.jpg)
+
+![](images/e7b8435da0c4cbe62f2e8d8cc0317ea4202a3afe0372c0b324bbe76068c871f6.jpg)  
+(b) Google  
+(c) Ours  
+Figure 5. Examples of Land Cover Mapping in China. The land cover products of ESA and Google are also generated from Sentinel-2 images but use different classification systems. In the black rectangles, our land cover results with fine-grained classes can provide more land cover information. In the red rectangles, our method recognizes the paddy field correctly, while the other two products classify it as cropland and water, respectively, due to the coarse classification system and the misleading spectral information of the paddy field.
+
+shows the results of an urban area located in the east of China, and that in the bottom row gives the results of a rural area located in the center of China. The land cover products of ESA and Google contain 11 and 9 classes, while our classification system includes 24 classes. From the black rectangles, the two comparison products mainly distinguish built areas, water, trees, and crops. Owing to the fine-grained classification system and the proposed class-aware SSL methods, our results can recognize more land cover categories, including industrial area, urban residential, rural residential, road, overpass, railway station, arbor forest, artificial meadow, irrigated field, bare land, lake, river, pond, and fish pond. Our land cover maps with fine-grained classes can provide more land cover information. Besides, from the red rectangles, our method recognizes the paddy field correctly, while the other two products classify it as cropland and water, respectively, due to the coarse classification system and the misleading spectral information of the paddy field. From the above qualitative comparison, our method provides land cover maps with fine-grained classes, significantly improving the land cover information.
+
+# 6. Conclusion
+
+In this work, we propose a unified Class-Aware Semi-Supervised Semantic Segmentation approach for large-
+
+scale land cover mapping with fine-grained classes. To address the class-imbalance issue, we propose a class-aware unlabeled data selection method to build an SSL dataset more balanced towards minority classes. Then, we propose a Class-Balanced Cross Entropy loss, considering the annotation and class biases on the built SSL dataset. Moreover, we propose the Class Center Contrast method to jointly use the labeled and unlabeled data for training. Experimental results validate the effectiveness of our method in large-scale land cover mapping with fine-grained classes.
+
+# Acknowledgements
+
+This research was supported in part by the National Natural Science Foundation of China (Grant No. T2125006), Jiangsu Innovation Capacity Building Program (Project No. BM2022028), China Postdoctoral Science Foundation (Grant No. 2023M731871), the German Federal Ministry of Education and Research (BMBF) in the framework of the international future AI lab "AI4EO - Artificial Intelligence for Earth Observation: Reasoning, Uncertainties, Ethics and Beyond" (Grant No. 01DD20001), and Shuimu Tsinghua Scholar Project.
+
+# References
+
+[1] Inigo Alonso, Alberto Sabater, David Ferstl, Luis Montesano, and Ana C Murillo. Semi-supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8219-8228, 2021. 2  
+[2] Christopher F Brown, Steven P Brumby, Brooke Guzder-Williams, Tanya Birch, Samantha Brooks Hyde, Joseph Mazzariello, Wanda Czerwinski, Valerie J Pasquarella, Robert Haertel, Simon Ilyushchenko, et al. Dynamic world, near real-time global  $10\mathrm{m}$  land use land cover mapping. Scientific Data, 9(1):251, 2022. 8  
+[3] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural networks, 106:249-259, 2018. 2  
+[4] Jonathon Byrd and Zachary Lipton. What is the effect of importance weighting in deep learning? In International Conference on Machine Learning, pages 872-881, 2019. 2, 3  
+[5] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 2  
+[6] Bingyu Chen, Min Xia, and Junqing Huang. Mfanet: A multi-level feature aggregation network for semantic segmentation of land cover. Remote Sensing, 13(4):731, 2021. 2  
+[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 6, 7  
+[8] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 2  
+[9] Foivos I Diakogiannis, François Waldner, Peter Caccetta, and Chen Wu. Resunet-a: A deep learning framework for semantic segmentation of remotely sensed data. ISPRS Journal of Photogrammetry and Remote Sensing, 162:94-114, 2020. 6  
+[10] Runmin Dong, Weizhen Fang, Haohuan Fu, Lin Gan, Jie Wang, and Peng Gong. High-resolution land cover mapping through learning with noise correction. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2021. 2  
+[11] Runmin Dong, Cong Li, Haohuan Fu, Jie Wang, Weijia Li, Yi Yao, Lin Gan, Le Yu, and Peng Gong. Improving 3-m resolution land cover mapping through efficient learning from an imperfect 10-m resolution map. Remote Sensing, 12(9):1418, 2020. 2  
+[12] Runmin Dong, Lixian Zhang, Weijia Li, Shuai Yuan, Lin Gan, Juepeng Zheng, Haohuan Fu, Lichao Mou, and Xiao Xiang Zhu. An adaptive image fusion method for sentinel-2 images and high-resolution images with long-time
+
+intervals. International Journal of Applied Earth Observation and Geoinformation, 121:103381, 2023. 5  
+[13] Madelyn Glymour, Judea Pearl, and Nicholas P Jewell. Causal inference in statistics: A primer. John Wiley & Sons, 2016. 3  
+[14] Qian Gui, Xinting Wu, and Baoning Niu. Class-aware pseudo labeling for non-random missing labels in semi-supervised learning. In 2022 IEEE Eighth International Conference on Multimedia Big Data, pages 138-143. IEEE, 2022. 2  
+[15] Lanzhe Guo and Yufeng Li. Class-imbalanced semi-supervised learning with adaptive thresholding. In International Conference on Machine Learning, pages 8082-8094, 2022. 2  
+[16] Hanzhe Hu, Jinshi Cui, and Liwei Wang. Region-aware contrastive learning for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16291-16301, 2021. 2  
+[17] Xinting Hu, Yulei Niu, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. On non-random missing labels in semi-supervised learning. In International Conference on Learning Representations, 2022. 2, 3, 4, 6, 7  
+[18] Muhammad Abdullah Jamal, Matthew Brown, Ming-Hsuan Yang, Liqiang Wang, and Boqing Gong. Rethinking class-balanced methods for long-tailed visual recognition from a domain adaptation perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7610-7619, 2020. 2  
+[19] Noah Johnson, Wayne Treible, and Daniel Crispell. Opensentinelmap: A large-scale land use dataset using openstreetmap and sentinel-2 imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1333–1341, 2022. 2  
+[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. 2019. 2  
+[21] Krishna Karra, Caitlin Kontgis, Zoe Statman-Weil, Joseph C Mazzariello, Mark Mathis, and Steven P Brumby. Global land use/land cover with sentinel 2 and deep learning. In 2021 IEEE international Geoscience and Remote Sensing Symposium IGARSS, pages 4704–4707, 2021. 1  
+[22] Salman H Khan, Munawar Hayat, Mohammed Bennamoun, Ferdous A Sohel, and Roberto Togneri. Cost-sensitive learning of deep feature representations from imbalanced data. IEEE Transactions on Neural Networks and Learning Systems, 29(8):3573-3587, 2017. 2  
+[23] Donghyeon Kwon and Suha Kwak. Semi-supervised semantic segmentation with error localization network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9957-9967, 2022. 2  
+[24] Charis Lanaras, José Bioucas-Dias, Silvano Galliani, Emmanuel Baltsavias, and Konrad Schindler. Super-resolution of sentinel-2 images: Learning a globally applicable deep neural network. ISPRS Journal of Photogrammetry and Remote Sensing, 146:305–319, 2018. 5  
+[25] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks.
+
+In Workshop on challenges in representation learning, International Conference on Machine Learning, volume 3, page 896, 2013. 6, 7  
+[26] Ruihuang Li, Shuai Li, Chenhang He, Yabin Zhang, Xu Jia, and Lei Zhang. Class-balanced pixel-level self-labeling for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11593–11603, 2022. 5  
+[27] Weijia Li, Runmin Dong, Haohuan Fu, Jie Wang, Le Yu, and Peng Gong. Integrating google earth imagery with landsat data to improve 30-m resolution land cover mapping. Remote Sensing of Environment, 237:111563, 2020. 1  
+[28] Zhuohong Li, Hongyan Zhang, Fangxiao Lu, Ruoyao Xue, Guangyi Yang, and Liangpei Zhang. Breaking the resolution barrier: A low-to-high network for large-scale high-resolution land-cover mapping using low-resolution labels. ISPRS Journal of Photogrammetry and Remote Sensing, 192:244-267, 2022. 7  
+[29] Xiaoqiang Lu, Licheng Jiao, Fang Liu, Shuyuan Yang, Xu Liu, Zhixi Feng, Lingling Li, and Puhua Chen. Simple and efficient: A semisupervised learning framework for remote sensing image semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 2  
+[30] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12674-12684, 2020. 2  
+[31] Prem Chandra Pandey, Nikos Koutsias, George P Petropoulos, Prashant K Srivastava, and Eyal Ben Dor. Land use/land cover in view of earth observation: data sources, input dimensions, and classifiers—a review of the state of the art. Geocarto International, 36(9):957-988, 2021. 1  
+[32] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In International Conference on Machine Learning, pages 4334-4343, 2018. 2  
+[33] Caleb Robinson, Le Hou, Kolya Malkin, Rachel Soobitsky, Jacob Czawlytko, Bistra Dilkina, and Nebojsa Jojic. Large scale high-resolution land cover mapping with multi-resolution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12726-12735, 2019. 2  
+[34] Vladan Stojnic and Vladimir Risojevic. Self-supervised learning of remote sensing scene representations using contrastive multiview coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1182-1191, 2021. 2  
+[35] Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11662-11671, 2020. 2  
+[36] Aysim Token, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Senaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite
+
+dataset for semantic change segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21158-21167, 2022. 2  
+[37] Xinyi Tong, Guisong Xia, and Xiao Xiang Zhu. Enabling country-scale land cover mapping with meter-resolution satellite imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 196:178-196, 2023. 1, 2, 5, 6, 7  
+[38] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2517-2526, 2019. 2, 6, 7  
+[39] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021. 2, 5  
+[40] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo-labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4248-4257, 2022. 2  
+[41] Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10857-10866, 2021. 2, 3  
+[42] Linshan Wu, Leyuan Fang, Xingxin He, Min He, Jiayi Ma, and Zhun Zhong. Querying labeled for unlabeled: Cross-image semantic consistency guided semi-supervised semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2  
+[43] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14421-14430, 2022. 2  
+[44] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4268-4277, 2022. 2  
+[45] D Zanaga, R Van De Kerchove, W De Keersmaecker, N Souverijns, C Brockmann, R Quast, J Wevers, A Grosu, A Paccini, S Vergnaud, et al. Esa worldwide 10 m 2020 v100. zenodo, 2021. 8  
+[46] Yuanyi Zhong, Bodi Yuan, Hong Wu, Zhiqiang Yuan, Jian Peng, and Yu-Xiong Wang. Pixel contrastive-consistent semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7273-7282, 2021. 2  
+[47] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, 2020. 2

2023/Large-Scale Land Cover Mapping with Fine-Grained Classes via Class-Aware Semi-Supervised Semantic Segmentation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48b414d8377cba02f13a5c16ac3ae686edec284aeba26489ae435f34781a7c2f
+size 810018

2023/Large-Scale Person Detection and Localization Using Overhead Fisheye Cameras/b691cb26-f3c1-4d93-ba53-9242b6258c4e_content_list.json

@@ -0,0 +1,1839 @@
+[
+    {
+        "type": "text",
+        "text": "Large-Scale Person Detection and Localization using Overhead Fisheye Cameras",
+        "text_level": 1,
+        "bbox": [
+            76,
+            130,
+            893,
+            154
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Lu Yang $^{1*}$ , Lulei Li $^{2*}$ , Xueshi Xin $^{1}$ , Yifan Sun $^{3}$ , Qing Song $^{1}$ , Wenguan Wang $^{4\\dagger}$",
+        "bbox": [
+            166,
+            179,
+            807,
+            200
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup> Beijing University of Posts and Telecommunications <sup>2</sup> ReLER, AAI, University of Technology Sydney",
+        "bbox": [
+            169,
+            200,
+            797,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>3</sup>Baidu <sup>4</sup>ReLU, CCAI, Zhejiang University",
+        "bbox": [
+            346,
+            217,
+            622,
+            233
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "https://LOAFisheye.github.io/",
+        "bbox": [
+            354,
+            237,
+            612,
+            250
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            286,
+            313,
+            301
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Location determination finds wide applications in daily life. Instead of existing efforts devoted to localizing tourist photos captured by perspective cameras, in this article, we focus on devising person positioning solutions using overhead fisheye cameras. Such solutions are advantageous in large field of view (FOV), low cost, anti-occlusion, and unaggressive work mode (without the necessity of cameras carried by persons). However, related studies are quite scarce, due to the paucity of data. To stimulate research in this exciting area, we present LOAF, the first large-scale overhead fisheye dataset for person detection and localization. LOAF is built with many essential features, e.g., i) the data cover abundant diversities in scenes, human pose, density, and location; ii) it contains currently the largest number of annotated pedestrian, i.e., 457K bounding boxes with ground-truth location information; iii) the body-boxes are labeled as radius-aligned so as to fully address the positioning challenge. To approach localization, we build a fisheye person detection network, which exploits the fisheye distortions by a rotation-equivariant training strategy and predict radius-aligned human boxes end-to-end. Then, the actual locations of the detected persons are calculated by a numerical solution on the fisheye model and camera altitude data. Extensive experiments on LOAF validate the superiority of our fisheye detector w.r.t. previous methods, and show that our whole fisheye positioning solution is able to locate all persons in FOV with an accuracy of  $0.5\\mathrm{m}$ , within  $0.1\\mathrm{s}$ .",
+        "bbox": [
+            75,
+            318,
+            473,
+            728
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            744,
+            209,
+            760
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Accurate position finding of persons attracts growing interest from both research and industrial communities, since it plays a crucial role in numerous location-sensitive application scenarios (e.g., surveillance, smart home, public health). Nevertheless, due to the line-of-sight (LOS) issue, GPS is unreliable in interior spaces and urban canyon. To overcome",
+        "bbox": [
+            75,
+            771,
+            468,
+            863
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/d9da8daa5ab5af46460ee1950a50f23e4df131a028e1412294e5dd9b1422a0c5.jpg",
+        "image_caption": [
+            "Figure 1: Person positioning using overhead fisheye camera: We detect humans on omnidirectional images and then project detects onto the real world coordinates to obtain physical locations. Compared with using perspective cameras, our fisheye camera based solution is favored in low cost, high accuracy, and fast speed."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            277,
+            885,
+            518
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "such limitation, various alternative solutions are investigated. Signal based solutions, including Bluetooth [13] and WiFi [74], are popular, but they are easily interfered by changing environments and nearby human bodies [73]. A complementary stream of work is vision based; they typically make use of traditional cameras, RGBD cameras, or in-built smartphone cameras, and enjoy the advantage of reliable services. To get location information, visual positioning solutions usually refer to a pre-acquired 3D map or a geotagged database as the scene representation [63], or directly utilize the captured image to estimate the camera pose [39].",
+        "bbox": [
+            496,
+            609,
+            893,
+            777
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Although visual localization has been a hotspot issue for many years, existing efforts are mainly dedicated to urban place recognition or indoor camera localization, based on perspective cameras [63, 39, 62, 43]. None of them addresses person positioning by using overhead fisheye cameras, even though fisheye cameras are widely used in visual surveillance applications. One possible reason is the lack of accessible datasets, compounded by considerable costs in-",
+        "bbox": [
+            496,
+            779,
+            895,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "The first two authors contribute equally to this work.",
+        "bbox": [
+            93,
+            875,
+            380,
+            887
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ Corresponding author: Wenguan Wang.",
+        "bbox": [
+            96,
+            887,
+            310,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "19961",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "volved in data collection. In this article, we provide a large-scale overhead fisheye dataset, LOAF, for person detection and localization in both indoor and outdoor scenes.",
+        "bbox": [
+            76,
+            90,
+            468,
+            135
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Compared with perspective cameras, overhead (top-view) fisheye cameras are promoted due to less occlusion among people and larger field of view (FOV) - allowing the coverage of a large space using a single, low-cost camera. Only few public datasets [21, 42, 24] provide top-view fisheye data. Unfortunately, they only cover quite few scenes and their people-box annotations do not adequately address the positioning challenge, negatively affecting location approximation (see §3.2 for detailed analysis). Differently, LOAF specifically targets at person localization in surveillance applications, and has the following appealing characteristics:",
+        "bbox": [
+            75,
+            137,
+            468,
+            303
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Large-scale: LOAF is the largest in the filed, to our best knowledge. It consists of over 70 videos, with more than 43K frame images, 457K person-detection annotations as well as corresponding location information.",
+            "- High diversity: LOAF contains a wide variety of surveillance scenarios; it includes a total of 11 indoor and 40 outdoor scenes, and the data are captured at different times of day and cover different illumination conditions.",
+            "- Positioning-aware person-box annotation: LOAF offers radius-aligned human-box annotations. Compared with other person representations, i.e., head center [21], axis-aligned [60] or body-aligned box [42, 24], used in previous fisheye datasets, the radius-aligned box is promoted, as it fits well radially-oriented bodies [42], and, more importantly, it is better aware of the positioning problem."
+        ],
+        "bbox": [
+            76,
+            306,
+            468,
+            532
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Moreover, we devise a person positioning system that first detects persons from raw fisheye images, and then calculates physical locations based on fisheye visual model and altitude information (see Fig. 1). Clearly, high-quality person detection is the crucial premise for precise localization using overhead fisheye cameras. Fisheye lenses provide large FOV, at the cost of strong radial distortion. This makes pedestrian detection in top-view fisheye images a much harder task, compared with using perspective cameras. Studies on overhead fisheye pedestrian detection are very scarce [60, 42, 51, 24, 64]; they rarely concern the link with person positioning, and many of them [60, 42] are even not trainable due to the lack of fisheye data. With our LOAF dataset, we develop a novel query based fisheye human detector. It explicitly exploits fisheye geometry by accommodating rotation equivariance into the matching between queries and human instances during network training. The insight here is intuitive: for a robust fisheye detector, rotation of a fisheye image should result in correspondingly rotated detections. In addition, our detection algorithm learns to predict radius-aligned person boxes, facilitating localization estimation.",
+        "bbox": [
+            75,
+            537,
+            468,
+            854
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We test our fisheye person detection algorithm as well as our whole positioning system over LOAF. We find that our detector significantly outperforms previous methods, and",
+        "bbox": [
+            76,
+            854,
+            468,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "our full system delivers precise localization results. We also empirically show that our algorithm generalizes well on previous top-view fisheye person detection datasets [24, 65].",
+        "bbox": [
+            498,
+            90,
+            890,
+            137
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            148,
+            640,
+            165
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Accurate Positioning. GPS is the most popular system for outdoor localization. As it requires LOS between the satellites and the handset, GPS does not function well in urban canyons, indoors and basements [30, 26]. This triggered the development of alternative positioning solutions, following a multi-disciplinary approach. Concretely, there are two main schemes of the alternatives: signal-based and vision-based.",
+        "bbox": [
+            496,
+            170,
+            890,
+            276
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Signal-based positioning systems typically lean on sound wave [50], geomagnetism [32], radio frequency (RF) [26, 74, 54, 13, 1], and infrared radiation (IR) [71], as well as different location determination techniques, such as TOA (time of arrival) [30] and RSS (received signal strength) [26]. The main challenge to signal-based systems is the sensitivity to environment changes, such as object moving, diffraction and reflection, which affect signal propagation [26].",
+        "bbox": [
+            496,
+            277,
+            890,
+            398
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Visual data is another potential information source for precise localization. Since put forward [61], visual positioning has become a hot topic in robotics and computer vision. Some methods adopt a pre-built geo-tagged database or a 3D scene model, as the reference for camera location estimation [33, 77, 66, 78, 76]. Some others rely on recognizing some deployed coded targets [27, 59, 20, 31, 36], e.g., concentric rings, barcodes, colored dots, etc. Some recent ones utilize deep learning techniques to replace some components (e.g., image retrieval, descriptor matching) in traditional systems [4, 16], or regress the camera pose directly [39, 38, 7]. Though promising, existing visual positioning systems are mostly founded on perspective cameras.",
+        "bbox": [
+            496,
+            400,
+            892,
+            595
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Due to the problem complexity, there is no persuasive solution for precise positioning yet, and hybrid schemes are often applied in practice. Fisheye cameras have advantage of providing wide FOV with low cost and reduced occlusion, while the research landscape is sparse for fisheye camera based localization [79, 21]. To foster research in this direction and facilitate practical deployment, we contribute the first overhead fisheye dataset, to our knowledge, that allows to conduct the task of person localization at large scale.",
+        "bbox": [
+            496,
+            597,
+            892,
+            734
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Deep Learning based Visual (Camera) Localization. Regarding application scenarios, scholars in computer vision community are mainly aware of city-scale location recognition [2, 72, 58, 57] and indoor camera localization [5, 67]. Popular visual localization approaches can be divided as retrieval-based, regression-based, and structure-based, according to the camera pose estimation strategy. Retrieval-based methods [58, 3, 23, 75, 56] represent a scene as a database of geo-tagged photos and use geo-tag of the most relevant database photo as an approximation to the camera position [58]. Regression-based methods learn to encode",
+        "bbox": [
+            496,
+            734,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "19962",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "the scene into a deep network and directly regress a 6DOF camera pose [39, 67, 38] from a captured image. Structure-based methods [6, 55, 29] pose the localization problem as a camera resection task [58]. They first represent scenes via 3D models and establish a set of 2D-3D matches, then recover the full camera pose by employing a PnP solver [34] inside a RANSAC [28] loop.",
+        "bbox": [
+            75,
+            90,
+            470,
+            196
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "These visual positioning techniques seek to localize images captured by handheld devices [43, 11] or vehicle cameras [48] and demand pre-created geo-tagged databases or 3D maps. And they work on an active mode – users need to carry the cameras. Differently, we address person localization in surveillance scenarios, by using stationary, overhead fisheye cameras. This yields a passive, low-cost scheme, where the localization is completed by a numerical solution on the fisheye model fused with altitude information. Thus our scheme faces a different challenge, i.e., conducting robust person detection from non-rectilinear fisheye images. Overall, our scheme fills the gap left by conventional studies and is complementary to existing positioning systems.",
+        "bbox": [
+            75,
+            199,
+            468,
+            396
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Person Detection in Overhead Fisheye Images. As a canonical subproblem of object detection, pedestrian detection has long received great interest owing to its broad applications such as intelligent surveillance and autonomous vehicles [12, 8]. By contrast, people detection in overhead fisheye images has been studied much less, due to the absence of such datasets. Moreover, adapting standard pedestrian detectors to top-view fisheye cameras is difficult [15]: First, the appearance of people in fisheye images is arbitrary-oriented, while in perspective images, people typically appear upright. Second, people suffer from severe geometric distortions, particularly in the fisheye image's periphery.",
+        "bbox": [
+            75,
+            397,
+            468,
+            580
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Faced with the above challenges, early attempts make use of handcrafted features (e.g., HOG, LBP) as well as standard pedestrian detectors with slight modifications to account for fisheye geometry [40, 14, 17, 69]. The common approach is dewarping fisheye images [14] or features [40] so as to approximate normal people' appearances from the deformed ones. Though this simplifies the classification of features, approximation errors are inevitable, causing performance degradation [64]. Later, a few CNN-based fisheye detectors were developed. Some of them are training-free. For example, [60] runs standard YOLOv2 [52] on dewarped versions of overlapped windows extracted from a fisheye image. Li et al. [42] rotate a fisheye image in  $15^{\\circ}$  increments, and apply off-the-shelf YOLOv3 [53] only to the top-center region of each rotated image, where people usually appear upright. Some more recent methods are trainable: [64] trains YOLOv2 with rotated perspective images so as to handle omnidirectional images without test-time transformation; [24] trains YOLOv3 with human-aligned boxes.",
+        "bbox": [
+            75,
+            580,
+            470,
+            868
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Although many prior arts [60, 42, 64] allowing to detect persons without any fisheye training data, they require a cer-",
+        "bbox": [
+            76,
+            869,
+            468,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "tain amount of computation time for post-processing: [60] carries out detection in multiple perspective images dewarped from one omnidirectional image; [42] applies YOLOv3 24 times to each fisheye image; [64] needs a grouping process to eliminate numerous redundant results, caused by a rotation-invariant training strategy. Hence, the utility of previous methods is limited in the context of visual positioning. Our analysis in §3.2 sheds light on the weaknesses of human representations adopted by existing fisheye detectors (e.g., head center [21], horizontal or body-aligned box [60, 42, 24]) in regard to localization. Thus we supply our large-scale dataset with positioning-aware person-detection annotation. As a result, our human detector can benefit from end-to-end, fisheye visual pattern learning and output radius-aligned body-boxes for precise positioning. Further, by regularizing the learning of image representations and instance queries with rotation equivariance, our algorithm naturally addresses the omnidirectional nature of fisheye images, yet using standard detection network architecture designed for perspective images.",
+        "bbox": [
+            496,
+            90,
+            893,
+            378
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. LOAF Dataset",
+        "text_level": 1,
+        "bbox": [
+            500,
+            387,
+            648,
+            402
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Dataset Acquisition",
+        "text_level": 1,
+        "bbox": [
+            500,
+            409,
+            686,
+            425
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We first describe how our fisheye images are collected. Apparatus and Technical Specifications. A TL-IPC59AE fisheye camera with  $1.1\\mathrm{mm}$  focal length is adopted for data recording. It has a wide FOV, reaching the full circle in the horizontal plane and  $180^{\\circ}$  in the vertical plane. This offers a clear advantage in reducing deployment cost – installing just one fisheye camera instead of multiple conventional cameras to monitor the same region. However, the severe geometric distortions introduced prevent the use of standard detectors, which are designed for conventional cameras [8]. Data Capturing. The fisheye camera is mounted on the ceiling (indoors) or poles (outdoors),  $2.5\\sim 4.0\\mathrm{m}$  from the ground with  $200\\sim 300\\mathrm{m}^2$  FOV. Considering the factors such as scenario diversity, pedestrians density and weather condition, we capture 110 fisheye image sequences in 80 realistic scenarios as the raw data pool. The recorded sequences span 14 hrs; the image resolution is  $2952\\times 2952$  pixels, and the frame rate is  $10\\sim 20$  fps. Eventually, 42,942 images, sampled at 1 fps, are collected to construct our LOAF dataset.",
+        "bbox": [
+            496,
+            430,
+            890,
+            718
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Dataset Annotation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            728,
+            684,
+            742
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We next describe how our fisheye images are annotated. Person Detection. As the performance of localization relies critically on the quality of person detections, significant effort should be spent on the fisheye person-detection annotation. Some human-detection representations were explored in previous datasets: [21] labels the center point of each human head; [42, 24] opt for human-aligned person-boxes. However, these representations have several shortcomings, especially with regards to the localization task. First, they suffer from some inherent limitations. The point-based rep",
+        "bbox": [
+            496,
+            750,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "19963",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/0a223e2c69ab10176ea19643a29696846404e75a7d4986b7ede427712ae801ed.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"2\">#Scene</td><td rowspan=\"2\">#Video</td><td rowspan=\"2\">#Image</td><td colspan=\"3\">#People</td><td rowspan=\"2\">Max Resolution</td><td rowspan=\"2\">FOV (m2)</td><td colspan=\"3\">Annotation</td><td rowspan=\"2\">FPS</td></tr><tr><td>Indoor</td><td>Outdoor</td><td>Total</td><td>Avg.</td><td>Max</td><td>People Detection</td><td>Location</td><td>Attribute</td></tr><tr><td>PIROPO [21]</td><td>2</td><td>0</td><td>27</td><td>-</td><td>-</td><td>-</td><td>-</td><td>800×600</td><td>&lt;50</td><td>Head Center</td><td></td><td></td><td>10</td></tr><tr><td>HABBOF [42]</td><td>2</td><td>0</td><td>4</td><td>5,837</td><td>20,466</td><td>3.5</td><td>5</td><td>2,048×2,048</td><td>&lt;36</td><td>Human-aligned box</td><td></td><td></td><td>12~30</td></tr><tr><td>MW-R [24]</td><td>6</td><td>0</td><td>19</td><td>8,752</td><td>22,825</td><td>2.6</td><td>6</td><td>1,488×1,488</td><td>&lt;36</td><td>Human-aligned box</td><td></td><td></td><td>15</td></tr><tr><td>CEPDOF [24]</td><td>5</td><td>0</td><td>8</td><td>25,504</td><td>173,073</td><td>6.8</td><td>13</td><td>2,048×2,048</td><td>&lt;36</td><td>Human-aligned box</td><td></td><td></td><td>1~10</td></tr><tr><td>WEPDTOF [65]</td><td>14</td><td>0</td><td>16</td><td>10,544</td><td>93,363</td><td>8.9</td><td>35</td><td>2,592×1,944</td><td>&lt;36</td><td>Human-aligned box</td><td></td><td></td><td>1~10</td></tr><tr><td>LOAF</td><td>11</td><td>39</td><td>74</td><td>42,942</td><td>457,762</td><td>10.5</td><td>65</td><td>2,952×2,952</td><td>200~300</td><td>Radius-aligned box</td><td>✓</td><td>✓</td><td>10~20</td></tr></table>",
+        "bbox": [
+            91,
+            82,
+            887,
+            185
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Overhead fisheye datasets comparison (§3.4). LOAF is the largest in terms of the total number of pedestrian and scene categories.",
+        "bbox": [
+            76,
+            186,
+            893,
+            202
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/e8ecf0dabec41382c1f4818a5fcc12928e67a8a026178700fedcc287896a15a8.jpg",
+        "image_caption": [
+            "Figure 2: Human annotation comparison for visual positioning (§3.2). Previous methods typically use the detected human head [21] ( $\\bullet$ ) or the center ( $\\bullet$ ) of human-aligned body-box ( $\\langle \\rangle$ ) [42, 24] to determine the physical location  $P$ , bringing positioning errors (i.e.,  $PP_{1}$ ,  $PP_{2}$ ). We instead leverage the radius-aligned box ( $\\square$ ) as human representation. The midpoint ( $\\bullet$ ) of side  $AB$ , i.e., the closest point to the image center on the radius-aligned box, better corresponds to the human actual position."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            210,
+            426,
+            380
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "resentation is not applicable in other analysis tasks (e.g., re-identification, attribute recognition); the human-aligned box has ambiguity [10, 68] – the ground-truths of human-aligned boxes are not uniquely determined [24]. Second, and most important, these human-detection representations cannot meet the need of precise localization. Clearly, based on the fisheye camera model, one can project 2D detections on the 3D world for localization. However, the center of human head or of human-aligned box is not the exact placement occupied by human on the image plane, causing errors to estimations of human physical position (see Fig. 2).",
+        "bbox": [
+            75,
+            507,
+            468,
+            672
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We instead label each person through a radius-aligned rectangular box. Such representation is favored as it: i) allows unique groundtruth box assignment; ii) fits well radially-oriented human bodies presented in fisheye images; and iii) better corresponds to the actual position of human on the image plane, facilitating physical localization. Notably, although the radially-oriented box constraint was explored in a prior fisheye detector [42], there is no previous datasets provide such kind of annotation, neither is there any literature points out the advantage of such representation in localization. Finally, around 457K human box annotations are obtained. High-quality annotation is ensured via a rigorous quality check, conducted by highly skilled reviewers.",
+        "bbox": [
+            75,
+            674,
+            470,
+            869
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Person Localization. For each scene, a  $10\\mathrm{m}$  ruler (with  $0.05\\mathrm{m}$  accuracy) is placed on the ground, and one end of the",
+        "bbox": [
+            76,
+            869,
+            470,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/e3f4fafde1e206055214c00a1cb3f0019fe43878e6a4b6156eb82f5574ed418b.jpg",
+        "image_caption": [
+            "(a)",
+            "Figure 3: Calibration for groundtruth person location annotation. (a) Ground makers (with  $0.05\\mathrm{m}$  precision). (b) Top-view fisheye image capture for calibration. See §3.2 for details."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            210,
+            666,
+            358
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/9af3d095353cd10aa67ed5028cc2c96b6e9c92936d06344f43b061731640c848.jpg",
+        "image_caption": [
+            "(b)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            681,
+            210,
+            870,
+            358
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ruler is directly under the fisheye lens (cf. Fig. 3 (a)). We take a fisheye picture and use this marked picture as the calibration for all the dataset images recorded in this scene (cf. Fig. 3(b)). Finally, for each annotated human, the physical location, at sub-decimeter precision, is provided (cf. Fig. 4).",
+        "bbox": [
+            498,
+            430,
+            890,
+            505
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Scene Attribute. To enable in-depth analysis, each image is annotated with multiple attributes, including day/night, outdoor/indoor, and sunny/rain/foggy/snow.",
+        "bbox": [
+            498,
+            506,
+            890,
+            551
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Dataset Design",
+        "text_level": 1,
+        "bbox": [
+            500,
+            559,
+            650,
+            574
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We then list several key aspects of our dataset design.",
+        "bbox": [
+            519,
+            580,
+            872,
+            595
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Privacy Protection. To protect personal information, we apply the gaussian filter to blur all the visible facial regions in our dataset, and conduct experiments on the blurred data. Dataset Splits. LOAF contains 29,569 training, 4,600 validation, and 8,773 testing images (approximately 7 train, 1 val, and 2 test). Moreover, to better evaluate models' generalization ability, LOAF is split into five sets: train (7/28 indoor/outdoor scenes, 29,569 images), val seen (1/2 indoor/outdoor scenes, 1,700 images), val unseen (2/3 indoor/outdoor scenes, 2,900 images), test seen (2/3 indoor/outdoor scenes, 2,774 images), test unseen (2/8 indoor/outdoor scenes, 5,999 images). There are no overlapping scenes between unseen and train sets.",
+        "bbox": [
+            498,
+            595,
+            890,
+            792
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Dataset Accessibility. Only the desensitized version of our dataset will be released online, under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 License [19].",
+        "bbox": [
+            498,
+            792,
+            890,
+            838
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4. Dataset Features and Analysis",
+        "text_level": 1,
+        "bbox": [
+            500,
+            845,
+            766,
+            861
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Finally, we present statistic analysis of LOAF in comparison with existing overhead fisheye datasets [21, 42, 24].",
+        "bbox": [
+            500,
+            869,
+            890,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "19964",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/3874faf83a1f87cb98734690d995bab419a5c4bb1b11818c892a2d4d0c5ea112.jpg",
+        "image_caption": [
+            "playground | winter | noon",
+            "(a) Our proposed LOAF"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            78,
+            222,
+            185
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/332e247dd4cce7a72e8f56e8da89acb758c6b3780779aaf43250a19015b116b1.jpg",
+        "image_caption": [
+            "hall | fall | afternoon"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            222,
+            85,
+            352,
+            185
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/0dd1005fd968575b6b564258bbd47967bdea814f2b54863d087f91051fbd04f7.jpg",
+        "image_caption": [
+            "street | spring | morning"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            352,
+            84,
+            480,
+            185
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/bb593de8269c91711ec122fba21d20b3c712e93d2109b6d365dc78eb87390715.jpg",
+        "image_caption": [
+            "(b) PIROPO [21]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            486,
+            78,
+            596,
+            185
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/4c39d8656378ebc5b2543cd9cef2012dc6d391329b153e7985a695ed7d3edc11.jpg",
+        "image_caption": [
+            "(c) HABBOF [42]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            604,
+            78,
+            741,
+            186
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/32395c7927e4611b23622ac2de31d9ebe2025fcce7026c2ffac12a12b84ee047.jpg",
+        "image_caption": [
+            "(d) CEPDOF [24]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            751,
+            78,
+            890,
+            186
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/f0a4041d8b3e2c825a8308c3dd8591271ef49dd0be458a169b8a4a5b487e57d6.jpg",
+        "image_caption": [
+            "Figure 4: Example images from different datasets ( $\\S 3.2-\\S 3.4$ ). Prior datasets [21, 42, 24] are restricted to few indoor scenes with person detection annotation only, i.e., head center  $(\\bullet)$  or human-aligned box  $(\\square)$ . In contrast, LOAF covers challenging indoor and outdoor scenes with human detection, localization and scene attribute annotations. The person-detection annotation is given as the radius-aligned box  $(- \\square)$ , which is more suitable for localization. For better visualization, we only present location ground-truths for some of persons.",
+            "(a) Person scale"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            277,
+            254,
+            426
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/72e18a3c6c6a1f9edf9be646ab56ad26f35e65015842e43e9b4f596851fb8106.jpg",
+        "image_caption": [
+            "(b) Person locations"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            267,
+            277,
+            452,
+            349
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/327c7873cd49ad6ddcd326dc87d87ad1e4a0e473f74ebe418369c540a32c6724.jpg",
+        "image_caption": [
+            "(c) Person density",
+            "Figure 5: Dataset Statistics (§3.4): We summarize LOAF with the distribution of (a) person scale (area in pixel), (b) person locations (horizontal distance between human and the fisheye lens in the real world), and (c) person density (number of person per image)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            267,
+            359,
+            452,
+            430
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "LOAF distinguishes itself from three aspects (cf. Table 1):",
+        "bbox": [
+            75,
+            506,
+            460,
+            521
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Large-scale: LOAF has 42,942 fisheye images with more than 457K person boxes. Moreover, LOAF data are captured by an advanced fisheye camera, which is capable of covering a larger area  $(200\\sim 300\\mathrm{m}^2)$  with higher pedestrian density (2~65 persons per scene, 10.5 in average). This makes LOAF the largest overhead fisheye dataset in terms of the total number of pedestrian and scene categories.",
+        "bbox": [
+            75,
+            521,
+            468,
+            627
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "High Diversity: Existing datasets limit in data diversities, i.e., only containing very few indoor scenes (2~14) and completely missing outdoor scenarios (cf. Fig. 4). In contrast, LOAF involves 51 realistic scenes, including 11 indoor scenes (e.g., lab, office, library, classroom) and 40 outdoor scenes (e.g., street, playground, parking lot, square). The recorded data cover four seasons under different illumination (e.g., morning, noon, afternoon) and weather (e.g., sunny, rain, snow) conditions, and involve vast variance of human pose (e.g., walking, standing, and sitting), scale, location, and density (cf. Fig. 5). Thus our dataset better reflects the distribution in real-world surveillance scenarios.",
+        "bbox": [
+            75,
+            628,
+            468,
+            808
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Rich and Positioning-aware Annotation: LOAF is provided with rich ground-truths for detection, localization, and scene attribute, which lays a solid foundation for fisheye camera based human-centric analysis. Hence, as demonstrated in §3.2, the radius-aligned human-box representation is adopted during our annotation. Compared with human-",
+        "bbox": [
+            75,
+            810,
+            468,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "head center based point annotation [21] and human-aligned person-boxes [42, 24] used in previous datasets, radius-aligned human-boxes are more suitable for the position task.",
+        "bbox": [
+            496,
+            275,
+            890,
+            321
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Our Approach",
+        "text_level": 1,
+        "bbox": [
+            498,
+            329,
+            648,
+            345
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our overhead fisheye camera based person localization solution comprises two parts. The former computes 2D detections that locate the people on the image plane (§4.1). The latter converts the 2D detections to 3D-world coordinates, obtaining the physical location of the people (§4.2).",
+        "bbox": [
+            496,
+            351,
+            890,
+            428
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Overhead Fisheye Person Detection",
+        "text_level": 1,
+        "bbox": [
+            498,
+            431,
+            807,
+            448
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Core Idea: Rotation Equivariance. One of the reasons for the tremendous success of CNNs is their equivariance to horizontal and vertical shifts and the resulting invariance to local deformations [41, 25]. This stimulates a line of efforts to learn robust representations equivariant to generic types of transformations [35, 18, 70]. Formally, a representation  $f$  is said to be equivariant with a geometric transformation  $g$  (e.g., cropping, flipping) for an input (image)  $I$  if:",
+        "bbox": [
+            496,
+            452,
+            892,
+            573
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf (g (I)) \\approx g (f (I)). \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            632,
+            578,
+            890,
+            594
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "That is to say, the output representation  $f(I)$  is changed in the same way as the transformation  $g$  imposed to the input  $I$ .",
+        "bbox": [
+            496,
+            598,
+            890,
+            628
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this work, we devise a query-based fisheye person detector that exhibits  $360^{\\circ}$ -rotational equivariance through an elaborately designed training protocol. Our key insight is derived from the omnidirectional nature of this task: if one rotates the input fisheye image by an arbitrary angle, then the outputs of a robust fisheye detector should change accordingly. We thus design a rotation equivariant training strategy, which forces the matching between object queries and image representations to be equivariant against  $360^{\\circ}$ -rotations. In this way, the rotational symmetry of omnidirectional images is explicitly addressed, without architectural modification of standard detectors developed for perspective images. Rotation Equivariant Training for Query-based Fisheye Detection. As shown in Fig. 6, our fisheye detector is built upon DETR [9], promoted by a rotation equivariant training strategy. Basically, DETR conducts detection in a query-based fashion. Denote  $\\mathcal{F}$  as a feature encoder that extracts representation  $I$  of image  $I$ , and  $\\kappa$  as a query creator",
+        "bbox": [
+            496,
+            628,
+            892,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "19965",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/84ddbf217cfe1d9f63999d77f1d13c59b4542d4725a8091df5c551f2f2bd612a.jpg",
+        "image_caption": [
+            "Figure 6: Our rotation equivariant training strategy for query-based fisheye person detection (§4.1). For equivariant training, object queries  $\\{\\pmb{q}_n^{g^r}\\}_{n=1}^N$  created for rotated image  $g^r(I)$  and the rotated representation  $g^r(I)$  are fed into the decoder  $\\mathcal{D}$ ; the rotated ground-truths  $\\{\\pmb{b}_{\\ell(n)}^{g^r}\\}_{n=1}^N$  are set as the training targets."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            80,
+            472,
+            244
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "that outputs a set of  $N$  object-aware descriptors  $\\{\\pmb{q}_n\\}_{n=1}^N$  from  $\\pmb{I}$ . DETR employs  $\\{\\pmb{q}_n\\}_{n=1}^N$  as queries to retrieve target objects from  $\\pmb{I}$ :",
+        "bbox": [
+            75,
+            338,
+            468,
+            385
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left\\{\\hat {\\boldsymbol {b}} _ {n} \\right\\} _ {n = 1} ^ {N} = \\mathcal {D} (\\boldsymbol {I}, \\left\\{\\boldsymbol {q} _ {n} \\right\\} _ {n = 1} ^ {N}), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            184,
+            388,
+            468,
+            407
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\pmb{I} = \\mathcal{F}(I)$ ,  $\\{\\pmb{q}_n\\}_{n=1}^N = \\kappa(\\pmb{I})$ , and  $\\mathcal{D}$  refers to the Transformer decoder which formulates the object-query matching process via neural cross-attention computation.  $\\{\\hat{\\pmb{b}}_n\\}_{n=1}^N$  are the set of predicted parameterized object bounding boxes.",
+        "bbox": [
+            76,
+            416,
+            468,
+            478
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Our rotation equivariant training further encourages robust object-query matching which is equivariant against rotations. First, given an input rotation transformation  $g^r$ , a robust fisheye detector is desired to be able to extract rotation-equivariant representation. Analogous to Eq. 1, we have:",
+        "bbox": [
+            75,
+            479,
+            468,
+            554
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {F} \\left(g ^ {r} (I)\\right) \\approx g (\\mathcal {F} (I)) = g ^ {r} (\\boldsymbol {I}). \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            174,
+            559,
+            468,
+            574
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Also, it is reasonable to assume that the final output of a robust fisheye detector should be changed in the same way to the rotation  $g^{r}$  applied to the input fisheye image  $I$ :",
+        "bbox": [
+            75,
+            579,
+            468,
+            625
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left\\{\\boldsymbol {b} _ {\\ell (n)} ^ {g ^ {r}} \\right\\} _ {n = 1} ^ {N} \\approx \\mathcal {D} \\left(\\boldsymbol {I} ^ {g ^ {r}}, \\left\\{\\boldsymbol {q} _ {n} ^ {g ^ {r}} \\right\\} _ {n = 1} ^ {N}\\right), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            631,
+            468,
+            651
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $I^{g^r}$  indicates the feature of rotated image  $g^{r}(I)$ , i.e.,  $\\pmb{I}^{g^{r}} = \\mathcal{F}(g^{r}(I))$ . Similarly,  $\\{\\pmb{q}_n^{g^r}\\}_{n=1}^N$  denote the object queries derived from  $g^{r}(I)$ , i.e.,  $\\{\\pmb{q}_n^{g^r}\\}_{n=1}^N = \\kappa(I^{g^r})$ .  $\\{\\pmb{b}_{\\ell(n)}^{g^{r}}\\}_{n=1}^N$  are the rotated groundtruth bounding boxes, where  $\\ell(n)$  returns the groundtruth index for  $n$ -th query.",
+        "bbox": [
+            76,
+            660,
+            468,
+            739
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Considering Eq.3 and Eq.4, we can have:  $\\{\\pmb{b}_{\\ell(n)}^{g^r}\\}_{n=1}^N \\approx \\mathcal{D}(g^r(\\pmb{I}), \\{\\pmb{q}_n^{g^r}\\}_{n=1}^N)$ . Hence our rotation equivariant training objective is given as:",
+        "bbox": [
+            76,
+            739,
+            468,
+            792
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {r o t a t - e q u i}} = \\mathcal {L} _ {\\det } \\left(\\left\\{\\boldsymbol {b} _ {\\ell (n)} ^ {g ^ {r}} \\right\\} _ {n = 1} ^ {N}, \\mathcal {D} \\left(g ^ {r} (\\boldsymbol {I}), \\left\\{\\boldsymbol {q} _ {n} ^ {g ^ {r}} \\right\\} _ {n = 1} ^ {N}\\right)\\right). \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            114,
+            797,
+            468,
+            816
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Here  $\\mathcal{L}_{\\mathrm{det}}$  is the standard detection loss in DETR [9]. As such, the rotation equivariance properties for both fisheye representation  $\\mathcal{F}$  (cf. Eq.3) and the query-based detection prediction  $\\mathcal{D}(\\boldsymbol {I},\\{\\pmb {q}_n\\}_{n = 1}^N)$  (cf. Eq.4) are sought in a single training target. This also allows us to effortlessly adapt standard",
+        "bbox": [
+            76,
+            824,
+            468,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d9515f35afae0c94c683b1ac86271a358c47042b91c37bf32e6d1296edf330e5.jpg",
+        "image_caption": [
+            "Figure 7: Our fisheye camera model based person localization (§4.2). Here  $\\pmb{p}$  and  $\\pmb{p}'$  denote the projection point of the distortion point  $P$  by fisheye lens and normal perspective lens, respectively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            87,
+            741,
+            252
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "DETR to our omnidirectional detection setting. Note that our rotation equivariant training differs from rotational data augmentation technique which only views rotated images as individual training samples (see §5.4 for detailed experiments).",
+        "bbox": [
+            498,
+            267,
+            890,
+            327
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. 2D-3D Projection based Person Localization",
+        "text_level": 1,
+        "bbox": [
+            498,
+            337,
+            874,
+            352
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "General Fisheye Camera Model. The perspective projection of a normal pinhole camera can be written as  $r = f \\tan \\theta$ , where  $r$  indicates the projection distance between the principal point and the image point,  $f$  is the focal length, and  $\\theta$  is the angle between the incident ray and the camera's optical axis. However, fisheye lens does not follow this perspective projection model [49], as the FOV equals to  $180^{\\circ}$  (cf. Fig. 7). Fortunately, the image formation of different kinds of fisheye lenses can be approximated by a general polynomial projection model [37], i.e.,",
+        "bbox": [
+            496,
+            359,
+            890,
+            511
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nr (\\theta) = \\sum_ {i = 1} ^ {n} k _ {i} \\theta^ {2 i - 1}, \\quad n = 1, 2, 3, 4, \\dots \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            560,
+            520,
+            890,
+            542
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "High distortions can be handled well when  $n = 5$  [37]. The coefficients  $k_{s}$  can be obtained from camera calibration.",
+        "bbox": [
+            496,
+            549,
+            890,
+            579
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "2D-3D Projection. Given a detected human location point  $\\pmb{p} = (u,v)^{\\top}$  in the fisheye image pixel coordinate system, our target is to calculate its 3D location  $\\pmb{P} = (X,Y,Z)^{\\top}$  in the camera coordinate system, where  $Z$  is the altitude of the fisheye camera and primarily known. To do so, one can first calculate  $r$  as:  $r = \\sqrt{(x^2 + y^2)}$ , where  $x = (u - u_0) / f$  and  $y = (v - v_0) / f$ , and  $(u_0,v_0)^{\\top}$  are the coordinates of the principal point in the fisheye image. Then  $\\theta$  can be obtained by solving Eq. 6 using a numerical means [37]. With the altitude  $Z$  of the camera, the actual location  $(X,Y)^{\\top}$  is given as:",
+        "bbox": [
+            496,
+            580,
+            890,
+            742
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\binom {X} {Y} = Z \\binom {\\tan \\theta \\cos \\phi} {\\tan \\theta \\sin \\phi}, \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            612,
+            739,
+            890,
+            776
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\phi = \\arctan (y,x)$  refers to the polar angle, which is shared by both  $\\pmb{p}$  and  $\\pmb{P}$ .",
+        "bbox": [
+            496,
+            780,
+            888,
+            809
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "It can be seen that precise determination of the person location point  $\\pmb{p}$  on the image plane is vital for estimating the corresponding physical position  $\\pmb{P}$ . However, most previous approaches, restricted to axis-aligned [79, 60] or person-aligned [24] human-box representation, use the center of the detection box to approximate  $\\pmb{p}$ . Few excep",
+        "bbox": [
+            496,
+            810,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "19966",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/d9366d37d95e9c6516309e1b7688079db46eddb141649ec873cb7425d391972c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"8\">val</td><td colspan=\"7\">test</td><td>FPS↑</td><td rowspan=\"2\"></td></tr><tr><td>mAP↑</td><td>AP50↑</td><td>AP75↑</td><td>APn↑</td><td>APm↑</td><td>APf↑</td><td>APseen↑</td><td>APunseen↑</td><td>mAP↑</td><td>AP50↑</td><td>AP75↑</td><td>APn↑</td><td>APm↑</td><td>APf↑</td><td>APseen↑</td><td>APunseen↑</td></tr><tr><td>Seide et al. [60]</td><td>21.8</td><td>59.8</td><td>7.6</td><td>32.8</td><td>28.5</td><td>2.3</td><td>23.2</td><td>19.5</td><td>20.2</td><td>58.2</td><td>7.1</td><td>32.2</td><td>28.3</td><td>3.9</td><td>22.4</td><td>18.9</td><td>10.2</td></tr><tr><td>Li et al. [42]</td><td>28.5</td><td>63.3</td><td>20.1</td><td>46.8</td><td>24.2</td><td>1.3</td><td>33.8</td><td>29.3</td><td>27.2</td><td>65.2</td><td>21.3</td><td>47.2</td><td>24.8</td><td>1.3</td><td>31.8</td><td>27.6</td><td>0.6</td></tr><tr><td>Tamura et al. [64]</td><td>34.8</td><td>72.1</td><td>27.7</td><td>51.7</td><td>38.8</td><td>8.7</td><td>38.8</td><td>32.9</td><td>34.2</td><td>72.8</td><td>28.7</td><td>53.7</td><td>37.3</td><td>6.5</td><td>39.5</td><td>33.2</td><td>10.2</td></tr><tr><td>RAPiD [24]</td><td>40.3</td><td>77.9</td><td>34.8</td><td>55.3</td><td>41.9</td><td>9.2</td><td>44.7</td><td>37.6</td><td>39.2</td><td>77.9</td><td>35.4</td><td>54.8</td><td>40.1</td><td>7.9</td><td>44.2</td><td>37.3</td><td>8.4</td></tr><tr><td>Ours</td><td>47.4</td><td>82.6</td><td>48.4</td><td>64.1</td><td>54.3</td><td>14.1</td><td>50.8</td><td>45.7</td><td>46.2</td><td>81.1</td><td>47.3</td><td>66.1</td><td>53.5</td><td>12.6</td><td>49.3</td><td>44.9</td><td>12.1</td></tr></table>",
+        "bbox": [
+            91,
+            83,
+            883,
+            176
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/cea33104a3275bb7c6dee67b760a6eeed7d43e0b6fa2c50010e32121d3523418.jpg",
+        "table_caption": [
+            "Table 2: Person detection results on val and test sets of LOAF (§5.2)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>mAP↑</td><td>AP50↑</td><td>AP75↑</td><td>Precision↑</td><td>Recall↑</td><td>F-Score↑</td></tr><tr><td>Seide et al. [60]</td><td>20.9</td><td>50.6</td><td>10.2</td><td>80.6</td><td>39.5</td><td>53.0</td></tr><tr><td>Li et al. [42]</td><td>34.2</td><td>75.7</td><td>28.6</td><td>86.3</td><td>65.4</td><td>74.4</td></tr><tr><td>Tamura et al. [64]</td><td>29.3</td><td>61.0</td><td>23.4</td><td>88.8</td><td>51.2</td><td>65.0</td></tr><tr><td>RAPiD [24]</td><td>39.3</td><td>85.4</td><td>26.0</td><td>89.2</td><td>78.7</td><td>83.6</td></tr><tr><td>Ours</td><td>47.2</td><td>88.5</td><td>38.2</td><td>90.3</td><td>87.6</td><td>88.9</td></tr></table>",
+        "bbox": [
+            88,
+            202,
+            475,
+            282
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "tions [21], built on point-based human representation, treat the center of the detected human head as  $\\pmb{p}$ . However, they suffer from a similar issue as the human head center is even often far from the exact position  $\\pmb{p}$  human stand on, w.r.t. the image plane. Instead, with our radius-aligned human-box representation, the closest point to the principal point on the bounding box better corresponds to  $\\pmb{p}$ .",
+        "bbox": [
+            75,
+            310,
+            468,
+            416
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3. Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            76,
+            425,
+            294,
+            441
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Network Architecture. Our fisheye person detector is built upon DAB-DETR [45], a prevalent variant of DETR [9] but converges much faster. Swin-T [46] is utilized as the backbone. For our LOAF with radius-aligned person boxes, the output of our detector is a 4D vector  $\\pmb{b} \\in [0,1]^4$  that parameterizes 2D center coordinates, height, and width. Note that, for traditional fisheye detection datasets like [24, 65] with arbitrary-oriented person box annotations, an extra output dimension is needed for rotation angle regression.",
+        "bbox": [
+            75,
+            448,
+            468,
+            585
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Training Objective. Our fisheye person detector is end-to-end trained by jointly optimizing the vanilla detection loss used in DETR [9, 45] (referred as  $\\mathcal{L}_{\\mathrm{Det}}$ ) and our proposed rotation-equivariant constraint (i.e.,  $\\mathcal{L}_{\\mathrm{rotat - equi}}$  in Eq. 5):",
+        "bbox": [
+            75,
+            585,
+            468,
+            646
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {d e t}} + \\lambda \\mathcal {L} _ {\\text {r o t a t - e q u i}}, \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            202,
+            652,
+            468,
+            667
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "where the coefficient  $\\lambda$  is empirically set to 0.5.",
+        "bbox": [
+            75,
+            674,
+            393,
+            688
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Training. Our fisheye detector is trained with a batch size of 8 for 50 epochs, where the AdamW [47] optimizer is employed with base learning rate 2e-4 and decayed by 0.1 at epoch 40. The remaining hyper-parameters are determined following [64, 24]. Specifically, we initialize backbones with ImageNet [22] pre-trained weights and adopt standard data augmentation techniques, i.e., color jitter, horizontal flip, and random scaling, with a base training size of  $608 \\times 608$ . For CEPDOF [24] and WEPDTOF [65], we use a batch size of 128 and pre-train the detector on COCO [44] for 50 epochs to prevent over-fitting, as in [24, 65]. For the computation of our rotation-equivariant loss  $\\mathcal{L}_{\\mathrm{rotat - equi}}$ , the training images are rotated by a degree randomly sampled from 0 to 360. Our detector is implemented in PyTorch and trained on eight",
+        "bbox": [
+            75,
+            688,
+            470,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/2fa67e3ee0acaec53ca67bbf1ddee0fd7571e712b6346cf3d20f534fee02d67f.jpg",
+        "table_caption": [
+            "Table 3: Person detection results on CEPDOF [24] (\\$5.2)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>mAP↑</td><td>AP50↑</td><td>AP75↑</td><td>Precision↑</td><td>Recall↑</td><td>F-Score↑</td></tr><tr><td>Seide et al. [60]</td><td>16.1</td><td>39.4</td><td>9.0</td><td>70.9</td><td>38.6</td><td>50.0</td></tr><tr><td>Li et al. [42]</td><td>25.2</td><td>69.9</td><td>30.2</td><td>81.4</td><td>64.5</td><td>72.0</td></tr><tr><td>Tamura et al. [64]</td><td>28.8</td><td>59.8</td><td>24.2</td><td>77.0</td><td>52.4</td><td>62.4</td></tr><tr><td>RAPiD [24]</td><td>37.7</td><td>72.0</td><td>26.8</td><td>73.3</td><td>67.8</td><td>70.4</td></tr><tr><td>Ours</td><td>45.8</td><td>85.6</td><td>36.6</td><td>84.9</td><td>74.6</td><td>79.8</td></tr></table>",
+        "bbox": [
+            511,
+            202,
+            898,
+            282
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Person detection results on WEPDTOF [65] (\\$5.2).",
+        "bbox": [
+            511,
+            285,
+            879,
+            299
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "NVIDIA Tesla V100 GPUs with a 32GB memory per-card. Inference. Once trained, our fisheye detector can be directly applied for locating persons on the omnidirectional image plane. After that, the physical locations of the detected persons can be easily obtained through the numerical solution described in §4.2. For fisheye person detection, we follow prior work [60, 42, 64, 24] to use  $1024 \\times 1024$  input resolution without any test-time augmentation or post-processing. Testing is conducted on a single NVIDIA V100 GPU with 16 GB memory.",
+        "bbox": [
+            496,
+            311,
+            890,
+            463
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5. Experiment",
+        "text_level": 1,
+        "bbox": [
+            500,
+            474,
+            624,
+            491
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1. Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            500,
+            497,
+            689,
+            513
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Protocol. On the top of LOAF, we conduct experiments for fisheye based person detection and localization. Following the dataset splitting (cf. §3.3), performance are reported on seen and unseen scenes, respectively, for both val and test sets. This allows us to assess the generalization ability over different surveillance scenarios. For comprehensive study, we further report person detection performance on two existing fisheye datasets, CEPDOF[24] and WEPDTOF[65]. Note that localization cannot be test since [24, 65] only provide person bounding box annotations. Evaluation Metrics. For fisheye person detection, we follow COCO [44] to report the mean average precision (mAP) for IoU  $\\in$  [0.5:0.05:0.95]. We also employ AP $_{50}$  and AP $_{75}$  for further analysis. For fisheye person localization, we measure positional error (PE) in meters, i.e., Euclidean distance of calculated position and ground-truth position, as the conventions in visual localization [43, 58, 73, 79]. For detailed evaluation, on our LOAF, we report performance w.r.t. horizontal distance between human and the fisheye lens, i.e., near (0~10 m), middle (10~20 m), and far (larger than 20 m). Hence we have AP $_{n,m,f}$  and PE $_{n,m,f}$  accordingly.",
+        "bbox": [
+            496,
+            518,
+            890,
+            839
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2. Performance on Person Detection",
+        "text_level": 1,
+        "bbox": [
+            500,
+            845,
+            792,
+            861
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "LOAF. We first report the person detection performance on LOAF. Specifically, four recent deep learning based fisheye",
+        "bbox": [
+            500,
+            869,
+            890,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "19967",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/09880f0396cde8565da3ece75e8f2537fa4beb87f3b4b85c1c5032add3283d97.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"6\">val</td><td colspan=\"6\">test</td></tr><tr><td>mPE↓</td><td>PEn↓</td><td>PEm↓</td><td>PEf↓</td><td>PEseen↓</td><td>PEunseen↓</td><td>mPE↓</td><td>PEn↓</td><td>PEm↓</td><td>PEf↓</td><td>PEseen↓</td><td>PEunseen↓</td></tr><tr><td>Seide et al. [60]</td><td>1.298</td><td>0.561</td><td>1.332</td><td>3.109</td><td>1.206</td><td>1.382</td><td>1.321</td><td>0.706</td><td>1.309</td><td>3.482</td><td>1.306</td><td>1.386</td></tr><tr><td>Li et al. [42]</td><td>0.898</td><td>0.502</td><td>0.871</td><td>2.650</td><td>0.832</td><td>0.962</td><td>0.913</td><td>0.543</td><td>0.884</td><td>2.780</td><td>0.904</td><td>0.998</td></tr><tr><td>Tamura et al. [64]</td><td>0.755</td><td>0.429</td><td>0.736</td><td>1.862</td><td>0.709</td><td>0.821</td><td>0.778</td><td>0.471</td><td>0.826</td><td>2.160</td><td>0.724</td><td>0.836</td></tr><tr><td>RAPiD [24]</td><td>0.674</td><td>0.426</td><td>0.623</td><td>1.403</td><td>0.625</td><td>0.757</td><td>0.682</td><td>0.461</td><td>0.664</td><td>1.445</td><td>0.638</td><td>0.776</td></tr><tr><td>Ours</td><td>0.368</td><td>0.144</td><td>0.363</td><td>0.775</td><td>0.318</td><td>0.402</td><td>0.374</td><td>0.148</td><td>0.375</td><td>0.819</td><td>0.326</td><td>0.398</td></tr></table>",
+        "bbox": [
+            91,
+            88,
+            885,
+            181
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "person detectors [60, 42, 64, 24] are involved for comparison. Among them, [60, 42] are training-free, thus their scores are obtained by directly running the algorithms on the val and test data. As [64, 24] are trainable methods, we first train them on the train set of LOAF following their official setups, and then report the scores on the val and test sets. Quantitative comparison results are presented in Table 2. Some essential observations are as follows:",
+        "bbox": [
+            75,
+            209,
+            468,
+            329
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Our fisheye person detector significantly outperforms existing methods across all the metrics. For example, our detector provides a considerable performance gain in mAP, i.e.,  $7.1\\%$  and  $7.0\\%$  higher than the second best, RAPiD [24], on the val and test sets, respectively.",
+            "- Our detector not only handles nearby persons but also approaches distant targets well. Specifically, for the targets at  $10\\sim 20\\mathrm{m}$  away from the fisheye lens, our performance gain over other methods is more significant than the improvement reported with the persons at  $0\\sim 10\\mathrm{m}$  distance.",
+            "- Our detector also yields small performance gap between the seen and unseen scenes (e.g.,  $49.3 \\rightarrow 44.9$  in terms of mAP), validating our good generalizability.",
+            "- For those persons at very far distance ( $>20\\mathrm{~m}$ ), our algorithm, though still giving higher scores than other competitors, suffers from great performance degradation. This sheds light on the direction of our future efforts."
+        ],
+        "bbox": [
+            76,
+            335,
+            467,
+            592
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "CEPDOF [24]. CEPDOF contains eight videos with 25,504 frames in total. Following the official setup, we train our fisheye person detector on HABBOF [42] and MW-R [24] datasets, and report performance on CEPDOF. As shown in Table 3, our fisheye detector delivers state-of-the-art performance: it significantly outperforms RAPiD [24], the current top-leading algorithm, by  $7.9\\%$  in terms of mAP.",
+        "bbox": [
+            75,
+            595,
+            467,
+            700
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "WEPDTOF [65]. WEPDTOF has 16 videos with 10,544 frames in total. Following the official setup, we use HAB-BOF [42], MW-R [24], and CEPDOF [24] for training and WEPDTOF for testing. Table 4 summarizes the results. Impressively, our detector greatly suppresses all the other competitors across all the evaluation metrics.",
+        "bbox": [
+            75,
+            700,
+            467,
+            791
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3. Performance on Person Localization",
+        "text_level": 1,
+        "bbox": [
+            76,
+            801,
+            390,
+            816
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Then we study the person localization performance on LOAF. None of previous fisheye person detectors [60, 42, 64, 24] are aware of the task of person localization. We therefore follow the common practice [79, 21] in the field of visual localization: for [60, 42, 64, 24], we project",
+        "bbox": [
+            75,
+            825,
+            467,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/6e87ce99dda350089a3e9385459bdfa3bf8606525e04e844a9190506454a97f6.jpg",
+        "table_caption": [
+            "Table 5: Person localization results on val and test sets of LOAF (§5.3)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">mAP↑</td><td colspan=\"3\">val seen</td><td colspan=\"3\">val unseen</td></tr><tr><td>mAP↑</td><td>AP50↑</td><td>AP75↑</td><td>mAP↑</td><td>AP50↑</td><td>AP75↑</td></tr><tr><td>Baseline [45]</td><td>43.1</td><td>46.4</td><td>81.2</td><td>48.4</td><td>40.9</td><td>76.2</td><td>38.9</td></tr><tr><td>+ Rotation Aug.</td><td>45.6</td><td>49.2</td><td>82.8</td><td>52.8</td><td>43.2</td><td>78.9</td><td>43.9</td></tr><tr><td>+ Rotation Equ.</td><td>47.4</td><td>50.8</td><td>83.8</td><td>54.8</td><td>45.7</td><td>80.5</td><td>47.3</td></tr></table>",
+        "bbox": [
+            511,
+            210,
+            898,
+            282
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Diagnostic results on val set of LOAF (§5.4).",
+        "bbox": [
+            526,
+            284,
+            861,
+            297
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "centers of their detected human boxes on the 3D world, based on the same strategy of ours (cf. §4.2), for physical position estimation. As shown in Table 5, our system produces much small localization errors in comparison with [60, 42, 64, 24], i.e.,  $\\mathbf{0.374m}$  vs  $0.682\\mathrm{m}[24]$  and 0.773 m of [64] on the test set of LOAF. In addition, for all the methods, the localization performance is declined as target distance increases, but our system suffers from the smallest drop. Moreover, the results suggest there is still large room for improvement, thus we hope that our dataset could encourage continuous efforts in this challenging task.",
+        "bbox": [
+            496,
+            310,
+            890,
+            476
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4. Diagnostic Study",
+        "text_level": 1,
+        "bbox": [
+            500,
+            493,
+            666,
+            508
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6 studies the efficacy of our rotation equivariant training strategy (§4.1), on the val set of LOAF. Our baseline model [45] (row #1) gains  $43.1\\%$  mAP. After adopting rotational data augmentation (row #2), the performance boosts by  $2.5\\%$  mAP. By contrast, our rotation equivariant training brings much larger improvements over the baseline, e.g.,  $4.3\\%$  mAP gain. It is remarkable that, training with standard rotational data augmentation technique can be viewed as a specific case of our equivariant training - only learning rotation-equivariant object-querying (cf. Eq. 4).",
+        "bbox": [
+            496,
+            518,
+            890,
+            671
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            691,
+            617,
+            707
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We presented LOAF – the first top-view fisheye dataset that supports large-scale study for person localization in realistic surveillance scenarios. With radius-aligned personbox annotations and precise location ground-truths, LOAF closes a crucial gap in the literature as these cases are not covered by previous datasets and annotation protocols. We further proposed an efficient fisheye person detection model that is equipped with a rotation-equivariant training strategy. The physical locations of detected persons are formulated based on the fisheye model and the altitude of the camera. We empirically verified the effectiveness and promising performance of our algorithm.",
+        "bbox": [
+            496,
+            719,
+            890,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "19968",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements This work was sponsored in part by the CCF-Tencent Open Fund, China Postdoctoral Science Foundation (2022M710467) and National Key Research and Development Program of China (Grant No. 2021YFF0500900).",
+        "bbox": [
+            76,
+            92,
+            468,
+            148
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            159,
+            173,
+            174
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Abdulrahman Alarifi, AbdulMalik Al-Salman, Mansour Alsaleh, Ahmad Alnafessah, Suheer Al-Hadhrami, Mai A Al-Ammar, and Hend S Al-Khalifa. Ultra wideband indoor positioning technologies: Analysis and recent advances. Sensors, 16(5):707, 2016. 2",
+            "[2] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. In CVPR, 2016. 2",
+            "[3] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In ECCV, 2018. 2",
+            "[4] Federico Boniardi, Abhinav Valada, Rohit Mohan, Tim Caselitz, and Wolfram Burgard. Robot localization in floor plans using a room layout edge extraction network. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2019. 2",
+            "[5] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In CVPR, 2017. 2",
+            "[6] Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In ICCV, 2019. 3",
+            "[7] Guillaume Bresson, Li Yu, Cyril Joly, and Fabien Moutarde. Urban localization with street views using a convolutional neural network for end-to-end camera pose regression. In IEEE Intelligent Vehicles Symposium, 2019. 2",
+            "[8] Jiale Cao, Yanwei Pang, Jin Xie, Fahad Shahbaz Khan, and Ling Shao. From handcrafted to deep features for pedestrian detection: a survey. IEEE TPAMI, 2021. 3",
+            "[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5, 6, 7",
+            "[10] Bao Xin Chen and John K Tsotsos. Fast visual object tracking with rotated bounding boxes. arXiv preprint arXiv:1907.03892, 2019. 4",
+            "[11] David M Chen, Georges Baatz, Kevin Koser, Sam S Tsai, Ramakrishna Vedantham, Timo Pylvanainen, Kimmo Roimela, Xin Chen, Jeff Bach, Marc Pollefeys, et al. City-scale landmark identification on mobile devices. In CVPR, 2011. 3",
+            "[12] Long Chen, Shaobo Lin, Xiankai Lu, Dongpu Cao, Hangbin Wu, Chi Guo, Chun Liu, and Fei-Yue Wang. Deep neural network based vehicle and pedestrian detection for autonomous driving: A survey. IEEE TITS, 2021. 3",
+            "[13] Liang Chen, Ling Pei, Heidi Kuusniemi, Yuwei Chen, Tuomo Kröger, and Ruizhi Chen. Bayesian fusion for indoor positioning using bluetooth fingerprints. Wireless Personal Communications, 2013. 1, 2",
+            "[14] An-Ti Chiang and Yao Wang. Human detection in fish-eye images using hog-based detectors over rotated windows. In ICME Workshop, 2014. 3",
+            "[15] Sheng-Ho Chiang, Tsaipei Wang, and Yi-Fu Chen. Effi-"
+        ],
+        "bbox": [
+            78,
+            184,
+            468,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "cient pedestrian detection in top-view fisheye images using compositions of perspective view patches. Image and Vision Computing, 2021. 3",
+            "[16] Shin-Fang Ch'ng, Naoya Sogi, Pulak Purkait, Tat-Jun Chin, and Kazuhiro Fukui. Resolving marker pose ambiguity by robust rotation averaging with clique constraints. In ICRA, 2020. 2",
+            "[17] Ibrahim Cinaroglu and Yalin Bastanlar. A direct approach for human detection with catadioptric omnidirectional cameras. In Signal Processing and Communications Applications Conference, 2014. 3",
+            "[18] Taco Cohen and Max Welling. Group equivariant convolutional networks. In ICLR, 2016. 5",
+            "[19] Creative Commons. Attribution-noncommercial-sharealike 4.0 international. https://creativecommons.org/licenses/by-nc-sa/4.0/.4",
+            "[20] Joseph DeGol, Timothy Bretl, and Derek Hoiem. Chromatag: A colored marker and fast detection algorithm. In ICCV, 2017. 2",
+            "[21] Carlos R del Blanco, Pablo Carballeira, Fernando Jaureguizar, and Narciso Garcia. Robust people indoor localization with omnidirectional cameras using a grid of spatial-aware classifiers. Signal Processing: Image Communication, 2021. 2, 3, 4, 5, 7, 8",
+            "[22] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 7",
+            "[23] Mingyu Ding, Zhe Wang, Jiankai Sun, Jianping Shi, and Ping Luo. Camnet: Coarse-to-fine retrieval for camera relocalization. In ICCV, 2019. 2",
+            "[24] Zhihao Duan, Ozan Tezcan, Hayato Nakamura, Prakash Ishwar, and Janusz Konrad. Rapid: rotation-aware people detection in overhead fisheye images. In CVPR Workshop, 2020. 2, 3, 4, 5, 6, 7, 8",
+            "[25] Carlos Esteves, Christine Allen-Blanchette, Ameesh Makadia, and Kostas Daniilidis. Learning SO(3) equivariant representations with spherical cnns. In ECCV, 2018. 5",
+            "[26] Zahid Farid, Rosdiadee Nordin, and Mahamod Ismail. Recent advances in wireless indoor localization techniques and system. Journal of Computer Networks and Communications, 2013. 2",
+            "[27] Mark Fiala. Artag, a fiducial marker system using digital techniques. In CVPR, 2005. 2",
+            "[28] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3",
+            "[29] Hugo Germain, Vincent Lepetit, and Guillaume Bourmaud. Neural reprojection error: Merging feature learning and camera pose estimation. In CVPR, 2021. 3",
+            "[30] Yanying Gu, Anthony Lo, and Ignas Niemegeers. A survey of indoor positioning systems for wireless personal networks. IEEE Communications surveys & tutorials, 2009. 2",
+            "[31] Mingfei Han, Yali Wang, Xiaojun Chang, and Yu Qiao. Mining inter-video proposal relations for video object detection. In ECCV, 2020. 2",
+            "[32] Janne Haverinen and Anssi Kemppainen. Global indoor self-localization based on the ambient magnetic field. Robotics and Autonomous Systems, 2009. 2",
+            "[33] James Hays and Alexei A Efros. Im2gps: estimating geo"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "19969",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "graphic information from a single image. In CVPR, 2008. 2",
+            "[34] Joel A Hesch and Stergios I Roumeliotis. A direct least-squares (dls) method for pnp. In ICCV, 2011. 3",
+            "[35] Geoffrey E Hinton, Alex Krizhevsky, and Sida D Wang. Transforming auto-encoders. In ICANN, 2011. 5",
+            "[36] Danying Hu, Daniel DeTone, and Tomasz Malisiewicz. Deep charuco: Dark charuco marker pose estimation. In CVPR, 2019. 2",
+            "[37] Juho Kannala and Sami S Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. IEEE TPAMI, 28(8):1335-1340, 2006. 6",
+            "[38] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In CVPR, 2017. 2, 3",
+            "[39] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In ICCV, 2015, 1, 2, 3",
+            "[40] Oded Krams and Nahum Kiryati. People detection in top-view fisheye imaging. In IEEE International Conference on Advanced Video and Signal based Surveillance, 2017. 3",
+            "[41] Karel Lenc and Andrea Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In CVPR, 2015. 5",
+            "[42] Shengye Li, M Ozan Tezcan, Prakash Ishwar, and Janusz Konrad. Supervised people counting using an overhead fisheye camera. In IEEE International Conference on Advanced Video and Signal Based Surveillance, 2019. 2, 3, 4, 5, 7, 8",
+            "[43] Yunpeng Li, Noah Snavely, Dan Huttenlocher, and Pascal Fua. Worldwide pose estimation using 3d point clouds. In ECCV, 2012. 1, 3, 7",
+            "[44] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 7",
+            "[45] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In ICLR, 2022. 7, 8",
+            "[46] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 7",
+            "[47] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7",
+            "[48] Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year,  $1000\\mathrm{km}$ : The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3-15, 2017. 3",
+            "[49] Kenro Miyamoto. Fish eye lens. JOSA, 54(8):1060-1061, 1964. 6",
+            "[50] Jun Qi and Guo-Ping Liu. A robust high-accuracy ultrasound indoor positioning system based on a wireless sensor network. Sensors, 17(11):2554, 2017. 2",
+            "[51] Hazem Rashed, Eslam Mohamed, Ganesh Sistu, Varun Ravi Kumar, Ciaran Eising, Ahmad El-Sallab, and Senthil Yogamani. Generalized object detection on fisheye cameras for autonomous driving: Dataset, representations and baseline. In WACV, 2021. 2",
+            "[52] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster,"
+        ],
+        "bbox": [
+            78,
+            90,
+            467,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "stronger. In CVPR, 2017. 3",
+            "[53] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 3",
+            "[54] Samer S Saab and Zahi S Nakad. A standalone RFID indoor positioning system using passive tags. IEEE Transactions on Industrial Electronics, 2010. 2",
+            "[55] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In CVPR, 2019. 3",
+            "[56] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In CVPR, 2021. 2",
+            "[57] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In CVPR, 2018. 2",
+            "[58] Torsten Sattler, Akihiko Torii, Josef Sivic, Marc Pollefeys, Hajime Taira, Masatoshi Okutomi, and Tomas Pajdla. Are large-scale 3d models really necessary for accurate visual localization? In CVPR, 2017. 2, 3, 7",
+            "[59] Gerald Schweighofer and Axel Pinz. Robust pose estimation from a planar target. IEEE TPAMI, 28(12):2024-2030, 2006. 2",
+            "[60] Roman Seidel, André Apitzsch, and Gangolf Hirtz. Improved person detection on omnidirectional images with non-maxima suppression. In International Conference on Computer Vision Theory and Applications, 2019. 2, 3, 6, 7, 8",
+            "[61] Steve Shafer, John Krumm, Barry Brumitt, Brian Meyers, Mary Czerwinski, and Daniel Robbins. The new easyliving project at microsoft research. In Proceedings of the 1998 DARPA/NIST Smart Spaces Workshop, 1998. 2",
+            "[62] Xun Sun, Yuanfan Xie, Pei Luo, and Liang Wang. A dataset for benchmarking image-based localization. In CVPR, 2017. 1",
+            "[63] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Akihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. In CVPR, 2018. 1",
+            "[64] Masato Tamura, Shota Horiguchi, and Tomokazu Murakami. Omnidirectional pedestrian detection by rotation invariant training. In WACV, 2019. 2, 3, 7, 8",
+            "[65] Ozan Tezcan, Zhihao Duan, Mertcan Cokbas, Prakash Ishwar, and Janusz Konrad. Wepdtof: A dataset and benchmark algorithms for in-the-wild people detection and tracking from overhead fisheye cameras. In WACV, 2022. 2, 4, 7, 8",
+            "[66] Gonzalo Vaca-Castano, Amir Roshan Zamir, and Mubarak Shah. City scale geo-spatial trajectory estimation of a moving camera. In CVPR, 2012. 2",
+            "[67] Florian Walch, Caner Hazirbas, Laura Leal-Taixe, Torsten Sattler, Sebastian Hilsenbeck, and Daniel Cremers. Image-based localization using lstms for structured feature correlation. In ICCV, 2017. 2, 3",
+            "[68] Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Philip HS Torr. Fast online object tracking and segmentation:"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            888
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "19970",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "A unifying approach. In CVPR, 2019. 4",
+            "[69] Tsaipei Wang, Chia-Wei Chang, and Yu-Shan Wu. Template-based people detection using a single downward-viewing fisheye camera. In International Symposium on Intelligent Signal Processing and Communication Systems, 2017. 3",
+            "[70] Wenguan Wang, James Chenhao Liang, and Dongfang Liu. Learning equivariant segmentation with instance-unique querying. In NeurIPS, 2013. 5",
+            "[71] Roy Want, Andy Hopper, Veronica Falcao, and Jonathan Gibbons. The active badge location system. ACM Transactions on Information Systems, 10(1):91-102, 1992. 2",
+            "[72] Tobias Weyand, Ilya Kostrikov, and James Philbin. Planetphoto geolocation with convolutional neural networks. In ECCV, 2016. 2",
+            "[73] Aoran Xiao, Ruizhi Chen, Deren Li, Yujin Chen, and Dewen Wu. An indoor positioning system based on static objects in large indoor scenes by using smartphone cameras. Sensors, 2018. 1, 7",
+            "[74] Chouchang Yang and Huai-Rong Shao. Wifi-based indoor positioning. IEEE Communications Magazine, 2015. 1, 2",
+            "[75] Luwei Yang, Ziqian Bai, Chengzhou Tang, Honghua Li, Yasutaka Furukawa, and Ping Tan. Sanet: Scene agnostic network for camera localization. In ICCV, 2019. 2",
+            "[76] Khalid Yousif, Yuichi Taguchi, and Srikumar Ramalingam. Monorgbd-slam: Simultaneous localization and mapping using both monocular and rgbd cameras. In ICRA, 2017. 2",
+            "[77] Amir Roshan Zamir and Mubarak Shah. Accurate image localization based on google maps street view. In ECCV, 2010. 2",
+            "[78] Amir Roshan Zamir and Mubarak Shah. Image geolocation based on multiplenearest neighbor feature matching using generalized graphs. IEEE TPAMI, 36(8):1546-1558, 2014. 2",
+            "[79] Jun Zhu, Jiangcheng Zhu, Xudong Wan, Chao Wu, and Chao Xu. Object detection and localization in 3d environment by fusing raw fisheye image and attitude data. Journal of Visual Communication and Image Representation, 59:128-139, 2019. 2, 6, 7, 8"
+        ],
+        "bbox": [
+            78,
+            92,
+            468,
+            617
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "19971",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2023/Large-Scale Person Detection and Localization Using Overhead Fisheye Cameras/b691cb26-f3c1-4d93-ba53-9242b6258c4e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79d95cbe36bc6141f304440235a7619be257b2a49c322e9686c28b04753108f6
+size 6491089

2023/Large-Scale Person Detection and Localization Using Overhead Fisheye Cameras/full.md

@@ -0,0 +1,387 @@
+# Large-Scale Person Detection and Localization using Overhead Fisheye Cameras
+
+Lu Yang $^{1*}$ , Lulei Li $^{2*}$ , Xueshi Xin $^{1}$ , Yifan Sun $^{3}$ , Qing Song $^{1}$ , Wenguan Wang $^{4\dagger}$
+
+<sup>1</sup> Beijing University of Posts and Telecommunications <sup>2</sup> ReLER, AAI, University of Technology Sydney
+
+<sup>3</sup>Baidu <sup>4</sup>ReLU, CCAI, Zhejiang University
+
+https://LOAFisheye.github.io/
+
+# Abstract
+
+Location determination finds wide applications in daily life. Instead of existing efforts devoted to localizing tourist photos captured by perspective cameras, in this article, we focus on devising person positioning solutions using overhead fisheye cameras. Such solutions are advantageous in large field of view (FOV), low cost, anti-occlusion, and unaggressive work mode (without the necessity of cameras carried by persons). However, related studies are quite scarce, due to the paucity of data. To stimulate research in this exciting area, we present LOAF, the first large-scale overhead fisheye dataset for person detection and localization. LOAF is built with many essential features, e.g., i) the data cover abundant diversities in scenes, human pose, density, and location; ii) it contains currently the largest number of annotated pedestrian, i.e., 457K bounding boxes with ground-truth location information; iii) the body-boxes are labeled as radius-aligned so as to fully address the positioning challenge. To approach localization, we build a fisheye person detection network, which exploits the fisheye distortions by a rotation-equivariant training strategy and predict radius-aligned human boxes end-to-end. Then, the actual locations of the detected persons are calculated by a numerical solution on the fisheye model and camera altitude data. Extensive experiments on LOAF validate the superiority of our fisheye detector w.r.t. previous methods, and show that our whole fisheye positioning solution is able to locate all persons in FOV with an accuracy of  $0.5\mathrm{m}$ , within  $0.1\mathrm{s}$ .
+
+# 1. Introduction
+
+Accurate position finding of persons attracts growing interest from both research and industrial communities, since it plays a crucial role in numerous location-sensitive application scenarios (e.g., surveillance, smart home, public health). Nevertheless, due to the line-of-sight (LOS) issue, GPS is unreliable in interior spaces and urban canyon. To overcome
+
+![](images/d9da8daa5ab5af46460ee1950a50f23e4df131a028e1412294e5dd9b1422a0c5.jpg)  
+Figure 1: Person positioning using overhead fisheye camera: We detect humans on omnidirectional images and then project detects onto the real world coordinates to obtain physical locations. Compared with using perspective cameras, our fisheye camera based solution is favored in low cost, high accuracy, and fast speed.
+
+such limitation, various alternative solutions are investigated. Signal based solutions, including Bluetooth [13] and WiFi [74], are popular, but they are easily interfered by changing environments and nearby human bodies [73]. A complementary stream of work is vision based; they typically make use of traditional cameras, RGBD cameras, or in-built smartphone cameras, and enjoy the advantage of reliable services. To get location information, visual positioning solutions usually refer to a pre-acquired 3D map or a geotagged database as the scene representation [63], or directly utilize the captured image to estimate the camera pose [39].
+
+Although visual localization has been a hotspot issue for many years, existing efforts are mainly dedicated to urban place recognition or indoor camera localization, based on perspective cameras [63, 39, 62, 43]. None of them addresses person positioning by using overhead fisheye cameras, even though fisheye cameras are widely used in visual surveillance applications. One possible reason is the lack of accessible datasets, compounded by considerable costs in-
+
+volved in data collection. In this article, we provide a large-scale overhead fisheye dataset, LOAF, for person detection and localization in both indoor and outdoor scenes.
+
+Compared with perspective cameras, overhead (top-view) fisheye cameras are promoted due to less occlusion among people and larger field of view (FOV) - allowing the coverage of a large space using a single, low-cost camera. Only few public datasets [21, 42, 24] provide top-view fisheye data. Unfortunately, they only cover quite few scenes and their people-box annotations do not adequately address the positioning challenge, negatively affecting location approximation (see §3.2 for detailed analysis). Differently, LOAF specifically targets at person localization in surveillance applications, and has the following appealing characteristics:
+
+- Large-scale: LOAF is the largest in the filed, to our best knowledge. It consists of over 70 videos, with more than 43K frame images, 457K person-detection annotations as well as corresponding location information.  
+- High diversity: LOAF contains a wide variety of surveillance scenarios; it includes a total of 11 indoor and 40 outdoor scenes, and the data are captured at different times of day and cover different illumination conditions.  
+- Positioning-aware person-box annotation: LOAF offers radius-aligned human-box annotations. Compared with other person representations, i.e., head center [21], axis-aligned [60] or body-aligned box [42, 24], used in previous fisheye datasets, the radius-aligned box is promoted, as it fits well radially-oriented bodies [42], and, more importantly, it is better aware of the positioning problem.
+
+Moreover, we devise a person positioning system that first detects persons from raw fisheye images, and then calculates physical locations based on fisheye visual model and altitude information (see Fig. 1). Clearly, high-quality person detection is the crucial premise for precise localization using overhead fisheye cameras. Fisheye lenses provide large FOV, at the cost of strong radial distortion. This makes pedestrian detection in top-view fisheye images a much harder task, compared with using perspective cameras. Studies on overhead fisheye pedestrian detection are very scarce [60, 42, 51, 24, 64]; they rarely concern the link with person positioning, and many of them [60, 42] are even not trainable due to the lack of fisheye data. With our LOAF dataset, we develop a novel query based fisheye human detector. It explicitly exploits fisheye geometry by accommodating rotation equivariance into the matching between queries and human instances during network training. The insight here is intuitive: for a robust fisheye detector, rotation of a fisheye image should result in correspondingly rotated detections. In addition, our detection algorithm learns to predict radius-aligned person boxes, facilitating localization estimation.
+
+We test our fisheye person detection algorithm as well as our whole positioning system over LOAF. We find that our detector significantly outperforms previous methods, and
+
+our full system delivers precise localization results. We also empirically show that our algorithm generalizes well on previous top-view fisheye person detection datasets [24, 65].
+
+# 2. Related Work
+
+Accurate Positioning. GPS is the most popular system for outdoor localization. As it requires LOS between the satellites and the handset, GPS does not function well in urban canyons, indoors and basements [30, 26]. This triggered the development of alternative positioning solutions, following a multi-disciplinary approach. Concretely, there are two main schemes of the alternatives: signal-based and vision-based.
+
+Signal-based positioning systems typically lean on sound wave [50], geomagnetism [32], radio frequency (RF) [26, 74, 54, 13, 1], and infrared radiation (IR) [71], as well as different location determination techniques, such as TOA (time of arrival) [30] and RSS (received signal strength) [26]. The main challenge to signal-based systems is the sensitivity to environment changes, such as object moving, diffraction and reflection, which affect signal propagation [26].
+
+Visual data is another potential information source for precise localization. Since put forward [61], visual positioning has become a hot topic in robotics and computer vision. Some methods adopt a pre-built geo-tagged database or a 3D scene model, as the reference for camera location estimation [33, 77, 66, 78, 76]. Some others rely on recognizing some deployed coded targets [27, 59, 20, 31, 36], e.g., concentric rings, barcodes, colored dots, etc. Some recent ones utilize deep learning techniques to replace some components (e.g., image retrieval, descriptor matching) in traditional systems [4, 16], or regress the camera pose directly [39, 38, 7]. Though promising, existing visual positioning systems are mostly founded on perspective cameras.
+
+Due to the problem complexity, there is no persuasive solution for precise positioning yet, and hybrid schemes are often applied in practice. Fisheye cameras have advantage of providing wide FOV with low cost and reduced occlusion, while the research landscape is sparse for fisheye camera based localization [79, 21]. To foster research in this direction and facilitate practical deployment, we contribute the first overhead fisheye dataset, to our knowledge, that allows to conduct the task of person localization at large scale.
+
+Deep Learning based Visual (Camera) Localization. Regarding application scenarios, scholars in computer vision community are mainly aware of city-scale location recognition [2, 72, 58, 57] and indoor camera localization [5, 67]. Popular visual localization approaches can be divided as retrieval-based, regression-based, and structure-based, according to the camera pose estimation strategy. Retrieval-based methods [58, 3, 23, 75, 56] represent a scene as a database of geo-tagged photos and use geo-tag of the most relevant database photo as an approximation to the camera position [58]. Regression-based methods learn to encode
+
+the scene into a deep network and directly regress a 6DOF camera pose [39, 67, 38] from a captured image. Structure-based methods [6, 55, 29] pose the localization problem as a camera resection task [58]. They first represent scenes via 3D models and establish a set of 2D-3D matches, then recover the full camera pose by employing a PnP solver [34] inside a RANSAC [28] loop.
+
+These visual positioning techniques seek to localize images captured by handheld devices [43, 11] or vehicle cameras [48] and demand pre-created geo-tagged databases or 3D maps. And they work on an active mode – users need to carry the cameras. Differently, we address person localization in surveillance scenarios, by using stationary, overhead fisheye cameras. This yields a passive, low-cost scheme, where the localization is completed by a numerical solution on the fisheye model fused with altitude information. Thus our scheme faces a different challenge, i.e., conducting robust person detection from non-rectilinear fisheye images. Overall, our scheme fills the gap left by conventional studies and is complementary to existing positioning systems.
+
+Person Detection in Overhead Fisheye Images. As a canonical subproblem of object detection, pedestrian detection has long received great interest owing to its broad applications such as intelligent surveillance and autonomous vehicles [12, 8]. By contrast, people detection in overhead fisheye images has been studied much less, due to the absence of such datasets. Moreover, adapting standard pedestrian detectors to top-view fisheye cameras is difficult [15]: First, the appearance of people in fisheye images is arbitrary-oriented, while in perspective images, people typically appear upright. Second, people suffer from severe geometric distortions, particularly in the fisheye image's periphery.
+
+Faced with the above challenges, early attempts make use of handcrafted features (e.g., HOG, LBP) as well as standard pedestrian detectors with slight modifications to account for fisheye geometry [40, 14, 17, 69]. The common approach is dewarping fisheye images [14] or features [40] so as to approximate normal people' appearances from the deformed ones. Though this simplifies the classification of features, approximation errors are inevitable, causing performance degradation [64]. Later, a few CNN-based fisheye detectors were developed. Some of them are training-free. For example, [60] runs standard YOLOv2 [52] on dewarped versions of overlapped windows extracted from a fisheye image. Li et al. [42] rotate a fisheye image in  $15^{\circ}$  increments, and apply off-the-shelf YOLOv3 [53] only to the top-center region of each rotated image, where people usually appear upright. Some more recent methods are trainable: [64] trains YOLOv2 with rotated perspective images so as to handle omnidirectional images without test-time transformation; [24] trains YOLOv3 with human-aligned boxes.
+
+Although many prior arts [60, 42, 64] allowing to detect persons without any fisheye training data, they require a cer-
+
+tain amount of computation time for post-processing: [60] carries out detection in multiple perspective images dewarped from one omnidirectional image; [42] applies YOLOv3 24 times to each fisheye image; [64] needs a grouping process to eliminate numerous redundant results, caused by a rotation-invariant training strategy. Hence, the utility of previous methods is limited in the context of visual positioning. Our analysis in §3.2 sheds light on the weaknesses of human representations adopted by existing fisheye detectors (e.g., head center [21], horizontal or body-aligned box [60, 42, 24]) in regard to localization. Thus we supply our large-scale dataset with positioning-aware person-detection annotation. As a result, our human detector can benefit from end-to-end, fisheye visual pattern learning and output radius-aligned body-boxes for precise positioning. Further, by regularizing the learning of image representations and instance queries with rotation equivariance, our algorithm naturally addresses the omnidirectional nature of fisheye images, yet using standard detection network architecture designed for perspective images.
+
+# 3. LOAF Dataset
+
+# 3.1. Dataset Acquisition
+
+We first describe how our fisheye images are collected. Apparatus and Technical Specifications. A TL-IPC59AE fisheye camera with  $1.1\mathrm{mm}$  focal length is adopted for data recording. It has a wide FOV, reaching the full circle in the horizontal plane and  $180^{\circ}$  in the vertical plane. This offers a clear advantage in reducing deployment cost – installing just one fisheye camera instead of multiple conventional cameras to monitor the same region. However, the severe geometric distortions introduced prevent the use of standard detectors, which are designed for conventional cameras [8]. Data Capturing. The fisheye camera is mounted on the ceiling (indoors) or poles (outdoors),  $2.5\sim 4.0\mathrm{m}$  from the ground with  $200\sim 300\mathrm{m}^2$  FOV. Considering the factors such as scenario diversity, pedestrians density and weather condition, we capture 110 fisheye image sequences in 80 realistic scenarios as the raw data pool. The recorded sequences span 14 hrs; the image resolution is  $2952\times 2952$  pixels, and the frame rate is  $10\sim 20$  fps. Eventually, 42,942 images, sampled at 1 fps, are collected to construct our LOAF dataset.
+
+# 3.2. Dataset Annotation
+
+We next describe how our fisheye images are annotated. Person Detection. As the performance of localization relies critically on the quality of person detections, significant effort should be spent on the fisheye person-detection annotation. Some human-detection representations were explored in previous datasets: [21] labels the center point of each human head; [42, 24] opt for human-aligned person-boxes. However, these representations have several shortcomings, especially with regards to the localization task. First, they suffer from some inherent limitations. The point-based rep
+
+<table><tr><td rowspan="2">Dataset</td><td colspan="2">#Scene</td><td rowspan="2">#Video</td><td rowspan="2">#Image</td><td colspan="3">#People</td><td rowspan="2">Max Resolution</td><td rowspan="2">FOV (m2)</td><td colspan="3">Annotation</td><td rowspan="2">FPS</td></tr><tr><td>Indoor</td><td>Outdoor</td><td>Total</td><td>Avg.</td><td>Max</td><td>People Detection</td><td>Location</td><td>Attribute</td></tr><tr><td>PIROPO [21]</td><td>2</td><td>0</td><td>27</td><td>-</td><td>-</td><td>-</td><td>-</td><td>800×600</td><td>&lt;50</td><td>Head Center</td><td></td><td></td><td>10</td></tr><tr><td>HABBOF [42]</td><td>2</td><td>0</td><td>4</td><td>5,837</td><td>20,466</td><td>3.5</td><td>5</td><td>2,048×2,048</td><td>&lt;36</td><td>Human-aligned box</td><td></td><td></td><td>12~30</td></tr><tr><td>MW-R [24]</td><td>6</td><td>0</td><td>19</td><td>8,752</td><td>22,825</td><td>2.6</td><td>6</td><td>1,488×1,488</td><td>&lt;36</td><td>Human-aligned box</td><td></td><td></td><td>15</td></tr><tr><td>CEPDOF [24]</td><td>5</td><td>0</td><td>8</td><td>25,504</td><td>173,073</td><td>6.8</td><td>13</td><td>2,048×2,048</td><td>&lt;36</td><td>Human-aligned box</td><td></td><td></td><td>1~10</td></tr><tr><td>WEPDTOF [65]</td><td>14</td><td>0</td><td>16</td><td>10,544</td><td>93,363</td><td>8.9</td><td>35</td><td>2,592×1,944</td><td>&lt;36</td><td>Human-aligned box</td><td></td><td></td><td>1~10</td></tr><tr><td>LOAF</td><td>11</td><td>39</td><td>74</td><td>42,942</td><td>457,762</td><td>10.5</td><td>65</td><td>2,952×2,952</td><td>200~300</td><td>Radius-aligned box</td><td>✓</td><td>✓</td><td>10~20</td></tr></table>
+
+Table 1: Overhead fisheye datasets comparison (§3.4). LOAF is the largest in terms of the total number of pedestrian and scene categories.
+
+![](images/e8ecf0dabec41382c1f4818a5fcc12928e67a8a026178700fedcc287896a15a8.jpg)  
+Figure 2: Human annotation comparison for visual positioning (§3.2). Previous methods typically use the detected human head [21] ( $\bullet$ ) or the center ( $\bullet$ ) of human-aligned body-box ( $\langle \rangle$ ) [42, 24] to determine the physical location  $P$ , bringing positioning errors (i.e.,  $PP_{1}$ ,  $PP_{2}$ ). We instead leverage the radius-aligned box ( $\square$ ) as human representation. The midpoint ( $\bullet$ ) of side  $AB$ , i.e., the closest point to the image center on the radius-aligned box, better corresponds to the human actual position.
+
+resentation is not applicable in other analysis tasks (e.g., re-identification, attribute recognition); the human-aligned box has ambiguity [10, 68] – the ground-truths of human-aligned boxes are not uniquely determined [24]. Second, and most important, these human-detection representations cannot meet the need of precise localization. Clearly, based on the fisheye camera model, one can project 2D detections on the 3D world for localization. However, the center of human head or of human-aligned box is not the exact placement occupied by human on the image plane, causing errors to estimations of human physical position (see Fig. 2).
+
+We instead label each person through a radius-aligned rectangular box. Such representation is favored as it: i) allows unique groundtruth box assignment; ii) fits well radially-oriented human bodies presented in fisheye images; and iii) better corresponds to the actual position of human on the image plane, facilitating physical localization. Notably, although the radially-oriented box constraint was explored in a prior fisheye detector [42], there is no previous datasets provide such kind of annotation, neither is there any literature points out the advantage of such representation in localization. Finally, around 457K human box annotations are obtained. High-quality annotation is ensured via a rigorous quality check, conducted by highly skilled reviewers.
+
+Person Localization. For each scene, a  $10\mathrm{m}$  ruler (with  $0.05\mathrm{m}$  accuracy) is placed on the ground, and one end of the
+
+![](images/e3f4fafde1e206055214c00a1cb3f0019fe43878e6a4b6156eb82f5574ed418b.jpg)  
+(a)  
+Figure 3: Calibration for groundtruth person location annotation. (a) Ground makers (with  $0.05\mathrm{m}$  precision). (b) Top-view fisheye image capture for calibration. See §3.2 for details.
+
+![](images/9af3d095353cd10aa67ed5028cc2c96b6e9c92936d06344f43b061731640c848.jpg)  
+(b)
+
+ruler is directly under the fisheye lens (cf. Fig. 3 (a)). We take a fisheye picture and use this marked picture as the calibration for all the dataset images recorded in this scene (cf. Fig. 3(b)). Finally, for each annotated human, the physical location, at sub-decimeter precision, is provided (cf. Fig. 4).
+
+Scene Attribute. To enable in-depth analysis, each image is annotated with multiple attributes, including day/night, outdoor/indoor, and sunny/rain/foggy/snow.
+
+# 3.3. Dataset Design
+
+We then list several key aspects of our dataset design.
+
+Privacy Protection. To protect personal information, we apply the gaussian filter to blur all the visible facial regions in our dataset, and conduct experiments on the blurred data. Dataset Splits. LOAF contains 29,569 training, 4,600 validation, and 8,773 testing images (approximately 7 train, 1 val, and 2 test). Moreover, to better evaluate models' generalization ability, LOAF is split into five sets: train (7/28 indoor/outdoor scenes, 29,569 images), val seen (1/2 indoor/outdoor scenes, 1,700 images), val unseen (2/3 indoor/outdoor scenes, 2,900 images), test seen (2/3 indoor/outdoor scenes, 2,774 images), test unseen (2/8 indoor/outdoor scenes, 5,999 images). There are no overlapping scenes between unseen and train sets.
+
+Dataset Accessibility. Only the desensitized version of our dataset will be released online, under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 License [19].
+
+# 3.4. Dataset Features and Analysis
+
+Finally, we present statistic analysis of LOAF in comparison with existing overhead fisheye datasets [21, 42, 24].
+
+![](images/3874faf83a1f87cb98734690d995bab419a5c4bb1b11818c892a2d4d0c5ea112.jpg)  
+playground | winter | noon  
+(a) Our proposed LOAF
+
+![](images/332e247dd4cce7a72e8f56e8da89acb758c6b3780779aaf43250a19015b116b1.jpg)  
+hall | fall | afternoon
+
+![](images/0dd1005fd968575b6b564258bbd47967bdea814f2b54863d087f91051fbd04f7.jpg)  
+street | spring | morning
+
+![](images/bb593de8269c91711ec122fba21d20b3c712e93d2109b6d365dc78eb87390715.jpg)  
+(b) PIROPO [21]
+
+![](images/4c39d8656378ebc5b2543cd9cef2012dc6d391329b153e7985a695ed7d3edc11.jpg)  
+(c) HABBOF [42]
+
+![](images/32395c7927e4611b23622ac2de31d9ebe2025fcce7026c2ffac12a12b84ee047.jpg)  
+(d) CEPDOF [24]
+
+![](images/f0a4041d8b3e2c825a8308c3dd8591271ef49dd0be458a169b8a4a5b487e57d6.jpg)  
+Figure 4: Example images from different datasets ( $\S 3.2-\S 3.4$ ). Prior datasets [21, 42, 24] are restricted to few indoor scenes with person detection annotation only, i.e., head center  $(\bullet)$  or human-aligned box  $(\square)$ . In contrast, LOAF covers challenging indoor and outdoor scenes with human detection, localization and scene attribute annotations. The person-detection annotation is given as the radius-aligned box  $(- \square)$ , which is more suitable for localization. For better visualization, we only present location ground-truths for some of persons.  
+(a) Person scale
+
+![](images/72e18a3c6c6a1f9edf9be646ab56ad26f35e65015842e43e9b4f596851fb8106.jpg)  
+(b) Person locations
+
+![](images/327c7873cd49ad6ddcd326dc87d87ad1e4a0e473f74ebe418369c540a32c6724.jpg)  
+(c) Person density  
+Figure 5: Dataset Statistics (§3.4): We summarize LOAF with the distribution of (a) person scale (area in pixel), (b) person locations (horizontal distance between human and the fisheye lens in the real world), and (c) person density (number of person per image).
+
+LOAF distinguishes itself from three aspects (cf. Table 1):
+
+Large-scale: LOAF has 42,942 fisheye images with more than 457K person boxes. Moreover, LOAF data are captured by an advanced fisheye camera, which is capable of covering a larger area  $(200\sim 300\mathrm{m}^2)$  with higher pedestrian density (2~65 persons per scene, 10.5 in average). This makes LOAF the largest overhead fisheye dataset in terms of the total number of pedestrian and scene categories.
+
+High Diversity: Existing datasets limit in data diversities, i.e., only containing very few indoor scenes (2~14) and completely missing outdoor scenarios (cf. Fig. 4). In contrast, LOAF involves 51 realistic scenes, including 11 indoor scenes (e.g., lab, office, library, classroom) and 40 outdoor scenes (e.g., street, playground, parking lot, square). The recorded data cover four seasons under different illumination (e.g., morning, noon, afternoon) and weather (e.g., sunny, rain, snow) conditions, and involve vast variance of human pose (e.g., walking, standing, and sitting), scale, location, and density (cf. Fig. 5). Thus our dataset better reflects the distribution in real-world surveillance scenarios.
+
+Rich and Positioning-aware Annotation: LOAF is provided with rich ground-truths for detection, localization, and scene attribute, which lays a solid foundation for fisheye camera based human-centric analysis. Hence, as demonstrated in §3.2, the radius-aligned human-box representation is adopted during our annotation. Compared with human-
+
+head center based point annotation [21] and human-aligned person-boxes [42, 24] used in previous datasets, radius-aligned human-boxes are more suitable for the position task.
+
+# 4. Our Approach
+
+Our overhead fisheye camera based person localization solution comprises two parts. The former computes 2D detections that locate the people on the image plane (§4.1). The latter converts the 2D detections to 3D-world coordinates, obtaining the physical location of the people (§4.2).
+
+# 4.1. Overhead Fisheye Person Detection
+
+Core Idea: Rotation Equivariance. One of the reasons for the tremendous success of CNNs is their equivariance to horizontal and vertical shifts and the resulting invariance to local deformations [41, 25]. This stimulates a line of efforts to learn robust representations equivariant to generic types of transformations [35, 18, 70]. Formally, a representation  $f$  is said to be equivariant with a geometric transformation  $g$  (e.g., cropping, flipping) for an input (image)  $I$  if:
+
+$$
+f (g (I)) \approx g (f (I)). \tag {1}
+$$
+
+That is to say, the output representation  $f(I)$  is changed in the same way as the transformation  $g$  imposed to the input  $I$ .
+
+In this work, we devise a query-based fisheye person detector that exhibits  $360^{\circ}$ -rotational equivariance through an elaborately designed training protocol. Our key insight is derived from the omnidirectional nature of this task: if one rotates the input fisheye image by an arbitrary angle, then the outputs of a robust fisheye detector should change accordingly. We thus design a rotation equivariant training strategy, which forces the matching between object queries and image representations to be equivariant against  $360^{\circ}$ -rotations. In this way, the rotational symmetry of omnidirectional images is explicitly addressed, without architectural modification of standard detectors developed for perspective images. Rotation Equivariant Training for Query-based Fisheye Detection. As shown in Fig. 6, our fisheye detector is built upon DETR [9], promoted by a rotation equivariant training strategy. Basically, DETR conducts detection in a query-based fashion. Denote  $\mathcal{F}$  as a feature encoder that extracts representation  $I$  of image  $I$ , and  $\kappa$  as a query creator
+
+![](images/84ddbf217cfe1d9f63999d77f1d13c59b4542d4725a8091df5c551f2f2bd612a.jpg)  
+Figure 6: Our rotation equivariant training strategy for query-based fisheye person detection (§4.1). For equivariant training, object queries  $\{\pmb{q}_n^{g^r}\}_{n=1}^N$  created for rotated image  $g^r(I)$  and the rotated representation  $g^r(I)$  are fed into the decoder  $\mathcal{D}$ ; the rotated ground-truths  $\{\pmb{b}_{\ell(n)}^{g^r}\}_{n=1}^N$  are set as the training targets.
+
+that outputs a set of  $N$  object-aware descriptors  $\{\pmb{q}_n\}_{n=1}^N$  from  $\pmb{I}$ . DETR employs  $\{\pmb{q}_n\}_{n=1}^N$  as queries to retrieve target objects from  $\pmb{I}$ :
+
+$$
+\left\{\hat {\boldsymbol {b}} _ {n} \right\} _ {n = 1} ^ {N} = \mathcal {D} (\boldsymbol {I}, \left\{\boldsymbol {q} _ {n} \right\} _ {n = 1} ^ {N}), \tag {2}
+$$
+
+where  $\pmb{I} = \mathcal{F}(I)$ ,  $\{\pmb{q}_n\}_{n=1}^N = \kappa(\pmb{I})$ , and  $\mathcal{D}$  refers to the Transformer decoder which formulates the object-query matching process via neural cross-attention computation.  $\{\hat{\pmb{b}}_n\}_{n=1}^N$  are the set of predicted parameterized object bounding boxes.
+
+Our rotation equivariant training further encourages robust object-query matching which is equivariant against rotations. First, given an input rotation transformation  $g^r$ , a robust fisheye detector is desired to be able to extract rotation-equivariant representation. Analogous to Eq. 1, we have:
+
+$$
+\mathcal {F} \left(g ^ {r} (I)\right) \approx g (\mathcal {F} (I)) = g ^ {r} (\boldsymbol {I}). \tag {3}
+$$
+
+Also, it is reasonable to assume that the final output of a robust fisheye detector should be changed in the same way to the rotation  $g^{r}$  applied to the input fisheye image  $I$ :
+
+$$
+\left\{\boldsymbol {b} _ {\ell (n)} ^ {g ^ {r}} \right\} _ {n = 1} ^ {N} \approx \mathcal {D} \left(\boldsymbol {I} ^ {g ^ {r}}, \left\{\boldsymbol {q} _ {n} ^ {g ^ {r}} \right\} _ {n = 1} ^ {N}\right), \tag {4}
+$$
+
+where  $I^{g^r}$  indicates the feature of rotated image  $g^{r}(I)$ , i.e.,  $\pmb{I}^{g^{r}} = \mathcal{F}(g^{r}(I))$ . Similarly,  $\{\pmb{q}_n^{g^r}\}_{n=1}^N$  denote the object queries derived from  $g^{r}(I)$ , i.e.,  $\{\pmb{q}_n^{g^r}\}_{n=1}^N = \kappa(I^{g^r})$ .  $\{\pmb{b}_{\ell(n)}^{g^{r}}\}_{n=1}^N$  are the rotated groundtruth bounding boxes, where  $\ell(n)$  returns the groundtruth index for  $n$ -th query.
+
+Considering Eq.3 and Eq.4, we can have:  $\{\pmb{b}_{\ell(n)}^{g^r}\}_{n=1}^N \approx \mathcal{D}(g^r(\pmb{I}), \{\pmb{q}_n^{g^r}\}_{n=1}^N)$ . Hence our rotation equivariant training objective is given as:
+
+$$
+\mathcal {L} _ {\text {r o t a t - e q u i}} = \mathcal {L} _ {\det } \left(\left\{\boldsymbol {b} _ {\ell (n)} ^ {g ^ {r}} \right\} _ {n = 1} ^ {N}, \mathcal {D} \left(g ^ {r} (\boldsymbol {I}), \left\{\boldsymbol {q} _ {n} ^ {g ^ {r}} \right\} _ {n = 1} ^ {N}\right)\right). \tag {5}
+$$
+
+Here  $\mathcal{L}_{\mathrm{det}}$  is the standard detection loss in DETR [9]. As such, the rotation equivariance properties for both fisheye representation  $\mathcal{F}$  (cf. Eq.3) and the query-based detection prediction  $\mathcal{D}(\boldsymbol {I},\{\pmb {q}_n\}_{n = 1}^N)$  (cf. Eq.4) are sought in a single training target. This also allows us to effortlessly adapt standard
+
+![](images/d9515f35afae0c94c683b1ac86271a358c47042b91c37bf32e6d1296edf330e5.jpg)  
+Figure 7: Our fisheye camera model based person localization (§4.2). Here  $\pmb{p}$  and  $\pmb{p}'$  denote the projection point of the distortion point  $P$  by fisheye lens and normal perspective lens, respectively.
+
+DETR to our omnidirectional detection setting. Note that our rotation equivariant training differs from rotational data augmentation technique which only views rotated images as individual training samples (see §5.4 for detailed experiments).
+
+# 4.2. 2D-3D Projection based Person Localization
+
+General Fisheye Camera Model. The perspective projection of a normal pinhole camera can be written as  $r = f \tan \theta$ , where  $r$  indicates the projection distance between the principal point and the image point,  $f$  is the focal length, and  $\theta$  is the angle between the incident ray and the camera's optical axis. However, fisheye lens does not follow this perspective projection model [49], as the FOV equals to  $180^{\circ}$  (cf. Fig. 7). Fortunately, the image formation of different kinds of fisheye lenses can be approximated by a general polynomial projection model [37], i.e.,
+
+$$
+r (\theta) = \sum_ {i = 1} ^ {n} k _ {i} \theta^ {2 i - 1}, \quad n = 1, 2, 3, 4, \dots \tag {6}
+$$
+
+High distortions can be handled well when  $n = 5$  [37]. The coefficients  $k_{s}$  can be obtained from camera calibration.
+
+2D-3D Projection. Given a detected human location point  $\pmb{p} = (u,v)^{\top}$  in the fisheye image pixel coordinate system, our target is to calculate its 3D location  $\pmb{P} = (X,Y,Z)^{\top}$  in the camera coordinate system, where  $Z$  is the altitude of the fisheye camera and primarily known. To do so, one can first calculate  $r$  as:  $r = \sqrt{(x^2 + y^2)}$ , where  $x = (u - u_0) / f$  and  $y = (v - v_0) / f$ , and  $(u_0,v_0)^{\top}$  are the coordinates of the principal point in the fisheye image. Then  $\theta$  can be obtained by solving Eq. 6 using a numerical means [37]. With the altitude  $Z$  of the camera, the actual location  $(X,Y)^{\top}$  is given as:
+
+$$
+\binom {X} {Y} = Z \binom {\tan \theta \cos \phi} {\tan \theta \sin \phi}, \tag {7}
+$$
+
+where  $\phi = \arctan (y,x)$  refers to the polar angle, which is shared by both  $\pmb{p}$  and  $\pmb{P}$ .
+
+It can be seen that precise determination of the person location point  $\pmb{p}$  on the image plane is vital for estimating the corresponding physical position  $\pmb{P}$ . However, most previous approaches, restricted to axis-aligned [79, 60] or person-aligned [24] human-box representation, use the center of the detection box to approximate  $\pmb{p}$ . Few excep
+
+<table><tr><td rowspan="2">Method</td><td colspan="8">val</td><td colspan="7">test</td><td>FPS↑</td><td rowspan="2"></td></tr><tr><td>mAP↑</td><td>AP50↑</td><td>AP75↑</td><td>APn↑</td><td>APm↑</td><td>APf↑</td><td>APseen↑</td><td>APunseen↑</td><td>mAP↑</td><td>AP50↑</td><td>AP75↑</td><td>APn↑</td><td>APm↑</td><td>APf↑</td><td>APseen↑</td><td>APunseen↑</td></tr><tr><td>Seide et al. [60]</td><td>21.8</td><td>59.8</td><td>7.6</td><td>32.8</td><td>28.5</td><td>2.3</td><td>23.2</td><td>19.5</td><td>20.2</td><td>58.2</td><td>7.1</td><td>32.2</td><td>28.3</td><td>3.9</td><td>22.4</td><td>18.9</td><td>10.2</td></tr><tr><td>Li et al. [42]</td><td>28.5</td><td>63.3</td><td>20.1</td><td>46.8</td><td>24.2</td><td>1.3</td><td>33.8</td><td>29.3</td><td>27.2</td><td>65.2</td><td>21.3</td><td>47.2</td><td>24.8</td><td>1.3</td><td>31.8</td><td>27.6</td><td>0.6</td></tr><tr><td>Tamura et al. [64]</td><td>34.8</td><td>72.1</td><td>27.7</td><td>51.7</td><td>38.8</td><td>8.7</td><td>38.8</td><td>32.9</td><td>34.2</td><td>72.8</td><td>28.7</td><td>53.7</td><td>37.3</td><td>6.5</td><td>39.5</td><td>33.2</td><td>10.2</td></tr><tr><td>RAPiD [24]</td><td>40.3</td><td>77.9</td><td>34.8</td><td>55.3</td><td>41.9</td><td>9.2</td><td>44.7</td><td>37.6</td><td>39.2</td><td>77.9</td><td>35.4</td><td>54.8</td><td>40.1</td><td>7.9</td><td>44.2</td><td>37.3</td><td>8.4</td></tr><tr><td>Ours</td><td>47.4</td><td>82.6</td><td>48.4</td><td>64.1</td><td>54.3</td><td>14.1</td><td>50.8</td><td>45.7</td><td>46.2</td><td>81.1</td><td>47.3</td><td>66.1</td><td>53.5</td><td>12.6</td><td>49.3</td><td>44.9</td><td>12.1</td></tr></table>
+
+Table 2: Person detection results on val and test sets of LOAF (§5.2).  
+
+<table><tr><td>Method</td><td>mAP↑</td><td>AP50↑</td><td>AP75↑</td><td>Precision↑</td><td>Recall↑</td><td>F-Score↑</td></tr><tr><td>Seide et al. [60]</td><td>20.9</td><td>50.6</td><td>10.2</td><td>80.6</td><td>39.5</td><td>53.0</td></tr><tr><td>Li et al. [42]</td><td>34.2</td><td>75.7</td><td>28.6</td><td>86.3</td><td>65.4</td><td>74.4</td></tr><tr><td>Tamura et al. [64]</td><td>29.3</td><td>61.0</td><td>23.4</td><td>88.8</td><td>51.2</td><td>65.0</td></tr><tr><td>RAPiD [24]</td><td>39.3</td><td>85.4</td><td>26.0</td><td>89.2</td><td>78.7</td><td>83.6</td></tr><tr><td>Ours</td><td>47.2</td><td>88.5</td><td>38.2</td><td>90.3</td><td>87.6</td><td>88.9</td></tr></table>
+
+tions [21], built on point-based human representation, treat the center of the detected human head as  $\pmb{p}$ . However, they suffer from a similar issue as the human head center is even often far from the exact position  $\pmb{p}$  human stand on, w.r.t. the image plane. Instead, with our radius-aligned human-box representation, the closest point to the principal point on the bounding box better corresponds to  $\pmb{p}$ .
+
+# 4.3. Implementation Details
+
+Network Architecture. Our fisheye person detector is built upon DAB-DETR [45], a prevalent variant of DETR [9] but converges much faster. Swin-T [46] is utilized as the backbone. For our LOAF with radius-aligned person boxes, the output of our detector is a 4D vector  $\pmb{b} \in [0,1]^4$  that parameterizes 2D center coordinates, height, and width. Note that, for traditional fisheye detection datasets like [24, 65] with arbitrary-oriented person box annotations, an extra output dimension is needed for rotation angle regression.
+
+Training Objective. Our fisheye person detector is end-to-end trained by jointly optimizing the vanilla detection loss used in DETR [9, 45] (referred as  $\mathcal{L}_{\mathrm{Det}}$ ) and our proposed rotation-equivariant constraint (i.e.,  $\mathcal{L}_{\mathrm{rotat - equi}}$  in Eq. 5):
+
+$$
+\mathcal {L} = \mathcal {L} _ {\mathrm {d e t}} + \lambda \mathcal {L} _ {\text {r o t a t - e q u i}}, \tag {8}
+$$
+
+where the coefficient  $\lambda$  is empirically set to 0.5.
+
+Training. Our fisheye detector is trained with a batch size of 8 for 50 epochs, where the AdamW [47] optimizer is employed with base learning rate 2e-4 and decayed by 0.1 at epoch 40. The remaining hyper-parameters are determined following [64, 24]. Specifically, we initialize backbones with ImageNet [22] pre-trained weights and adopt standard data augmentation techniques, i.e., color jitter, horizontal flip, and random scaling, with a base training size of  $608 \times 608$ . For CEPDOF [24] and WEPDTOF [65], we use a batch size of 128 and pre-train the detector on COCO [44] for 50 epochs to prevent over-fitting, as in [24, 65]. For the computation of our rotation-equivariant loss  $\mathcal{L}_{\mathrm{rotat - equi}}$ , the training images are rotated by a degree randomly sampled from 0 to 360. Our detector is implemented in PyTorch and trained on eight
+
+Table 3: Person detection results on CEPDOF [24] (\$5.2).  
+
+<table><tr><td>Method</td><td>mAP↑</td><td>AP50↑</td><td>AP75↑</td><td>Precision↑</td><td>Recall↑</td><td>F-Score↑</td></tr><tr><td>Seide et al. [60]</td><td>16.1</td><td>39.4</td><td>9.0</td><td>70.9</td><td>38.6</td><td>50.0</td></tr><tr><td>Li et al. [42]</td><td>25.2</td><td>69.9</td><td>30.2</td><td>81.4</td><td>64.5</td><td>72.0</td></tr><tr><td>Tamura et al. [64]</td><td>28.8</td><td>59.8</td><td>24.2</td><td>77.0</td><td>52.4</td><td>62.4</td></tr><tr><td>RAPiD [24]</td><td>37.7</td><td>72.0</td><td>26.8</td><td>73.3</td><td>67.8</td><td>70.4</td></tr><tr><td>Ours</td><td>45.8</td><td>85.6</td><td>36.6</td><td>84.9</td><td>74.6</td><td>79.8</td></tr></table>
+
+Table 4: Person detection results on WEPDTOF [65] (\$5.2).
+
+NVIDIA Tesla V100 GPUs with a 32GB memory per-card. Inference. Once trained, our fisheye detector can be directly applied for locating persons on the omnidirectional image plane. After that, the physical locations of the detected persons can be easily obtained through the numerical solution described in §4.2. For fisheye person detection, we follow prior work [60, 42, 64, 24] to use  $1024 \times 1024$  input resolution without any test-time augmentation or post-processing. Testing is conducted on a single NVIDIA V100 GPU with 16 GB memory.
+
+# 5. Experiment
+
+# 5.1. Experimental Setup
+
+Evaluation Protocol. On the top of LOAF, we conduct experiments for fisheye based person detection and localization. Following the dataset splitting (cf. §3.3), performance are reported on seen and unseen scenes, respectively, for both val and test sets. This allows us to assess the generalization ability over different surveillance scenarios. For comprehensive study, we further report person detection performance on two existing fisheye datasets, CEPDOF[24] and WEPDTOF[65]. Note that localization cannot be test since [24, 65] only provide person bounding box annotations. Evaluation Metrics. For fisheye person detection, we follow COCO [44] to report the mean average precision (mAP) for IoU  $\in$  [0.5:0.05:0.95]. We also employ AP $_{50}$  and AP $_{75}$  for further analysis. For fisheye person localization, we measure positional error (PE) in meters, i.e., Euclidean distance of calculated position and ground-truth position, as the conventions in visual localization [43, 58, 73, 79]. For detailed evaluation, on our LOAF, we report performance w.r.t. horizontal distance between human and the fisheye lens, i.e., near (0~10 m), middle (10~20 m), and far (larger than 20 m). Hence we have AP $_{n,m,f}$  and PE $_{n,m,f}$  accordingly.
+
+# 5.2. Performance on Person Detection
+
+LOAF. We first report the person detection performance on LOAF. Specifically, four recent deep learning based fisheye
+
+<table><tr><td rowspan="2">Method</td><td colspan="6">val</td><td colspan="6">test</td></tr><tr><td>mPE↓</td><td>PEn↓</td><td>PEm↓</td><td>PEf↓</td><td>PEseen↓</td><td>PEunseen↓</td><td>mPE↓</td><td>PEn↓</td><td>PEm↓</td><td>PEf↓</td><td>PEseen↓</td><td>PEunseen↓</td></tr><tr><td>Seide et al. [60]</td><td>1.298</td><td>0.561</td><td>1.332</td><td>3.109</td><td>1.206</td><td>1.382</td><td>1.321</td><td>0.706</td><td>1.309</td><td>3.482</td><td>1.306</td><td>1.386</td></tr><tr><td>Li et al. [42]</td><td>0.898</td><td>0.502</td><td>0.871</td><td>2.650</td><td>0.832</td><td>0.962</td><td>0.913</td><td>0.543</td><td>0.884</td><td>2.780</td><td>0.904</td><td>0.998</td></tr><tr><td>Tamura et al. [64]</td><td>0.755</td><td>0.429</td><td>0.736</td><td>1.862</td><td>0.709</td><td>0.821</td><td>0.778</td><td>0.471</td><td>0.826</td><td>2.160</td><td>0.724</td><td>0.836</td></tr><tr><td>RAPiD [24]</td><td>0.674</td><td>0.426</td><td>0.623</td><td>1.403</td><td>0.625</td><td>0.757</td><td>0.682</td><td>0.461</td><td>0.664</td><td>1.445</td><td>0.638</td><td>0.776</td></tr><tr><td>Ours</td><td>0.368</td><td>0.144</td><td>0.363</td><td>0.775</td><td>0.318</td><td>0.402</td><td>0.374</td><td>0.148</td><td>0.375</td><td>0.819</td><td>0.326</td><td>0.398</td></tr></table>
+
+person detectors [60, 42, 64, 24] are involved for comparison. Among them, [60, 42] are training-free, thus their scores are obtained by directly running the algorithms on the val and test data. As [64, 24] are trainable methods, we first train them on the train set of LOAF following their official setups, and then report the scores on the val and test sets. Quantitative comparison results are presented in Table 2. Some essential observations are as follows:
+
+- Our fisheye person detector significantly outperforms existing methods across all the metrics. For example, our detector provides a considerable performance gain in mAP, i.e.,  $7.1\%$  and  $7.0\%$  higher than the second best, RAPiD [24], on the val and test sets, respectively.  
+- Our detector not only handles nearby persons but also approaches distant targets well. Specifically, for the targets at  $10\sim 20\mathrm{m}$  away from the fisheye lens, our performance gain over other methods is more significant than the improvement reported with the persons at  $0\sim 10\mathrm{m}$  distance.  
+- Our detector also yields small performance gap between the seen and unseen scenes (e.g.,  $49.3 \rightarrow 44.9$  in terms of mAP), validating our good generalizability.  
+- For those persons at very far distance ( $>20\mathrm{~m}$ ), our algorithm, though still giving higher scores than other competitors, suffers from great performance degradation. This sheds light on the direction of our future efforts.
+
+CEPDOF [24]. CEPDOF contains eight videos with 25,504 frames in total. Following the official setup, we train our fisheye person detector on HABBOF [42] and MW-R [24] datasets, and report performance on CEPDOF. As shown in Table 3, our fisheye detector delivers state-of-the-art performance: it significantly outperforms RAPiD [24], the current top-leading algorithm, by  $7.9\%$  in terms of mAP.
+
+WEPDTOF [65]. WEPDTOF has 16 videos with 10,544 frames in total. Following the official setup, we use HAB-BOF [42], MW-R [24], and CEPDOF [24] for training and WEPDTOF for testing. Table 4 summarizes the results. Impressively, our detector greatly suppresses all the other competitors across all the evaluation metrics.
+
+# 5.3. Performance on Person Localization
+
+Then we study the person localization performance on LOAF. None of previous fisheye person detectors [60, 42, 64, 24] are aware of the task of person localization. We therefore follow the common practice [79, 21] in the field of visual localization: for [60, 42, 64, 24], we project
+
+Table 5: Person localization results on val and test sets of LOAF (§5.3).  
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">mAP↑</td><td colspan="3">val seen</td><td colspan="3">val unseen</td></tr><tr><td>mAP↑</td><td>AP50↑</td><td>AP75↑</td><td>mAP↑</td><td>AP50↑</td><td>AP75↑</td></tr><tr><td>Baseline [45]</td><td>43.1</td><td>46.4</td><td>81.2</td><td>48.4</td><td>40.9</td><td>76.2</td><td>38.9</td></tr><tr><td>+ Rotation Aug.</td><td>45.6</td><td>49.2</td><td>82.8</td><td>52.8</td><td>43.2</td><td>78.9</td><td>43.9</td></tr><tr><td>+ Rotation Equ.</td><td>47.4</td><td>50.8</td><td>83.8</td><td>54.8</td><td>45.7</td><td>80.5</td><td>47.3</td></tr></table>
+
+Table 6: Diagnostic results on val set of LOAF (§5.4).
+
+centers of their detected human boxes on the 3D world, based on the same strategy of ours (cf. §4.2), for physical position estimation. As shown in Table 5, our system produces much small localization errors in comparison with [60, 42, 64, 24], i.e.,  $\mathbf{0.374m}$  vs  $0.682\mathrm{m}[24]$  and 0.773 m of [64] on the test set of LOAF. In addition, for all the methods, the localization performance is declined as target distance increases, but our system suffers from the smallest drop. Moreover, the results suggest there is still large room for improvement, thus we hope that our dataset could encourage continuous efforts in this challenging task.
+
+# 5.4. Diagnostic Study
+
+Table 6 studies the efficacy of our rotation equivariant training strategy (§4.1), on the val set of LOAF. Our baseline model [45] (row #1) gains  $43.1\%$  mAP. After adopting rotational data augmentation (row #2), the performance boosts by  $2.5\%$  mAP. By contrast, our rotation equivariant training brings much larger improvements over the baseline, e.g.,  $4.3\%$  mAP gain. It is remarkable that, training with standard rotational data augmentation technique can be viewed as a specific case of our equivariant training - only learning rotation-equivariant object-querying (cf. Eq. 4).
+
+# 6. Conclusion
+
+We presented LOAF – the first top-view fisheye dataset that supports large-scale study for person localization in realistic surveillance scenarios. With radius-aligned personbox annotations and precise location ground-truths, LOAF closes a crucial gap in the literature as these cases are not covered by previous datasets and annotation protocols. We further proposed an efficient fisheye person detection model that is equipped with a rotation-equivariant training strategy. The physical locations of detected persons are formulated based on the fisheye model and the altitude of the camera. We empirically verified the effectiveness and promising performance of our algorithm.
+
+Acknowledgements This work was sponsored in part by the CCF-Tencent Open Fund, China Postdoctoral Science Foundation (2022M710467) and National Key Research and Development Program of China (Grant No. 2021YFF0500900).
+
+# References
+
+[1] Abdulrahman Alarifi, AbdulMalik Al-Salman, Mansour Alsaleh, Ahmad Alnafessah, Suheer Al-Hadhrami, Mai A Al-Ammar, and Hend S Al-Khalifa. Ultra wideband indoor positioning technologies: Analysis and recent advances. Sensors, 16(5):707, 2016. 2  
+[2] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. In CVPR, 2016. 2  
+[3] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In ECCV, 2018. 2  
+[4] Federico Boniardi, Abhinav Valada, Rohit Mohan, Tim Caselitz, and Wolfram Burgard. Robot localization in floor plans using a room layout edge extraction network. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2019. 2  
+[5] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In CVPR, 2017. 2  
+[6] Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In ICCV, 2019. 3  
+[7] Guillaume Bresson, Li Yu, Cyril Joly, and Fabien Moutarde. Urban localization with street views using a convolutional neural network for end-to-end camera pose regression. In IEEE Intelligent Vehicles Symposium, 2019. 2  
+[8] Jiale Cao, Yanwei Pang, Jin Xie, Fahad Shahbaz Khan, and Ling Shao. From handcrafted to deep features for pedestrian detection: a survey. IEEE TPAMI, 2021. 3  
+[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 5, 6, 7  
+[10] Bao Xin Chen and John K Tsotsos. Fast visual object tracking with rotated bounding boxes. arXiv preprint arXiv:1907.03892, 2019. 4  
+[11] David M Chen, Georges Baatz, Kevin Koser, Sam S Tsai, Ramakrishna Vedantham, Timo Pylvanainen, Kimmo Roimela, Xin Chen, Jeff Bach, Marc Pollefeys, et al. City-scale landmark identification on mobile devices. In CVPR, 2011. 3  
+[12] Long Chen, Shaobo Lin, Xiankai Lu, Dongpu Cao, Hangbin Wu, Chi Guo, Chun Liu, and Fei-Yue Wang. Deep neural network based vehicle and pedestrian detection for autonomous driving: A survey. IEEE TITS, 2021. 3  
+[13] Liang Chen, Ling Pei, Heidi Kuusniemi, Yuwei Chen, Tuomo Kröger, and Ruizhi Chen. Bayesian fusion for indoor positioning using bluetooth fingerprints. Wireless Personal Communications, 2013. 1, 2  
+[14] An-Ti Chiang and Yao Wang. Human detection in fish-eye images using hog-based detectors over rotated windows. In ICME Workshop, 2014. 3  
+[15] Sheng-Ho Chiang, Tsaipei Wang, and Yi-Fu Chen. Effi-
+
+cient pedestrian detection in top-view fisheye images using compositions of perspective view patches. Image and Vision Computing, 2021. 3  
+[16] Shin-Fang Ch'ng, Naoya Sogi, Pulak Purkait, Tat-Jun Chin, and Kazuhiro Fukui. Resolving marker pose ambiguity by robust rotation averaging with clique constraints. In ICRA, 2020. 2  
+[17] Ibrahim Cinaroglu and Yalin Bastanlar. A direct approach for human detection with catadioptric omnidirectional cameras. In Signal Processing and Communications Applications Conference, 2014. 3  
+[18] Taco Cohen and Max Welling. Group equivariant convolutional networks. In ICLR, 2016. 5  
+[19] Creative Commons. Attribution-noncommercial-sharealike 4.0 international. https://creativecommons.org/licenses/by-nc-sa/4.0/.4  
+[20] Joseph DeGol, Timothy Bretl, and Derek Hoiem. Chromatag: A colored marker and fast detection algorithm. In ICCV, 2017. 2  
+[21] Carlos R del Blanco, Pablo Carballeira, Fernando Jaureguizar, and Narciso Garcia. Robust people indoor localization with omnidirectional cameras using a grid of spatial-aware classifiers. Signal Processing: Image Communication, 2021. 2, 3, 4, 5, 7, 8  
+[22] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 7  
+[23] Mingyu Ding, Zhe Wang, Jiankai Sun, Jianping Shi, and Ping Luo. Camnet: Coarse-to-fine retrieval for camera relocalization. In ICCV, 2019. 2  
+[24] Zhihao Duan, Ozan Tezcan, Hayato Nakamura, Prakash Ishwar, and Janusz Konrad. Rapid: rotation-aware people detection in overhead fisheye images. In CVPR Workshop, 2020. 2, 3, 4, 5, 6, 7, 8  
+[25] Carlos Esteves, Christine Allen-Blanchette, Ameesh Makadia, and Kostas Daniilidis. Learning SO(3) equivariant representations with spherical cnns. In ECCV, 2018. 5  
+[26] Zahid Farid, Rosdiadee Nordin, and Mahamod Ismail. Recent advances in wireless indoor localization techniques and system. Journal of Computer Networks and Communications, 2013. 2  
+[27] Mark Fiala. Artag, a fiducial marker system using digital techniques. In CVPR, 2005. 2  
+[28] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3  
+[29] Hugo Germain, Vincent Lepetit, and Guillaume Bourmaud. Neural reprojection error: Merging feature learning and camera pose estimation. In CVPR, 2021. 3  
+[30] Yanying Gu, Anthony Lo, and Ignas Niemegeers. A survey of indoor positioning systems for wireless personal networks. IEEE Communications surveys & tutorials, 2009. 2  
+[31] Mingfei Han, Yali Wang, Xiaojun Chang, and Yu Qiao. Mining inter-video proposal relations for video object detection. In ECCV, 2020. 2  
+[32] Janne Haverinen and Anssi Kemppainen. Global indoor self-localization based on the ambient magnetic field. Robotics and Autonomous Systems, 2009. 2  
+[33] James Hays and Alexei A Efros. Im2gps: estimating geo
+
+graphic information from a single image. In CVPR, 2008. 2  
+[34] Joel A Hesch and Stergios I Roumeliotis. A direct least-squares (dls) method for pnp. In ICCV, 2011. 3  
+[35] Geoffrey E Hinton, Alex Krizhevsky, and Sida D Wang. Transforming auto-encoders. In ICANN, 2011. 5  
+[36] Danying Hu, Daniel DeTone, and Tomasz Malisiewicz. Deep charuco: Dark charuco marker pose estimation. In CVPR, 2019. 2  
+[37] Juho Kannala and Sami S Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. IEEE TPAMI, 28(8):1335-1340, 2006. 6  
+[38] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In CVPR, 2017. 2, 3  
+[39] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In ICCV, 2015, 1, 2, 3  
+[40] Oded Krams and Nahum Kiryati. People detection in top-view fisheye imaging. In IEEE International Conference on Advanced Video and Signal based Surveillance, 2017. 3  
+[41] Karel Lenc and Andrea Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In CVPR, 2015. 5  
+[42] Shengye Li, M Ozan Tezcan, Prakash Ishwar, and Janusz Konrad. Supervised people counting using an overhead fisheye camera. In IEEE International Conference on Advanced Video and Signal Based Surveillance, 2019. 2, 3, 4, 5, 7, 8  
+[43] Yunpeng Li, Noah Snavely, Dan Huttenlocher, and Pascal Fua. Worldwide pose estimation using 3d point clouds. In ECCV, 2012. 1, 3, 7  
+[44] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 7  
+[45] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In ICLR, 2022. 7, 8  
+[46] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 7  
+[47] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7  
+[48] Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year,  $1000\mathrm{km}$ : The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3-15, 2017. 3  
+[49] Kenro Miyamoto. Fish eye lens. JOSA, 54(8):1060-1061, 1964. 6  
+[50] Jun Qi and Guo-Ping Liu. A robust high-accuracy ultrasound indoor positioning system based on a wireless sensor network. Sensors, 17(11):2554, 2017. 2  
+[51] Hazem Rashed, Eslam Mohamed, Ganesh Sistu, Varun Ravi Kumar, Ciaran Eising, Ahmad El-Sallab, and Senthil Yogamani. Generalized object detection on fisheye cameras for autonomous driving: Dataset, representations and baseline. In WACV, 2021. 2  
+[52] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster,
+
+stronger. In CVPR, 2017. 3  
+[53] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 3  
+[54] Samer S Saab and Zahi S Nakad. A standalone RFID indoor positioning system using passive tags. IEEE Transactions on Industrial Electronics, 2010. 2  
+[55] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In CVPR, 2019. 3  
+[56] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In CVPR, 2021. 2  
+[57] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In CVPR, 2018. 2  
+[58] Torsten Sattler, Akihiko Torii, Josef Sivic, Marc Pollefeys, Hajime Taira, Masatoshi Okutomi, and Tomas Pajdla. Are large-scale 3d models really necessary for accurate visual localization? In CVPR, 2017. 2, 3, 7  
+[59] Gerald Schweighofer and Axel Pinz. Robust pose estimation from a planar target. IEEE TPAMI, 28(12):2024-2030, 2006. 2  
+[60] Roman Seidel, André Apitzsch, and Gangolf Hirtz. Improved person detection on omnidirectional images with non-maxima suppression. In International Conference on Computer Vision Theory and Applications, 2019. 2, 3, 6, 7, 8  
+[61] Steve Shafer, John Krumm, Barry Brumitt, Brian Meyers, Mary Czerwinski, and Daniel Robbins. The new easyliving project at microsoft research. In Proceedings of the 1998 DARPA/NIST Smart Spaces Workshop, 1998. 2  
+[62] Xun Sun, Yuanfan Xie, Pei Luo, and Liang Wang. A dataset for benchmarking image-based localization. In CVPR, 2017. 1  
+[63] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Akihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. In CVPR, 2018. 1  
+[64] Masato Tamura, Shota Horiguchi, and Tomokazu Murakami. Omnidirectional pedestrian detection by rotation invariant training. In WACV, 2019. 2, 3, 7, 8  
+[65] Ozan Tezcan, Zhihao Duan, Mertcan Cokbas, Prakash Ishwar, and Janusz Konrad. Wepdtof: A dataset and benchmark algorithms for in-the-wild people detection and tracking from overhead fisheye cameras. In WACV, 2022. 2, 4, 7, 8  
+[66] Gonzalo Vaca-Castano, Amir Roshan Zamir, and Mubarak Shah. City scale geo-spatial trajectory estimation of a moving camera. In CVPR, 2012. 2  
+[67] Florian Walch, Caner Hazirbas, Laura Leal-Taixe, Torsten Sattler, Sebastian Hilsenbeck, and Daniel Cremers. Image-based localization using lstms for structured feature correlation. In ICCV, 2017. 2, 3  
+[68] Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Philip HS Torr. Fast online object tracking and segmentation:
+
+A unifying approach. In CVPR, 2019. 4  
+[69] Tsaipei Wang, Chia-Wei Chang, and Yu-Shan Wu. Template-based people detection using a single downward-viewing fisheye camera. In International Symposium on Intelligent Signal Processing and Communication Systems, 2017. 3  
+[70] Wenguan Wang, James Chenhao Liang, and Dongfang Liu. Learning equivariant segmentation with instance-unique querying. In NeurIPS, 2013. 5  
+[71] Roy Want, Andy Hopper, Veronica Falcao, and Jonathan Gibbons. The active badge location system. ACM Transactions on Information Systems, 10(1):91-102, 1992. 2  
+[72] Tobias Weyand, Ilya Kostrikov, and James Philbin. Planetphoto geolocation with convolutional neural networks. In ECCV, 2016. 2  
+[73] Aoran Xiao, Ruizhi Chen, Deren Li, Yujin Chen, and Dewen Wu. An indoor positioning system based on static objects in large indoor scenes by using smartphone cameras. Sensors, 2018. 1, 7  
+[74] Chouchang Yang and Huai-Rong Shao. Wifi-based indoor positioning. IEEE Communications Magazine, 2015. 1, 2  
+[75] Luwei Yang, Ziqian Bai, Chengzhou Tang, Honghua Li, Yasutaka Furukawa, and Ping Tan. Sanet: Scene agnostic network for camera localization. In ICCV, 2019. 2  
+[76] Khalid Yousif, Yuichi Taguchi, and Srikumar Ramalingam. Monorgbd-slam: Simultaneous localization and mapping using both monocular and rgbd cameras. In ICRA, 2017. 2  
+[77] Amir Roshan Zamir and Mubarak Shah. Accurate image localization based on google maps street view. In ECCV, 2010. 2  
+[78] Amir Roshan Zamir and Mubarak Shah. Image geolocation based on multiplenearest neighbor feature matching using generalized graphs. IEEE TPAMI, 36(8):1546-1558, 2014. 2  
+[79] Jun Zhu, Jiangcheng Zhu, Xudong Wan, Chao Wu, and Chao Xu. Object detection and localization in 3d environment by fusing raw fisheye image and attitude data. Journal of Visual Communication and Image Representation, 59:128-139, 2019. 2, 6, 7, 8

2023/Large-Scale Person Detection and Localization Using Overhead Fisheye Cameras/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9531747fcfc12bb6145794f71b4dfe728e7a85c3a4e85a355a72f35bcc4d8e6b
+size 562472

2023/Late Stopping_ Avoiding Confidently Learning from Mislabeled Examples/c0b4cc8e-d6dd-4adb-8592-11f986f8c0fc_content_list.json

@@ -0,0 +1,1360 @@
+[
+    {
+        "type": "text",
+        "text": "Late Stopping: Avoiding Confidently Learning from Mislabeled Examples",
+        "text_level": 1,
+        "bbox": [
+            112,
+            130,
+            856,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Suqin Yuan $^{1}$  Lei Feng $^{2*}$  Tongliang Liu $^{1*}$ $^{1}$ The University of Sydney  $^{2}$ Nanyang Technological University",
+        "bbox": [
+            225,
+            179,
+            743,
+            217
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            251,
+            312,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sample selection is a prevalent method in learning with noisy labels, where small-loss data are typically considered as correctly labeled data. However, this method may not effectively identify clean hard examples with large losses, which are critical for achieving the model's close-to-optimal generalization performance. In this paper, we propose a new framework, Late Stopping, which leverages the intrinsic robust learning ability of DNNs through a prolonged training process. Specifically, Late Stopping gradually shrinks the noisy dataset by removing high-probability mislabeled examples while retaining the majority of clean hard examples in the training set throughout the learning process. We empirically observe that mislabeled and clean examples exhibit differences in the number of epochs required for them to be consistently and correctly classified, and thus high-probability mislabeled examples can be removed. Experimental results on benchmark-simulated and real-world noisy datasets demonstrate that the proposed method outperforms state-of-the-art counterparts.",
+        "bbox": [
+            75,
+            282,
+            472,
+            570
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            589,
+            209,
+            604
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Deep Neural Networks (DNNs) have achieved outstanding success in various tasks, while the success largely relies on data with high-quality annotations [13, 55, 41, 19]. In many real-world scenarios, it would be quite difficult to collect large-scale accurately labeled data, which could inevitably contain noisy labels. Unfortunately, previous studies [1, 56] showed that DNNs can easily overfit random labels, resulting in poor generalization performance. Therefore, an increasing number of methods have been proposed for Learning with Noisy Labels (LNL).",
+        "bbox": [
+            75,
+            614,
+            468,
+            766
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "One mainstream solution in existing methods of LNL is to train the classifier with confident examples [20, 14, 46, 34, 36], which is based on the memorization effect of DNNs, i.e., DNNs learn example with dominant patterns first and then overfit rare ones [1]. Given only noisy data, to exploit the memorization effect, the typical strategy starts from a small confident clean dataset and then",
+        "bbox": [
+            75,
+            767,
+            468,
+            875
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "gradually expands the dataset, which prevents DNNs from over-fitting noisy data. In general, there are two primary approaches to exploit confident examples in the learning process. The first approach involves identifying examples with high-probability clean labels and training the classifier based on these examples, which is commonly referred to as the \"small-loss trick\" [20, 14, 12, 28, 52]. The second approach involves controlling the learning process of the classifier to primarily learn high-probability clean examples in noisy datasets, which is commonly referred to as \"early stopping\" [37, 42, 17, 25, 3].",
+        "bbox": [
+            496,
+            252,
+            893,
+            417
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite providing satisfactory performance, these approaches present an unintended consequence in their methods to prevent over-fitting noise. Specifically, to reduce the impact of mislabeled examples, these approaches limit the capability of the model to effectively learn clean hard examples (CHEs) in noisy datasets, where CHEs are defined as clean examples that are close to the decision boundary, and a significant proportion of CHEs are non-dominated sub-population examples [9]. Identifying CHEs in noisy data is quite challenging [2, 22], as both the CHEs and the mislabeled examples are often characterized by large losses [52, 11, 6], causing them to become entangled. To maintain the purity of the dataset of confident examples, existing LNL methods [34, 2] normally try to eliminate potential examples that are likely to be mislabeled, which inevitably contain many CHEs. However, it is necessary to consider the positive impact of memorization effects on underrepresented sub-populations [10] (i.e., CHEs) when learning from natural datasets, as this is crucial for achieving close-to-optimal generalization performance.",
+        "bbox": [
+            496,
+            417,
+            893,
+            719
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this work, our goal is to enable the classifier to learn as many useful non-dominated sub-population examples in the training set as possible during the process of learning with noisy labels. This entails selecting as many clean examples as possible, particularly the clean hard examples, during the sample selection process. In relation to this context, we introduce a novel concept termed First-time  $k$ -epoch Learning (FkL), which is defined as the index of the epoch during the training procedure where an example has been predicted to its given label for consecutive  $k$  epochs for the first time, as shown in Figure 1(a).",
+        "bbox": [
+            496,
+            720,
+            893,
+            886
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            805,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding authors.",
+        "bbox": [
+            94,
+            887,
+            227,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "16079",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/8447846622861a9cee5d4f28f3f3d1bbc44d3f359d99f06c02ddc5be8aa762b3.jpg",
+        "image_caption": [
+            "(a)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            112,
+            85,
+            315,
+            328
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/3ee4c5efc137a5350c63ffbd5465764dbc038deaa61fe00f82238f8df48c3782.jpg",
+        "image_caption": [
+            "(b)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            338,
+            89,
+            602,
+            321
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/d99cd1818e22dd2eb077d300100d3b39ea0745cc934bed036a57620be3aaa668.jpg",
+        "image_caption": [
+            "Figure 1. (a) We propose the First-time  $k$ -epoch Learning (FkL) metric, which determines the minimum index of the training epoch until which an example has been predicted as its given label for consecutive  $k$  epochs. (b) The normalised histogram of CIFAR-10 examples with  $40\\%$  symmetric label noise w.r.t. the sequence they meet the FkL metric during training procedure. The horizontal axis represents the sequential order in which training examples meet the FkL metric. The vertical axis represents the normalised histogram of examples. (c) Rather than the methods that start from a small yet clean training set ('Early' Stopping), our proposed framework starts from a large training set ('Late' Stopping) that retains as many clean examples as possible."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            629,
+            90,
+            736,
+            325
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/92929d3565c32fbf4e1d033d9a5b330e55d91b52dab584bce45e04747d74cd59.jpg",
+        "image_caption": [
+            "(c)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            750,
+            90,
+            854,
+            325
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Using First-time  $k$ -epoch Learning (FkL) as a metric, we sequence the examples in the noisy dataset according to the order they meet the FkL metric during the training procedure, as shown in Figure 1(b). As shown in the same figure, most unlabeled examples can only be classified into their given labels for consecutive  $k$  epochs in the later stages of training (i.e., larger in the sequence), which implies a relatively large FkL value. This observation suggests that the examples with large FkL values (i.e., those examples that are classified to their given labels for consecutive  $k$  epochs only in the late training stage) are predominantly those with incorrect labels. Therefore, the FkL values of different training examples can be used to distinguish whether an example is unlabeled or not.",
+        "bbox": [
+            75,
+            445,
+            468,
+            655
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Motivated by the above observations, we propose a novel method based on reverse thinking of the conventional confident example selection strategy. Our proposed method, called Late Stopping, employs an iterative sample selection process that gradually reduces the noise rate of the dataset, leading to a positive feedback loop. Instead of selecting high-probability clean examples in the early stage, our method focuses on selecting high-probability mislabeled examples in the late stage to retain as many CHEs as possible in the training set throughout the learning process, even though this method may lead to the retention of an acceptable level of mislabeled examples, as shown in 1(c). Building on this, we introduce a novel sample selection criterion, FkL, which works effectively in selecting mislabeled examples under the Late Stopping framework, and empirical results show that the FkL is more ef",
+        "bbox": [
+            75,
+            659,
+            470,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "fective than the loss criterion. We evaluate our method on benchmark-simulated and real-world noisy datasets. Empirical results demonstrate that the Late Stopping method achieves superior performance compared with state-of-the-art counterparts of learning with noisy labels.",
+        "bbox": [
+            496,
+            445,
+            890,
+            521
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            539,
+            640,
+            555
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Existing methods of learning with noisy labels can be broadly categorized into two types. The first category includes classifier-consistent algorithms that use the noise transition matrix [27, 29, 54, 35, 8], which indicates the probabilities of clean labels flipping to noisy labels. The second category focuses on heuristic methods to mitigate the impact of label noise [36, 50, 39, 45, 30, 49, 36], which is also the primary focus of our work. Many existing heuristic methods in LNL are based on the memorization effect, where the algorithm attempts to only learn from confident clean examples [20, 14, 46, 34]. Alternatively optimizing the classifier and updating the training set is not a new idea in LNL. For instance, Joint Optim [42], Co-teaching [14], SELF [34], and Me-Momentum [2] employ a similar positive feedback loop to our proposed Late Stopping.",
+        "bbox": [
+            496,
+            566,
+            890,
+            792
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Previous research on the dynamics of DNNs has demonstrated that hard examples are typically learned during the late stage of the learning process [1, 43, 31]. Furthermore, many recent studies have employed various training-time metrics to quantify the \"hardness\" of examples [32, 16, 7, 4, 21, 31], leading to an increase in LNL approaches that use learning dynamics to select clean examples. For instance,",
+        "bbox": [
+            496,
+            795,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "16080",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [
+            "Algorithm 1 Late Stopping"
+        ],
+        "code_body": "Input: Original noisy training set  $\\mathcal{D}_1$ , iteration rate  $m\\%$ , noise rate  $n\\%$ , epoch  $T_{max}$  and iteration  $I_{max}$ .  \nOutput: Extracted final training set and the classifier.  \n1: for  $I = 1, \\dots, I_{max}$  do  \n2: Initialize  $\\mathcal{D}_i$ ,  $S_{F_i} = 0$  and  $\\mathcal{D}_{F_i} = [\\ldots]$ .  \n3: //Initialize new training set, the number/dataset of examples that meet the First-time k-epoch Learning, respectively.  \n4: for  $T = 1, \\dots, T_{max}$  do  \n5: Train  $f_i$  on  $\\mathcal{D}_i$ .  \n6: Update  $\\mathcal{D}_{F_i}$ ,  $S_{F_i}$ .  \n7: Break and Output  $\\mathcal{D}_{i+1} = \\mathcal{D}_{F_i}$ , if  $S_{F_i} > S_{F_{i-1}} \\times (1 - m\\%)$ .  \n8: end for  \n9: Break and Output  $f_i$  and  $\\mathcal{D}_{F_i}$ , if  $m \\times i > n$ .  \n10: end for",
+        "bbox": [
+            76,
+            108,
+            883,
+            292
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "FSLT&SSFT [31], Self-Filtering [48], SELFIE [40], and RoCL [57] adopt a criterion similar to our proposed FkL criterion. Specifically, FSLT&SSFT and Self-Filtering use the classifier's error prediction to pinpoint clean examples, SELFIE employs entropy values from prediction histories for selection, and RoCL leverages the variance of training losses to select clean examples. Instead of quantifying the \"hardness\" of examples based on the characteristics of individual examples in the dataset, our proposed method focuses on extracting the intrinsic robust learning ability of DNNs from the model trained on noisy datasets.",
+        "bbox": [
+            75,
+            319,
+            473,
+            487
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Late Stopping",
+        "text_level": 1,
+        "bbox": [
+            76,
+            503,
+            220,
+            521
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we elaborate on our proposed Late Stopping method. As shown in Figure 1, the primary goal of the Late Stopping framework is to maximize the retention of clean examples in the training set throughout the learning process, enabling our method to learn clean hard examples (CHEs) from a noisy training set. Therefore, we halt the training process in the late training stage, leveraging First-time  $k$ -epoch Learning (FkL) to exploit the results of the intrinsic robust learning ability of DNNs for sample selection. Iteratively performing this operation allows for a gradual reduction of noise while maximizing the retention of clean samples in the training dataset.",
+        "bbox": [
+            75,
+            529,
+            470,
+            710
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Algorithm Flow",
+        "text_level": 1,
+        "bbox": [
+            76,
+            723,
+            235,
+            739
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For a comprehensive explanation of Late Stopping, we begin by formally defining First-time  $k$ -epoch Learning (FkL). In learning with noisy labels, let us consider a model  $f_{i}$  trained on a noisy training set  $\\mathcal{D}_i$  composed of  $n$  training examples  $\\{\\mathbf{x}_j,y_j\\}_{j = 1}^n$  where  $y_{j}$  denotes the given label (which may not be true) of  $\\mathbf{x}_j$ . After  $t$  training epochs, the predicted label  $\\hat{y}_j^t$  for the instance  $\\mathbf{x}_j$  can be obtained. Let  $\\mathrm{acc}_j^t = \\mathbb{1}_{\\hat{y}_j^t = y_j}$  denote a binary variable indicating whether  $f_{i}$  predicts the given label  $y_{j}$  of the instance  $\\mathbf{x}_j$  at epoch  $t$ . The FkL for the instance  $\\mathbf{x}_j$  is defined as the minimum",
+        "bbox": [
+            75,
+            748,
+            468,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "index of the training epoch that the instance  $\\mathbf{x}_j$  has been predicted to its given label  $y_{j}$  for  $k$  consecutive epochs:",
+        "bbox": [
+            496,
+            320,
+            890,
+            351
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathrm {F k L} _ {j} = \\underset {t ^ {*} \\in t} {\\operatorname {a r g m i n}} \\left(\\mathrm {a c c} _ {j} ^ {t ^ {*}} \\wedge \\ldots \\wedge \\mathrm {a c c} _ {j} ^ {(t ^ {*} - k + 1)} = 1\\right).\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            526,
+            354,
+            864,
+            383
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "If  $\\mathrm{FkL}_j = t^*$ , it implies that the classifier  $f_{i}$  \"learns\" the instance  $\\mathbf{x}_j$  after  $t^*$  epochs. Notably, if  $\\mathbf{x}_j$  is \"learned\" by the classifier early in the training, it will have a small FkL value, however, if it is \"learned\" in the late stage, it will have a large FkL value. Based on our FkL definition, we observed that the majority of examples with large FkL values are mislabeled, as shown in Figure 1(a). This observation provides a rationale for selecting examples with large FkL values, those learned in the late training stage, as high-probability mislabeled examples.",
+        "bbox": [
+            496,
+            386,
+            890,
+            537
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "With the FkL selection criterion, we present the algorithmic framework of Late Stopping. At the low level (Steps 4-8 of Algorithm 1), the sample selection strategy follows a \"from easy to hard\" curriculum learning procedure [5, 18]. During the  $i$ -th iteration, the classifier  $f_{i}$  is trained on a new training set  $\\mathcal{D}_i$  (Step 5), and newly identified FkL examples are added to  $\\mathcal{D}_{F_i}$  (Step 6). If the size of  $\\mathcal{D}_{F_i}$ , represented as  $S_{F_i}$ , exceeds a predefined threshold, the training in this iteration will halt, and the output becomes  $\\mathcal{D}_{F_i}$  (Step 7). At the higher level (Steps 1-10 of Algorithm 1), our method learns through a \"from hard to easy\" positive feedback loop, distinguishing it from most existing LNL approaches. The algorithm starts with initializing  $\\mathcal{D}_1$  and training  $f_{1}$  (Iteration 1). In subsequent iterations, the new training set  $\\mathcal{D}_i$  (Step 2, Iteration  $i$ ) is updated from the previous iteration's  $\\mathcal{D}_{F_{i - 1}}$  (Step 8, Iteration  $i - 1$ ). When the stopping condition is met, the output is  $f_{i}$  and  $\\mathcal{D}_{F_i}$  (Step 10).",
+        "bbox": [
+            496,
+            537,
+            892,
+            795
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Learning CHEs by Positive Feedback Loop",
+        "text_level": 1,
+        "bbox": [
+            500,
+            801,
+            867,
+            818
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Here, we delve deeper into the typical positive feedback loop in LNL, highlighting its limitations in effectively learning CHEs. We then elucidate how our proposed method overcomes these limitations and maximizes the retention of clean examples throughout the learning process.",
+        "bbox": [
+            496,
+            824,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "16081",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Typical positive feedback loop in LNL. According to statistical learning theory [33], training data with better quality yields a better classifier. Therefore, we can improve the classifier's generalization performance in LNL through the construction of a positive feedback loop. The goal of the positive feedback loop is that, after each training step, the new classifier showcases enhanced generalization performance compared with its predecessor. By doing so, it can more robustly guide the sample selection process, enabling the use of improved training data for classifier training in the succeeding steps [2].",
+        "bbox": [
+            76,
+            90,
+            472,
+            257
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Most LNL methods, relying on confident examples, initiate with either a small yet clean training set or a weak yet reliable classifier. To enhance the classifier's generalization performance during initialization, they often rigorously limit memorization effects and eliminate potential mislabeled examples, even if this entails removing many CHEs. By restricting the memorization of mislabeled examples, they achieve superior generalization performance compared with unrestricted training on the original set. Then, they either iteratively or incrementally, train classifiers, ultimately obtaining either a larger clean training set or a more robust classifier. This strategy is effective when the classes or subclasses within the dataset are well-balanced and the examples within each class are interrelated, allowing the classifier to learn harder examples from easier ones.",
+        "bbox": [
+            76,
+            260,
+            472,
+            487
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Limitations of typical positive feedback loop. Although LNL methods that rely on a typical positive feedback loop achieve good generalization performance in a variety of benchmark tasks, natural datasets are often highly imbalanced. This imbalance poses a challenge to achieving close-to-optimal generalization performance with existing LNL methods. Specifically, such challenges are because CHEs are often characterized by large losses and non-dominant patterns. As a result, substantial learning about CHEs predominantly transpires during the late stages of training, driven by memorization effects [2, 52]. The conventional positive feedback loop, aimed at preventing overfitting, often strictly limits the memorization effects of DNNs. Consequently, this strict constraint reduces a classifier's ability to learn from under-represented subpopulations in datasets [9], and thus limits its learning primarily to typical (dominant) classes and subclasses. Such constraints ultimately lead to a situation where the learning ability of rare classes and subclasses, i.e., CHEs, is continuously neglected, while the learning ability of the classes and subclasses representing the dominant pattern is continuously enhanced.",
+        "bbox": [
+            76,
+            489,
+            472,
+            821
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Positive feedback loop in Late Stopping. Based on our previous discussions, to achieve close-to-optimal generalization performance on natural datasets, we need to focus on learning from CHEs throughout the training process. To achieve this, we prolong the training process for each train-",
+        "bbox": [
+            76,
+            825,
+            472,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ing iteration to make use of the memorization effects to help memorize CHEs. With a loose sample selection process, we can ensure that we maximize the retention of clean examples in the training set throughout the learning process. However, retaining as many clean examples as possible during the training initialization requires accepting a significant amount of label noise in the initial training set. Additionally, relying on the \"memorization effects in the late training stage\" to learn CHEs might result in poor generalization performance, as the classifier may seriously overfit label noise in this stage.",
+        "bbox": [
+            496,
+            90,
+            893,
+            257
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To counter these negative effects, we propose a \"from hard to easy\" positive feedback loop within the Late Stopping framework. While we prolong the training process in each training iteration, we simultaneously use the classifier's robust learning capabilities as the criterion for sample selection, i.e., FkL. In each iteration, we train a new classifier on a new training set, using it solely to guide the sample selection for the current iteration. In this positive feedback loop, we opt to disregard the generalization performance of the classifier in each iteration. Instead, we exploit the robust learning capabilities of each new classifier,  $f_{i}$ , in its  $i$  iteration on  $\\mathcal{D}_i$  from its learning dynamics, i.e., using the FkL criterion to select examples for the next iteration.",
+        "bbox": [
+            496,
+            260,
+            893,
+            455
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The new training set, obtained by the FkL criterion from the previous iteration, will have less noise. This enables the training of a better new classifier, which can learn more from the CHEs before it begins to memorize mislabeled examples. Therefore, tracking learning dynamics using the FkL criterion can better distinguish between CHEs and mislabeled examples, resulting in a new training set with even less noise than the previous one but contains CHEs. By observing each classifier's learning dynamics through FkL in every iteration, we can iteratively refine our training set, reducing noise while retaining the majority of clean examples, as shown in Figure 2.",
+        "bbox": [
+            496,
+            459,
+            893,
+            641
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Discussion. Our proposed method is motivated by the seemingly conflicting views on the ability of DNNs to handle label noise as presented in recent studies: \"Deep neural networks easily fit random labels [56]\" and \"Deep learning is robust to massive label noises [37]\". Empirical evidence [1] suggests that DNNs first learn simple patterns before over-fitting in the late training stage. However, there is a scarcity of studies on the behavior of DNNs during the intermediate stages of training, which may be a critical factor for understanding the robustness of DNNs to label noise - DNNs might exhibit spontaneously robust learning to massive label noises in all periods of training, though the generalization performance is over-fitting noisy labels. Thus, we aim to separate the robust learning ability of DNNs from their generalization performance. To this end, we leverage the intrinsic robust learning ability of DNNs by continuing the training process until the generalization performance se",
+        "bbox": [
+            496,
+            643,
+            893,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "16082",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/0bb0cf552779ea2d7d338e7ad3a121554314ee42301076f0542b61a316dd11ab.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            101,
+            95,
+            295,
+            305
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/8fad7cead5903932755b2f00bc7dd9c76e89556aec67dd378b77e65d22dba3bb.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            305,
+            95,
+            483,
+            304
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/35340cd50c450ee8c74a49181468d1b2650c7db2c36740a911b14ceb1e96ad41.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            495,
+            95,
+            673,
+            304
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/d5116a105fc06502ef1c75a082b10053ab028240597fe2ef44cf1ccdebe98cee.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            686,
+            95,
+            864,
+            304
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/a6a93e6d9452aafb9b41fe113129918007448a49fe81938f85cbceb640ac6010.jpg",
+        "image_caption": [
+            "Figure 2. We refer to one update of the classifier and the training set as one iteration. Top subfigures: We illustrate the changes in the Test-set Accuracy and the Training-set Accuracy of the classifier during the training process of Late Stopping. There are ten peaks in each figure because we set the number of iterations  $= 10$  and the classifiers are re-initialized at the beginning of each new iteration. Bottom subfigures: We illustrate the changes in the noise rate of the training set in each iteration. There are ten bars in each figure because we set the number of iterations  $= 10$  and the training set of each iteration is obtained from the previous iteration."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            112,
+            309,
+            292,
+            422
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/7fccf7d2d3a6a3bbe2811305228f9405440e05bba75c70001bbf14ae7446220f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            303,
+            311,
+            480,
+            422
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/b9be724f8ac70093670aa00831aa917c795c01e624a5767017bae51d2f58f3bb.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            491,
+            311,
+            669,
+            422
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/67f65b7a20cff59ecadc867180a36b27facc844bff3eea3aa0c89169e852d1d3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            687,
+            311,
+            864,
+            422
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "riously deteriorates (i.e., Late Stopping) and we extract the robust learning results during this process (i.e., FkL).",
+        "bbox": [
+            75,
+            523,
+            470,
+            555
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            569,
+            209,
+            588
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we conduct experiments on both synthetic and real-world datasets containing label noise to validate the effectiveness of our proposed Late Stopping.",
+        "bbox": [
+            75,
+            595,
+            468,
+            642
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets. We use two popular benchmark datasets i.e., CIFAR-10 and CIFAR-100 [23], to test the generalization performance of our proposed method. Following previous works [2, 51], we manually corrupt CIFAR-10 and CIFAR-100 with Symmetric noise (abbreviated as Sym.) [44] and Instance-dependent (abbreviated as Ins.) noise [53]. For Sym. noise and Ins. noise, the noise rate is set to  $20\\%$  and  $40\\%$ . For a fair comparison, we leave out  $10\\%$  of the noisy training data. We employ the real-world noisy dataset CIFAR-10N [47], which equips the training datasets of CIFAR-10 with human-annotated real-world noisy labels. More details about the above datasets and noise types are provided in Appendix A1.",
+        "bbox": [
+            75,
+            643,
+            470,
+            838
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Compared methods. We compare the Late Stopping method (Algorithm 1) with the following well-known methods: (1) Statistically inconsistent approaches (small-loss trick): MentorNet [20], Co-teaching [14], and JoCoR [46];",
+        "bbox": [
+            75,
+            839,
+            470,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "(2) Statistically inconsistent approaches (other tricks): Joint Optim [42], CDR [51], and Me-Momentum [2]; (3) Statistically consistent approaches: Forward [35] and DMI [54]. Note that drawing a direct comparison with certain semi-supervised approaches, especially those with multiple technical aggregations such as SELF [34], DivideMix [24], ELR+ [26], and Unicon [22]. The results for baselines are copied from original papers and [2, 47].",
+        "bbox": [
+            496,
+            523,
+            890,
+            643
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Network structure and experimental setup. We utilize a typical warming-up strategy to ensure stable predictions of the DNNs. The optimizer settings used in our experiments are as follows: SGD with a momentum of 0.9, weight decay of 5e-4, batch size of 128, and initial learning rate of 0.02. In our experiments, we applied typical data augmentations such as horizontal flipping and random cropping. For the experiments on synthetic noisy datasets, ResNet-18 and ResNet-34 [15] are used for CIFAR-10 and CIFAR-100, respectively. For the experiments on real-world noisy datasets, ResNet-34 is used for CIFAR-10N. More details about our experimental setup are provided in Appendix A2.",
+        "bbox": [
+            496,
+            643,
+            892,
+            825
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Effectiveness of Late Stopping",
+        "text_level": 1,
+        "bbox": [
+            500,
+            832,
+            769,
+            849
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Here, we aim to provide empirical evidence to verify the effectiveness of Late Stopping. We conduct experiments on the CIFAR-10 and CIFAR-100 datasets, as their",
+        "bbox": [
+            496,
+            854,
+            893,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "16083",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/15304dee3243407b58892187e9aa4e0abd280f9a9e4e627067536ca30318710f.jpg",
+        "table_caption": [
+            "Table 1. Label precision of the selected examples (45k in total, with 27k clean examples)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Data&amp;Noise</td><td rowspan=\"2\">criterion</td><td colspan=\"3\">Selecting Clean Examples</td><td colspan=\"3\">Selecting Mislabeled Examples</td></tr><tr><td>0-10k</td><td>0-15k</td><td>0-25k</td><td>30k-45k</td><td>35k-45k</td><td>40k-45k</td></tr><tr><td>CIFAR-10</td><td>loss</td><td>92.73%</td><td>89.56%</td><td>82.89%</td><td>75.67%</td><td>83.49%</td><td>91.54%</td></tr><tr><td>Sym. 40%</td><td>FkL(Ours)</td><td>95.49%</td><td>95.59%</td><td>93.56%</td><td>96.39%</td><td>99.81%</td><td>99.94%</td></tr><tr><td>CIFAR-10</td><td>loss</td><td>75.55%</td><td>72.33%</td><td>71.38%</td><td>60.27%</td><td>68.69%</td><td>77.28%</td></tr><tr><td>Ins. 40%</td><td>FkL(Ours)</td><td>81.88%</td><td>81.67%</td><td>81.51%</td><td>78.51%</td><td>88.85%</td><td>99.30%</td></tr><tr><td>CIFAR-100</td><td>loss</td><td>88.43%</td><td>84.72%</td><td>78.49%</td><td>68.02%</td><td>75.73%</td><td>84.68%</td></tr><tr><td>Sym. 40%</td><td>FkL(Ours)</td><td>97.45%</td><td>96.78%</td><td>95.06%</td><td>94.63%</td><td>97.91%</td><td>98.86%</td></tr><tr><td>CIFAR-100</td><td>loss</td><td>71.30%</td><td>68.23%</td><td>65.29%</td><td>45.91%</td><td>44.23%</td><td>42.56%</td></tr><tr><td>Ins. 40%</td><td>FkL(Ours)</td><td>89.25%</td><td>88.56%</td><td>84.40%</td><td>78.22%</td><td>86.28%</td><td>91.76%</td></tr></table>",
+        "bbox": [
+            86,
+            101,
+            883,
+            284
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/b287010e5edc6dadd6e08c0380aeacc50bf98552a4b763d313566128bfa7169d.jpg",
+        "table_caption": [
+            "Table 2. Numbers of clean examples of the final training set (27k in total)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>CIFAR-10 (Sym. 40%)</td><td>CIFAR-10 (Ins. 40%)</td><td>CIFAR-100 (Sym. 40%)</td><td>CIFAR-100 (Ins. 40%)</td></tr><tr><td>Me-Momentum [2]</td><td>25,791</td><td>25,694</td><td>22,244</td><td>20,779</td></tr><tr><td>Late Stopping (Ours)</td><td>26,562</td><td>25,915</td><td>26,120</td><td>23,979</td></tr></table>",
+        "bbox": [
+            86,
+            305,
+            883,
+            363
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "ground-truth labels are available. The average results of five runs of our experiments are presented in Figure 2, which demonstrates the iterative improvement in the quality of the training set when Late Stopping is applied in the learning process. And a classifier with better generalization performance is obtained. We demonstrate the ability of our proposed FkL criterion to select mislabeled examples under the Late Stopping framework, compared with the loss criterion. Additionally, we show the effectiveness of our proposed Late Stopping in retaining clean examples in the training set by comparing it with Me-Momentum [2], which is also an sample selection method that focuses on extracting clean hard examples.",
+        "bbox": [
+            75,
+            385,
+            468,
+            582
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Comparison with loss criterion. The selection of confident clean examples based on the small-loss criterion is a commonly used approach in LNL and has been shown to be effective [20, 14, 46]. However, the approach of selecting mislabeled examples based on the large-loss [6] can be tricky since it fails to distinguish between mislabeled and clean hard examples [52]. We conducted experiments on the CIFAR-10 and CIFAR-100 datasets to compare the effectiveness of the FkL and loss criterion for selecting examples during the same training process. Specifically, we ranked all examples (45k in total, with 27k clean examples) using the FkL criterion and the loss criterion, respectively, from the smallest 0k to the largest 45k. We then compared the precision of the two criteria in selecting clean examples and mislabeled examples in different ranges. The details of the experiments and the results are presented in Table 1. Our experimental results demonstrate that the FkL criterion outperforms the loss criterion in selecting both clean and mislabeled examples under the Late Stopping framework. The largest performance gap was observed in the  $40k - 45k$  range. These results provide empirical evidence that the FkL cri",
+        "bbox": [
+            75,
+            583,
+            470,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "terion is more effective than the loss criterion for selecting mislabeled examples in the Late Stopping framework.",
+        "bbox": [
+            496,
+            385,
+            890,
+            414
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Comparison with hard example selection approach. In Table 2, we compare our proposed method with Me-Momentum [2] which also focuses on selecting clean hard examples. Since there are no clear definitions of clean hard examples, we cannot directly compare the effectiveness of selecting such examples. Nevertheless, we can make an indirect comparison by comparing the ability to select clean examples. Our experimental results demonstrate that Late Stopping is an effective approach for retaining clean hard examples. On the CIFAR-10 dataset, both Late Stopping and Me-Momentum showed comparable performance. However, on the more challenging CIFAR-100 dataset, Late Stopping significantly achieved better performance in selecting clean hard examples.",
+        "bbox": [
+            496,
+            415,
+            892,
+            627
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Classification Accuracy",
+        "text_level": 1,
+        "bbox": [
+            500,
+            635,
+            718,
+            651
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Synthetic datasets. The experimental results on test accuracy on synthetic datasets with class-dependent and instance-dependent label noise are provided in Table 3 and 4. Each trial is repeated five times and the mean value and standard deviation are recorded. On CIFAR-10, our method achieves varying degrees of lead over baselines. In the  $20\\%$  symmetric noise task, which is the simplest task in all cases, the Me-Momentum baseline outperforms ours, which illustrates the effectiveness of conventional sample selection based on confident clean examples for simple LNL. For the more challenging CIFAR-100 dataset, our proposed method consistently achieved the best results. The size of each class in CIFAR-100 is ten times smaller than that of CIFAR-10, making it difficult to retain sufficient CHEs while maintaining a small yet clean training set using conventional sample selection methods. Methods that focus on improving the",
+        "bbox": [
+            496,
+            657,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "16084",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/bb51523ffb97f7b2aa4a3a85ec020d7c7f9a7e33cf12d47c9a6fd199cc068955.jpg",
+        "table_caption": [
+            "Table 3. Test performance (mean±std) of each approach using ResNet-18 on CIFAR-10."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Sym. 20%</td><td>Sym. 40%</td><td>Ins. 20%</td><td>Ins. 40%</td></tr><tr><td>Late Stopping(Ours)</td><td>91.06±0.22%</td><td>88.92±0.38%</td><td>91.08±0.23%</td><td>87.41±0.38%</td></tr><tr><td>Cross-Entropy [38]</td><td>85.00±0.43%</td><td>79.59±1.31%</td><td>85.92±1.09%</td><td>79.91±1.41%</td></tr><tr><td>MentorNet [20]</td><td>80.49±0.11%</td><td>77.48±3.45%</td><td>79.12±0.42%</td><td>70.27±1.52%</td></tr><tr><td>Forward [35]</td><td>85.63±0.11%</td><td>74.30±0.26%</td><td>85.29±0.38%</td><td>74.72±3.24%</td></tr><tr><td>Co-teaching [14]</td><td>87.16±0.52%</td><td>83.59±0.28%</td><td>86.54±0.11%</td><td>80.98±0.39%</td></tr><tr><td>JoCoR [46]</td><td>88.69±0.19%</td><td>85.44±0.29%</td><td>87.31±0.27%</td><td>82.49±0.57%</td></tr><tr><td>DMI [54]</td><td>88.18±0.13%</td><td>83.98±0.48%</td><td>89.14±0.36%</td><td>84.78±1.97%</td></tr><tr><td>Joint Optim [42]</td><td>89.70±0.36%</td><td>87.79±0.20%</td><td>89.69±0.42%</td><td>82.62±0.57%</td></tr><tr><td>CDR [51]</td><td>89.68±0.38%</td><td>86.13±0.44%</td><td>90.24±0.39%</td><td>83.07±1.33%</td></tr><tr><td>Me-Momentum [2]</td><td>91.44±0.33%</td><td>88.39±0.34%</td><td>90.86±0.21%</td><td>86.66±0.91%</td></tr></table>",
+        "bbox": [
+            86,
+            101,
+            883,
+            284
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/a2f8a0f48b28f0c7201c92c4c1df03b20c6c335d2ed80467e32aa6c32065b7e7.jpg",
+        "table_caption": [
+            "Table 4. Test performance (mean±std) of each approach using ResNet-34 on CIFAR-100."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Sym. 20%</td><td>Sym. 40%</td><td>Ins. 20%</td><td>Ins. 40%</td></tr><tr><td>Late Stopping(Ours)</td><td>68.67±0.67%</td><td>64.10±0.40%</td><td>68.59±0.70%</td><td>59.28±0.46%</td></tr><tr><td>Cross-Entropy [38]</td><td>57.59±2.55%</td><td>45.74±2.61%</td><td>59.85±1.56%</td><td>43.74±1.54%</td></tr><tr><td>MentorNet [20]</td><td>52.11±0.10%</td><td>35.12±1.13%</td><td>51.73±0.17%</td><td>40.90±0.45%</td></tr><tr><td>Forward [35]</td><td>57.75±0.37%</td><td>38.59±1.62%</td><td>58.76±0.66%</td><td>44.50±0.72%</td></tr><tr><td>Co-teaching [14]</td><td>59.28±0.47%</td><td>51.60±0.49%</td><td>57.24±0.69%</td><td>45.69±0.99%</td></tr><tr><td>JoCoR [46]</td><td>64.17±0.19%</td><td>55.97±0.46%</td><td>61.98±0.39%</td><td>50.59±0.71%</td></tr><tr><td>DMI [54]</td><td>58.73±0.70%</td><td>49.81±1.22%</td><td>58.05±0.20%</td><td>47.36±0.68%</td></tr><tr><td>Joint Optim [42]</td><td>64.55±0.38%</td><td>57.97±0.67%</td><td>65.15±0.31%</td><td>55.57±0.41%</td></tr><tr><td>CDR [51]</td><td>66.52±0.24%</td><td>60.18±0.22%</td><td>67.06±0.50%</td><td>56.86±0.62%</td></tr><tr><td>Me-Momentum [2]</td><td>68.03±0.53%</td><td>63.48±0.72%</td><td>68.11±0.57%</td><td>58.38±1.28%</td></tr></table>",
+        "bbox": [
+            86,
+            309,
+            883,
+            494
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/537bdfeb6899fe709f7f8ffa939c9fab15ea18a19a3223a703446f7d757f11ed.jpg",
+        "table_caption": [
+            "Table 5. Test accuracy of each approach using ResNet-34 on CIFAR-10N (Worst)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Cross-Entropy [38]</td><td>Forward [35]</td><td>Co-teaching [14]</td><td>JoCoR [46]</td><td>Me-Momentum [2]</td><td>Late Stopping(Ours)</td></tr><tr><td>77.69±1.55%</td><td>79.79±0.46%</td><td>83.83±0.13%</td><td>83.37±0.30%</td><td>84.71±0.37%</td><td>85.24±0.38%</td></tr></table>",
+        "bbox": [
+            86,
+            518,
+            883,
+            561
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "number of CHEs in the training set can achieve better generalization performance in this task. For instance, DMI and Co-teaching show a much larger performance gap with our proposed method on CIFAR-100 compared with CIFAR-10. Furthermore, Table 2 validates our method's superior capability to retain clean examples, resulting in a larger performance gap with Me-Momentum on CIFAR-100.",
+        "bbox": [
+            75,
+            585,
+            468,
+            691
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Real-world dataset. To validate the efficacy of our proposed method on real-world datasets, we assess our method using the CIFAR-10N dataset in Table 4, a recognized benchmark in learning with noisy labels tasks with both ground-truth labels and human-annotated real-world noisy labels. We utilized the most challenging real-world noisy labels from CIFAR-10N, i.e., the \"Worst\" setting. In this setting, for each image, if there are any incorrectly labeled examples, the given label is randomly selected from human-annotated false labels. Each trial is repeated five times and the mean value and standard deviation are recorded. As shown in Table 4, our method achieves the best performance compared with other methods.",
+        "bbox": [
+            75,
+            703,
+            470,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3. Further analysis",
+        "text_level": 1,
+        "bbox": [
+            500,
+            585,
+            663,
+            602
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Here, we conduct further analysis of our proposed method with the same experimental settings as those used in Section 4.2 unless stated otherwise.",
+        "bbox": [
+            498,
+            609,
+            890,
+            655
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Limitations. While our method achieves good robustness by simply removing the lately learned examples, it is not immune to errors. Some clean examples might be incorrectly removed from the training set, i.e., falsely removed examples, and some mislabeled examples are incorrectly retained in the training set, i.e., falsely retained examples. We visualize such typical examples in Figure 3 and provide the following empirical analysis.",
+        "bbox": [
+            496,
+            657,
+            890,
+            777
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Falsely retained examples are mislabeled with the wrong label  $L_{w}$ , however, such examples often possess patterns similar to the dominant subclasses in the  $L_{w}$  class, making it difficult to identify them as high probability mislabeled examples by the FkL criterion. As shown in Table 6, based on the FkL criterion and the loss criterion, the \"hardness\" of these examples decreases after being mislabeled. In particular, we named the examples with a decrease in",
+        "bbox": [
+            496,
+            779,
+            892,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "16085",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/bb131935bcded0d1eb98a2305e21cd862c8533fcdcbe93795ec04ac5c2e49096.jpg",
+        "image_caption": [
+            "Figure 3. The labels in the figure are ground-truth labels for each example. Falsely Removed (1-3) are typical examples that are incorrectly removed in the first iteration of Late Stopping; Falsely Retained (1-3) are typical examples that are incorrectly marked by FkL in the last iteration of Late Stopping. In the training set, Falsely Retained (1) is incorrectly labeled as \"deer\", Falsely Retained (2) is incorrectly labeled as \"truck\", and Falsely Retained (3) is incorrectly labeled as \"frog\"."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            104,
+            99,
+            272,
+            250
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/34340c5b7fdfaff1f2a036501159a9ecad1650462804b4a3d0b6478abda668ce.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            277,
+            99,
+            356,
+            250
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/d1fe7ce62347848503949b305d2a414cfc467b064263998ee36fac42839e70b7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            359,
+            99,
+            437,
+            250
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/4f12463c5d7bf1bee38740de148e269351f4e68317fc1f4208326fc69a0e7a79.jpg",
+        "table_caption": [
+            "Table 6. The comparison of the average ranking of falsely retained examples using the loss criterion and the FkL criterion before and after fixing the noisy labels (CIFAR-10, Before: Sym.  $40\\%$  noise)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Label</td><td>Criterion</td><td>Avg. Ranking</td></tr><tr><td rowspan=\"2\">Before fixing (Given label)</td><td>FkL</td><td>22340.66</td></tr><tr><td>loss</td><td>23673.50</td></tr><tr><td rowspan=\"2\">After fixing (Ground-truth label)</td><td>FkL</td><td>28452.08 (+27.36%)</td></tr><tr><td>loss</td><td>29476.46 (+24.51%)</td></tr></table>",
+        "bbox": [
+            84,
+            434,
+            460,
+            520
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/1775aba9b41305aef3d313a10049072f7e00abf1d76be5af1255e13dfe05a671.jpg",
+        "table_caption": [
+            "Table 7. Comparison of the number of clean examples in datasets before and after applying our method as a pre-processing approach to decrease the noise level of a  $40\\%$  noisy training set to  $20\\%$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Training set</td><td>Before (40% noise)</td><td>After (20% noise)</td></tr><tr><td>CIFAR-10 (Sym.)</td><td>27228</td><td>27139 (-0.33%)</td></tr><tr><td>CIFAR-10 (Ins.)</td><td>27094</td><td>26710 (-1.42%)</td></tr><tr><td>CIFAR-100 (Sym.)</td><td>27341</td><td>27037 (-1.11%)</td></tr><tr><td>CIFAR-100 (Ins.)</td><td>27249</td><td>25079 (-7.96%)</td></tr><tr><td>CIFAR-10N (Worst)</td><td>29896</td><td>29613 (-0.95%)</td></tr></table>",
+        "bbox": [
+            84,
+            578,
+            460,
+            672
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "training losses after being mislabeled as \"hard mislabeled examples\". Our experimental results indicate that hard mislabeled examples pose a challenge for all sample selection-based LNL methods. More details about Table 6 are provided in Appendix B1.",
+        "bbox": [
+            75,
+            702,
+            468,
+            777
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Falsely removed examples refer to the examples that are too hard to remove in practice. To limit the computational resources consumed by our proposed method, we set the removal rate (Algorithm 1, iteration rate  $m\\%$ ) to a large value. Thus, our proposed method cannot completely distinguish between rare patterns and mislabeled examples, which can result in a small number of examples from rare subclasses being mistakenly removed from the training set.",
+        "bbox": [
+            75,
+            779,
+            470,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/2efc3115f78fce9d87ca5a0555f44dc8464716ac4ec32c955a3abca30183633d.jpg",
+        "table_caption": [
+            "Table 8. Comparison of the total training hours on CIFAR-100."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Co-teaching [14]</td><td>Me-Momentum [2]</td><td>Late Stopping (Ours)</td></tr><tr><td>3.3h</td><td>7.1h</td><td>9.5h</td></tr></table>",
+        "bbox": [
+            508,
+            101,
+            883,
+            140
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/70c7769ec338a69881711bf58e86cba66bebdc465732d24fead3c802151bc71f.jpg",
+        "image_caption": [
+            "Figure 4. Illustrates the performance of classifiers with the change in iteration rate  $m$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            550,
+            157,
+            834,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Versatility. We are focusing on demonstrating the concept, i.e., using Late Stopping to retain clean hard examples in the training set, but not on boosting the classification performance. However, to cope with even more complex noisy scenarios, our method works well in combination with other advanced techniques. In particular, our approach has excellent performance as a pre-processing approach for reducing the noise rate of the noisy training set. As shown in Table 7, we evaluate Late Stopping as a pre-processing approach to decrease the noise level of a  $40\\%$  noisy training set to  $20\\%$ . Our experimental results indicate that our proposed method performs exceptionally well in reducing label noise. It can substantially reduce the noise level of the training set with minimal loss of clean examples.",
+        "bbox": [
+            496,
+            344,
+            890,
+            556
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Training time. Our proposed method involves training the classifier on a larger training set for a longer period, which may not be advantageous in terms of efficiency. Despite this, as the noise rate decreases, the size of the training set decreases, and the speed of obtaining FkL-examples is accelerated (see Figure 2). As shown in Table 8, we compare the training time with typical baseline methods on CIFAR-100 (Sym.  $40\\%$ ). Our experimental results indicate that the training time of our proposed method is in a comparable range with other approaches.",
+        "bbox": [
+            496,
+            561,
+            890,
+            713
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Sensitivity of the iteration rate. The iteration rate  $m$  (see Algorithm 1) determines the iteration number for Late Stopping and the number of examples removed in each iteration. To investigate the sensitivity of  $m$ , we conducted experiments on CIFAR10 with  $40\\%$  Sym. noise and Ins. noise by varying  $m$  in the range  $\\{10,5,4,3,2\\}$ . Figure 4 shows the performance of classifiers, which gradually improves as  $m$  decreases. This observation aligns with the findings presented in Table 1, where a decrease in the range of selected high-probability unlabeled examples leads to higher accuracy in selecting such examples, ultimately benefiting the classifier's performance.",
+        "bbox": [
+            496,
+            719,
+            890,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "16086",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            76,
+            89,
+            194,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we focused on the challenge of the inability of existing sample selection methods to effectively select clean hard examples. To address this challenge, we proposed a novel method called Late Stopping. In contrast to traditional early-stopping strategies, our method is rooted in a prolonged training process that distinguishes between mislabeled and clean examples by pinpointing the number of training epochs required for each example to be consistently classified to its given label for the first time. We coin this new sample selection criterion as First-time  $k$ -epoch Learning. Experimental results on synthetic and real-world datasets demonstrate that our proposed method is both straightforward and effective in handling learning with noisy labels. In future work, we plan to enhance the FkL criterion to more accurately capture the intrinsic robust learning ability of DNNs, which could potentially bolster the effectiveness of our Late Stopping framework, especially under more intricate noise conditions.",
+        "bbox": [
+            75,
+            114,
+            472,
+            387
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            76,
+            400,
+            236,
+            417
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Lei Feng was supported by the National Natural Science Foundation of China (Grant No. 62106028), Chongqing Overseas Chinese Entrepreneurship and Innovation Support Program, CAAI-Huawei MindSpore Open Fund, and Chongqing Artificial Intelligence Innovation Center. Tongliang Liu was partially supported by Australian Research Council Projects IC-190100031, LP-220100527, DP-220102121, and FT-220100318. We would like to thank anonymous reviewers for their constructive feedback that improved our paper.",
+        "bbox": [
+            75,
+            426,
+            472,
+            580
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            76,
+            590,
+            174,
+            607
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In ICML, 2017.",
+            "[2] Yingbin Bai and Tongliang Liu. Me-momentum: Extracting hard confident examples from noisily labeled data. In ICCV, 2021.",
+            "[3] Yingbin Bai, Erkun Yang, Bo Han, Yanhua Yang, Jiatong Li, Yinian Mao, Gang Niu, and Tongliang Liu. Understanding and improving early stopping for learning with noisy labels. NeurIPS, 2021.",
+            "[4] Robert Baldock, Hartmut Maennel, and Behnam Neyshabur. Deep learning through the lens of example difficulty. NeurIPS, 2021.",
+            "[5] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ICML, 2009.",
+            "[6] Kaidi Cao, Yining Chen, Junwei Lu, Nikos Arechiga, Adrien Gaidon, and Tengyu Ma. Heteroskedastic and imbalanced deep learning with adaptive regularization. arXiv, 2020."
+        ],
+        "bbox": [
+            86,
+            616,
+            468,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[7] Nicholas Carlini, Ulfar Erlingsson, and Nicolas Papernot. Distribution density, tails, and outliers in machine learning: Metrics and applications. arXiv, 2019.",
+            "[8] Pengfei Chen, Ben Ben Liao, Guangyong Chen, and Shengyu Zhang. Understanding and utilizing deep neural networks trained with noisy labels. In ICML, 2019.",
+            "[9] Vitaly Feldman. Does learning require memorization? a short tale about a long tail. In STOC, 2020.",
+            "[10] Vitaly Feldman and Chiyuan Zhang. What neural networks memorize and why: Discovering the long tail via influence estimation. NeurIPS, 2020.",
+            "[11] Guy Hacohen and Daphna Weinshall. On the power of curriculum learning in training deep networks. In ICML, 2019.",
+            "[12] Bo Han, Gang Niu, Xingrui Yu, Quanming Yao, Miao Xu, Ivor Tsang, and Masashi Sugiyama. *Siga: Forgetting may make learning with noisy labels more robust.* In ICML, 2020.",
+            "[13] Bo Han, Quanming Yao, Tongliang Liu, Gang Niu, Ivor W Tsang, James T Kwok, and Masashi Sugiyama. A survey of label-noise representation learning: Past, present and future. arXiv, 2020.",
+            "[14] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Coteaching: Robust training of deep neural networks with extremely noisy labels. NeurIPS, 2018.",
+            "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016.",
+            "[16] Sara Hooker, Aaron Courville, Gregory Clark, Yann Dauphin, and Andrea Frome. What do compressed deep neural networks forget? arXiv, 2019.",
+            "[17] Wei Hu, Zhiyuan Li, and Dingli Yu. Simple and effective regularization methods for training on noisily labeled data with generalization guarantee. arXiv, 2019.",
+            "[18] Jinchi Huang, Lie Qu, Rongfei Jia, and Binqiang Zhao. O2unet: A simple noisy label detection approach for deep neural networks. In ICCV, 2019.",
+            "[19] Zhuo Huang, Xiaobo Xia, Li Shen, Bo Han, Mingming Gong, Chen Gong, and Tongliang Liu. Harnessing out-of-distribution examples via augmenting content and style. arXiv, 2022.",
+            "[20] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, 2018.",
+            "[21] Ziheng Jiang, Chiyuan Zhang, Kunal Talwar, and Michael C Mozer. Characterizing structural regularities of labeled data in overparameterized models. In ICML, 2021.",
+            "[22] Nazmul Karim, Mamshad Nayeem Rizve, Nazanin Rahnavard, Ajmal Mian, and Mubarak Shah. Unicon: Combating label noise through uniform selection and contrastive learning. In CVPR, 2022.",
+            "[23] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009.",
+            "[24] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv, 2020."
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "16087",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[25] Mingchen Li, Mahdi Soltanolkotabi, and Samet Oymak. Gradient descent with early stopping is provably robust to label noise for overparameterized neural networks. In AISTATS, 2020.",
+            "[26] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. NeurIPS, 2020.",
+            "[27] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015.",
+            "[28] Yang Liu and Hongyi Guo. Peer loss functions: Learning from noisy labels without knowing noise rates. In ICML, 2020.",
+            "[29] Bingfeng Luo, Yansong Feng, Zheng Wang, Zhanxing Zhu, Songfang Huang, Rui Yan, and Dongyan Zhao. Learning with noise: Enhance distantly supervised relation extraction with dynamic transition matrix. arXiv, 2017.",
+            "[30] Yueming Lyu and Ivor W Tsang. Curriculum loss: Robust learning and generalization against label corruption. arXiv, 2019.",
+            "[31] Pratyush Maini, Saurabh Garg, Zachary Chase Lipton, and J Zico Kolter. Characterizing datapoints via second-split forgetting. In ICML 2022 Workshop on Spurious Correlations, Invariance and Stability, 2022.",
+            "[32] Karttikeya Mangalam and Vinay Uday Prabhu. Do deep neural networks learn shallow learnable examples first? In ICML 2019 Workshop on Identifying and Understanding Deep Learning Phenomena, 2019.",
+            "[33] Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar. Foundations of machine learning. MIT press, 2018.",
+            "[34] Duc Tam Nguyen, Chaithanya Kumar Mummadi, Thi Phuong Nhung Ngo, Thi Hoai Phuong Nguyen, Laura Beggel, and Thomas Brox. Self: Learning to filter noisy labels with self-ensembling. arXiv, 2019.",
+            "[35] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In CVPR, 2017.",
+            "[36] Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. Identifying mislabeled data using the area under the margin ranking. NeurIPS, 2020.",
+            "[37] David Rolnick, Andreas Veit, Serge Belongie, and Nir Shavit. Deep learning is robust to massive label noise. arXiv, 2017.",
+            "[38] Reuven Rubinstein. The cross-entropy method for combinatorial and continuous optimization. Methodology and computing in applied probability, 1999.",
+            "[39] Yanyao Shen and Sujay Sanghavi. Learning with bad training data via iterative trimmed loss minimization. In ICML, 2019.",
+            "[40] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In ICML, 2019.",
+            "[41] Hwanjun Song, Minseok Kim, Dongmin Park, Yooju Shin, and Jae-Gil Lee. Learning from noisy labels with deep neural networks: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022."
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[42] Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework for learning with noisy labels. In CVPR, 2018.",
+            "[43] Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J Gordon. An empirical study of example forgetting during deep neural network learning. In ICLR, 2018.",
+            "[44] Brendan Van Rooyen, Aditya Menon, and Robert C Williamson. Learning with symmetric label noise: The importance of being unhinged. NeurIPS, 2015.",
+            "[45] Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative learning with open-set noisy labels. In CVPR, 2018.",
+            "[46] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In CVPR, 2020.",
+            "[47] Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. Learning with noisy labels revisited: A study using real-world human annotations. arXiv, 2021.",
+            "[48] Qi Wei, Haoliang Sun, Xiankai Lu, and Yilong Yin. Self-filtering: A noise-aware sample selection for label noise with confidence penalization. In ECCV, 2022.",
+            "[49] Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris Metaxas, and Chao Chen. A topological filter for learning with label noise. NeurIPS, 2020.",
+            "[50] Zhi-Fan Wu, Tong Wei, Jianwen Jiang, Chaojie Mao, Mingqian Tang, and Yu-Feng Li. Ngc: A unified framework for learning with open-world noisy data. In ICCV, 2021.",
+            "[51] Xiaobo Xia, Tongliang Liu, Bo Han, Chen Gong, Nannan Wang, Zongyuan Ge, and Yi Chang. Robust early-learning: Hinding the memorization of noisy labels. In ICLR, 2020.",
+            "[52] Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Jun Yu, Gang Niu, and Masashi Sugiyama. Sample selection with uncertainty of losses for learning with noisy labels. arXiv, 2021.",
+            "[53] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. NeurIPS, 2020.",
+            "[54] Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang. L_dmi: A novel information-theoretic loss function for training deep nets robust to label noise. NeurIPS, 2019.",
+            "[55] Xiangli Yang, Zixing Song, Irwin King, and Zenglin Xu. A survey on deep semi-supervised learning. IEEE Transactions on Knowledge and Data Engineering, 2022.",
+            "[56] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 2021.",
+            "[57] Tianyi Zhou, Shengjie Wang, and Jeff Bilmes. Robust curriculum learning: from clean label detection to noisy label self-correction. In ICLR, 2021."
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            843
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "16088",
+        "bbox": [
+            480,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 9
+    }
+]