Add Batch 035e1327-0f4a-4357-b490-1e570ad95485 data

.gitattributes

@@ -3833,3 +3833,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/Self-supervised[[:space:]]Learning[[:space:]]to[[:space:]]Bring[[:space:]]Dual[[:space:]]Reversed[[:space:]]Rolling[[:space:]]Shutter[[:space:]]Images[[:space:]]Alive/0f7c9f43-96dc-4fd0-a054-e650a421a3a0_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Self-supervised[[:space:]]Monocular[[:space:]]Depth[[:space:]]Estimation_[[:space:]]Let's[[:space:]]Talk[[:space:]]About[[:space:]]The[[:space:]]Weather/7432c306-22bf-4ee4-b730-8425345482aa_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Self-supervised[[:space:]]Monocular[[:space:]]Underwater[[:space:]]Depth[[:space:]]Recovery,[[:space:]]Image[[:space:]]Restoration,[[:space:]]and[[:space:]]a[[:space:]]Real-sea[[:space:]]Video[[:space:]]Dataset/e11d4111-32f8-407a-a327-9e6fbe49fac5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Self-supervised[[:space:]]Pre-training[[:space:]]for[[:space:]]Mirror[[:space:]]Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SemARFlow_[[:space:]]Injecting[[:space:]]Semantics[[:space:]]into[[:space:]]Unsupervised[[:space:]]Optical[[:space:]]Flow[[:space:]]Estimation[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/fa7524e1-a76d-4844-9361-43c8b8046bf9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semantic[[:space:]]Attention[[:space:]]Flow[[:space:]]Fields[[:space:]]for[[:space:]]Monocular[[:space:]]Dynamic[[:space:]]Scene[[:space:]]Decomposition/68ea1ef8-9cb5-44da-894d-92e02632af66_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semantic[[:space:]]Information[[:space:]]in[[:space:]]Contrastive[[:space:]]Learning/c3144656-62d8-4975-97d6-ef4630f410df_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semantic-Aware[[:space:]]Dynamic[[:space:]]Parameter[[:space:]]for[[:space:]]Video[[:space:]]Inpainting[[:space:]]Transformer/706f36ef-c7c5-4ea7-8481-e3a3bb79a94c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semantic-Aware[[:space:]]Implicit[[:space:]]Template[[:space:]]Learning[[:space:]]via[[:space:]]Part[[:space:]]Deformation[[:space:]]Consistency/840dd92d-8817-48f0-ac2c-586a3d27151d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semantically[[:space:]]Structured[[:space:]]Image[[:space:]]Compression[[:space:]]via[[:space:]]Irregular[[:space:]]Group-Based[[:space:]]Decoupling/e3cccf23-d23a-4b30-90e1-8c04059b69dd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semantics[[:space:]]Meets[[:space:]]Temporal[[:space:]]Correspondence_[[:space:]]Self-supervised[[:space:]]Object-centric[[:space:]]Learning[[:space:]]in[[:space:]]Videos/0da0877f-f78d-4f97-9a49-55c50b2292de_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semantics-Consistent[[:space:]]Feature[[:space:]]Search[[:space:]]for[[:space:]]Self-Supervised[[:space:]]Visual[[:space:]]Representation[[:space:]]Learning/fd85c784-c77b-49cb-91b7-f3a6f94e50c3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semantify_[[:space:]]Simplifying[[:space:]]the[[:space:]]Control[[:space:]]of[[:space:]]3D[[:space:]]Morphable[[:space:]]Models[[:space:]]Using[[:space:]]CLIP/ff2d3341-5fbc-47bb-9ce8-0e16c61108bc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semi-Supervised[[:space:]]Learning[[:space:]]via[[:space:]]Weight-Aware[[:space:]]Distillation[[:space:]]under[[:space:]]Class[[:space:]]Distribution[[:space:]]Mismatch/ab302b84-1623-4a04-b35b-84d3ca9ad536_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semi-Supervised[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]under[[:space:]]Label[[:space:]]Noise[[:space:]]via[[:space:]]Diverse[[:space:]]Learning[[:space:]]Groups/8e5c15a5-f270-46c8-829c-81448a8fde6f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semi-supervised[[:space:]]Semantics-guided[[:space:]]Adversarial[[:space:]]Training[[:space:]]for[[:space:]]Robust[[:space:]]Trajectory[[:space:]]Prediction/487dfb22-6094-4f31-976c-310213635421_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Semi-supervised[[:space:]]Speech-driven[[:space:]]3D[[:space:]]Facial[[:space:]]Animation[[:space:]]via[[:space:]]Cross-modal[[:space:]]Encoding/ebc67249-8be2-4ca5-975d-54a569184984_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sensitivity-Aware[[:space:]]Visual[[:space:]]Parameter-Efficient[[:space:]]Fine-Tuning/219bbdca-cfe8-433f-a6db-97c4773a1a10_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sentence[[:space:]]Attention[[:space:]]Blocks[[:space:]]for[[:space:]]Answer[[:space:]]Grounding/0622d151-144e-4e04-b267-1cccf0dba507_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sequential[[:space:]]Texts[[:space:]]Driven[[:space:]]Cohesive[[:space:]]Motions[[:space:]]Synthesis[[:space:]]with[[:space:]]Natural[[:space:]]Transitions/2c5b9c34-9105-42fb-b593-c16dc3db7c25_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Set-level[[:space:]]Guidance[[:space:]]Attack_[[:space:]]Boosting[[:space:]]Adversarial[[:space:]]Transferability[[:space:]]of[[:space:]]Vision-Language[[:space:]]Pre-training[[:space:]]Models/4d2215bf-24a1-45b3-bcea-24623e459934_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Shape[[:space:]]Analysis[[:space:]]of[[:space:]]Euclidean[[:space:]]Curves[[:space:]]under[[:space:]]Frenet-Serret[[:space:]]Framework/235fdcf4-691b-437b-8ee8-c7b76395f3ab_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Shape[[:space:]]Anchor[[:space:]]Guided[[:space:]]Holistic[[:space:]]Indoor[[:space:]]Scene[[:space:]]Understanding/4f878220-d6cc-48e2-9d1e-82c1cf350095_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ShapeScaffolder_[[:space:]]Structure-Aware[[:space:]]3D[[:space:]]Shape[[:space:]]Generation[[:space:]]from[[:space:]]Text/3212f743-a520-438c-b4e2-ff41229e0a86_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Shatter[[:space:]]and[[:space:]]Gather_[[:space:]]Learning[[:space:]]Referring[[:space:]]Image[[:space:]]Segmentation[[:space:]]with[[:space:]]Text[[:space:]]Supervision/eba869e5-3a94-47e8-afcc-2f1d3dc46bfb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Shift[[:space:]]from[[:space:]]Texture-bias[[:space:]]to[[:space:]]Shape-bias_[[:space:]]Edge[[:space:]]Deformation-based[[:space:]]Augmentation[[:space:]]for[[:space:]]Robust[[:space:]]Object[[:space:]]Recognition/e071fbaf-3036-4ac7-9bfd-fb6983ccf03f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ShiftNAS_[[:space:]]Improving[[:space:]]One-shot[[:space:]]NAS[[:space:]]via[[:space:]]Probability[[:space:]]Shift/9d0cd6fb-8343-4e85-b19e-ff450eb2e5e4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Shortcut-V2V_[[:space:]]Compression[[:space:]]Framework[[:space:]]for[[:space:]]Video-to-Video[[:space:]]Translation[[:space:]]Based[[:space:]]on[[:space:]]Temporal[[:space:]]Redundancy[[:space:]]Reduction/99da4ac0-6f5c-433f-bc7b-caa55aa672d5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Shrinking[[:space:]]Class[[:space:]]Space[[:space:]]for[[:space:]]Enhanced[[:space:]]Certainty[[:space:]]in[[:space:]]Semi-Supervised[[:space:]]Learning/34917a9e-765e-47a5-bd3d-8044ab9de911_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SiLK_[[:space:]]Simple[[:space:]]Learned[[:space:]]Keypoints/545a50e4-a6f6-42d1-849e-c613959c21fd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sigmoid[[:space:]]Loss[[:space:]]for[[:space:]]Language[[:space:]]Image[[:space:]]Pre-Training/85b123ac-a2cf-49ce-9282-93b24961a80b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sign[[:space:]]Language[[:space:]]Translation[[:space:]]with[[:space:]]Iterative[[:space:]]Prototype/f2c1bf6d-f55b-40d7-9372-42bff493c6d3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SimFIR_[[:space:]]A[[:space:]]Simple[[:space:]]Framework[[:space:]]for[[:space:]]Fisheye[[:space:]]Image[[:space:]]Rectification[[:space:]]with[[:space:]]Self-supervised[[:space:]]Representation[[:space:]]Learning/e93fe221-1f98-47dd-bbe6-409ea24f23b2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SimMatchV2_[[:space:]]Semi-Supervised[[:space:]]Learning[[:space:]]with[[:space:]]Graph[[:space:]]Consistency/4de24035-8115-4285-8e43-089c94dd354f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SimNP_[[:space:]]Learning[[:space:]]Self-Similarity[[:space:]]Priors[[:space:]]Between[[:space:]]Neural[[:space:]]Points/54ca6d2a-b8a1-4a4e-a6f5-57eb0fddb4dd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Similarity[[:space:]]Min-Max_[[:space:]]Zero-Shot[[:space:]]Day-Night[[:space:]]Domain[[:space:]]Adaptation/d2d1b76d-0c03-43ab-96b8-c3ce7c667abf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Simoun_[[:space:]]Synergizing[[:space:]]Interactive[[:space:]]Motion-appearance[[:space:]]Understanding[[:space:]]for[[:space:]]Vision-based[[:space:]]Reinforcement[[:space:]]Learning/17dc049b-d1dc-4e17-921e-1317f662d966_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Simple[[:space:]]Baselines[[:space:]]for[[:space:]]Interactive[[:space:]]Video[[:space:]]Retrieval[[:space:]]with[[:space:]]Questions[[:space:]]and[[:space:]]Answers/149533a1-79f1-428b-885a-6c528f5bd0a8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Simple[[:space:]]and[[:space:]]Effective[[:space:]]Out-of-Distribution[[:space:]]Detection[[:space:]]via[[:space:]]Cosine-based[[:space:]]Softmax[[:space:]]Loss/081c8308-e453-453c-8d9f-1e2b9b04b642_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SimpleClick_[[:space:]]Interactive[[:space:]]Image[[:space:]]Segmentation[[:space:]]with[[:space:]]Simple[[:space:]]Vision[[:space:]]Transformers/2b87aeaa-0b31-4e94-84d5-d9bfaea7e328_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Simulating[[:space:]]Fluids[[:space:]]in[[:space:]]Real-World[[:space:]]Still[[:space:]]Images/d51a02b7-7fc4-4f85-a456-fd173559877f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Single[[:space:]]Depth-image[[:space:]]3D[[:space:]]Reflection[[:space:]]Symmetry[[:space:]]and[[:space:]]Shape[[:space:]]Prediction/fcff96a5-45a1-4f7e-ab9f-0c309af710cd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Single[[:space:]]Image[[:space:]]Deblurring[[:space:]]with[[:space:]]Row-dependent[[:space:]]Blur[[:space:]]Magnitude/05f13a7b-32f2-44f9-b585-5bd4c1beae6a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Single[[:space:]]Image[[:space:]]Defocus[[:space:]]Deblurring[[:space:]]via[[:space:]]Implicit[[:space:]]Neural[[:space:]]Inverse[[:space:]]Kernels/bee3d863-7fc6-4e0b-a339-87a4afbe8bb4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Single[[:space:]]Image[[:space:]]Reflection[[:space:]]Separation[[:space:]]via[[:space:]]Component[[:space:]]Synergy/6169238f-a2a5-4863-b901-a9125cd1dec8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Single-Stage[[:space:]]Diffusion[[:space:]]NeRF_[[:space:]]A[[:space:]]Unified[[:space:]]Approach[[:space:]]to[[:space:]]3D[[:space:]]Generation[[:space:]]and[[:space:]]Reconstruction/06f68dbb-5c23-4704-ae55-c667675f6d9f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Size[[:space:]]Does[[:space:]]Matter_[[:space:]]Size-aware[[:space:]]Virtual[[:space:]]Try-on[[:space:]]via[[:space:]]Clothing-oriented[[:space:]]Transformation[[:space:]]Try-on[[:space:]]Network/2e5e4941-f171-4e8c-be33-35affe248d0e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SkeleTR_[[:space:]]Towards[[:space:]]Skeleton-based[[:space:]]Action[[:space:]]Recognition[[:space:]]in[[:space:]]the[[:space:]]Wild/27665a84-18aa-44bc-b25b-e6429d7734c6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SkeletonMAE_[[:space:]]Graph-based[[:space:]]Masked[[:space:]]Autoencoder[[:space:]]for[[:space:]]Skeleton[[:space:]]Sequence[[:space:]]Pre-training/dcc19c04-02ae-4f8d-8f44-a86cdd91e8a4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sketch[[:space:]]and[[:space:]]Text[[:space:]]Guided[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]Colored[[:space:]]Point[[:space:]]Cloud[[:space:]]Generation/6a47e5bb-0a8f-4500-ae70-d8f17ed82ef9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Skill[[:space:]]Transformer_[[:space:]]A[[:space:]]Monolithic[[:space:]]Policy[[:space:]]for[[:space:]]Mobile[[:space:]]Manipulation/e273608a-9944-4287-91dc-881cf0220c1e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Skip-Plan_[[:space:]]Procedure[[:space:]]Planning[[:space:]]in[[:space:]]Instructional[[:space:]]Videos[[:space:]]via[[:space:]]Condensed[[:space:]]Action[[:space:]]Space[[:space:]]Learning/c1a3a0ec-3885-4f2f-93dd-62793c50f314_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SlaBins_[[:space:]]Fisheye[[:space:]]Depth[[:space:]]Estimation[[:space:]]using[[:space:]]Slanted[[:space:]]Bins[[:space:]]on[[:space:]]Road[[:space:]]Environments/2f72858d-bf2e-45dd-b589-9c5d6578e810_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Small[[:space:]]Object[[:space:]]Detection[[:space:]]via[[:space:]]Coarse-to-fine[[:space:]]Proposal[[:space:]]Generation[[:space:]]and[[:space:]]Imitation[[:space:]]Learning/231f8e4e-8ec6-40cc-a455-514ae44e6fa1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Smoothness[[:space:]]Similarity[[:space:]]Regularization[[:space:]]for[[:space:]]Few-Shot[[:space:]]GAN[[:space:]]Adaptation/e521c479-803a-4c4d-bf48-e946e8872ca9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Snow[[:space:]]Removal[[:space:]]in[[:space:]]Video_[[:space:]]A[[:space:]]New[[:space:]]Dataset[[:space:]]and[[:space:]]A[[:space:]]Novel[[:space:]]Method/5ae50b63-d59f-4dc4-bf7f-4d6ac34210b4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SoDaCam_[[:space:]]Software-defined[[:space:]]Cameras[[:space:]]via[[:space:]]Single-Photon[[:space:]]Imaging/ce1d43a9-f5c5-4d36-93a7-015029e926c0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Social[[:space:]]Diffusion_[[:space:]]Long-term[[:space:]]Multiple[[:space:]]Human[[:space:]]Motion[[:space:]]Anticipation/e840bc61-69a6-4f9d-9e0d-f25143903533_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sound[[:space:]]Localization[[:space:]]from[[:space:]]Motion_[[:space:]]Jointly[[:space:]]Learning[[:space:]]Sound[[:space:]]Direction[[:space:]]and[[:space:]]Camera[[:space:]]Rotation/ebfa790a-8a91-4c3c-9e52-be53fe87e14f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sound[[:space:]]Source[[:space:]]Localization[[:space:]]is[[:space:]]All[[:space:]]about[[:space:]]Cross-Modal[[:space:]]Alignment/7b57a83c-7e70-4280-a6ee-f0f550384305_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Source-free[[:space:]]Depth[[:space:]]for[[:space:]]Object[[:space:]]Pop-out/1ac228ed-431b-40e3-b444-e5a574a99abb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Source-free[[:space:]]Domain[[:space:]]Adaptive[[:space:]]Human[[:space:]]Pose[[:space:]]Estimation/6db7e049-3b2d-4942-adce-fb8455dd18b2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Space[[:space:]]Engage_[[:space:]]Collaborative[[:space:]]Space[[:space:]]Supervision[[:space:]]for[[:space:]]Contrastive-Based[[:space:]]Semi-Supervised[[:space:]]Semantic[[:space:]]Segmentation/5953e602-104c-4dc8-b226-b90796937563_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Space-time[[:space:]]Prompting[[:space:]]for[[:space:]]Video[[:space:]]Class-incremental[[:space:]]Learning/23f1a364-1b5a-487e-9a38-2d1ef0112aa7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/SpaceEvo_[[:space:]]Hardware-Friendly[[:space:]]Search[[:space:]]Space[[:space:]]Design[[:space:]]for[[:space:]]Efficient[[:space:]]INT8[[:space:]]Inference/7d8d9b36-b64d-45d0-8aa5-d17e8abbee45_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Spacetime[[:space:]]Surface[[:space:]]Regularization[[:space:]]for[[:space:]]Neural[[:space:]]Dynamic[[:space:]]Scene[[:space:]]Reconstruction/11a08db4-fb69-460a-a705-16e05f3f48ac_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Sparse[[:space:]]Instance[[:space:]]Conditioned[[:space:]]Multimodal[[:space:]]Trajectory[[:space:]]Prediction/5a87e2e4-dac0-41f3-9966-85985369c6c6_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/Self-supervised Pre-training for Mirror Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_content_list.json

@@ -0,0 +1,1586 @@
+[
+    {
+        "type": "text",
+        "text": "Self-supervised Pre-training for Mirror Detection",
+        "text_level": 1,
+        "bbox": [
+            233,
+            130,
+            736,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jiaying Lin* Rynson W.H. Lau* City University of Hong Kong",
+        "bbox": [
+            341,
+            181,
+            625,
+            217
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{jiayinlin5, rynson.lau}@cityu.edu.hk",
+        "bbox": [
+            323,
+            219,
+            647,
+            233
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            268,
+            313,
+            284
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Existing mirror detection methods require supervised ImageNet pre-training to obtain good general-purpose image features. However, supervised ImageNet pre-training focuses on category-level discrimination and may not be suitable for downstream tasks like mirror detection, due to the overfitting upstream tasks (e.g., supervised image classification). We observe that mirror reflection is crucial to how people perceive the presence of mirrors, and such mid-level features can be better transferred from self-supervised pretrained models. Inspired by this observation, in this paper we aim to improve mirror detection methods by proposing a new self-supervised learning (SSL) pre-training framework for modeling the representation of mirror reflection progressively in the pre-training process. Our framework consists of three pre-training stages at different levels: 1) an image-level pre-training stage to globally incorporate mirror reflection features into the pre-trained model; 2) a patch-level pre-training stage to spatially simulate and learn local mirror reflection from image patches; and 3) a pixel-level pre-training stage to pixel-wisely capture mirror reflection via reconstructing corrupted mirror images based on the relationship between the inside and outside of mirrors. Extensive experiments show that our SSL pre-training framework significantly outperforms previous state-of-the-art CNN-based SSL pre-training frameworks and even outperforms supervised ImageNet pre-training when transferred to the mirror detection task. Code and models are available at https://jiaying.link/iccv2023-sslmirror/",
+        "bbox": [
+            73,
+            301,
+            473,
+            724
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            756,
+            209,
+            771
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Mirrors are prevalent in our daily lives. As their appearances are largely determined by their surroundings, they generally lack a consistent appearance, making it difficult to separate them from their surroundings. This may affect many computer vision tasks such as object detection [2], vision-language navigation [1] and depth estimation [29].",
+        "bbox": [
+            75,
+            782,
+            468,
+            875
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/2053fb7e33566aa3d281812f2faf69bdad556d6356cac893b22a3c6a21bc08bc.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            267,
+            583,
+            340
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/97624ca6d1125c6f4390b003b014011a351f9c1950c7ccbc8c0f5ed0ce450959.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            586,
+            267,
+            658,
+            340
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/a3a1688cab67f58dac910e5e2accc57c4180feddaabc8e18cc4be9edbfbf4b69.jpg",
+        "image_caption": [
+            "(b) MirrorNet on ImageNet"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            586,
+            342,
+            658,
+            415
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/9c79b20acd129e22e321256d41d612df587f83459c20148214c01a2b2a65136b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            661,
+            267,
+            735,
+            340
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/21f231f0b30b96f0819c87b4f82bdc5afed12d33bcd81d816967afb479a37ca7.jpg",
+        "image_caption": [
+            "(c)VCNet on ImageNet"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            661,
+            342,
+            735,
+            415
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/2407085591717f6000629a100ebbac3458bd38d6891f0485e3353e069cd334da.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            736,
+            267,
+            810,
+            340
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/1d85123e767780ec94c3702dff904630533d38df76179b68b715bd5920b81888.jpg",
+        "image_caption": [
+            "Figure 1. State-of-the-art mirror detection methods [37, 30] are based on costly supervised ImageNet pre-training. They may fail even in obvious cases, e.g., (a) when the mirror clearly reflects a real object outside of the mirror. (b) and (c) are MirrorNet [37] and VCNet [30], respectively, pre-trained on ImageNet with full supervision. (c) is MirrorNet with our proposed SSL framework and without supervised ImageNet pre-training. Our SSL scheme leverages the mirror reflection cue and avoids feature redundancy in the pre-training stage. It outperforms those pre-trained on ImageNet with full supervision (i.e., (b) and (c))."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            736,
+            342,
+            810,
+            415
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/172cb65753d1987a963d6fe5c399475d5ab9a638956f1ba00e1256d1d7e58382.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            813,
+            267,
+            887,
+            340
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/db1d2d05c41cec15303148981a731908325a4eedbf1269610181f34c7e074c40.jpg",
+        "image_caption": [
+            "1) Ours (MirrorNet on our SSL)",
+            "(e) GT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            813,
+            342,
+            887,
+            415
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "It is therefore crucial to design high-performance mirror detection methods to facilitate computer vision applications.",
+        "bbox": [
+            498,
+            613,
+            890,
+            643
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recently, a few methods [37, 23, 30] have been proposed for mirror detection, but they all require to be initialized with supervised ImageNet [8] pre-trained weights and then fine-tuned on mirror detection datasets. While such a transfer learning approach is common in general computer vision tasks (e.g., object detection [24] and semantic segmentation [7]) to utilize category-oriented features from large-scale image datasets, we challenge if this strategy is necessary for the mirror detection task for two reasons. First, the supervised ImageNet pre-training process involves high labelling costs, as it is very time-consuming and labor-intensive to label such a large-scale image dataset ( $\\sim$ 1.3M images). Second, unlike general vision tasks, mirror detection does not require category-level object understanding. Instead, it requires an understanding of the relationship between inside and outside of mirrors (e.g., contrast [37], similarity [23], and visual chirality [30]). Thus,",
+        "bbox": [
+            496,
+            643,
+            892,
+            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": "*Corresponding authors.",
+        "bbox": [
+            94,
+            886,
+            227,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "12227",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "while pre-training on ImageNet requires expensive labels, it causes feature redundancy for the mirror detection task.",
+        "bbox": [
+            76,
+            90,
+            468,
+            119
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Figure 1 shows that existing models pre-trained on ImageNet tend to over-detect the mirror regions. Although the two SOTA methods [37, 30] can correctly locate the mirror regions, they are unable to distinguish between the mirrored objects inside the mirrors and the real objects outside, even though these methods contain well-designed modules to learn to address this problem via the fine-tuning stage. This motivates us to investigate whether the current pre-training approach may affect the final detection performance, and whether we may incorporate some intrinsic mirror properties in the pre-training process to assist mirror detection.",
+        "bbox": [
+            75,
+            121,
+            467,
+            286
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "An interesting observation is that humans do not need to be \"trained\" to recognize mirrors. Instead, we learn to recognize them implicitly from young. This aligns with the key idea of self-supervised learning (SSL). Works from neuroscience [31, 20] have shown that humans use mid-level visual cortex to recognize mirror reflection. According to [44], mid-level features are hard to be transferred from supervised pre-training on ImageNet, and SSL is suitable for learning mid-level representations, which inspires us to develop an SSL pre-training framework for modeling mirror reflection.",
+        "bbox": [
+            75,
+            287,
+            467,
+            450
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a new SSL pre-training framework for mirror detection, which explicitly considers mirror reflection during the pre-training process. Our framework does not require human-annotated labels from large-scale image datasets. It consists of three stages to progressively pre-train the backbone network from global to local: 1) an image-level pre-training stage to obtain the representation of mirror reflection globally by recognizing the geometric transformation applied to the image; 2) a patch-level pretraining stage to mimic patch-wise mirror reflection and then learn the spatial correlation between the original object patches and the corresponding mirrored object patches; and 3) a pixel-level pre-training stage to extract the pixel-to-pixel relationship of mirror reflection by image reconstruction; Under such progressive pre-training scheme, our SSL pre-training framework can learn the representation of mirror reflection, and then effectively transfer this knowledge to the subsequent mirror detection process for better detection performances. We conduct comprehensive experiments to demonstrate the effectiveness of our SSL pretraining framework. We show that it can significantly boost the performances of existing mirror detection methods, and outperform other CNN-based SSL pre-training frameworks on the mirror detection task.",
+        "bbox": [
+            75,
+            454,
+            467,
+            813
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To conclude, this paper makes three key contributions:",
+        "bbox": [
+            96,
+            816,
+            455,
+            830
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- To the best of our knowledge, we are the first to investigate how existing SSL pre-training frameworks perform on the mirror detection task, compared with supervised ImageNet pre-training.",
+        "bbox": [
+            94,
+            839,
+            467,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose a new SSL pre-training framework that consists of three stages at different levels to progressively learn the representation of mirror reflection. Compared with the features from a supervised ImageNet pre-trained model, our representation is better due to the reduced gap between the pre-training task and the target downstream task (i.e., mirror detection).",
+            "- Extensive experiments show that our SSL pre-training framework performs the best among all state-of-the-art SSL methods, and even better than models with supervised ImageNet pre-training."
+        ],
+        "bbox": [
+            517,
+            90,
+            890,
+            270
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            285,
+            640,
+            301
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Mirror Detection",
+        "text_level": 1,
+        "bbox": [
+            500,
+            310,
+            666,
+            325
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Mirror detection is an essential task in computer vision, as mirrors are everywhere these days. It is also a challenging task, since mirrors do not possess a consistent appearance, but reflect those of their surroundings. Other vision tasks such as generic object detection and segmentation would tend to detect and segment objects outside as well as inside of the mirror, providing incorrect information on the mirror region, and a depth prediction task may return the depths of the reflected objects inside the mirror region instead of the depth of the mirror.",
+        "bbox": [
+            496,
+            335,
+            890,
+            484
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address the mirror detection problem, Yang et al. [37] propose the first large-scale mirror detection dataset and the first deep-learning model, MirrorNet, for mirror detection by learning the contextual contrast between the inside and outside of mirrors. Lin et al. [23] further propose a new model with similarity and edge learning to progressively detect mirrors. Most recently, Guan et al. [15] employ graph convolutional networks to model contextual associations for mirror detection, while Tan et al. propose VCNet [30] to exploit the visual chirality cue at a feature level.",
+        "bbox": [
+            496,
+            486,
+            890,
+            636
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "However, existing mirror detection methods are all fully-supervised. They also directly adopt the supervised ImageNet [8] pre-training to obtain generic image features. Unlike recent works [19, 18, 22] that propose novel network architectures for mirror detection, we propose in this paper to tackle the mirror detection problem from a learning-based perspective, offering the first self-supervised pre-training framework for mirror detection.",
+        "bbox": [
+            496,
+            638,
+            890,
+            758
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Self-supervised Learning (SSL)",
+        "text_level": 1,
+        "bbox": [
+            500,
+            770,
+            774,
+            786
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "SSL is a popular research topic in computer vision [38, 39], and has attracted a lot of research interests. It aims to learn good feature representations without needing labeled data. Early SSL methods focus on proposing novel pretext tasks, such as relative location prediction [10] and rotation prediction [21]. Recently, most methods are based on the utilization of contrastive learning [16, 4, 14, 3, 32, 5, 41,",
+        "bbox": [
+            496,
+            794,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "12228",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/908689b51dc7553c0ccc26fde83f0bd65eb6d802e9a616f0f113c004d699d59b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            88,
+            459,
+            183
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/855243fd694ad6ad2d7363759b0686b9465af05eab1878a3b52b8f769d8e9a32.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            464,
+            88,
+            890,
+            183
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/d955a6d1249ed61c65caf126c7f5730a39a2581f3573b017e7bde9fc5f13c937.jpg",
+        "image_caption": [
+            "Figure 2. Our SSL pre-training framework. It consists of three pre-training stages at different levels: an image-level pre-training stage, a patch-level pre-training stage, and a pixel-level pre-training stage. It pre-trains the backbone network from global to local (image-level  $\\rightarrow$  patch-level  $\\rightarrow$  pixel-level) progressively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            163,
+            183,
+            834,
+            340
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "40]. Most recently, [44] shows that some SSL methods perform even better than supervised ImageNet pre-training in downstream tasks, such as object detection [11, 24] and semantic segmentation [7, 34], with the help of instance discrimination during SSL pre-training. Recent transformer-based SSL frameworks [36] also involve masked image modeling to obtain better self-supervised representations.",
+        "bbox": [
+            75,
+            412,
+            468,
+            518
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we observe that unlike general vision tasks, mirror detection is not benefited from supervised ImageNet pre-training since it does not require category-level object understanding. Hence, we propose a novel SSL pre-training framework specifically for the mirror detection task. The proposed framework aims to simulate and learn the relationship between mirrors and their surrounding during the pre-training stage.",
+        "bbox": [
+            75,
+            518,
+            470,
+            642
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            76,
+            656,
+            168,
+            672
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Figure 2 shows the overall architecture of the proposed SSL pre-training framework. Our SSL pre-training framework consists of three pre-training stages at different levels: an image-level pre-training, a patch-level pre-training, and a pixel-level pre-training. The pre-training at each level is conducted sequentially and independently. The progressive and sequential training strategy can help the pre-trained model obtain a global-to-local representation of mirrors.",
+        "bbox": [
+            75,
+            681,
+            468,
+            805
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Image-level Pre-training",
+        "text_level": 1,
+        "bbox": [
+            76,
+            816,
+            302,
+            832
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Inspired by the definition of visual chirality [26] and its potential applications [46, 30] in various computer vision problems, we formulate the image-level pre-training as a binary classification problem: Given an input image and its",
+        "bbox": [
+            75,
+            839,
+            470,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "corresponding horizontal flipped image, can the backbone network tell which one is the original image? To achieve this objective, we attach a classification head to predict the probability that the input image is flipped. The classification head consists of a convolution layer followed by a sigmoid activation function to project the output value into  $[0,1]$ . In this pre-training stage, we use binary cross-entropy (BCE) as our loss function  $\\mathcal{L}_f$  to optimize the backbone network. Formally, we have:",
+        "bbox": [
+            496,
+            412,
+            892,
+            549
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {f} = - \\left[ y \\log (\\hat {y}) + (1 - y) \\log (1 - \\hat {y}) \\right], \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            563,
+            556,
+            890,
+            575
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $y$  is the ground truth label and  $\\hat{y}$  is the predicted probability that the input image is horizontally flipped.",
+        "bbox": [
+            496,
+            582,
+            890,
+            613
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Patch-level Pre-training",
+        "text_level": 1,
+        "bbox": [
+            500,
+            619,
+            720,
+            636
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The image-level pre-training pretext task alone cannot bring much help in locating the mirror inside an image since mirror detection is a pixel-level problem. However, directly connecting the image-level and the pixel-level pretext tasks is challenging due to the gap between the scales of these two pretext tasks. Thus, we design a patch-level pre-training stage as a bridge to cover the feature representations of mirrors at different scales. Besides, the image-level pre-training does not benefit relational spatial understanding, which has been proven important in mirror detection [23]. Our patch-level pre-training consists of two sub-tasks to both obtain the patch-wise relationship of mirror reflection and simulate the mirrored region with pseudo labels. We first split the input image into a grid of  $3 \\times 3$  image patches and randomly shuffle them. After that, we randomly apply a flipping operation to one of the image patches, to simulate a mirrored region. The first sub-task",
+        "bbox": [
+            496,
+            643,
+            892,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "12229",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "is to predict the location index from the shuffled grid by the backbone network. We append a location predictor head to the backbone network to predict the location index of input patches. We apply cross-entropy loss as  $\\mathcal{L}_{loc}$  to optimize this sub-task. The second sub-task is to predict the flipped patch from the nine randomly shuffled image patches. Similar to our image-level pre-training, we use BCE loss as our loss function  $\\mathcal{L}_{cls}$  in this sub-task.",
+        "bbox": [
+            75,
+            90,
+            472,
+            212
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {l o c} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} y _ {i} \\log \\left(\\hat {y} _ {i}\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            94,
+            223,
+            468,
+            263
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {c l s} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} y _ {j} \\log \\left(\\hat {y _ {j}}\\right) + (1 - y _ {j}) \\log \\left(1 - \\hat {y _ {j}}\\right), \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            94,
+            268,
+            468,
+            310
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\hat{y}_i$  is the predicted location index, and  $y_i$  is the corresponding ground truth label.  $\\hat{y}_j$  is the predicted probability that the input patch is horizontally flipped and  $y_j$  is the corresponding ground truth probability. Finally, we jointly optimize these two sub-tasks and the final loss function is:",
+        "bbox": [
+            75,
+            321,
+            468,
+            398
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {p a t c h}} = \\mathcal {L} _ {\\text {l o c}} + \\mathcal {L} _ {\\text {c l s}}. \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            196,
+            425,
+            468,
+            444
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Pixel-level Pre-training",
+        "text_level": 1,
+        "bbox": [
+            76,
+            453,
+            292,
+            470
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Extracting the pixel-level relationship of mirror reflection is a challenging task, as studied by previous related work [12] for symmetry detection. It requires point-level GT labels to represent symmetry centers. However, obtaining such labels requires huge human annotation effort and they are not applicable for SSL pre-training. Instead of generating pseudo labels for symmetry detection, our solution is that we formulate this task as an image reconstruction problem on images containing mirrors. An ideal pre-trained network should be able to reconstruct the input image containing mirrors with the pixel-level understanding of mirror reflection. We randomly mask out the input image by cutout augmentation [9] with the ratio of 0.3 and then feed it to the backbone network with a reconstruction head. The details of the reconstruction head we used are listed in Table 1. The objective of our pixel-level pre-training is defined as:",
+        "bbox": [
+            75,
+            477,
+            472,
+            720
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {p i x} = \\| x - \\hat {x} \\| _ {1}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            210,
+            732,
+            468,
+            750
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{L}_{pix}$  is an L1 loss.  $x$  and  $\\hat{x}$  are the input image and the reconstructed image, respectively.",
+        "bbox": [
+            75,
+            761,
+            468,
+            792
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            806,
+            209,
+            823
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1. Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            76,
+            830,
+            284,
+            848
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Dataset and Evaluation Metrics. The image-level and patch-level pre-training experiments are conducted on the training set of ImageNet-100 [28] and MS COCO [25],",
+        "bbox": [
+            75,
+            854,
+            468,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1. The architecture of the reconstruction head used in the pixel-level pre-training. Note that each \"conv-IR\" corresponds to a sequence of convolution layer, InstanceNorm layer and ReLU activation.  $K$ ,  $S$  and  $P$  denote the number of kernels, the number of strides and the padding size, respectively, used in the convolution layer.",
+        "bbox": [
+            496,
+            85,
+            893,
+            170
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/d5f1df4fcdc5de47ce0dcf08320a1def3b98b1b81fdc75f45f072bacb7572ba9.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Layer Details</td></tr><tr><td>3 × 3 conv-IR, K = 1024, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>3 × 3 conv-IR, K = 512, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>3 × 3 conv-IR, K = 256, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>3 × 3 conv-IR, K = 128, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>3 × 3 conv-IR, K = 128, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>1 × 1 conv-IR, K = 3, S = 1, P = 1</td></tr></table>",
+        "bbox": [
+            558,
+            171,
+            834,
+            367
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "which contains  $130\\mathrm{K}$  and  $\\sim 118\\mathrm{k}$  images respectively. In the pixel-level pre-training and fine-tuning stage, our experiments are conducted on MSD [37], the first large-scale dataset for mirror detection. MSD [37] provides 4,018 mirror images, which are divided into 3,063 images for training and 955 images for testing. We adopt two popular metrics namely, F-measure  $(\\mathbf{F}_{\\beta})$  and mean absolute error (MAE) to evaluate the performance of our SSL pre-training framework. F-measure can reflect the overall model performance and is calculated by the weighted harmonic mean of the precision and recall:  $F_{\\beta} = \\frac{(1 + \\beta^{2})\\cdot\\text{precision}\\cdot\\text{recall}}{\\beta^{2}\\cdot\\text{precision} + \\text{recall}}$ , where  $\\beta^2$  is set to 0.3. MAE can measure the average pixel-wise absolute disparity between the predicted mirror maps  $M$  and the binary ground truth  $G$ .",
+        "bbox": [
+            496,
+            393,
+            890,
+            611
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. We use Pytorch to implement our proposed SSL pre-training framework. All experiments are conducted on a GPU server with eight NVIDIA RTX 3090 GPUs. Following existing SSL pre-training frameworks, we adopt ResNet-50 [17] as our backbone network. All pre-trained model weights of existing SSL pretraining frameworks are obtained from the released models of mmselfsup [6] for fair comparisons. For each pretext task in the pre-training stage, all images are resized to  $256 \\times 256$ . The batch size is set to 352. We use AdamW [27] with an initial learning rate of 1e-4 to optimize our SSL pre-training framework and stop pre-training after 20,000 iterations. We only adopt random color jittering for data augmentation. In the fine-tuning stage, we use the same hyperparameters reported in the original papers of the adopted",
+        "bbox": [
+            496,
+            611,
+            892,
+            839
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>It is worth noting that the datasets that we used (about 250K training images) are much smaller than the pre-trained dataset ImageNet (14 million images) used by other baseline SSL methods. We adopted ImageNet-100 to pre-train our network due to the limited computational resources.",
+        "bbox": [
+            500,
+            849,
+            893,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "12230",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/fb1c2b369ed7aed8c0640a1bc2c105e7f93e01f49c8804770f246f9699221c5a.jpg",
+        "image_caption": [
+            "Figure 3. Qualitative Comparison of our proposed method with NPID [33], MocoV2 [4], SwAV [3], Barlow Twins [41], ImageNet pretraining (Supervised). We use the released pre-trained models of these existing SSL methods from mmselfsup [6] for fair comparison."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            84,
+            87,
+            890,
+            592
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "base methods MirrorNet [37] and VCNet [30].",
+        "bbox": [
+            76,
+            657,
+            383,
+            672
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Comparison with State-of-the-Arts",
+        "text_level": 1,
+        "bbox": [
+            76,
+            686,
+            382,
+            702
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We compare against 11 state-of-the-art CNN-based frameworks in self-supervised learning, including Relative Location (Rel. Loc.) [10], Rotation Prediction (Rot. Pred.) [21], NPID [33], ODC [43], Moco V2 [16], SimCLR [4], BYOL [14], SwAV [3], DenseCL [32], SimSiam [5] and Barlow Twins (Bar. Twins) [41]. We select MirrorNet [37] and VCNet [30], which are two representative mirror detection methods, as base methods to conduct experiments in the fine-tuning stage.",
+        "bbox": [
+            75,
+            710,
+            468,
+            847
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Qualitative Evaluation. We further show the performance of our SSL pre-training framework in Figure 3. We compare our framework with four state-of-the-art SSL pre-training",
+        "bbox": [
+            75,
+            854,
+            470,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "frameworks (including the two best-performing frameworks NPID [33] and Moco V2 [16] based on MirrorNet and the two best-performing frameworks SwAV [3] and Barlow Twins [41] based on VCNet, according to their performance in Table 2) and supervised ImageNet pre-training due to the limitation of space. Our framework can correctly detect the mirror regions by utilizing the relationship of mirror reflection caused by mirrors, even though the reflected regions are pure texture (i.e., lack of object keypoints in the second row) and relatively small in the image (e.g., the last three rows). We attribute the superior performance of our framework to the utilization of mirror reflection in the pre-training stage.",
+        "bbox": [
+            496,
+            657,
+            890,
+            839
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Quantitative Evaluation. Table 2 shows the experimentation results comparing the proposed SSL pre-training framework with the state-of-the-arts. Besides, we also re",
+        "bbox": [
+            498,
+            854,
+            892,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "12231",
+        "bbox": [
+            480,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/2f5fd76e72a5d854ad1405ab978ab00692d8bc861f3d00afd8860386de67b80f.jpg",
+        "table_caption": [
+            "Table 2. Quantitative comparison between our SSL pre-training framework and existing CNN-based SSL methods. We also compare it with random initialization (Random) and ImageNet pretraining (Supervised)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td rowspan=\"2\">Venue</td><td colspan=\"2\">MirrorNet</td><td colspan=\"2\">VCNet</td></tr><tr><td>Fβ↑</td><td>MAE↓</td><td>Fβ↑</td><td>MAE↓</td></tr><tr><td>Rel. Loc.</td><td>ICCV &#x27;15</td><td>0.582</td><td>0.254</td><td>0.842</td><td>0.072</td></tr><tr><td>Rot. Pred.</td><td>ICLR &#x27;18</td><td>0.537</td><td>0.212</td><td>0.851</td><td>0.069</td></tr><tr><td>NPID</td><td>CVPR &#x27;18</td><td>0.739</td><td>0.132</td><td>0.832</td><td>0.076</td></tr><tr><td>ODC</td><td>CVPR &#x27;20</td><td>0.558</td><td>0.223</td><td>0.836</td><td>0.077</td></tr><tr><td>Moco V2</td><td>CVPR &#x27;20</td><td>0.707</td><td>0.177</td><td>0.835</td><td>0.072</td></tr><tr><td>SimCLR</td><td>ICML &#x27;20</td><td>0.594</td><td>0.244</td><td>0.837</td><td>0.071</td></tr><tr><td>BYOL</td><td>NeurIPS &#x27;20</td><td>0.704</td><td>0.158</td><td>0.832</td><td>0.077</td></tr><tr><td>SwAV</td><td>NeurIPS &#x27;20</td><td>0.578</td><td>0.224</td><td>0.856</td><td>0.071</td></tr><tr><td>DenseCL</td><td>CVPR &#x27;21</td><td>0.694</td><td>0.240</td><td>0.844</td><td>0.072</td></tr><tr><td>SimSiam</td><td>CVPR &#x27;21</td><td>0.672</td><td>0.198</td><td>0.818</td><td>0.080</td></tr><tr><td>Bar. Twins</td><td>ICML &#x27;21</td><td>0.655</td><td>0.178</td><td>0.857</td><td>0.069</td></tr><tr><td>Random</td><td></td><td>0.586</td><td>0.200</td><td>0.400</td><td>0.103</td></tr><tr><td>Supervised2</td><td></td><td>0.727</td><td>0.170</td><td>0.871</td><td>0.062</td></tr><tr><td>Ours</td><td></td><td>0.763</td><td>0.116</td><td>0.886</td><td>0.057</td></tr></table>",
+        "bbox": [
+            76,
+            142,
+            490,
+            411
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "port the results from models adopting Xavier initialization [13] (Random) or supervised ImageNet pre-training (Supervised). Our SSL pre-training framework (Ours) achieves the best performance across both metrics when applied on MirrorNet and VCNet and shows a good generalization of different mirror detection methods. It is worth noting that our framework even outperforms the supervised ImageNet pre-training.",
+        "bbox": [
+            75,
+            436,
+            468,
+            556
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            76,
+            566,
+            228,
+            583
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 3 demonstrates the effectiveness of each pretraining stage in our framework. We analyze its results point by point as follows.",
+        "bbox": [
+            75,
+            590,
+            468,
+            636
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Image-level Pre-training vs. Random Initialization. To evaluate the effectiveness of the proposed image-level pretraining, we compare the model only adopting image-level pre-training (B2) with the model initialized randomly without any pre-training (B1). We can see that only adopting image-level pre-training (B2) will even cause a performance drop compared with the one without pre-training (B1). This somehow supports the finding from VCNet [30] that image-level flipping cannot provide sufficient information to pixelwisely locate the mirror inside an image.",
+        "bbox": [
+            75,
+            642,
+            468,
+            792
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/2f7cfdf59eca0a1b160578d6d2e9538726cbed7619ecb9b41733fca790aad8b2.jpg",
+        "table_caption": [
+            "Table 3. Ablation study of the proposed SSL pre-training framework. We use MirrorNet [37] trained on the MSD dataset as our base method to conduct the experiments in this ablation study. Best results are shown in bold."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Rand. Init.</td><td>Image-lev. Pretrain</td><td>Patch-lev. Pretrain</td><td>Pixel-lev. Pretrain</td><td>Fβ↑</td><td>MAE↓</td></tr><tr><td>B1</td><td>✓</td><td></td><td></td><td></td><td>0.601</td><td>0.195</td></tr><tr><td>B2</td><td></td><td>✓</td><td></td><td></td><td>0.623</td><td>0.191</td></tr><tr><td>B3</td><td></td><td></td><td>✓</td><td></td><td>0.689</td><td>0.162</td></tr><tr><td>B4</td><td></td><td></td><td></td><td>✓</td><td>0.620</td><td>0.198</td></tr><tr><td>B5</td><td></td><td>✓</td><td>✓</td><td></td><td>0.723</td><td>0.166</td></tr><tr><td>B6</td><td></td><td></td><td>✓</td><td>✓</td><td>0.758</td><td>0.124</td></tr><tr><td>B7</td><td></td><td>✓</td><td></td><td>✓</td><td>0.666</td><td>0.178</td></tr><tr><td>Ours</td><td></td><td>✓</td><td>✓</td><td>✓</td><td>0.763</td><td>0.116</td></tr></table>",
+        "bbox": [
+            500,
+            141,
+            898,
+            280
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Patch-level Pre-training vs. Random Initialization. We compare our patch-level pre-training (B3) with random initialization (B1) to analyze if patch-level pre-training is useful in our SSL pre-training framework. Our patch-level pretraining outperforms the model without pre-training on both two metrics (14.64% improvement on  $F_{\\beta}$  and 16.32% improvement on MAE), which indicates the effectiveness of our patch-level pre-training. It also demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level.",
+        "bbox": [
+            498,
+            311,
+            893,
+            465
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Pixel-level Pre-training vs. Random Initialization. Similarly, we conduct the ablation experiment on our pixel-level pre-training. The model adopting pixel-level pre-training performs better than the model with image pre-training. However, the performance gain is not as much as the model with patch-level pre-training (B3). One possible reason is that directly optimizing the backbone network for the representation of mirrors at a pixel level is challenging and may not produce satisfactory results for mirror detection.",
+        "bbox": [
+            498,
+            473,
+            890,
+            609
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Location Predictor in Patch-level Pre-training. The location predictor in our patch-level pre-training is designed to capture the relational spatial information, which has been proven important in mirror detection [23]. We conduct an ablation experiment to remove the location predictor and the performance of our methods drops from 0.763 to 0.731. This shows the importance of learning spatial information in the mirror detection task.",
+        "bbox": [
+            498,
+            618,
+            890,
+            739
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Combinations of Different Pre-training Stages. Based on the above three comparisons, we find that the patch-level pre-training contributes the most improvement. It demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level. We then analyze the results from the model with different combinations of our pretraining stages. The model with image-level pre-training and patch-level pre-training (B5) performs better than the model that only adopts patch-level pre-training (B3), which shows that the global information provided by image-level",
+        "bbox": [
+            496,
+            750,
+            893,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "2Note that the original backbone networks reported in their papers are ResNext-101 [35]. However, the model weights of ResNext-101 pre-trained by our baseline SSL frameworks are not available in mmselfsup [6]. Thus, we switch the backbone network of these two methods to ResNet-50 [17], which is a common backbone network used in evaluating the performances of SSL frameworks. We also directly use the released ResNet-50 SSL pre-trained weights in our experiments for fair comparisons.",
+        "bbox": [
+            75,
+            801,
+            468,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "12232",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/c94e799004490cc55ca02611f6e1d3d0a18c07e817a95c713108ccb4eda94f38.jpg",
+        "image_caption": [
+            "Image"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            107,
+            103,
+            215,
+            207
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/adb43e744b866dd7ca54d2d708480002a84938a4c0f338e7cd4663448de446fd.jpg",
+        "image_caption": [
+            "Image-level Pretraining"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            217,
+            103,
+            323,
+            207
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/ef97824b33362b7b6edd9ca6e0ece5d9cda25eeb9aba72ad576370e87fd8e77c.jpg",
+        "image_caption": [
+            "Patch-level Pretraining"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            325,
+            103,
+            433,
+            207
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/6155df1df3d0122b4be72a7609d83ada630161ef92e3acf130d06f647cb0f2fd.jpg",
+        "image_caption": [
+            "Pixel-level Pretraining"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            434,
+            103,
+            540,
+            207
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/db77c22435167185c10f3cc363d5ff96d45d33e0f592430360e6dbc4e66205d5.jpg",
+        "image_caption": [
+            "Figure 4. Visualization of the effects of the proposed pre-training stages."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            542,
+            103,
+            648,
+            207
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/20de36026fd80a4e087ead7e427f1f5f25a35580a4efbb87fc1d63e9b09084d1.jpg",
+        "image_caption": [
+            "(Image+Patch)-level Pretraining",
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            651,
+            103,
+            756,
+            207
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2247c9baf0804f7f3164b3547a535be04149800f888792e44bb69438800898a8.jpg",
+        "image_caption": [
+            "GT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            759,
+            103,
+            867,
+            207
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "pre-training can benefit the model with patch-level pretraining. Similarly, as shown in B7, B4 (i.e., only pixel-level pretraining) can boost the performance of the pre-trained model after image-level pre-training. These two comparisons indicate that image-level pre-training should be used with other pre-training stages to benefit the overall performance and it should not be solely adopted to our SSL pretraining framework. Besides, we find that combining patch-level and pixel-level pre-training (B6) has superior performance when compared with B3, especially on MAE with about  $23.4\\%$  improvement. This combination also outperforms supervised ImageNet pre-training, according to the results from Table 2. We attribute its superior performance to the pixel-level representation of mirror reflection. Our final model with all pre-training stages outperforms other ablation models, showing the effectiveness of adopting pretraining stages progressively.",
+        "bbox": [
+            75,
+            285,
+            472,
+            542
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Visual Comparison of Our Pre-training Stages. Figure 4 shows a visual example of our ablation study. We can see that only adopting image-level pre-training on our SSL pre-training framework can approximately locate the mirror region but fails to produce a precise boundary for the predicted mask of the mirror. Patch-level pre-training can help predict mirrors more accurately with clearer boundaries but over-predict the non-mirror regions as mirrors. In comparison to predictions from all other ablation models, the prediction from the model with pixel-level pretraining has the clearest boundary but contains the largest over-predicted regions, primarily because of the insufficient global-level learning on mirrors. Our framework can significantly reduce over-predictions when image-level and patch-level pre-training are used, as opposed to when patch-level pre-training is used only, and performs the best with the help of the global-to-local pre-training process.",
+        "bbox": [
+            75,
+            549,
+            468,
+            806
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4. Discussions",
+        "text_level": 1,
+        "bbox": [
+            76,
+            816,
+            202,
+            830
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Impacts of Pre-training Strategy. Some previous SSL works [42, 45], especially for those focusing on self-supervised pre-training for downstream tasks like segmentation [42] and salient object detection [45], would adopt",
+        "bbox": [
+            75,
+            839,
+            468,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/9739fadea13fe3780833f9878be510cf87ad0a19c2692b950d139ef45f588e77.jpg",
+        "table_caption": [
+            "Table 4. The results of MirrorNet [37] when using different pretraining strategies. Best results are shown in bold."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Fβ↑</td><td>MAE↓</td></tr><tr><td>Random Initialized</td><td>0.586</td><td>0.200</td></tr><tr><td>ImageNet Supervised</td><td>0.727</td><td>0.170</td></tr><tr><td>Ours (In parallel)</td><td>0.708</td><td>0.331</td></tr><tr><td>Ours (Sequential)</td><td>0.763</td><td>0.116</td></tr></table>",
+        "bbox": [
+            563,
+            308,
+            828,
+            406
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "different pretext tasks sequentially to pre-train their backbone networks. We also adopt this common strategy in our proposed SSL framework. To verify if employing different pre-training strategies may affect the performance of the proposed SSL framework, we have tried training the three stages sequentially as well as in parallel. We find that when pre-training the three stages in parallel, the backbone network fails to converge and its final fine-tuning results are much worse than those using sequential pre-training. Table 4 shows the results of this experiment. We can see that the MAE performance of pre-training the different stages in parallel is even worse than that of random initialization. A possible reason for this is that directly pre-training different pretext tasks in parallel is difficult to transfer the learned knowledge from the pre-trained network to the target downstream task (i.e., pixel-level mirror detection).",
+        "bbox": [
+            496,
+            435,
+            890,
+            678
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Apart from the superior performance made by the sequential pre-training strategy, we also notice that the number of pre-training iterations for each pretext task can heavily affect the performance of our SSL framework. To avoid heavy hyperparameter tuning and increase the reproducibility of our pre-training framework, we adopt the same number of pre-training iterations for all stages. Although this strategy of using the same number of pre-training iterations is likely not the optimal design for individual stages, we believe that it can better reflect the real performance of our framework with less stochasticity and more reactivity. We provide a more detailed discussion on this issue next.",
+        "bbox": [
+            496,
+            680,
+            890,
+            861
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Impacts of Pre-training Iterations. SSL pre-training frameworks usually require huge pre-training iterations to",
+        "bbox": [
+            500,
+            869,
+            890,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "12233",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/eb08feac410142ebb0149c6e0bde878812dd728c60a1ad1abf429cb263c5f165.jpg",
+        "image_caption": [
+            "Figure 5. Ablation study on the number of pre-training iterations for our SSL pre-training framework. We find that our SSL pretraining framework continues to improve in performance until  $\\sim 20,000$  pre-training iterations. After that, the performance starts to decline, mainly due to overfitting."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            133,
+            90,
+            410,
+            275
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "converge. For example, SimSiam [5] requires about 250K iterations to pre-train their model on ImageNet, which is very time-consuming. To validate how this issue may affect our SSL pre-training framework, we conduct the ablation study on the number of pre-training iterations for our SSL pre-training framework, and the results are shown in Figure 5. Note that we use the same number of iterations for all pre-training stages for convenience and to get rid of heavy hyperparameter tuning.",
+        "bbox": [
+            75,
+            372,
+            467,
+            508
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "From the results, we find that the performance of our SSL pre-training framework improves rapidly as the number of pre-training iterations increases, up to around 20,000, which is a relatively small number compared with those used in existing SSL pre-training frameworks. The performance of our SSL pre-training framework gradually drops as the number of pre-training iterations exceeds 20,000, mainly due to overfitting. One possible explanation of this is that the model pre-trained by our SSL pre-training framework can efficiently learn the representation of mirror reflection under a short pre-training period. Another possible explanation is that unlike generic SSL pre-training frameworks that attempt to extract general-propose image representation, our SSL pre-training framework is developed specially for a specific task, i.e., mirror detection. While generic SSL pre-training may cause feature redundancy for our task, having too many pre-training iterations (i.e., over 20,000) causes overfitting.",
+        "bbox": [
+            75,
+            511,
+            467,
+            785
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Failure Cases. Our SSL pre-training framework does have limitations. It continues to struggle with the limitations of the base method [37], as the base method adopted can also significantly affect the final detection performances. For example, as shown in first row of Figure 6, MirrorNet with our SSL pre-training may still fail to handle some mirror-like regions (e.g., the top-right region). It may also fail to de",
+        "bbox": [
+            75,
+            794,
+            467,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5e0fa381c7da1010282d6fc2a226bfa379b1d1dfb405d95dcf84c3caa7ac4974.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            511,
+            87,
+            635,
+            165
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a21f8058477846417d25926fda83d6439c42b732b24f2a0af94409428e39c431.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            637,
+            88,
+            759,
+            166
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5739f80c78e3835cac5bc18142f5c73c76e0a05db49d5b10faeba1c9b4c9370a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            761,
+            88,
+            883,
+            166
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/7d9e1db27b0ef5915888bd41c21fdb0bf768ae131bbe9dc9f456e9e524bf3e5c.jpg",
+        "image_caption": [
+            "Image"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            511,
+            167,
+            635,
+            287
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/646ac3372c6f1b8574c145ef0ee9f0430b5de75bf66388c8b1ba8e482a373c38.jpg",
+        "image_caption": [
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            637,
+            167,
+            759,
+            287
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ef33e88d90d37b68820d4534a4f8cef4e29bf268f405ad1c5d3162e47e1c4c44.jpg",
+        "image_caption": [
+            "GT",
+            "Figure 6. Failure cases of our SSL pre-training framework based on MirrorNet. Our SSL pre-training framework fails to tackle the intrinsic limitation of existing mirror detection methods. The model pre-trained by our SSL pre-training framework still has wrong predictions when the input images are challenging (e.g., containing mirror-like regions in the first row, and insufficient relationship between mirror/non-mirror regions in the second row."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            761,
+            167,
+            883,
+            287
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tect mirrors in some challenging cases when the input image contains ambiguous mirrors without the relationship of mirror reflection and sufficient contextual contrast, as shown in the second row of Figure 6.",
+        "bbox": [
+            498,
+            441,
+            890,
+            503
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            521,
+            617,
+            536
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we have investigated how self-supervised learning (SSL) works with the mirror detection task. To the best of our knowledge, we are the first to explore SSL pre-training frameworks applied to mirror detection. We have found that the supervised ImageNet pre-training might not be the ideal way to extract backbone image features for mirror detection due to the discrepancy between general-propose representation and mirror-specific representation. We have also proposed a new SSL pre-training framework to pre-train the backbone network for mirror detection. Our SSL pre-training framework does not require any labeled data. It progressively (image-, patch- and pixel-level pre-training) models the relationship of mirror reflection, and obtains a better representation of mirrors in the pre-training stage. Experimental results show that the proposed SSL pre-training framework outperforms CNN-based state-of-the-art SSL pre-training frameworks, and even achieves better results compared with supervised ImageNet pre-training.",
+        "bbox": [
+            496,
+            547,
+            890,
+            835
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments: The work described in this paper was partially supported by a GRF grant from the Research Grants Council of Hong Kong (Project No.: CityU 11211223).",
+        "bbox": [
+            498,
+            839,
+            890,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "12234",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            173,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018.",
+            "[2] M. Braun, S. Krebs, F. Flohr, and D. M. Gavrila. Eurocity persons: A novel benchmark for person detection in traffic scenes. IEEE TPAMI, 41(8):1844-1861, Aug 2019.",
+            "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020.",
+            "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020.",
+            "[5] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021.",
+            "[6] MMSelfSup Contributors. MMSelfSup: Openmmlab self-supervised learning toolbox and benchmark. https://github.com/open-mmlab/mmselfsup, 2021.",
+            "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016.",
+            "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009.",
+            "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017.",
+            "[10] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015.",
+            "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010.",
+            "[12] Christopher Funk and Yanxi Liu. Beyond planar symmetry: Modeling human perception of reflection and rotation symmetries in the wild. In ICCV, pages 793-803, 2017.",
+            "[13] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010.",
+            "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020.",
+            "[15] Huankang Guan, Jiaying Lin, and Rynson W.H. Lau. Learning semantic associations for mirror detection. In CVPR, pages 5941-5950, June 2022."
+        ],
+        "bbox": [
+            78,
+            116,
+            468,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020.",
+            "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016.",
+            "[18] Ruozhen He, Jiaying Lin, and Rynson WH Lau. Efficient mirror detection via multi-level heterogeneous learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 790-798, 2023.",
+            "[19] Tianyu Huang, Bowen Dong, Jiaying Lin, Xiaohui Liu, Rynson WH Lau, and Wangmeng Zuo. Symmetry-aware transformer-based mirror detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 935–943, 2023.",
+            "[20] Peter Kohler, Alexandra Yakovleva, Alasdair Clarke, Yanxi Liu, and Anthony Norcia. Parametric responses to rotation symmetry in mid-level visual cortex. Journal of Vision, 15(12):1122-1122, 2015.",
+            "[21] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In ICLR, 2018.",
+            "[22] Jiaying Lin, Xin Tan, and Rynson WH Lau. Learning to detect mirrors from videos via dual correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9109-9118, 2023.",
+            "[23] Jiaying Lin, Guodong Wang, and Rynson W. H. Lau. Progressive mirror detection. In CVPR, 2020.",
+            "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014.",
+            "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014.",
+            "[26] Zhiqiu Lin, Jin Sun, Abe Davis, and Noah Snavely. Visual chirality. In CVPR, pages 12295-12303, 2020.",
+            "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018.",
+            "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015.",
+            "[29] Jiaqi Tan, Weijie Lin, Angel X Chang, and Manolis Savva. Mirror3d: Depth refinement for mirror surfaces. In CVPR, pages 15990-15999, 2021.",
+            "[30] Xin Tan, Jiaying Lin, Ke Xu, Chen Pan, Lizhuang Ma, and Rynson W. H. Lau. Mirror detection with the visual chirality cue. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.",
+            "[31] Christopher W Tyler. Human symmetry perception and its computational analysis. Psychology Press, 2003.",
+            "[32] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, 2021."
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "12235",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[33] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018.",
+            "[34] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, pages 418-434, 2018.",
+            "[35] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 1492-1500, 2017.",
+            "[36] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022.",
+            "[37] Xin Yang, Haiyang Mei, Ke Xu, Xiaopeng Wei, Baocai Yin, and Rynson W.H. Lau. Where is my mirror? In ICCV, 2019.",
+            "[38] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Learning with noisy labels for robust point cloud segmentation. In ICCV, pages 6443-6452, 2021.",
+            "[39] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Robust point cloud segmentation with noisy annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7696-7710, 2022.",
+            "[40] Shuquan Ye, Yujia Xie, Dongdong Chen, Yichong Xu, Lu Yuan, Chenguang Zhu, and Jing Liao. Improving commonsense in vision-language models via knowledge graph riddles. In CVPR, pages 2634-2645, 2023.",
+            "[41] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021.",
+            "[42] Xiaohang Zhan, Ziwei Liu, Ping Luo, Xiaou Tang, and Chen Loy. Mix-and-match tuning for self-supervised semantic segmentation. In AAAI, volume 32, 2018.",
+            "[43] Xiaohang Zhan, Jiahao Xie, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Online deep clustering for unsupervised representation learning. In CVPR, 2020.",
+            "[44] Nanxuan Zhao, Zhirong Wu, Rynson WH Lau, and Stephen Lin. What makes instance discrimination good for transfer learning? In ICLR, 2021.",
+            "[45] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, , Hutchuan Lu, and Xiang Ruan. Self-supervised pretraining for rgb-d salient object detection. In AAAI, 2022.",
+            "[46] Ying Zheng, Yiyi Zhang, Xiaogang Xu, Jun Wang, and Hongxun Yao. Visual chirality meets freehand sketches. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1544-1548. IEEE, 2021."
+        ],
+        "bbox": [
+            78,
+            90,
+            467,
+            729
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "12236",
+        "bbox": [
+            480,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2023/Self-supervised Pre-training for Mirror Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_model.json

@@ -0,0 +1,2200 @@
+[
+    [
+        {
+            "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": [
+                0.235,
+                0.131,
+                0.737,
+                0.154
+            ],
+            "angle": 0,
+            "content": "Self-supervised Pre-training for Mirror Detection"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.342,
+                0.182,
+                0.626,
+                0.218
+            ],
+            "angle": 0,
+            "content": "Jiaying Lin* Rynson W.H. Lau* City University of Hong Kong"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.324,
+                0.22,
+                0.648,
+                0.234
+            ],
+            "angle": 0,
+            "content": "{jiayinlin5, rynson.lau}@cityu.edu.hk"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.235,
+                0.269,
+                0.314,
+                0.285
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.075,
+                0.302,
+                0.474,
+                0.726
+            ],
+            "angle": 0,
+            "content": "Existing mirror detection methods require supervised ImageNet pre-training to obtain good general-purpose image features. However, supervised ImageNet pre-training focuses on category-level discrimination and may not be suitable for downstream tasks like mirror detection, due to the overfitting upstream tasks (e.g., supervised image classification). We observe that mirror reflection is crucial to how people perceive the presence of mirrors, and such mid-level features can be better transferred from self-supervised pretrained models. Inspired by this observation, in this paper we aim to improve mirror detection methods by proposing a new self-supervised learning (SSL) pre-training framework for modeling the representation of mirror reflection progressively in the pre-training process. Our framework consists of three pre-training stages at different levels: 1) an image-level pre-training stage to globally incorporate mirror reflection features into the pre-trained model; 2) a patch-level pre-training stage to spatially simulate and learn local mirror reflection from image patches; and 3) a pixel-level pre-training stage to pixel-wisely capture mirror reflection via reconstructing corrupted mirror images based on the relationship between the inside and outside of mirrors. Extensive experiments show that our SSL pre-training framework significantly outperforms previous state-of-the-art CNN-based SSL pre-training frameworks and even outperforms supervised ImageNet pre-training when transferred to the mirror detection task. Code and models are available at https://jiaying.link/iccv2023-sslmirror/"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.757,
+                0.21,
+                0.772
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.783,
+                0.47,
+                0.875
+            ],
+            "angle": 0,
+            "content": "Mirrors are prevalent in our daily lives. As their appearances are largely determined by their surroundings, they generally lack a consistent appearance, making it difficult to separate them from their surroundings. This may affect many computer vision tasks such as object detection [2], vision-language navigation [1] and depth estimation [29]."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.51,
+                0.268,
+                0.584,
+                0.342
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.587,
+                0.268,
+                0.659,
+                0.341
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.587,
+                0.343,
+                0.659,
+                0.416
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.587,
+                0.418,
+                0.655,
+                0.439
+            ],
+            "angle": 0,
+            "content": "(b) MirrorNet on ImageNet"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.662,
+                0.268,
+                0.736,
+                0.341
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.662,
+                0.343,
+                0.736,
+                0.416
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.669,
+                0.418,
+                0.73,
+                0.439
+            ],
+            "angle": 0,
+            "content": "(c)VCNet on ImageNet"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.738,
+                0.268,
+                0.812,
+                0.341
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.738,
+                0.343,
+                0.812,
+                0.416
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.737,
+                0.418,
+                0.822,
+                0.439
+            ],
+            "angle": 0,
+            "content": "1) Ours (MirrorNet on our SSL)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.815,
+                0.268,
+                0.888,
+                0.341
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.815,
+                0.343,
+                0.888,
+                0.416
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.835,
+                0.424,
+                0.867,
+                0.434
+            ],
+            "angle": 0,
+            "content": "(e) GT"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.499,
+                0.45,
+                0.893,
+                0.59
+            ],
+            "angle": 0,
+            "content": "Figure 1. State-of-the-art mirror detection methods [37, 30] are based on costly supervised ImageNet pre-training. They may fail even in obvious cases, e.g., (a) when the mirror clearly reflects a real object outside of the mirror. (b) and (c) are MirrorNet [37] and VCNet [30], respectively, pre-trained on ImageNet with full supervision. (c) is MirrorNet with our proposed SSL framework and without supervised ImageNet pre-training. Our SSL scheme leverages the mirror reflection cue and avoids feature redundancy in the pre-training stage. It outperforms those pre-trained on ImageNet with full supervision (i.e., (b) and (c))."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.614,
+                0.892,
+                0.644
+            ],
+            "angle": 0,
+            "content": "It is therefore crucial to design high-performance mirror detection methods to facilitate computer vision applications."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.645,
+                0.893,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Recently, a few methods [37, 23, 30] have been proposed for mirror detection, but they all require to be initialized with supervised ImageNet [8] pre-trained weights and then fine-tuned on mirror detection datasets. While such a transfer learning approach is common in general computer vision tasks (e.g., object detection [24] and semantic segmentation [7]) to utilize category-oriented features from large-scale image datasets, we challenge if this strategy is necessary for the mirror detection task for two reasons. First, the supervised ImageNet pre-training process involves high labelling costs, as it is very time-consuming and labor-intensive to label such a large-scale image dataset (\\(\\sim\\)1.3M images). Second, unlike general vision tasks, mirror detection does not require category-level object understanding. Instead, it requires an understanding of the relationship between inside and outside of mirrors (e.g., contrast [37], similarity [23], and visual chirality [30]). Thus,"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.096,
+                0.887,
+                0.228,
+                0.9
+            ],
+            "angle": 0,
+            "content": "*Corresponding authors."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "12227"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.092,
+                0.47,
+                0.121
+            ],
+            "angle": 0,
+            "content": "while pre-training on ImageNet requires expensive labels, it causes feature redundancy for the mirror detection task."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.122,
+                0.468,
+                0.287
+            ],
+            "angle": 0,
+            "content": "Figure 1 shows that existing models pre-trained on ImageNet tend to over-detect the mirror regions. Although the two SOTA methods [37, 30] can correctly locate the mirror regions, they are unable to distinguish between the mirrored objects inside the mirrors and the real objects outside, even though these methods contain well-designed modules to learn to address this problem via the fine-tuning stage. This motivates us to investigate whether the current pre-training approach may affect the final detection performance, and whether we may incorporate some intrinsic mirror properties in the pre-training process to assist mirror detection."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.288,
+                0.468,
+                0.452
+            ],
+            "angle": 0,
+            "content": "An interesting observation is that humans do not need to be \"trained\" to recognize mirrors. Instead, we learn to recognize them implicitly from young. This aligns with the key idea of self-supervised learning (SSL). Works from neuroscience [31, 20] have shown that humans use mid-level visual cortex to recognize mirror reflection. According to [44], mid-level features are hard to be transferred from supervised pre-training on ImageNet, and SSL is suitable for learning mid-level representations, which inspires us to develop an SSL pre-training framework for modeling mirror reflection."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.455,
+                0.468,
+                0.814
+            ],
+            "angle": 0,
+            "content": "In this paper, we propose a new SSL pre-training framework for mirror detection, which explicitly considers mirror reflection during the pre-training process. Our framework does not require human-annotated labels from large-scale image datasets. It consists of three stages to progressively pre-train the backbone network from global to local: 1) an image-level pre-training stage to obtain the representation of mirror reflection globally by recognizing the geometric transformation applied to the image; 2) a patch-level pretraining stage to mimic patch-wise mirror reflection and then learn the spatial correlation between the original object patches and the corresponding mirrored object patches; and 3) a pixel-level pre-training stage to extract the pixel-to-pixel relationship of mirror reflection by image reconstruction; Under such progressive pre-training scheme, our SSL pre-training framework can learn the representation of mirror reflection, and then effectively transfer this knowledge to the subsequent mirror detection process for better detection performances. We conduct comprehensive experiments to demonstrate the effectiveness of our SSL pretraining framework. We show that it can significantly boost the performances of existing mirror detection methods, and outperform other CNN-based SSL pre-training frameworks on the mirror detection task."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.097,
+                0.817,
+                0.457,
+                0.832
+            ],
+            "angle": 0,
+            "content": "To conclude, this paper makes three key contributions:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.096,
+                0.84,
+                0.468,
+                0.901
+            ],
+            "angle": 0,
+            "content": "- To the best of our knowledge, we are the first to investigate how existing SSL pre-training frameworks perform on the mirror detection task, compared with supervised ImageNet pre-training."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.518,
+                0.092,
+                0.892,
+                0.198
+            ],
+            "angle": 0,
+            "content": "- We propose a new SSL pre-training framework that consists of three stages at different levels to progressively learn the representation of mirror reflection. Compared with the features from a supervised ImageNet pre-trained model, our representation is better due to the reduced gap between the pre-training task and the target downstream task (i.e., mirror detection)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.519,
+                0.21,
+                0.892,
+                0.271
+            ],
+            "angle": 0,
+            "content": "- Extensive experiments show that our SSL pre-training framework performs the best among all state-of-the-art SSL methods, and even better than models with supervised ImageNet pre-training."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.518,
+                0.092,
+                0.892,
+                0.271
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.286,
+                0.642,
+                0.302
+            ],
+            "angle": 0,
+            "content": "2. Related Work"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.311,
+                0.668,
+                0.326
+            ],
+            "angle": 0,
+            "content": "2.1. Mirror Detection"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.336,
+                0.892,
+                0.486
+            ],
+            "angle": 0,
+            "content": "Mirror detection is an essential task in computer vision, as mirrors are everywhere these days. It is also a challenging task, since mirrors do not possess a consistent appearance, but reflect those of their surroundings. Other vision tasks such as generic object detection and segmentation would tend to detect and segment objects outside as well as inside of the mirror, providing incorrect information on the mirror region, and a depth prediction task may return the depths of the reflected objects inside the mirror region instead of the depth of the mirror."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.487,
+                0.892,
+                0.637
+            ],
+            "angle": 0,
+            "content": "To address the mirror detection problem, Yang et al. [37] propose the first large-scale mirror detection dataset and the first deep-learning model, MirrorNet, for mirror detection by learning the contextual contrast between the inside and outside of mirrors. Lin et al. [23] further propose a new model with similarity and edge learning to progressively detect mirrors. Most recently, Guan et al. [15] employ graph convolutional networks to model contextual associations for mirror detection, while Tan et al. propose VCNet [30] to exploit the visual chirality cue at a feature level."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.64,
+                0.892,
+                0.759
+            ],
+            "angle": 0,
+            "content": "However, existing mirror detection methods are all fully-supervised. They also directly adopt the supervised ImageNet [8] pre-training to obtain generic image features. Unlike recent works [19, 18, 22] that propose novel network architectures for mirror detection, we propose in this paper to tackle the mirror detection problem from a learning-based perspective, offering the first self-supervised pre-training framework for mirror detection."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.771,
+                0.776,
+                0.787
+            ],
+            "angle": 0,
+            "content": "2.2. Self-supervised Learning (SSL)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.795,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "SSL is a popular research topic in computer vision [38, 39], and has attracted a lot of research interests. It aims to learn good feature representations without needing labeled data. Early SSL methods focus on proposing novel pretext tasks, such as relative location prediction [10] and rotation prediction [21]. Recently, most methods are based on the utilization of contrastive learning [16, 4, 14, 3, 32, 5, 41,"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "12228"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.082,
+                0.089,
+                0.46,
+                0.184
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.465,
+                0.089,
+                0.891,
+                0.184
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.164,
+                0.184,
+                0.835,
+                0.341
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.347,
+                0.895,
+                0.391
+            ],
+            "angle": 0,
+            "content": "Figure 2. Our SSL pre-training framework. It consists of three pre-training stages at different levels: an image-level pre-training stage, a patch-level pre-training stage, and a pixel-level pre-training stage. It pre-trains the backbone network from global to local (image-level \\(\\rightarrow\\) patch-level \\(\\rightarrow\\) pixel-level) progressively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.413,
+                0.47,
+                0.52
+            ],
+            "angle": 0,
+            "content": "40]. Most recently, [44] shows that some SSL methods perform even better than supervised ImageNet pre-training in downstream tasks, such as object detection [11, 24] and semantic segmentation [7, 34], with the help of instance discrimination during SSL pre-training. Recent transformer-based SSL frameworks [36] also involve masked image modeling to obtain better self-supervised representations."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.52,
+                0.471,
+                0.643
+            ],
+            "angle": 0,
+            "content": "In this paper, we observe that unlike general vision tasks, mirror detection is not benefited from supervised ImageNet pre-training since it does not require category-level object understanding. Hence, we propose a novel SSL pre-training framework specifically for the mirror detection task. The proposed framework aims to simulate and learn the relationship between mirrors and their surrounding during the pre-training stage."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.657,
+                0.169,
+                0.673
+            ],
+            "angle": 0,
+            "content": "3. Method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.683,
+                0.47,
+                0.806
+            ],
+            "angle": 0,
+            "content": "Figure 2 shows the overall architecture of the proposed SSL pre-training framework. Our SSL pre-training framework consists of three pre-training stages at different levels: an image-level pre-training, a patch-level pre-training, and a pixel-level pre-training. The pre-training at each level is conducted sequentially and independently. The progressive and sequential training strategy can help the pre-trained model obtain a global-to-local representation of mirrors."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.817,
+                0.303,
+                0.833
+            ],
+            "angle": 0,
+            "content": "3.1. Image-level Pre-training"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.84,
+                0.471,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Inspired by the definition of visual chirality [26] and its potential applications [46, 30] in various computer vision problems, we formulate the image-level pre-training as a binary classification problem: Given an input image and its"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.413,
+                0.893,
+                0.55
+            ],
+            "angle": 0,
+            "content": "corresponding horizontal flipped image, can the backbone network tell which one is the original image? To achieve this objective, we attach a classification head to predict the probability that the input image is flipped. The classification head consists of a convolution layer followed by a sigmoid activation function to project the output value into \\([0,1]\\). In this pre-training stage, we use binary cross-entropy (BCE) as our loss function \\(\\mathcal{L}_f\\) to optimize the backbone network. Formally, we have:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.564,
+                0.558,
+                0.892,
+                0.576
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {f} = - \\left[ y \\log (\\hat {y}) + (1 - y) \\log (1 - \\hat {y}) \\right], \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.583,
+                0.892,
+                0.614
+            ],
+            "angle": 0,
+            "content": "where \\(y\\) is the ground truth label and \\(\\hat{y}\\) is the predicted probability that the input image is horizontally flipped."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.621,
+                0.722,
+                0.637
+            ],
+            "angle": 0,
+            "content": "3.2. Patch-level Pre-training"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.644,
+                0.893,
+                0.902
+            ],
+            "angle": 0,
+            "content": "The image-level pre-training pretext task alone cannot bring much help in locating the mirror inside an image since mirror detection is a pixel-level problem. However, directly connecting the image-level and the pixel-level pretext tasks is challenging due to the gap between the scales of these two pretext tasks. Thus, we design a patch-level pre-training stage as a bridge to cover the feature representations of mirrors at different scales. Besides, the image-level pre-training does not benefit relational spatial understanding, which has been proven important in mirror detection [23]. Our patch-level pre-training consists of two sub-tasks to both obtain the patch-wise relationship of mirror reflection and simulate the mirrored region with pseudo labels. We first split the input image into a grid of \\( 3 \\times 3 \\) image patches and randomly shuffle them. After that, we randomly apply a flipping operation to one of the image patches, to simulate a mirrored region. The first sub-task"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "12229"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.091,
+                0.473,
+                0.213
+            ],
+            "angle": 0,
+            "content": "is to predict the location index from the shuffled grid by the backbone network. We append a location predictor head to the backbone network to predict the location index of input patches. We apply cross-entropy loss as \\(\\mathcal{L}_{loc}\\) to optimize this sub-task. The second sub-task is to predict the flipped patch from the nine randomly shuffled image patches. Similar to our image-level pre-training, we use BCE loss as our loss function \\(\\mathcal{L}_{cls}\\) in this sub-task."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.096,
+                0.224,
+                0.47,
+                0.265
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {l o c} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} y _ {i} \\log \\left(\\hat {y} _ {i}\\right), \\tag {2}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.096,
+                0.269,
+                0.47,
+                0.311
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {c l s} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} y _ {j} \\log \\left(\\hat {y _ {j}}\\right) + (1 - y _ {j}) \\log \\left(1 - \\hat {y _ {j}}\\right), \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.322,
+                0.47,
+                0.399
+            ],
+            "angle": 0,
+            "content": "where \\(\\hat{y}_i\\) is the predicted location index, and \\(y_i\\) is the corresponding ground truth label. \\(\\hat{y}_j\\) is the predicted probability that the input patch is horizontally flipped and \\(y_j\\) is the corresponding ground truth probability. Finally, we jointly optimize these two sub-tasks and the final loss function is:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.197,
+                0.426,
+                0.47,
+                0.445
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\text {p a t c h}} = \\mathcal {L} _ {\\text {l o c}} + \\mathcal {L} _ {\\text {c l s}}. \\tag {4}\n\\]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.454,
+                0.294,
+                0.471
+            ],
+            "angle": 0,
+            "content": "3.3. Pixel-level Pre-training"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.478,
+                0.473,
+                0.721
+            ],
+            "angle": 0,
+            "content": "Extracting the pixel-level relationship of mirror reflection is a challenging task, as studied by previous related work [12] for symmetry detection. It requires point-level GT labels to represent symmetry centers. However, obtaining such labels requires huge human annotation effort and they are not applicable for SSL pre-training. Instead of generating pseudo labels for symmetry detection, our solution is that we formulate this task as an image reconstruction problem on images containing mirrors. An ideal pre-trained network should be able to reconstruct the input image containing mirrors with the pixel-level understanding of mirror reflection. We randomly mask out the input image by cutout augmentation [9] with the ratio of 0.3 and then feed it to the backbone network with a reconstruction head. The details of the reconstruction head we used are listed in Table 1. The objective of our pixel-level pre-training is defined as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.211,
+                0.733,
+                0.47,
+                0.751
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {p i x} = \\| x - \\hat {x} \\| _ {1}, \\tag {5}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.762,
+                0.47,
+                0.793
+            ],
+            "angle": 0,
+            "content": "where \\(\\mathcal{L}_{pix}\\) is an L1 loss. \\(x\\) and \\(\\hat{x}\\) are the input image and the reconstructed image, respectively."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.807,
+                0.21,
+                0.824
+            ],
+            "angle": 0,
+            "content": "4. Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.832,
+                0.285,
+                0.849
+            ],
+            "angle": 0,
+            "content": "4.1. Experimental Settings"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.856,
+                0.47,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Dataset and Evaluation Metrics. The image-level and patch-level pre-training experiments are conducted on the training set of ImageNet-100 [28] and MS COCO [25],"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.087,
+                0.894,
+                0.171
+            ],
+            "angle": 0,
+            "content": "Table 1. The architecture of the reconstruction head used in the pixel-level pre-training. Note that each \"conv-IR\" corresponds to a sequence of convolution layer, InstanceNorm layer and ReLU activation. \\( K \\), \\( S \\) and \\( P \\) denote the number of kernels, the number of strides and the padding size, respectively, used in the convolution layer."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.559,
+                0.172,
+                0.835,
+                0.368
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Layer Details</td></tr><tr><td>3 × 3 conv-IR, K = 1024, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>3 × 3 conv-IR, K = 512, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>3 × 3 conv-IR, K = 256, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>3 × 3 conv-IR, K = 128, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>3 × 3 conv-IR, K = 128, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>1 × 1 conv-IR, K = 3, S = 1, P = 1</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.394,
+                0.892,
+                0.612
+            ],
+            "angle": 0,
+            "content": "which contains \\(130\\mathrm{K}\\) and \\(\\sim 118\\mathrm{k}\\) images respectively. In the pixel-level pre-training and fine-tuning stage, our experiments are conducted on MSD [37], the first large-scale dataset for mirror detection. MSD [37] provides 4,018 mirror images, which are divided into 3,063 images for training and 955 images for testing. We adopt two popular metrics namely, F-measure \\((\\mathbf{F}_{\\beta})\\) and mean absolute error (MAE) to evaluate the performance of our SSL pre-training framework. F-measure can reflect the overall model performance and is calculated by the weighted harmonic mean of the precision and recall: \\(F_{\\beta} = \\frac{(1 + \\beta^{2})\\cdot\\text{precision}\\cdot\\text{recall}}{\\beta^{2}\\cdot\\text{precision} + \\text{recall}}\\), where \\(\\beta^2\\) is set to 0.3. MAE can measure the average pixel-wise absolute disparity between the predicted mirror maps \\(M\\) and the binary ground truth \\(G\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.612,
+                0.893,
+                0.84
+            ],
+            "angle": 0,
+            "content": "Implementation Details. We use Pytorch to implement our proposed SSL pre-training framework. All experiments are conducted on a GPU server with eight NVIDIA RTX 3090 GPUs. Following existing SSL pre-training frameworks, we adopt ResNet-50 [17] as our backbone network. All pre-trained model weights of existing SSL pretraining frameworks are obtained from the released models of mmselfsup [6] for fair comparisons. For each pretext task in the pre-training stage, all images are resized to \\(256 \\times 256\\). The batch size is set to 352. We use AdamW [27] with an initial learning rate of 1e-4 to optimize our SSL pre-training framework and stop pre-training after 20,000 iterations. We only adopt random color jittering for data augmentation. In the fine-tuning stage, we use the same hyperparameters reported in the original papers of the adopted"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.5,
+                0.851,
+                0.894,
+                0.901
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>It is worth noting that the datasets that we used (about 250K training images) are much smaller than the pre-trained dataset ImageNet (14 million images) used by other baseline SSL methods. We adopted ImageNet-100 to pre-train our network due to the limited computational resources."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "12230"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.086,
+                0.088,
+                0.891,
+                0.593
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.605,
+                0.893,
+                0.634
+            ],
+            "angle": 0,
+            "content": "Figure 3. Qualitative Comparison of our proposed method with NPID [33], MocoV2 [4], SwAV [3], Barlow Twins [41], ImageNet pretraining (Supervised). We use the released pre-trained models of these existing SSL methods from mmselfsup [6] for fair comparison."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.658,
+                0.385,
+                0.673
+            ],
+            "angle": 0,
+            "content": "base methods MirrorNet [37] and VCNet [30]."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.687,
+                0.383,
+                0.703
+            ],
+            "angle": 0,
+            "content": "4.2. Comparison with State-of-the-Arts"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.711,
+                0.47,
+                0.848
+            ],
+            "angle": 0,
+            "content": "We compare against 11 state-of-the-art CNN-based frameworks in self-supervised learning, including Relative Location (Rel. Loc.) [10], Rotation Prediction (Rot. Pred.) [21], NPID [33], ODC [43], Moco V2 [16], SimCLR [4], BYOL [14], SwAV [3], DenseCL [32], SimSiam [5] and Barlow Twins (Bar. Twins) [41]. We select MirrorNet [37] and VCNet [30], which are two representative mirror detection methods, as base methods to conduct experiments in the fine-tuning stage."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.856,
+                0.471,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Qualitative Evaluation. We further show the performance of our SSL pre-training framework in Figure 3. We compare our framework with four state-of-the-art SSL pre-training"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.658,
+                0.892,
+                0.84
+            ],
+            "angle": 0,
+            "content": "frameworks (including the two best-performing frameworks NPID [33] and Moco V2 [16] based on MirrorNet and the two best-performing frameworks SwAV [3] and Barlow Twins [41] based on VCNet, according to their performance in Table 2) and supervised ImageNet pre-training due to the limitation of space. Our framework can correctly detect the mirror regions by utilizing the relationship of mirror reflection caused by mirrors, even though the reflected regions are pure texture (i.e., lack of object keypoints in the second row) and relatively small in the image (e.g., the last three rows). We attribute the superior performance of our framework to the utilization of mirror reflection in the pre-training stage."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.856,
+                0.893,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Quantitative Evaluation. Table 2 shows the experimentation results comparing the proposed SSL pre-training framework with the state-of-the-arts. Besides, we also re"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.946,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "12231"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.077,
+                0.088,
+                0.47,
+                0.143
+            ],
+            "angle": 0,
+            "content": "Table 2. Quantitative comparison between our SSL pre-training framework and existing CNN-based SSL methods. We also compare it with random initialization (Random) and ImageNet pretraining (Supervised)."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.078,
+                0.143,
+                0.491,
+                0.412
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Methods</td><td rowspan=\"2\">Venue</td><td colspan=\"2\">MirrorNet</td><td colspan=\"2\">VCNet</td></tr><tr><td>Fβ↑</td><td>MAE↓</td><td>Fβ↑</td><td>MAE↓</td></tr><tr><td>Rel. Loc.</td><td>ICCV &#x27;15</td><td>0.582</td><td>0.254</td><td>0.842</td><td>0.072</td></tr><tr><td>Rot. Pred.</td><td>ICLR &#x27;18</td><td>0.537</td><td>0.212</td><td>0.851</td><td>0.069</td></tr><tr><td>NPID</td><td>CVPR &#x27;18</td><td>0.739</td><td>0.132</td><td>0.832</td><td>0.076</td></tr><tr><td>ODC</td><td>CVPR &#x27;20</td><td>0.558</td><td>0.223</td><td>0.836</td><td>0.077</td></tr><tr><td>Moco V2</td><td>CVPR &#x27;20</td><td>0.707</td><td>0.177</td><td>0.835</td><td>0.072</td></tr><tr><td>SimCLR</td><td>ICML &#x27;20</td><td>0.594</td><td>0.244</td><td>0.837</td><td>0.071</td></tr><tr><td>BYOL</td><td>NeurIPS &#x27;20</td><td>0.704</td><td>0.158</td><td>0.832</td><td>0.077</td></tr><tr><td>SwAV</td><td>NeurIPS &#x27;20</td><td>0.578</td><td>0.224</td><td>0.856</td><td>0.071</td></tr><tr><td>DenseCL</td><td>CVPR &#x27;21</td><td>0.694</td><td>0.240</td><td>0.844</td><td>0.072</td></tr><tr><td>SimSiam</td><td>CVPR &#x27;21</td><td>0.672</td><td>0.198</td><td>0.818</td><td>0.080</td></tr><tr><td>Bar. Twins</td><td>ICML &#x27;21</td><td>0.655</td><td>0.178</td><td>0.857</td><td>0.069</td></tr><tr><td>Random</td><td></td><td>0.586</td><td>0.200</td><td>0.400</td><td>0.103</td></tr><tr><td>Supervised2</td><td></td><td>0.727</td><td>0.170</td><td>0.871</td><td>0.062</td></tr><tr><td>Ours</td><td></td><td>0.763</td><td>0.116</td><td>0.886</td><td>0.057</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.438,
+                0.47,
+                0.558
+            ],
+            "angle": 0,
+            "content": "port the results from models adopting Xavier initialization [13] (Random) or supervised ImageNet pre-training (Supervised). Our SSL pre-training framework (Ours) achieves the best performance across both metrics when applied on MirrorNet and VCNet and shows a good generalization of different mirror detection methods. It is worth noting that our framework even outperforms the supervised ImageNet pre-training."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.568,
+                0.23,
+                0.584
+            ],
+            "angle": 0,
+            "content": "4.3. Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.591,
+                0.47,
+                0.637
+            ],
+            "angle": 0,
+            "content": "Table 3 demonstrates the effectiveness of each pretraining stage in our framework. We analyze its results point by point as follows."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.643,
+                0.47,
+                0.794
+            ],
+            "angle": 0,
+            "content": "Image-level Pre-training vs. Random Initialization. To evaluate the effectiveness of the proposed image-level pretraining, we compare the model only adopting image-level pre-training (B2) with the model initialized randomly without any pre-training (B1). We can see that only adopting image-level pre-training (B2) will even cause a performance drop compared with the one without pre-training (B1). This somehow supports the finding from VCNet [30] that image-level flipping cannot provide sufficient information to pixelwisely locate the mirror inside an image."
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.5,
+                0.088,
+                0.894,
+                0.142
+            ],
+            "angle": 0,
+            "content": "Table 3. Ablation study of the proposed SSL pre-training framework. We use MirrorNet [37] trained on the MSD dataset as our base method to conduct the experiments in this ablation study. Best results are shown in bold."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.5,
+                0.142,
+                0.9,
+                0.281
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Rand. Init.</td><td>Image-lev. Pretrain</td><td>Patch-lev. Pretrain</td><td>Pixel-lev. Pretrain</td><td>Fβ↑</td><td>MAE↓</td></tr><tr><td>B1</td><td>✓</td><td></td><td></td><td></td><td>0.601</td><td>0.195</td></tr><tr><td>B2</td><td></td><td>✓</td><td></td><td></td><td>0.623</td><td>0.191</td></tr><tr><td>B3</td><td></td><td></td><td>✓</td><td></td><td>0.689</td><td>0.162</td></tr><tr><td>B4</td><td></td><td></td><td></td><td>✓</td><td>0.620</td><td>0.198</td></tr><tr><td>B5</td><td></td><td>✓</td><td>✓</td><td></td><td>0.723</td><td>0.166</td></tr><tr><td>B6</td><td></td><td></td><td>✓</td><td>✓</td><td>0.758</td><td>0.124</td></tr><tr><td>B7</td><td></td><td>✓</td><td></td><td>✓</td><td>0.666</td><td>0.178</td></tr><tr><td>Ours</td><td></td><td>✓</td><td>✓</td><td>✓</td><td>0.763</td><td>0.116</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.313,
+                0.894,
+                0.466
+            ],
+            "angle": 0,
+            "content": "Patch-level Pre-training vs. Random Initialization. We compare our patch-level pre-training (B3) with random initialization (B1) to analyze if patch-level pre-training is useful in our SSL pre-training framework. Our patch-level pretraining outperforms the model without pre-training on both two metrics (14.64% improvement on \\( F_{\\beta} \\) and 16.32% improvement on MAE), which indicates the effectiveness of our patch-level pre-training. It also demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.474,
+                0.892,
+                0.61
+            ],
+            "angle": 0,
+            "content": "Pixel-level Pre-training vs. Random Initialization. Similarly, we conduct the ablation experiment on our pixel-level pre-training. The model adopting pixel-level pre-training performs better than the model with image pre-training. However, the performance gain is not as much as the model with patch-level pre-training (B3). One possible reason is that directly optimizing the backbone network for the representation of mirrors at a pixel level is challenging and may not produce satisfactory results for mirror detection."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.619,
+                0.892,
+                0.74
+            ],
+            "angle": 0,
+            "content": "Location Predictor in Patch-level Pre-training. The location predictor in our patch-level pre-training is designed to capture the relational spatial information, which has been proven important in mirror detection [23]. We conduct an ablation experiment to remove the location predictor and the performance of our methods drops from 0.763 to 0.731. This shows the importance of learning spatial information in the mirror detection task."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.75,
+                0.894,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Combinations of Different Pre-training Stages. Based on the above three comparisons, we find that the patch-level pre-training contributes the most improvement. It demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level. We then analyze the results from the model with different combinations of our pretraining stages. The model with image-level pre-training and patch-level pre-training (B5) performs better than the model that only adopts patch-level pre-training (B3), which shows that the global information provided by image-level"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.076,
+                0.803,
+                0.47,
+                0.901
+            ],
+            "angle": 0,
+            "content": "2Note that the original backbone networks reported in their papers are ResNext-101 [35]. However, the model weights of ResNext-101 pre-trained by our baseline SSL frameworks are not available in mmselfsup [6]. Thus, we switch the backbone network of these two methods to ResNet-50 [17], which is a common backbone network used in evaluating the performances of SSL frameworks. We also directly use the released ResNet-50 SSL pre-trained weights in our experiments for fair comparisons."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "12232"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.108,
+                0.104,
+                0.216,
+                0.208
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.142,
+                0.217,
+                0.179,
+                0.23
+            ],
+            "angle": 0,
+            "content": "Image"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.218,
+                0.104,
+                0.324,
+                0.208
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.238,
+                0.211,
+                0.304,
+                0.235
+            ],
+            "angle": 0,
+            "content": "Image-level Pretraining"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.326,
+                0.104,
+                0.434,
+                0.208
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.349,
+                0.211,
+                0.41,
+                0.235
+            ],
+            "angle": 0,
+            "content": "Patch-level Pretraining"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.436,
+                0.104,
+                0.542,
+                0.208
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.459,
+                0.211,
+                0.517,
+                0.235
+            ],
+            "angle": 0,
+            "content": "Pixel-level Pretraining"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.543,
+                0.104,
+                0.65,
+                0.208
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.542,
+                0.211,
+                0.65,
+                0.235
+            ],
+            "angle": 0,
+            "content": "(Image+Patch)-level Pretraining"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.652,
+                0.104,
+                0.758,
+                0.208
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.691,
+                0.218,
+                0.719,
+                0.228
+            ],
+            "angle": 0,
+            "content": "Ours"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.76,
+                0.104,
+                0.868,
+                0.208
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.803,
+                0.218,
+                0.823,
+                0.227
+            ],
+            "angle": 0,
+            "content": "GT"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.269,
+                0.247,
+                0.699,
+                0.262
+            ],
+            "angle": 0,
+            "content": "Figure 4. Visualization of the effects of the proposed pre-training stages."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.286,
+                0.473,
+                0.543
+            ],
+            "angle": 0,
+            "content": "pre-training can benefit the model with patch-level pretraining. Similarly, as shown in B7, B4 (i.e., only pixel-level pretraining) can boost the performance of the pre-trained model after image-level pre-training. These two comparisons indicate that image-level pre-training should be used with other pre-training stages to benefit the overall performance and it should not be solely adopted to our SSL pretraining framework. Besides, we find that combining patch-level and pixel-level pre-training (B6) has superior performance when compared with B3, especially on MAE with about \\(23.4\\%\\) improvement. This combination also outperforms supervised ImageNet pre-training, according to the results from Table 2. We attribute its superior performance to the pixel-level representation of mirror reflection. Our final model with all pre-training stages outperforms other ablation models, showing the effectiveness of adopting pretraining stages progressively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.55,
+                0.47,
+                0.807
+            ],
+            "angle": 0,
+            "content": "Visual Comparison of Our Pre-training Stages. Figure 4 shows a visual example of our ablation study. We can see that only adopting image-level pre-training on our SSL pre-training framework can approximately locate the mirror region but fails to produce a precise boundary for the predicted mask of the mirror. Patch-level pre-training can help predict mirrors more accurately with clearer boundaries but over-predict the non-mirror regions as mirrors. In comparison to predictions from all other ablation models, the prediction from the model with pixel-level pretraining has the clearest boundary but contains the largest over-predicted regions, primarily because of the insufficient global-level learning on mirrors. Our framework can significantly reduce over-predictions when image-level and patch-level pre-training are used, as opposed to when patch-level pre-training is used only, and performs the best with the help of the global-to-local pre-training process."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.817,
+                0.203,
+                0.832
+            ],
+            "angle": 0,
+            "content": "4.4. Discussions"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.84,
+                0.47,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Impacts of Pre-training Strategy. Some previous SSL works [42, 45], especially for those focusing on self-supervised pre-training for downstream tasks like segmentation [42] and salient object detection [45], would adopt"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.499,
+                0.282,
+                0.892,
+                0.309
+            ],
+            "angle": 0,
+            "content": "Table 4. The results of MirrorNet [37] when using different pretraining strategies. Best results are shown in bold."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.564,
+                0.309,
+                0.83,
+                0.407
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Fβ↑</td><td>MAE↓</td></tr><tr><td>Random Initialized</td><td>0.586</td><td>0.200</td></tr><tr><td>ImageNet Supervised</td><td>0.727</td><td>0.170</td></tr><tr><td>Ours (In parallel)</td><td>0.708</td><td>0.331</td></tr><tr><td>Ours (Sequential)</td><td>0.763</td><td>0.116</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.436,
+                0.892,
+                0.679
+            ],
+            "angle": 0,
+            "content": "different pretext tasks sequentially to pre-train their backbone networks. We also adopt this common strategy in our proposed SSL framework. To verify if employing different pre-training strategies may affect the performance of the proposed SSL framework, we have tried training the three stages sequentially as well as in parallel. We find that when pre-training the three stages in parallel, the backbone network fails to converge and its final fine-tuning results are much worse than those using sequential pre-training. Table 4 shows the results of this experiment. We can see that the MAE performance of pre-training the different stages in parallel is even worse than that of random initialization. A possible reason for this is that directly pre-training different pretext tasks in parallel is difficult to transfer the learned knowledge from the pre-trained network to the target downstream task (i.e., pixel-level mirror detection)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.681,
+                0.892,
+                0.862
+            ],
+            "angle": 0,
+            "content": "Apart from the superior performance made by the sequential pre-training strategy, we also notice that the number of pre-training iterations for each pretext task can heavily affect the performance of our SSL framework. To avoid heavy hyperparameter tuning and increase the reproducibility of our pre-training framework, we adopt the same number of pre-training iterations for all stages. Although this strategy of using the same number of pre-training iterations is likely not the optimal design for individual stages, we believe that it can better reflect the real performance of our framework with less stochasticity and more reactivity. We provide a more detailed discussion on this issue next."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.871,
+                0.892,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Impacts of Pre-training Iterations. SSL pre-training frameworks usually require huge pre-training iterations to"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "12233"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.134,
+                0.091,
+                0.411,
+                0.276
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.282,
+                0.47,
+                0.351
+            ],
+            "angle": 0,
+            "content": "Figure 5. Ablation study on the number of pre-training iterations for our SSL pre-training framework. We find that our SSL pretraining framework continues to improve in performance until \\(\\sim 20,000\\) pre-training iterations. After that, the performance starts to decline, mainly due to overfitting."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.373,
+                0.468,
+                0.509
+            ],
+            "angle": 0,
+            "content": "converge. For example, SimSiam [5] requires about 250K iterations to pre-train their model on ImageNet, which is very time-consuming. To validate how this issue may affect our SSL pre-training framework, we conduct the ablation study on the number of pre-training iterations for our SSL pre-training framework, and the results are shown in Figure 5. Note that we use the same number of iterations for all pre-training stages for convenience and to get rid of heavy hyperparameter tuning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.512,
+                0.468,
+                0.786
+            ],
+            "angle": 0,
+            "content": "From the results, we find that the performance of our SSL pre-training framework improves rapidly as the number of pre-training iterations increases, up to around 20,000, which is a relatively small number compared with those used in existing SSL pre-training frameworks. The performance of our SSL pre-training framework gradually drops as the number of pre-training iterations exceeds 20,000, mainly due to overfitting. One possible explanation of this is that the model pre-trained by our SSL pre-training framework can efficiently learn the representation of mirror reflection under a short pre-training period. Another possible explanation is that unlike generic SSL pre-training frameworks that attempt to extract general-propose image representation, our SSL pre-training framework is developed specially for a specific task, i.e., mirror detection. While generic SSL pre-training may cause feature redundancy for our task, having too many pre-training iterations (i.e., over 20,000) causes overfitting."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.795,
+                0.468,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Failure Cases. Our SSL pre-training framework does have limitations. It continues to struggle with the limitations of the base method [37], as the base method adopted can also significantly affect the final detection performances. For example, as shown in first row of Figure 6, MirrorNet with our SSL pre-training may still fail to handle some mirror-like regions (e.g., the top-right region). It may also fail to de"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.513,
+                0.088,
+                0.637,
+                0.166
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.638,
+                0.089,
+                0.761,
+                0.167
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.762,
+                0.089,
+                0.885,
+                0.167
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.512,
+                0.168,
+                0.636,
+                0.288
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.555,
+                0.293,
+                0.59,
+                0.306
+            ],
+            "angle": 0,
+            "content": "Image"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.638,
+                0.168,
+                0.761,
+                0.288
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.685,
+                0.293,
+                0.712,
+                0.305
+            ],
+            "angle": 0,
+            "content": "Ours"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.762,
+                0.168,
+                0.885,
+                0.288
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.814,
+                0.293,
+                0.834,
+                0.304
+            ],
+            "angle": 0,
+            "content": "GT"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.499,
+                0.318,
+                0.892,
+                0.415
+            ],
+            "angle": 0,
+            "content": "Figure 6. Failure cases of our SSL pre-training framework based on MirrorNet. Our SSL pre-training framework fails to tackle the intrinsic limitation of existing mirror detection methods. The model pre-trained by our SSL pre-training framework still has wrong predictions when the input images are challenging (e.g., containing mirror-like regions in the first row, and insufficient relationship between mirror/non-mirror regions in the second row."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.443,
+                0.892,
+                0.504
+            ],
+            "angle": 0,
+            "content": "tect mirrors in some challenging cases when the input image contains ambiguous mirrors without the relationship of mirror reflection and sufficient contextual contrast, as shown in the second row of Figure 6."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.522,
+                0.618,
+                0.537
+            ],
+            "angle": 0,
+            "content": "5. Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.548,
+                0.892,
+                0.836
+            ],
+            "angle": 0,
+            "content": "In this paper, we have investigated how self-supervised learning (SSL) works with the mirror detection task. To the best of our knowledge, we are the first to explore SSL pre-training frameworks applied to mirror detection. We have found that the supervised ImageNet pre-training might not be the ideal way to extract backbone image features for mirror detection due to the discrepancy between general-propose representation and mirror-specific representation. We have also proposed a new SSL pre-training framework to pre-train the backbone network for mirror detection. Our SSL pre-training framework does not require any labeled data. It progressively (image-, patch- and pixel-level pre-training) models the relationship of mirror reflection, and obtains a better representation of mirrors in the pre-training stage. Experimental results show that the proposed SSL pre-training framework outperforms CNN-based state-of-the-art SSL pre-training frameworks, and even achieves better results compared with supervised ImageNet pre-training."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.84,
+                0.892,
+                0.9
+            ],
+            "angle": 0,
+            "content": "Acknowledgments: The work described in this paper was partially supported by a GRF grant from the Research Grants Council of Hong Kong (Project No.: CityU 11211223)."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "12234"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.08,
+                0.09,
+                0.174,
+                0.106
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.117,
+                0.47,
+                0.186
+            ],
+            "angle": 0,
+            "content": "[1] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.188,
+                0.47,
+                0.231
+            ],
+            "angle": 0,
+            "content": "[2] M. Braun, S. Krebs, F. Flohr, and D. M. Gavrila. Eurocity persons: A novel benchmark for person detection in traffic scenes. IEEE TPAMI, 41(8):1844-1861, Aug 2019."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.233,
+                0.47,
+                0.288
+            ],
+            "angle": 0,
+            "content": "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.291,
+                0.47,
+                0.332
+            ],
+            "angle": 0,
+            "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.334,
+                0.47,
+                0.363
+            ],
+            "angle": 0,
+            "content": "[5] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.365,
+                0.47,
+                0.407
+            ],
+            "angle": 0,
+            "content": "[6] MMSelfSup Contributors. MMSelfSup: Openmmlab self-supervised learning toolbox and benchmark. https://github.com/open-mmlab/mmselfsup, 2021."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.409,
+                0.47,
+                0.478
+            ],
+            "angle": 0,
+            "content": "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.481,
+                0.47,
+                0.523
+            ],
+            "angle": 0,
+            "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.525,
+                0.47,
+                0.567
+            ],
+            "angle": 0,
+            "content": "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.569,
+                0.469,
+                0.61
+            ],
+            "angle": 0,
+            "content": "[10] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.613,
+                0.47,
+                0.669
+            ],
+            "angle": 0,
+            "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.671,
+                0.469,
+                0.713
+            ],
+            "angle": 0,
+            "content": "[12] Christopher Funk and Yanxi Liu. Beyond planar symmetry: Modeling human perception of reflection and rotation symmetries in the wild. In ICCV, pages 793-803, 2017."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.715,
+                0.47,
+                0.785
+            ],
+            "angle": 0,
+            "content": "[13] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.787,
+                0.47,
+                0.857
+            ],
+            "angle": 0,
+            "content": "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.859,
+                0.47,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[15] Huankang Guan, Jiaying Lin, and Rynson W.H. Lau. Learning semantic associations for mirror detection. In CVPR, pages 5941-5950, June 2022."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.117,
+                0.47,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.093,
+                0.892,
+                0.135
+            ],
+            "angle": 0,
+            "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.136,
+                0.892,
+                0.178
+            ],
+            "angle": 0,
+            "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.179,
+                0.892,
+                0.234
+            ],
+            "angle": 0,
+            "content": "[18] Ruozhen He, Jiaying Lin, and Rynson WH Lau. Efficient mirror detection via multi-level heterogeneous learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 790-798, 2023."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.235,
+                0.892,
+                0.303
+            ],
+            "angle": 0,
+            "content": "[19] Tianyu Huang, Bowen Dong, Jiaying Lin, Xiaohui Liu, Rynson WH Lau, and Wangmeng Zuo. Symmetry-aware transformer-based mirror detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 935–943, 2023."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.305,
+                0.892,
+                0.359
+            ],
+            "angle": 0,
+            "content": "[20] Peter Kohler, Alexandra Yakovleva, Alasdair Clarke, Yanxi Liu, and Anthony Norcia. Parametric responses to rotation symmetry in mid-level visual cortex. Journal of Vision, 15(12):1122-1122, 2015."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.361,
+                0.892,
+                0.402
+            ],
+            "angle": 0,
+            "content": "[21] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In ICLR, 2018."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.404,
+                0.892,
+                0.46
+            ],
+            "angle": 0,
+            "content": "[22] Jiaying Lin, Xin Tan, and Rynson WH Lau. Learning to detect mirrors from videos via dual correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9109-9118, 2023."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.461,
+                0.892,
+                0.488
+            ],
+            "angle": 0,
+            "content": "[23] Jiaying Lin, Guodong Wang, and Rynson W. H. Lau. Progressive mirror detection. In CVPR, 2020."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.49,
+                0.892,
+                0.546
+            ],
+            "angle": 0,
+            "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.547,
+                0.892,
+                0.602
+            ],
+            "angle": 0,
+            "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.603,
+                0.892,
+                0.631
+            ],
+            "angle": 0,
+            "content": "[26] Zhiqiu Lin, Jin Sun, Abe Davis, and Noah Snavely. Visual chirality. In CVPR, pages 12295-12303, 2020."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.632,
+                0.892,
+                0.659
+            ],
+            "angle": 0,
+            "content": "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.661,
+                0.892,
+                0.729
+            ],
+            "angle": 0,
+            "content": "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.731,
+                0.892,
+                0.773
+            ],
+            "angle": 0,
+            "content": "[29] Jiaqi Tan, Weijie Lin, Angel X Chang, and Manolis Savva. Mirror3d: Depth refinement for mirror surfaces. In CVPR, pages 15990-15999, 2021."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.774,
+                0.892,
+                0.829
+            ],
+            "angle": 0,
+            "content": "[30] Xin Tan, Jiaying Lin, Ke Xu, Chen Pan, Lizhuang Ma, and Rynson W. H. Lau. Mirror detection with the visual chirality cue. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.831,
+                0.892,
+                0.859
+            ],
+            "angle": 0,
+            "content": "[31] Christopher W Tyler. Human symmetry perception and its computational analysis. Psychology Press, 2003."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.86,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[32] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, 2021."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.503,
+                0.093,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "12235"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.092,
+                0.468,
+                0.133
+            ],
+            "angle": 0,
+            "content": "[33] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.136,
+                0.468,
+                0.177
+            ],
+            "angle": 0,
+            "content": "[34] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, pages 418-434, 2018."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.179,
+                0.468,
+                0.22
+            ],
+            "angle": 0,
+            "content": "[35] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 1492-1500, 2017."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.222,
+                0.468,
+                0.275
+            ],
+            "angle": 0,
+            "content": "[36] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.278,
+                0.468,
+                0.305
+            ],
+            "angle": 0,
+            "content": "[37] Xin Yang, Haiyang Mei, Ke Xu, Xiaopeng Wei, Baocai Yin, and Rynson W.H. Lau. Where is my mirror? In ICCV, 2019."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.307,
+                0.468,
+                0.348
+            ],
+            "angle": 0,
+            "content": "[38] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Learning with noisy labels for robust point cloud segmentation. In ICCV, pages 6443-6452, 2021."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.35,
+                0.468,
+                0.404
+            ],
+            "angle": 0,
+            "content": "[39] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Robust point cloud segmentation with noisy annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7696-7710, 2022."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.407,
+                0.468,
+                0.46
+            ],
+            "angle": 0,
+            "content": "[40] Shuquan Ye, Yujia Xie, Dongdong Chen, Yichong Xu, Lu Yuan, Chenguang Zhu, and Jing Liao. Improving commonsense in vision-language models via knowledge graph riddles. In CVPR, pages 2634-2645, 2023."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.463,
+                0.468,
+                0.503
+            ],
+            "angle": 0,
+            "content": "[41] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.506,
+                0.468,
+                0.546
+            ],
+            "angle": 0,
+            "content": "[42] Xiaohang Zhan, Ziwei Liu, Ping Luo, Xiaou Tang, and Chen Loy. Mix-and-match tuning for self-supervised semantic segmentation. In AAAI, volume 32, 2018."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.549,
+                0.468,
+                0.589
+            ],
+            "angle": 0,
+            "content": "[43] Xiaohang Zhan, Jiahao Xie, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Online deep clustering for unsupervised representation learning. In CVPR, 2020."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.592,
+                0.468,
+                0.632
+            ],
+            "angle": 0,
+            "content": "[44] Nanxuan Zhao, Zhirong Wu, Rynson WH Lau, and Stephen Lin. What makes instance discrimination good for transfer learning? In ICLR, 2021."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.635,
+                0.468,
+                0.674
+            ],
+            "angle": 0,
+            "content": "[45] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, , Hutchuan Lu, and Xiang Ruan. Self-supervised pretraining for rgb-d salient object detection. In AAAI, 2022."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.677,
+                0.468,
+                0.731
+            ],
+            "angle": 0,
+            "content": "[46] Ying Zheng, Yiyi Zhang, Xiaogang Xu, Jun Wang, and Hongxun Yao. Visual chirality meets freehand sketches. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1544-1548. IEEE, 2021."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.092,
+                0.468,
+                0.731
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.946,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "12236"
+        }
+    ]
+]

2023/Self-supervised Pre-training for Mirror Detection/9faec277-19c4-4583-a3c8-84bf53084f2f_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a188edcbb1d8c7cbd200d7f57d0fd77c56d829b51bf4014d19bce8299d48cedb
+size 1583664

2023/Self-supervised Pre-training for Mirror Detection/full.md

@@ -0,0 +1,303 @@
+# Self-supervised Pre-training for Mirror Detection
+
+Jiaying Lin* Rynson W.H. Lau* City University of Hong Kong
+
+{jiayinlin5, rynson.lau}@cityu.edu.hk
+
+# Abstract
+
+Existing mirror detection methods require supervised ImageNet pre-training to obtain good general-purpose image features. However, supervised ImageNet pre-training focuses on category-level discrimination and may not be suitable for downstream tasks like mirror detection, due to the overfitting upstream tasks (e.g., supervised image classification). We observe that mirror reflection is crucial to how people perceive the presence of mirrors, and such mid-level features can be better transferred from self-supervised pretrained models. Inspired by this observation, in this paper we aim to improve mirror detection methods by proposing a new self-supervised learning (SSL) pre-training framework for modeling the representation of mirror reflection progressively in the pre-training process. Our framework consists of three pre-training stages at different levels: 1) an image-level pre-training stage to globally incorporate mirror reflection features into the pre-trained model; 2) a patch-level pre-training stage to spatially simulate and learn local mirror reflection from image patches; and 3) a pixel-level pre-training stage to pixel-wisely capture mirror reflection via reconstructing corrupted mirror images based on the relationship between the inside and outside of mirrors. Extensive experiments show that our SSL pre-training framework significantly outperforms previous state-of-the-art CNN-based SSL pre-training frameworks and even outperforms supervised ImageNet pre-training when transferred to the mirror detection task. Code and models are available at https://jiaying.link/iccv2023-sslmirror/
+
+# 1. Introduction
+
+Mirrors are prevalent in our daily lives. As their appearances are largely determined by their surroundings, they generally lack a consistent appearance, making it difficult to separate them from their surroundings. This may affect many computer vision tasks such as object detection [2], vision-language navigation [1] and depth estimation [29].
+
+![](images/2053fb7e33566aa3d281812f2faf69bdad556d6356cac893b22a3c6a21bc08bc.jpg)
+
+![](images/97624ca6d1125c6f4390b003b014011a351f9c1950c7ccbc8c0f5ed0ce450959.jpg)
+
+![](images/a3a1688cab67f58dac910e5e2accc57c4180feddaabc8e18cc4be9edbfbf4b69.jpg)  
+(b) MirrorNet on ImageNet
+
+![](images/9c79b20acd129e22e321256d41d612df587f83459c20148214c01a2b2a65136b.jpg)
+
+![](images/21f231f0b30b96f0819c87b4f82bdc5afed12d33bcd81d816967afb479a37ca7.jpg)  
+(c)VCNet on ImageNet
+
+![](images/2407085591717f6000629a100ebbac3458bd38d6891f0485e3353e069cd334da.jpg)
+
+![](images/1d85123e767780ec94c3702dff904630533d38df76179b68b715bd5920b81888.jpg)  
+Figure 1. State-of-the-art mirror detection methods [37, 30] are based on costly supervised ImageNet pre-training. They may fail even in obvious cases, e.g., (a) when the mirror clearly reflects a real object outside of the mirror. (b) and (c) are MirrorNet [37] and VCNet [30], respectively, pre-trained on ImageNet with full supervision. (c) is MirrorNet with our proposed SSL framework and without supervised ImageNet pre-training. Our SSL scheme leverages the mirror reflection cue and avoids feature redundancy in the pre-training stage. It outperforms those pre-trained on ImageNet with full supervision (i.e., (b) and (c)).
+
+![](images/172cb65753d1987a963d6fe5c399475d5ab9a638956f1ba00e1256d1d7e58382.jpg)
+
+![](images/db1d2d05c41cec15303148981a731908325a4eedbf1269610181f34c7e074c40.jpg)  
+1) Ours (MirrorNet on our SSL)  
+(e) GT
+
+It is therefore crucial to design high-performance mirror detection methods to facilitate computer vision applications.
+
+Recently, a few methods [37, 23, 30] have been proposed for mirror detection, but they all require to be initialized with supervised ImageNet [8] pre-trained weights and then fine-tuned on mirror detection datasets. While such a transfer learning approach is common in general computer vision tasks (e.g., object detection [24] and semantic segmentation [7]) to utilize category-oriented features from large-scale image datasets, we challenge if this strategy is necessary for the mirror detection task for two reasons. First, the supervised ImageNet pre-training process involves high labelling costs, as it is very time-consuming and labor-intensive to label such a large-scale image dataset ( $\sim$ 1.3M images). Second, unlike general vision tasks, mirror detection does not require category-level object understanding. Instead, it requires an understanding of the relationship between inside and outside of mirrors (e.g., contrast [37], similarity [23], and visual chirality [30]). Thus,
+
+while pre-training on ImageNet requires expensive labels, it causes feature redundancy for the mirror detection task.
+
+Figure 1 shows that existing models pre-trained on ImageNet tend to over-detect the mirror regions. Although the two SOTA methods [37, 30] can correctly locate the mirror regions, they are unable to distinguish between the mirrored objects inside the mirrors and the real objects outside, even though these methods contain well-designed modules to learn to address this problem via the fine-tuning stage. This motivates us to investigate whether the current pre-training approach may affect the final detection performance, and whether we may incorporate some intrinsic mirror properties in the pre-training process to assist mirror detection.
+
+An interesting observation is that humans do not need to be "trained" to recognize mirrors. Instead, we learn to recognize them implicitly from young. This aligns with the key idea of self-supervised learning (SSL). Works from neuroscience [31, 20] have shown that humans use mid-level visual cortex to recognize mirror reflection. According to [44], mid-level features are hard to be transferred from supervised pre-training on ImageNet, and SSL is suitable for learning mid-level representations, which inspires us to develop an SSL pre-training framework for modeling mirror reflection.
+
+In this paper, we propose a new SSL pre-training framework for mirror detection, which explicitly considers mirror reflection during the pre-training process. Our framework does not require human-annotated labels from large-scale image datasets. It consists of three stages to progressively pre-train the backbone network from global to local: 1) an image-level pre-training stage to obtain the representation of mirror reflection globally by recognizing the geometric transformation applied to the image; 2) a patch-level pretraining stage to mimic patch-wise mirror reflection and then learn the spatial correlation between the original object patches and the corresponding mirrored object patches; and 3) a pixel-level pre-training stage to extract the pixel-to-pixel relationship of mirror reflection by image reconstruction; Under such progressive pre-training scheme, our SSL pre-training framework can learn the representation of mirror reflection, and then effectively transfer this knowledge to the subsequent mirror detection process for better detection performances. We conduct comprehensive experiments to demonstrate the effectiveness of our SSL pretraining framework. We show that it can significantly boost the performances of existing mirror detection methods, and outperform other CNN-based SSL pre-training frameworks on the mirror detection task.
+
+To conclude, this paper makes three key contributions:
+
+- To the best of our knowledge, we are the first to investigate how existing SSL pre-training frameworks perform on the mirror detection task, compared with supervised ImageNet pre-training.
+
+- We propose a new SSL pre-training framework that consists of three stages at different levels to progressively learn the representation of mirror reflection. Compared with the features from a supervised ImageNet pre-trained model, our representation is better due to the reduced gap between the pre-training task and the target downstream task (i.e., mirror detection).  
+- Extensive experiments show that our SSL pre-training framework performs the best among all state-of-the-art SSL methods, and even better than models with supervised ImageNet pre-training.
+
+# 2. Related Work
+
+# 2.1. Mirror Detection
+
+Mirror detection is an essential task in computer vision, as mirrors are everywhere these days. It is also a challenging task, since mirrors do not possess a consistent appearance, but reflect those of their surroundings. Other vision tasks such as generic object detection and segmentation would tend to detect and segment objects outside as well as inside of the mirror, providing incorrect information on the mirror region, and a depth prediction task may return the depths of the reflected objects inside the mirror region instead of the depth of the mirror.
+
+To address the mirror detection problem, Yang et al. [37] propose the first large-scale mirror detection dataset and the first deep-learning model, MirrorNet, for mirror detection by learning the contextual contrast between the inside and outside of mirrors. Lin et al. [23] further propose a new model with similarity and edge learning to progressively detect mirrors. Most recently, Guan et al. [15] employ graph convolutional networks to model contextual associations for mirror detection, while Tan et al. propose VCNet [30] to exploit the visual chirality cue at a feature level.
+
+However, existing mirror detection methods are all fully-supervised. They also directly adopt the supervised ImageNet [8] pre-training to obtain generic image features. Unlike recent works [19, 18, 22] that propose novel network architectures for mirror detection, we propose in this paper to tackle the mirror detection problem from a learning-based perspective, offering the first self-supervised pre-training framework for mirror detection.
+
+# 2.2. Self-supervised Learning (SSL)
+
+SSL is a popular research topic in computer vision [38, 39], and has attracted a lot of research interests. It aims to learn good feature representations without needing labeled data. Early SSL methods focus on proposing novel pretext tasks, such as relative location prediction [10] and rotation prediction [21]. Recently, most methods are based on the utilization of contrastive learning [16, 4, 14, 3, 32, 5, 41,
+
+![](images/908689b51dc7553c0ccc26fde83f0bd65eb6d802e9a616f0f113c004d699d59b.jpg)
+
+![](images/855243fd694ad6ad2d7363759b0686b9465af05eab1878a3b52b8f769d8e9a32.jpg)
+
+![](images/d955a6d1249ed61c65caf126c7f5730a39a2581f3573b017e7bde9fc5f13c937.jpg)  
+Figure 2. Our SSL pre-training framework. It consists of three pre-training stages at different levels: an image-level pre-training stage, a patch-level pre-training stage, and a pixel-level pre-training stage. It pre-trains the backbone network from global to local (image-level  $\rightarrow$  patch-level  $\rightarrow$  pixel-level) progressively.
+
+40]. Most recently, [44] shows that some SSL methods perform even better than supervised ImageNet pre-training in downstream tasks, such as object detection [11, 24] and semantic segmentation [7, 34], with the help of instance discrimination during SSL pre-training. Recent transformer-based SSL frameworks [36] also involve masked image modeling to obtain better self-supervised representations.
+
+In this paper, we observe that unlike general vision tasks, mirror detection is not benefited from supervised ImageNet pre-training since it does not require category-level object understanding. Hence, we propose a novel SSL pre-training framework specifically for the mirror detection task. The proposed framework aims to simulate and learn the relationship between mirrors and their surrounding during the pre-training stage.
+
+# 3. Method
+
+Figure 2 shows the overall architecture of the proposed SSL pre-training framework. Our SSL pre-training framework consists of three pre-training stages at different levels: an image-level pre-training, a patch-level pre-training, and a pixel-level pre-training. The pre-training at each level is conducted sequentially and independently. The progressive and sequential training strategy can help the pre-trained model obtain a global-to-local representation of mirrors.
+
+# 3.1. Image-level Pre-training
+
+Inspired by the definition of visual chirality [26] and its potential applications [46, 30] in various computer vision problems, we formulate the image-level pre-training as a binary classification problem: Given an input image and its
+
+corresponding horizontal flipped image, can the backbone network tell which one is the original image? To achieve this objective, we attach a classification head to predict the probability that the input image is flipped. The classification head consists of a convolution layer followed by a sigmoid activation function to project the output value into  $[0,1]$ . In this pre-training stage, we use binary cross-entropy (BCE) as our loss function  $\mathcal{L}_f$  to optimize the backbone network. Formally, we have:
+
+$$
+\mathcal {L} _ {f} = - \left[ y \log (\hat {y}) + (1 - y) \log (1 - \hat {y}) \right], \tag {1}
+$$
+
+where  $y$  is the ground truth label and  $\hat{y}$  is the predicted probability that the input image is horizontally flipped.
+
+# 3.2. Patch-level Pre-training
+
+The image-level pre-training pretext task alone cannot bring much help in locating the mirror inside an image since mirror detection is a pixel-level problem. However, directly connecting the image-level and the pixel-level pretext tasks is challenging due to the gap between the scales of these two pretext tasks. Thus, we design a patch-level pre-training stage as a bridge to cover the feature representations of mirrors at different scales. Besides, the image-level pre-training does not benefit relational spatial understanding, which has been proven important in mirror detection [23]. Our patch-level pre-training consists of two sub-tasks to both obtain the patch-wise relationship of mirror reflection and simulate the mirrored region with pseudo labels. We first split the input image into a grid of  $3 \times 3$  image patches and randomly shuffle them. After that, we randomly apply a flipping operation to one of the image patches, to simulate a mirrored region. The first sub-task
+
+is to predict the location index from the shuffled grid by the backbone network. We append a location predictor head to the backbone network to predict the location index of input patches. We apply cross-entropy loss as  $\mathcal{L}_{loc}$  to optimize this sub-task. The second sub-task is to predict the flipped patch from the nine randomly shuffled image patches. Similar to our image-level pre-training, we use BCE loss as our loss function  $\mathcal{L}_{cls}$  in this sub-task.
+
+$$
+\mathcal {L} _ {l o c} = - \frac {1}{N} \sum_ {i = 1} ^ {N} y _ {i} \log \left(\hat {y} _ {i}\right), \tag {2}
+$$
+
+$$
+\mathcal {L} _ {c l s} = - \frac {1}{N} \sum_ {j = 1} ^ {N} y _ {j} \log \left(\hat {y _ {j}}\right) + (1 - y _ {j}) \log \left(1 - \hat {y _ {j}}\right), \tag {3}
+$$
+
+where  $\hat{y}_i$  is the predicted location index, and  $y_i$  is the corresponding ground truth label.  $\hat{y}_j$  is the predicted probability that the input patch is horizontally flipped and  $y_j$  is the corresponding ground truth probability. Finally, we jointly optimize these two sub-tasks and the final loss function is:
+
+$$
+\mathcal {L} _ {\text {p a t c h}} = \mathcal {L} _ {\text {l o c}} + \mathcal {L} _ {\text {c l s}}. \tag {4}
+$$
+
+# 3.3. Pixel-level Pre-training
+
+Extracting the pixel-level relationship of mirror reflection is a challenging task, as studied by previous related work [12] for symmetry detection. It requires point-level GT labels to represent symmetry centers. However, obtaining such labels requires huge human annotation effort and they are not applicable for SSL pre-training. Instead of generating pseudo labels for symmetry detection, our solution is that we formulate this task as an image reconstruction problem on images containing mirrors. An ideal pre-trained network should be able to reconstruct the input image containing mirrors with the pixel-level understanding of mirror reflection. We randomly mask out the input image by cutout augmentation [9] with the ratio of 0.3 and then feed it to the backbone network with a reconstruction head. The details of the reconstruction head we used are listed in Table 1. The objective of our pixel-level pre-training is defined as:
+
+$$
+\mathcal {L} _ {p i x} = \| x - \hat {x} \| _ {1}, \tag {5}
+$$
+
+where  $\mathcal{L}_{pix}$  is an L1 loss.  $x$  and  $\hat{x}$  are the input image and the reconstructed image, respectively.
+
+# 4. Experiments
+
+# 4.1. Experimental Settings
+
+Dataset and Evaluation Metrics. The image-level and patch-level pre-training experiments are conducted on the training set of ImageNet-100 [28] and MS COCO [25],
+
+Table 1. The architecture of the reconstruction head used in the pixel-level pre-training. Note that each "conv-IR" corresponds to a sequence of convolution layer, InstanceNorm layer and ReLU activation.  $K$ ,  $S$  and  $P$  denote the number of kernels, the number of strides and the padding size, respectively, used in the convolution layer.
+
+<table><tr><td>Layer Details</td></tr><tr><td>3 × 3 conv-IR, K = 1024, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>3 × 3 conv-IR, K = 512, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>3 × 3 conv-IR, K = 256, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>3 × 3 conv-IR, K = 128, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>3 × 3 conv-IR, K = 128, S = 1, P = 1</td></tr><tr><td>2 × Upsample</td></tr><tr><td>1 × 1 conv-IR, K = 3, S = 1, P = 1</td></tr></table>
+
+which contains  $130\mathrm{K}$  and  $\sim 118\mathrm{k}$  images respectively. In the pixel-level pre-training and fine-tuning stage, our experiments are conducted on MSD [37], the first large-scale dataset for mirror detection. MSD [37] provides 4,018 mirror images, which are divided into 3,063 images for training and 955 images for testing. We adopt two popular metrics namely, F-measure  $(\mathbf{F}_{\beta})$  and mean absolute error (MAE) to evaluate the performance of our SSL pre-training framework. F-measure can reflect the overall model performance and is calculated by the weighted harmonic mean of the precision and recall:  $F_{\beta} = \frac{(1 + \beta^{2})\cdot\text{precision}\cdot\text{recall}}{\beta^{2}\cdot\text{precision} + \text{recall}}$ , where  $\beta^2$  is set to 0.3. MAE can measure the average pixel-wise absolute disparity between the predicted mirror maps  $M$  and the binary ground truth  $G$ .
+
+Implementation Details. We use Pytorch to implement our proposed SSL pre-training framework. All experiments are conducted on a GPU server with eight NVIDIA RTX 3090 GPUs. Following existing SSL pre-training frameworks, we adopt ResNet-50 [17] as our backbone network. All pre-trained model weights of existing SSL pretraining frameworks are obtained from the released models of mmselfsup [6] for fair comparisons. For each pretext task in the pre-training stage, all images are resized to  $256 \times 256$ . The batch size is set to 352. We use AdamW [27] with an initial learning rate of 1e-4 to optimize our SSL pre-training framework and stop pre-training after 20,000 iterations. We only adopt random color jittering for data augmentation. In the fine-tuning stage, we use the same hyperparameters reported in the original papers of the adopted
+
+![](images/fb1c2b369ed7aed8c0640a1bc2c105e7f93e01f49c8804770f246f9699221c5a.jpg)  
+Figure 3. Qualitative Comparison of our proposed method with NPID [33], MocoV2 [4], SwAV [3], Barlow Twins [41], ImageNet pretraining (Supervised). We use the released pre-trained models of these existing SSL methods from mmselfsup [6] for fair comparison.
+
+base methods MirrorNet [37] and VCNet [30].
+
+# 4.2. Comparison with State-of-the-Arts
+
+We compare against 11 state-of-the-art CNN-based frameworks in self-supervised learning, including Relative Location (Rel. Loc.) [10], Rotation Prediction (Rot. Pred.) [21], NPID [33], ODC [43], Moco V2 [16], SimCLR [4], BYOL [14], SwAV [3], DenseCL [32], SimSiam [5] and Barlow Twins (Bar. Twins) [41]. We select MirrorNet [37] and VCNet [30], which are two representative mirror detection methods, as base methods to conduct experiments in the fine-tuning stage.
+
+Qualitative Evaluation. We further show the performance of our SSL pre-training framework in Figure 3. We compare our framework with four state-of-the-art SSL pre-training
+
+frameworks (including the two best-performing frameworks NPID [33] and Moco V2 [16] based on MirrorNet and the two best-performing frameworks SwAV [3] and Barlow Twins [41] based on VCNet, according to their performance in Table 2) and supervised ImageNet pre-training due to the limitation of space. Our framework can correctly detect the mirror regions by utilizing the relationship of mirror reflection caused by mirrors, even though the reflected regions are pure texture (i.e., lack of object keypoints in the second row) and relatively small in the image (e.g., the last three rows). We attribute the superior performance of our framework to the utilization of mirror reflection in the pre-training stage.
+
+Quantitative Evaluation. Table 2 shows the experimentation results comparing the proposed SSL pre-training framework with the state-of-the-arts. Besides, we also re
+
+Table 2. Quantitative comparison between our SSL pre-training framework and existing CNN-based SSL methods. We also compare it with random initialization (Random) and ImageNet pretraining (Supervised).  
+
+<table><tr><td rowspan="2">Methods</td><td rowspan="2">Venue</td><td colspan="2">MirrorNet</td><td colspan="2">VCNet</td></tr><tr><td>Fβ↑</td><td>MAE↓</td><td>Fβ↑</td><td>MAE↓</td></tr><tr><td>Rel. Loc.</td><td>ICCV &#x27;15</td><td>0.582</td><td>0.254</td><td>0.842</td><td>0.072</td></tr><tr><td>Rot. Pred.</td><td>ICLR &#x27;18</td><td>0.537</td><td>0.212</td><td>0.851</td><td>0.069</td></tr><tr><td>NPID</td><td>CVPR &#x27;18</td><td>0.739</td><td>0.132</td><td>0.832</td><td>0.076</td></tr><tr><td>ODC</td><td>CVPR &#x27;20</td><td>0.558</td><td>0.223</td><td>0.836</td><td>0.077</td></tr><tr><td>Moco V2</td><td>CVPR &#x27;20</td><td>0.707</td><td>0.177</td><td>0.835</td><td>0.072</td></tr><tr><td>SimCLR</td><td>ICML &#x27;20</td><td>0.594</td><td>0.244</td><td>0.837</td><td>0.071</td></tr><tr><td>BYOL</td><td>NeurIPS &#x27;20</td><td>0.704</td><td>0.158</td><td>0.832</td><td>0.077</td></tr><tr><td>SwAV</td><td>NeurIPS &#x27;20</td><td>0.578</td><td>0.224</td><td>0.856</td><td>0.071</td></tr><tr><td>DenseCL</td><td>CVPR &#x27;21</td><td>0.694</td><td>0.240</td><td>0.844</td><td>0.072</td></tr><tr><td>SimSiam</td><td>CVPR &#x27;21</td><td>0.672</td><td>0.198</td><td>0.818</td><td>0.080</td></tr><tr><td>Bar. Twins</td><td>ICML &#x27;21</td><td>0.655</td><td>0.178</td><td>0.857</td><td>0.069</td></tr><tr><td>Random</td><td></td><td>0.586</td><td>0.200</td><td>0.400</td><td>0.103</td></tr><tr><td>Supervised2</td><td></td><td>0.727</td><td>0.170</td><td>0.871</td><td>0.062</td></tr><tr><td>Ours</td><td></td><td>0.763</td><td>0.116</td><td>0.886</td><td>0.057</td></tr></table>
+
+port the results from models adopting Xavier initialization [13] (Random) or supervised ImageNet pre-training (Supervised). Our SSL pre-training framework (Ours) achieves the best performance across both metrics when applied on MirrorNet and VCNet and shows a good generalization of different mirror detection methods. It is worth noting that our framework even outperforms the supervised ImageNet pre-training.
+
+# 4.3. Ablation Study
+
+Table 3 demonstrates the effectiveness of each pretraining stage in our framework. We analyze its results point by point as follows.
+
+Image-level Pre-training vs. Random Initialization. To evaluate the effectiveness of the proposed image-level pretraining, we compare the model only adopting image-level pre-training (B2) with the model initialized randomly without any pre-training (B1). We can see that only adopting image-level pre-training (B2) will even cause a performance drop compared with the one without pre-training (B1). This somehow supports the finding from VCNet [30] that image-level flipping cannot provide sufficient information to pixelwisely locate the mirror inside an image.
+
+Table 3. Ablation study of the proposed SSL pre-training framework. We use MirrorNet [37] trained on the MSD dataset as our base method to conduct the experiments in this ablation study. Best results are shown in bold.  
+
+<table><tr><td></td><td>Rand. Init.</td><td>Image-lev. Pretrain</td><td>Patch-lev. Pretrain</td><td>Pixel-lev. Pretrain</td><td>Fβ↑</td><td>MAE↓</td></tr><tr><td>B1</td><td>✓</td><td></td><td></td><td></td><td>0.601</td><td>0.195</td></tr><tr><td>B2</td><td></td><td>✓</td><td></td><td></td><td>0.623</td><td>0.191</td></tr><tr><td>B3</td><td></td><td></td><td>✓</td><td></td><td>0.689</td><td>0.162</td></tr><tr><td>B4</td><td></td><td></td><td></td><td>✓</td><td>0.620</td><td>0.198</td></tr><tr><td>B5</td><td></td><td>✓</td><td>✓</td><td></td><td>0.723</td><td>0.166</td></tr><tr><td>B6</td><td></td><td></td><td>✓</td><td>✓</td><td>0.758</td><td>0.124</td></tr><tr><td>B7</td><td></td><td>✓</td><td></td><td>✓</td><td>0.666</td><td>0.178</td></tr><tr><td>Ours</td><td></td><td>✓</td><td>✓</td><td>✓</td><td>0.763</td><td>0.116</td></tr></table>
+
+Patch-level Pre-training vs. Random Initialization. We compare our patch-level pre-training (B3) with random initialization (B1) to analyze if patch-level pre-training is useful in our SSL pre-training framework. Our patch-level pretraining outperforms the model without pre-training on both two metrics (14.64% improvement on  $F_{\beta}$  and 16.32% improvement on MAE), which indicates the effectiveness of our patch-level pre-training. It also demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level.
+
+Pixel-level Pre-training vs. Random Initialization. Similarly, we conduct the ablation experiment on our pixel-level pre-training. The model adopting pixel-level pre-training performs better than the model with image pre-training. However, the performance gain is not as much as the model with patch-level pre-training (B3). One possible reason is that directly optimizing the backbone network for the representation of mirrors at a pixel level is challenging and may not produce satisfactory results for mirror detection.
+
+Location Predictor in Patch-level Pre-training. The location predictor in our patch-level pre-training is designed to capture the relational spatial information, which has been proven important in mirror detection [23]. We conduct an ablation experiment to remove the location predictor and the performance of our methods drops from 0.763 to 0.731. This shows the importance of learning spatial information in the mirror detection task.
+
+Combinations of Different Pre-training Stages. Based on the above three comparisons, we find that the patch-level pre-training contributes the most improvement. It demonstrates the importance of modeling the spatial relationship of mirror reflection at a patch level. We then analyze the results from the model with different combinations of our pretraining stages. The model with image-level pre-training and patch-level pre-training (B5) performs better than the model that only adopts patch-level pre-training (B3), which shows that the global information provided by image-level
+
+![](images/c94e799004490cc55ca02611f6e1d3d0a18c07e817a95c713108ccb4eda94f38.jpg)  
+Image
+
+![](images/adb43e744b866dd7ca54d2d708480002a84938a4c0f338e7cd4663448de446fd.jpg)  
+Image-level Pretraining
+
+![](images/ef97824b33362b7b6edd9ca6e0ece5d9cda25eeb9aba72ad576370e87fd8e77c.jpg)  
+Patch-level Pretraining
+
+![](images/6155df1df3d0122b4be72a7609d83ada630161ef92e3acf130d06f647cb0f2fd.jpg)  
+Pixel-level Pretraining
+
+![](images/db77c22435167185c10f3cc363d5ff96d45d33e0f592430360e6dbc4e66205d5.jpg)  
+Figure 4. Visualization of the effects of the proposed pre-training stages.
+
+![](images/20de36026fd80a4e087ead7e427f1f5f25a35580a4efbb87fc1d63e9b09084d1.jpg)  
+(Image+Patch)-level Pretraining  
+Ours
+
+![](images/2247c9baf0804f7f3164b3547a535be04149800f888792e44bb69438800898a8.jpg)  
+GT
+
+pre-training can benefit the model with patch-level pretraining. Similarly, as shown in B7, B4 (i.e., only pixel-level pretraining) can boost the performance of the pre-trained model after image-level pre-training. These two comparisons indicate that image-level pre-training should be used with other pre-training stages to benefit the overall performance and it should not be solely adopted to our SSL pretraining framework. Besides, we find that combining patch-level and pixel-level pre-training (B6) has superior performance when compared with B3, especially on MAE with about  $23.4\%$  improvement. This combination also outperforms supervised ImageNet pre-training, according to the results from Table 2. We attribute its superior performance to the pixel-level representation of mirror reflection. Our final model with all pre-training stages outperforms other ablation models, showing the effectiveness of adopting pretraining stages progressively.
+
+Visual Comparison of Our Pre-training Stages. Figure 4 shows a visual example of our ablation study. We can see that only adopting image-level pre-training on our SSL pre-training framework can approximately locate the mirror region but fails to produce a precise boundary for the predicted mask of the mirror. Patch-level pre-training can help predict mirrors more accurately with clearer boundaries but over-predict the non-mirror regions as mirrors. In comparison to predictions from all other ablation models, the prediction from the model with pixel-level pretraining has the clearest boundary but contains the largest over-predicted regions, primarily because of the insufficient global-level learning on mirrors. Our framework can significantly reduce over-predictions when image-level and patch-level pre-training are used, as opposed to when patch-level pre-training is used only, and performs the best with the help of the global-to-local pre-training process.
+
+# 4.4. Discussions
+
+Impacts of Pre-training Strategy. Some previous SSL works [42, 45], especially for those focusing on self-supervised pre-training for downstream tasks like segmentation [42] and salient object detection [45], would adopt
+
+Table 4. The results of MirrorNet [37] when using different pretraining strategies. Best results are shown in bold.  
+
+<table><tr><td></td><td>Fβ↑</td><td>MAE↓</td></tr><tr><td>Random Initialized</td><td>0.586</td><td>0.200</td></tr><tr><td>ImageNet Supervised</td><td>0.727</td><td>0.170</td></tr><tr><td>Ours (In parallel)</td><td>0.708</td><td>0.331</td></tr><tr><td>Ours (Sequential)</td><td>0.763</td><td>0.116</td></tr></table>
+
+different pretext tasks sequentially to pre-train their backbone networks. We also adopt this common strategy in our proposed SSL framework. To verify if employing different pre-training strategies may affect the performance of the proposed SSL framework, we have tried training the three stages sequentially as well as in parallel. We find that when pre-training the three stages in parallel, the backbone network fails to converge and its final fine-tuning results are much worse than those using sequential pre-training. Table 4 shows the results of this experiment. We can see that the MAE performance of pre-training the different stages in parallel is even worse than that of random initialization. A possible reason for this is that directly pre-training different pretext tasks in parallel is difficult to transfer the learned knowledge from the pre-trained network to the target downstream task (i.e., pixel-level mirror detection).
+
+Apart from the superior performance made by the sequential pre-training strategy, we also notice that the number of pre-training iterations for each pretext task can heavily affect the performance of our SSL framework. To avoid heavy hyperparameter tuning and increase the reproducibility of our pre-training framework, we adopt the same number of pre-training iterations for all stages. Although this strategy of using the same number of pre-training iterations is likely not the optimal design for individual stages, we believe that it can better reflect the real performance of our framework with less stochasticity and more reactivity. We provide a more detailed discussion on this issue next.
+
+Impacts of Pre-training Iterations. SSL pre-training frameworks usually require huge pre-training iterations to
+
+![](images/eb08feac410142ebb0149c6e0bde878812dd728c60a1ad1abf429cb263c5f165.jpg)  
+Figure 5. Ablation study on the number of pre-training iterations for our SSL pre-training framework. We find that our SSL pretraining framework continues to improve in performance until  $\sim 20,000$  pre-training iterations. After that, the performance starts to decline, mainly due to overfitting.
+
+converge. For example, SimSiam [5] requires about 250K iterations to pre-train their model on ImageNet, which is very time-consuming. To validate how this issue may affect our SSL pre-training framework, we conduct the ablation study on the number of pre-training iterations for our SSL pre-training framework, and the results are shown in Figure 5. Note that we use the same number of iterations for all pre-training stages for convenience and to get rid of heavy hyperparameter tuning.
+
+From the results, we find that the performance of our SSL pre-training framework improves rapidly as the number of pre-training iterations increases, up to around 20,000, which is a relatively small number compared with those used in existing SSL pre-training frameworks. The performance of our SSL pre-training framework gradually drops as the number of pre-training iterations exceeds 20,000, mainly due to overfitting. One possible explanation of this is that the model pre-trained by our SSL pre-training framework can efficiently learn the representation of mirror reflection under a short pre-training period. Another possible explanation is that unlike generic SSL pre-training frameworks that attempt to extract general-propose image representation, our SSL pre-training framework is developed specially for a specific task, i.e., mirror detection. While generic SSL pre-training may cause feature redundancy for our task, having too many pre-training iterations (i.e., over 20,000) causes overfitting.
+
+Failure Cases. Our SSL pre-training framework does have limitations. It continues to struggle with the limitations of the base method [37], as the base method adopted can also significantly affect the final detection performances. For example, as shown in first row of Figure 6, MirrorNet with our SSL pre-training may still fail to handle some mirror-like regions (e.g., the top-right region). It may also fail to de
+
+![](images/5e0fa381c7da1010282d6fc2a226bfa379b1d1dfb405d95dcf84c3caa7ac4974.jpg)
+
+![](images/a21f8058477846417d25926fda83d6439c42b732b24f2a0af94409428e39c431.jpg)
+
+![](images/5739f80c78e3835cac5bc18142f5c73c76e0a05db49d5b10faeba1c9b4c9370a.jpg)
+
+![](images/7d9e1db27b0ef5915888bd41c21fdb0bf768ae131bbe9dc9f456e9e524bf3e5c.jpg)  
+Image
+
+![](images/646ac3372c6f1b8574c145ef0ee9f0430b5de75bf66388c8b1ba8e482a373c38.jpg)  
+Ours
+
+![](images/ef33e88d90d37b68820d4534a4f8cef4e29bf268f405ad1c5d3162e47e1c4c44.jpg)  
+GT  
+Figure 6. Failure cases of our SSL pre-training framework based on MirrorNet. Our SSL pre-training framework fails to tackle the intrinsic limitation of existing mirror detection methods. The model pre-trained by our SSL pre-training framework still has wrong predictions when the input images are challenging (e.g., containing mirror-like regions in the first row, and insufficient relationship between mirror/non-mirror regions in the second row.
+
+tect mirrors in some challenging cases when the input image contains ambiguous mirrors without the relationship of mirror reflection and sufficient contextual contrast, as shown in the second row of Figure 6.
+
+# 5. Conclusion
+
+In this paper, we have investigated how self-supervised learning (SSL) works with the mirror detection task. To the best of our knowledge, we are the first to explore SSL pre-training frameworks applied to mirror detection. We have found that the supervised ImageNet pre-training might not be the ideal way to extract backbone image features for mirror detection due to the discrepancy between general-propose representation and mirror-specific representation. We have also proposed a new SSL pre-training framework to pre-train the backbone network for mirror detection. Our SSL pre-training framework does not require any labeled data. It progressively (image-, patch- and pixel-level pre-training) models the relationship of mirror reflection, and obtains a better representation of mirrors in the pre-training stage. Experimental results show that the proposed SSL pre-training framework outperforms CNN-based state-of-the-art SSL pre-training frameworks, and even achieves better results compared with supervised ImageNet pre-training.
+
+Acknowledgments: The work described in this paper was partially supported by a GRF grant from the Research Grants Council of Hong Kong (Project No.: CityU 11211223).
+
+# References
+
+[1] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018.  
+[2] M. Braun, S. Krebs, F. Flohr, and D. M. Gavrila. Eurocity persons: A novel benchmark for person detection in traffic scenes. IEEE TPAMI, 41(8):1844-1861, Aug 2019.  
+[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020.  
+[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020.  
+[5] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021.  
+[6] MMSelfSup Contributors. MMSelfSup: Openmmlab self-supervised learning toolbox and benchmark. https://github.com/open-mmlab/mmselfsup, 2021.  
+[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016.  
+[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009.  
+[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017.  
+[10] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015.  
+[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010.  
+[12] Christopher Funk and Yanxi Liu. Beyond planar symmetry: Modeling human perception of reflection and rotation symmetries in the wild. In ICCV, pages 793-803, 2017.  
+[13] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010.  
+[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020.  
+[15] Huankang Guan, Jiaying Lin, and Rynson W.H. Lau. Learning semantic associations for mirror detection. In CVPR, pages 5941-5950, June 2022.
+
+[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020.  
+[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016.  
+[18] Ruozhen He, Jiaying Lin, and Rynson WH Lau. Efficient mirror detection via multi-level heterogeneous learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 790-798, 2023.  
+[19] Tianyu Huang, Bowen Dong, Jiaying Lin, Xiaohui Liu, Rynson WH Lau, and Wangmeng Zuo. Symmetry-aware transformer-based mirror detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 935–943, 2023.  
+[20] Peter Kohler, Alexandra Yakovleva, Alasdair Clarke, Yanxi Liu, and Anthony Norcia. Parametric responses to rotation symmetry in mid-level visual cortex. Journal of Vision, 15(12):1122-1122, 2015.  
+[21] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In ICLR, 2018.  
+[22] Jiaying Lin, Xin Tan, and Rynson WH Lau. Learning to detect mirrors from videos via dual correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9109-9118, 2023.  
+[23] Jiaying Lin, Guodong Wang, and Rynson W. H. Lau. Progressive mirror detection. In CVPR, 2020.  
+[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014.  
+[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014.  
+[26] Zhiqiu Lin, Jin Sun, Abe Davis, and Noah Snavely. Visual chirality. In CVPR, pages 12295-12303, 2020.  
+[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018.  
+[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015.  
+[29] Jiaqi Tan, Weijie Lin, Angel X Chang, and Manolis Savva. Mirror3d: Depth refinement for mirror surfaces. In CVPR, pages 15990-15999, 2021.  
+[30] Xin Tan, Jiaying Lin, Ke Xu, Chen Pan, Lizhuang Ma, and Rynson W. H. Lau. Mirror detection with the visual chirality cue. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.  
+[31] Christopher W Tyler. Human symmetry perception and its computational analysis. Psychology Press, 2003.  
+[32] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, 2021.
+
+[33] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018.  
+[34] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, pages 418-434, 2018.  
+[35] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 1492-1500, 2017.  
+[36] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, pages 9653-9663, 2022.  
+[37] Xin Yang, Haiyang Mei, Ke Xu, Xiaopeng Wei, Baocai Yin, and Rynson W.H. Lau. Where is my mirror? In ICCV, 2019.  
+[38] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Learning with noisy labels for robust point cloud segmentation. In ICCV, pages 6443-6452, 2021.  
+[39] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. Robust point cloud segmentation with noisy annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7696-7710, 2022.  
+[40] Shuquan Ye, Yujia Xie, Dongdong Chen, Yichong Xu, Lu Yuan, Chenguang Zhu, and Jing Liao. Improving commonsense in vision-language models via knowledge graph riddles. In CVPR, pages 2634-2645, 2023.  
+[41] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021.  
+[42] Xiaohang Zhan, Ziwei Liu, Ping Luo, Xiaou Tang, and Chen Loy. Mix-and-match tuning for self-supervised semantic segmentation. In AAAI, volume 32, 2018.  
+[43] Xiaohang Zhan, Jiahao Xie, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Online deep clustering for unsupervised representation learning. In CVPR, 2020.  
+[44] Nanxuan Zhao, Zhirong Wu, Rynson WH Lau, and Stephen Lin. What makes instance discrimination good for transfer learning? In ICLR, 2021.  
+[45] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, , Hutchuan Lu, and Xiang Ruan. Self-supervised pretraining for rgb-d salient object detection. In AAAI, 2022.  
+[46] Ying Zheng, Yiyi Zhang, Xiaogang Xu, Jun Wang, and Hongxun Yao. Visual chirality meets freehand sketches. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1544-1548. IEEE, 2021.

2023/Self-supervised Pre-training for Mirror Detection/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b294b6b18d975d2749654590fca9fde243f2fc9872987b9bd9676078393ce52
+size 421414

2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/fa7524e1-a76d-4844-9361-43c8b8046bf9_content_list.json

@@ -0,0 +1,1876 @@
+[
+    {
+        "type": "text",
+        "text": "SemARFlow: Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving",
+        "text_level": 1,
+        "bbox": [
+            94,
+            128,
+            875,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Shuai Yuan, Shuzhi Yu, Hannah Kim, and Carlo Tomasi  \nDuke University",
+        "bbox": [
+            261,
+            203,
+            707,
+            239
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{shuai, shuzhiyu, hannah, tomasi}@cs.duke.edu",
+        "bbox": [
+            285,
+            241,
+            683,
+            255
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            291,
+            313,
+            306
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Unsupervised optical flow estimation is especially hard near occlusions and motion boundaries and in low-texture regions. We show that additional information such as semantics and domain knowledge can help better constrain this problem. We introduce SemARFlow, an unsupervised optical flow network designed for autonomous driving data that takes estimated semantic segmentation masks as additional inputs. This additional information is injected into the encoder and into a learned upsampler that refines the flow output. In addition, a simple yet effective semantic augmentation module provides self-supervision when learning flow and its boundaries for vehicles, poles, and sky. Together, these injections of semantic information improve the KITTI-2015 optical flow test error rate from  $11.80\\%$  to  $8.38\\%$ . We also show visible improvements around object boundaries as well as a greater ability to generalize across datasets. Code is available at https://github.com/duke-vision/semantic-unsup-flow-release.",
+        "bbox": [
+            75,
+            324,
+            473,
+            612
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            643,
+            209,
+            659
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Optical flow estimation, i.e., pixel-level motion tracking across video frames, has broad applications in many computer vision tasks that include object tracking [61], video editing [14, 31], and autonomous driving [6, 58].",
+        "bbox": [
+            75,
+            671,
+            468,
+            733
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Thanks to the success of deep convolutional neural networks [34, 17, 36] and transformer networks [33, 69, 88, 28] in computer vision, many top-performing supervised optical flow networks have been proposed in recent years [10, 65, 22, 68, 85], in which ground-truth labels supervise training. However, real optical flow is hard to label, so most supervised methods train (or at least pre-train) on synthetic datasets [10, 44, 4, 56] which makes them hard to adapt to real applications due to the significant gap between synthetic and real data [70, 29].",
+        "bbox": [
+            75,
+            733,
+            468,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Due to label scarcity, unsupervised training of optical",
+        "bbox": [
+            96,
+            885,
+            468,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/9e65a6cce7dfec507031a1407f7e49721073b9180ae490256c7b968ba3697ddc.jpg",
+        "image_caption": [
+            "Figure 1. An example on KITTI-2015 test set (sample #63). Only one input frame is shown for conciseness. Our SemARFlow takes additional semantic segmentation inputs (estimated by an off-the-shelf model) and outputs much sharper flow around semantic boundaries."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            290,
+            671,
+            415
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/343f36ac10bb396cab4abe94b28fbb130cf086e612f0ae4d6d44fd01bbc59452.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            674,
+            290,
+            727,
+            414
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/81f42782b8b7e34d6a081d106bfccb2d41438995034e911df052f59146cd50c3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            728,
+            291,
+            779,
+            414
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/85dc6b3c2243ce70f0fa7c131cc53de504af6f1b5e13f883b0520d0b57d3ac48.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            781,
+            291,
+            834,
+            414
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/b36e7fc92d3acc591e035ba081195e897be82f0c0d8276200009904df42750ff.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            834,
+            291,
+            888,
+            414
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "flow estimators [81] instead uses loss terms based on the assumptions of constant brightness and smooth flow [45, 27, 42]. Some self-supervision techniques have also been studied to enhance model performance [38, 39, 37]. Unsupervised training makes it possible to train flow networks directly on large real datasets from the target domain.",
+        "bbox": [
+            496,
+            529,
+            893,
+            621
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Even so, unsupervised optical flow estimation is a poorly constrained problem. Brightness is not constant in regions with occlusions [72] or on shiny surfaces [43], nor is it smooth across motion boundaries [30, 82]. Moreover, points in regions with poor texture content [59] or in dark shadows [58] are difficult to track as they are easily confused with neighboring points (the so-called aperture problem). This dearth of reliable constraints makes the unsupervised training of flow networks especially challenging.",
+        "bbox": [
+            496,
+            623,
+            893,
+            760
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "A natural way to address this issue is to inject additional constraints in the form of semantics and domain knowledge. For example, in autonomous driving applications [84, 6], we have clear expectations about object types and layouts of the scene, as well as prior knowledge of how each type of object typically moves. We focus on this application domain and show that this additional information helps the network achieve better flow results.",
+        "bbox": [
+            496,
+            761,
+            893,
+            882
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To make an autonomous driving system work well in",
+        "bbox": [
+            519,
+            885,
+            890,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "9566",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "reality, it is best to train it on real data. However, annotating real driving video with optical flow labels is expensive [83], as it requires careful synchronization and calibration of diverse sensors including cameras, LiDAR, and GPS/IMU [16], aided by some manual annotation and curation based on CAD models of moving objects [46].",
+        "bbox": [
+            75,
+            90,
+            468,
+            181
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In contrast, annotating semantic labels seems much more feasible, and indeed semantic labels are available in most (if not all) existing driving datasets. We consider semantic segmentation because it provides semantics at the pixel level, the same level as optical flow. As one of the most popular and well-studied tasks in modern computer vision, semantic segmentation [40, 47, 87] has been extensively adopted for autonomous driving systems. In this paper, we show that adding semantic segmentation inputs helps improve unsupervised optical flow performance significantly.",
+        "bbox": [
+            75,
+            184,
+            470,
+            335
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we first infer semantic segmentation maps using an off-the-shelf model [89], which of course is trained with semantic labels. An encoder with semantic map input is used to aggregate image and semantic features (Sec. 3.1), and a learned upsampler is added into the iterative decoder to refine flow around object boundaries given semantic inputs (Sec. 3.2). We also propose a simple yet effective semantic augmentation module for self-supervision, which provides realistic augmentations specific to the vehicles, poles, and sky classes based on domain knowledge and segmentation maps (Sec. 3.3). An occluder cache is implemented to improve efficiency (Sec. 3.4). Semantic augmentation provides challenging samples for self-supervision, which help train flow better in occluded regions and around foreground object boundaries.",
+        "bbox": [
+            75,
+            337,
+            470,
+            564
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Overall, by injecting semantic segmentation inputs, our SemARFlow network achieves significantly better unsupervised flow results both quantitatively (Sec. 4.3) and qualitatively (Sec. 4.4). Adapted from ARFlow [37], SemARFlow reduces KITTI-2015 [46] test error from  $11.80\\%$  to  $8.38\\%$  and out-performs the current unsupervised state-of-the-art UPFlow [42]  $(9.38\\%)$  by a clear margin. The example in Fig. 1 also demonstrates visible improvements around object boundaries. The effectiveness of each module is justified through an extensive ablation study (Sec. 4.5) and improvement analysis (Sec. 4.6). In addition to performance boost, unsupervised flow networks with additional semantic inputs generalize better across different datasets (Sec. 4.7).",
+        "bbox": [
+            75,
+            566,
+            468,
+            763
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our research is essentially novel compared to previous approaches. Some early work incorporates semantics in traditional energy-minimization methods [41, 18] for optical flow through geometric constraints such as piece-wise rigid motion and planar surface motion [57, 21, 1, 75]. In comparison, to the best of our knowledge, we are the first to inject semantics into the unsupervised training of recent optical flow networks. Some research also trains a network for segmentation and optical flow jointly [9]. In contrast, we",
+        "bbox": [
+            75,
+            763,
+            470,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "leverage existing, separately trained segmentation systems both because they are available and because modularity—separating segmentation from flow estimation—is important for the development of large real-application systems such as autonomous driving.",
+        "bbox": [
+            496,
+            90,
+            890,
+            166
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In summary, our contributions are as follows.",
+        "bbox": [
+            517,
+            167,
+            818,
+            181
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- To the best of our knowledge, we are the first to explore adding semantic inputs to assist the unsupervised training of deep optical flow networks.",
+            "- We propose a simple yet effective network called SemARFlow that achieves state-of-the-art results both quantitatively and qualitatively. Our model works well on real-life occlusions and yields sharp motion boundaries around objects.",
+            "- We provide full training and inference code as well as trained models to encourage follow-up research."
+        ],
+        "bbox": [
+            517,
+            190,
+            890,
+            361
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            375,
+            637,
+            388
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Supervised optical flow Since the introduction of the pioneering deep optical flow network FlowNet [10], more and more top-performing CNN-based flow estimators have been proposed over the years [23, 49, 65, 20, 22, 68]. Recently, vision transformer networks and attention mechanism have also been applied to this problem, and these have achieved state-of-the-art performance on benchmark datasets [76, 26, 85, 64, 77]. The supervised methods are often pre-trained on large synthetic datasets such as FlyingChairs [10] and FlyingThings3D [44] before fine-tuning on the target dataset. However, there is a clear gap between the artificially generated data and real scenarios.",
+        "bbox": [
+            496,
+            398,
+            890,
+            580
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Unsupervised optical flow Due to the lack of ground-truth labels, unsupervised optical flow estimation uses surrogate losses such as photometric loss and smoothness loss to supervise training [81, 55]. To tackle the issues around occlusion regions, various methods have been proposed including estimated occlusion masks [72], forward-backward consistency [45], and multi-frame fusion [25, 53]. To better upsample the flow in the decoder, UPFlow [42] additionally predicts a confidence map and an interpolation flow to guide flow refinement and has become the state-of-the-art unsupervised flow method.",
+        "bbox": [
+            496,
+            598,
+            890,
+            763
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Some latest research has also explored the use of self-supervision to enhance flow prediction. Early methods apply the knowledge distillation technique to train a two-stage teacher-student network [38, 39]. ARFlow [37] further improves this idea by generating reliable self-supervision signals from data transformations, while merging the two training stages into single-stage training with one added loss term. SimFlow [24] replaces the handcrafted features with deep self-supervised features to measure similarity in the",
+        "bbox": [
+            496,
+            763,
+            890,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "9567",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "unsupervised losses. SMURF [63] utilizes a RAFT-like structure [68] and applies multi-frame self-supervised training with many technical improvements. Our SemARFlow also uses self-supervision but with the guidance of semantic segmentation, which is much more realistic than guidance without semantics.",
+        "bbox": [
+            75,
+            90,
+            470,
+            181
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Semantic segmentation Semantic segmentation classifies each pixel of the given image into semantic objects. Fully Convolutional Network (FCN) [40] is one of the early CNN-based segmentation methods. It takes inputs of arbitrary sizes and outputs dense pixel-level predictions, becoming one of the popular backbone architectures for follow-up work. Deconvolution network [47] is also proposed to better recover low-level details of the prediction. One main challenge for these systems is that they lack global scene information. Subsequent work addresses this issue by enlarging the receptive field of the network with global pyramid pooling layers as in PSPNet [86], hybrid dilated convolutions [71], and a fast-down-sampling strategy [80, 79]. Attention modules have also shown to help in semantic segmentation by capturing full-image dependencies of all pixels [19, 87] or the semantic inter-dependencies across spatial and channel dimensions [12].",
+        "bbox": [
+            75,
+            200,
+            468,
+            455
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "There also exists extensive work on semantic segmentation in the context of autonomous driving [2, 67, 13, 5, 89], thanks to the publication of large-scale driving datasets [8, 46]. To better train the network on a coarsely labeled dataset like KITTI [46], Zhu et al. use a video prediction network SDCNet [52] to synthesize new training samples with relaxed label propagation [89]. Due to their good performance on KITTI [46], we utilize their network models to infer semantic inputs for all our experiments.",
+        "bbox": [
+            75,
+            458,
+            468,
+            594
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Combining semantics and optical flow Though there has been much progress on both semantic segmentation and optical flow estimation, the semantic optical flow problem (how to exploit semantics to help optical flow estimation) has received limited attention in recent years, and the current best results are thus much outdated. Some early methods incorporate semantics through geometric constraints to refine flow on various semantic regions, such as planar regions (using homographies) [57], static regions (using rigid camera motion and epipolar constraints) [21, 75], and rigid objects (estimating rigid motion for each object instance) [1]. However, most methods are traditional flow methods based on energy minimization, where an initial flow estimate is usually needed and semantics is mostly used for refinement. In comparison, we explore adapting latest unsupervised optical flow networks to leverage semantic inputs in one single stage of estimation.",
+        "bbox": [
+            75,
+            613,
+            468,
+            869
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Apart from using semantics to help flow, some research has also explored using optical flow to help seman",
+        "bbox": [
+            76,
+            869,
+            468,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "tic segmentation [51, 35], or to train both tasks jointly [9]. There are also studies on exploiting semantics on some other correspondence matching tasks such as stereo matching [74, 78] and 3D scene flow estimation, where some additional depth cues such as stereo camera inputs [54, 11] or point clouds [60] are needed.",
+        "bbox": [
+            496,
+            90,
+            890,
+            181
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            500,
+            196,
+            589,
+            210
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our network is adapted from the two-frame version of ARFlow [37], which uses a light-weight PwCNet [65] as its backbone. The inputs are two consecutive frames  $I_1, I_2 \\in \\mathbb{R}^{H \\times W \\times 3}$  as well as their semantic segmentation maps  $S_1, S_2 \\in \\{0, 1, \\dots, c\\}^{H \\times W}$ , where the number of classes  $c$  is 19 as we use the Cityscapes format [8]. A detailed diagram of the structure of our network can be found in the Appendix.",
+        "bbox": [
+            496,
+            222,
+            890,
+            343
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Semantic encoder",
+        "text_level": 1,
+        "bbox": [
+            500,
+            354,
+            671,
+            369
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We first inject semantic information into the encoder. As shown in Fig. 2(a), shared by each frame  $i \\in \\{1,2\\}$ , separate convolutional layers extract features from image  $I_{i}$  and semantics  $S_{i}$  (one-hot encoded). Features from the two pipelines are concatenated and fed to additional layers. Features at different resolutions  $(H / 2^{l}, W / 2^{l})$  form a pyramid  $\\{f_{i}^{(l)} \\mid 2 \\leq l \\leq 6\\}$ . The semantic information in these features helps delineate objects in dark shadows, where appearance is more homogeneous.",
+        "bbox": [
+            496,
+            377,
+            890,
+            516
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Iterative decoder with a learned upsampler",
+        "text_level": 1,
+        "bbox": [
+            500,
+            526,
+            867,
+            542
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Following [22, 37], an iterative residual refinement decoder starts from zero estimate  $F_{1\\rightarrow 2}^{(7)} = 0$ . For iteration  $l \\in \\{6,5,4,3,2\\}$ , the decoder refines feature map  $F_{1\\rightarrow 2}^{(l + 1)}$  into  $F_{1\\rightarrow 2}^{(l)}$  based on features  $f_{1}^{(l)}, f_{2}^{(l)}$  (Fig. 2(b)).",
+        "bbox": [
+            496,
+            550,
+            890,
+            619
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "More specifically,  $F_{1\\rightarrow 2}^{(l + 1)}$  is upsampled to match the resolution and used to warp  $f_2^{(l)}$  to yield warped feature  $\\hat{f}_1^{(l)}$ . Correlation volumes are computed between  $f_1^{(l)}$  and  $\\hat{f}_1^{(l)}$ . A one-by-one convolutional layer  $C^{(l)}$  compresses the number of channels to a fixed number so that the same layer can be reused across all iterations as proposed by [22]. A flow estimator network predicts a flow residual to be added to the current estimate, and a context network then aggregates flow information spatially and refines the current flow again.",
+        "bbox": [
+            496,
+            619,
+            890,
+            776
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "A learned upsampler network upsamples the final output  $F_{1\\rightarrow 2}^{(2)}$  in our system. This is different from ARFlow [37], which simply uses four-fold bilinear interpolation, making the final flow boundaries blurry. In contrast, our model learns to sharpen flow boundaries based on the semantic inputs, which have clear boundaries around moving objects. To this end, we add a convex upsampler network similar to the one in RAFT [68]. Different from UPFlow [42],",
+        "bbox": [
+            496,
+            777,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "9568",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/1f578db99d3ccda59acd52464574df776530c318c409b21f7fd6dae0bee7b43c.jpg",
+        "image_caption": [
+            "(a) shared pyramidal encoder  $(i\\in \\{1,2\\})$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            97,
+            343,
+            223
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/c30e1a5c14116a469523b3f00f526371d8f8a59ad87c073d3cadb07e3d7d3548.jpg",
+        "image_caption": [
+            "(b) one iteration (at the  $l$ th-level) of the iterative decoder  $(l\\in \\{6,5,4,3,2\\})$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            375,
+            98,
+            870,
+            219
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/4b0eca463840993e1e56b7ce31b8aa97cb087ccf3349b518dc7346f392bc4ee8.jpg",
+        "image_caption": [
+            "Figure 2. Network structure. See text in Sec. 3.1 and Sec. 3.2 for explanations. More detailed diagrams are in appendix.",
+            "Figure 3. Illustration of semantic augmentation as self-supervision. See text in Sec. 3.3 for details."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            84,
+            284,
+            888,
+            542
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "our learned upsampler is only used to upsample from internal estimate  $F_{1\\rightarrow 2}^{(l)}$  to output flow  $U_{1\\rightarrow 2}^{(l)}$ , but not when we upsample  $F_{1\\rightarrow 2}^{(l + 1)}$  at the first step in the decoder iteration. The upsampled  $U_{1\\rightarrow 2}^{(l)}$  has resolution  $(H / 2^{l - 2},W / 2^{l - 2})$ , so  $U_{1\\rightarrow 2} = U_{1\\rightarrow 2}^{(2)}$  (with the original resolution) is the final flow prediction of the network.",
+        "bbox": [
+            75,
+            604,
+            468,
+            705
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Semantic augmentation as self-supervision",
+        "text_level": 1,
+        "bbox": [
+            76,
+            720,
+            439,
+            739
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ARFlow has a very effective in-network augmentation module that samples random transformations  $\\mathcal{T}_{\\theta_1},\\mathcal{T}_{\\theta_2}$  of the flow prediction  $U_{1\\rightarrow 2}$  in the first pass of the network and then uses the transformed images  $\\hat{I}_1 = \\mathcal{T}_{\\theta_1}(I_1),\\hat{I}_2 =$ $\\mathcal{T}_{\\theta_2}(I_2)$  in a second pass. The prediction  $U_{1\\rightarrow 2}$  is also transformed accordingly and used to self-supervise the output of the second pass. See [37] for details.",
+        "bbox": [
+            75,
+            747,
+            468,
+            853
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We retain the augmentation module but make a third pass of the network using semantics-transformed inputs for self-supervision in addition to the ARFlow augmentations of ap",
+        "bbox": [
+            75,
+            854,
+            470,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "pearance (e.g., color jitter, random noise) and spatial transformations (e.g., random rotation, random rescaling). The idea behind semantic augmentation is to blend in real object motions across samples.",
+        "bbox": [
+            496,
+            604,
+            890,
+            665
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To this end (see Fig. 3), we carve objects from other samples based on their semantic maps and paste them as moving foreground objects into the current sample  $I_{1}, I_{2}$ , thereby producing augmented images  $\\tilde{I}_{1}, \\tilde{I}_{2}$ . The segmentation maps  $S_{1}, S_{2}$  and the first pass output flow  $U_{1 \\to 2}$  are transformed accordingly, and the transformed  $\\tilde{U}_{1 \\to 2}$  self-supervises the third-pass network outputs  $\\tilde{U}_{1 \\to 2}' = \\mathcal{F}(\\tilde{I}_{1}, \\tilde{I}_{2}, \\tilde{S}_{1}, \\tilde{S}_{2})$ , where  $\\mathcal{F}$  is the flow network.",
+        "bbox": [
+            496,
+            667,
+            892,
+            789
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Similar to occlusion hallucination in SelFlow [39], our semantic augmentation also creates new occlusions and uses the reliable non-occluded flow from the first pass to self-supervise the flow on those newly occluded pixels. However, rather than superpixels filled with noise, we create occlusions with real objects, which are much more realistic.",
+        "bbox": [
+            496,
+            791,
+            890,
+            882
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Moreover, since we determine the motion of each oc",
+        "bbox": [
+            517,
+            885,
+            890,
+            898
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "9569",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "cluder, we can use its known optical flow for motion self-supervision. As illustrated in Fig. 3, the occluder motion can be very different from that of the background. This trains the model to assume less smoothness around highly-dynamic objects and to output sharper motion estimates near motion boundaries. These hallucinated but realistic occluders also provide more examples of actively moving objects, which are relatively scarce in the original datasets.",
+        "bbox": [
+            75,
+            90,
+            472,
+            213
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "**Vehicles** We augment vehicle classes (car, truck, bus, train) as they are major sources of errors in previous models (see Tab. 5 for statistics). We group these four classes because they are easily confused with each other in the semantic input. To carve out a single car instance with high probability, we find connected components of vehicle segmentation regions using OpenCV [3] and accept masks whose bounding boxes are 50-300 pixels wide and 50-150 pixels high. We also drop masks that occupy less than  $60\\%$  area of their bounding boxes to remove largely occluded vehicles.",
+        "bbox": [
+            75,
+            244,
+            472,
+            397
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Poles We augment pole classes (pole, traffic light, traffic sign) not only because they are common foreground objects that cause occlusions, but also because their motions are usually poorly estimated due to their thin shapes. We use a 200-pixel wide, image-height sliding window to scan the segmentation of these pole classes and find the window with the most pole pixels. If more than  $10\\%$  pole pixels occur in it, that object mask is stored. This produces groups of nearby poles rather than just individual poles, which may not provide enough augmentation by themselves.",
+        "bbox": [
+            75,
+            426,
+            472,
+            580
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Sky We also augment the sky through prior knowledge that flow in the sky should be very small. We shrink the sky flow estimate by half before using it for self-supervision. As shown in Fig. 3, the middle tree part in  $U_{1\\rightarrow 2}$  is very blurry, with a large part of sky mixed in there. After shrinking sky flow by half, the tree in  $\\tilde{U}_{1\\rightarrow 2}$  now has sharper flow around the boundary between tree and sky. Using this as self-supervision helps the model sharpen the motion boundaries between sky and other regions.",
+        "bbox": [
+            75,
+            611,
+            468,
+            747
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Another way to apply this prior knowledge is to postprocess, i.e., shrink the final output flow in the sky region, instead of using the shrunk flow to self-supervise. However, the latter approach is preferable because it allows the network to balance prior knowledge (self-supervision) with the current sample observation (photometric loss). Our experiments do show cases where some small objects (such as some overhead electric power lines) are misclassified as part of the sky in the semantic input. In these cases, the network learns to rely less on prior knowledge.",
+        "bbox": [
+            75,
+            750,
+            472,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4.Occluder cache",
+        "text_level": 1,
+        "bbox": [
+            500,
+            90,
+            656,
+            106
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "During training, we maintain a cache of occluding objects for semantic augmentation. We sample objects from the cache to augment the current batch and then push any new objects found in the batch into the cache, with random replacement if the cache is full.",
+        "bbox": [
+            496,
+            113,
+            893,
+            188
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Specifically, after we load a sample for training, we first use the current model to infer its optical flow and then search within the sample for any cars (or poles) that meet the standards defined in Sec. 3.3 as \"cutouts\". For each cutout, we first find the flow vectors on all pixels that belong to that cutout and then compute the mean flow for later reference. We store the current sample, the cutout mask, and the mean flow together in the cache.",
+        "bbox": [
+            496,
+            189,
+            893,
+            309
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Later on, when we randomly select a cutout to augment another sample, we also retrieve the stored mean flow of this cutout, which is then augmented by random rescaling (by 0.8-1.5 times) and reversing (with  $50\\%$  probability). We use the augmented mean flow to translate the entire cutout object as an occluder to generate a new sample. For realism, occluders are pasted to the same location they occupied in their image of origin. Holding occluders in a cache provides a random mixture of samples across multiple batches and makes semantic augmentation very efficient.",
+        "bbox": [
+            496,
+            310,
+            893,
+            462
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.5. Loss functions",
+        "text_level": 1,
+        "bbox": [
+            500,
+            469,
+            648,
+            484
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Photometric loss In the first pass, forward and backward flow  $U_{1\\rightarrow 2}^{(l)}, U_{2\\rightarrow 1}^{(l)}$  at each level are predicted and occlusion masks  $O_{1\\rightarrow 2}^{(l)}, O_{2\\rightarrow 1}^{(l)}$  are computed by forward-backward consistency [45]. Frames are warped by",
+        "bbox": [
+            496,
+            492,
+            893,
+            559
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nI _ {i} ^ {\\prime (l)} (\\pmb {p}) = I _ {j} ^ {(l)} (\\pmb {p} + U _ {i \\rightarrow j} ^ {(l)} (\\pmb {p})), \\quad (i, j \\in \\{1, 2 \\})\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            535,
+            565,
+            854,
+            588
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $I_{i}^{(l)}$  is  $I_{i}$  down-sampled to the  $l$ -th scale, and  $\\pmb{p}$  denotes pixel coordinates at that scale. Following [37], three measures, namely  $L_{1}$ -distance  $(\\rho_{1})$ , structural similarity (SSIM) [73]  $(\\rho_{2})$ , and census loss [45]  $(\\rho_{3})$  are linearly combined to measure photometric differences between  $I_{i}^{(l)}$  and  $I_{i}^{\\prime(l)}$ . Occlusion regions are masked out and both forward and backward directions are taken into account at each level. The final photometric loss is",
+        "bbox": [
+            496,
+            595,
+            893,
+            726
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\ell_ {\\mathrm {p h}} = \\frac {1}{2} \\sum_ {(i, j) \\in \\{(1, 2), (2, 1) \\}} \\sum_ {l = 2} ^ {6} a _ {l} \\sum_ {k = 1} ^ {3} c _ {k} \\rho_ {k} \\left(I _ {i} ^ {(l)}, I _ {i} ^ {\\prime (l)}, O _ {i \\rightarrow j} ^ {(l)}\\right). \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            500,
+            732,
+            890,
+            790
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Smoothness loss Unlike most previous work, we do not include a smoothness loss because we find it to conflict with our learned upsampler. See ablation study in Sec. 4.5.",
+        "bbox": [
+            496,
+            806,
+            893,
+            853
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Augmentation loss As in ARFlow [37], our second forward pass computes the  $\\mathrm{L}_1$ -distance between the trans-",
+        "bbox": [
+            496,
+            869,
+            893,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "9570",
+        "bbox": [
+            480,
+            944,
+            517,
+            957
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/a0fa49c82b26e757e90ee9676028000104af60fc54e15a78edce27fa6d41aa86.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 1. KITTI benchmark results (EPE/px and Fl/%). Metrics evaluated at 'all' (all pixels, default for EPE), 'noc' (non-occlusions), 'bg' (background), and 'fg' (foreground). Key metrics (used in official ranking) are underlined. Our 'baseline' is an adapted ARFlow with added learned upsampler and no smoothness loss. +enc means adding semantic encoder; +aug means adding semantic augmentation. () means evaluation data used in training. - means unavailable.  $\\dagger$  denotes models with semantic inputs. For all metrics, lower is better."
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\" colspan=\"2\">Method</td><td colspan=\"2\">Train</td><td colspan=\"6\">Test</td></tr><tr><td>2012 EPE</td><td>2015 EPE</td><td>2012 Fl-noc</td><td>EPE</td><td>Fl-all</td><td>2015 Fl-noc</td><td>Fl-bg</td><td>Fl-fg</td></tr><tr><td rowspan=\"4\">supervised</td><td>PWC-Net+ [66]</td><td>-</td><td>(1.50)</td><td>3.36</td><td>1.4</td><td>7.72</td><td>4.91</td><td>7.69</td><td>7.88</td></tr><tr><td>IRR-PWC [22]</td><td>-</td><td>(1.63)</td><td>3.21</td><td>1.6</td><td>7.65</td><td>4.86</td><td>7.68</td><td>7.52</td></tr><tr><td>RAFT [68]</td><td>-</td><td>(0.63)</td><td>-</td><td>-</td><td>5.10</td><td>3.07</td><td>4.74</td><td>6.87</td></tr><tr><td>Separable Flow [85]</td><td>-</td><td>(0.69)</td><td>-</td><td>-</td><td>4.53</td><td>2.78</td><td>4.25</td><td>5.92</td></tr><tr><td rowspan=\"8\">unsupervised</td><td>SelFlow [39]</td><td>1.69</td><td>4.84</td><td>4.31</td><td>2.2</td><td>14.19</td><td>9.65</td><td>12.68</td><td>21.74</td></tr><tr><td>SimFlow [24]</td><td>-</td><td>5.19</td><td>-</td><td>-</td><td>13.38</td><td>8.21</td><td>12.60</td><td>17.27</td></tr><tr><td>ARFlow [37]</td><td>1.44</td><td>2.85</td><td>-</td><td>1.8</td><td>11.80</td><td>-</td><td>-</td><td>-</td></tr><tr><td>UFlow [27]</td><td>1.68</td><td>2.71</td><td>4.26</td><td>1.9</td><td>11.13</td><td>8.41</td><td>9.78</td><td>17.87</td></tr><tr><td>UPFlow [42]</td><td>1.27</td><td>2.45</td><td>-</td><td>1.4</td><td>9.38</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Ours (baseline)</td><td>1.39</td><td>2.61</td><td>4.30</td><td>1.7</td><td>9.89</td><td>6.98</td><td>8.82</td><td>15.21</td></tr><tr><td>Ours (+enc)†</td><td>1.29</td><td>2.42</td><td>3.97</td><td>1.5</td><td>8.99</td><td>3.97</td><td>8.19</td><td>13.01</td></tr><tr><td>Ours (+enc +aug)†</td><td>1.28</td><td>2.18</td><td>3.90</td><td>1.5</td><td>8.38</td><td>3.90</td><td>7.48</td><td>12.91</td></tr></table>",
+        "bbox": [
+            163,
+            87,
+            807,
+            321
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "formed flow  $\\hat{U}_{1\\rightarrow 2}$  and the second pass output  $\\hat{U}_{1\\rightarrow 2}^{\\prime} = \\mathcal{F}(\\hat{I}_1,\\hat{I}_2,\\hat{S}_1,\\hat{S}_2)$  for each pixel  $\\pmb{p}$  as",
+        "bbox": [
+            75,
+            409,
+            468,
+            441
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {D} (\\boldsymbol {p}) = \\left\\| \\hat {U} _ {1 \\rightarrow 2} (\\boldsymbol {p}) - \\hat {U} _ {1 \\rightarrow 2} ^ {\\prime} (\\boldsymbol {p}) \\right\\| _ {1}.\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            156,
+            449,
+            388,
+            469
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "This loss is then averaged over the transformed non-occluded region where self-supervision  $\\hat{U}_{1\\rightarrow 2}$  is accurate",
+        "bbox": [
+            75,
+            478,
+            468,
+            508
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\ell_ {\\mathrm {a r}} = \\frac {\\sum_ {\\boldsymbol {p}} (1 - \\hat {O} _ {1 \\rightarrow 2} (\\boldsymbol {p})) \\hat {D} (\\boldsymbol {p})}{\\sum_ {\\boldsymbol {p}} (1 - \\hat {O} _ {1 \\rightarrow 2} (\\boldsymbol {p}))}. \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            163,
+            518,
+            468,
+            560
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For semantic augmentation (third pass), since we are also doing self-supervision with a new generated sample, we use a similar loss definition",
+        "bbox": [
+            75,
+            569,
+            468,
+            613
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\ell_ {\\mathrm {a u g}} = \\frac {\\sum_ {\\boldsymbol {p}} \\left(1 - \\tilde {O} _ {1 \\rightarrow 2} (\\boldsymbol {p})\\right) \\tilde {D} (\\boldsymbol {p})}{\\sum_ {\\boldsymbol {p}} \\left(1 - \\tilde {O} _ {1 \\rightarrow 2} (\\boldsymbol {p})\\right)}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            160,
+            621,
+            468,
+            662
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\tilde{D} (\\pmb {p})$  is the distance between the semanticaugmented flow  $\\tilde{U}_{1\\rightarrow 2}$  and the third-pass output  $\\tilde{U}_{1\\rightarrow 2}^{\\prime}$",
+        "bbox": [
+            75,
+            672,
+            468,
+            705
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\tilde {D} (\\boldsymbol {p}) = \\left\\| \\tilde {U} _ {1 \\rightarrow 2} (\\boldsymbol {p}) - \\tilde {U} _ {1 \\rightarrow 2} ^ {\\prime} (\\boldsymbol {p}) \\right\\| _ {1},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            156,
+            713,
+            388,
+            733
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "and the augmented mask  $\\tilde{O}_{1\\rightarrow 2}$  is now computed by",
+        "bbox": [
+            75,
+            741,
+            423,
+            758
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\tilde {O} _ {1 \\rightarrow 2} = 1 - \\max (1 - O _ {1 \\rightarrow 2}, M)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            156,
+            766,
+            387,
+            785
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "because we penalize on both the originally non-occluded region  $1 - O_{1\\rightarrow 2}$  and the pasted foreground object mask  $M$  (of which we know the true motion).",
+        "bbox": [
+            75,
+            795,
+            468,
+            840
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Final loss (with  $\\lambda = 0.02$  as in ARFlow [37]):",
+        "bbox": [
+            75,
+            858,
+            398,
+            875
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\ell = \\ell_ {\\mathrm {p h}} + \\lambda \\left(\\ell_ {\\mathrm {a r}} + \\ell_ {\\mathrm {a u g}}\\right). \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            883,
+            468,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            409,
+            632,
+            426
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Datasets",
+        "text_level": 1,
+        "bbox": [
+            500,
+            433,
+            601,
+            448
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We mainly use KITTI [15, 46] and Cityscapes [8] datasets for experiments. Following [37], we train our model first on KITTI raw sequences [46] (55.7k samples) and then fine-tune on KITTI-2015 multi-view extension [46] (11.8k samples). We validate our model using KITTI-2015 train [46] (200 samples) and KITTI-2012 train set [15] (194 samples) since they are the only sets that have optical flow labels. We also train on Cityscapes [8] sequences (83.3k samples) to test model generalization ability, and we sample every other frame to match its frame rate with KITTI. All semantic segmentation inputs are estimated using an off-the-shelf model [89] with a DeepLabV3Plus [7] backbone, which achieves  $83.45\\%$  mean IoU on Cityscapes and  $72.82\\%$  mean IoU on KITTI.",
+        "bbox": [
+            496,
+            455,
+            890,
+            669
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Implementation details",
+        "text_level": 1,
+        "bbox": [
+            500,
+            676,
+            714,
+            693
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We implement the model in PyTorch [48]  $^{1}$ . We use the Adam optimizer [32] with  $\\beta_{1} = 0.9$ ,  $\\beta_{2} = 0.999$  and batch size 4. We first train on KITTI raw sequences [46] for 100k iterations with a fixed learning rate 0.0002 and then train on KITTI multi-view extension set [15, 46] for another 100k iterations using OneCycleLR schedule [62] with maximum learning rate 0.0004 and linear annealing.",
+        "bbox": [
+            496,
+            699,
+            890,
+            805
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For data augmentation, we include random horizontal flipping and swapping of the input frames. We resize the inputs to  $256 \\times 832$  before feeding into the network. The photometric loss weight for each scale  $a_{l}$  ( $2 \\leq l \\leq 6$ ) in",
+        "bbox": [
+            496,
+            805,
+            890,
+            867
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{1}$ Code and instructions are available at https://github.com/duke-vision/semantic-unsup-flow-release.",
+        "bbox": [
+            500,
+            875,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "9571",
+        "bbox": [
+            480,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/1837df44ffbaed2651e485dc2dbce8ffd2943490d301b3338d54190c2c707589.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Fl-all</td><td>Fl-noc</td><td>Fl-bg</td><td>Fl-fg</td></tr><tr><td>JFS [21]</td><td>17.07</td><td>9.81</td><td>15.9</td><td>22.92</td></tr><tr><td>SOF [57]</td><td>16.81</td><td>10.86</td><td>14.63</td><td>27.73</td></tr><tr><td>MRFlow [75]</td><td>12.19</td><td>8.86</td><td>10.13</td><td>22.52</td></tr><tr><td>Bai et al. [1]</td><td>11.62</td><td>8.75</td><td>8.61</td><td>26.69</td></tr><tr><td>Ours (final)</td><td>8.38</td><td>3.90</td><td>7.48</td><td>12.91</td></tr></table>",
+        "bbox": [
+            101,
+            88,
+            444,
+            181
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Eq. (1) are set as 1, 1, 1, 1, 0. The weights for three photometric distance measures  $\\rho_{k}$  ( $1 \\leq k \\leq 3$ ) in Eq. (1) are set as 0.15, 0.85, 0 for the first 50k iterations and 0, 0, 1 afterwards. We start the appearance and spatial augmentation (second pass as in ARFlow [37]) at 50k iterations, and semantic augmentation (third pass) after 150k iterations.",
+        "bbox": [
+            76,
+            267,
+            467,
+            357
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3. Benchmark testing results",
+        "text_level": 1,
+        "bbox": [
+            76,
+            378,
+            313,
+            395
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As common practice, we evaluate optical flow predictions based on two error measurements, Fl (error rate) and EPE (mean  $\\mathrm{L}_2$  distance). When computing Fl, the estimate of each pixel is considered correct if the error is smaller than 3 pixels or  $5\\%$  of the magnitude of ground-truth flow [46].",
+        "bbox": [
+            76,
+            406,
+            467,
+            482
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As shown in Tab. 1, our semantic modules (both semantic encoder and semantic augmentation) improve performance on both KITTI-2012 and KITTI-2015 sets on all metrics. Our final model achieves  $8.38\\%$  F1-all error rate on KITTI-2015 test set, which is significantly better than ARFlow [37]  $(11.80\\%)$ , from which we adapt. We also outperform the current state-of-the-art UPFlow [42]  $(9.38\\%)$  by a clear margin. All these results strongly suggest that adding semantic inputs can help improve unsupervised flow estimation significantly.",
+        "bbox": [
+            75,
+            484,
+            467,
+            636
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In addition, Tab. 2 shows that our test results significantly outperform all previous semantic optical flow methods, most of which are based on traditional energy minimization. We are the first to apply semantic inputs to recent unsupervised flow networks, so we are able to push the state-of-the-art by a clear margin.",
+        "bbox": [
+            75,
+            640,
+            467,
+            731
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4. Qualitative results",
+        "text_level": 1,
+        "bbox": [
+            76,
+            752,
+            253,
+            767
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Some qualitative results are shown in Fig. 4. We can see that our flow outputs have very sharp boundaries and very clean object motion, especially around cars and poles. Our model successfully learns to adapt flow estimation to the semantic map input. Moreover, our model is able to handle very challenging samples where the foreground car motion is drastically different from the background due to the challenging samples we create in semantic augmentation.",
+        "bbox": [
+            75,
+            779,
+            467,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/8e9b8866ad2171ccb810955a26d057356e880f8be6848a0e401c5ed180d3de35.jpg",
+        "table_caption": [
+            "Table 2. KITTI-2015 test results (Fl/\\%) compared with other semantic optical flow methods. Metrics evaluated at 'all' (all pixels), 'noc' (non-occlusions), 'bg' (background), and 'fg' (foreground)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>up</td><td>no sm</td><td>enc</td><td>aug</td><td>Fl-all</td><td>EPE-all</td><td>EPE-noc</td><td>EPE-occ</td></tr><tr><td></td><td></td><td></td><td></td><td>10.36</td><td>2.90</td><td>2.07</td><td>6.97</td></tr><tr><td></td><td></td><td>1</td><td></td><td>9.92</td><td>2.69</td><td>1.89</td><td>6.57</td></tr><tr><td></td><td></td><td>2</td><td></td><td>9.83</td><td>2.65</td><td>1.86</td><td>6.40</td></tr><tr><td></td><td></td><td>3</td><td></td><td>9.75</td><td>2.61</td><td>1.85</td><td>6.54</td></tr><tr><td></td><td></td><td>4</td><td></td><td>9.75</td><td>2.64</td><td>1.85</td><td>6.55</td></tr><tr><td>✓</td><td></td><td></td><td></td><td>10.22</td><td>2.83</td><td>1.97</td><td>6.75</td></tr><tr><td>✓</td><td>✓</td><td></td><td></td><td>8.87</td><td>2.61</td><td>1.85</td><td>6.04</td></tr><tr><td>✓</td><td>✓</td><td>3</td><td></td><td>8.26</td><td>2.42</td><td>1.73</td><td>5.79</td></tr><tr><td>✓</td><td>✓</td><td></td><td>✓</td><td>8.80</td><td>2.48</td><td>1.60</td><td>6.64</td></tr><tr><td>✓</td><td>✓</td><td>3</td><td>✓</td><td>7.79</td><td>2.18</td><td>1.40</td><td>5.64</td></tr></table>",
+        "bbox": [
+            506,
+            88,
+            885,
+            244
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/d6899810fa46e2fa8b8110cd20f6e09521321eab8a3028129ff4d8c259abe500.jpg",
+        "table_caption": [
+            "Table 3. Ablation study on KITTI-2015 train set (EPE/px and F1%). 'up' means the learned upsampler; 'no sm' means turning off smoothness loss; 'enc' means the level at which we merge semantic features and image features in the encoder; 'aug' means semantic augmentation."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Options of aug</td><td>Fl-all</td><td>EPE-all</td><td>EPE-noc</td><td>EPE-occ</td></tr><tr><td>Ours (final)</td><td>7.79</td><td>2.18</td><td>1.40</td><td>5.64</td></tr><tr><td>start from 100k</td><td>7.92</td><td>2.19</td><td>1.40</td><td>5.41</td></tr><tr><td>vehicles only</td><td>7.94</td><td>2.21</td><td>1.44</td><td>5.78</td></tr><tr><td>loss on new occ</td><td>8.15</td><td>2.27</td><td>1.42</td><td>5.62</td></tr></table>",
+        "bbox": [
+            501,
+            348,
+            890,
+            429
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Ablation study of different semantic augmentation options on KITTI-2015 train (EPE/px and F1/%). See text for explanations.",
+        "bbox": [
+            498,
+            439,
+            890,
+            467
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.5. Ablation study",
+        "text_level": 1,
+        "bbox": [
+            500,
+            503,
+            648,
+            518
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We also do ablation study to show the effectiveness of each of our proposed modules. We can see in Tab. 3 that adding a semantic encoder to the vanilla ARFlow does help, and the optimal number of encoder layer groups added is 3. Also, our learned upsampler improves results significantly once we turn off the smoothness loss. This makes sense because the upsampler learns to adapt to boundaries, which can be distracted by the smoothness loss. Moreover, both our semantic encoder and semantic augmentation module help improve the results further, which suggests the effectiveness of adding semantic inputs.",
+        "bbox": [
+            496,
+            532,
+            890,
+            698
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Another ablation study is on different options for the semantic augmentation (Tab. 4). We try starting augmentation earlier from 100k iterations, which works slightly worse. There are mainly two reasons: (1) we use the first forward output to self-supervise our augmented output, so if we start early, the model will use poor output to self-supervise the augmented pass; and (2) our semantic augmentation creates very challenging samples with objects moving very differently from the background, so exposing these hard samples to the model too early may make it hard to train. Apart from this, we also try only using vehicles to augment as well as focusing loss on the newly occluded region, which are both inferior to our final version.",
+        "bbox": [
+            496,
+            704,
+            890,
+            898
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "9572",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/07c59636c83be6777bb7d26b345ec42e13f6245ab9cf5eed0a798a8bc2146132.jpg",
+        "image_caption": [
+            "Image Input"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            88,
+            241,
+            284
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/cbb5321d0383d9cd403ec9c926150713f4acbaadab3bca47ced4520392966d5e.jpg",
+        "image_caption": [
+            "Semantics Input"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            241,
+            88,
+            401,
+            284
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b18ba7a222c78a7ae580df67155847eef12a2edad21db45d5541879811b4cb0a.jpg",
+        "image_caption": [
+            "ARFlow",
+            "Figure 4. Qualitative results on KITTI test set (sample #7, 20, 38, 112, 183) compared with ARFlow [37] and UPFlow [42]. We only show first frame input for conciseness. Our outputs have much clearer boundaries around objects. For more qualitative examples, see Appendix."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            403,
+            88,
+            563,
+            284
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/81bafe3a42b8dceeb4b7fce071a2fb656aab904f459b663294b6eca2000eff17.jpg",
+        "image_caption": [
+            "UPFlow"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            565,
+            88,
+            727,
+            284
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/02089204af00bcd73238d308a144146cad31925f78d7c8c61735a06124c2374c.jpg",
+        "image_caption": [
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            727,
+            88,
+            890,
+            284
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/be5ecba44ee485cd5a6314673f47cc7873dfc49f9180708a710e28cff251dc48.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>road</td><td>car</td><td>terrain</td><td>vegetation</td><td>sidewalk</td><td>building</td><td>wall</td><td>pole</td></tr><tr><td>Proportion</td><td>42.4%</td><td>17.6%</td><td>14.0%</td><td>11.6%</td><td>6.2%</td><td>4.1%</td><td>1.3%</td><td>1.1%</td></tr><tr><td>ARFlow [37]</td><td>4.57</td><td>15.79</td><td>9.42</td><td>15.34</td><td>4.61</td><td>6.00</td><td>16.34</td><td>10.75</td></tr><tr><td>Ours (final)</td><td>3.67</td><td>10.17</td><td>8.74</td><td>13.47</td><td>3.09</td><td>4.27</td><td>11.18</td><td>9.44</td></tr><tr><td>Relative improvement</td><td>19.6%</td><td>35.6%</td><td>7.3%</td><td>12.2%</td><td>33.0%</td><td>28.7%</td><td>31.6%</td><td>12.2%</td></tr><tr><td>Reweighed contribution</td><td>19.0%</td><td>49.4%</td><td>4.8%</td><td>10.9%</td><td>4.7%</td><td>3.5%</td><td>3.4%</td><td>0.7%</td></tr></table>",
+        "bbox": [
+            122,
+            356,
+            848,
+            454
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.6. Improvement analysis",
+        "text_level": 1,
+        "bbox": [
+            76,
+            515,
+            284,
+            531
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "It may be interesting to understand where our improvements come from in terms of semantic classes, so we compute the KITTI-2015 train set error for each class. As shown in Tab. 5, our model improves the flow on every class. After reweighing the absolute improvement of each class based on their proportion in the evaluated pixels, we find that car flow accounts for nearly half of our improvement on Fl-all, indicating the effectiveness of augmenting vehicle objects.",
+        "bbox": [
+            75,
+            539,
+            468,
+            660
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.7. Generalization ability",
+        "text_level": 1,
+        "bbox": [
+            76,
+            670,
+            282,
+            686
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We also investigate the generalization ability of our semantic-aided flow models. We train flow models on Cityscapes and directly test them on KITTI-2015 train set without fine-tuning. Following [50], we crop the bottom  $25\\%$  of the frame to remove the car logo and resize the frame to size  $256 \\times 704$ . All other experiment settings are exactly the same as for KITTI. As we can see from Tab. 6, adding semantic inputs significantly helps our unsupervised flow model generalize and adapt better across datasets.",
+        "bbox": [
+            75,
+            694,
+            468,
+            830
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion and future work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            844,
+            336,
+            859
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we show that adding semantic segmentation inputs can help significantly improve the performance",
+        "bbox": [
+            76,
+            869,
+            468,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/4685a30e710e8fd8d831d674fe3ac4d8359dba10f7cdf0baff8260804b356c97.jpg",
+        "table_caption": [
+            "Table 5. KITTI-2015 flow error (Fl-all/%) for each semantic class. The first row shows the proportion of each class among evaluated pixels (KITTI flow does not evaluate on every pixel). We only show classes that account for at least  $1\\%$  of the evaluated pixels."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Fl-all</td><td>EPE-all</td><td>EPE-noc</td><td>EPE-occ</td></tr><tr><td>ARFlow (our impl.)</td><td>13.21</td><td>4.08</td><td>2.88</td><td>9.40</td></tr><tr><td>Ours (baseline)</td><td>12.27</td><td>3.81</td><td>2.43</td><td>9.91</td></tr><tr><td>Ours (+enc)†</td><td>11.28</td><td>3.33</td><td>2.12</td><td>8.75</td></tr><tr><td>Ours (+enc +aug)†</td><td>10.32</td><td>2.64</td><td>1.56</td><td>7.12</td></tr></table>",
+        "bbox": [
+            503,
+            512,
+            888,
+            593
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6. Generalization results (train on Cityscapes, and test on KITTI-2015 train). We use our implementation of ARFlow [37] to run this experiment. Our 'baseline' is an adapted ARFlow with added learned upsampler and no smoothness loss.  $^+$  enc means adding semantic encoder;  $^+$  aug means adding semantic augmentation.  $\\dagger$  denotes models with semantic inputs.",
+        "bbox": [
+            496,
+            603,
+            892,
+            686
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "of unsupervised optical flow networks on autonomous driving datasets. We propose a novel network model called SemARFlow, a semantic adaptation of ARFlow [37], with a semantic encoder, a learned upsampler, and a semantic augmentation module, where some domain-specific motion prior has been taken into account. Our network better predicts flow at occlusion regions and effectively sharpens flow estimates around object boundaries, even for very challenging samples. The additional semantic inputs also make our network generalize better across datasets.",
+        "bbox": [
+            496,
+            715,
+            890,
+            867
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "One direction for future work is motivated by the example illustrated in Fig. 5. Our model does output sharp flow",
+        "bbox": [
+            498,
+            869,
+            890,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "9573",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/1163bd8c232875b98653a17d74378498bfe4ef69b40c6cd4a3f95ac58e27a07a.jpg",
+        "image_caption": [
+            "Figure 5. An example on KITTI-2015 train set (sample #164). See the discussion in Sec. 5 for details."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            89,
+            207,
+            148
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/d3f5dfb78ff75b20c7534879e6a07823a7a3f0014c676ba3b095e4f7f671dee7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            210,
+            90,
+            336,
+            148
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/a7bdac7bbcf7508d90aa410a55955595fee7f51f1cf2bf6e69324c96fcd7f87f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            339,
+            90,
+            465,
+            148
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "boundaries between the cars and the background, thanks to good boundary information from the semantic maps. However, the boundary between these two fast-moving cars is inferred from image information only, because their semantic maps merge into one. As a result, that part of the flow boundary is less accurate. An interesting direction for future work is to use instance-level semantics as input for more detailed object masks.",
+        "bbox": [
+            76,
+            218,
+            470,
+            338
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments. This research is based upon work supported by the National Science Foundation under Grant No. 1909821.",
+        "bbox": [
+            76,
+            359,
+            468,
+            405
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            419,
+            173,
+            435
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Min Bai, Wenjie Luo, Kaustav Kundu, and Raquel Urtasun. Exploiting semantic information and deep matching for optical flow. In Proceedings of the European Conference on Computer Vision, pages 154-170. Springer, 2016. 2, 3, 7",
+            "[2] Shubhankar Borse, Ying Wang, Yizhe Zhang, and Fatih Porikli. Inverseform: A loss function for structured boundary-aware segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5901-5911, 2021. 3",
+            "[3] Gary Bradski. The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer, 25(11):120-123, 2000. 5",
+            "[4] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Proceedings of the European Conference on Computer Vision, pages 611-625. Springer, 2012. 1",
+            "[5] Yingfeng Cai, Lei Dai, Hai Wang, and Zhixiong Li. Multi-target pan-class intrinsic relevance driven model for improving semantic segmentation in autonomous driving. IEEE Transactions on Image Processing, 30:9069-9084, 2021. 3",
+            "[6] Linda Capito, Umit Ozguner, and Keith Redmill. Optical flow based visual potential field for autonomous driving. In IEEE Intelligent Vehicles Symposium, pages 885-891. IEEE, 2020. 1",
+            "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision, pages 801-818, 2018. 6",
+            "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes"
+        ],
+        "bbox": [
+            86,
+            445,
+            468,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016. 3, 6",
+            "[9] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10713-10720, 2020. 2, 3",
+            "[10] Alexey Dosovitskiy, Philipp Fischery, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision, ICCV '15, pages 2758-2766, Washington, DC, USA, 2015. IEEE Computer Society. 1, 2",
+            "[11] Cheng Feng, Long Ma, Congxuan Zhang, Zhen Chen, Liyue Ge, and Shaofeng Jiang. Ss-sf: Piecewise 3d scene flow estimation with semantic segmentation. IEEE Access, 9:22745-22759, 2021. 3",
+            "[12] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3146-3154, 2019. 3",
+            "[13] Aditya Ganeshan, Alexis Vallet, Yasunori Kudo, Shin-ichi Maeda, Tommi Kerola, Rares Ambrus, Dennis Park, and Adrien Gaidon. Warp-refine propagation: Semi-supervised auto-labeling via cycle-consistency. In Proceedings of the IEEE International Conference on Computer Vision, pages 15499–15509, 2021. 3",
+            "[14] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In Proceedings of the European Conference on Computer Vision, pages 713-729. Springer, 2020. 1",
+            "[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research, 32(11):1231-1237, 2013. 6",
+            "[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 2",
+            "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1",
+            "[18] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial Intelligence, 17(1-3):185–203, 1981. 2",
+            "[19] 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 Conference on Computer Vision and Pattern Recognition, pages 603-612, 2019. 3",
+            "[20] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "9574",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "on Computer Vision and Pattern Recognition, pages 8981-8989, 2018. 2",
+            "[21] Junhwa Hur and Stefan Roth. Joint optical flow and temporally consistent semantic segmentation. In Proceedings of the European Conference on Computer Vision, pages 163-177. Springer, 2016. 2, 3, 7",
+            "[22] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2019. 1, 2, 3, 6",
+            "[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1647-1655, 2017. 2",
+            "[24] Woobin Im, Tae-Kyun Kim, and Sung-Eui Yoon. Unsupervised learning of optical flow with deep feature similarity. In Proceedings of the European Conference on Computer Vision, pages 172-188. Springer, 2020. 2, 6",
+            "[25] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European Conference on Computer Vision, pages 690-706, 2018. 2",
+            "[26] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE International Conference on Computer Vision, pages 9772-9781, 2021. 2",
+            "[27] Rico Jonschkowski, Austin Stone, Jonathan T Barron, Ariel Gordon, Kurt Konolige, and Anelia Angelova. What matters in unsupervised optical flow. In Proceedings of the European Conference on Computer Vision, pages 557-572. Springer, 2020. 1, 6",
+            "[28] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys, 54(10s):1-41, 2022. 1",
+            "[29] Hannah H Kim, Celia Cintas, Girmaw Abebe Tadesse, and Skyler Speakman. Spatially constrained adversarial attack detection and localization in the representation space of optical flow networks. In International Joint Conference on Artificial Intelligence, 2023. 1",
+            "[30] Hannah Halin Kim, Shuzhi Yu, and Carlo Tomasi. Joint detection of motion boundaries and occlusions. In Proceedings of the British Machine Vision Conference, 2021. 1",
+            "[31] Hannah Halin Kim, Shuzhi Yu, Shuai Yuan, and Carlo Tomasi. Cross-attention transformer for video interpolation. In Proceedings of the Asian Conference on Computer Vision Workshops, pages 320–337, 2022. 1",
+            "[32] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Proceedings of the International Conference on Learning Representations, 2014. 6",
+            "[33] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recogni"
+        ],
+        "bbox": [
+            78,
+            92,
+            468,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "tion at scale. In Proceedings of the International Conference on Learning Representations, 2021. 1",
+            "[34] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1",
+            "[35] Xiangtai Li, Jiangang Bai, Kuiyuan Yang, and Yunhai Tong. Flow2seg: Motion-aided semantic segmentation. In Proceedings of the International Conference on Artificial Neural Networks, pages 225–237. Springer, 2019. 3",
+            "[36] Zewen Li, Fan Liu, Wenjie Yang, Shouheng Peng, and Jun Zhou. A survey of convolutional neural networks: analysis, applications, and prospects. IEEE transactions on neural networks and learning systems, 2021. 1",
+            "[37] Liang Liu, Jiangning Zhang, Ruifei He, Yong Liu, Yabiao Wang, Ying Tai, Donghao Luo, Chengjie Wang, Jilin Li, and Feiyue Huang. Learning by analogy: Reliable supervision from transformations for unsupervised optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6489-6498, 2020. 1, 2, 3, 4, 5, 6, 7, 8",
+            "[38] Pengpeng Liu, Irwin King, Michael R Lyu, and Jia Xu. Ddflow: Learning optical flow with unlabeled data distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8770-8777, 2019. 1, 2",
+            "[39] Pengpeng Liu, Michael Lyu, Irwin King, and Jia Xu. Selfflow: Self-supervised learning of optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4571-4580, 2019. 1, 2, 4, 6",
+            "[40] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431-3440, 2015. 2, 3",
+            "[41] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In International Joint Conference on Artificial Intelligence, volume 2, pages 674-679, 1981. 2",
+            "[42] Kunming Luo, Chuan Wang, Shuaicheng Liu, Haoqiang Fan, Jue Wang, and Jian Sun. Upflow: Upsampling pyramid for unsupervised optical flow learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1045-1054, 2021. 1, 2, 3, 6, 7, 8",
+            "[43] Rémi Marsal, Florian Chabot, Angélique Loesch, and Hichem Sahbi. Brightflow: Brightness-change-aware unsupervised learning of optical flow. In Proceedings of the IEEE winter conference on applications of computer vision, pages 2061-2070, 2023. 1",
+            "[44] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4040-4048, 2016. 1, 2",
+            "[45] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5",
+            "[46] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE Conference"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "9575",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "on Computer Vision and Pattern Recognition, 2015. 2, 3, 6, 7",
+            "[47] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1520-1528, 2015. 2, 3",
+            "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035. Curran Associates, Inc., 2019. 6",
+            "[49] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2720-2729, 2017. 2",
+            "[50] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12240–12249, 2019. 8",
+            "[51] Hazem Rashed, Senthil Yogamani, Ahmad El-Sallab, Pavel Krizek, and Mohamed El-Helw. Optical flow augmented semantic segmentation networks for automated driving. In International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 165-172, 2019. 3",
+            "[52] Fitsum A Reda, Guilin Liu, Kevin J Shih, Robert Kirby, Jon Barker, David Tarjan, Andrew Tao, and Bryan Catanzaro. Sdc-net: Video prediction using spatially-displaced convolution. In Proceedings of the European Conference on Computer Vision, pages 718-733, 2018. 3",
+            "[53] Zhile Ren, Orazio Gallo, Deqing Sun, Ming-Hsuan Yang, Erik B Sudderth, and Jan Kautz. A fusion approach for multiframe optical flow estimation. In Proceedings of the IEEE winter conference on applications of computer vision, pages 2077-2086. IEEE, 2019. 2",
+            "[54] Zhile Ren, Deqing Sun, Jan Kautz, and Erik Sudderth. Cascaded scene flow prediction using semantic segmentation. In Proceedings of the International Conference on 3D Vision, pages 225-233. IEEE, 2017. 3",
+            "[55] Zhe Ren, Junchi Yan, Bingbing Ni, Bin Liu, Xiaokang Yang, and Hongyuan Zha. Unsupervised deep learning for optical flow estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 2",
+            "[56] Stephan R. Richter, Zeeshan Hayder, and Vladlen Koltun. Playing for benchmarks. In Proceedings of the IEEE International Conference on Computer Vision, pages 2232-2241, 2017. 1",
+            "[57] Laura Sevilla-Lara, Deqing Sun, Varun Jampani, and Michael J Black. Optical flow with semantic segmentation and localized layers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3889-3898, 2016. 2, 3, 7"
+        ],
+        "bbox": [
+            78,
+            90,
+            467,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[58] Shihao Shen, Louis Kerofsky, and Senthil Yogamani. Optical flow for autonomous driving: Applications, challenges and improvements. arXiv preprint arXiv:2301.04422, 2023. 1",
+            "[59] Jianbo Shi and Carlo Tomasi. Good features to track. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 593-600, 1994. 1",
+            "[60] Yukang Shi and Kaiheng Ma. Safit: Segmentation-aware scene flow with improved transformer. In Proceedings of the International Conference on Robotics and Automation, pages 10648-10655. IEEE, 2022. 3",
+            "[61] Jeongho Shin, Sangjin Kim, Sangkyu Kang, Seong-Won Lee, Joonki Paik, Besma Abidi, and Mongi Abidi. Optical flow-based real-time object tracking using non-prior training active feature model. Real-time imaging, 11(3):204-218, 2005. 1",
+            "[62] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 6",
+            "[63] Austin Stone, Daniel Maurer, Alper Ayvaci, Anelia Angelova, and Rico Jonschkowski. Smurf: Self-teaching multiframe unsupervised raft with full-image warping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3887-3896, 2021. 3",
+            "[64] Xiuchao Sui, Shaohua Li, Xue Geng, Yan Wu, Xinxing Xu, Yong Liu, Rick Goh, and Hongyuan Zhu. Craft: Cross-attentional flow transformer for robust optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17602-17611, 2022. 2",
+            "[65] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8934-8943, 2018. 1, 2, 3",
+            "[66] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Models matter, so does training: An empirical study of cnns for optical flow estimation. IEEE transactions on pattern analysis and machine intelligence, 42(6):1408-1423, 2019. 6",
+            "[67] Andrew Tao, Karan Sapra, and Bryan Catanzaro. Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821, 2020. 3",
+            "[68] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision, pages 402-419. Springer, 2020. 1, 2, 3, 6",
+            "[69] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1",
+            "[70] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE Con"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "9576",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "ference on Computer Vision and Pattern Recognition Workshops, pages 969-977, 2018. 1",
+            "[71] Panqu Wang, Pengfei Chen, Ye Yuan, Ding Liu, Zehua Huang, Xiaodi Hou, and Garrison Cottrell. Understanding convolution for semantic segmentation. In Proceedings of the IEEE winter conference on applications of computer vision, pages 1451-1460. IEEE, 2018. 3",
+            "[72] Yang Wang, Yi Yang, Zhenheng Yang, Liang Zhao, Peng Wang, and Wei Xu. Occlusion aware unsupervised learning of optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4884-4893, 2018. 1, 2",
+            "[73] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5",
+            "[74] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Song Wang, and Lili Ju. Semantic stereo matching with pyramid cost volumes. In Proceedings of the IEEE International Conference on Computer Vision, pages 7484-7493, 2019. 3",
+            "[75] Jonas Wulff, Laura Sevilla-Lara, and Michael J Black. Optical flow in mostly rigid scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4671-4680, 2017. 2, 3, 7",
+            "[76] Haofei Xu, Jiaolong Yang, Jianfei Cai, Juyong Zhang, and Xin Tong. High-resolution optical flow from 1d attention and correlation. In Proceedings of the IEEE International Conference on Computer Vision, pages 10498-10507, 2021. 2",
+            "[77] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 2",
+            "[78] Guorun Yang, Hengshuang Zhao, Jianping Shi, Zhidong Deng, and Jiaya Jia. Segstereo: Exploiting semantic information for disparity estimation. In Proceedings of the European Conference on Computer Vision, pages 636-651, 2018. 3",
+            "[79] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129:3051-3068, 2021. 3",
+            "[80] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European Conference on Computer Vision, pages 325-341, 2018. 3",
+            "[81] Jason J. Yu, Adam W. Harley, and Konstantinos G. Derpanis. Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In Gang Hua and Hervé Jégou, editors, Proceedings of the European Conference on Computer Vision Workshops, pages 3-10, Cham, 2016. Springer International Publishing. 1, 2",
+            "[82] Shuzhi Yu, Hannah H Kim, Shuai Yuan, and Carlo Tomasi. Unsupervised flow refinement near motion boundaries. In"
+        ],
+        "bbox": [
+            78,
+            92,
+            468,
+            900
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Proceedings of the British Machine Vision Conference. BMVA Press, 2022. 1",
+            "[83] Shuai Yuan, Xian Sun, Hannah Kim, Shuzhi Yu, and Carlo Tomasi. Optical flow training under limited label budget via active learning. In Proceedings of the European Conference on Computer Vision, pages 410-427, 2022. 2",
+            "[84] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1",
+            "[85] Feihu Zhang, Oliver J Woodford, Victor Adrian Prisacariu, and Philip HS Torr. Separable flow: Learning motion cost volumes for optical flow estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 10807-10817, 2021. 1, 2, 6",
+            "[86] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2881-2890, 2017. 3",
+            "[87] 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 Conference on Computer Vision and Pattern Recognition, pages 6881-6890, 2021. 2, 3",
+            "[88] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In Proceedings of the International Conference on Learning Representations, 2021. 1",
+            "[89] Yi Zhu, Karan Sapra, Fitsum A Reda, Kevin J Shih, Shawn Newsam, Andrew Tao, and Bryan Catanzaro. Improving semantic segmentation via video propagation and label relaxation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8856-8865, 2019. 2, 3, 6"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            612
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "9577",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 11
+    }
+]

2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/fa7524e1-a76d-4844-9361-43c8b8046bf9_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc15db8dd049638629bf26f8c28502c4c186f10218a672b3100206f81601f834
+size 5219511

2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/full.md

@@ -0,0 +1,395 @@
+# SemARFlow: Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving
+
+Shuai Yuan, Shuzhi Yu, Hannah Kim, and Carlo Tomasi  
+Duke University
+
+{shuai, shuzhiyu, hannah, tomasi}@cs.duke.edu
+
+# Abstract
+
+Unsupervised optical flow estimation is especially hard near occlusions and motion boundaries and in low-texture regions. We show that additional information such as semantics and domain knowledge can help better constrain this problem. We introduce SemARFlow, an unsupervised optical flow network designed for autonomous driving data that takes estimated semantic segmentation masks as additional inputs. This additional information is injected into the encoder and into a learned upsampler that refines the flow output. In addition, a simple yet effective semantic augmentation module provides self-supervision when learning flow and its boundaries for vehicles, poles, and sky. Together, these injections of semantic information improve the KITTI-2015 optical flow test error rate from  $11.80\%$  to  $8.38\%$ . We also show visible improvements around object boundaries as well as a greater ability to generalize across datasets. Code is available at https://github.com/duke-vision/semantic-unsup-flow-release.
+
+# 1. Introduction
+
+Optical flow estimation, i.e., pixel-level motion tracking across video frames, has broad applications in many computer vision tasks that include object tracking [61], video editing [14, 31], and autonomous driving [6, 58].
+
+Thanks to the success of deep convolutional neural networks [34, 17, 36] and transformer networks [33, 69, 88, 28] in computer vision, many top-performing supervised optical flow networks have been proposed in recent years [10, 65, 22, 68, 85], in which ground-truth labels supervise training. However, real optical flow is hard to label, so most supervised methods train (or at least pre-train) on synthetic datasets [10, 44, 4, 56] which makes them hard to adapt to real applications due to the significant gap between synthetic and real data [70, 29].
+
+Due to label scarcity, unsupervised training of optical
+
+![](images/9e65a6cce7dfec507031a1407f7e49721073b9180ae490256c7b968ba3697ddc.jpg)  
+Figure 1. An example on KITTI-2015 test set (sample #63). Only one input frame is shown for conciseness. Our SemARFlow takes additional semantic segmentation inputs (estimated by an off-the-shelf model) and outputs much sharper flow around semantic boundaries.
+
+![](images/343f36ac10bb396cab4abe94b28fbb130cf086e612f0ae4d6d44fd01bbc59452.jpg)
+
+![](images/81f42782b8b7e34d6a081d106bfccb2d41438995034e911df052f59146cd50c3.jpg)
+
+![](images/85dc6b3c2243ce70f0fa7c131cc53de504af6f1b5e13f883b0520d0b57d3ac48.jpg)
+
+![](images/b36e7fc92d3acc591e035ba081195e897be82f0c0d8276200009904df42750ff.jpg)
+
+flow estimators [81] instead uses loss terms based on the assumptions of constant brightness and smooth flow [45, 27, 42]. Some self-supervision techniques have also been studied to enhance model performance [38, 39, 37]. Unsupervised training makes it possible to train flow networks directly on large real datasets from the target domain.
+
+Even so, unsupervised optical flow estimation is a poorly constrained problem. Brightness is not constant in regions with occlusions [72] or on shiny surfaces [43], nor is it smooth across motion boundaries [30, 82]. Moreover, points in regions with poor texture content [59] or in dark shadows [58] are difficult to track as they are easily confused with neighboring points (the so-called aperture problem). This dearth of reliable constraints makes the unsupervised training of flow networks especially challenging.
+
+A natural way to address this issue is to inject additional constraints in the form of semantics and domain knowledge. For example, in autonomous driving applications [84, 6], we have clear expectations about object types and layouts of the scene, as well as prior knowledge of how each type of object typically moves. We focus on this application domain and show that this additional information helps the network achieve better flow results.
+
+To make an autonomous driving system work well in
+
+reality, it is best to train it on real data. However, annotating real driving video with optical flow labels is expensive [83], as it requires careful synchronization and calibration of diverse sensors including cameras, LiDAR, and GPS/IMU [16], aided by some manual annotation and curation based on CAD models of moving objects [46].
+
+In contrast, annotating semantic labels seems much more feasible, and indeed semantic labels are available in most (if not all) existing driving datasets. We consider semantic segmentation because it provides semantics at the pixel level, the same level as optical flow. As one of the most popular and well-studied tasks in modern computer vision, semantic segmentation [40, 47, 87] has been extensively adopted for autonomous driving systems. In this paper, we show that adding semantic segmentation inputs helps improve unsupervised optical flow performance significantly.
+
+Specifically, we first infer semantic segmentation maps using an off-the-shelf model [89], which of course is trained with semantic labels. An encoder with semantic map input is used to aggregate image and semantic features (Sec. 3.1), and a learned upsampler is added into the iterative decoder to refine flow around object boundaries given semantic inputs (Sec. 3.2). We also propose a simple yet effective semantic augmentation module for self-supervision, which provides realistic augmentations specific to the vehicles, poles, and sky classes based on domain knowledge and segmentation maps (Sec. 3.3). An occluder cache is implemented to improve efficiency (Sec. 3.4). Semantic augmentation provides challenging samples for self-supervision, which help train flow better in occluded regions and around foreground object boundaries.
+
+Overall, by injecting semantic segmentation inputs, our SemARFlow network achieves significantly better unsupervised flow results both quantitatively (Sec. 4.3) and qualitatively (Sec. 4.4). Adapted from ARFlow [37], SemARFlow reduces KITTI-2015 [46] test error from  $11.80\%$  to  $8.38\%$  and out-performs the current unsupervised state-of-the-art UPFlow [42]  $(9.38\%)$  by a clear margin. The example in Fig. 1 also demonstrates visible improvements around object boundaries. The effectiveness of each module is justified through an extensive ablation study (Sec. 4.5) and improvement analysis (Sec. 4.6). In addition to performance boost, unsupervised flow networks with additional semantic inputs generalize better across different datasets (Sec. 4.7).
+
+Our research is essentially novel compared to previous approaches. Some early work incorporates semantics in traditional energy-minimization methods [41, 18] for optical flow through geometric constraints such as piece-wise rigid motion and planar surface motion [57, 21, 1, 75]. In comparison, to the best of our knowledge, we are the first to inject semantics into the unsupervised training of recent optical flow networks. Some research also trains a network for segmentation and optical flow jointly [9]. In contrast, we
+
+leverage existing, separately trained segmentation systems both because they are available and because modularity—separating segmentation from flow estimation—is important for the development of large real-application systems such as autonomous driving.
+
+In summary, our contributions are as follows.
+
+- To the best of our knowledge, we are the first to explore adding semantic inputs to assist the unsupervised training of deep optical flow networks.  
+- We propose a simple yet effective network called SemARFlow that achieves state-of-the-art results both quantitatively and qualitatively. Our model works well on real-life occlusions and yields sharp motion boundaries around objects.  
+- We provide full training and inference code as well as trained models to encourage follow-up research.
+
+# 2. Related work
+
+Supervised optical flow Since the introduction of the pioneering deep optical flow network FlowNet [10], more and more top-performing CNN-based flow estimators have been proposed over the years [23, 49, 65, 20, 22, 68]. Recently, vision transformer networks and attention mechanism have also been applied to this problem, and these have achieved state-of-the-art performance on benchmark datasets [76, 26, 85, 64, 77]. The supervised methods are often pre-trained on large synthetic datasets such as FlyingChairs [10] and FlyingThings3D [44] before fine-tuning on the target dataset. However, there is a clear gap between the artificially generated data and real scenarios.
+
+Unsupervised optical flow Due to the lack of ground-truth labels, unsupervised optical flow estimation uses surrogate losses such as photometric loss and smoothness loss to supervise training [81, 55]. To tackle the issues around occlusion regions, various methods have been proposed including estimated occlusion masks [72], forward-backward consistency [45], and multi-frame fusion [25, 53]. To better upsample the flow in the decoder, UPFlow [42] additionally predicts a confidence map and an interpolation flow to guide flow refinement and has become the state-of-the-art unsupervised flow method.
+
+Some latest research has also explored the use of self-supervision to enhance flow prediction. Early methods apply the knowledge distillation technique to train a two-stage teacher-student network [38, 39]. ARFlow [37] further improves this idea by generating reliable self-supervision signals from data transformations, while merging the two training stages into single-stage training with one added loss term. SimFlow [24] replaces the handcrafted features with deep self-supervised features to measure similarity in the
+
+unsupervised losses. SMURF [63] utilizes a RAFT-like structure [68] and applies multi-frame self-supervised training with many technical improvements. Our SemARFlow also uses self-supervision but with the guidance of semantic segmentation, which is much more realistic than guidance without semantics.
+
+Semantic segmentation Semantic segmentation classifies each pixel of the given image into semantic objects. Fully Convolutional Network (FCN) [40] is one of the early CNN-based segmentation methods. It takes inputs of arbitrary sizes and outputs dense pixel-level predictions, becoming one of the popular backbone architectures for follow-up work. Deconvolution network [47] is also proposed to better recover low-level details of the prediction. One main challenge for these systems is that they lack global scene information. Subsequent work addresses this issue by enlarging the receptive field of the network with global pyramid pooling layers as in PSPNet [86], hybrid dilated convolutions [71], and a fast-down-sampling strategy [80, 79]. Attention modules have also shown to help in semantic segmentation by capturing full-image dependencies of all pixels [19, 87] or the semantic inter-dependencies across spatial and channel dimensions [12].
+
+There also exists extensive work on semantic segmentation in the context of autonomous driving [2, 67, 13, 5, 89], thanks to the publication of large-scale driving datasets [8, 46]. To better train the network on a coarsely labeled dataset like KITTI [46], Zhu et al. use a video prediction network SDCNet [52] to synthesize new training samples with relaxed label propagation [89]. Due to their good performance on KITTI [46], we utilize their network models to infer semantic inputs for all our experiments.
+
+Combining semantics and optical flow Though there has been much progress on both semantic segmentation and optical flow estimation, the semantic optical flow problem (how to exploit semantics to help optical flow estimation) has received limited attention in recent years, and the current best results are thus much outdated. Some early methods incorporate semantics through geometric constraints to refine flow on various semantic regions, such as planar regions (using homographies) [57], static regions (using rigid camera motion and epipolar constraints) [21, 75], and rigid objects (estimating rigid motion for each object instance) [1]. However, most methods are traditional flow methods based on energy minimization, where an initial flow estimate is usually needed and semantics is mostly used for refinement. In comparison, we explore adapting latest unsupervised optical flow networks to leverage semantic inputs in one single stage of estimation.
+
+Apart from using semantics to help flow, some research has also explored using optical flow to help seman
+
+tic segmentation [51, 35], or to train both tasks jointly [9]. There are also studies on exploiting semantics on some other correspondence matching tasks such as stereo matching [74, 78] and 3D scene flow estimation, where some additional depth cues such as stereo camera inputs [54, 11] or point clouds [60] are needed.
+
+# 3. Method
+
+Our network is adapted from the two-frame version of ARFlow [37], which uses a light-weight PwCNet [65] as its backbone. The inputs are two consecutive frames  $I_1, I_2 \in \mathbb{R}^{H \times W \times 3}$  as well as their semantic segmentation maps  $S_1, S_2 \in \{0, 1, \dots, c\}^{H \times W}$ , where the number of classes  $c$  is 19 as we use the Cityscapes format [8]. A detailed diagram of the structure of our network can be found in the Appendix.
+
+# 3.1. Semantic encoder
+
+We first inject semantic information into the encoder. As shown in Fig. 2(a), shared by each frame  $i \in \{1,2\}$ , separate convolutional layers extract features from image  $I_{i}$  and semantics  $S_{i}$  (one-hot encoded). Features from the two pipelines are concatenated and fed to additional layers. Features at different resolutions  $(H / 2^{l}, W / 2^{l})$  form a pyramid  $\{f_{i}^{(l)} \mid 2 \leq l \leq 6\}$ . The semantic information in these features helps delineate objects in dark shadows, where appearance is more homogeneous.
+
+# 3.2. Iterative decoder with a learned upsampler
+
+Following [22, 37], an iterative residual refinement decoder starts from zero estimate  $F_{1\rightarrow 2}^{(7)} = 0$ . For iteration  $l \in \{6,5,4,3,2\}$ , the decoder refines feature map  $F_{1\rightarrow 2}^{(l + 1)}$  into  $F_{1\rightarrow 2}^{(l)}$  based on features  $f_{1}^{(l)}, f_{2}^{(l)}$  (Fig. 2(b)).
+
+More specifically,  $F_{1\rightarrow 2}^{(l + 1)}$  is upsampled to match the resolution and used to warp  $f_2^{(l)}$  to yield warped feature  $\hat{f}_1^{(l)}$ . Correlation volumes are computed between  $f_1^{(l)}$  and  $\hat{f}_1^{(l)}$ . A one-by-one convolutional layer  $C^{(l)}$  compresses the number of channels to a fixed number so that the same layer can be reused across all iterations as proposed by [22]. A flow estimator network predicts a flow residual to be added to the current estimate, and a context network then aggregates flow information spatially and refines the current flow again.
+
+A learned upsampler network upsamples the final output  $F_{1\rightarrow 2}^{(2)}$  in our system. This is different from ARFlow [37], which simply uses four-fold bilinear interpolation, making the final flow boundaries blurry. In contrast, our model learns to sharpen flow boundaries based on the semantic inputs, which have clear boundaries around moving objects. To this end, we add a convex upsampler network similar to the one in RAFT [68]. Different from UPFlow [42],
+
+![](images/1f578db99d3ccda59acd52464574df776530c318c409b21f7fd6dae0bee7b43c.jpg)  
+(a) shared pyramidal encoder  $(i\in \{1,2\})$
+
+![](images/c30e1a5c14116a469523b3f00f526371d8f8a59ad87c073d3cadb07e3d7d3548.jpg)  
+(b) one iteration (at the  $l$ th-level) of the iterative decoder  $(l\in \{6,5,4,3,2\})$
+
+![](images/4b0eca463840993e1e56b7ce31b8aa97cb087ccf3349b518dc7346f392bc4ee8.jpg)  
+Figure 2. Network structure. See text in Sec. 3.1 and Sec. 3.2 for explanations. More detailed diagrams are in appendix.  
+Figure 3. Illustration of semantic augmentation as self-supervision. See text in Sec. 3.3 for details.
+
+our learned upsampler is only used to upsample from internal estimate  $F_{1\rightarrow 2}^{(l)}$  to output flow  $U_{1\rightarrow 2}^{(l)}$ , but not when we upsample  $F_{1\rightarrow 2}^{(l + 1)}$  at the first step in the decoder iteration. The upsampled  $U_{1\rightarrow 2}^{(l)}$  has resolution  $(H / 2^{l - 2},W / 2^{l - 2})$ , so  $U_{1\rightarrow 2} = U_{1\rightarrow 2}^{(2)}$  (with the original resolution) is the final flow prediction of the network.
+
+# 3.3. Semantic augmentation as self-supervision
+
+ARFlow has a very effective in-network augmentation module that samples random transformations  $\mathcal{T}_{\theta_1},\mathcal{T}_{\theta_2}$  of the flow prediction  $U_{1\rightarrow 2}$  in the first pass of the network and then uses the transformed images  $\hat{I}_1 = \mathcal{T}_{\theta_1}(I_1),\hat{I}_2 =$ $\mathcal{T}_{\theta_2}(I_2)$  in a second pass. The prediction  $U_{1\rightarrow 2}$  is also transformed accordingly and used to self-supervise the output of the second pass. See [37] for details.
+
+We retain the augmentation module but make a third pass of the network using semantics-transformed inputs for self-supervision in addition to the ARFlow augmentations of ap
+
+pearance (e.g., color jitter, random noise) and spatial transformations (e.g., random rotation, random rescaling). The idea behind semantic augmentation is to blend in real object motions across samples.
+
+To this end (see Fig. 3), we carve objects from other samples based on their semantic maps and paste them as moving foreground objects into the current sample  $I_{1}, I_{2}$ , thereby producing augmented images  $\tilde{I}_{1}, \tilde{I}_{2}$ . The segmentation maps  $S_{1}, S_{2}$  and the first pass output flow  $U_{1 \to 2}$  are transformed accordingly, and the transformed  $\tilde{U}_{1 \to 2}$  self-supervises the third-pass network outputs  $\tilde{U}_{1 \to 2}' = \mathcal{F}(\tilde{I}_{1}, \tilde{I}_{2}, \tilde{S}_{1}, \tilde{S}_{2})$ , where  $\mathcal{F}$  is the flow network.
+
+Similar to occlusion hallucination in SelFlow [39], our semantic augmentation also creates new occlusions and uses the reliable non-occluded flow from the first pass to self-supervise the flow on those newly occluded pixels. However, rather than superpixels filled with noise, we create occlusions with real objects, which are much more realistic.
+
+Moreover, since we determine the motion of each oc
+
+cluder, we can use its known optical flow for motion self-supervision. As illustrated in Fig. 3, the occluder motion can be very different from that of the background. This trains the model to assume less smoothness around highly-dynamic objects and to output sharper motion estimates near motion boundaries. These hallucinated but realistic occluders also provide more examples of actively moving objects, which are relatively scarce in the original datasets.
+
+**Vehicles** We augment vehicle classes (car, truck, bus, train) as they are major sources of errors in previous models (see Tab. 5 for statistics). We group these four classes because they are easily confused with each other in the semantic input. To carve out a single car instance with high probability, we find connected components of vehicle segmentation regions using OpenCV [3] and accept masks whose bounding boxes are 50-300 pixels wide and 50-150 pixels high. We also drop masks that occupy less than  $60\%$  area of their bounding boxes to remove largely occluded vehicles.
+
+Poles We augment pole classes (pole, traffic light, traffic sign) not only because they are common foreground objects that cause occlusions, but also because their motions are usually poorly estimated due to their thin shapes. We use a 200-pixel wide, image-height sliding window to scan the segmentation of these pole classes and find the window with the most pole pixels. If more than  $10\%$  pole pixels occur in it, that object mask is stored. This produces groups of nearby poles rather than just individual poles, which may not provide enough augmentation by themselves.
+
+Sky We also augment the sky through prior knowledge that flow in the sky should be very small. We shrink the sky flow estimate by half before using it for self-supervision. As shown in Fig. 3, the middle tree part in  $U_{1\rightarrow 2}$  is very blurry, with a large part of sky mixed in there. After shrinking sky flow by half, the tree in  $\tilde{U}_{1\rightarrow 2}$  now has sharper flow around the boundary between tree and sky. Using this as self-supervision helps the model sharpen the motion boundaries between sky and other regions.
+
+Another way to apply this prior knowledge is to postprocess, i.e., shrink the final output flow in the sky region, instead of using the shrunk flow to self-supervise. However, the latter approach is preferable because it allows the network to balance prior knowledge (self-supervision) with the current sample observation (photometric loss). Our experiments do show cases where some small objects (such as some overhead electric power lines) are misclassified as part of the sky in the semantic input. In these cases, the network learns to rely less on prior knowledge.
+
+# 3.4.Occluder cache
+
+During training, we maintain a cache of occluding objects for semantic augmentation. We sample objects from the cache to augment the current batch and then push any new objects found in the batch into the cache, with random replacement if the cache is full.
+
+Specifically, after we load a sample for training, we first use the current model to infer its optical flow and then search within the sample for any cars (or poles) that meet the standards defined in Sec. 3.3 as "cutouts". For each cutout, we first find the flow vectors on all pixels that belong to that cutout and then compute the mean flow for later reference. We store the current sample, the cutout mask, and the mean flow together in the cache.
+
+Later on, when we randomly select a cutout to augment another sample, we also retrieve the stored mean flow of this cutout, which is then augmented by random rescaling (by 0.8-1.5 times) and reversing (with  $50\%$  probability). We use the augmented mean flow to translate the entire cutout object as an occluder to generate a new sample. For realism, occluders are pasted to the same location they occupied in their image of origin. Holding occluders in a cache provides a random mixture of samples across multiple batches and makes semantic augmentation very efficient.
+
+# 3.5. Loss functions
+
+Photometric loss In the first pass, forward and backward flow  $U_{1\rightarrow 2}^{(l)}, U_{2\rightarrow 1}^{(l)}$  at each level are predicted and occlusion masks  $O_{1\rightarrow 2}^{(l)}, O_{2\rightarrow 1}^{(l)}$  are computed by forward-backward consistency [45]. Frames are warped by
+
+$$
+I _ {i} ^ {\prime (l)} (\pmb {p}) = I _ {j} ^ {(l)} (\pmb {p} + U _ {i \rightarrow j} ^ {(l)} (\pmb {p})), \quad (i, j \in \{1, 2 \})
+$$
+
+where  $I_{i}^{(l)}$  is  $I_{i}$  down-sampled to the  $l$ -th scale, and  $\pmb{p}$  denotes pixel coordinates at that scale. Following [37], three measures, namely  $L_{1}$ -distance  $(\rho_{1})$ , structural similarity (SSIM) [73]  $(\rho_{2})$ , and census loss [45]  $(\rho_{3})$  are linearly combined to measure photometric differences between  $I_{i}^{(l)}$  and  $I_{i}^{\prime(l)}$ . Occlusion regions are masked out and both forward and backward directions are taken into account at each level. The final photometric loss is
+
+$$
+\ell_ {\mathrm {p h}} = \frac {1}{2} \sum_ {(i, j) \in \{(1, 2), (2, 1) \}} \sum_ {l = 2} ^ {6} a _ {l} \sum_ {k = 1} ^ {3} c _ {k} \rho_ {k} \left(I _ {i} ^ {(l)}, I _ {i} ^ {\prime (l)}, O _ {i \rightarrow j} ^ {(l)}\right). \tag {1}
+$$
+
+Smoothness loss Unlike most previous work, we do not include a smoothness loss because we find it to conflict with our learned upsampler. See ablation study in Sec. 4.5.
+
+Augmentation loss As in ARFlow [37], our second forward pass computes the  $\mathrm{L}_1$ -distance between the trans-
+
+<table><tr><td rowspan="2" colspan="2">Method</td><td colspan="2">Train</td><td colspan="6">Test</td></tr><tr><td>2012 EPE</td><td>2015 EPE</td><td>2012 Fl-noc</td><td>EPE</td><td>Fl-all</td><td>2015 Fl-noc</td><td>Fl-bg</td><td>Fl-fg</td></tr><tr><td rowspan="4">supervised</td><td>PWC-Net+ [66]</td><td>-</td><td>(1.50)</td><td>3.36</td><td>1.4</td><td>7.72</td><td>4.91</td><td>7.69</td><td>7.88</td></tr><tr><td>IRR-PWC [22]</td><td>-</td><td>(1.63)</td><td>3.21</td><td>1.6</td><td>7.65</td><td>4.86</td><td>7.68</td><td>7.52</td></tr><tr><td>RAFT [68]</td><td>-</td><td>(0.63)</td><td>-</td><td>-</td><td>5.10</td><td>3.07</td><td>4.74</td><td>6.87</td></tr><tr><td>Separable Flow [85]</td><td>-</td><td>(0.69)</td><td>-</td><td>-</td><td>4.53</td><td>2.78</td><td>4.25</td><td>5.92</td></tr><tr><td rowspan="8">unsupervised</td><td>SelFlow [39]</td><td>1.69</td><td>4.84</td><td>4.31</td><td>2.2</td><td>14.19</td><td>9.65</td><td>12.68</td><td>21.74</td></tr><tr><td>SimFlow [24]</td><td>-</td><td>5.19</td><td>-</td><td>-</td><td>13.38</td><td>8.21</td><td>12.60</td><td>17.27</td></tr><tr><td>ARFlow [37]</td><td>1.44</td><td>2.85</td><td>-</td><td>1.8</td><td>11.80</td><td>-</td><td>-</td><td>-</td></tr><tr><td>UFlow [27]</td><td>1.68</td><td>2.71</td><td>4.26</td><td>1.9</td><td>11.13</td><td>8.41</td><td>9.78</td><td>17.87</td></tr><tr><td>UPFlow [42]</td><td>1.27</td><td>2.45</td><td>-</td><td>1.4</td><td>9.38</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Ours (baseline)</td><td>1.39</td><td>2.61</td><td>4.30</td><td>1.7</td><td>9.89</td><td>6.98</td><td>8.82</td><td>15.21</td></tr><tr><td>Ours (+enc)†</td><td>1.29</td><td>2.42</td><td>3.97</td><td>1.5</td><td>8.99</td><td>3.97</td><td>8.19</td><td>13.01</td></tr><tr><td>Ours (+enc +aug)†</td><td>1.28</td><td>2.18</td><td>3.90</td><td>1.5</td><td>8.38</td><td>3.90</td><td>7.48</td><td>12.91</td></tr></table>
+
+Table 1. KITTI benchmark results (EPE/px and Fl/%). Metrics evaluated at 'all' (all pixels, default for EPE), 'noc' (non-occlusions), 'bg' (background), and 'fg' (foreground). Key metrics (used in official ranking) are underlined. Our 'baseline' is an adapted ARFlow with added learned upsampler and no smoothness loss. +enc means adding semantic encoder; +aug means adding semantic augmentation. () means evaluation data used in training. - means unavailable.  $\dagger$  denotes models with semantic inputs. For all metrics, lower is better.
+
+formed flow  $\hat{U}_{1\rightarrow 2}$  and the second pass output  $\hat{U}_{1\rightarrow 2}^{\prime} = \mathcal{F}(\hat{I}_1,\hat{I}_2,\hat{S}_1,\hat{S}_2)$  for each pixel  $\pmb{p}$  as
+
+$$
+\hat {D} (\boldsymbol {p}) = \left\| \hat {U} _ {1 \rightarrow 2} (\boldsymbol {p}) - \hat {U} _ {1 \rightarrow 2} ^ {\prime} (\boldsymbol {p}) \right\| _ {1}.
+$$
+
+This loss is then averaged over the transformed non-occluded region where self-supervision  $\hat{U}_{1\rightarrow 2}$  is accurate
+
+$$
+\ell_ {\mathrm {a r}} = \frac {\sum_ {\boldsymbol {p}} (1 - \hat {O} _ {1 \rightarrow 2} (\boldsymbol {p})) \hat {D} (\boldsymbol {p})}{\sum_ {\boldsymbol {p}} (1 - \hat {O} _ {1 \rightarrow 2} (\boldsymbol {p}))}. \tag {2}
+$$
+
+For semantic augmentation (third pass), since we are also doing self-supervision with a new generated sample, we use a similar loss definition
+
+$$
+\ell_ {\mathrm {a u g}} = \frac {\sum_ {\boldsymbol {p}} \left(1 - \tilde {O} _ {1 \rightarrow 2} (\boldsymbol {p})\right) \tilde {D} (\boldsymbol {p})}{\sum_ {\boldsymbol {p}} \left(1 - \tilde {O} _ {1 \rightarrow 2} (\boldsymbol {p})\right)}, \tag {3}
+$$
+
+where  $\tilde{D} (\pmb {p})$  is the distance between the semanticaugmented flow  $\tilde{U}_{1\rightarrow 2}$  and the third-pass output  $\tilde{U}_{1\rightarrow 2}^{\prime}$
+
+$$
+\tilde {D} (\boldsymbol {p}) = \left\| \tilde {U} _ {1 \rightarrow 2} (\boldsymbol {p}) - \tilde {U} _ {1 \rightarrow 2} ^ {\prime} (\boldsymbol {p}) \right\| _ {1},
+$$
+
+and the augmented mask  $\tilde{O}_{1\rightarrow 2}$  is now computed by
+
+$$
+\tilde {O} _ {1 \rightarrow 2} = 1 - \max (1 - O _ {1 \rightarrow 2}, M)
+$$
+
+because we penalize on both the originally non-occluded region  $1 - O_{1\rightarrow 2}$  and the pasted foreground object mask  $M$  (of which we know the true motion).
+
+Final loss (with  $\lambda = 0.02$  as in ARFlow [37]):
+
+$$
+\ell = \ell_ {\mathrm {p h}} + \lambda \left(\ell_ {\mathrm {a r}} + \ell_ {\mathrm {a u g}}\right). \tag {4}
+$$
+
+# 4. Experiments
+
+# 4.1. Datasets
+
+We mainly use KITTI [15, 46] and Cityscapes [8] datasets for experiments. Following [37], we train our model first on KITTI raw sequences [46] (55.7k samples) and then fine-tune on KITTI-2015 multi-view extension [46] (11.8k samples). We validate our model using KITTI-2015 train [46] (200 samples) and KITTI-2012 train set [15] (194 samples) since they are the only sets that have optical flow labels. We also train on Cityscapes [8] sequences (83.3k samples) to test model generalization ability, and we sample every other frame to match its frame rate with KITTI. All semantic segmentation inputs are estimated using an off-the-shelf model [89] with a DeepLabV3Plus [7] backbone, which achieves  $83.45\%$  mean IoU on Cityscapes and  $72.82\%$  mean IoU on KITTI.
+
+# 4.2. Implementation details
+
+We implement the model in PyTorch [48]  $^{1}$ . We use the Adam optimizer [32] with  $\beta_{1} = 0.9$ ,  $\beta_{2} = 0.999$  and batch size 4. We first train on KITTI raw sequences [46] for 100k iterations with a fixed learning rate 0.0002 and then train on KITTI multi-view extension set [15, 46] for another 100k iterations using OneCycleLR schedule [62] with maximum learning rate 0.0004 and linear annealing.
+
+For data augmentation, we include random horizontal flipping and swapping of the input frames. We resize the inputs to  $256 \times 832$  before feeding into the network. The photometric loss weight for each scale  $a_{l}$  ( $2 \leq l \leq 6$ ) in
+
+<table><tr><td>Method</td><td>Fl-all</td><td>Fl-noc</td><td>Fl-bg</td><td>Fl-fg</td></tr><tr><td>JFS [21]</td><td>17.07</td><td>9.81</td><td>15.9</td><td>22.92</td></tr><tr><td>SOF [57]</td><td>16.81</td><td>10.86</td><td>14.63</td><td>27.73</td></tr><tr><td>MRFlow [75]</td><td>12.19</td><td>8.86</td><td>10.13</td><td>22.52</td></tr><tr><td>Bai et al. [1]</td><td>11.62</td><td>8.75</td><td>8.61</td><td>26.69</td></tr><tr><td>Ours (final)</td><td>8.38</td><td>3.90</td><td>7.48</td><td>12.91</td></tr></table>
+
+Eq. (1) are set as 1, 1, 1, 1, 0. The weights for three photometric distance measures  $\rho_{k}$  ( $1 \leq k \leq 3$ ) in Eq. (1) are set as 0.15, 0.85, 0 for the first 50k iterations and 0, 0, 1 afterwards. We start the appearance and spatial augmentation (second pass as in ARFlow [37]) at 50k iterations, and semantic augmentation (third pass) after 150k iterations.
+
+# 4.3. Benchmark testing results
+
+As common practice, we evaluate optical flow predictions based on two error measurements, Fl (error rate) and EPE (mean  $\mathrm{L}_2$  distance). When computing Fl, the estimate of each pixel is considered correct if the error is smaller than 3 pixels or  $5\%$  of the magnitude of ground-truth flow [46].
+
+As shown in Tab. 1, our semantic modules (both semantic encoder and semantic augmentation) improve performance on both KITTI-2012 and KITTI-2015 sets on all metrics. Our final model achieves  $8.38\%$  F1-all error rate on KITTI-2015 test set, which is significantly better than ARFlow [37]  $(11.80\%)$ , from which we adapt. We also outperform the current state-of-the-art UPFlow [42]  $(9.38\%)$  by a clear margin. All these results strongly suggest that adding semantic inputs can help improve unsupervised flow estimation significantly.
+
+In addition, Tab. 2 shows that our test results significantly outperform all previous semantic optical flow methods, most of which are based on traditional energy minimization. We are the first to apply semantic inputs to recent unsupervised flow networks, so we are able to push the state-of-the-art by a clear margin.
+
+# 4.4. Qualitative results
+
+Some qualitative results are shown in Fig. 4. We can see that our flow outputs have very sharp boundaries and very clean object motion, especially around cars and poles. Our model successfully learns to adapt flow estimation to the semantic map input. Moreover, our model is able to handle very challenging samples where the foreground car motion is drastically different from the background due to the challenging samples we create in semantic augmentation.
+
+Table 2. KITTI-2015 test results (Fl/\%) compared with other semantic optical flow methods. Metrics evaluated at 'all' (all pixels), 'noc' (non-occlusions), 'bg' (background), and 'fg' (foreground).  
+
+<table><tr><td>up</td><td>no sm</td><td>enc</td><td>aug</td><td>Fl-all</td><td>EPE-all</td><td>EPE-noc</td><td>EPE-occ</td></tr><tr><td></td><td></td><td></td><td></td><td>10.36</td><td>2.90</td><td>2.07</td><td>6.97</td></tr><tr><td></td><td></td><td>1</td><td></td><td>9.92</td><td>2.69</td><td>1.89</td><td>6.57</td></tr><tr><td></td><td></td><td>2</td><td></td><td>9.83</td><td>2.65</td><td>1.86</td><td>6.40</td></tr><tr><td></td><td></td><td>3</td><td></td><td>9.75</td><td>2.61</td><td>1.85</td><td>6.54</td></tr><tr><td></td><td></td><td>4</td><td></td><td>9.75</td><td>2.64</td><td>1.85</td><td>6.55</td></tr><tr><td>✓</td><td></td><td></td><td></td><td>10.22</td><td>2.83</td><td>1.97</td><td>6.75</td></tr><tr><td>✓</td><td>✓</td><td></td><td></td><td>8.87</td><td>2.61</td><td>1.85</td><td>6.04</td></tr><tr><td>✓</td><td>✓</td><td>3</td><td></td><td>8.26</td><td>2.42</td><td>1.73</td><td>5.79</td></tr><tr><td>✓</td><td>✓</td><td></td><td>✓</td><td>8.80</td><td>2.48</td><td>1.60</td><td>6.64</td></tr><tr><td>✓</td><td>✓</td><td>3</td><td>✓</td><td>7.79</td><td>2.18</td><td>1.40</td><td>5.64</td></tr></table>
+
+Table 3. Ablation study on KITTI-2015 train set (EPE/px and F1%). 'up' means the learned upsampler; 'no sm' means turning off smoothness loss; 'enc' means the level at which we merge semantic features and image features in the encoder; 'aug' means semantic augmentation.  
+
+<table><tr><td>Options of aug</td><td>Fl-all</td><td>EPE-all</td><td>EPE-noc</td><td>EPE-occ</td></tr><tr><td>Ours (final)</td><td>7.79</td><td>2.18</td><td>1.40</td><td>5.64</td></tr><tr><td>start from 100k</td><td>7.92</td><td>2.19</td><td>1.40</td><td>5.41</td></tr><tr><td>vehicles only</td><td>7.94</td><td>2.21</td><td>1.44</td><td>5.78</td></tr><tr><td>loss on new occ</td><td>8.15</td><td>2.27</td><td>1.42</td><td>5.62</td></tr></table>
+
+Table 4. Ablation study of different semantic augmentation options on KITTI-2015 train (EPE/px and F1/%). See text for explanations.
+
+# 4.5. Ablation study
+
+We also do ablation study to show the effectiveness of each of our proposed modules. We can see in Tab. 3 that adding a semantic encoder to the vanilla ARFlow does help, and the optimal number of encoder layer groups added is 3. Also, our learned upsampler improves results significantly once we turn off the smoothness loss. This makes sense because the upsampler learns to adapt to boundaries, which can be distracted by the smoothness loss. Moreover, both our semantic encoder and semantic augmentation module help improve the results further, which suggests the effectiveness of adding semantic inputs.
+
+Another ablation study is on different options for the semantic augmentation (Tab. 4). We try starting augmentation earlier from 100k iterations, which works slightly worse. There are mainly two reasons: (1) we use the first forward output to self-supervise our augmented output, so if we start early, the model will use poor output to self-supervise the augmented pass; and (2) our semantic augmentation creates very challenging samples with objects moving very differently from the background, so exposing these hard samples to the model too early may make it hard to train. Apart from this, we also try only using vehicles to augment as well as focusing loss on the newly occluded region, which are both inferior to our final version.
+
+![](images/07c59636c83be6777bb7d26b345ec42e13f6245ab9cf5eed0a798a8bc2146132.jpg)  
+Image Input
+
+![](images/cbb5321d0383d9cd403ec9c926150713f4acbaadab3bca47ced4520392966d5e.jpg)  
+Semantics Input
+
+![](images/b18ba7a222c78a7ae580df67155847eef12a2edad21db45d5541879811b4cb0a.jpg)  
+ARFlow  
+Figure 4. Qualitative results on KITTI test set (sample #7, 20, 38, 112, 183) compared with ARFlow [37] and UPFlow [42]. We only show first frame input for conciseness. Our outputs have much clearer boundaries around objects. For more qualitative examples, see Appendix.
+
+![](images/81bafe3a42b8dceeb4b7fce071a2fb656aab904f459b663294b6eca2000eff17.jpg)  
+UPFlow
+
+![](images/02089204af00bcd73238d308a144146cad31925f78d7c8c61735a06124c2374c.jpg)  
+Ours
+
+<table><tr><td></td><td>road</td><td>car</td><td>terrain</td><td>vegetation</td><td>sidewalk</td><td>building</td><td>wall</td><td>pole</td></tr><tr><td>Proportion</td><td>42.4%</td><td>17.6%</td><td>14.0%</td><td>11.6%</td><td>6.2%</td><td>4.1%</td><td>1.3%</td><td>1.1%</td></tr><tr><td>ARFlow [37]</td><td>4.57</td><td>15.79</td><td>9.42</td><td>15.34</td><td>4.61</td><td>6.00</td><td>16.34</td><td>10.75</td></tr><tr><td>Ours (final)</td><td>3.67</td><td>10.17</td><td>8.74</td><td>13.47</td><td>3.09</td><td>4.27</td><td>11.18</td><td>9.44</td></tr><tr><td>Relative improvement</td><td>19.6%</td><td>35.6%</td><td>7.3%</td><td>12.2%</td><td>33.0%</td><td>28.7%</td><td>31.6%</td><td>12.2%</td></tr><tr><td>Reweighed contribution</td><td>19.0%</td><td>49.4%</td><td>4.8%</td><td>10.9%</td><td>4.7%</td><td>3.5%</td><td>3.4%</td><td>0.7%</td></tr></table>
+
+# 4.6. Improvement analysis
+
+It may be interesting to understand where our improvements come from in terms of semantic classes, so we compute the KITTI-2015 train set error for each class. As shown in Tab. 5, our model improves the flow on every class. After reweighing the absolute improvement of each class based on their proportion in the evaluated pixels, we find that car flow accounts for nearly half of our improvement on Fl-all, indicating the effectiveness of augmenting vehicle objects.
+
+# 4.7. Generalization ability
+
+We also investigate the generalization ability of our semantic-aided flow models. We train flow models on Cityscapes and directly test them on KITTI-2015 train set without fine-tuning. Following [50], we crop the bottom  $25\%$  of the frame to remove the car logo and resize the frame to size  $256 \times 704$ . All other experiment settings are exactly the same as for KITTI. As we can see from Tab. 6, adding semantic inputs significantly helps our unsupervised flow model generalize and adapt better across datasets.
+
+# 5. Conclusion and future work
+
+In this paper, we show that adding semantic segmentation inputs can help significantly improve the performance
+
+Table 5. KITTI-2015 flow error (Fl-all/%) for each semantic class. The first row shows the proportion of each class among evaluated pixels (KITTI flow does not evaluate on every pixel). We only show classes that account for at least  $1\%$  of the evaluated pixels.  
+
+<table><tr><td>Method</td><td>Fl-all</td><td>EPE-all</td><td>EPE-noc</td><td>EPE-occ</td></tr><tr><td>ARFlow (our impl.)</td><td>13.21</td><td>4.08</td><td>2.88</td><td>9.40</td></tr><tr><td>Ours (baseline)</td><td>12.27</td><td>3.81</td><td>2.43</td><td>9.91</td></tr><tr><td>Ours (+enc)†</td><td>11.28</td><td>3.33</td><td>2.12</td><td>8.75</td></tr><tr><td>Ours (+enc +aug)†</td><td>10.32</td><td>2.64</td><td>1.56</td><td>7.12</td></tr></table>
+
+Table 6. Generalization results (train on Cityscapes, and test on KITTI-2015 train). We use our implementation of ARFlow [37] to run this experiment. Our 'baseline' is an adapted ARFlow with added learned upsampler and no smoothness loss.  $^+$  enc means adding semantic encoder;  $^+$  aug means adding semantic augmentation.  $\dagger$  denotes models with semantic inputs.
+
+of unsupervised optical flow networks on autonomous driving datasets. We propose a novel network model called SemARFlow, a semantic adaptation of ARFlow [37], with a semantic encoder, a learned upsampler, and a semantic augmentation module, where some domain-specific motion prior has been taken into account. Our network better predicts flow at occlusion regions and effectively sharpens flow estimates around object boundaries, even for very challenging samples. The additional semantic inputs also make our network generalize better across datasets.
+
+One direction for future work is motivated by the example illustrated in Fig. 5. Our model does output sharp flow
+
+![](images/1163bd8c232875b98653a17d74378498bfe4ef69b40c6cd4a3f95ac58e27a07a.jpg)  
+Figure 5. An example on KITTI-2015 train set (sample #164). See the discussion in Sec. 5 for details.
+
+![](images/d3f5dfb78ff75b20c7534879e6a07823a7a3f0014c676ba3b095e4f7f671dee7.jpg)
+
+![](images/a7bdac7bbcf7508d90aa410a55955595fee7f51f1cf2bf6e69324c96fcd7f87f.jpg)
+
+boundaries between the cars and the background, thanks to good boundary information from the semantic maps. However, the boundary between these two fast-moving cars is inferred from image information only, because their semantic maps merge into one. As a result, that part of the flow boundary is less accurate. An interesting direction for future work is to use instance-level semantics as input for more detailed object masks.
+
+Acknowledgments. This research is based upon work supported by the National Science Foundation under Grant No. 1909821.
+
+# References
+
+[1] Min Bai, Wenjie Luo, Kaustav Kundu, and Raquel Urtasun. Exploiting semantic information and deep matching for optical flow. In Proceedings of the European Conference on Computer Vision, pages 154-170. Springer, 2016. 2, 3, 7  
+[2] Shubhankar Borse, Ying Wang, Yizhe Zhang, and Fatih Porikli. Inverseform: A loss function for structured boundary-aware segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5901-5911, 2021. 3  
+[3] Gary Bradski. The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer, 25(11):120-123, 2000. 5  
+[4] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Proceedings of the European Conference on Computer Vision, pages 611-625. Springer, 2012. 1  
+[5] Yingfeng Cai, Lei Dai, Hai Wang, and Zhixiong Li. Multi-target pan-class intrinsic relevance driven model for improving semantic segmentation in autonomous driving. IEEE Transactions on Image Processing, 30:9069-9084, 2021. 3  
+[6] Linda Capito, Umit Ozguner, and Keith Redmill. Optical flow based visual potential field for autonomous driving. In IEEE Intelligent Vehicles Symposium, pages 885-891. IEEE, 2020. 1  
+[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision, pages 801-818, 2018. 6  
+[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes
+
+dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016. 3, 6  
+[9] Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10713-10720, 2020. 2, 3  
+[10] Alexey Dosovitskiy, Philipp Fischery, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision, ICCV '15, pages 2758-2766, Washington, DC, USA, 2015. IEEE Computer Society. 1, 2  
+[11] Cheng Feng, Long Ma, Congxuan Zhang, Zhen Chen, Liyue Ge, and Shaofeng Jiang. Ss-sf: Piecewise 3d scene flow estimation with semantic segmentation. IEEE Access, 9:22745-22759, 2021. 3  
+[12] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3146-3154, 2019. 3  
+[13] Aditya Ganeshan, Alexis Vallet, Yasunori Kudo, Shin-ichi Maeda, Tommi Kerola, Rares Ambrus, Dennis Park, and Adrien Gaidon. Warp-refine propagation: Semi-supervised auto-labeling via cycle-consistency. In Proceedings of the IEEE International Conference on Computer Vision, pages 15499–15509, 2021. 3  
+[14] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In Proceedings of the European Conference on Computer Vision, pages 713-729. Springer, 2020. 1  
+[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research, 32(11):1231-1237, 2013. 6  
+[16] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 2  
+[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1  
+[18] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial Intelligence, 17(1-3):185–203, 1981. 2  
+[19] 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 Conference on Computer Vision and Pattern Recognition, pages 603-612, 2019. 3  
+[20] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE Conference
+
+on Computer Vision and Pattern Recognition, pages 8981-8989, 2018. 2  
+[21] Junhwa Hur and Stefan Roth. Joint optical flow and temporally consistent semantic segmentation. In Proceedings of the European Conference on Computer Vision, pages 163-177. Springer, 2016. 2, 3, 7  
+[22] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2019. 1, 2, 3, 6  
+[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1647-1655, 2017. 2  
+[24] Woobin Im, Tae-Kyun Kim, and Sung-Eui Yoon. Unsupervised learning of optical flow with deep feature similarity. In Proceedings of the European Conference on Computer Vision, pages 172-188. Springer, 2020. 2, 6  
+[25] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European Conference on Computer Vision, pages 690-706, 2018. 2  
+[26] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE International Conference on Computer Vision, pages 9772-9781, 2021. 2  
+[27] Rico Jonschkowski, Austin Stone, Jonathan T Barron, Ariel Gordon, Kurt Konolige, and Anelia Angelova. What matters in unsupervised optical flow. In Proceedings of the European Conference on Computer Vision, pages 557-572. Springer, 2020. 1, 6  
+[28] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys, 54(10s):1-41, 2022. 1  
+[29] Hannah H Kim, Celia Cintas, Girmaw Abebe Tadesse, and Skyler Speakman. Spatially constrained adversarial attack detection and localization in the representation space of optical flow networks. In International Joint Conference on Artificial Intelligence, 2023. 1  
+[30] Hannah Halin Kim, Shuzhi Yu, and Carlo Tomasi. Joint detection of motion boundaries and occlusions. In Proceedings of the British Machine Vision Conference, 2021. 1  
+[31] Hannah Halin Kim, Shuzhi Yu, Shuai Yuan, and Carlo Tomasi. Cross-attention transformer for video interpolation. In Proceedings of the Asian Conference on Computer Vision Workshops, pages 320–337, 2022. 1  
+[32] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Proceedings of the International Conference on Learning Representations, 2014. 6  
+[33] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recogni
+
+tion at scale. In Proceedings of the International Conference on Learning Representations, 2021. 1  
+[34] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1  
+[35] Xiangtai Li, Jiangang Bai, Kuiyuan Yang, and Yunhai Tong. Flow2seg: Motion-aided semantic segmentation. In Proceedings of the International Conference on Artificial Neural Networks, pages 225–237. Springer, 2019. 3  
+[36] Zewen Li, Fan Liu, Wenjie Yang, Shouheng Peng, and Jun Zhou. A survey of convolutional neural networks: analysis, applications, and prospects. IEEE transactions on neural networks and learning systems, 2021. 1  
+[37] Liang Liu, Jiangning Zhang, Ruifei He, Yong Liu, Yabiao Wang, Ying Tai, Donghao Luo, Chengjie Wang, Jilin Li, and Feiyue Huang. Learning by analogy: Reliable supervision from transformations for unsupervised optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6489-6498, 2020. 1, 2, 3, 4, 5, 6, 7, 8  
+[38] Pengpeng Liu, Irwin King, Michael R Lyu, and Jia Xu. Ddflow: Learning optical flow with unlabeled data distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8770-8777, 2019. 1, 2  
+[39] Pengpeng Liu, Michael Lyu, Irwin King, and Jia Xu. Selfflow: Self-supervised learning of optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4571-4580, 2019. 1, 2, 4, 6  
+[40] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431-3440, 2015. 2, 3  
+[41] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In International Joint Conference on Artificial Intelligence, volume 2, pages 674-679, 1981. 2  
+[42] Kunming Luo, Chuan Wang, Shuaicheng Liu, Haoqiang Fan, Jue Wang, and Jian Sun. Upflow: Upsampling pyramid for unsupervised optical flow learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1045-1054, 2021. 1, 2, 3, 6, 7, 8  
+[43] Rémi Marsal, Florian Chabot, Angélique Loesch, and Hichem Sahbi. Brightflow: Brightness-change-aware unsupervised learning of optical flow. In Proceedings of the IEEE winter conference on applications of computer vision, pages 2061-2070, 2023. 1  
+[44] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4040-4048, 2016. 1, 2  
+[45] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5  
+[46] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE Conference
+
+on Computer Vision and Pattern Recognition, 2015. 2, 3, 6, 7  
+[47] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1520-1528, 2015. 2, 3  
+[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035. Curran Associates, Inc., 2019. 6  
+[49] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2720-2729, 2017. 2  
+[50] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12240–12249, 2019. 8  
+[51] Hazem Rashed, Senthil Yogamani, Ahmad El-Sallab, Pavel Krizek, and Mohamed El-Helw. Optical flow augmented semantic segmentation networks for automated driving. In International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 165-172, 2019. 3  
+[52] Fitsum A Reda, Guilin Liu, Kevin J Shih, Robert Kirby, Jon Barker, David Tarjan, Andrew Tao, and Bryan Catanzaro. Sdc-net: Video prediction using spatially-displaced convolution. In Proceedings of the European Conference on Computer Vision, pages 718-733, 2018. 3  
+[53] Zhile Ren, Orazio Gallo, Deqing Sun, Ming-Hsuan Yang, Erik B Sudderth, and Jan Kautz. A fusion approach for multiframe optical flow estimation. In Proceedings of the IEEE winter conference on applications of computer vision, pages 2077-2086. IEEE, 2019. 2  
+[54] Zhile Ren, Deqing Sun, Jan Kautz, and Erik Sudderth. Cascaded scene flow prediction using semantic segmentation. In Proceedings of the International Conference on 3D Vision, pages 225-233. IEEE, 2017. 3  
+[55] Zhe Ren, Junchi Yan, Bingbing Ni, Bin Liu, Xiaokang Yang, and Hongyuan Zha. Unsupervised deep learning for optical flow estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 2  
+[56] Stephan R. Richter, Zeeshan Hayder, and Vladlen Koltun. Playing for benchmarks. In Proceedings of the IEEE International Conference on Computer Vision, pages 2232-2241, 2017. 1  
+[57] Laura Sevilla-Lara, Deqing Sun, Varun Jampani, and Michael J Black. Optical flow with semantic segmentation and localized layers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3889-3898, 2016. 2, 3, 7
+
+[58] Shihao Shen, Louis Kerofsky, and Senthil Yogamani. Optical flow for autonomous driving: Applications, challenges and improvements. arXiv preprint arXiv:2301.04422, 2023. 1  
+[59] Jianbo Shi and Carlo Tomasi. Good features to track. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 593-600, 1994. 1  
+[60] Yukang Shi and Kaiheng Ma. Safit: Segmentation-aware scene flow with improved transformer. In Proceedings of the International Conference on Robotics and Automation, pages 10648-10655. IEEE, 2022. 3  
+[61] Jeongho Shin, Sangjin Kim, Sangkyu Kang, Seong-Won Lee, Joonki Paik, Besma Abidi, and Mongi Abidi. Optical flow-based real-time object tracking using non-prior training active feature model. Real-time imaging, 11(3):204-218, 2005. 1  
+[62] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 6  
+[63] Austin Stone, Daniel Maurer, Alper Ayvaci, Anelia Angelova, and Rico Jonschkowski. Smurf: Self-teaching multiframe unsupervised raft with full-image warping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3887-3896, 2021. 3  
+[64] Xiuchao Sui, Shaohua Li, Xue Geng, Yan Wu, Xinxing Xu, Yong Liu, Rick Goh, and Hongyuan Zhu. Craft: Cross-attentional flow transformer for robust optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17602-17611, 2022. 2  
+[65] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8934-8943, 2018. 1, 2, 3  
+[66] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Models matter, so does training: An empirical study of cnns for optical flow estimation. IEEE transactions on pattern analysis and machine intelligence, 42(6):1408-1423, 2019. 6  
+[67] Andrew Tao, Karan Sapra, and Bryan Catanzaro. Hierarchical multi-scale attention for semantic segmentation. arXiv preprint arXiv:2005.10821, 2020. 3  
+[68] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision, pages 402-419. Springer, 2020. 1, 2, 3, 6  
+[69] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1  
+[70] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE Con
+
+ference on Computer Vision and Pattern Recognition Workshops, pages 969-977, 2018. 1  
+[71] Panqu Wang, Pengfei Chen, Ye Yuan, Ding Liu, Zehua Huang, Xiaodi Hou, and Garrison Cottrell. Understanding convolution for semantic segmentation. In Proceedings of the IEEE winter conference on applications of computer vision, pages 1451-1460. IEEE, 2018. 3  
+[72] Yang Wang, Yi Yang, Zhenheng Yang, Liang Zhao, Peng Wang, and Wei Xu. Occlusion aware unsupervised learning of optical flow. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4884-4893, 2018. 1, 2  
+[73] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5  
+[74] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Song Wang, and Lili Ju. Semantic stereo matching with pyramid cost volumes. In Proceedings of the IEEE International Conference on Computer Vision, pages 7484-7493, 2019. 3  
+[75] Jonas Wulff, Laura Sevilla-Lara, and Michael J Black. Optical flow in mostly rigid scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4671-4680, 2017. 2, 3, 7  
+[76] Haofei Xu, Jiaolong Yang, Jianfei Cai, Juyong Zhang, and Xin Tong. High-resolution optical flow from 1d attention and correlation. In Proceedings of the IEEE International Conference on Computer Vision, pages 10498-10507, 2021. 2  
+[77] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 2  
+[78] Guorun Yang, Hengshuang Zhao, Jianping Shi, Zhidong Deng, and Jiaya Jia. Segstereo: Exploiting semantic information for disparity estimation. In Proceedings of the European Conference on Computer Vision, pages 636-651, 2018. 3  
+[79] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129:3051-3068, 2021. 3  
+[80] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European Conference on Computer Vision, pages 325-341, 2018. 3  
+[81] Jason J. Yu, Adam W. Harley, and Konstantinos G. Derpanis. Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In Gang Hua and Hervé Jégou, editors, Proceedings of the European Conference on Computer Vision Workshops, pages 3-10, Cham, 2016. Springer International Publishing. 1, 2  
+[82] Shuzhi Yu, Hannah H Kim, Shuai Yuan, and Carlo Tomasi. Unsupervised flow refinement near motion boundaries. In
+
+Proceedings of the British Machine Vision Conference. BMVA Press, 2022. 1  
+[83] Shuai Yuan, Xian Sun, Hannah Kim, Shuzhi Yu, and Carlo Tomasi. Optical flow training under limited label budget via active learning. In Proceedings of the European Conference on Computer Vision, pages 410-427, 2022. 2  
+[84] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1  
+[85] Feihu Zhang, Oliver J Woodford, Victor Adrian Prisacariu, and Philip HS Torr. Separable flow: Learning motion cost volumes for optical flow estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 10807-10817, 2021. 1, 2, 6  
+[86] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2881-2890, 2017. 3  
+[87] 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 Conference on Computer Vision and Pattern Recognition, pages 6881-6890, 2021. 2, 3  
+[88] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In Proceedings of the International Conference on Learning Representations, 2021. 1  
+[89] Yi Zhu, Karan Sapra, Fitsum A Reda, Kevin J Shih, Shawn Newsam, Andrew Tao, and Bryan Catanzaro. Improving semantic segmentation via video propagation and label relaxation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8856-8865, 2019. 2, 3, 6

2023/SemARFlow_ Injecting Semantics into Unsupervised Optical Flow Estimation for Autonomous Driving/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebb38f954c58a16387c5b5fae12db4f4d8f8b030471693e8c4b95bdb3c8e6628
+size 551527

2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/68ea1ef8-9cb5-44da-894d-92e02632af66_content_list.json

@@ -0,0 +1,2254 @@
+[
+    {
+        "type": "text",
+        "text": "Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition",
+        "text_level": 1,
+        "bbox": [
+            96,
+            102,
+            870,
+            125
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yiqing Liang",
+        "bbox": [
+            114,
+            143,
+            222,
+            162
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Eliot Laidlaw",
+        "bbox": [
+            243,
+            143,
+            354,
+            161
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Alexander Meyerowitz Brown University",
+        "bbox": [
+            377,
+            143,
+            560,
+            180
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Srinath Sridhar",
+        "bbox": [
+            581,
+            143,
+            705,
+            160
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "James Tompkin",
+        "bbox": [
+            727,
+            143,
+            854,
+            162
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            209,
+            313,
+            226
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "From video, we reconstruct a neural volume that captures time-varying color, density, scene flow, semantics, and attention information. The semantics and attention let us identify salient foreground objects separately from the background across spacetime. To mitigate low resolution semantic and attention features, we compute pyramids that trade detail with whole-image context. After optimization, we perform a saliency-aware clustering to decompose the scene. To evaluate real-world scenes, we annotate object masks in the NVIDIA Dynamic Scene and DyCheck datasets. We demonstrate that this method can decompose dynamic scenes in an unsupervised way with competitive performance to a supervised method, and that it improves foreground/background segmentation over recent static/dynamic split methods. Project webpage: https://visual.cs.brown.edu/saff",
+        "bbox": [
+            75,
+            238,
+            473,
+            465
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            479,
+            207,
+            496
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Given a casually-captured monocular RGB video of a dynamic scene, decomposing it into foreground objects and background is an important task in computer vision, with downstream applications in segmentation and video editing. An ideal method would also reconstruct the geometry and appearance over time including frame correspondences.",
+        "bbox": [
+            75,
+            503,
+            470,
+            595
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Previous methods have made great progress but there are many limitations. Some works assume that there is no object motion in the scene [32, 46, 31, 16, 4], take input from multi-view cameras [16, 46], or do not explicitly reconstruct the underlying 3D structure of the scene [4, 7]. For objects, some works rely on masks or user input to aid segmentation [42, 15, 7], or use task-specific training datasets [7]. Sometimes, works assume the number of foreground objects [4, 20]. Given the challenges, many works train and test on synthetic data [14, 38].",
+        "bbox": [
+            75,
+            597,
+            470,
+            747
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We present Semantic Attention Flow Fields (SAFF): A method to overcome these limitations by integrating low-level reconstruction cues with high-level pretrained information—both bottom-up and top-down—into a neural volume. With this, we demonstrate that embedded semantic and saliency (attention) information is useful for unsupervised dynamic scene decomposition. SAFF builds upon neural scene flow fields [18], an approach that reconstructs appearance, geometry, and motion. This uses frame interpolation rather than explicit canonicalization [35] or a latent hyperspace [26],",
+        "bbox": [
+            75,
+            750,
+            472,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/01fefb4a9e45aeceefff10089d5bae4b9533e86d532a1f53df7e2cc944385a37.jpg",
+        "image_caption": [
+            "Density  $\\sigma$ , Saliency  $a$ , Semantics  $s$ , Color  $c$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            511,
+            198,
+            879,
+            364
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/17f14ffcf60fce11f834a50b881e8ff330d777688d5b719bbc157a87104f5d21.jpg",
+        "image_caption": [
+            "Novel view renderings with effects (depth of field, color adjustment)",
+            "Figure 1: We decompose a dynamic 3D scene from monocular input into time-varying color, density, scene flow, semantics, and attention information. This could be used to provide precise editing of objects or effects focused around objects."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            513,
+            382,
+            694,
+            455
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/41eff40600d65b7aa1eeea11fea2b17c3386a06b748cc4ca9b57cb24afc10184.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            694,
+            382,
+            875,
+            455
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "which lets it more easily apply to casual videos. For optimization, we supervise two network heads with pretrained DINO-ViT [6] semantic features and attention. Naively supervising high-resolution SAFF with low-resolution DINO-ViT output reduces reconstruction quality. To mitigate the mismatch, we build a semantic attention pyramid that trades detail with whole-image context. Having optimized a SAFF representation for a dynamic scene, we perform a saliency-aware clustering both in 3D and on rendered feature images to describe objects and their background. Given the volume reconstruction, the clustering generalizes to novel spacetime views.",
+        "bbox": [
+            496,
+            551,
+            893,
+            718
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To evaluate SAFF's dynamic scene decomposition capacity, we expand the NVIDIA Dynamic Scene [45] and DyCheck [10] datasets by manually annotating object masks across input and hold-out views. We demonstrate that SAFF outperforms 2D DINO-ViT baselines and is comparable to a state-of-the-art video segmentation method ProposeReduce[19] on our data. Existing monocular video dynamic volume reconstruction methods typically separate static and dynamic parts, but these often do not represent meaningful foreground. We show improved foreground segmentation over NSFF and the current  $\\mathrm{D}^2\\mathrm{NeRF}$  [40] method for downstream tasks like editing.",
+        "bbox": [
+            496,
+            719,
+            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_number",
+        "text": "21797",
+        "bbox": [
+            478,
+            924,
+            519,
+            936
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            89,
+            217,
+            104
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Decomposing a scene into regions of interest is a long-studied task in computer vision [5], including class, instance, and panoptic segmentation in high-level or top-down vision, and use of low-level or bottom-up cues like motion. Recent progress has considered images [12, 4, 20], videos [14, 15], layer decomposition [44, 24], and in-the-wild databases using deep generative models [25]. One example,  $\\mathrm{SAVi}++[7]$ , uses slot attention [20] to define 2D objects in real-world videos. Providing first-frame segmentation masks achieves stable performance, with validation on the driving Waymo Open Dataset [33]. Our work attempts 3D scene decomposition for a casual monocular video without initial masks.",
+        "bbox": [
+            75,
+            112,
+            470,
+            292
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Scene Decomposition with NeRFs Neural Radiance Fields (NeRF) [41] have spurred new scene decomposition research through volumes. ObSuRF [32] and uORF [46] are unsupervised slot attention works that bind a latent code to each object. Unsupervised decomposition is also possible on light fields [31]. For dynamic scenes, works like NeuralDiff [37] and  $\\mathrm{D}^2\\mathrm{NeRF}$  [40] focus on foreground separation, where foreground is defined to contain moving objects. Other works like N3F [36] and occlusions-4d [13] also decompose foregrounds into individual objects. N3F requires user input to specify which object to segment, and occlusions-4d takes RGB point clouds as input. Our work attempts to recover a segmented dynamic scene from a monocular RGB video without added markup or masks.",
+        "bbox": [
+            75,
+            310,
+            470,
+            521
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Neural Fields Beyond Radiance Research has begun to add non-color information to neural volumes to aid decomposition via additional feature heads on the MLP. iLabel [47] adds a semantic head to propagate user-provided segmentations in the volume. PNF [17] and Panoptic-NeRF [9] attempt panoptic segmentation within neural fields, and Object-NeRF integrates instance segmentation masks into the field during optimization [42]. Research also investigates how to apply generic pretrained features to neural fields, like DINO-ViT.",
+        "bbox": [
+            75,
+            537,
+            468,
+            675
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "DINO-ViT for Semantics and Saliency DINO-ViT is a self-supervised transformer that, after pretraining, extracts generic semantic information [6]. Amir et al. [1] use DINO-ViT features with  $k$ -means clustering to achieve co-segmentation across a video. Seitzer et al. [29] combine slot attention and DINO-ViT features for object-centric learning on real-world 2D data. TokenCut [39] performs normalized cuts on DINO-ViT features for foreground segmentation on natural images. Deep Spectral Segmentation [22] show that graph Laplacian processing of DINO-ViT features provides unsupervised foreground segmentation, and Selfmask [30] shows that these features can provide object saliency masks. Our approach considers these clustering and saliency findings for the setting of 3D decomposition from monocular video.",
+        "bbox": [
+            75,
+            691,
+            470,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "DINO-ViT Fields Concurrent works have integrated DINO-ViT features into neural fields. DFF [16] distills features for dense multi-view static scenes with user input for segmentation. N3F [36] expands NeuralDiff to dynamic scenes, and relies on user input for segmentation. AutoLabel [3] uses DINO-ViT features to accelerate segmentation in static scenes given a ground truth segmentation mask. Other works use DINO-ViT differently. FeatureRealistic-Fusion [21] uses DINO-ViT in an online feature fusion task, focusing on propagating user input, and NeRF-SOS [8] uses 2D DINO-ViT to process NeRF-rendered multi-view RGB images of a static scene. In contrast to these works, we consider real-world casual monocular videos, recover and decompose a 3D scene, then explore whether saliency can avoid the need for masks or user input in segmenting objects.",
+        "bbox": [
+            496,
+            90,
+            893,
+            319
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            500,
+            329,
+            589,
+            344
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "For a baseline dynamic scene reconstruction method, we begin with NSFF from Li et al. [18] (Sec. 3.1), which builds upon NeRF [23]. NSFF's low-level scene flow frame-to-frame approach provides better reconstructions for real-world casual monocular videos than deformation-based methods [35, 26]. We modify the architecture to integrate higher-level semantic and saliency (or attention) features (Sec. 3.2). After optimizing a SAFF for each scene, we perform saliency-aware clustering of the field (Sec. 3.4). All implementation details are in our supplemental material.",
+        "bbox": [
+            496,
+            349,
+            893,
+            501
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Input Our method takes in a single RGB video over time  $i$  as an ordered set of images  $I \\in \\mathcal{I}$  and camera poses. We use COLMAP to recover camera poses [28]. From all poses, we define an NDC-like space that bounds the scene, and a set of rays  $\\mathbf{r} \\in \\mathcal{R}$ , one per image pixel with color  $\\hat{\\mathbf{c}}^\\dagger$ . Here,  $\\hat{\\mathbf{c}}$  denotes a 2D pixel value in contrast to a 3D field value, and  $\\hat{\\mathbf{c}}^\\dagger$  denotes an input value in contrast to an estimated value.",
+        "bbox": [
+            496,
+            512,
+            893,
+            619
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "From pretrained networks, we estimate single-frame monocular depth  $\\hat{d}^{\\dagger}$  (MiDaSv2 [27]), optical flow  $\\hat{\\mathbf{p}}_i^\\dagger$  (RAFT [34]), and semantic features  $\\hat{\\mathbf{s}}^\\dagger$  and attention  $\\hat{\\mathbf{a}}^\\dagger$  (DINO-ViT [6]) after important preprocessing (Sec. 3.3).",
+        "bbox": [
+            496,
+            619,
+            890,
+            679
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1. Initial Dynamic Neural Volume Representation",
+        "text_level": 1,
+        "bbox": [
+            498,
+            686,
+            890,
+            704
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The initial representation comprises a static NeRF  $F_{\\theta}^{\\mathrm{st}}$  and a dynamic NeRF  $F_{\\theta}^{\\mathrm{dy}}$ . The static network predicts a color  $\\mathbf{c}$ , density  $\\sigma$ , and blending weight  $v$  (Eq. (1)), and the dynamic network predicts time-varying color  $\\mathbf{c}_i$ , density  $\\sigma_i$ , scene flow  $\\mathbf{f}_i$ , and occlusion weights  $w_i$  (Eq. (2)). In both network architectures, other than position  $\\mathbf{x}$ , we add direction  $\\boldsymbol{\\omega}$  to a late separate head to only condition the estimation of color  $\\mathbf{c}$ .",
+        "bbox": [
+            496,
+            707,
+            893,
+            814
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {\\boldsymbol {\\theta}} ^ {\\mathrm {s t}}: (\\mathbf {x}, \\omega) \\rightarrow \\left(\\mathbf {c} ^ {\\mathrm {s t}}, \\sigma^ {\\mathrm {s t}}, v\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            581,
+            823,
+            890,
+            840
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {\\boldsymbol {\\theta}} ^ {\\mathrm {d y}}: (\\mathbf {x}, \\boldsymbol {\\omega}, i) \\rightarrow \\left(\\mathbf {c} _ {i} ^ {\\mathrm {d y}}, \\sigma_ {i} ^ {\\mathrm {d y}}, \\mathbf {f} _ {i}, w _ {i}\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            581,
+            843,
+            890,
+            862
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To produce a pixel's color, we sample points at distances  $t$  along the ray  $\\mathbf{x}_t = \\mathbf{x} - \\omega t$  between near and far planes  $t_n$",
+        "bbox": [
+            496,
+            869,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "21798",
+        "bbox": [
+            478,
+            924,
+            519,
+            936
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/f11cce526596a99c53cb277dfee5c4ba6fe1f8e2e53d703a83c76fedd620f2ad.jpg",
+        "image_caption": [
+            "(a) SAFF video inputs"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            106,
+            349,
+            303
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/08f48acf6d9bc6c16bccd6de7bb8553c7cb55375f61b01b26ae38212d33b08ce.jpg",
+        "image_caption": [
+            "(b) SAFF model"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            375,
+            112,
+            573,
+            300
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/8f5ce03d1d46c9edbba3fab841f598fd38e818f838815b8436f0ed4d1f04ab7e.jpg",
+        "image_caption": [
+            "(c) Volume rendered outputs",
+            "Figure 2: Overview. From monocular video, SAFF builds a neural field of scene-flowed 3D density, radiance, semantics, and attention (b). This is guided by semantic attention pyramids that increase resolution, plus depth and optical flow priors (a). We can render new spacetime views of any channel, and use saliency-aware clustering to decompose objects and background (c)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            598,
+            108,
+            879,
+            303
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "to  $t_f$ , query each network, then integrate transmittance  $T$ , density  $\\sigma$ , and color along the ray from these samples [23]. For brevity, we omit evaluating at  $\\mathbf{x}_t, \\boldsymbol{\\omega}$ , e.g.,  $\\sigma^{\\mathrm{st}}(\\mathbf{x}_t)$  is simply  $\\sigma^{\\mathrm{st}}$ ;  $\\mathbf{c}^{\\mathrm{st}}(\\mathbf{x}_t, \\boldsymbol{\\omega})$  is simply  $\\mathbf{c}^{\\mathrm{st}}$ .",
+        "bbox": [
+            75,
+            364,
+            470,
+            424
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We produce a combined color from the static and dynamic colors by multiplication with their densities:",
+        "bbox": [
+            76,
+            424,
+            468,
+            454
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\sigma_ {i} \\mathbf {c} _ {i} = v \\sigma^ {\\mathrm {s t}} \\mathbf {c} ^ {\\mathrm {s t}} + (1 - v) \\sigma_ {i} ^ {\\mathrm {d y}} \\mathbf {c} _ {i} ^ {\\mathrm {d y}} \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            460,
+            468,
+            479
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given that transmittance integrates density up to the current sampled point under Beer-Lambert volume attenuation, the rendered pixel color for the ray is computed as:",
+        "bbox": [
+            76,
+            491,
+            468,
+            537
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {c}} _ {i} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i} \\sigma_ {i} \\mathbf {c} _ {i} d t \\text {w h e r e} T _ {i} = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma_ {i} d t\\right) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            86,
+            542,
+            468,
+            579
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To optimize the volume to reconstruct input images, we compute a photometric loss  $L_{\\hat{\\mathbf{c}}}$  between rendered and input colors:",
+        "bbox": [
+            75,
+            589,
+            468,
+            633
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\hat {\\mathbf {c}}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\left\\| \\hat {\\mathbf {c}} _ {i} \\left(\\mathbf {r} _ {i}\\right) - \\hat {\\mathbf {c}} _ {i} ^ {\\dagger} \\left(\\mathbf {r} _ {i}\\right) \\right\\| _ {2} ^ {2} \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            155,
+            642,
+            468,
+            679
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Scene Flow Defining correspondence over time is important for monocular input as it lets us penalize a reprojection error with neighboring frames  $j \\in \\mathcal{N}$ , e.g., where  $j = i + 1$  or  $j = i - 1$ . We denote  $i \\rightarrow j$  for the projection of frame  $i$  onto frame  $j$  by scene flow.  $F_{\\theta}^{\\mathrm{dy}}$  estimates both forwards and backwards scene flow at every point to penalize a bi-direction loss.",
+        "bbox": [
+            75,
+            693,
+            468,
+            786
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Thus, we approximate color output at time step  $j$  by flowing the queried network values at time step  $i$ :",
+        "bbox": [
+            76,
+            786,
+            468,
+            816
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {c}} _ {i \\rightarrow j} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i \\rightarrow j} \\sigma_ {i \\rightarrow j} \\mathbf {c} _ {i \\rightarrow j} d t \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            825,
+            468,
+            859
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Reprojection must account for occlusion and disocclusion by motion. As such,  $F_{\\theta}^{\\mathrm{dy}}$  also predicts forwards and backwards",
+        "bbox": [
+            76,
+            869,
+            468,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "scene occlusion weights  $w_{i + 1}$  and  $w_{i - 1} \\in [0,1]$ , where a point with  $w_{i + 1} = 0$  means that occlusion status is changed one step forwards in time. We can integrate  $w$  to a pixel:",
+        "bbox": [
+            496,
+            364,
+            892,
+            410
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {w} _ {i \\rightarrow j} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i \\rightarrow j} \\sigma_ {i \\rightarrow j} w _ {j} d t \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            596,
+            415,
+            892,
+            450
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Then, this pixel weight modulates the color loss such that occluded pixels are ignored. (8)",
+        "bbox": [
+            496,
+            457,
+            892,
+            491
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\hat {\\mathbf {c}} _ {i \\rightarrow j}} = \\frac {1}{| \\mathcal {R} | | \\mathcal {N} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\sum_ {j \\in \\mathcal {N}} \\hat {w} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) | | \\hat {\\mathbf {c}} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) - \\hat {\\mathbf {c}} _ {j} ^ {\\dagger} (\\mathbf {r} _ {j}) | | _ {2} ^ {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            500,
+            493,
+            892,
+            530
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Prior losses We use the pretrained depth and optical flow map losses to help overcome the ill-posed monocular reconstruction problem. These losses decay as optimization progresses to rely more and more on the optimized self-consistent geometry and scene flow. For geometry, we estimate a depth  $\\hat{d}_i$  for each ray  $\\mathbf{r}_i$  by replacing  $\\mathbf{c}_i$  in Eq. (4) by the distance  $t$  along the ray. Transform  $z$  estimates a scale and shift as the pretrained network produces only relative depth.",
+        "bbox": [
+            496,
+            545,
+            893,
+            667
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\hat {d}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\left| \\left| \\hat {d} _ {i} - z \\left(\\hat {d} _ {i} ^ {\\dagger}\\right) \\right| \\right| _ {1} \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            593,
+            670,
+            892,
+            708
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For motion, projecting scene flow to a camera lets us compare to the estimated optical flow. Each sample point along a ray  $\\mathbf{x}_i$  is advected to a point in the neighboring frame  $\\mathbf{x}_{i\\rightarrow j}$ , then integrated to the neighboring camera plane to produce a 2D point offset  $\\hat{\\mathbf{p}}_i(\\mathbf{r}_i)$ . Then, we expect the difference in the start and end positions to match the prior:",
+        "bbox": [
+            496,
+            714,
+            892,
+            806
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\hat {\\mathbf {p}}} = \\frac {1}{| \\mathcal {R} | | \\mathcal {N} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\sum_ {j \\in \\mathcal {N} (i)} | | \\hat {\\mathbf {p}} _ {i} (\\mathbf {r} _ {i}) - \\hat {\\mathbf {p}} _ {i} ^ {\\dagger} (\\mathbf {r} _ {i}) | | _ {1} \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            527,
+            809,
+            892,
+            848
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Additional regularizations encourage occlusion weights to be close to one, scene flow to be small, locally constant, and cyclically consistent, and blending weight  $v$  to be sparse.",
+        "bbox": [
+            496,
+            854,
+            893,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "21799",
+        "bbox": [
+            478,
+            924,
+            517,
+            936
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Semantic Attention Flow Fields",
+        "text_level": 1,
+        "bbox": [
+            76,
+            90,
+            354,
+            104
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Beyond low-level or bottom-up features, high-level or top-down features are also useful to define objects and help down-stream tasks like segmentation. For example, methods like NSFF or  $\\mathrm{D}^2\\mathrm{NeRF}$  struggle to provide useful separation of static and dynamic parts because blend weight  $v$  estimates whether the volume appears to be occupied by some moving entity. This is not the same as objectness; tasks like video editing could benefit from accurate dynamic object masks.",
+        "bbox": [
+            75,
+            113,
+            468,
+            234
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As such, we extract 2D semantic features and attention (or saliency) values from a pretrained DINO-ViT network, then optimize the SAFF such that unknown 3D semantic and attention features over time can be projected to recreate their 2D complements. This helps us to ascribe semantic meaning to the volume and to identify objects. As semantics/attention are integrated into the 4D volume, we can render them from novel spacetime views without further DINO-ViT computation.",
+        "bbox": [
+            75,
+            234,
+            470,
+            354
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To estimate semantic features  $\\mathbf{s}$  and attention  $\\mathbf{a}$  at 3D points in the volume at time  $i$ , we add two new heads to both the static  $F_{\\theta}^{\\mathrm{st}}$  and the dynamic  $F_{\\theta}^{\\mathrm{dy}}$  networks:",
+        "bbox": [
+            76,
+            354,
+            468,
+            402
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {\\boldsymbol {\\theta}} ^ {\\mathrm {s t}}: (\\mathbf {x}, \\omega) \\quad \\rightarrow \\left(\\mathbf {c} ^ {\\mathrm {s t}}, \\sigma^ {\\mathrm {s t}}, v, \\mathbf {s} ^ {\\mathrm {s t}}, \\mathbf {a} ^ {\\mathrm {s t}}\\right) \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            133,
+            407,
+            468,
+            425
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {\\boldsymbol {\\theta}} ^ {\\mathrm {d y}}: (\\mathbf {x}, \\boldsymbol {\\omega}, i) \\rightarrow \\left(\\mathbf {c} _ {i} ^ {\\mathrm {d y}}, \\sigma_ {i} ^ {\\mathrm {d y}}, \\mathbf {f} _ {i}, w _ {i}, \\mathbf {s} _ {i} ^ {\\mathrm {d y}}, \\mathbf {a} _ {i} ^ {\\mathrm {d y}}\\right) \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            130,
+            428,
+            468,
+            448
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As semantic features have been demonstrated to be somewhat robust to view dependence [1], in our architectures both heads for s, a are taken off the backbone before  $\\omega$  is injected.",
+        "bbox": [
+            76,
+            454,
+            468,
+            500
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To render semantics from the volume, we replace the color term  $\\mathbf{c}$  in Eq. (4) with  $\\mathbf{s}$  and equivalently for  $\\mathbf{a}$ :",
+        "bbox": [
+            76,
+            500,
+            468,
+            529
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\sigma_ {i} \\mathbf {s} _ {i} = v \\sigma^ {\\mathrm {s t}} \\mathbf {s} ^ {\\mathrm {s t}} + (1 - v) \\sigma_ {i} ^ {\\mathrm {d y}} \\mathbf {s} _ {i} ^ {\\mathrm {d y}}, \\hat {\\mathbf {s}} _ {i} = \\int_ {t _ {n}} ^ {t _ {f}} T _ {i} \\sigma_ {i} \\mathbf {s} _ {i} d t \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            86,
+            534,
+            468,
+            568
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To encourage scene flow to respect semantics over time, we penalize complementary losses on s and a (showing s only):",
+        "bbox": [
+            76,
+            575,
+            468,
+            604
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\hat {\\mathbf {s}} _ {i \\rightarrow j}} = \\frac {1}{| \\mathcal {R} | | \\mathcal {N} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\sum_ {j \\in \\mathcal {N}} \\hat {w} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) \\| \\hat {\\mathbf {s}} _ {i \\rightarrow j} (\\mathbf {r} _ {i}) - \\hat {\\mathbf {s}} _ {j} ^ {\\dagger} (\\mathbf {r} _ {j}) \\| _ {2} ^ {2} \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            76,
+            607,
+            468,
+            652
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Finally, as supervision, we add respective losses on the reconstruction of the 2D semantic and attention features from projected 3D volume points (showing s only):",
+        "bbox": [
+            76,
+            660,
+            468,
+            705
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\hat {\\mathbf {s}}} = \\frac {1}{| \\mathcal {R} |} \\sum_ {\\mathbf {r} _ {i} \\in \\mathcal {R}} \\left\\| \\hat {\\mathbf {s}} _ {i} \\left(\\mathbf {r} _ {i}\\right) - \\hat {\\mathbf {s}} _ {i} ^ {\\dagger} \\left(\\mathbf {r} _ {i}\\right) \\right\\| _ {2} ^ {2} \\tag {15}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            156,
+            708,
+            468,
+            744
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Unlike depth and scene flow priors, these are not priors—there is no self-consistency for semantics to constrain their values. Thus, we do not decay their contribution. While decaying avoids disagreements between semantic and attention features and color-enforced scene geometry, it also leads to a loss of useful meaning (please see supplemental).",
+        "bbox": [
+            76,
+            751,
+            470,
+            842
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Thus our final loss becomes:",
+        "bbox": [
+            96,
+            842,
+            284,
+            856
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {S A F F}} = \\mathcal {L} _ {\\hat {\\mathbf {c}}} + \\lambda_ {\\hat {\\mathbf {c}} _ {i \\rightarrow j}} \\mathcal {L} _ {\\hat {\\mathbf {c}} _ {i \\rightarrow j}} + \\lambda_ {\\hat {d}} \\mathcal {L} _ {\\hat {d}} + \\lambda_ {\\hat {\\mathbf {p}}} \\mathcal {L} _ {\\hat {\\mathbf {p}}} \\tag {16} \\\\ + \\lambda_ {\\hat {\\mathbf {s}} _ {i \\rightarrow j}} \\mathcal {L} _ {\\hat {\\mathbf {s}} _ {i \\rightarrow j}} + \\lambda_ {\\hat {\\mathbf {a}} _ {i \\rightarrow j}} \\mathcal {L} _ {\\hat {\\mathbf {a}} _ {i \\rightarrow j}} + \\lambda_ {\\hat {\\mathbf {s}}} \\mathcal {L} _ {\\hat {\\mathbf {s}}} + \\lambda_ {\\hat {\\mathbf {a}}} \\mathcal {L} _ {\\hat {\\mathbf {a}}} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            109,
+            862,
+            468,
+            898
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Semantic Attention Pyramids",
+        "text_level": 1,
+        "bbox": [
+            500,
+            90,
+            763,
+            106
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "When thinking about scenes, we might argue that semantics from an ideal extractor should be scale invariant, as distant objects have the same class as close objects. We might also argue that saliency (or attention features) may not be scale invariant, as small details in a scene should only be salient when viewed close up. In practice, both extracted features vary across scale and have limited resolution, e.g., DINO-ViT [6] produces one output for each  $8 \\times 8$  patch. But, from this, we want semantic features and saliency for every RGB pixel that still respects scene boundaries.",
+        "bbox": [
+            496,
+            113,
+            893,
+            263
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Thus far, work on static scenes has ignored the input/feature resolution mismatch [16] as multi-view constraints provide improved localization within the volume. For monocular video, this approach has limitations [36]. Forming many constraints on dynamic objects requires long-term motion correspondence—a tricky task—and so we want to maximize the resolution of any input features where possible without changing their meaning.",
+        "bbox": [
+            496,
+            265,
+            893,
+            385
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "One way may be through a pyramid of semantic and attention features that uses a sliding window approach at finer resolutions. Averaging features could increase detail around edges, but we must overcome the practical limit that these features are not stable across scales. This is especially important for saliency: unlike typical RGB pyramids that must preserve energy in an alias-free way [2], saliency changes significantly over scales and does not preserve energy.",
+        "bbox": [
+            496,
+            385,
+            893,
+            506
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Consider a feature pyramid  $\\mathcal{P}$  with loss weights per level:",
+        "bbox": [
+            517,
+            506,
+            892,
+            521
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathcal {P} \\hat {\\mathbf {s}}} = \\sum_ {i \\in \\mathcal {P}} \\lambda_ {\\hat {\\mathbf {s}}} ^ {i} \\mathcal {L} _ {\\hat {\\mathbf {s}}} ^ {i} \\quad \\mathcal {L} _ {\\mathcal {P} \\hat {\\mathbf {a}}} = \\sum_ {i \\in \\mathcal {P}} \\lambda_ {\\hat {\\mathbf {a}}} ^ {i} \\mathcal {L} _ {\\hat {\\mathbf {a}}} ^ {i} \\tag {17}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            568,
+            541,
+            892,
+            574
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Naively encouraging scale-consistent semantics and whole-image saliency, e.g.,  $\\lambda_{\\hat{\\mathbf{s}}} = \\{1/3, 1/3, 1/3\\}$  with  $\\lambda_{\\hat{\\mathbf{a}}} = \\{1, 0, 0\\}$ , empirically leads to poor recovered object edges because the balanced semantics and coarse saliency compete over where the underlying geometry is. Instead, we weight both equally  $\\lambda_{\\hat{\\mathbf{s}}} = \\lambda_{\\hat{\\mathbf{a}}} = \\{1/9, 4/9, 4/9\\}$ . Even though the coarse layer has smaller weight, it is sufficient to guide the overall result. This balances high resolution edges from fine layers and whole object features from coarse layers while reducing geometry conflicts, and leads to improved features (Fig. 3).",
+        "bbox": [
+            496,
+            583,
+            893,
+            734
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Of course, any sliding window must contain an object to extract reliable features for that object. At coarse levels, an object is always in view. At fine levels, an object is only captured in some windows. Objects of interest tend to be near the middle of the frame, meaning that boundary windows at finer pyramid levels contain features that less reliably capture those objects. This can cause spurious connections in clustering. To cope with this, we relatively decrease finer level boundary window weights: We upsample all levels to the finest level, then increase the coarsest level weight towards the frame boundary to  $\\lambda_{\\hat{\\mathbf{s}}} = \\lambda_{\\hat{\\mathbf{a}}} = \\{1/3, 1/3, 1/3\\}$ .",
+        "bbox": [
+            496,
+            734,
+            893,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "21800",
+        "bbox": [
+            478,
+            924,
+            519,
+            936
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/f178d6f08ab968f5f60529102575cf05b82ce1c9023a855566a217b585eb38d3.jpg",
+        "image_caption": [
+            "(a) Input"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            109,
+            99,
+            258,
+            174
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/9b0a274bb513cdbaff27d2e781ba72fa72821d8b34d8ed9cb154a712427de7bd.jpg",
+        "image_caption": [
+            "(b) Semantics"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            259,
+            99,
+            408,
+            174
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/ca4d6cc4b08415640f4ab2ce014bf87ff955f0bdf83c5692ffb80dd75522697d.jpg",
+        "image_caption": [
+            "(c) Semantics (volume)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            410,
+            99,
+            558,
+            174
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/286f833094b4816d0f26ea379f6c302a9a40006555badeca5a9734de5a57cb54.jpg",
+        "image_caption": [
+            "(d) Saliency"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            560,
+            99,
+            707,
+            174
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/6073298e54a3d420c359a7e00e8a065e7a90a15f4532e6dd95c82ee33325699a.jpg",
+        "image_caption": [
+            "(e) Saliency (volume)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            709,
+            99,
+            856,
+            174
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/9a772ca97f6d34a23e449345bc01ca9b625edcc4b2d9d0c0e4c5afca1890d8ad.jpg",
+        "image_caption": [
+            "(f) Pyr. semantics"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            259,
+            186,
+            408,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/b1d357ef4844ecadfd229b10f4610f2f43a0e7ebec2c3b41c59afbd833772873.jpg",
+        "image_caption": [
+            "(g) Our pyr. semantics (volume)",
+            "Figure 3: Semantics and saliency improve by both volume integration and by our pyramid. On Balloon NBoard, blocky artifacts are removed. Semantics are visualized as most significant three PCA dimensions; specific colors are less meaningful."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            410,
+            186,
+            557,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/db2932d48e70bef0de273f03533759cd8762fef47e784615bda46b636c792bcb.jpg",
+        "image_caption": [
+            "(h) Pyr. saliency"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            560,
+            186,
+            707,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/d9cad011c3b648a0848faa5ed6182fbf8a504b14c318a601d52790f011237b06.jpg",
+        "image_caption": [
+            "(i) Our pyr. saliency (volume)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            710,
+            186,
+            856,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4. Using SAFF for Saliency-aware Clustering",
+        "text_level": 1,
+        "bbox": [
+            76,
+            314,
+            439,
+            330
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We now wish to isolate salient objects. Even in dynamic scenes, relevant objects may not move, so analyzing dynamic elements is insufficient (cf. [40]). One approach predicts segmentation end-to-end [19]. However, end-to-end learning requires priors about the scene provided by supervision, and even large-scale pretraining might fail given unseen test scenes. To achieve scene-specific decompositions from per-video semantics, inspired by Amir et al. [1], we design clustering for spacetime volumes that allows segmenting novel spatio-temporal views. While DINO-ViT is trained on images, its features are loosely temporally consistent [6].",
+        "bbox": [
+            75,
+            338,
+            468,
+            505
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Some works optimize a representation with a fixed number of clusters, e.g., via slot attention [20] in NeRFs [32, 46]. Instead, we cluster using elbow  $k$ -means, letting us adaptively find the number of clusters after optimization. This is more flexible than baking an anticipated number of slots (sometimes with fixed semantics), and lets us cluster and segment in novel spatio-temporal viewpoints.",
+        "bbox": [
+            75,
+            506,
+            470,
+            611
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We demonstrate clustering results in both 3D over time and on rendered volume 2D projections over time. Given the volume reconstruction, we might think that clustering directly in 3D would be better. But, monocular input with narrow baselines makes it challenging to precisely reconstruct geometries: Consider that depth integrated a long a ray can still be accurate even though geometry at specific 3D points may be inaccurate or 'fluffy'. As such, we use the 2D volume projection clustering results in 2D comparisons.",
+        "bbox": [
+            75,
+            612,
+            468,
+            750
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Method For 3D over time, we sample points from the SAFF uniformly along input rays  $(128\\times H\\times W)$  and treat each pixel as a feature point, e.g., semantics are 64 dim. and saliency is 1 dim. For volume projection, we render to  $N$  input poses  $(N\\times H\\times W)$ , and treat each pixel as a feature point instead. In either case, we cluster all feature points together using elbow  $k$ -means to produce an initial set of separate regions. For each cluster  $c$ , for each image, we calculate the mean attention of all feature points within the cluster  $\\bar{\\mathbf{a}}_c$ . If  $\\bar{\\mathbf{a}}_c > 0.07$ , then this",
+        "bbox": [
+            75,
+            763,
+            470,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "cluster is salient for this image. Finally, all feature points vote on saliency: if more than  $70\\%$  agree, the cluster is salient.",
+        "bbox": [
+            496,
+            315,
+            890,
+            345
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Salient objects may still be split into semantic parts: e.g., in Fig. 4, the person's head/body are separated. Plus, unwanted background saliency may exist, e.g., input  $\\hat{\\mathbf{a}}^\\dagger$  is high for the teal graphic on the wall. As such, before saliency voting, we merge clusters whose centroids have a cosine similarity  $>0.5$ . This reduces the first problem as heads and bodies are similar, and reduces the second problem as merging the graphic cluster into the background reduces its average saliency (Fig. 4).",
+        "bbox": [
+            496,
+            345,
+            892,
+            467
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To extract an object from the 3D volume, we sample 3D points along each input ray, then ascribe the label from the semantically-closest centroid. All clusters not similar to the stored salient clusters are marked as background with zero density. For novel space-time views, we render feature images from the volume, then assign cluster labels to each pixel according to its similarity with stored input view centroids.",
+        "bbox": [
+            496,
+            467,
+            893,
+            573
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            587,
+            630,
+            604
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We show the impact of adding semantic and saliency features through scene decomposition and foreground experiments. Our website contains supplemental videos.",
+        "bbox": [
+            496,
+            613,
+            890,
+            657
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Data: NVIDIA Dynamic Scene Dataset (Masked) This data [45] has 8 sequences of 24 time steps formed from 12 cameras simultaneously capturing video. We manually annotate object masks for view and time step splits; we will release this data publicly. We define three data splits per sequence:",
+        "bbox": [
+            496,
+            670,
+            893,
+            747
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Input: A monocular camera that moves position for every timestep is simulated from the input sequences; we use Yoon et al.'s input sequences [45].",
+            "2. Fix Cam 0 (hold out): We fix the camera at position 0 as time plays, requiring novel view and time synthesis.  $\\{(\\mathrm{cam}_0,\\mathrm{time}_i),i\\in [1,2,\\dots 23]\\}$",
+            "3. Fix Time 0 (hold out): We fix time at step 0 as the camera moves, requiring novel view and time synthesis.  $\\{(\\mathrm{cam}_i,\\mathrm{time}_0),i\\in [1,2,\\ldots 11]\\}$"
+        ],
+        "bbox": [
+            500,
+            752,
+            893,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "21801",
+        "bbox": [
+            478,
+            924,
+            517,
+            936
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/0916a3dac95539f8c91aff80d1f8a07ab4e1ba67931550ee0b2cae13ffc89eda.jpg",
+        "image_caption": [
+            "(i) Input frame",
+            "(ii) Input semantic/attention"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            104,
+            99,
+            251,
+            172
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/c5ce24e922c13f4821e3240ca0f0021f73e48b716f94b30e793395bc9e3a343b.jpg",
+        "image_caption": [
+            "(a) Our pyr. semantic (volume)",
+            "(d) Our pyr. saliency (volume)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            253,
+            99,
+            401,
+            172
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5d567c20160ac212f558b1a90c90f18dce7c52124b47c9276e2fdd6d4ca11381.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            403,
+            89,
+            563,
+            172
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/f462e2415f3c4072ff3dd65adef4b2717184e0446ea7a0119253bc18cc96ab9d.jpg",
+        "image_caption": [
+            "(e) Our merged clustering",
+            "Figure 4: Saliency-aware clustering improves decomposition. On Dynamic Face, the head and body are semantically and saliently different, but are mutually different from the background. This allows us to extract a time-varying 3D field of the object."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            410,
+            184,
+            558,
+            258
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/83fe39f2429a8082a84ae5b94ad37683b9859938142cb82fa7d83523de749201.jpg",
+        "image_caption": [
+            "(c) Salient clusters (pre-merge)",
+            "(f) Our final foreground",
+            "(iii) Extraction (fixed time)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            566,
+            99,
+            715,
+            172
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5cc523c6ad484004e2d5657e5123e5dbf9683ce2a9478a19116031c1ea14a0d0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            566,
+            185,
+            715,
+            258
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5cbf3a9ab54ff469420060cb15eb9ca62dcd439ff5e73fb19525cc2d36915a10.jpg",
+        "image_caption": [
+            "(iv) Extraction (fixed view)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            718,
+            119,
+            866,
+            172
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d2014859e5f8596c4f049fd7a5542a994321f44e6e401f6f045a53168d008f72.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            718,
+            190,
+            866,
+            220
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/e20bdacede3a7a5868b7cef4fcdcf593b23fa2655dfd89336a1d4e0a92985e35.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            718,
+            227,
+            866,
+            258
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/09c1891a830a23fbc620a22ac5e9c58856b78f5efaea2cf8f1da9005971890f1.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            720,
+            301,
+            867,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/7def0160a5af8dfbc2da02f54623763411d8065e9bb925784ad0a2b1dd18007b.jpg",
+        "image_caption": [
+            "(a) Input",
+            "Figure 5: SAFF object segmentations show balanced quality while recovering a volumetric scene representation (e). Basic DINO-ViT produces low-quality segmentations and misses objects. A state-of-the-art 2D video learning method [19] sometimes has edge detail (Umbrella, legs) but othertimes misses detail and objects (Balloon NBoard)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            316,
+            261,
+            508
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/4b13ac5c9c77ea60444552e5524c3e891235297addb0747db1403a807f583b6e.jpg",
+        "image_caption": [
+            "(b) DINO-ViT (2D)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            264,
+            316,
+            411,
+            508
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/6e89c0f7d9b9ee8d30bfbc5bec0ae618274ab0497d01755188af532b00ba8da6.jpg",
+        "image_caption": [
+            "(c) ProposeReduce [19]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            416,
+            316,
+            563,
+            508
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/a9ed21c21fb0dcc4c27f0d7b78c66d1cf4381b857645386272644700ab7fdba2.jpg",
+        "image_caption": [
+            "(d) SAFF (Ours; input)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            568,
+            316,
+            715,
+            508
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/8810bee4ce7e1f93f0f11ff2df60cc4844cb666696cdb3af51a551bec96b38da.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            720,
+            383,
+            866,
+            508
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Data: DyCheck Dataset (Masked) For single-camera casual monocular data, we select a subset of Gao et al.'s DyCheck dataset [10]. We remove scenes without an obvious subject, e.g., where telling foreground from background is hard even for a human. We select haru-sit, space-out, spin, paper-windmill, and mochi-high-five. We uniformly sample 48 frames and manually annotate object masks. We take even frames as the input set and odd frames as the hold out set. We use the same hyperparameters between both datasets.",
+        "bbox": [
+            75,
+            587,
+            468,
+            723
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Metrics To assess clustering performance, we use the Adjusted Rand Index (ARI;  $[-1, 1]$ ). This compares the similarity of two assignments without label matching, where random assignment would score  $\\approx 0$ . For foreground segmentation, we compute IoU (Jaccard), and for RGB quality we use PSNR, SSIM, and LPIPS.",
+        "bbox": [
+            75,
+            731,
+            468,
+            821
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Comparisons including ablations",
+        "text_level": 1,
+        "bbox": [
+            76,
+            832,
+            366,
+            848
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "SAFF (ours) We optimize upon the input split of each scene, and perform clustering to obtain object segmentations. To produce a foreground, we merge all salient objects.",
+        "bbox": [
+            75,
+            854,
+            468,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "SAFF (3D) The same as above, but processed in 3D rather than on 2D volume projections.",
+        "bbox": [
+            498,
+            587,
+            890,
+            618
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "-  $\\mathbf{w} / \\mathbf{pyr}\\lambda_{\\hat{\\mathbf{a}}} = \\{1,0,0\\}$  Pyramid with only coarse saliency (Sec. 3.3) and balanced semantic weight across levels.",
+            "—w/o pyr No pyramid (Sec. 3.3); we optimize with features and saliency extracted from the input image only.",
+            "— w/o merge With pyramid, but we remove cluster merging inside the saliency-aware clustering algorithm.",
+            "- w/blend  $v$  To compare generic dynamic segmentation to saliency segmentation, we use the static/dynamic weight instead of volume saliency to segment foreground objects. We set every pixel below the  $80\\%$ $v$  quantile in each image to be background, or otherwise foreground.",
+            "— w/post process We add a step after the saliency-aware clustering to refine edges using a conditional random field (please see supplemental material for details). This gains significantly from the depth estimated via volume reconstruction, producing sharp and detailed edges."
+        ],
+        "bbox": [
+            496,
+            625,
+            893,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "21802",
+        "bbox": [
+            478,
+            924,
+            519,
+            936
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/36a63135f22edbeb9eb8b71e22f73a3ba4d8b599837e8409cfe50dd8f4c40fc9.jpg",
+        "image_caption": [
+            "(a) Input",
+            "Figure 6: Saliency improves foreground segmentation. Static/dynamic separations are not foreground segmentations, leading to limited use of dynamic NeRF models for downstream tasks. Minor improvements to dynamic blending weight  $v$  are seen in some sequences (Jumping) by adding the saliency head to the shared backbone."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            90,
+            259,
+            222
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2cc18da6cca93a8b5c3df31bac7ab4722ea3c3d49eb0f8b46f3dc1a1f05dfb45.jpg",
+        "image_caption": [
+            "(b) NSFF [18] blend  $v$  (c)  $\\mathrm{D}^2\\mathrm{NeRF}$  [40] blend  $v$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            261,
+            90,
+            411,
+            222
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "c)  $\\mathrm{D}^2\\mathrm{NeRF}$  [40] blend  $v$",
+        "bbox": [
+            416,
+            75,
+            573,
+            89
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/21cc1dd0431ac7f807f9d4cb3470d324905107b22fdef2f00b6fcc51388a164d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            416,
+            90,
+            566,
+            151
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/9e957d9085e676a91bfd7c9c567126cb875382a76823f9a1b0396fca531b832e.jpg",
+        "image_caption": [
+            "(d) SAFF blend  $v$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            573,
+            90,
+            720,
+            151
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2fa19228ccb41766b7bf0778ab8bac6cf30332c73aef241eaefc80e232489399.jpg",
+        "image_caption": [
+            "(e) SAFF foregrounds"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            723,
+            90,
+            870,
+            151
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "NSFF [18] This method cannot produce semantic clusterings. While saliency and blend weight  $v$  have different meanings, if we compare our  $v$  to NSFF's, then we can see any impact of a shared backbone with attention heads upon the static/dynamic separation. We disable the supervised motion mask initialization in NSFF as our method does not use such information.",
+        "bbox": [
+            75,
+            286,
+            470,
+            377
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "$\\mathbf{D}^{2}\\mathbf{N}\\mathbf{e}\\mathbf{R}\\mathbf{F}$  [40] This method also cannot produce semantic clusterings. Over HyperNeRF [26], it adds a shadow field network and further losses to try to isolate objects into the dynamic NeRF over the separate static one. The paper also compares to NSFF without motion mask initialization.",
+        "bbox": [
+            75,
+            381,
+            470,
+            458
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "DINO-ViT (2D) [1] We ignore the volume and pass 2D semantic and attention features into the clustering algorithm. This cannot apply to novel viewpoints. Instead, we evaluate the approach upon all multi-view color images—input and hold-out—whereas other methods must render hold-out views. With our added pyramid processing (Sec. 3.3).",
+        "bbox": [
+            75,
+            465,
+            470,
+            556
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "- w/o pyr No pyramid; upsample to input RGB size.",
+        "bbox": [
+            76,
+            563,
+            429,
+            579
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "ProposeReduce (2D) [19] As a general comparison point, we apply this state-of-the-art 2D video segmentation method. For object segmentation, we use a ProposeReduce network that was pretrained with supervision on YouTube-VIS 2019 [43] for instance segmentation. For foreground segmentation, we use weights pretrained on UVOS [5] intended specifically for unsupervised foreground segmentation. As ProposeReduce is only a 2D method, we provide it with hold-out images for splits with novel views rather than our approach that must render novel views at hold-out poses.",
+        "bbox": [
+            75,
+            585,
+            470,
+            737
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2. Findings",
+        "text_level": 1,
+        "bbox": [
+            76,
+            746,
+            181,
+            762
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Dynamic scene decomposition We separate the background and each foreground object individually (Tab. 1). The baseline 2D DINO-ViT method is improved by our pyramid approach. But, being only 2D, this fails to produce a consistent decomposition across novel spacetime views even when given ground truth hold-out images. This shows the value of a volume integration. Next, supervised ProposeReduce produces good results (Fig. 5), but sometimes misses salient objects or fails to join them behind occluding objects, and only",
+        "bbox": [
+            75,
+            763,
+            470,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/350ded1779f0cc54162f52260ead563617104562086209c2929a161f8df6657e.jpg",
+        "table_caption": [
+            "Table 1: Spacetime volume integration improves dynamic scene decomposition. Pyramid construction and cluster merging help quantitatively, and ours is comparable to SOTA supervised 2D video segmentation network ProposeReduce.",
+            "Metric: Adjusted Rand Index  $[-1, 1]$ , higher is better."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Input</td><td>Fix Cam 0</td><td>Fix Time 0</td></tr><tr><td>ProposeReduce (2D)</td><td>0.725</td><td>0.736</td><td>0.742</td></tr><tr><td>DINO-ViT (2D)</td><td>0.501</td><td>0.495</td><td>0.321</td></tr><tr><td>w/o pyr帽子,hat</td><td>0.470</td><td>0.464</td><td>0.346</td></tr><tr><td>SAFF (3D)</td><td>0.594</td><td>0.578</td><td>0.566</td></tr><tr><td>SAFF (ours)</td><td>0.653</td><td>0.634</td><td>0.625</td></tr><tr><td>w/ pyr λa = {1,0,0}</td><td>0.620</td><td>0.598</td><td>0.592</td></tr><tr><td>w/o pyr帽子,hat</td><td>0.545</td><td>0.532</td><td>0.521</td></tr><tr><td>w/o merge cluster</td><td>0.593</td><td>0.574</td><td>0.563</td></tr><tr><td>w/post process</td><td>0.759</td><td>0.733</td><td>0.735</td></tr><tr><td>w/oracle</td><td>0.834</td><td>0.806</td><td>0.800</td></tr><tr><td>w/oracle + post process</td><td>0.922</td><td>0.890</td><td>0.880</td></tr></table>",
+        "bbox": [
+            504,
+            359,
+            888,
+            545
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "sometimes produces better edges than our method without post-processing as it can oversmooth edges. ProposeReduce also receives ground truth images in hold-out sets.",
+        "bbox": [
+            496,
+            556,
+            892,
+            603
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Instead, our approach must render hold-out views via the volume reconstruction. This produces more consistent segmentations through spacetime manipulations—this is the added value of volume integration through learned saliency and attention heads. Ablated components show the value of our pyramid step, its coarse-saliency-only variant, and the cluster merge and image post processing steps. Qualitatively, we see good detail (Fig. 5); post-processing additionally improves edge quality and removes small unwanted regions.",
+        "bbox": [
+            496,
+            604,
+            893,
+            741
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Oracle saliency We also include an oracle experiment where saliency voting is replaced by cluster selection based on ground truth masks. This experiment tells us what part of the performance gap lies with saliency itself, and what remains due to volume integration and cluster boundary issues. With oracle clusters, our decomposition performance is 0.8 ARI (Tab. 1) even in hold-out views. This shows that our existing cluster boundaries are accurate, and that accurate saliency for object selection is the larger remaining problem.",
+        "bbox": [
+            496,
+            763,
+            893,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "21803",
+        "bbox": [
+            478,
+            924,
+            519,
+            936
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/cc037c42e5797e3aadb14db85e9ca407d4c278629fbb159978018f9ae1fcf99d.jpg",
+        "table_caption": [
+            "Table 2: Saliency improves foreground segmentation. Adding saliency also slightly aids how much static/dynamic blend weight  $v$  represents the foreground (cf. NSFF blend  $v$  to SAFF's). Here, ProposeReduce uses unsupervised training and data [5]. Metric: IoU/Jaccard ([0, 1], higher is better)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Input</td><td>Fix Cam 0</td><td>Fix Time 0</td></tr><tr><td>ProposeReduce (2D)</td><td>0.609</td><td>0.464</td><td>0.591</td></tr><tr><td>DINO-ViT (2D)</td><td>0.381</td><td>0.382</td><td>0.357</td></tr><tr><td>NSFF — blend v</td><td>0.322</td><td>0.309</td><td>0.268</td></tr><tr><td>D2NeRF— blend v</td><td>0.470</td><td>0.334</td><td>0.269</td></tr><tr><td>SAFF — saliency</td><td>0.609</td><td>0.589</td><td>0.572</td></tr><tr><td>— blend v</td><td>0.388</td><td>0.380</td><td>0.329</td></tr><tr><td>— post process v</td><td>0.720</td><td>0.694</td><td>0.679</td></tr></table>",
+        "bbox": [
+            94,
+            169,
+            449,
+            297
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/38078e65d1eed18ee574ddb69b341f59d29e0a4b8545ee092f7d422d3faf9930.jpg",
+        "table_caption": [
+            "Table 3: SAFF semantics generalize to DyCheck Dataset (Masked) Oracle saliency. ProposeReduce is given ground truth test frames, while SAFF must render them. SAFF is comparable to SOTA supervised 2D video segmentation network ProposeReduce. Metric: Adjusted Rand Index  $[-1, 1]$ , higher is better); IoU/Jaccard index  $[-1, 1]$ , higher is better)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td colspan=\"2\">ARI</td><td colspan=\"2\">IoU</td></tr><tr><td></td><td>Input</td><td>Test</td><td>Input</td><td>Test</td></tr><tr><td>ProposeReduce (2D)</td><td>0.761</td><td>0.762</td><td>0.761</td><td>0.762</td></tr><tr><td>DINO-ViT (2D)</td><td>0.801</td><td>0.800</td><td>0.797</td><td>0.797</td></tr><tr><td>SAFF (3D)</td><td>0.902</td><td>0.820</td><td>0.910</td><td>0.812</td></tr></table>",
+        "bbox": [
+            81,
+            407,
+            465,
+            494
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Foreground segmentation We simplify the problem and consider all objects as simply 'foreground' to compare to methods that do not produce per object masks. Here, the same trend continues (Tab. 2). We note more subtle improvements to static/dynamic blending weights when adding our additional feature heads to the backbone MLP, and overall show that adding top-down information helps produce more useful object masks. Qualitative results show whole objects in the foreground rather than broad regions of possible dynamics (NSFF) or broken objects  $(\\mathrm{D}^2\\mathrm{NeRF};$  Fig. 6).",
+        "bbox": [
+            75,
+            511,
+            468,
+            662
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "DyCheck evaluation DyCheck often has close-up objects and, even with our selected sequences, saliency struggles to find foreground objects. Thus, we use oracle saliency. DINOViT and ProposeReduce are given test views while SAFF must render them. Quantitative (Tab. 3) and qualitative (Fig. 7) experiments show similar trends as before: ProposeReduce is good but may miss objects and fine details; SAFF may produce finer details and is more geometrically consistent.",
+        "bbox": [
+            75,
+            671,
+            470,
+            792
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Discussion and Limitations",
+        "text_level": 1,
+        "bbox": [
+            76,
+            806,
+            326,
+            821
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "DINO-ViT features are not instance-aware (Fig. 8). This is in contrast to object-centric learning approaches that aim to identify individual objects. To represent these different approaches, we compare to a result from slot-attention based  $\\mathrm{SAVi}++$  [7]. This method trains on thousands of supervised",
+        "bbox": [
+            75,
+            824,
+            468,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/cdd7b3c6f9d1e1e1425caa90ae6f392fa17fe43167b54091676d627962928cfa.jpg",
+        "image_caption": [
+            "Input ProposeReduce [19] SAFF (Ours; input) SAFF (Ours; novel)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            99,
+            598,
+            195
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/273f11278c90df905bb04230318be2162cc45f2dd6ab17510c1db4b067fbc468.jpg",
+        "image_caption": [
+            "Figure 7: SAFF can apply to DyCheck Dataset (Masked). SAFF is comparable to supervised large-scale 2D video learning [19]. Some fine details are improved (bottom, fingers) and novel views maintain their geometry (bottom, gap between legs occluded by head at input timestep)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            196,
+            599,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e822e330344030109a8f27f3b32ab14a8a9a04c53f39ea6bf92763540d178a21.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            602,
+            99,
+            697,
+            195
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/f118d418e44feb85ea722ffb224829fa95d77cb850e703874a4d2dcaedeb8466.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            602,
+            196,
+            697,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/41e9bf5dedef7a9aa4433f49ac7d51e411c80fd64d4ff68c6fd4a6c02c4dc620.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            700,
+            99,
+            795,
+            195
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/d63c51f4f565fe03ce474325545808eb92b7f3beafc4605e703a8166445f1dad.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            700,
+            196,
+            795,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/662699dfb3e9528689d2f33db4a4c8da2c57f93c1e930c0ca44483f2db93a7b7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            799,
+            99,
+            895,
+            195
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/c894697ca2b290957c8e39c04bb2ec7be403cd2d62bcad781bbbe2c4acfe88a0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            799,
+            196,
+            895,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6729bdccffeb5d82405d77c40eecf02828d678f74eb3d84859773e7e5f4a06ba.jpg",
+        "image_caption": [
+            "(a) Input"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            393,
+            627,
+            463
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/988c7a0f5473de6d3d76cb7af8e50e1eb2966cae1592553828303c6e8169f2ae.jpg",
+        "image_caption": [
+            "(b)  $\\mathrm{SAV_i + + [7]}$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            632,
+            393,
+            756,
+            463
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e8844aa33ab7268e2befb4420a8bdff4f44183a0e86e2b9956937aa9eb001cd7.jpg",
+        "image_caption": [
+            "(c) SAFF"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            761,
+            393,
+            885,
+            462
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/cc9c34111c72b5cd2873e5abb927357374f13c87078284374297bf3b4b5a6620.jpg",
+        "image_caption": [
+            "(d) Input",
+            "Figure 8: Limitations. DINO-ViT is not instance-aware, causing unwanted grouping (c). Unwanted static objects may be salient; assuming salient objects are dynamic fixes this (f)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            481,
+            627,
+            532
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/d0b34e16db9794f6f8c60fa715c6dfee06a9337e01c9fc2539e48867a77d9b04.jpg",
+        "image_caption": [
+            "(e) SAFF salient"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            629,
+            481,
+            753,
+            532
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/bfdbd6632afcaeb6b3475a898124129ca7c741b11ecfe7f74ba05018b82b489f.jpg",
+        "image_caption": [
+            "(f) Salient+dynamic"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            758,
+            481,
+            883,
+            532
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "MOVi [11] sequences with per-object masks, whereas we use generic pre-trained features and gain better edges from volume integration. Combining these two approaches could give accurate instance-level scene objects.",
+        "bbox": [
+            498,
+            594,
+            890,
+            655
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "DINO-ViT saliency may attend to unwanted regions. In Figure 8d-e, the static pillars could be isolated using scene flow. But, often our desired subjects do not move (cf. people in Umbrella or Balloon NBoard). For tasks or data that can assume that salient objects are dynamic, we use SAFF's 4D scene reconstruction to reject static-but-salient objects by merging clusters via scene flow: First, we project  $\\mathbf{f}$  over each timestep into each input camera pose—this simulates optical flow with a static camera. Clusters are marked as salient per image if mean flow magnitude per cluster  $|\\bar{\\mathbf{p}}| > 0.07$  and mean attention  $\\bar{\\mathbf{a}}_c > 0.07$ . Finally, as before, a cluster is globally salient if  $70\\%$  of images agree (Fig. 8f).",
+        "bbox": [
+            496,
+            655,
+            892,
+            837
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements The CV community in New England for feedback, and funding from NSF CNS-2038897 and an Amazon Research Award. Eliot thanks a Randy F. Pausch '82 Undergraduate Summer Research Award at Brown CS.",
+        "bbox": [
+            498,
+            839,
+            893,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "21804",
+        "bbox": [
+            478,
+            924,
+            517,
+            936
+        ],
+        "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] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2021. 2, 4, 5, 7",
+            "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 4",
+            "[3] Kenneth Blomqvist, Lionel Ott, Jen Jen Chung, and Roland Y. Siegwart. Baking in the feature: Accelerating volumetric segmentation by rendering feature maps. ArXiv, abs/2209.12744, 2022. 2",
+            "[4] Christopher P. Burgess, Loic Matthews, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matthew M. Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. ArXiv, abs/1901.11390, 2019. 1, 2",
+            "[5] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv:1905.00737, 2019. 2, 7, 8",
+            "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 1, 2, 4, 5",
+            "[7] Gamaeldin F. Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C. Mozer, and Thomas Kipf.  $\\mathrm{SAVi}++$ : Towards end-to-end object-centric learning from real-world videos. In Advances in Neural Information Processing Systems, 2022. 1, 2, 8",
+            "[8] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. ArXiv, abs/2209.08776, 2022. 2",
+            "[9] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In International Conference on 3D Vision (3DV), 2022. 2",
+            "[10] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In NeurIPS, 2022. 1, 6",
+            "[11] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam H. Laradji, Hsueh-Ti Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3739-3751, 2022. 8",
+            "[12] Klaus Greff, Raphael Lopez Kaufman, Rishabh Kabra, Nicholas Watters, Christopher P. Burgess, Daniel Zoran, Loic"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Matthew, Matthew M. Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In ICML, 2019. 2",
+            "[13] Basile Van Hoorick, Purva Tendulka, Dídac Surís, Dennis Park, Simon Stent, and Carl Vondrick. Revealing occlusions with 4d neural fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3001-3011, 2022. 2",
+            "[14] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthey, Antonia Creswell, Matthew M. Botvinick, Alexander Lerchner, and Christopher P. Burgess. Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. ArXiv, abs/2106.03849, 2021. 1, 2",
+            "[15] Thomas Kipf, Gamaleldin F. Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional Object-Centric Learning from Video. In International Conference on Learning Representations (ICLR), 2022. 1, 2",
+            "[16] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Advances in Neural Information Processing Systems, volume 35, 2022. 1, 2, 4",
+            "[17] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J. Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12871-12881, June 2022. 2",
+            "[18] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 7",
+            "[19] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1719-1728, 2021. 1, 5, 6, 7, 8",
+            "[20] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. arXiv preprint arXiv:2006.15055, 2020.1,2,5",
+            "[21] Kirill Mazur, Edgar Sucar, and Andrew J. Davison. Feature-realistic neural fusion for real-time, open set scene understanding. ArXiv, abs/2210.03043, 2022. 2",
+            "[22] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, 2022. 2",
+            "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3",
+            "[24] Tom Monnier, Elliot Vincent, Jean Ponce, and Mathieu Aubry. Unsupervised layered image decomposition into object prototypes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8640-8650, October 2021. 2"
+        ],
+        "bbox": [
+            503,
+            92,
+            893,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "21805",
+        "bbox": [
+            478,
+            924,
+            517,
+            936
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[25] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2",
+            "[26] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021. 1, 2, 7",
+            "[27] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2",
+            "[28] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2",
+            "[29] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. ArXiv, abs/2209.14860, 2022. 2",
+            "[30] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In CVPRW, 2022. 2",
+            "[31] Cameron Smith, Hong-Xing Yu, Sergey Zakharov, Frédo Durand, Joshua B. Tenenbaum, Jiajun Wu, and Vincent Sitzmann. Unsupervised discovery and composition of object light fields. ArXiv, abs/2205.03923, 2022. 1, 2",
+            "[32] Karl Stelzner, Kristian Kersting, and Adam R. Kosiorek. Decomposing 3d scenes into objects via unsupervised volume segmentation. ArXiv, abs/2104.01148, 2021. 1, 2, 5",
+            "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2",
+            "[34] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 2",
+            "[35] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12939-12950, 2021. 1, 2",
+            "[36] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural Feature Fusion Fields: 3D distillation of self-supervised 2D image representations. In Proceedings of the International Conference on 3D Vision (3DV), 2022. 2, 4",
+            "[37] Vadim Tschernezki, Diane Larlus, and Andrea Vedaldi. NeuralDiff: Segmenting 3D objects that move in egocentric videos. In Proceedings of the International Conference on 3D Vision (3DV), 2021. 2"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[38] Rishi Veerapaneni, John D. Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity abstraction in visual model-based reinforcement learning. In Leslie Pack Kaelbling, Danica Kragic, and Komei Sugiura, editors, Proceedings of the Conference on Robot Learning, volume 100 of Proceedings of Machine Learning Research, pages 1439-1456. PMLR, 30 Oct-01 Nov 2020. 1",
+            "[39] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L. Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Conference on Computer Vision and Pattern Recognition, 2022. 2",
+            "[40] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Öz Tireli. D2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. ArXiv, abs/2205.15838, 2022. 1, 2, 5, 7",
+            "[41] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 2",
+            "[42] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editable scene rendering. In International Conference on Computer Vision (ICCV), October 2021. 1, 2",
+            "[43] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019. 7",
+            "[44] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 2",
+            "[45] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 5",
+            "[46] Hong-Xing Yu, Leonidas J. Guibas, and Jiajun Wu. Unsupervised discovery of object radiance fields. In International Conference on Learning Representations, 2022. 1, 2, 5",
+            "[47] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            726
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "21806",
+        "bbox": [
+            478,
+            924,
+            519,
+            936
+        ],
+        "page_idx": 9
+    }
+]

2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/68ea1ef8-9cb5-44da-894d-92e02632af66_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f077e0f38009353dabe84ae3054bdcfb093a393d4411d5040ebf97c7968ddb5
+size 7001137

2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/full.md

@@ -0,0 +1,468 @@
+# Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition
+
+Yiqing Liang
+
+Eliot Laidlaw
+
+Alexander Meyerowitz Brown University
+
+Srinath Sridhar
+
+James Tompkin
+
+# Abstract
+
+From video, we reconstruct a neural volume that captures time-varying color, density, scene flow, semantics, and attention information. The semantics and attention let us identify salient foreground objects separately from the background across spacetime. To mitigate low resolution semantic and attention features, we compute pyramids that trade detail with whole-image context. After optimization, we perform a saliency-aware clustering to decompose the scene. To evaluate real-world scenes, we annotate object masks in the NVIDIA Dynamic Scene and DyCheck datasets. We demonstrate that this method can decompose dynamic scenes in an unsupervised way with competitive performance to a supervised method, and that it improves foreground/background segmentation over recent static/dynamic split methods. Project webpage: https://visual.cs.brown.edu/saff
+
+# 1. Introduction
+
+Given a casually-captured monocular RGB video of a dynamic scene, decomposing it into foreground objects and background is an important task in computer vision, with downstream applications in segmentation and video editing. An ideal method would also reconstruct the geometry and appearance over time including frame correspondences.
+
+Previous methods have made great progress but there are many limitations. Some works assume that there is no object motion in the scene [32, 46, 31, 16, 4], take input from multi-view cameras [16, 46], or do not explicitly reconstruct the underlying 3D structure of the scene [4, 7]. For objects, some works rely on masks or user input to aid segmentation [42, 15, 7], or use task-specific training datasets [7]. Sometimes, works assume the number of foreground objects [4, 20]. Given the challenges, many works train and test on synthetic data [14, 38].
+
+We present Semantic Attention Flow Fields (SAFF): A method to overcome these limitations by integrating low-level reconstruction cues with high-level pretrained information—both bottom-up and top-down—into a neural volume. With this, we demonstrate that embedded semantic and saliency (attention) information is useful for unsupervised dynamic scene decomposition. SAFF builds upon neural scene flow fields [18], an approach that reconstructs appearance, geometry, and motion. This uses frame interpolation rather than explicit canonicalization [35] or a latent hyperspace [26],
+
+![](images/01fefb4a9e45aeceefff10089d5bae4b9533e86d532a1f53df7e2cc944385a37.jpg)  
+Density  $\sigma$ , Saliency  $a$ , Semantics  $s$ , Color  $c$
+
+![](images/17f14ffcf60fce11f834a50b881e8ff330d777688d5b719bbc157a87104f5d21.jpg)  
+Novel view renderings with effects (depth of field, color adjustment)  
+Figure 1: We decompose a dynamic 3D scene from monocular input into time-varying color, density, scene flow, semantics, and attention information. This could be used to provide precise editing of objects or effects focused around objects.
+
+![](images/41eff40600d65b7aa1eeea11fea2b17c3386a06b748cc4ca9b57cb24afc10184.jpg)
+
+which lets it more easily apply to casual videos. For optimization, we supervise two network heads with pretrained DINO-ViT [6] semantic features and attention. Naively supervising high-resolution SAFF with low-resolution DINO-ViT output reduces reconstruction quality. To mitigate the mismatch, we build a semantic attention pyramid that trades detail with whole-image context. Having optimized a SAFF representation for a dynamic scene, we perform a saliency-aware clustering both in 3D and on rendered feature images to describe objects and their background. Given the volume reconstruction, the clustering generalizes to novel spacetime views.
+
+To evaluate SAFF's dynamic scene decomposition capacity, we expand the NVIDIA Dynamic Scene [45] and DyCheck [10] datasets by manually annotating object masks across input and hold-out views. We demonstrate that SAFF outperforms 2D DINO-ViT baselines and is comparable to a state-of-the-art video segmentation method ProposeReduce[19] on our data. Existing monocular video dynamic volume reconstruction methods typically separate static and dynamic parts, but these often do not represent meaningful foreground. We show improved foreground segmentation over NSFF and the current  $\mathrm{D}^2\mathrm{NeRF}$  [40] method for downstream tasks like editing.
+
+# 2. Related Work
+
+Decomposing a scene into regions of interest is a long-studied task in computer vision [5], including class, instance, and panoptic segmentation in high-level or top-down vision, and use of low-level or bottom-up cues like motion. Recent progress has considered images [12, 4, 20], videos [14, 15], layer decomposition [44, 24], and in-the-wild databases using deep generative models [25]. One example,  $\mathrm{SAVi}++[7]$ , uses slot attention [20] to define 2D objects in real-world videos. Providing first-frame segmentation masks achieves stable performance, with validation on the driving Waymo Open Dataset [33]. Our work attempts 3D scene decomposition for a casual monocular video without initial masks.
+
+Scene Decomposition with NeRFs Neural Radiance Fields (NeRF) [41] have spurred new scene decomposition research through volumes. ObSuRF [32] and uORF [46] are unsupervised slot attention works that bind a latent code to each object. Unsupervised decomposition is also possible on light fields [31]. For dynamic scenes, works like NeuralDiff [37] and  $\mathrm{D}^2\mathrm{NeRF}$  [40] focus on foreground separation, where foreground is defined to contain moving objects. Other works like N3F [36] and occlusions-4d [13] also decompose foregrounds into individual objects. N3F requires user input to specify which object to segment, and occlusions-4d takes RGB point clouds as input. Our work attempts to recover a segmented dynamic scene from a monocular RGB video without added markup or masks.
+
+Neural Fields Beyond Radiance Research has begun to add non-color information to neural volumes to aid decomposition via additional feature heads on the MLP. iLabel [47] adds a semantic head to propagate user-provided segmentations in the volume. PNF [17] and Panoptic-NeRF [9] attempt panoptic segmentation within neural fields, and Object-NeRF integrates instance segmentation masks into the field during optimization [42]. Research also investigates how to apply generic pretrained features to neural fields, like DINO-ViT.
+
+DINO-ViT for Semantics and Saliency DINO-ViT is a self-supervised transformer that, after pretraining, extracts generic semantic information [6]. Amir et al. [1] use DINO-ViT features with  $k$ -means clustering to achieve co-segmentation across a video. Seitzer et al. [29] combine slot attention and DINO-ViT features for object-centric learning on real-world 2D data. TokenCut [39] performs normalized cuts on DINO-ViT features for foreground segmentation on natural images. Deep Spectral Segmentation [22] show that graph Laplacian processing of DINO-ViT features provides unsupervised foreground segmentation, and Selfmask [30] shows that these features can provide object saliency masks. Our approach considers these clustering and saliency findings for the setting of 3D decomposition from monocular video.
+
+DINO-ViT Fields Concurrent works have integrated DINO-ViT features into neural fields. DFF [16] distills features for dense multi-view static scenes with user input for segmentation. N3F [36] expands NeuralDiff to dynamic scenes, and relies on user input for segmentation. AutoLabel [3] uses DINO-ViT features to accelerate segmentation in static scenes given a ground truth segmentation mask. Other works use DINO-ViT differently. FeatureRealistic-Fusion [21] uses DINO-ViT in an online feature fusion task, focusing on propagating user input, and NeRF-SOS [8] uses 2D DINO-ViT to process NeRF-rendered multi-view RGB images of a static scene. In contrast to these works, we consider real-world casual monocular videos, recover and decompose a 3D scene, then explore whether saliency can avoid the need for masks or user input in segmenting objects.
+
+# 3. Method
+
+For a baseline dynamic scene reconstruction method, we begin with NSFF from Li et al. [18] (Sec. 3.1), which builds upon NeRF [23]. NSFF's low-level scene flow frame-to-frame approach provides better reconstructions for real-world casual monocular videos than deformation-based methods [35, 26]. We modify the architecture to integrate higher-level semantic and saliency (or attention) features (Sec. 3.2). After optimizing a SAFF for each scene, we perform saliency-aware clustering of the field (Sec. 3.4). All implementation details are in our supplemental material.
+
+Input Our method takes in a single RGB video over time  $i$  as an ordered set of images  $I \in \mathcal{I}$  and camera poses. We use COLMAP to recover camera poses [28]. From all poses, we define an NDC-like space that bounds the scene, and a set of rays  $\mathbf{r} \in \mathcal{R}$ , one per image pixel with color  $\hat{\mathbf{c}}^\dagger$ . Here,  $\hat{\mathbf{c}}$  denotes a 2D pixel value in contrast to a 3D field value, and  $\hat{\mathbf{c}}^\dagger$  denotes an input value in contrast to an estimated value.
+
+From pretrained networks, we estimate single-frame monocular depth  $\hat{d}^{\dagger}$  (MiDaSv2 [27]), optical flow  $\hat{\mathbf{p}}_i^\dagger$  (RAFT [34]), and semantic features  $\hat{\mathbf{s}}^\dagger$  and attention  $\hat{\mathbf{a}}^\dagger$  (DINO-ViT [6]) after important preprocessing (Sec. 3.3).
+
+# 3.1. Initial Dynamic Neural Volume Representation
+
+The initial representation comprises a static NeRF  $F_{\theta}^{\mathrm{st}}$  and a dynamic NeRF  $F_{\theta}^{\mathrm{dy}}$ . The static network predicts a color  $\mathbf{c}$ , density  $\sigma$ , and blending weight  $v$  (Eq. (1)), and the dynamic network predicts time-varying color  $\mathbf{c}_i$ , density  $\sigma_i$ , scene flow  $\mathbf{f}_i$ , and occlusion weights  $w_i$  (Eq. (2)). In both network architectures, other than position  $\mathbf{x}$ , we add direction  $\boldsymbol{\omega}$  to a late separate head to only condition the estimation of color  $\mathbf{c}$ .
+
+$$
+F _ {\boldsymbol {\theta}} ^ {\mathrm {s t}}: (\mathbf {x}, \omega) \rightarrow \left(\mathbf {c} ^ {\mathrm {s t}}, \sigma^ {\mathrm {s t}}, v\right) \tag {1}
+$$
+
+$$
+F _ {\boldsymbol {\theta}} ^ {\mathrm {d y}}: (\mathbf {x}, \boldsymbol {\omega}, i) \rightarrow \left(\mathbf {c} _ {i} ^ {\mathrm {d y}}, \sigma_ {i} ^ {\mathrm {d y}}, \mathbf {f} _ {i}, w _ {i}\right) \tag {2}
+$$
+
+To produce a pixel's color, we sample points at distances  $t$  along the ray  $\mathbf{x}_t = \mathbf{x} - \omega t$  between near and far planes  $t_n$
+
+![](images/f11cce526596a99c53cb277dfee5c4ba6fe1f8e2e53d703a83c76fedd620f2ad.jpg)  
+(a) SAFF video inputs
+
+![](images/08f48acf6d9bc6c16bccd6de7bb8553c7cb55375f61b01b26ae38212d33b08ce.jpg)  
+(b) SAFF model
+
+![](images/8f5ce03d1d46c9edbba3fab841f598fd38e818f838815b8436f0ed4d1f04ab7e.jpg)  
+(c) Volume rendered outputs  
+Figure 2: Overview. From monocular video, SAFF builds a neural field of scene-flowed 3D density, radiance, semantics, and attention (b). This is guided by semantic attention pyramids that increase resolution, plus depth and optical flow priors (a). We can render new spacetime views of any channel, and use saliency-aware clustering to decompose objects and background (c).
+
+to  $t_f$ , query each network, then integrate transmittance  $T$ , density  $\sigma$ , and color along the ray from these samples [23]. For brevity, we omit evaluating at  $\mathbf{x}_t, \boldsymbol{\omega}$ , e.g.,  $\sigma^{\mathrm{st}}(\mathbf{x}_t)$  is simply  $\sigma^{\mathrm{st}}$ ;  $\mathbf{c}^{\mathrm{st}}(\mathbf{x}_t, \boldsymbol{\omega})$  is simply  $\mathbf{c}^{\mathrm{st}}$ .
+
+We produce a combined color from the static and dynamic colors by multiplication with their densities:
+
+$$
+\sigma_ {i} \mathbf {c} _ {i} = v \sigma^ {\mathrm {s t}} \mathbf {c} ^ {\mathrm {s t}} + (1 - v) \sigma_ {i} ^ {\mathrm {d y}} \mathbf {c} _ {i} ^ {\mathrm {d y}} \tag {3}
+$$
+
+Given that transmittance integrates density up to the current sampled point under Beer-Lambert volume attenuation, the rendered pixel color for the ray is computed as:
+
+$$
+\hat {\mathbf {c}} _ {i} = \int_ {t _ {n}} ^ {t _ {f}} T _ {i} \sigma_ {i} \mathbf {c} _ {i} d t \text {w h e r e} T _ {i} = \exp \left(- \int_ {t _ {n}} ^ {t} \sigma_ {i} d t\right) \tag {4}
+$$
+
+To optimize the volume to reconstruct input images, we compute a photometric loss  $L_{\hat{\mathbf{c}}}$  between rendered and input colors:
+
+$$
+\mathcal {L} _ {\hat {\mathbf {c}}} = \frac {1}{| \mathcal {R} |} \sum_ {\mathbf {r} _ {i} \in \mathcal {R}} \left\| \hat {\mathbf {c}} _ {i} \left(\mathbf {r} _ {i}\right) - \hat {\mathbf {c}} _ {i} ^ {\dagger} \left(\mathbf {r} _ {i}\right) \right\| _ {2} ^ {2} \tag {5}
+$$
+
+Scene Flow Defining correspondence over time is important for monocular input as it lets us penalize a reprojection error with neighboring frames  $j \in \mathcal{N}$ , e.g., where  $j = i + 1$  or  $j = i - 1$ . We denote  $i \rightarrow j$  for the projection of frame  $i$  onto frame  $j$  by scene flow.  $F_{\theta}^{\mathrm{dy}}$  estimates both forwards and backwards scene flow at every point to penalize a bi-direction loss.
+
+Thus, we approximate color output at time step  $j$  by flowing the queried network values at time step  $i$ :
+
+$$
+\hat {\mathbf {c}} _ {i \rightarrow j} = \int_ {t _ {n}} ^ {t _ {f}} T _ {i \rightarrow j} \sigma_ {i \rightarrow j} \mathbf {c} _ {i \rightarrow j} d t \tag {6}
+$$
+
+Reprojection must account for occlusion and disocclusion by motion. As such,  $F_{\theta}^{\mathrm{dy}}$  also predicts forwards and backwards
+
+scene occlusion weights  $w_{i + 1}$  and  $w_{i - 1} \in [0,1]$ , where a point with  $w_{i + 1} = 0$  means that occlusion status is changed one step forwards in time. We can integrate  $w$  to a pixel:
+
+$$
+\hat {w} _ {i \rightarrow j} = \int_ {t _ {n}} ^ {t _ {f}} T _ {i \rightarrow j} \sigma_ {i \rightarrow j} w _ {j} d t \tag {7}
+$$
+
+Then, this pixel weight modulates the color loss such that occluded pixels are ignored. (8)
+
+$$
+\mathcal {L} _ {\hat {\mathbf {c}} _ {i \rightarrow j}} = \frac {1}{| \mathcal {R} | | \mathcal {N} |} \sum_ {\mathbf {r} _ {i} \in \mathcal {R}} \sum_ {j \in \mathcal {N}} \hat {w} _ {i \rightarrow j} (\mathbf {r} _ {i}) | | \hat {\mathbf {c}} _ {i \rightarrow j} (\mathbf {r} _ {i}) - \hat {\mathbf {c}} _ {j} ^ {\dagger} (\mathbf {r} _ {j}) | | _ {2} ^ {2}
+$$
+
+Prior losses We use the pretrained depth and optical flow map losses to help overcome the ill-posed monocular reconstruction problem. These losses decay as optimization progresses to rely more and more on the optimized self-consistent geometry and scene flow. For geometry, we estimate a depth  $\hat{d}_i$  for each ray  $\mathbf{r}_i$  by replacing  $\mathbf{c}_i$  in Eq. (4) by the distance  $t$  along the ray. Transform  $z$  estimates a scale and shift as the pretrained network produces only relative depth.
+
+$$
+\mathcal {L} _ {\hat {d}} = \frac {1}{| \mathcal {R} |} \sum_ {\mathbf {r} _ {i} \in \mathcal {R}} \left| \left| \hat {d} _ {i} - z \left(\hat {d} _ {i} ^ {\dagger}\right) \right| \right| _ {1} \tag {9}
+$$
+
+For motion, projecting scene flow to a camera lets us compare to the estimated optical flow. Each sample point along a ray  $\mathbf{x}_i$  is advected to a point in the neighboring frame  $\mathbf{x}_{i\rightarrow j}$ , then integrated to the neighboring camera plane to produce a 2D point offset  $\hat{\mathbf{p}}_i(\mathbf{r}_i)$ . Then, we expect the difference in the start and end positions to match the prior:
+
+$$
+\mathcal {L} _ {\hat {\mathbf {p}}} = \frac {1}{| \mathcal {R} | | \mathcal {N} |} \sum_ {\mathbf {r} _ {i} \in \mathcal {R}} \sum_ {j \in \mathcal {N} (i)} | | \hat {\mathbf {p}} _ {i} (\mathbf {r} _ {i}) - \hat {\mathbf {p}} _ {i} ^ {\dagger} (\mathbf {r} _ {i}) | | _ {1} \tag {10}
+$$
+
+Additional regularizations encourage occlusion weights to be close to one, scene flow to be small, locally constant, and cyclically consistent, and blending weight  $v$  to be sparse.
+
+# 3.2. Semantic Attention Flow Fields
+
+Beyond low-level or bottom-up features, high-level or top-down features are also useful to define objects and help down-stream tasks like segmentation. For example, methods like NSFF or  $\mathrm{D}^2\mathrm{NeRF}$  struggle to provide useful separation of static and dynamic parts because blend weight  $v$  estimates whether the volume appears to be occupied by some moving entity. This is not the same as objectness; tasks like video editing could benefit from accurate dynamic object masks.
+
+As such, we extract 2D semantic features and attention (or saliency) values from a pretrained DINO-ViT network, then optimize the SAFF such that unknown 3D semantic and attention features over time can be projected to recreate their 2D complements. This helps us to ascribe semantic meaning to the volume and to identify objects. As semantics/attention are integrated into the 4D volume, we can render them from novel spacetime views without further DINO-ViT computation.
+
+To estimate semantic features  $\mathbf{s}$  and attention  $\mathbf{a}$  at 3D points in the volume at time  $i$ , we add two new heads to both the static  $F_{\theta}^{\mathrm{st}}$  and the dynamic  $F_{\theta}^{\mathrm{dy}}$  networks:
+
+$$
+F _ {\boldsymbol {\theta}} ^ {\mathrm {s t}}: (\mathbf {x}, \omega) \quad \rightarrow \left(\mathbf {c} ^ {\mathrm {s t}}, \sigma^ {\mathrm {s t}}, v, \mathbf {s} ^ {\mathrm {s t}}, \mathbf {a} ^ {\mathrm {s t}}\right) \tag {11}
+$$
+
+$$
+F _ {\boldsymbol {\theta}} ^ {\mathrm {d y}}: (\mathbf {x}, \boldsymbol {\omega}, i) \rightarrow \left(\mathbf {c} _ {i} ^ {\mathrm {d y}}, \sigma_ {i} ^ {\mathrm {d y}}, \mathbf {f} _ {i}, w _ {i}, \mathbf {s} _ {i} ^ {\mathrm {d y}}, \mathbf {a} _ {i} ^ {\mathrm {d y}}\right) \tag {12}
+$$
+
+As semantic features have been demonstrated to be somewhat robust to view dependence [1], in our architectures both heads for s, a are taken off the backbone before  $\omega$  is injected.
+
+To render semantics from the volume, we replace the color term  $\mathbf{c}$  in Eq. (4) with  $\mathbf{s}$  and equivalently for  $\mathbf{a}$ :
+
+$$
+\sigma_ {i} \mathbf {s} _ {i} = v \sigma^ {\mathrm {s t}} \mathbf {s} ^ {\mathrm {s t}} + (1 - v) \sigma_ {i} ^ {\mathrm {d y}} \mathbf {s} _ {i} ^ {\mathrm {d y}}, \hat {\mathbf {s}} _ {i} = \int_ {t _ {n}} ^ {t _ {f}} T _ {i} \sigma_ {i} \mathbf {s} _ {i} d t \tag {13}
+$$
+
+To encourage scene flow to respect semantics over time, we penalize complementary losses on s and a (showing s only):
+
+$$
+\mathcal {L} _ {\hat {\mathbf {s}} _ {i \rightarrow j}} = \frac {1}{| \mathcal {R} | | \mathcal {N} |} \sum_ {\mathbf {r} _ {i} \in \mathcal {R}} \sum_ {j \in \mathcal {N}} \hat {w} _ {i \rightarrow j} (\mathbf {r} _ {i}) \| \hat {\mathbf {s}} _ {i \rightarrow j} (\mathbf {r} _ {i}) - \hat {\mathbf {s}} _ {j} ^ {\dagger} (\mathbf {r} _ {j}) \| _ {2} ^ {2} \tag {14}
+$$
+
+Finally, as supervision, we add respective losses on the reconstruction of the 2D semantic and attention features from projected 3D volume points (showing s only):
+
+$$
+\mathcal {L} _ {\hat {\mathbf {s}}} = \frac {1}{| \mathcal {R} |} \sum_ {\mathbf {r} _ {i} \in \mathcal {R}} \left\| \hat {\mathbf {s}} _ {i} \left(\mathbf {r} _ {i}\right) - \hat {\mathbf {s}} _ {i} ^ {\dagger} \left(\mathbf {r} _ {i}\right) \right\| _ {2} ^ {2} \tag {15}
+$$
+
+Unlike depth and scene flow priors, these are not priors—there is no self-consistency for semantics to constrain their values. Thus, we do not decay their contribution. While decaying avoids disagreements between semantic and attention features and color-enforced scene geometry, it also leads to a loss of useful meaning (please see supplemental).
+
+Thus our final loss becomes:
+
+$$
+\begin{array}{l} \mathcal {L} _ {\mathrm {S A F F}} = \mathcal {L} _ {\hat {\mathbf {c}}} + \lambda_ {\hat {\mathbf {c}} _ {i \rightarrow j}} \mathcal {L} _ {\hat {\mathbf {c}} _ {i \rightarrow j}} + \lambda_ {\hat {d}} \mathcal {L} _ {\hat {d}} + \lambda_ {\hat {\mathbf {p}}} \mathcal {L} _ {\hat {\mathbf {p}}} \tag {16} \\ + \lambda_ {\hat {\mathbf {s}} _ {i \rightarrow j}} \mathcal {L} _ {\hat {\mathbf {s}} _ {i \rightarrow j}} + \lambda_ {\hat {\mathbf {a}} _ {i \rightarrow j}} \mathcal {L} _ {\hat {\mathbf {a}} _ {i \rightarrow j}} + \lambda_ {\hat {\mathbf {s}}} \mathcal {L} _ {\hat {\mathbf {s}}} + \lambda_ {\hat {\mathbf {a}}} \mathcal {L} _ {\hat {\mathbf {a}}} \\ \end{array}
+$$
+
+# 3.3. Semantic Attention Pyramids
+
+When thinking about scenes, we might argue that semantics from an ideal extractor should be scale invariant, as distant objects have the same class as close objects. We might also argue that saliency (or attention features) may not be scale invariant, as small details in a scene should only be salient when viewed close up. In practice, both extracted features vary across scale and have limited resolution, e.g., DINO-ViT [6] produces one output for each  $8 \times 8$  patch. But, from this, we want semantic features and saliency for every RGB pixel that still respects scene boundaries.
+
+Thus far, work on static scenes has ignored the input/feature resolution mismatch [16] as multi-view constraints provide improved localization within the volume. For monocular video, this approach has limitations [36]. Forming many constraints on dynamic objects requires long-term motion correspondence—a tricky task—and so we want to maximize the resolution of any input features where possible without changing their meaning.
+
+One way may be through a pyramid of semantic and attention features that uses a sliding window approach at finer resolutions. Averaging features could increase detail around edges, but we must overcome the practical limit that these features are not stable across scales. This is especially important for saliency: unlike typical RGB pyramids that must preserve energy in an alias-free way [2], saliency changes significantly over scales and does not preserve energy.
+
+Consider a feature pyramid  $\mathcal{P}$  with loss weights per level:
+
+$$
+\mathcal {L} _ {\mathcal {P} \hat {\mathbf {s}}} = \sum_ {i \in \mathcal {P}} \lambda_ {\hat {\mathbf {s}}} ^ {i} \mathcal {L} _ {\hat {\mathbf {s}}} ^ {i} \quad \mathcal {L} _ {\mathcal {P} \hat {\mathbf {a}}} = \sum_ {i \in \mathcal {P}} \lambda_ {\hat {\mathbf {a}}} ^ {i} \mathcal {L} _ {\hat {\mathbf {a}}} ^ {i} \tag {17}
+$$
+
+Naively encouraging scale-consistent semantics and whole-image saliency, e.g.,  $\lambda_{\hat{\mathbf{s}}} = \{1/3, 1/3, 1/3\}$  with  $\lambda_{\hat{\mathbf{a}}} = \{1, 0, 0\}$ , empirically leads to poor recovered object edges because the balanced semantics and coarse saliency compete over where the underlying geometry is. Instead, we weight both equally  $\lambda_{\hat{\mathbf{s}}} = \lambda_{\hat{\mathbf{a}}} = \{1/9, 4/9, 4/9\}$ . Even though the coarse layer has smaller weight, it is sufficient to guide the overall result. This balances high resolution edges from fine layers and whole object features from coarse layers while reducing geometry conflicts, and leads to improved features (Fig. 3).
+
+Of course, any sliding window must contain an object to extract reliable features for that object. At coarse levels, an object is always in view. At fine levels, an object is only captured in some windows. Objects of interest tend to be near the middle of the frame, meaning that boundary windows at finer pyramid levels contain features that less reliably capture those objects. This can cause spurious connections in clustering. To cope with this, we relatively decrease finer level boundary window weights: We upsample all levels to the finest level, then increase the coarsest level weight towards the frame boundary to  $\lambda_{\hat{\mathbf{s}}} = \lambda_{\hat{\mathbf{a}}} = \{1/3, 1/3, 1/3\}$ .
+
+![](images/f178d6f08ab968f5f60529102575cf05b82ce1c9023a855566a217b585eb38d3.jpg)  
+(a) Input
+
+![](images/9b0a274bb513cdbaff27d2e781ba72fa72821d8b34d8ed9cb154a712427de7bd.jpg)  
+(b) Semantics
+
+![](images/ca4d6cc4b08415640f4ab2ce014bf87ff955f0bdf83c5692ffb80dd75522697d.jpg)  
+(c) Semantics (volume)
+
+![](images/286f833094b4816d0f26ea379f6c302a9a40006555badeca5a9734de5a57cb54.jpg)  
+(d) Saliency
+
+![](images/6073298e54a3d420c359a7e00e8a065e7a90a15f4532e6dd95c82ee33325699a.jpg)  
+(e) Saliency (volume)
+
+![](images/9a772ca97f6d34a23e449345bc01ca9b625edcc4b2d9d0c0e4c5afca1890d8ad.jpg)  
+(f) Pyr. semantics
+
+![](images/b1d357ef4844ecadfd229b10f4610f2f43a0e7ebec2c3b41c59afbd833772873.jpg)  
+(g) Our pyr. semantics (volume)  
+Figure 3: Semantics and saliency improve by both volume integration and by our pyramid. On Balloon NBoard, blocky artifacts are removed. Semantics are visualized as most significant three PCA dimensions; specific colors are less meaningful.
+
+![](images/db2932d48e70bef0de273f03533759cd8762fef47e784615bda46b636c792bcb.jpg)  
+(h) Pyr. saliency
+
+![](images/d9cad011c3b648a0848faa5ed6182fbf8a504b14c318a601d52790f011237b06.jpg)  
+(i) Our pyr. saliency (volume)
+
+# 3.4. Using SAFF for Saliency-aware Clustering
+
+We now wish to isolate salient objects. Even in dynamic scenes, relevant objects may not move, so analyzing dynamic elements is insufficient (cf. [40]). One approach predicts segmentation end-to-end [19]. However, end-to-end learning requires priors about the scene provided by supervision, and even large-scale pretraining might fail given unseen test scenes. To achieve scene-specific decompositions from per-video semantics, inspired by Amir et al. [1], we design clustering for spacetime volumes that allows segmenting novel spatio-temporal views. While DINO-ViT is trained on images, its features are loosely temporally consistent [6].
+
+Some works optimize a representation with a fixed number of clusters, e.g., via slot attention [20] in NeRFs [32, 46]. Instead, we cluster using elbow  $k$ -means, letting us adaptively find the number of clusters after optimization. This is more flexible than baking an anticipated number of slots (sometimes with fixed semantics), and lets us cluster and segment in novel spatio-temporal viewpoints.
+
+We demonstrate clustering results in both 3D over time and on rendered volume 2D projections over time. Given the volume reconstruction, we might think that clustering directly in 3D would be better. But, monocular input with narrow baselines makes it challenging to precisely reconstruct geometries: Consider that depth integrated a long a ray can still be accurate even though geometry at specific 3D points may be inaccurate or 'fluffy'. As such, we use the 2D volume projection clustering results in 2D comparisons.
+
+Method For 3D over time, we sample points from the SAFF uniformly along input rays  $(128\times H\times W)$  and treat each pixel as a feature point, e.g., semantics are 64 dim. and saliency is 1 dim. For volume projection, we render to  $N$  input poses  $(N\times H\times W)$ , and treat each pixel as a feature point instead. In either case, we cluster all feature points together using elbow  $k$ -means to produce an initial set of separate regions. For each cluster  $c$ , for each image, we calculate the mean attention of all feature points within the cluster  $\bar{\mathbf{a}}_c$ . If  $\bar{\mathbf{a}}_c > 0.07$ , then this
+
+cluster is salient for this image. Finally, all feature points vote on saliency: if more than  $70\%$  agree, the cluster is salient.
+
+Salient objects may still be split into semantic parts: e.g., in Fig. 4, the person's head/body are separated. Plus, unwanted background saliency may exist, e.g., input  $\hat{\mathbf{a}}^\dagger$  is high for the teal graphic on the wall. As such, before saliency voting, we merge clusters whose centroids have a cosine similarity  $>0.5$ . This reduces the first problem as heads and bodies are similar, and reduces the second problem as merging the graphic cluster into the background reduces its average saliency (Fig. 4).
+
+To extract an object from the 3D volume, we sample 3D points along each input ray, then ascribe the label from the semantically-closest centroid. All clusters not similar to the stored salient clusters are marked as background with zero density. For novel space-time views, we render feature images from the volume, then assign cluster labels to each pixel according to its similarity with stored input view centroids.
+
+# 4. Experiments
+
+We show the impact of adding semantic and saliency features through scene decomposition and foreground experiments. Our website contains supplemental videos.
+
+Data: NVIDIA Dynamic Scene Dataset (Masked) This data [45] has 8 sequences of 24 time steps formed from 12 cameras simultaneously capturing video. We manually annotate object masks for view and time step splits; we will release this data publicly. We define three data splits per sequence:
+
+1. Input: A monocular camera that moves position for every timestep is simulated from the input sequences; we use Yoon et al.'s input sequences [45].  
+2. Fix Cam 0 (hold out): We fix the camera at position 0 as time plays, requiring novel view and time synthesis.  $\{(\mathrm{cam}_0,\mathrm{time}_i),i\in [1,2,\dots 23]\}$  
+3. Fix Time 0 (hold out): We fix time at step 0 as the camera moves, requiring novel view and time synthesis.  $\{(\mathrm{cam}_i,\mathrm{time}_0),i\in [1,2,\ldots 11]\}$
+
+![](images/0916a3dac95539f8c91aff80d1f8a07ab4e1ba67931550ee0b2cae13ffc89eda.jpg)  
+(i) Input frame  
+(ii) Input semantic/attention
+
+![](images/c5ce24e922c13f4821e3240ca0f0021f73e48b716f94b30e793395bc9e3a343b.jpg)  
+(a) Our pyr. semantic (volume)  
+(d) Our pyr. saliency (volume)
+
+![](images/5d567c20160ac212f558b1a90c90f18dce7c52124b47c9276e2fdd6d4ca11381.jpg)
+
+![](images/f462e2415f3c4072ff3dd65adef4b2717184e0446ea7a0119253bc18cc96ab9d.jpg)  
+(e) Our merged clustering  
+Figure 4: Saliency-aware clustering improves decomposition. On Dynamic Face, the head and body are semantically and saliently different, but are mutually different from the background. This allows us to extract a time-varying 3D field of the object.
+
+![](images/83fe39f2429a8082a84ae5b94ad37683b9859938142cb82fa7d83523de749201.jpg)  
+(c) Salient clusters (pre-merge)  
+(f) Our final foreground  
+(iii) Extraction (fixed time)
+
+![](images/5cc523c6ad484004e2d5657e5123e5dbf9683ce2a9478a19116031c1ea14a0d0.jpg)
+
+![](images/5cbf3a9ab54ff469420060cb15eb9ca62dcd439ff5e73fb19525cc2d36915a10.jpg)  
+(iv) Extraction (fixed view)
+
+![](images/d2014859e5f8596c4f049fd7a5542a994321f44e6e401f6f045a53168d008f72.jpg)
+
+![](images/e20bdacede3a7a5868b7cef4fcdcf593b23fa2655dfd89336a1d4e0a92985e35.jpg)
+
+![](images/09c1891a830a23fbc620a22ac5e9c58856b78f5efaea2cf8f1da9005971890f1.jpg)
+
+![](images/7def0160a5af8dfbc2da02f54623763411d8065e9bb925784ad0a2b1dd18007b.jpg)  
+(a) Input  
+Figure 5: SAFF object segmentations show balanced quality while recovering a volumetric scene representation (e). Basic DINO-ViT produces low-quality segmentations and misses objects. A state-of-the-art 2D video learning method [19] sometimes has edge detail (Umbrella, legs) but othertimes misses detail and objects (Balloon NBoard).
+
+![](images/4b13ac5c9c77ea60444552e5524c3e891235297addb0747db1403a807f583b6e.jpg)  
+(b) DINO-ViT (2D)
+
+![](images/6e89c0f7d9b9ee8d30bfbc5bec0ae618274ab0497d01755188af532b00ba8da6.jpg)  
+(c) ProposeReduce [19]
+
+![](images/a9ed21c21fb0dcc4c27f0d7b78c66d1cf4381b857645386272644700ab7fdba2.jpg)  
+(d) SAFF (Ours; input)
+
+![](images/8810bee4ce7e1f93f0f11ff2df60cc4844cb666696cdb3af51a551bec96b38da.jpg)
+
+Data: DyCheck Dataset (Masked) For single-camera casual monocular data, we select a subset of Gao et al.'s DyCheck dataset [10]. We remove scenes without an obvious subject, e.g., where telling foreground from background is hard even for a human. We select haru-sit, space-out, spin, paper-windmill, and mochi-high-five. We uniformly sample 48 frames and manually annotate object masks. We take even frames as the input set and odd frames as the hold out set. We use the same hyperparameters between both datasets.
+
+Metrics To assess clustering performance, we use the Adjusted Rand Index (ARI;  $[-1, 1]$ ). This compares the similarity of two assignments without label matching, where random assignment would score  $\approx 0$ . For foreground segmentation, we compute IoU (Jaccard), and for RGB quality we use PSNR, SSIM, and LPIPS.
+
+# 4.1. Comparisons including ablations
+
+SAFF (ours) We optimize upon the input split of each scene, and perform clustering to obtain object segmentations. To produce a foreground, we merge all salient objects.
+
+SAFF (3D) The same as above, but processed in 3D rather than on 2D volume projections.
+
+-  $\mathbf{w} / \mathbf{pyr}\lambda_{\hat{\mathbf{a}}} = \{1,0,0\}$  Pyramid with only coarse saliency (Sec. 3.3) and balanced semantic weight across levels.  
+—w/o pyr No pyramid (Sec. 3.3); we optimize with features and saliency extracted from the input image only.  
+— w/o merge With pyramid, but we remove cluster merging inside the saliency-aware clustering algorithm.  
+- w/blend  $v$  To compare generic dynamic segmentation to saliency segmentation, we use the static/dynamic weight instead of volume saliency to segment foreground objects. We set every pixel below the  $80\%$ $v$  quantile in each image to be background, or otherwise foreground.  
+— w/post process We add a step after the saliency-aware clustering to refine edges using a conditional random field (please see supplemental material for details). This gains significantly from the depth estimated via volume reconstruction, producing sharp and detailed edges.
+
+![](images/36a63135f22edbeb9eb8b71e22f73a3ba4d8b599837e8409cfe50dd8f4c40fc9.jpg)  
+(a) Input  
+Figure 6: Saliency improves foreground segmentation. Static/dynamic separations are not foreground segmentations, leading to limited use of dynamic NeRF models for downstream tasks. Minor improvements to dynamic blending weight  $v$  are seen in some sequences (Jumping) by adding the saliency head to the shared backbone.
+
+![](images/2cc18da6cca93a8b5c3df31bac7ab4722ea3c3d49eb0f8b46f3dc1a1f05dfb45.jpg)  
+(b) NSFF [18] blend  $v$  (c)  $\mathrm{D}^2\mathrm{NeRF}$  [40] blend  $v$
+
+c)  $\mathrm{D}^2\mathrm{NeRF}$  [40] blend  $v$
+
+![](images/21cc1dd0431ac7f807f9d4cb3470d324905107b22fdef2f00b6fcc51388a164d.jpg)
+
+![](images/9e957d9085e676a91bfd7c9c567126cb875382a76823f9a1b0396fca531b832e.jpg)  
+(d) SAFF blend  $v$
+
+![](images/2fa19228ccb41766b7bf0778ab8bac6cf30332c73aef241eaefc80e232489399.jpg)  
+(e) SAFF foregrounds
+
+NSFF [18] This method cannot produce semantic clusterings. While saliency and blend weight  $v$  have different meanings, if we compare our  $v$  to NSFF's, then we can see any impact of a shared backbone with attention heads upon the static/dynamic separation. We disable the supervised motion mask initialization in NSFF as our method does not use such information.
+
+$\mathbf{D}^{2}\mathbf{N}\mathbf{e}\mathbf{R}\mathbf{F}$  [40] This method also cannot produce semantic clusterings. Over HyperNeRF [26], it adds a shadow field network and further losses to try to isolate objects into the dynamic NeRF over the separate static one. The paper also compares to NSFF without motion mask initialization.
+
+DINO-ViT (2D) [1] We ignore the volume and pass 2D semantic and attention features into the clustering algorithm. This cannot apply to novel viewpoints. Instead, we evaluate the approach upon all multi-view color images—input and hold-out—whereas other methods must render hold-out views. With our added pyramid processing (Sec. 3.3).
+
+- w/o pyr No pyramid; upsample to input RGB size.
+
+ProposeReduce (2D) [19] As a general comparison point, we apply this state-of-the-art 2D video segmentation method. For object segmentation, we use a ProposeReduce network that was pretrained with supervision on YouTube-VIS 2019 [43] for instance segmentation. For foreground segmentation, we use weights pretrained on UVOS [5] intended specifically for unsupervised foreground segmentation. As ProposeReduce is only a 2D method, we provide it with hold-out images for splits with novel views rather than our approach that must render novel views at hold-out poses.
+
+# 4.2. Findings
+
+Dynamic scene decomposition We separate the background and each foreground object individually (Tab. 1). The baseline 2D DINO-ViT method is improved by our pyramid approach. But, being only 2D, this fails to produce a consistent decomposition across novel spacetime views even when given ground truth hold-out images. This shows the value of a volume integration. Next, supervised ProposeReduce produces good results (Fig. 5), but sometimes misses salient objects or fails to join them behind occluding objects, and only
+
+Table 1: Spacetime volume integration improves dynamic scene decomposition. Pyramid construction and cluster merging help quantitatively, and ours is comparable to SOTA supervised 2D video segmentation network ProposeReduce.  
+Metric: Adjusted Rand Index  $[-1, 1]$ , higher is better.  
+
+<table><tr><td></td><td>Input</td><td>Fix Cam 0</td><td>Fix Time 0</td></tr><tr><td>ProposeReduce (2D)</td><td>0.725</td><td>0.736</td><td>0.742</td></tr><tr><td>DINO-ViT (2D)</td><td>0.501</td><td>0.495</td><td>0.321</td></tr><tr><td>w/o pyr帽子,hat</td><td>0.470</td><td>0.464</td><td>0.346</td></tr><tr><td>SAFF (3D)</td><td>0.594</td><td>0.578</td><td>0.566</td></tr><tr><td>SAFF (ours)</td><td>0.653</td><td>0.634</td><td>0.625</td></tr><tr><td>w/ pyr λa = {1,0,0}</td><td>0.620</td><td>0.598</td><td>0.592</td></tr><tr><td>w/o pyr帽子,hat</td><td>0.545</td><td>0.532</td><td>0.521</td></tr><tr><td>w/o merge cluster</td><td>0.593</td><td>0.574</td><td>0.563</td></tr><tr><td>w/post process</td><td>0.759</td><td>0.733</td><td>0.735</td></tr><tr><td>w/oracle</td><td>0.834</td><td>0.806</td><td>0.800</td></tr><tr><td>w/oracle + post process</td><td>0.922</td><td>0.890</td><td>0.880</td></tr></table>
+
+sometimes produces better edges than our method without post-processing as it can oversmooth edges. ProposeReduce also receives ground truth images in hold-out sets.
+
+Instead, our approach must render hold-out views via the volume reconstruction. This produces more consistent segmentations through spacetime manipulations—this is the added value of volume integration through learned saliency and attention heads. Ablated components show the value of our pyramid step, its coarse-saliency-only variant, and the cluster merge and image post processing steps. Qualitatively, we see good detail (Fig. 5); post-processing additionally improves edge quality and removes small unwanted regions.
+
+Oracle saliency We also include an oracle experiment where saliency voting is replaced by cluster selection based on ground truth masks. This experiment tells us what part of the performance gap lies with saliency itself, and what remains due to volume integration and cluster boundary issues. With oracle clusters, our decomposition performance is 0.8 ARI (Tab. 1) even in hold-out views. This shows that our existing cluster boundaries are accurate, and that accurate saliency for object selection is the larger remaining problem.
+
+Table 2: Saliency improves foreground segmentation. Adding saliency also slightly aids how much static/dynamic blend weight  $v$  represents the foreground (cf. NSFF blend  $v$  to SAFF's). Here, ProposeReduce uses unsupervised training and data [5]. Metric: IoU/Jaccard ([0, 1], higher is better).  
+
+<table><tr><td></td><td>Input</td><td>Fix Cam 0</td><td>Fix Time 0</td></tr><tr><td>ProposeReduce (2D)</td><td>0.609</td><td>0.464</td><td>0.591</td></tr><tr><td>DINO-ViT (2D)</td><td>0.381</td><td>0.382</td><td>0.357</td></tr><tr><td>NSFF — blend v</td><td>0.322</td><td>0.309</td><td>0.268</td></tr><tr><td>D2NeRF— blend v</td><td>0.470</td><td>0.334</td><td>0.269</td></tr><tr><td>SAFF — saliency</td><td>0.609</td><td>0.589</td><td>0.572</td></tr><tr><td>— blend v</td><td>0.388</td><td>0.380</td><td>0.329</td></tr><tr><td>— post process v</td><td>0.720</td><td>0.694</td><td>0.679</td></tr></table>
+
+Table 3: SAFF semantics generalize to DyCheck Dataset (Masked) Oracle saliency. ProposeReduce is given ground truth test frames, while SAFF must render them. SAFF is comparable to SOTA supervised 2D video segmentation network ProposeReduce. Metric: Adjusted Rand Index  $[-1, 1]$ , higher is better); IoU/Jaccard index  $[-1, 1]$ , higher is better).  
+
+<table><tr><td></td><td colspan="2">ARI</td><td colspan="2">IoU</td></tr><tr><td></td><td>Input</td><td>Test</td><td>Input</td><td>Test</td></tr><tr><td>ProposeReduce (2D)</td><td>0.761</td><td>0.762</td><td>0.761</td><td>0.762</td></tr><tr><td>DINO-ViT (2D)</td><td>0.801</td><td>0.800</td><td>0.797</td><td>0.797</td></tr><tr><td>SAFF (3D)</td><td>0.902</td><td>0.820</td><td>0.910</td><td>0.812</td></tr></table>
+
+Foreground segmentation We simplify the problem and consider all objects as simply 'foreground' to compare to methods that do not produce per object masks. Here, the same trend continues (Tab. 2). We note more subtle improvements to static/dynamic blending weights when adding our additional feature heads to the backbone MLP, and overall show that adding top-down information helps produce more useful object masks. Qualitative results show whole objects in the foreground rather than broad regions of possible dynamics (NSFF) or broken objects  $(\mathrm{D}^2\mathrm{NeRF};$  Fig. 6).
+
+DyCheck evaluation DyCheck often has close-up objects and, even with our selected sequences, saliency struggles to find foreground objects. Thus, we use oracle saliency. DINOViT and ProposeReduce are given test views while SAFF must render them. Quantitative (Tab. 3) and qualitative (Fig. 7) experiments show similar trends as before: ProposeReduce is good but may miss objects and fine details; SAFF may produce finer details and is more geometrically consistent.
+
+# 5. Discussion and Limitations
+
+DINO-ViT features are not instance-aware (Fig. 8). This is in contrast to object-centric learning approaches that aim to identify individual objects. To represent these different approaches, we compare to a result from slot-attention based  $\mathrm{SAVi}++$  [7]. This method trains on thousands of supervised
+
+![](images/cdd7b3c6f9d1e1e1425caa90ae6f392fa17fe43167b54091676d627962928cfa.jpg)  
+Input ProposeReduce [19] SAFF (Ours; input) SAFF (Ours; novel)
+
+![](images/273f11278c90df905bb04230318be2162cc45f2dd6ab17510c1db4b067fbc468.jpg)  
+Figure 7: SAFF can apply to DyCheck Dataset (Masked). SAFF is comparable to supervised large-scale 2D video learning [19]. Some fine details are improved (bottom, fingers) and novel views maintain their geometry (bottom, gap between legs occluded by head at input timestep).
+
+![](images/e822e330344030109a8f27f3b32ab14a8a9a04c53f39ea6bf92763540d178a21.jpg)
+
+![](images/f118d418e44feb85ea722ffb224829fa95d77cb850e703874a4d2dcaedeb8466.jpg)
+
+![](images/41e9bf5dedef7a9aa4433f49ac7d51e411c80fd64d4ff68c6fd4a6c02c4dc620.jpg)
+
+![](images/d63c51f4f565fe03ce474325545808eb92b7f3beafc4605e703a8166445f1dad.jpg)
+
+![](images/662699dfb3e9528689d2f33db4a4c8da2c57f93c1e930c0ca44483f2db93a7b7.jpg)
+
+![](images/c894697ca2b290957c8e39c04bb2ec7be403cd2d62bcad781bbbe2c4acfe88a0.jpg)
+
+![](images/6729bdccffeb5d82405d77c40eecf02828d678f74eb3d84859773e7e5f4a06ba.jpg)  
+(a) Input
+
+![](images/988c7a0f5473de6d3d76cb7af8e50e1eb2966cae1592553828303c6e8169f2ae.jpg)  
+(b)  $\mathrm{SAV_i + + [7]}$
+
+![](images/e8844aa33ab7268e2befb4420a8bdff4f44183a0e86e2b9956937aa9eb001cd7.jpg)  
+(c) SAFF
+
+![](images/cc9c34111c72b5cd2873e5abb927357374f13c87078284374297bf3b4b5a6620.jpg)  
+(d) Input  
+Figure 8: Limitations. DINO-ViT is not instance-aware, causing unwanted grouping (c). Unwanted static objects may be salient; assuming salient objects are dynamic fixes this (f).
+
+![](images/d0b34e16db9794f6f8c60fa715c6dfee06a9337e01c9fc2539e48867a77d9b04.jpg)  
+(e) SAFF salient
+
+![](images/bfdbd6632afcaeb6b3475a898124129ca7c741b11ecfe7f74ba05018b82b489f.jpg)  
+(f) Salient+dynamic
+
+MOVi [11] sequences with per-object masks, whereas we use generic pre-trained features and gain better edges from volume integration. Combining these two approaches could give accurate instance-level scene objects.
+
+DINO-ViT saliency may attend to unwanted regions. In Figure 8d-e, the static pillars could be isolated using scene flow. But, often our desired subjects do not move (cf. people in Umbrella or Balloon NBoard). For tasks or data that can assume that salient objects are dynamic, we use SAFF's 4D scene reconstruction to reject static-but-salient objects by merging clusters via scene flow: First, we project  $\mathbf{f}$  over each timestep into each input camera pose—this simulates optical flow with a static camera. Clusters are marked as salient per image if mean flow magnitude per cluster  $|\bar{\mathbf{p}}| > 0.07$  and mean attention  $\bar{\mathbf{a}}_c > 0.07$ . Finally, as before, a cluster is globally salient if  $70\%$  of images agree (Fig. 8f).
+
+Acknowledgements The CV community in New England for feedback, and funding from NSF CNS-2038897 and an Amazon Research Award. Eliot thanks a Randy F. Pausch '82 Undergraduate Summer Research Award at Brown CS.
+
+# References
+
+[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2021. 2, 4, 5, 7  
+[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 4  
+[3] Kenneth Blomqvist, Lionel Ott, Jen Jen Chung, and Roland Y. Siegwart. Baking in the feature: Accelerating volumetric segmentation by rendering feature maps. ArXiv, abs/2209.12744, 2022. 2  
+[4] Christopher P. Burgess, Loic Matthews, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matthew M. Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. ArXiv, abs/1901.11390, 2019. 1, 2  
+[5] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv:1905.00737, 2019. 2, 7, 8  
+[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 1, 2, 4, 5  
+[7] Gamaeldin F. Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C. Mozer, and Thomas Kipf.  $\mathrm{SAVi}++$ : Towards end-to-end object-centric learning from real-world videos. In Advances in Neural Information Processing Systems, 2022. 1, 2, 8  
+[8] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. ArXiv, abs/2209.08776, 2022. 2  
+[9] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In International Conference on 3D Vision (3DV), 2022. 2  
+[10] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In NeurIPS, 2022. 1, 6  
+[11] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam H. Laradji, Hsueh-Ti Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3739-3751, 2022. 8  
+[12] Klaus Greff, Raphael Lopez Kaufman, Rishabh Kabra, Nicholas Watters, Christopher P. Burgess, Daniel Zoran, Loic
+
+Matthew, Matthew M. Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In ICML, 2019. 2  
+[13] Basile Van Hoorick, Purva Tendulka, Dídac Surís, Dennis Park, Simon Stent, and Carl Vondrick. Revealing occlusions with 4d neural fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3001-3011, 2022. 2  
+[14] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthey, Antonia Creswell, Matthew M. Botvinick, Alexander Lerchner, and Christopher P. Burgess. Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. ArXiv, abs/2106.03849, 2021. 1, 2  
+[15] Thomas Kipf, Gamaleldin F. Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional Object-Centric Learning from Video. In International Conference on Learning Representations (ICLR), 2022. 1, 2  
+[16] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Advances in Neural Information Processing Systems, volume 35, 2022. 1, 2, 4  
+[17] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J. Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12871-12881, June 2022. 2  
+[18] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 7  
+[19] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1719-1728, 2021. 1, 5, 6, 7, 8  
+[20] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. arXiv preprint arXiv:2006.15055, 2020.1,2,5  
+[21] Kirill Mazur, Edgar Sucar, and Andrew J. Davison. Feature-realistic neural fusion for real-time, open set scene understanding. ArXiv, abs/2210.03043, 2022. 2  
+[22] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, 2022. 2  
+[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3  
+[24] Tom Monnier, Elliot Vincent, Jean Ponce, and Mathieu Aubry. Unsupervised layered image decomposition into object prototypes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8640-8650, October 2021. 2
+
+[25] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2  
+[26] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021. 1, 2, 7  
+[27] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2  
+[28] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2  
+[29] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. ArXiv, abs/2209.14860, 2022. 2  
+[30] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In CVPRW, 2022. 2  
+[31] Cameron Smith, Hong-Xing Yu, Sergey Zakharov, Frédo Durand, Joshua B. Tenenbaum, Jiajun Wu, and Vincent Sitzmann. Unsupervised discovery and composition of object light fields. ArXiv, abs/2205.03923, 2022. 1, 2  
+[32] Karl Stelzner, Kristian Kersting, and Adam R. Kosiorek. Decomposing 3d scenes into objects via unsupervised volume segmentation. ArXiv, abs/2104.01148, 2021. 1, 2, 5  
+[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2  
+[34] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 2  
+[35] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12939-12950, 2021. 1, 2  
+[36] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural Feature Fusion Fields: 3D distillation of self-supervised 2D image representations. In Proceedings of the International Conference on 3D Vision (3DV), 2022. 2, 4  
+[37] Vadim Tschernezki, Diane Larlus, and Andrea Vedaldi. NeuralDiff: Segmenting 3D objects that move in egocentric videos. In Proceedings of the International Conference on 3D Vision (3DV), 2021. 2
+
+[38] Rishi Veerapaneni, John D. Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity abstraction in visual model-based reinforcement learning. In Leslie Pack Kaelbling, Danica Kragic, and Komei Sugiura, editors, Proceedings of the Conference on Robot Learning, volume 100 of Proceedings of Machine Learning Research, pages 1439-1456. PMLR, 30 Oct-01 Nov 2020. 1  
+[39] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L. Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Conference on Computer Vision and Pattern Recognition, 2022. 2  
+[40] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Öz Tireli. D2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. ArXiv, abs/2205.15838, 2022. 1, 2, 5, 7  
+[41] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 2  
+[42] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editable scene rendering. In International Conference on Computer Vision (ICCV), October 2021. 1, 2  
+[43] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019. 7  
+[44] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 2  
+[45] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 5  
+[46] Hong-Xing Yu, Leonidas J. Guibas, and Jiajun Wu. Unsupervised discovery of object radiance fields. In International Conference on Learning Representations, 2022. 1, 2, 5  
+[47] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2

2023/Semantic Attention Flow Fields for Monocular Dynamic Scene Decomposition/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fc7d860933a12e1a8ce3b81aa3c2f6be10b5151457d0988f0cbd17074d573cd
+size 672479

2023/Semantic Information in Contrastive Learning/c3144656-62d8-4975-97d6-ef4630f410df_content_list.json

@@ -0,0 +1,1699 @@
+[
+    {
+        "type": "text",
+        "text": "Semantic Information in Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            248,
+            130,
+            722,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Shengjiang Quan $^{1*}$ , Masahiro Hirano $^{2}$  and Yuji Yamakawa $^{3}$ \\\n $^{1}$ Graduate School of Engineering, The University of Tokyo, Japan\\\n $^{2}$ Institute of Industrial Science, The University of Tokyo, Japan\\\n $^{3}$ Interfaculty Initiative in Information Studies, The University of Tokyo, Japan\\\nquan@g.ecc.u-tokyo.ac.jp, {mhirano, y-ymkw}@iis.u-tokyo.ac.jp",
+        "bbox": [
+            176,
+            179,
+            792,
+            271
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            304,
+            313,
+            319
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "This work investigates the functionality of Semantic information in Contrastive Learning (SemCL). An advanced pretext task is designed: a contrast is performed between each object and its environment, taken from a scene. This allows the SemCL pretrained model to extract objects from their environment in an image, significantly improving the spatial understanding of the pretrained models. Downstream tasks of semantic/instance segmentation, object detection and depth estimation are implemented on PASCAL VOC, Cityscapes, COCO, KITTI, etc. SemCL pretrained models substantially outperform ImageNet pretrained counterparts and are competitive with well-known works on downstream tasks. The results suggest that a dedicated pretext task leveraging semantic information can be powerful in benchmarks related to spatial understanding. The code is available at https://github.com/sjiang95/semcl.",
+        "bbox": [
+            75,
+            335,
+            473,
+            579
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            607,
+            209,
+            625
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Within the field of visual representation learning, contrastive learning attracts special attention for its inspiring performance in transfer learning [57, 49, 31, 22, 8, 9]. The concept of contrastive learning can literally be explained as discovering the difference between positive and negative samples. The definition of positive-negative samples is one of the main subjects of contrastive learning, which directly determines the pretext task and the corresponding loss function.",
+        "bbox": [
+            75,
+            633,
+            470,
+            768
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We humans can recognize an object from its even complicated environment because we have learned and bound the typical geometric feature and the corresponding semantic concept of the object class. For example, we can recognize a cat in a scene thanks to our knowledge of its appearance (typical geometric feature) and the concept of cat (semantic information). Similarly, the target of prevalent",
+        "bbox": [
+            75,
+            770,
+            470,
+            877
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/bab89fd9041f4a80e562caeb78e23e0262a1f6e2656bcaff36d7a6fdd19d7503.jpg",
+        "image_caption": [
+            "Figure 1: Perceptual cognition of SemCL. Leveraging semantic information, SemCL separates an object from its environment. Each sample is divided into objects and surroundings and considered as a contrastive pair. The task is to maximize the distance between the object and its environment in the embedding space."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            306,
+            893,
+            452
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "visual tasks, such as semantic/instance segmentation, object detection, etc., can be summarized as distinguishing and localizing subjects from surroundings. Inspired by and to extend the spatial recognition mechanism to CV, we propose our method SemCL which directly teaches models to extract a subject from its environment. Utilizing publicly off-the-shelf datasets providing semantic labels, SemCL can be considered as a supervised contrastive learning approach whose samples consist of subjects and their corresponding surroundings (semantically not the subject). The pretext task of SemCL is to tell the difference between one subject and its surroundings. From a representation learning perspective, the pretext task is to maximize the distance  $(L^2$  norm) between a subject and its environment in the embedding space (see Figure 1). The pretext task mimics the spatial recognition mechanism to discriminate objects and surroundings, which significantly improves the spatial information understanding of pretrained models.",
+        "bbox": [
+            496,
+            583,
+            893,
+            853
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In SemCL, contrast is performed pairwise between positive-negative pairs from one image/scene. Inter-scene contrast is considered inappropriate: only the contrasts",
+        "bbox": [
+            498,
+            854,
+            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,
+            220,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "5686",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "among subject-surrounding pairs from one scene are considered valid. Therefore, paired InfoNCE loss is adopted. With the mechanism of contrasting only paired samples, SemCL is characterized by decoupling the number of negative samples from the batch size. Compared to MoCo [22], whose performance is positively correlated with batch size and thus increases hardware cost (e.g. batchsize  $= 4096$  on 128 GPUs for ViT-B in MoCo v3 [9]) for satisfactory results, SemCL can achieve consistent yet competitive results on downstream tasks with relatively small batch size (e.g., batchsize  $= 64 \\ @ 224 \\times 224$ ). We adopt the MoCo v3 [9] as the pretraining framework.",
+        "bbox": [
+            75,
+            90,
+            472,
+            272
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The primary motivation of representation learning is to pretrain general representations that can be fine-tuned and transferred to downstream tasks. In this work, the SemCL pretrained models are benchmarked on semantic segmentation, object detection and instance segmentation, and depth estimation tasks. SemCL models substantially outperform their ImageNet pretrained counterparts due to the improved ability of spacial information understanding. In a system-level comparison, SemCL models also show gains over previous well-known works (e.g. Deeplabv3+ [7] in semantic segmentation, Mask2Former [10] in instance segmentation and Binsformer [35] in depth estimation).",
+        "bbox": [
+            75,
+            273,
+            472,
+            455
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            470,
+            218,
+            486
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            76,
+            497,
+            277,
+            513
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The seminal idea of contrastive learning dates back at least to the 1990s [2, 3, 31], but it has become prominent in recent years thanks to the large pretrained models in the fields of NLP and CV [31]. As the name suggests, contrastive learning mines abstract representations by comparing between \"similar\" and \"dissimilar\" samples: minimizing the representation of \"similar\" samples and maximizing that of \"dissimilar\" samples in the embedding space. Formally, the \"similar\" inputs are called positive samples, and the \"dissimilar\" inputs are called negative samples.",
+        "bbox": [
+            75,
+            521,
+            472,
+            672
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Defining \"similar\" and \"dissimilar\" samples is an important topic in contrastive learning. In instance discrimination [57], which is a ubiquitous approach to define positive-negative samples in CV, each sample in a minibatch  $x \\in \\{x_1, \\ldots, x_n\\}$  is considered mutually exclusive to the rest. The encoded output  $\\theta(\\cdot)$  of each sample is pushed away from the others in the embedding space, denoted as  $\\theta(x_1) \\Leftrightarrow \\{\\theta(x_2), \\ldots, \\theta(x_n)\\} \\Leftrightarrow \\text{push}$ . Contrastive Predictive Coding (CPC) [49] defines positive-negative samples in a generative way: Given a sequential input  $\\{x_{t-3}, x_{t-2}, x_{t-1}, x_t\\}$ , a context latent representation  $C_t$  is computed to predict the encoded outputs of future inputs  $\\{\\theta(\\mathbb{X}_{t+1}), \\theta(\\mathbb{X}_{t+2}), \\theta(\\mathbb{X}_{t+3}), \\ldots\\}$ . Those predicted future embeddings are considered as positive samples to real future embeddings, and exclusive to some other",
+        "bbox": [
+            75,
+            674,
+            472,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "irrelevant inputs.",
+        "bbox": [
+            500,
+            92,
+            614,
+            106
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Momentum Contrast",
+        "text_level": 1,
+        "bbox": [
+            500,
+            114,
+            697,
+            128
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "He et al. [22] proposed a momentum contrast framework, MoCo, for contrastive representation learning. MoCo summed unsupervised visual representation learning [57, 49, 26, 69, 25, 48, 1] up as dictionary look-up: the input data are represented and sampled by an encoder network  $\\theta_{k}$  as \"keys\"  $k = \\theta_{k}(x^{k})$ . The goal of contrastive learning is to train encoders  $\\theta_{q}$  by which a \"query\"  $q = \\theta_{q}(x^{q})$  should be similar to its matching key and dissimilar to others. Since dictionary consistency is considered crucial for unsupervised learning with contrastive loss [22], MoCo introduces momentum updating. In the MoCo framework, the key coder  $\\theta_{k}$  is progressively updated by the weighted average of  $\\theta_{k}$  and the query coder  $\\theta_{q}$ .",
+        "bbox": [
+            496,
+            138,
+            893,
+            335
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\theta_ {k} \\leftarrow m \\theta_ {k} + (1 - m) \\theta_ {q}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            607,
+            344,
+            893,
+            362
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where the momentum coefficient  $m \\in [0,1)$  is set quite high ( $m = 0.999$  in [22]) for slower key encoder update. The MoCo series has proven the power of the Momentum Contrast framework on feature transferring and is therefore employed as the basis for this work.",
+        "bbox": [
+            496,
+            369,
+            893,
+            445
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3. Labeled Data In Unsupervised Framework",
+        "text_level": 1,
+        "bbox": [
+            500,
+            453,
+            864,
+            469
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Large networks are quite thirsty for training data, while labelling the large amount of collected data is a human resource consuming challenge. Apart from turning to unsupervised learning, which inherently avoids using labelled data, using unlabeled data in supervised learning is also a solution. Pseudo-Label [44, 38, 56] is a trick that uses (usually supervised) pretrained networks to generate features that can play the role of labels. And the generated pseudo-labels participate in the iterations of the target network just like annotations.",
+        "bbox": [
+            496,
+            476,
+            893,
+            626
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In contrast, Khosla et al. [28] also investigated the function of labelled data in unsupervised learning. As a contrastive learning research, Khosla et al. [28] intuitively uses the classification label to define samples within the same class are mutually positive, and those from other categories are negative. To exploit the label information, clusters of points belonging to the same class are pulled together in the embedding space, while clusters of samples from different classes are pushed apart [28]:  $\\left\\{\\theta\\left(x_1^{c0}\\right) \\rightleftharpoons \\theta\\left(x_2^{c0}\\right) \\rightleftharpoons \\theta\\left(x_{n}^{c0}\\right) \\rightleftharpoons \\theta\\left(x_n^0\\right)$ $\\Leftrightarrow$ $\\left\\{\\theta\\left(x_1^{c1}\\right) \\rightleftharpoons \\theta\\left(x_2^{c1}\\right) \\rightleftharpoons \\theta\\left(x_{m}^{c1}\\right)$  ( $\\rightleftharpoons$  for pull). The experimental result shows that supervised contrastive learning can improve both the accuracy and robustness of classifiers - [28] achieves a top-1 accuracy of  $81.4\\%$  on the ImageNet (IN) dataset, which is  $0.8\\%$  above the best result reported by ResNet-200 architecture [24]. Likewise, our work investigates the effectiveness of semantic information in unsupervised framework contrastive learning.",
+        "bbox": [
+            496,
+            627,
+            895,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "5687",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.4. Semantic Information in Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            76,
+            90,
+            470,
+            107
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Other than [28], there are works that use small amounts of labelled data in unsupervised representation learning. Wang et al. [54] proposed a pixel-wise contrastive algorithm for semantic segmentation: for a pixel  $i$ , it is pulled towards pixels belonging to the same category and pushed away from pixels in other categories in the embedding space. This contrast is performed across images from the same dataset. In the Cityscapes test benchmark, [54] R101 achieves 79.2 mIoU, which is 1.1 points better than Deeplabv3 [6].",
+        "bbox": [
+            75,
+            113,
+            468,
+            250
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The attempt and results of works using labels in unsupervised representation learning, such as [28, 54], indicate that label information can bring gains to a fully unsupervised representation. It is promising to explore advanced pretext tasks to improve the performance of pretrained models in downstream tasks.",
+        "bbox": [
+            75,
+            250,
+            470,
+            339
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            76,
+            352,
+            168,
+            368
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. SemCL Dataset",
+        "text_level": 1,
+        "bbox": [
+            76,
+            376,
+            233,
+            392
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The core of this work is to define semantically contrastive pairs by utilizing off-the-shelf semantic labels, which are mappings between categories of objects and their coordinates on the raw image at the pixel level. As decided by the authors of a dataset, semantic labels can be color or grayscale images. Each predefined color or grayscale denotes a category, and the distribution of label pixels reflects the location of pixels belonging to a particular category in the raw image.",
+        "bbox": [
+            75,
+            400,
+            468,
+            521
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The SemCL dataset is generated by the mechanism shown in Figure 2, where the raw image and its semantic label are selected from the PASCAL VOC2012 dataset [14]. From the semantic label annotating two objects - a motorcycle and a rider - two binary masks BM0 and BM1 can be extracted to separately represent two annotated objects. Their inverted counterparts  $\\neg$ BM0 and  $\\neg$ BM1 represent pixels other than the target object. These binary masks are then bitwise applied to the raw image to separate an object (Anchor) from its environment ( $\\neg$ Anchor), forming a semantically contrastive pair. Each element in a raw Anchor- $\\neg$ Anchor pair is considered to be semantically opposite to the other.",
+        "bbox": [
+            75,
+            521,
+            468,
+            703
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In practice, the SemCL-1M dataset is produced from training sets of augmented PASCAL VOC2012 [14] (also known as SBD [20]), Cityscapes [11], ADE20K [67, 68] and COCO [36] (stuff+thing 2017 [4]) datasets. The components of the SemCL dataset are illustrated in Table 1. When generating the dataset, a threshold  $t = 0.01$  is set to filter out too small (less than  $[\\sqrt{t} \\times \\text{height}] \\times [\\sqrt{t} \\times \\text{width}]$ ) objects that are considered semantically unsaturated.",
+        "bbox": [
+            75,
+            702,
+            470,
+            824
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Paired InfoNCE Loss",
+        "text_level": 1,
+        "bbox": [
+            76,
+            830,
+            277,
+            847
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Since contrastive learning can be considered as a dictionary look-up task [22], it is assumed that within a set of keys  $\\{k_0, k_1, k_2, \\ldots\\}$  of a dictionary, a query  $q$  matches only one",
+        "bbox": [
+            76,
+            854,
+            470,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/4f0e152aa9bdfd04cd415a97c9454b917d14019e511435d7455f721b3c307df8.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>VOC2012</td><td>Cityscapes</td><td>ADE20K</td><td>COCO</td><td>Total</td></tr><tr><td># training samples</td><td>10,582</td><td>2,975</td><td>25,574</td><td>118,287</td><td>157,418</td></tr><tr><td># SemCL pairs</td><td>14,203</td><td>20,280</td><td>270,218</td><td>712,212</td><td>1,016,913</td></tr></table>",
+        "bbox": [
+            501,
+            88,
+            893,
+            161
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Components of the SemCL dataset. The SemCL datasets spawned from VOC2012(aug), Cityscapes, ADE20k and COCO 2017 are named SemCL- $\\{\\mathrm{VOC}$ , City, ADE, COCO\\} respectively.",
+        "bbox": [
+            496,
+            170,
+            893,
+            232
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "key  $k_{+}$  of the dictionary [22]. The value of the objective function should be low if  $q$  is similar to its positive key  $k_{+}$  and dissimilar to all negative keys  $\\{k_{0}, k_{1}, k_{2}, \\ldots\\} \\setminus \\{k_{+}\\}$ . In MoCo [22], the form of a contrastive loss function InfoNCE [49] is adopted",
+        "bbox": [
+            496,
+            258,
+            893,
+            335
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {q} = - \\log \\frac {\\exp (q \\cdot k _ {+} / \\tau)}{\\sum_ {i = 0} ^ {K} \\exp (q \\cdot k _ {i} / \\tau)} \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            584,
+            345,
+            892,
+            383
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\tau$  is a scalar temperature hyperparameter controlling the strength of penalties on hard negative samples [50], e.g. in the instance discrimination task [49, 22], there may be potential positive samples in the randomly chosen negative samples. Summing over one positive and  $K$  negative samples, Equation (2) is the log loss of a  $(\\mathbf{K} + 1)$ -way softmax-based classifier trying to classify  $q$  as  $k_{+}$  [22]. While in our case an anchor instance  $q$  compares only with its own  $n$  augmentations (positive)  $K^{+} := \\{k_{1}^{+}, k_{2}^{+}, k_{3}^{+}, \\ldots, k_{n}^{+}\\}$  and paired  $n + 1$  negatives  $K^{-} := \\{k_{0}^{-}, k_{1}^{-}, k_{2}^{-}, k_{3}^{-}, \\ldots, k_{n}^{-}\\}$  using paired InfoNCE",
+        "bbox": [
+            496,
+            392,
+            893,
+            559
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {q} = - \\sum_ {j = t + 1} ^ {n} \\log \\frac {\\exp \\left(q _ {t} \\cdot k _ {j} ^ {+} / \\tau\\right)}{\\sum_ {i = 0} ^ {K ^ {-}} \\exp \\left(q _ {t} \\cdot k _ {i} ^ {-} / \\tau\\right) + \\exp \\left(q _ {t} \\cdot k _ {j} ^ {+} / \\tau\\right)}. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            500,
+            571,
+            901,
+            625
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For one anchor as query  $q_{0}(t = 0)$ , the paired InfoNCE loss tries to classify it as  $k_{j}^{+}, j \\in [1, n]$ . The next step is to go forward: classify the first positive  $q_{1}(t = 1)$  of the anchor as  $k_{j}^{+}, j \\in [2, n]$ , and so on, as shown in the algorithm 1, Appendix A.1. To contrast with the paired InfoNCE, we refer to the original InfoNCE loss as unpaired InfoNCE hereafter. The behavior difference between unpaired and paired InfoNCE losses is compared in Figure 3. The paired InfoNCE only maximizes the distance between anchors and -anchors in the embedding space, neglecting irrelevant samples. The implementation of the paired InfoNCE loss is introduced in detail in Appendix A.2.",
+        "bbox": [
+            496,
+            626,
+            893,
+            806
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3. Pretext Task",
+        "text_level": 1,
+        "bbox": [
+            500,
+            816,
+            633,
+            830
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Since the aim of this work is to enable networks to strip targets (anchors) from their surroundings ( $\\neg$ anchors), we use a restricted instance discrimination task [57]: a query and a key are positive pairs if they are different views (by",
+        "bbox": [
+            496,
+            839,
+            893,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "5688",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/94bcab4c34e62920d5c29550c8ba28bbd49c686ee525a78e3d55635c96dac62c.jpg",
+        "image_caption": [
+            "Figure 2: SemCL dataset. The semantic label of a sample is utilized to generate binary masks. Each binary mask depicts the pixel-level information of one object (anchor) in the raw image. Those binary masks, together with their inverts, are applied on the raw image to extract objects (anchors) and corresponding surroundings ( $\\neg$ anchors)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            76,
+            88,
+            890,
+            335
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/f00a92c7ece71a2863a0c55f5a16a3c78bc0764ebe9e51e37957ba554ae0cf66.jpg",
+        "image_caption": [
+            "Figure 3: Comparison between unpaired (left) and paired InfoNCE losses (right). Left: a sample (anchor) is positive only to its own augmentations, while negative to others in the same mini-batch. Right: A sample (anchor) is positive only to its own augmentations, while negative only to its paired counterparts."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            76,
+            405,
+            519,
+            676
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/562f206f886369d73ce424b9fecb105a9fbdc964e9552b00571342f39e1ab72a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            405,
+            893,
+            676
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "data augmentation) of the same anchor, and are negative pairs if they originate from an anchor and its corresponding surroundings ( $\\neg$ anchors). A PyTorch-style pseudocode is given in the Appendix A.2. The inputs are encoded by base encoder  $f_{b}$  and momentum encoder  $f_{m}$ , producing queries and keys. Theoretically, the backbone of  $f_{b}$  and  $f_{m}$  can be any architecture. In practice, we use ResNet [23] and Swin Transformer [37] because they are the most representative architectures in recent years. All keys are stacked along",
+        "bbox": [
+            75,
+            758,
+            472,
+            896
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "dimension 1 for paired InfoNCE loss.",
+        "bbox": [
+            500,
+            758,
+            750,
+            773
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            796,
+            633,
+            813
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Considering the enormous scale difference between ImageNet and SemCL dataset, all SemCL models are initialized by corresponding IN pretrained weights, which are compared as counterparts in downstream tasks. The total number of pretraining iterations is set to  $30k$  to minimize the significant",
+        "bbox": [
+            496,
+            825,
+            895,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "5689",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "variation in training time caused by scale differences among SemCL sub-datasets. We adopt AdamW [40] with  $wd = 0.1$ , and the OneCycle [46] learning rate schedule consisting of warmup [17] for the initial 3, 750 (warmup ratio 0.125) steps and cosine annealing [9, 39]. Following [9], the base  $lr$  is set to  $1.5 \\times 10^{-4}$  and scaled linearly by  $lr \\times batchsize / 256$  and temperature  $\\tau = 0.2$ . The initial MoCo momentum is 0.99 and is gradually increased to 1 with a half cycle cosine schedule. The data augmentation strategy includes  $224 \\times 224$  random crops after random rescaling the original image with an area ratio in the range [0.08, 1.0], followed by random color jitter, random grayscale, random Gaussian blur, and random horizontal flip. All batch norm (BN) layers are synchronized across GPUs (SyncBN [41]). Limited by the two dual-GPU nodes (RTX3090  $\\times$  2+Titan RTX  $\\times$  2) we use, a native batch size of 64 can be applied to all ResNets and up to Swin-B for Swin Transformers with crop size  $224 \\times 224$ . For Swin-L in particular, we compromise by using a batch size of 32 with gradient accumulation over 2 iterations to achieve an equivalent batch size of 64.",
+        "bbox": [
+            76,
+            90,
+            472,
+            391
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "SemCL aims to improve the understanding of spatial information of models for which downstream tasks of semantic segmentation, object detection and depth estimation are conducted. Unless otherwise stated, all SemCL backbones are pretrained on the corresponding SemCL sub-datasets. The implementation details of downstream tasks are listed in Appendix A.3.",
+        "bbox": [
+            75,
+            393,
+            472,
+            500
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Semantic Segmentation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            508,
+            294,
+            526
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Backbone comparison. In the semantic segmentation benchmark on validation sets of VOC2012, Cityscapes, ADE20K and COCO 2017, SemCL models pretrained on the corresponding SemCL sub dataset are compared with ImageNet pretrained counterparts in Table 2. For VOC2012 and COCO, SemCL outperforms its IN counterparts by a maximum of 0.83 and 0.66 points respectively. SemCL can comprehensively outperform its IN counterparts on all backbones in the Cityscapes and ADE20k benchmarks.",
+        "bbox": [
+            75,
+            532,
+            472,
+            669
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "System-level comparison. SemCL pretrained backbones are compared with other methods on semantic segmentation tasks in Table 3. On VOC2012, SemCL R50/R101 outperform Deeplabv3+ [7] by 2.37/3.12 points. The Deeplabv3+ [7] X-71 is 0.55 points behind SemCL R101 on Cityscapes. For  $\\mathrm{CP}^2$  [51] ViT-S/16, SemCL Swin-S significantly outperforms it on all benchmarks by 4.97, 4.53 and 6.71 points. Compared to MAE [21] ViT-B, SemCL Swin-B achieves a gain of 0.58 points. A qualitative comparison between MAE [21] ViT-B and SemCL Swin-B is shown in Figure 4.",
+        "bbox": [
+            75,
+            670,
+            472,
+            821
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Object Detection and Instance Segmentation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            830,
+            455,
+            848
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Backbone comparison. In the benchmark for object detection and instance segmentation on the test set of VOC07 and validation sets of Cityscapes and COCO 2017, SemCL",
+        "bbox": [
+            76,
+            854,
+            472,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/4464714652d0d386d11cba4348eea09574b3c6db395b35b868060c353a60eee7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            101,
+            733,
+            189
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/3f932164755daac84f3361f1fbec4912e01ed15afe9034f869ae861a85472cbb.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            733,
+            101,
+            888,
+            188
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/1eb08784a2efb5c367b9822502a620995202afc14e03ad7395adfa34a6525eb2.jpg",
+        "image_caption": [
+            "Figure 4: Qualitative comparison of semantic segmentation between MAE [21] ViT-B (top) and SemCL Swin-B (bottom) on ADE20K semantic segmentation validation set. Both decoders are UPerNet[58]."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            503,
+            191,
+            733,
+            281
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/f2f480a4cf239efbceb69cbd05d3d16e03ca52ff8605d6ace2cba415ac391135.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            733,
+            191,
+            888,
+            281
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "models pretrained on the corresponding SemCL subset are compared with IN pretrained counterparts in Table 4. SemCL comprehensively outperforms its IN counterparts on the Cityscapes and COCO benchmarks. In particular, on the instance segmentation task, SemCL improves by a maximum of 0.5 and 0.3 points on Cityscapes and COCO respectively. However, SemCL performs worse than its IN counterparts in VOC detection, which is considered to be a result of insufficient training samples in the SemCL-VOC dataset and is further investigated in the ablation study (see Appendix A.4).",
+        "bbox": [
+            496,
+            378,
+            893,
+            545
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/881e02686d5b8419b9404f054eedae8382a4634bf805d80c8a92204695daeeaa.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">VOC2007 \nAP50</td><td colspan=\"2\">Cityscapes†</td><td colspan=\"2\">COCO</td></tr><tr><td>APbox</td><td>APmask</td><td>APbox</td><td>APmask</td></tr><tr><td>MoCo v3[9] IN-1k</td><td>81.88</td><td>42.7</td><td>37.5</td><td>42.6</td><td>37.9</td></tr><tr><td>SemCL</td><td>81.57(-0.31)</td><td>42.9(+0.2)</td><td>38.0(+0.5)</td><td>42.8(+0.2)</td><td>38.0(+0.1)</td></tr></table>",
+        "bbox": [
+            504,
+            556,
+            895,
+            614
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/b66851922523b6f50ca535f239e5dc7ddc52139143ad1dda8bb753c92e07334c.jpg",
+        "table_caption": [
+            "(a) R50"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">VOC2007 \nAP50</td><td colspan=\"2\">Cityscapes†</td><td colspan=\"2\">COCO</td></tr><tr><td>APbox</td><td>APmask</td><td>APbox</td><td>APmask</td></tr><tr><td>Swin[37] IN-22k</td><td>86.58</td><td>45.9</td><td>39.8</td><td>49.2</td><td>43.0</td></tr><tr><td>SemCL</td><td>86.53(-0.05)</td><td>46.0(+0.1)</td><td>40.2(+0.4)</td><td>49.0(-0.2)</td><td>43.3(+0.3)</td></tr></table>",
+        "bbox": [
+            504,
+            643,
+            895,
+            702
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "(b) Swin-T",
+        "bbox": [
+            669,
+            705,
+            730,
+            718
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Results of object detection and instance segmentation on various backbones w/ Cascade Mask R-CNN. All results are from our implementation.  $\\dagger$ : Sync BN is disabled due to gradient accumulation.",
+        "bbox": [
+            496,
+            729,
+            893,
+            791
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "System-level comparison. Table 5 compares the object detection and instance segmentation results of SemCL backbones with other models. On Cityscapes and COCO, SemCL backbones comprehensively outperform their twinborn MoCo Specifically, SemCL ResNet50 backbones achieve  $+1.7$  over MoCo IG-1B [22] in COCO detection and",
+        "bbox": [
+            496,
+            809,
+            895,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "5690",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/082d3671feab936b337929b5aef38900265e6cd95b0e1b788214a2974c3c3f21.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "$\\dagger$  Although MoCo v3 [9] is originally based on ViT, R50 pretrained models are provided in their repo https://github. com/facebookresearch/moco-v3."
+        ],
+        "table_body": "<table><tr><td>Backbone</td><td>Method</td><td>VOC2012</td><td>Cityscapes</td><td>ADE20K</td><td>COCO</td></tr><tr><td rowspan=\"2\">R50</td><td>MoCo v3[9] IN-1k†</td><td>79.65</td><td>79.01</td><td>42.71</td><td>40.59</td></tr><tr><td>SemCL</td><td>79.57(-0.08)</td><td>79.13(+0.12)</td><td>42.96(+0.25)</td><td>40.80(+0.21)</td></tr><tr><td rowspan=\"2\">R101</td><td>timm[55] IN-1k</td><td>81.82</td><td>79.87</td><td>44.43</td><td>41.43</td></tr><tr><td>SemCL</td><td>81.97(+0.15)</td><td>80.10(+0.23)</td><td>44.66(+0.23)</td><td>41.58(+0.15)</td></tr><tr><td rowspan=\"2\">Swin-T</td><td>Swin[37] IN-22k</td><td>80.90</td><td>78.67</td><td>44.60</td><td>42.39</td></tr><tr><td>SemCL</td><td>81.73(+0.83)</td><td>78.91(+0.26)</td><td>44.72(+0.12)</td><td>42.53(+0.14)</td></tr><tr><td rowspan=\"2\">Swin-S</td><td>Swin[37] IN-22k</td><td>83.71</td><td>81.25</td><td>47.62</td><td>44.75</td></tr><tr><td>SemCL</td><td>84.47(+0.76)</td><td>81.53(+0.28)</td><td>47.91(+0.29)</td><td>44.57(-0.18)</td></tr><tr><td rowspan=\"2\">Swin-B</td><td>Swin[37] IN-22k</td><td>84.64</td><td>81.61</td><td>48.21</td><td>44.81</td></tr><tr><td>SemCL</td><td>84.67(+0.03)</td><td>81.94(+0.33)</td><td>48.68(+0.47)</td><td>45.47(+0.66)</td></tr></table>",
+        "bbox": [
+            148,
+            88,
+            823,
+            296
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/a88f6f47b2805d265dad4e1af595c06f4e49492eed648c230cdfb35669b82287.jpg",
+        "table_caption": [
+            "Table 2: SemCL vs. ImageNet pretrained backbones on VOC2012, Cityscapes, ADE20K and COCO semantic segmentation (mIoU). All results are from our implementation."
+        ],
+        "table_footnote": [
+            "(a) VOC2012"
+        ],
+        "table_body": "<table><tr><td>Method</td><td>Backbone</td><td>mIoU</td></tr><tr><td>Deeplabv3+[7]</td><td>R50</td><td>77.2</td></tr><tr><td>Supervised-IN[22]</td><td>R50</td><td>74.4</td></tr><tr><td>MoCo IG-1B[22]</td><td>R50</td><td>75.5</td></tr><tr><td>BYOL[18]</td><td>R50</td><td>76.3</td></tr><tr><td>SemCL</td><td>R50</td><td>79.57</td></tr><tr><td>Deeplabv3+[7]</td><td>R101</td><td>78.85</td></tr><tr><td>Res2Net[15]</td><td>Res2Net-101</td><td>80.2</td></tr><tr><td>DFN[61]</td><td>R101</td><td>80.60</td></tr><tr><td>SemCL</td><td>R101</td><td>81.97</td></tr><tr><td>CP2[51]</td><td>ViT-S/16</td><td>79.5</td></tr><tr><td>SemCL</td><td>Swin-S</td><td>84.47</td></tr><tr><td>Leopart[70]</td><td>ViT-B/8</td><td>76.3</td></tr><tr><td>SemCL</td><td>Swin-B</td><td>84.67</td></tr></table>",
+        "bbox": [
+            148,
+            388,
+            370,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/7ea98b2dbfb2b38ce71183e887d7da52cb8af53865505b61f735ea0865ed291f.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "(b) Cityscapes"
+        ],
+        "table_body": "<table><tr><td>Method</td><td>Backbone</td><td>mIoU</td></tr><tr><td>MoCo IN-1M[22]</td><td>R50</td><td>72.5</td></tr><tr><td>MoCo IG-1B[22]</td><td>R50</td><td>73.6</td></tr><tr><td>Supervised-IN[22]</td><td>R50</td><td>74.6</td></tr><tr><td>CP2[51]</td><td>R50</td><td>76.5</td></tr><tr><td>SemCL</td><td>R50</td><td>79.13</td></tr><tr><td>Deeplabv3+[7]</td><td>X-71</td><td>79.55</td></tr><tr><td>SemCL</td><td>R101</td><td>80.10</td></tr><tr><td>Trans4Trans-T[62]</td><td>PVTv2-B1</td><td>78.23</td></tr><tr><td>SemCL</td><td>Swin-T</td><td>78.91</td></tr><tr><td>CP2[51]</td><td>ViT-S/16</td><td>77.0</td></tr><tr><td>Trans4Trans-S[62]</td><td>PVTv2-B2</td><td>80.02</td></tr><tr><td>SemCL</td><td>Swin-S</td><td>81.53</td></tr></table>",
+        "bbox": [
+            375,
+            388,
+            607,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/109acbebfc1947ec79c70f17f2f219f2b04e15e3050ce38334797cd51d81bc36.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "(c) ADE20K"
+        ],
+        "table_body": "<table><tr><td>Method</td><td>Backbone</td><td>mIoU</td></tr><tr><td>CP2[51]</td><td>R50</td><td>39.2</td></tr><tr><td>PSPNet[66]</td><td>R50</td><td>42.78</td></tr><tr><td>SemCL</td><td>R50</td><td>42.96</td></tr><tr><td>PSPNet[66]</td><td>R101</td><td>43.29</td></tr><tr><td>SemCL</td><td>R101</td><td>44.66</td></tr><tr><td>CP2[51]</td><td>ViT-S/16</td><td>41.2</td></tr><tr><td>Seg-S/16[47]</td><td>ViT-S</td><td>45.37</td></tr><tr><td>SemCL</td><td>Swin-S</td><td>47.91</td></tr><tr><td>Seg-B/16[47]</td><td>Vi-B/16</td><td>48.06</td></tr><tr><td>MAE[21]</td><td>ViT-B</td><td>48.1</td></tr><tr><td>SemCL</td><td>Swin-B</td><td>48.68</td></tr></table>",
+        "bbox": [
+            611,
+            388,
+            823,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Comparison with previous best results on validation sets of VOC2012, Cityscapes, and ADE20K semantic segmentation (mIoU). Best results are bolded.",
+        "bbox": [
+            73,
+            597,
+            895,
+            628
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "+5.1/+0.6 in Cityscapes/COCO instance segmentation tasks. For SemCL Swin-T backbones, SemCL achieves +1.2/+0.1 over Focal [60] on COCO detection/instance segmentation. Figure 5 demonstrates a qualitative comparison between Mask2Former [10] and SemCL Swin-T on Cityscapes validation set. With the improved spatial information understanding, SemCL model clearly discriminates (even overlapped) instances in Figures 5a and 5b and successfully classify part of an instance to a defined category as in Figure 5c.",
+        "bbox": [
+            73,
+            654,
+            470,
+            791
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Depth Estimation",
+        "text_level": 1,
+        "bbox": [
+            75,
+            801,
+            250,
+            816
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Backbone comparison. SemCL pretrained backbones are compared with corresponding IN pretrained ones in Table 6. For cityscapes, SemCL comprehensively outperforms its IN pretrained counterparts. In particular, the ResNet50 backbones improve the RMSE by 0.138 points. For KITTI",
+        "bbox": [
+            75,
+            824,
+            472,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "and NTUv2, all SemCL backbones show improvements over the IN pretrained ones. The SemCL pretrained SwinL model achieves improvements of 0.001/0.026 points on KITTI AbsRel/RMSE, and the SemCL pretrained ResNet50 model achieves improvements of 0.003/0.005 points on NYUv2.",
+        "bbox": [
+            496,
+            654,
+            893,
+            744
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "System-level comparison. We report the comparison of SemCL pretrained backbones with the previous state-of-the-art method Binsformer[35] and other works in Table 7. As for the Cityscapes results shown in Table 7a, SemCL pretrained ResNet50 significantly outperforms SDC [53] by  $-0.089$  and  $-2.218$  points on AbsRel and RMSE respectively. And SD-SSMDE [43] lags behind SemCL by  $0.005 / 3.742$  points. Meanwhile, the test on KITTI (Eigen split [13]) is shown in Table 7b, SemCL pretrained ResNet50 and Swin-T comprehensively outperform the previous SOTA",
+        "bbox": [
+            496,
+            750,
+            895,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "5691",
+        "bbox": [
+            482,
+            944,
+            513,
+            957
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/f0aea2d67442d3453f0fbb5112f6eacee395552c46628977e1c37de8d71aa79e.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "(a) Object det. on VOC2007 test."
+        ],
+        "table_body": "<table><tr><td>Method</td><td>Backbone</td><td>AP50</td></tr><tr><td>Supervised-IN[22]</td><td>R50</td><td>81.3</td></tr><tr><td>MoCo IN-1M[22]</td><td>R50</td><td>81.5</td></tr><tr><td>MoCo IG-1B[22]</td><td>R50</td><td>82.2</td></tr><tr><td>MoCo v2 800ep[8]</td><td>R50</td><td>82.5</td></tr><tr><td>SemCL</td><td>R50</td><td>81.57</td></tr><tr><td>UP-DETR/300[12]</td><td>*</td><td>80.1</td></tr><tr><td>OW-DETR[19]</td><td>DDETR</td><td>82.1</td></tr><tr><td>SemCL</td><td>Swin-T</td><td>86.53</td></tr></table>",
+        "bbox": [
+            106,
+            99,
+            331,
+            224
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/205c13fa2f3d33f22861b55e37ffc24e67b41f95f70bbaef87e043681cae870d.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "(b) Instance seg. on Cityscapes val."
+        ],
+        "table_body": "<table><tr><td>Method</td><td>Backbone</td><td>APmask</td></tr><tr><td>MoCo IN-1M[22]</td><td>R50</td><td>32.3</td></tr><tr><td>MoCo IG-1B[22]</td><td>R50</td><td>32.9</td></tr><tr><td>Supervised-IN[22]</td><td>R50</td><td>32.9</td></tr><tr><td>Mask2Former[10]</td><td>R50</td><td>37.4</td></tr><tr><td>SemCL</td><td>R50</td><td>38.0</td></tr><tr><td>BoundaryFormer[30]</td><td>*</td><td>38.3</td></tr><tr><td>Mask2Former[10]</td><td>Swin-T</td><td>39.7</td></tr><tr><td>SemCL</td><td>Swin-T</td><td>40.2</td></tr></table>",
+        "bbox": [
+            334,
+            99,
+            578,
+            224
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/b0115554c9ab6aba34934dba7a61a1063d74f807831ed54f6be7bf3d84103743.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "(c) Object det. and instance seg. on COCO val2017."
+        ],
+        "table_body": "<table><tr><td>Method</td><td>Backbone</td><td>APbox</td><td>APmask</td></tr><tr><td>Supervised-IN[22]</td><td>R50</td><td>40.6</td><td>36.8</td></tr><tr><td>MoCo IN-1M[22]</td><td>R50</td><td>40.8</td><td>36.9</td></tr><tr><td>MoCo IG-1B[22]</td><td>R50</td><td>41.1</td><td>37.4</td></tr><tr><td>SemCL</td><td>R50</td><td>42.8</td><td>38.0</td></tr><tr><td>MST[34]</td><td>Swin-T</td><td>42.7</td><td>38.8</td></tr><tr><td>VSA[63]</td><td>Swin-T</td><td>46.9</td><td>42.1</td></tr><tr><td>Focal[60]</td><td>Focal-Base</td><td>47.8</td><td>43.2</td></tr><tr><td>SemCL</td><td>Swin-T</td><td>49.0</td><td>43.3</td></tr></table>",
+        "bbox": [
+            583,
+            99,
+            864,
+            224
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/8bf50fb1d870c899a12d71afc831c2a9dc3a804607e544327d9a9c0fa27a4ca5.jpg",
+        "table_caption": [
+            "Table 5: System-level comparison on object detection and instance segmentation. The best results are **bolded**. * claims their backbones are CNN-based, but are considered and compared as transformer-based networks, since transformer blocks are massively used."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Backbone</td><td rowspan=\"2\">Method</td><td colspan=\"2\">Cityscapes</td><td colspan=\"2\">KITTI</td><td colspan=\"2\">NYUv2</td></tr><tr><td>AbsRel</td><td>RMSE</td><td>AbsRel</td><td>RMSE</td><td>AbsRel</td><td>RMSE</td></tr><tr><td rowspan=\"2\">R50</td><td>MoCo v3[9] IN-1k</td><td>0.138</td><td>4.837</td><td>0.059</td><td>2.337</td><td>0.120</td><td>0.405</td></tr><tr><td>SemCL</td><td>0.138</td><td>4.699(-0.138)</td><td>0.058(-0.001)</td><td>2.337</td><td>0.117(-0.003)</td><td>0.400(-0.005)</td></tr><tr><td rowspan=\"2\">Swin-T</td><td>Swin[37] IN-22k</td><td>0.134</td><td>4.643</td><td>0.057</td><td>2.247</td><td>0.114</td><td>0.386</td></tr><tr><td>SemCL</td><td>0.133(-0.001)</td><td>4.577(-0.066)</td><td>0.057</td><td>2.237(-0.010)</td><td>0.112(-0.002)</td><td>0.381(-0.005)</td></tr><tr><td rowspan=\"2\">Swin-L</td><td>Swin[37] IN-22k</td><td>0.130</td><td>4.536</td><td>0.053</td><td>2.130</td><td>0.095</td><td>0.329</td></tr><tr><td>SemCL</td><td>0.128(-0.002)</td><td>4.426(-0.110)</td><td>0.052(-0.001)</td><td>2.104(-0.026)</td><td>0.094(-0.001)</td><td>0.329</td></tr></table>",
+        "bbox": [
+            78,
+            311,
+            895,
+            445
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 6: SemCL vs. ImageNet pretraining on Cityscapes, KITTI and NYUv2 depth estimation (lower is better). All results are from our implementation.",
+        "bbox": [
+            75,
+            453,
+            892,
+            484
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4f828fa5596d72bf9b809e81a5743ab171f28b17213a00bec4ae5cf72cb1236d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            508,
+            467,
+            582
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/9eebcdaf5d930380d32a1ce1228b380344b5f5c2f25b0450afc3010ebbb72469.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            583,
+            467,
+            654
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d14dc188fc07519c4c2d193ac715f644034b4cf0ec9484d690735ac98a5ff5cd.jpg",
+        "image_caption": [
+            "Figure 5: Qualitative comparison of instance segmentation between Mask2Former [10] (left) and SemCL (right) Swin-T on Cityscapes validation set."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            655,
+            467,
+            729
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Binsformer. In particular, SemCL ResNet50 achieves a gain of  $-0.003 / - 0.089$  over Binsformer, and SemCL Swin-L is on par with Binsformer Swin-L. On NYUv2 (see Table 7c), the gains of SemCL pretrained ResNet50 are also high with  $-0.010$  AbsRel and  $-0.173$  RMSE. The SemCL",
+        "bbox": [
+            75,
+            824,
+            472,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/eb131c095b7ec1c7b169c6f45ea819e017f9722d5d9360e6ab2443f96f6cf025.jpg",
+        "image_caption": [
+            "Figure 6: Qualitative comparison of depth estimation between Binsformer[35] and SemCL Swin-T on the NYUv2 test set."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            521,
+            890,
+            768
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Swin-L RMSE is 0.001 points lower than the SOTA Binsformer. Qualitative comparison between Binsformer [35] and SemCL Swin-T is given in Figure 6.",
+        "bbox": [
+            498,
+            854,
+            893,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "5692",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/4a2a00e6430d21ed3b2c4f48b4c6933169b22d91c026c758bf0265e2e6acd86b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Backbone</td><td>AbsRel</td><td>RMSE</td></tr><tr><td>UpProj[29]</td><td>R50</td><td>0.257[53]</td><td>7.273[53]</td></tr><tr><td>Pad-net[59]</td><td>R50</td><td>0.246[53]</td><td>7.117[53]</td></tr><tr><td>TRL[64]</td><td>R50</td><td>0.234[53]</td><td>7.104[53]</td></tr><tr><td>Pilzer, et al.[45]</td><td>-</td><td>0.440</td><td>5.443</td></tr><tr><td>SDC-Depth[53]</td><td>R50</td><td>0.227</td><td>6.917</td></tr><tr><td>DTS-Depth[27]</td><td>MobileNetv2</td><td>0.167</td><td>7.785</td></tr><tr><td>struct2depth[5]</td><td>-</td><td>0.151</td><td>7.024</td></tr><tr><td>SD-SSMDE[43]</td><td>R50</td><td>0.143</td><td>8.441</td></tr><tr><td>SemCL</td><td>R50</td><td>0.138</td><td>4.699</td></tr></table>",
+        "bbox": [
+            99,
+            99,
+            397,
+            224
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/8351eebc395c62ca76c3574320eaf60e2e6803d09c41d05afa5ff76716febdd8.jpg",
+        "table_caption": [
+            "(a) Cityscapes"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Backbone</td><td>AbsRel</td><td>RMSE</td></tr><tr><td>EPCDepth [42]</td><td>R50</td><td>0.091</td><td>4.207</td></tr><tr><td>Binsformer[35]</td><td>R50</td><td>0.061</td><td>2.426</td></tr><tr><td>SemCL</td><td>R50</td><td>0.058</td><td>2.337</td></tr><tr><td>MonoViT[65]</td><td>-</td><td>0.098</td><td>4.333</td></tr><tr><td>Binsformer [35]</td><td>Swin-T</td><td>0.058</td><td>2.286</td></tr><tr><td>SemCL</td><td>Swin-T</td><td>0.057</td><td>2.237</td></tr><tr><td>DepthFormer [33]</td><td>Swin-L +R50CI</td><td>0.052</td><td>2.143</td></tr><tr><td>Binsformer[35]</td><td>Swin-L</td><td>0.052</td><td>2.098</td></tr><tr><td>SemCL</td><td>Swin-L</td><td>0.052</td><td>2.104</td></tr></table>",
+        "bbox": [
+            400,
+            99,
+            629,
+            224
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/71e30f17f65e81b2c4e177c6ae6216589702ba758817f108caa56853fec9b7b7.jpg",
+        "table_caption": [
+            "(b) KITTI"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Backbone</td><td>AbsRel</td><td>RMSE</td></tr><tr><td>Pad-net [59]</td><td>R50</td><td>0.214</td><td>0.792</td></tr><tr><td>TRL [64]</td><td>R50</td><td>0.144</td><td>0.501</td></tr><tr><td>SDC-Depth [53]</td><td>R50</td><td>0.128</td><td>0.497</td></tr><tr><td>UpProj [29]</td><td>R50</td><td>0.127</td><td>0.573</td></tr><tr><td>SemCL</td><td>R50</td><td>0.117</td><td>0.400</td></tr><tr><td>Binsformer [35]</td><td>Swin-T</td><td>0.113</td><td>0.379</td></tr><tr><td>SemCL</td><td>Swin-T</td><td>0.112</td><td>0.381</td></tr><tr><td>Binsformer [35]</td><td>Swin-L</td><td>0.094</td><td>0.330</td></tr><tr><td>SemCL</td><td>Swin-L</td><td>0.094</td><td>0.329</td></tr></table>",
+        "bbox": [
+            632,
+            99,
+            872,
+            224
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "(c) NYUv2",
+        "bbox": [
+            720,
+            227,
+            781,
+            239
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 7: System-level comparison on depth estimation. Best results (lower is better) are bolded.",
+        "bbox": [
+            166,
+            252,
+            797,
+            267
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.4. Attention Map Visualization",
+        "text_level": 1,
+        "bbox": [
+            76,
+            294,
+            330,
+            310
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To improve the qualitative understanding of the SemCL representation, we visualize attention maps of Swin-T pretrained on SemCL-VOC and ImageNet using PyTorch library for CAM methods [16] with the Score-CAM method [52]. These samples in Figure 7 selected from VOC2012 val and have not been seen by the pretrained models are characterized by simple centered subjects (Figure 7a), complex scene with occlusion (Figure 7b) and multiple subjects (Figures 7c and 7d).",
+        "bbox": [
+            75,
+            316,
+            470,
+            452
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In Figure 7a, the IN pretrained model failed to focus on the subject, while the SemCL model accurately depicted the area of the main subject. Meanwhile, SemCL performs better in complex scenes containing foreground-background relationships, as shown in Figure 7b. The IN pretrained model pays no attention to an occluded object, but the SemCL model recognizes that the occluded part also belongs to the horse, as shown in Figure 7b. In Figure 7c, the IN model directly ignores the distant person, but the SemCL model not only annotates two people but also focuses more on the nearby one, suggesting that the SemCL model can understand the spatial distribution of subjects in an image. Also, in Figure 7d, the SemCL model successfully recognizes one plant and its pot as one object.",
+        "bbox": [
+            75,
+            453,
+            468,
+            664
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Summary. SemCL pretrained backbones can substantially outperform ImageNet pretrained counterparts on semantic segmentation, object detection, instance segmentation and depth estimation tasks. Using semantically contrastive pairs generated from off-the-shelf datasets, SemCL significantly improves the spatial information understanding of pretrained models. Such an improvement is also promising to benefit more challenging tasks including generative ones [32]. Ablation study concerning paired/unpaired InfoNCE loss, pretraining batch size, training length and dataset scale are given in Appendix A.4.",
+        "bbox": [
+            75,
+            665,
+            470,
+            832
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            76,
+            844,
+            194,
+            859
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We investigate the effectiveness of semantic information in the contrastive learning pretext task on the spatial un",
+        "bbox": [
+            76,
+            869,
+            470,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5226e012f04effbc40caf2ee71320be0dd930e0f6780ee6e035ff270d3c00d18.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            292,
+            638,
+            383
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/bc20e95d32622b1d719688fb7165e8bcba5717039984cbeea7d857567e6845de.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            383,
+            638,
+            474
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/dde65f9160e7393015f2b97881d5058707c118ec74bc9b84b93e11c0565da30e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            474,
+            638,
+            564
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/7919bb40da1c18c0f38cacca0e4346a87da9f4a3fa4da570a4ea3730d441e554.jpg",
+        "image_caption": [
+            "Figure 7: Swin-T attention map visualization. Left: raw image. Middle: attention maps of the IN pretrained model. Right: attention maps of the SemCL pretrained model."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            565,
+            638,
+            655
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/1fc2ba11d96a5b524ff739c0e9eb982a41fd4562c0c91572ad37b10b6e55bb2b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            645,
+            292,
+            764,
+            383
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/693d5bde387383fc97b44205cca0feb880c9879a5708401aa62593fb1774048e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            645,
+            383,
+            764,
+            473
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/648068eebffd0c743bfea2d1146cabbd5569dc9fa89dc62364f507530818824e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            645,
+            473,
+            764,
+            564
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/231350e5e0a7098c25bc15d484f0c3ff1aca3dd2c956250a23babc7618d29816.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            645,
+            564,
+            764,
+            655
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/537e899522c70cbcab38245ca5a0863705d05756a4dcac98dba417070839438f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            292,
+            890,
+            383
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a6caed9c7280655bf1a27b87c9a3c67be42a401e0db67034708997b7285fbc54.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            383,
+            890,
+            474
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/689106e8ce1df92ee8ea95243925d7e07effab326b0a18106a216c5342d7e681.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            474,
+            890,
+            564
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/4f2b7f5cdecb8ffc041152283ccb5171172d6cb5b0e25a425fada50073325794.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            564,
+            890,
+            655
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "derstanding ability of models. Tests are conducted on semantic/instance segmentation, object detection and depth estimation tasks with both ResNets and Swin Transformers, and SemCL pretrained backbones substantially outperform their ImageNet pretrained counterparts. By supporting small batch sizes and fast pretraining, SemCL is a lightweight yet effective approach. In contrastive representation learning, if dedicated pretext task is designed properly, four ounces can move a thousand pounds. We hope that SemCL will inspire more advanced pretext tasks for contrastive learning.",
+        "bbox": [
+            496,
+            739,
+            893,
+            890
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "5693",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            173,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2",
+            "[2] Suzanne Becker and Geoffrey E Hinton. Self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355(6356):161-163, 1992. 2",
+            "[3] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a “siamese” time delay neural network. Advances in neural information processing systems, 6, 1993. 2",
+            "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Computer vision and pattern recognition (CVPR), 2018 IEEE conference on. IEEE, 2018. 3",
+            "[5] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Unsupervised monocular depth and ego-motion learning with structure and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 8",
+            "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 3",
+            "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2, 5, 6",
+            "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 7",
+            "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 1, 2, 5, 6, 7",
+            "[10] 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. 2, 6, 7",
+            "[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3",
+            "[12] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Updetr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1601-1610, 2021. 7",
+            "[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 6"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[14] 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. 3",
+            "[15] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multiscale backbone architecture. IEEE transactions on pattern analysis and machine intelligence, 43(2):652-662, 2019. 6",
+            "[16] Jacob Gildenblat and contributors. Pytorch library for cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021.8",
+            "[17] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: TrainingImagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5",
+            "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020. 6",
+            "[19] Akshita Gupta, Sanath Narayan, KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Ow-detr: Open-world detection transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9235-9244, 2022. 7",
+            "[20] Bharath Hariharan, Pablo Arbelaez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In International Conference on Computer Vision (ICCV), 2011. 3",
+            "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 5, 6",
+            "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 3, 5, 6, 7",
+            "[23] 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. 4",
+            "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 2",
+            "[25] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International Conference on Machine Learning, pages 4182-4192. PMLR, 2020. 2",
+            "[26] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In International Conference on Learning Representations, 2019. 2"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "5694",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[27] Hatem Ibrahim, Ahmed Salem, and Hyun-Soo Kang. Dts-depth: Real-time single-image depth estimation using depth-to-space image construction. Sensors, 22(5):1914, 2022. 8",
+            "[28] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 3",
+            "[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 8",
+            "[30] Justin Lazarow, Weijian Xu, and Zhuowen Tu. Instance segmentation with mask-supervised polygonal boundary transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4382-4391, 2022. 7",
+            "[31] Phuc H. Le-Khac, Graham Healy, and Alan F. Smeaton. Contrastive representation learning: A framework and review. IEEE Access, 8:193907-193934, 2020. 1, 2",
+            "[32] Yuheng Li, Yijun Li, Jingwan Lu, Eli Shechtman, Yong Jae Lee, and Krishna Kumar Singh. Contrastive learning for diverse disentangled foreground generation. In European Conference on Computer Vision, pages 334-351. Springer, 2022. 8",
+            "[33] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. arXiv preprint arXiv:2203.14211, 2022. 8",
+            "[34] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. Mst: Masked self-supervised transformer for visual representation. Advances in Neural Information Processing Systems, 34:13165-13176, 2021. 7",
+            "[35] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987, 2022. 2, 6, 7, 8",
+            "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 3",
+            "[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 4, 5, 6, 7",
+            "[38] Vishnu Suresh Lokhande, Songwong Tasneeyapant, Abhay Venkatesh, Sathya N Ravi, and Vikas Singh. Generating accurate pseudo-labels in semi-supervised learning and avoiding overconfident predictions via hermite polynomial activations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11435-11443, 2020. 2",
+            "[39] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017.5",
+            "[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5",
+            "[41] Chao Peng, Tete Xiao, Zeming Li, Yuning Jiang, Xiangyu Zhang, Kai Jia, Gang Yu, and Jian Sun. Megdet: A large minibatch object detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6181-6189, 2018. 5",
+            "[42] Rui Peng, Ronggang Wang, Yawen Lai, Luyang Tang, and Yangang Cai. Excavating the potential capacity of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15560-15569, 2021. 8",
+            "[43] Andra Petrovai and Sergiu Nedevschi. Exploiting pseudo labels in a self-supervised learning framework for improved monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1578-1588, 2022. 6, 8",
+            "[44] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11557-11568, 2021. 2",
+            "[45] Andrea Pilzer, Dan Xu, Mihai Puscas, Elisa Ricci, and Nicu Sebe. Unsupervised adversarial depth estimation using cycled generative networks. In 2018 international conference on 3D vision (3DV), pages 587-595. IEEE, 2018. 8",
+            "[46] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 5",
+            "[47] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 6",
+            "[48] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2",
+            "[49] Aaron Van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv e-prints, pages arXiv-1807, 2018. 1, 2, 3",
+            "[50] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3",
+            "[51] Feng Wang, Huiyu Wang, Chen Wei, Alan Yuille, and Wei Shen. Cp 2: Copy-paste contrastive pretraining for semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXX, pages 499-515. Springer, 2022. 5, 6",
+            "[52] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 8"
+        ],
+        "bbox": [
+            503,
+            92,
+            893,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "5695",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[53] Lijun Wang, Jianming Zhang, Oliver Wang, Zhe Lin, and Huchuan Lu. Sdc-depth: Semantic divide-and-conquer network for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 541-550, 2020. 6, 8",
+            "[54] 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. 3",
+            "[55] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476, 2021. 6",
+            "[56] Hao Wu and Saurabh Prasad. Semi-supervised deep learning using pseudo labels for hyperspectral image classification. IEEE Transactions on Image Processing, 27(3):1259-1270, 2017. 2",
+            "[57] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 1, 2, 3",
+            "[58] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 5",
+            "[59] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Padnet: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 675-684, 2018. 8",
+            "[60] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. Advances in Neural Information Processing Systems, 34:30008-30022, 2021. 6, 7",
+            "[61] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Learning a discriminative feature network for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1857-1866, 2018. 6",
+            "[62] Jiaming Zhang, Kailun Yang, Angela Constantinescu, Kunyu Peng, Karin Müller, and Rainer Stiefelhagen. Trans4trans: Efficient transformer for transparent object and semantic scene segmentation in real-world navigation assistance. IEEE Transactions on Intelligent Transportation Systems, 23(10):19173-19186, 2022. 6",
+            "[63] Qiming Zhang, Yufei Xu, Jing Zhang, and Dacheng Tao. Vsa: learning varied-size window attention in vision transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV, pages 466-483. Springer, 2022. 7",
+            "[64] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Zequn Jie, Xiang Li, and Jian Yang. Joint task-recursive learning for semantic segmentation and depth estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 235–251, 2018. 8",
+            "[65] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            901
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. arXiv preprint arXiv:2208.03543, 2022. 8",
+            "[66] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 6",
+            "[67] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 3",
+            "[68] 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. 3",
+            "[69] Chengxu Zhuang, Alex Lin Zhai, and Daniel Yamins. Local aggregation for unsupervised learning of visual embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6002-6012, 2019. 2",
+            "[70] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022. 6"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            416
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "5696",
+        "bbox": [
+            482,
+            944,
+            516,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2023/Semantic Information in Contrastive Learning/c3144656-62d8-4975-97d6-ef4630f410df_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e218b4fccb8a5db52bc216fe320d7f3b8cd1d5776971d71e082d902d83667a89
+size 8409101

2023/Semantic Information in Contrastive Learning/full.md

@@ -0,0 +1,346 @@
+# Semantic Information in Contrastive Learning
+
+Shengjiang Quan $^{1*}$ , Masahiro Hirano $^{2}$  and Yuji Yamakawa $^{3}$ \
+ $^{1}$ Graduate School of Engineering, The University of Tokyo, Japan\
+ $^{2}$ Institute of Industrial Science, The University of Tokyo, Japan\
+ $^{3}$ Interfaculty Initiative in Information Studies, The University of Tokyo, Japan\
+quan@g.ecc.u-tokyo.ac.jp, {mhirano, y-ymkw}@iis.u-tokyo.ac.jp
+
+# Abstract
+
+This work investigates the functionality of Semantic information in Contrastive Learning (SemCL). An advanced pretext task is designed: a contrast is performed between each object and its environment, taken from a scene. This allows the SemCL pretrained model to extract objects from their environment in an image, significantly improving the spatial understanding of the pretrained models. Downstream tasks of semantic/instance segmentation, object detection and depth estimation are implemented on PASCAL VOC, Cityscapes, COCO, KITTI, etc. SemCL pretrained models substantially outperform ImageNet pretrained counterparts and are competitive with well-known works on downstream tasks. The results suggest that a dedicated pretext task leveraging semantic information can be powerful in benchmarks related to spatial understanding. The code is available at https://github.com/sjiang95/semcl.
+
+# 1. Introduction
+
+Within the field of visual representation learning, contrastive learning attracts special attention for its inspiring performance in transfer learning [57, 49, 31, 22, 8, 9]. The concept of contrastive learning can literally be explained as discovering the difference between positive and negative samples. The definition of positive-negative samples is one of the main subjects of contrastive learning, which directly determines the pretext task and the corresponding loss function.
+
+We humans can recognize an object from its even complicated environment because we have learned and bound the typical geometric feature and the corresponding semantic concept of the object class. For example, we can recognize a cat in a scene thanks to our knowledge of its appearance (typical geometric feature) and the concept of cat (semantic information). Similarly, the target of prevalent
+
+![](images/bab89fd9041f4a80e562caeb78e23e0262a1f6e2656bcaff36d7a6fdd19d7503.jpg)  
+Figure 1: Perceptual cognition of SemCL. Leveraging semantic information, SemCL separates an object from its environment. Each sample is divided into objects and surroundings and considered as a contrastive pair. The task is to maximize the distance between the object and its environment in the embedding space.
+
+visual tasks, such as semantic/instance segmentation, object detection, etc., can be summarized as distinguishing and localizing subjects from surroundings. Inspired by and to extend the spatial recognition mechanism to CV, we propose our method SemCL which directly teaches models to extract a subject from its environment. Utilizing publicly off-the-shelf datasets providing semantic labels, SemCL can be considered as a supervised contrastive learning approach whose samples consist of subjects and their corresponding surroundings (semantically not the subject). The pretext task of SemCL is to tell the difference between one subject and its surroundings. From a representation learning perspective, the pretext task is to maximize the distance  $(L^2$  norm) between a subject and its environment in the embedding space (see Figure 1). The pretext task mimics the spatial recognition mechanism to discriminate objects and surroundings, which significantly improves the spatial information understanding of pretrained models.
+
+In SemCL, contrast is performed pairwise between positive-negative pairs from one image/scene. Inter-scene contrast is considered inappropriate: only the contrasts
+
+among subject-surrounding pairs from one scene are considered valid. Therefore, paired InfoNCE loss is adopted. With the mechanism of contrasting only paired samples, SemCL is characterized by decoupling the number of negative samples from the batch size. Compared to MoCo [22], whose performance is positively correlated with batch size and thus increases hardware cost (e.g. batchsize  $= 4096$  on 128 GPUs for ViT-B in MoCo v3 [9]) for satisfactory results, SemCL can achieve consistent yet competitive results on downstream tasks with relatively small batch size (e.g., batchsize  $= 64 \ @ 224 \times 224$ ). We adopt the MoCo v3 [9] as the pretraining framework.
+
+The primary motivation of representation learning is to pretrain general representations that can be fine-tuned and transferred to downstream tasks. In this work, the SemCL pretrained models are benchmarked on semantic segmentation, object detection and instance segmentation, and depth estimation tasks. SemCL models substantially outperform their ImageNet pretrained counterparts due to the improved ability of spacial information understanding. In a system-level comparison, SemCL models also show gains over previous well-known works (e.g. Deeplabv3+ [7] in semantic segmentation, Mask2Former [10] in instance segmentation and Binsformer [35] in depth estimation).
+
+# 2. Related Work
+
+# 2.1. Contrastive Learning
+
+The seminal idea of contrastive learning dates back at least to the 1990s [2, 3, 31], but it has become prominent in recent years thanks to the large pretrained models in the fields of NLP and CV [31]. As the name suggests, contrastive learning mines abstract representations by comparing between "similar" and "dissimilar" samples: minimizing the representation of "similar" samples and maximizing that of "dissimilar" samples in the embedding space. Formally, the "similar" inputs are called positive samples, and the "dissimilar" inputs are called negative samples.
+
+Defining "similar" and "dissimilar" samples is an important topic in contrastive learning. In instance discrimination [57], which is a ubiquitous approach to define positive-negative samples in CV, each sample in a minibatch  $x \in \{x_1, \ldots, x_n\}$  is considered mutually exclusive to the rest. The encoded output  $\theta(\cdot)$  of each sample is pushed away from the others in the embedding space, denoted as  $\theta(x_1) \Leftrightarrow \{\theta(x_2), \ldots, \theta(x_n)\} \Leftrightarrow \text{push}$ . Contrastive Predictive Coding (CPC) [49] defines positive-negative samples in a generative way: Given a sequential input  $\{x_{t-3}, x_{t-2}, x_{t-1}, x_t\}$ , a context latent representation  $C_t$  is computed to predict the encoded outputs of future inputs  $\{\theta(\mathbb{X}_{t+1}), \theta(\mathbb{X}_{t+2}), \theta(\mathbb{X}_{t+3}), \ldots\}$ . Those predicted future embeddings are considered as positive samples to real future embeddings, and exclusive to some other
+
+irrelevant inputs.
+
+# 2.2. Momentum Contrast
+
+He et al. [22] proposed a momentum contrast framework, MoCo, for contrastive representation learning. MoCo summed unsupervised visual representation learning [57, 49, 26, 69, 25, 48, 1] up as dictionary look-up: the input data are represented and sampled by an encoder network  $\theta_{k}$  as "keys"  $k = \theta_{k}(x^{k})$ . The goal of contrastive learning is to train encoders  $\theta_{q}$  by which a "query"  $q = \theta_{q}(x^{q})$  should be similar to its matching key and dissimilar to others. Since dictionary consistency is considered crucial for unsupervised learning with contrastive loss [22], MoCo introduces momentum updating. In the MoCo framework, the key coder  $\theta_{k}$  is progressively updated by the weighted average of  $\theta_{k}$  and the query coder  $\theta_{q}$ .
+
+$$
+\theta_ {k} \leftarrow m \theta_ {k} + (1 - m) \theta_ {q}, \tag {1}
+$$
+
+where the momentum coefficient  $m \in [0,1)$  is set quite high ( $m = 0.999$  in [22]) for slower key encoder update. The MoCo series has proven the power of the Momentum Contrast framework on feature transferring and is therefore employed as the basis for this work.
+
+# 2.3. Labeled Data In Unsupervised Framework
+
+Large networks are quite thirsty for training data, while labelling the large amount of collected data is a human resource consuming challenge. Apart from turning to unsupervised learning, which inherently avoids using labelled data, using unlabeled data in supervised learning is also a solution. Pseudo-Label [44, 38, 56] is a trick that uses (usually supervised) pretrained networks to generate features that can play the role of labels. And the generated pseudo-labels participate in the iterations of the target network just like annotations.
+
+In contrast, Khosla et al. [28] also investigated the function of labelled data in unsupervised learning. As a contrastive learning research, Khosla et al. [28] intuitively uses the classification label to define samples within the same class are mutually positive, and those from other categories are negative. To exploit the label information, clusters of points belonging to the same class are pulled together in the embedding space, while clusters of samples from different classes are pushed apart [28]:  $\left\{\theta\left(x_1^{c0}\right) \rightleftharpoons \theta\left(x_2^{c0}\right) \rightleftharpoons \theta\left(x_{n}^{c0}\right) \rightleftharpoons \theta\left(x_n^0\right)$ $\Leftrightarrow$ $\left\{\theta\left(x_1^{c1}\right) \rightleftharpoons \theta\left(x_2^{c1}\right) \rightleftharpoons \theta\left(x_{m}^{c1}\right)$  ( $\rightleftharpoons$  for pull). The experimental result shows that supervised contrastive learning can improve both the accuracy and robustness of classifiers - [28] achieves a top-1 accuracy of  $81.4\%$  on the ImageNet (IN) dataset, which is  $0.8\%$  above the best result reported by ResNet-200 architecture [24]. Likewise, our work investigates the effectiveness of semantic information in unsupervised framework contrastive learning.
+
+# 2.4. Semantic Information in Contrastive Learning
+
+Other than [28], there are works that use small amounts of labelled data in unsupervised representation learning. Wang et al. [54] proposed a pixel-wise contrastive algorithm for semantic segmentation: for a pixel  $i$ , it is pulled towards pixels belonging to the same category and pushed away from pixels in other categories in the embedding space. This contrast is performed across images from the same dataset. In the Cityscapes test benchmark, [54] R101 achieves 79.2 mIoU, which is 1.1 points better than Deeplabv3 [6].
+
+The attempt and results of works using labels in unsupervised representation learning, such as [28, 54], indicate that label information can bring gains to a fully unsupervised representation. It is promising to explore advanced pretext tasks to improve the performance of pretrained models in downstream tasks.
+
+# 3. Method
+
+# 3.1. SemCL Dataset
+
+The core of this work is to define semantically contrastive pairs by utilizing off-the-shelf semantic labels, which are mappings between categories of objects and their coordinates on the raw image at the pixel level. As decided by the authors of a dataset, semantic labels can be color or grayscale images. Each predefined color or grayscale denotes a category, and the distribution of label pixels reflects the location of pixels belonging to a particular category in the raw image.
+
+The SemCL dataset is generated by the mechanism shown in Figure 2, where the raw image and its semantic label are selected from the PASCAL VOC2012 dataset [14]. From the semantic label annotating two objects - a motorcycle and a rider - two binary masks BM0 and BM1 can be extracted to separately represent two annotated objects. Their inverted counterparts  $\neg$ BM0 and  $\neg$ BM1 represent pixels other than the target object. These binary masks are then bitwise applied to the raw image to separate an object (Anchor) from its environment ( $\neg$ Anchor), forming a semantically contrastive pair. Each element in a raw Anchor- $\neg$ Anchor pair is considered to be semantically opposite to the other.
+
+In practice, the SemCL-1M dataset is produced from training sets of augmented PASCAL VOC2012 [14] (also known as SBD [20]), Cityscapes [11], ADE20K [67, 68] and COCO [36] (stuff+thing 2017 [4]) datasets. The components of the SemCL dataset are illustrated in Table 1. When generating the dataset, a threshold  $t = 0.01$  is set to filter out too small (less than  $[\sqrt{t} \times \text{height}] \times [\sqrt{t} \times \text{width}]$ ) objects that are considered semantically unsaturated.
+
+# 3.2. Paired InfoNCE Loss
+
+Since contrastive learning can be considered as a dictionary look-up task [22], it is assumed that within a set of keys  $\{k_0, k_1, k_2, \ldots\}$  of a dictionary, a query  $q$  matches only one
+
+<table><tr><td></td><td>VOC2012</td><td>Cityscapes</td><td>ADE20K</td><td>COCO</td><td>Total</td></tr><tr><td># training samples</td><td>10,582</td><td>2,975</td><td>25,574</td><td>118,287</td><td>157,418</td></tr><tr><td># SemCL pairs</td><td>14,203</td><td>20,280</td><td>270,218</td><td>712,212</td><td>1,016,913</td></tr></table>
+
+Table 1: Components of the SemCL dataset. The SemCL datasets spawned from VOC2012(aug), Cityscapes, ADE20k and COCO 2017 are named SemCL- $\{\mathrm{VOC}$ , City, ADE, COCO\} respectively.
+
+key  $k_{+}$  of the dictionary [22]. The value of the objective function should be low if  $q$  is similar to its positive key  $k_{+}$  and dissimilar to all negative keys  $\{k_{0}, k_{1}, k_{2}, \ldots\} \setminus \{k_{+}\}$ . In MoCo [22], the form of a contrastive loss function InfoNCE [49] is adopted
+
+$$
+\mathcal {L} _ {q} = - \log \frac {\exp (q \cdot k _ {+} / \tau)}{\sum_ {i = 0} ^ {K} \exp (q \cdot k _ {i} / \tau)} \tag {2}
+$$
+
+where  $\tau$  is a scalar temperature hyperparameter controlling the strength of penalties on hard negative samples [50], e.g. in the instance discrimination task [49, 22], there may be potential positive samples in the randomly chosen negative samples. Summing over one positive and  $K$  negative samples, Equation (2) is the log loss of a  $(\mathbf{K} + 1)$ -way softmax-based classifier trying to classify  $q$  as  $k_{+}$  [22]. While in our case an anchor instance  $q$  compares only with its own  $n$  augmentations (positive)  $K^{+} := \{k_{1}^{+}, k_{2}^{+}, k_{3}^{+}, \ldots, k_{n}^{+}\}$  and paired  $n + 1$  negatives  $K^{-} := \{k_{0}^{-}, k_{1}^{-}, k_{2}^{-}, k_{3}^{-}, \ldots, k_{n}^{-}\}$  using paired InfoNCE
+
+$$
+\mathcal {L} _ {q} = - \sum_ {j = t + 1} ^ {n} \log \frac {\exp \left(q _ {t} \cdot k _ {j} ^ {+} / \tau\right)}{\sum_ {i = 0} ^ {K ^ {-}} \exp \left(q _ {t} \cdot k _ {i} ^ {-} / \tau\right) + \exp \left(q _ {t} \cdot k _ {j} ^ {+} / \tau\right)}. \tag {3}
+$$
+
+For one anchor as query  $q_{0}(t = 0)$ , the paired InfoNCE loss tries to classify it as  $k_{j}^{+}, j \in [1, n]$ . The next step is to go forward: classify the first positive  $q_{1}(t = 1)$  of the anchor as  $k_{j}^{+}, j \in [2, n]$ , and so on, as shown in the algorithm 1, Appendix A.1. To contrast with the paired InfoNCE, we refer to the original InfoNCE loss as unpaired InfoNCE hereafter. The behavior difference between unpaired and paired InfoNCE losses is compared in Figure 3. The paired InfoNCE only maximizes the distance between anchors and -anchors in the embedding space, neglecting irrelevant samples. The implementation of the paired InfoNCE loss is introduced in detail in Appendix A.2.
+
+# 3.3. Pretext Task
+
+Since the aim of this work is to enable networks to strip targets (anchors) from their surroundings ( $\neg$ anchors), we use a restricted instance discrimination task [57]: a query and a key are positive pairs if they are different views (by
+
+![](images/94bcab4c34e62920d5c29550c8ba28bbd49c686ee525a78e3d55635c96dac62c.jpg)  
+Figure 2: SemCL dataset. The semantic label of a sample is utilized to generate binary masks. Each binary mask depicts the pixel-level information of one object (anchor) in the raw image. Those binary masks, together with their inverts, are applied on the raw image to extract objects (anchors) and corresponding surroundings ( $\neg$ anchors).
+
+![](images/f00a92c7ece71a2863a0c55f5a16a3c78bc0764ebe9e51e37957ba554ae0cf66.jpg)  
+Figure 3: Comparison between unpaired (left) and paired InfoNCE losses (right). Left: a sample (anchor) is positive only to its own augmentations, while negative to others in the same mini-batch. Right: A sample (anchor) is positive only to its own augmentations, while negative only to its paired counterparts.
+
+![](images/562f206f886369d73ce424b9fecb105a9fbdc964e9552b00571342f39e1ab72a.jpg)
+
+data augmentation) of the same anchor, and are negative pairs if they originate from an anchor and its corresponding surroundings ( $\neg$ anchors). A PyTorch-style pseudocode is given in the Appendix A.2. The inputs are encoded by base encoder  $f_{b}$  and momentum encoder  $f_{m}$ , producing queries and keys. Theoretically, the backbone of  $f_{b}$  and  $f_{m}$  can be any architecture. In practice, we use ResNet [23] and Swin Transformer [37] because they are the most representative architectures in recent years. All keys are stacked along
+
+dimension 1 for paired InfoNCE loss.
+
+# 4. Experiments
+
+Considering the enormous scale difference between ImageNet and SemCL dataset, all SemCL models are initialized by corresponding IN pretrained weights, which are compared as counterparts in downstream tasks. The total number of pretraining iterations is set to  $30k$  to minimize the significant
+
+variation in training time caused by scale differences among SemCL sub-datasets. We adopt AdamW [40] with  $wd = 0.1$ , and the OneCycle [46] learning rate schedule consisting of warmup [17] for the initial 3, 750 (warmup ratio 0.125) steps and cosine annealing [9, 39]. Following [9], the base  $lr$  is set to  $1.5 \times 10^{-4}$  and scaled linearly by  $lr \times batchsize / 256$  and temperature  $\tau = 0.2$ . The initial MoCo momentum is 0.99 and is gradually increased to 1 with a half cycle cosine schedule. The data augmentation strategy includes  $224 \times 224$  random crops after random rescaling the original image with an area ratio in the range [0.08, 1.0], followed by random color jitter, random grayscale, random Gaussian blur, and random horizontal flip. All batch norm (BN) layers are synchronized across GPUs (SyncBN [41]). Limited by the two dual-GPU nodes (RTX3090  $\times$  2+Titan RTX  $\times$  2) we use, a native batch size of 64 can be applied to all ResNets and up to Swin-B for Swin Transformers with crop size  $224 \times 224$ . For Swin-L in particular, we compromise by using a batch size of 32 with gradient accumulation over 2 iterations to achieve an equivalent batch size of 64.
+
+SemCL aims to improve the understanding of spatial information of models for which downstream tasks of semantic segmentation, object detection and depth estimation are conducted. Unless otherwise stated, all SemCL backbones are pretrained on the corresponding SemCL sub-datasets. The implementation details of downstream tasks are listed in Appendix A.3.
+
+# 4.1. Semantic Segmentation
+
+Backbone comparison. In the semantic segmentation benchmark on validation sets of VOC2012, Cityscapes, ADE20K and COCO 2017, SemCL models pretrained on the corresponding SemCL sub dataset are compared with ImageNet pretrained counterparts in Table 2. For VOC2012 and COCO, SemCL outperforms its IN counterparts by a maximum of 0.83 and 0.66 points respectively. SemCL can comprehensively outperform its IN counterparts on all backbones in the Cityscapes and ADE20k benchmarks.
+
+System-level comparison. SemCL pretrained backbones are compared with other methods on semantic segmentation tasks in Table 3. On VOC2012, SemCL R50/R101 outperform Deeplabv3+ [7] by 2.37/3.12 points. The Deeplabv3+ [7] X-71 is 0.55 points behind SemCL R101 on Cityscapes. For  $\mathrm{CP}^2$  [51] ViT-S/16, SemCL Swin-S significantly outperforms it on all benchmarks by 4.97, 4.53 and 6.71 points. Compared to MAE [21] ViT-B, SemCL Swin-B achieves a gain of 0.58 points. A qualitative comparison between MAE [21] ViT-B and SemCL Swin-B is shown in Figure 4.
+
+# 4.2. Object Detection and Instance Segmentation
+
+Backbone comparison. In the benchmark for object detection and instance segmentation on the test set of VOC07 and validation sets of Cityscapes and COCO 2017, SemCL
+
+![](images/4464714652d0d386d11cba4348eea09574b3c6db395b35b868060c353a60eee7.jpg)
+
+![](images/3f932164755daac84f3361f1fbec4912e01ed15afe9034f869ae861a85472cbb.jpg)
+
+![](images/1eb08784a2efb5c367b9822502a620995202afc14e03ad7395adfa34a6525eb2.jpg)  
+Figure 4: Qualitative comparison of semantic segmentation between MAE [21] ViT-B (top) and SemCL Swin-B (bottom) on ADE20K semantic segmentation validation set. Both decoders are UPerNet[58].
+
+![](images/f2f480a4cf239efbceb69cbd05d3d16e03ca52ff8605d6ace2cba415ac391135.jpg)
+
+models pretrained on the corresponding SemCL subset are compared with IN pretrained counterparts in Table 4. SemCL comprehensively outperforms its IN counterparts on the Cityscapes and COCO benchmarks. In particular, on the instance segmentation task, SemCL improves by a maximum of 0.5 and 0.3 points on Cityscapes and COCO respectively. However, SemCL performs worse than its IN counterparts in VOC detection, which is considered to be a result of insufficient training samples in the SemCL-VOC dataset and is further investigated in the ablation study (see Appendix A.4).
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">VOC2007 
+AP50</td><td colspan="2">Cityscapes†</td><td colspan="2">COCO</td></tr><tr><td>APbox</td><td>APmask</td><td>APbox</td><td>APmask</td></tr><tr><td>MoCo v3[9] IN-1k</td><td>81.88</td><td>42.7</td><td>37.5</td><td>42.6</td><td>37.9</td></tr><tr><td>SemCL</td><td>81.57(-0.31)</td><td>42.9(+0.2)</td><td>38.0(+0.5)</td><td>42.8(+0.2)</td><td>38.0(+0.1)</td></tr></table>
+
+(a) R50  
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">VOC2007 
+AP50</td><td colspan="2">Cityscapes†</td><td colspan="2">COCO</td></tr><tr><td>APbox</td><td>APmask</td><td>APbox</td><td>APmask</td></tr><tr><td>Swin[37] IN-22k</td><td>86.58</td><td>45.9</td><td>39.8</td><td>49.2</td><td>43.0</td></tr><tr><td>SemCL</td><td>86.53(-0.05)</td><td>46.0(+0.1)</td><td>40.2(+0.4)</td><td>49.0(-0.2)</td><td>43.3(+0.3)</td></tr></table>
+
+(b) Swin-T
+
+Table 4: Results of object detection and instance segmentation on various backbones w/ Cascade Mask R-CNN. All results are from our implementation.  $\dagger$ : Sync BN is disabled due to gradient accumulation.
+
+System-level comparison. Table 5 compares the object detection and instance segmentation results of SemCL backbones with other models. On Cityscapes and COCO, SemCL backbones comprehensively outperform their twinborn MoCo Specifically, SemCL ResNet50 backbones achieve  $+1.7$  over MoCo IG-1B [22] in COCO detection and
+
+<table><tr><td>Backbone</td><td>Method</td><td>VOC2012</td><td>Cityscapes</td><td>ADE20K</td><td>COCO</td></tr><tr><td rowspan="2">R50</td><td>MoCo v3[9] IN-1k†</td><td>79.65</td><td>79.01</td><td>42.71</td><td>40.59</td></tr><tr><td>SemCL</td><td>79.57(-0.08)</td><td>79.13(+0.12)</td><td>42.96(+0.25)</td><td>40.80(+0.21)</td></tr><tr><td rowspan="2">R101</td><td>timm[55] IN-1k</td><td>81.82</td><td>79.87</td><td>44.43</td><td>41.43</td></tr><tr><td>SemCL</td><td>81.97(+0.15)</td><td>80.10(+0.23)</td><td>44.66(+0.23)</td><td>41.58(+0.15)</td></tr><tr><td rowspan="2">Swin-T</td><td>Swin[37] IN-22k</td><td>80.90</td><td>78.67</td><td>44.60</td><td>42.39</td></tr><tr><td>SemCL</td><td>81.73(+0.83)</td><td>78.91(+0.26)</td><td>44.72(+0.12)</td><td>42.53(+0.14)</td></tr><tr><td rowspan="2">Swin-S</td><td>Swin[37] IN-22k</td><td>83.71</td><td>81.25</td><td>47.62</td><td>44.75</td></tr><tr><td>SemCL</td><td>84.47(+0.76)</td><td>81.53(+0.28)</td><td>47.91(+0.29)</td><td>44.57(-0.18)</td></tr><tr><td rowspan="2">Swin-B</td><td>Swin[37] IN-22k</td><td>84.64</td><td>81.61</td><td>48.21</td><td>44.81</td></tr><tr><td>SemCL</td><td>84.67(+0.03)</td><td>81.94(+0.33)</td><td>48.68(+0.47)</td><td>45.47(+0.66)</td></tr></table>
+
+$\dagger$  Although MoCo v3 [9] is originally based on ViT, R50 pretrained models are provided in their repo https://github. com/facebookresearch/moco-v3.
+
+Table 2: SemCL vs. ImageNet pretrained backbones on VOC2012, Cityscapes, ADE20K and COCO semantic segmentation (mIoU). All results are from our implementation.  
+
+<table><tr><td>Method</td><td>Backbone</td><td>mIoU</td></tr><tr><td>Deeplabv3+[7]</td><td>R50</td><td>77.2</td></tr><tr><td>Supervised-IN[22]</td><td>R50</td><td>74.4</td></tr><tr><td>MoCo IG-1B[22]</td><td>R50</td><td>75.5</td></tr><tr><td>BYOL[18]</td><td>R50</td><td>76.3</td></tr><tr><td>SemCL</td><td>R50</td><td>79.57</td></tr><tr><td>Deeplabv3+[7]</td><td>R101</td><td>78.85</td></tr><tr><td>Res2Net[15]</td><td>Res2Net-101</td><td>80.2</td></tr><tr><td>DFN[61]</td><td>R101</td><td>80.60</td></tr><tr><td>SemCL</td><td>R101</td><td>81.97</td></tr><tr><td>CP2[51]</td><td>ViT-S/16</td><td>79.5</td></tr><tr><td>SemCL</td><td>Swin-S</td><td>84.47</td></tr><tr><td>Leopart[70]</td><td>ViT-B/8</td><td>76.3</td></tr><tr><td>SemCL</td><td>Swin-B</td><td>84.67</td></tr></table>
+
+(a) VOC2012
+
+<table><tr><td>Method</td><td>Backbone</td><td>mIoU</td></tr><tr><td>MoCo IN-1M[22]</td><td>R50</td><td>72.5</td></tr><tr><td>MoCo IG-1B[22]</td><td>R50</td><td>73.6</td></tr><tr><td>Supervised-IN[22]</td><td>R50</td><td>74.6</td></tr><tr><td>CP2[51]</td><td>R50</td><td>76.5</td></tr><tr><td>SemCL</td><td>R50</td><td>79.13</td></tr><tr><td>Deeplabv3+[7]</td><td>X-71</td><td>79.55</td></tr><tr><td>SemCL</td><td>R101</td><td>80.10</td></tr><tr><td>Trans4Trans-T[62]</td><td>PVTv2-B1</td><td>78.23</td></tr><tr><td>SemCL</td><td>Swin-T</td><td>78.91</td></tr><tr><td>CP2[51]</td><td>ViT-S/16</td><td>77.0</td></tr><tr><td>Trans4Trans-S[62]</td><td>PVTv2-B2</td><td>80.02</td></tr><tr><td>SemCL</td><td>Swin-S</td><td>81.53</td></tr></table>
+
+(b) Cityscapes
+
+<table><tr><td>Method</td><td>Backbone</td><td>mIoU</td></tr><tr><td>CP2[51]</td><td>R50</td><td>39.2</td></tr><tr><td>PSPNet[66]</td><td>R50</td><td>42.78</td></tr><tr><td>SemCL</td><td>R50</td><td>42.96</td></tr><tr><td>PSPNet[66]</td><td>R101</td><td>43.29</td></tr><tr><td>SemCL</td><td>R101</td><td>44.66</td></tr><tr><td>CP2[51]</td><td>ViT-S/16</td><td>41.2</td></tr><tr><td>Seg-S/16[47]</td><td>ViT-S</td><td>45.37</td></tr><tr><td>SemCL</td><td>Swin-S</td><td>47.91</td></tr><tr><td>Seg-B/16[47]</td><td>Vi-B/16</td><td>48.06</td></tr><tr><td>MAE[21]</td><td>ViT-B</td><td>48.1</td></tr><tr><td>SemCL</td><td>Swin-B</td><td>48.68</td></tr></table>
+
+(c) ADE20K
+
+Table 3: Comparison with previous best results on validation sets of VOC2012, Cityscapes, and ADE20K semantic segmentation (mIoU). Best results are bolded.
+
++5.1/+0.6 in Cityscapes/COCO instance segmentation tasks. For SemCL Swin-T backbones, SemCL achieves +1.2/+0.1 over Focal [60] on COCO detection/instance segmentation. Figure 5 demonstrates a qualitative comparison between Mask2Former [10] and SemCL Swin-T on Cityscapes validation set. With the improved spatial information understanding, SemCL model clearly discriminates (even overlapped) instances in Figures 5a and 5b and successfully classify part of an instance to a defined category as in Figure 5c.
+
+# 4.3. Depth Estimation
+
+Backbone comparison. SemCL pretrained backbones are compared with corresponding IN pretrained ones in Table 6. For cityscapes, SemCL comprehensively outperforms its IN pretrained counterparts. In particular, the ResNet50 backbones improve the RMSE by 0.138 points. For KITTI
+
+and NTUv2, all SemCL backbones show improvements over the IN pretrained ones. The SemCL pretrained SwinL model achieves improvements of 0.001/0.026 points on KITTI AbsRel/RMSE, and the SemCL pretrained ResNet50 model achieves improvements of 0.003/0.005 points on NYUv2.
+
+System-level comparison. We report the comparison of SemCL pretrained backbones with the previous state-of-the-art method Binsformer[35] and other works in Table 7. As for the Cityscapes results shown in Table 7a, SemCL pretrained ResNet50 significantly outperforms SDC [53] by  $-0.089$  and  $-2.218$  points on AbsRel and RMSE respectively. And SD-SSMDE [43] lags behind SemCL by  $0.005 / 3.742$  points. Meanwhile, the test on KITTI (Eigen split [13]) is shown in Table 7b, SemCL pretrained ResNet50 and Swin-T comprehensively outperform the previous SOTA
+
+<table><tr><td>Method</td><td>Backbone</td><td>AP50</td></tr><tr><td>Supervised-IN[22]</td><td>R50</td><td>81.3</td></tr><tr><td>MoCo IN-1M[22]</td><td>R50</td><td>81.5</td></tr><tr><td>MoCo IG-1B[22]</td><td>R50</td><td>82.2</td></tr><tr><td>MoCo v2 800ep[8]</td><td>R50</td><td>82.5</td></tr><tr><td>SemCL</td><td>R50</td><td>81.57</td></tr><tr><td>UP-DETR/300[12]</td><td>*</td><td>80.1</td></tr><tr><td>OW-DETR[19]</td><td>DDETR</td><td>82.1</td></tr><tr><td>SemCL</td><td>Swin-T</td><td>86.53</td></tr></table>
+
+(a) Object det. on VOC2007 test.
+
+<table><tr><td>Method</td><td>Backbone</td><td>APmask</td></tr><tr><td>MoCo IN-1M[22]</td><td>R50</td><td>32.3</td></tr><tr><td>MoCo IG-1B[22]</td><td>R50</td><td>32.9</td></tr><tr><td>Supervised-IN[22]</td><td>R50</td><td>32.9</td></tr><tr><td>Mask2Former[10]</td><td>R50</td><td>37.4</td></tr><tr><td>SemCL</td><td>R50</td><td>38.0</td></tr><tr><td>BoundaryFormer[30]</td><td>*</td><td>38.3</td></tr><tr><td>Mask2Former[10]</td><td>Swin-T</td><td>39.7</td></tr><tr><td>SemCL</td><td>Swin-T</td><td>40.2</td></tr></table>
+
+(b) Instance seg. on Cityscapes val.
+
+<table><tr><td>Method</td><td>Backbone</td><td>APbox</td><td>APmask</td></tr><tr><td>Supervised-IN[22]</td><td>R50</td><td>40.6</td><td>36.8</td></tr><tr><td>MoCo IN-1M[22]</td><td>R50</td><td>40.8</td><td>36.9</td></tr><tr><td>MoCo IG-1B[22]</td><td>R50</td><td>41.1</td><td>37.4</td></tr><tr><td>SemCL</td><td>R50</td><td>42.8</td><td>38.0</td></tr><tr><td>MST[34]</td><td>Swin-T</td><td>42.7</td><td>38.8</td></tr><tr><td>VSA[63]</td><td>Swin-T</td><td>46.9</td><td>42.1</td></tr><tr><td>Focal[60]</td><td>Focal-Base</td><td>47.8</td><td>43.2</td></tr><tr><td>SemCL</td><td>Swin-T</td><td>49.0</td><td>43.3</td></tr></table>
+
+(c) Object det. and instance seg. on COCO val2017.
+
+Table 5: System-level comparison on object detection and instance segmentation. The best results are **bolded**. * claims their backbones are CNN-based, but are considered and compared as transformer-based networks, since transformer blocks are massively used.  
+
+<table><tr><td rowspan="2">Backbone</td><td rowspan="2">Method</td><td colspan="2">Cityscapes</td><td colspan="2">KITTI</td><td colspan="2">NYUv2</td></tr><tr><td>AbsRel</td><td>RMSE</td><td>AbsRel</td><td>RMSE</td><td>AbsRel</td><td>RMSE</td></tr><tr><td rowspan="2">R50</td><td>MoCo v3[9] IN-1k</td><td>0.138</td><td>4.837</td><td>0.059</td><td>2.337</td><td>0.120</td><td>0.405</td></tr><tr><td>SemCL</td><td>0.138</td><td>4.699(-0.138)</td><td>0.058(-0.001)</td><td>2.337</td><td>0.117(-0.003)</td><td>0.400(-0.005)</td></tr><tr><td rowspan="2">Swin-T</td><td>Swin[37] IN-22k</td><td>0.134</td><td>4.643</td><td>0.057</td><td>2.247</td><td>0.114</td><td>0.386</td></tr><tr><td>SemCL</td><td>0.133(-0.001)</td><td>4.577(-0.066)</td><td>0.057</td><td>2.237(-0.010)</td><td>0.112(-0.002)</td><td>0.381(-0.005)</td></tr><tr><td rowspan="2">Swin-L</td><td>Swin[37] IN-22k</td><td>0.130</td><td>4.536</td><td>0.053</td><td>2.130</td><td>0.095</td><td>0.329</td></tr><tr><td>SemCL</td><td>0.128(-0.002)</td><td>4.426(-0.110)</td><td>0.052(-0.001)</td><td>2.104(-0.026)</td><td>0.094(-0.001)</td><td>0.329</td></tr></table>
+
+Table 6: SemCL vs. ImageNet pretraining on Cityscapes, KITTI and NYUv2 depth estimation (lower is better). All results are from our implementation.
+
+![](images/4f828fa5596d72bf9b809e81a5743ab171f28b17213a00bec4ae5cf72cb1236d.jpg)
+
+![](images/9eebcdaf5d930380d32a1ce1228b380344b5f5c2f25b0450afc3010ebbb72469.jpg)
+
+![](images/d14dc188fc07519c4c2d193ac715f644034b4cf0ec9484d690735ac98a5ff5cd.jpg)  
+Figure 5: Qualitative comparison of instance segmentation between Mask2Former [10] (left) and SemCL (right) Swin-T on Cityscapes validation set.
+
+Binsformer. In particular, SemCL ResNet50 achieves a gain of  $-0.003 / - 0.089$  over Binsformer, and SemCL Swin-L is on par with Binsformer Swin-L. On NYUv2 (see Table 7c), the gains of SemCL pretrained ResNet50 are also high with  $-0.010$  AbsRel and  $-0.173$  RMSE. The SemCL
+
+![](images/eb131c095b7ec1c7b169c6f45ea819e017f9722d5d9360e6ab2443f96f6cf025.jpg)  
+Figure 6: Qualitative comparison of depth estimation between Binsformer[35] and SemCL Swin-T on the NYUv2 test set.
+
+Swin-L RMSE is 0.001 points lower than the SOTA Binsformer. Qualitative comparison between Binsformer [35] and SemCL Swin-T is given in Figure 6.
+
+<table><tr><td>Method</td><td>Backbone</td><td>AbsRel</td><td>RMSE</td></tr><tr><td>UpProj[29]</td><td>R50</td><td>0.257[53]</td><td>7.273[53]</td></tr><tr><td>Pad-net[59]</td><td>R50</td><td>0.246[53]</td><td>7.117[53]</td></tr><tr><td>TRL[64]</td><td>R50</td><td>0.234[53]</td><td>7.104[53]</td></tr><tr><td>Pilzer, et al.[45]</td><td>-</td><td>0.440</td><td>5.443</td></tr><tr><td>SDC-Depth[53]</td><td>R50</td><td>0.227</td><td>6.917</td></tr><tr><td>DTS-Depth[27]</td><td>MobileNetv2</td><td>0.167</td><td>7.785</td></tr><tr><td>struct2depth[5]</td><td>-</td><td>0.151</td><td>7.024</td></tr><tr><td>SD-SSMDE[43]</td><td>R50</td><td>0.143</td><td>8.441</td></tr><tr><td>SemCL</td><td>R50</td><td>0.138</td><td>4.699</td></tr></table>
+
+(a) Cityscapes  
+
+<table><tr><td>Method</td><td>Backbone</td><td>AbsRel</td><td>RMSE</td></tr><tr><td>EPCDepth [42]</td><td>R50</td><td>0.091</td><td>4.207</td></tr><tr><td>Binsformer[35]</td><td>R50</td><td>0.061</td><td>2.426</td></tr><tr><td>SemCL</td><td>R50</td><td>0.058</td><td>2.337</td></tr><tr><td>MonoViT[65]</td><td>-</td><td>0.098</td><td>4.333</td></tr><tr><td>Binsformer [35]</td><td>Swin-T</td><td>0.058</td><td>2.286</td></tr><tr><td>SemCL</td><td>Swin-T</td><td>0.057</td><td>2.237</td></tr><tr><td>DepthFormer [33]</td><td>Swin-L +R50CI</td><td>0.052</td><td>2.143</td></tr><tr><td>Binsformer[35]</td><td>Swin-L</td><td>0.052</td><td>2.098</td></tr><tr><td>SemCL</td><td>Swin-L</td><td>0.052</td><td>2.104</td></tr></table>
+
+(b) KITTI  
+
+<table><tr><td>Method</td><td>Backbone</td><td>AbsRel</td><td>RMSE</td></tr><tr><td>Pad-net [59]</td><td>R50</td><td>0.214</td><td>0.792</td></tr><tr><td>TRL [64]</td><td>R50</td><td>0.144</td><td>0.501</td></tr><tr><td>SDC-Depth [53]</td><td>R50</td><td>0.128</td><td>0.497</td></tr><tr><td>UpProj [29]</td><td>R50</td><td>0.127</td><td>0.573</td></tr><tr><td>SemCL</td><td>R50</td><td>0.117</td><td>0.400</td></tr><tr><td>Binsformer [35]</td><td>Swin-T</td><td>0.113</td><td>0.379</td></tr><tr><td>SemCL</td><td>Swin-T</td><td>0.112</td><td>0.381</td></tr><tr><td>Binsformer [35]</td><td>Swin-L</td><td>0.094</td><td>0.330</td></tr><tr><td>SemCL</td><td>Swin-L</td><td>0.094</td><td>0.329</td></tr></table>
+
+(c) NYUv2
+
+Table 7: System-level comparison on depth estimation. Best results (lower is better) are bolded.
+
+# 4.4. Attention Map Visualization
+
+To improve the qualitative understanding of the SemCL representation, we visualize attention maps of Swin-T pretrained on SemCL-VOC and ImageNet using PyTorch library for CAM methods [16] with the Score-CAM method [52]. These samples in Figure 7 selected from VOC2012 val and have not been seen by the pretrained models are characterized by simple centered subjects (Figure 7a), complex scene with occlusion (Figure 7b) and multiple subjects (Figures 7c and 7d).
+
+In Figure 7a, the IN pretrained model failed to focus on the subject, while the SemCL model accurately depicted the area of the main subject. Meanwhile, SemCL performs better in complex scenes containing foreground-background relationships, as shown in Figure 7b. The IN pretrained model pays no attention to an occluded object, but the SemCL model recognizes that the occluded part also belongs to the horse, as shown in Figure 7b. In Figure 7c, the IN model directly ignores the distant person, but the SemCL model not only annotates two people but also focuses more on the nearby one, suggesting that the SemCL model can understand the spatial distribution of subjects in an image. Also, in Figure 7d, the SemCL model successfully recognizes one plant and its pot as one object.
+
+Summary. SemCL pretrained backbones can substantially outperform ImageNet pretrained counterparts on semantic segmentation, object detection, instance segmentation and depth estimation tasks. Using semantically contrastive pairs generated from off-the-shelf datasets, SemCL significantly improves the spatial information understanding of pretrained models. Such an improvement is also promising to benefit more challenging tasks including generative ones [32]. Ablation study concerning paired/unpaired InfoNCE loss, pretraining batch size, training length and dataset scale are given in Appendix A.4.
+
+# 5. Conclusion
+
+We investigate the effectiveness of semantic information in the contrastive learning pretext task on the spatial un
+
+![](images/5226e012f04effbc40caf2ee71320be0dd930e0f6780ee6e035ff270d3c00d18.jpg)
+
+![](images/bc20e95d32622b1d719688fb7165e8bcba5717039984cbeea7d857567e6845de.jpg)
+
+![](images/dde65f9160e7393015f2b97881d5058707c118ec74bc9b84b93e11c0565da30e.jpg)
+
+![](images/7919bb40da1c18c0f38cacca0e4346a87da9f4a3fa4da570a4ea3730d441e554.jpg)  
+Figure 7: Swin-T attention map visualization. Left: raw image. Middle: attention maps of the IN pretrained model. Right: attention maps of the SemCL pretrained model.
+
+![](images/1fc2ba11d96a5b524ff739c0e9eb982a41fd4562c0c91572ad37b10b6e55bb2b.jpg)
+
+![](images/693d5bde387383fc97b44205cca0feb880c9879a5708401aa62593fb1774048e.jpg)
+
+![](images/648068eebffd0c743bfea2d1146cabbd5569dc9fa89dc62364f507530818824e.jpg)
+
+![](images/231350e5e0a7098c25bc15d484f0c3ff1aca3dd2c956250a23babc7618d29816.jpg)
+
+![](images/537e899522c70cbcab38245ca5a0863705d05756a4dcac98dba417070839438f.jpg)
+
+![](images/a6caed9c7280655bf1a27b87c9a3c67be42a401e0db67034708997b7285fbc54.jpg)
+
+![](images/689106e8ce1df92ee8ea95243925d7e07effab326b0a18106a216c5342d7e681.jpg)
+
+![](images/4f2b7f5cdecb8ffc041152283ccb5171172d6cb5b0e25a425fada50073325794.jpg)
+
+derstanding ability of models. Tests are conducted on semantic/instance segmentation, object detection and depth estimation tasks with both ResNets and Swin Transformers, and SemCL pretrained backbones substantially outperform their ImageNet pretrained counterparts. By supporting small batch sizes and fast pretraining, SemCL is a lightweight yet effective approach. In contrastive representation learning, if dedicated pretext task is designed properly, four ounces can move a thousand pounds. We hope that SemCL will inspire more advanced pretext tasks for contrastive learning.
+
+# References
+
+[1] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2  
+[2] Suzanne Becker and Geoffrey E Hinton. Self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355(6356):161-163, 1992. 2  
+[3] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a “siamese” time delay neural network. Advances in neural information processing systems, 6, 1993. 2  
+[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Computer vision and pattern recognition (CVPR), 2018 IEEE conference on. IEEE, 2018. 3  
+[5] Vincent Casser, Soeren Pirk, Reza Mahjourian, and Anelia Angelova. Unsupervised monocular depth and ego-motion learning with structure and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 8  
+[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 3  
+[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2, 5, 6  
+[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 7  
+[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 1, 2, 5, 6, 7  
+[10] 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. 2, 6, 7  
+[11] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3  
+[12] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Updetr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1601-1610, 2021. 7  
+[13] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 6
+
+[14] 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. 3  
+[15] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multiscale backbone architecture. IEEE transactions on pattern analysis and machine intelligence, 43(2):652-662, 2019. 6  
+[16] Jacob Gildenblat and contributors. Pytorch library for cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021.8  
+[17] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: TrainingImagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5  
+[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020. 6  
+[19] Akshita Gupta, Sanath Narayan, KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Ow-detr: Open-world detection transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9235-9244, 2022. 7  
+[20] Bharath Hariharan, Pablo Arbelaez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In International Conference on Computer Vision (ICCV), 2011. 3  
+[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 5, 6  
+[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 3, 5, 6, 7  
+[23] 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. 4  
+[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 2  
+[25] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International Conference on Machine Learning, pages 4182-4192. PMLR, 2020. 2  
+[26] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In International Conference on Learning Representations, 2019. 2
+
+[27] Hatem Ibrahim, Ahmed Salem, and Hyun-Soo Kang. Dts-depth: Real-time single-image depth estimation using depth-to-space image construction. Sensors, 22(5):1914, 2022. 8  
+[28] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2, 3  
+[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 8  
+[30] Justin Lazarow, Weijian Xu, and Zhuowen Tu. Instance segmentation with mask-supervised polygonal boundary transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4382-4391, 2022. 7  
+[31] Phuc H. Le-Khac, Graham Healy, and Alan F. Smeaton. Contrastive representation learning: A framework and review. IEEE Access, 8:193907-193934, 2020. 1, 2  
+[32] Yuheng Li, Yijun Li, Jingwan Lu, Eli Shechtman, Yong Jae Lee, and Krishna Kumar Singh. Contrastive learning for diverse disentangled foreground generation. In European Conference on Computer Vision, pages 334-351. Springer, 2022. 8  
+[33] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. arXiv preprint arXiv:2203.14211, 2022. 8  
+[34] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. Mst: Masked self-supervised transformer for visual representation. Advances in Neural Information Processing Systems, 34:13165-13176, 2021. 7  
+[35] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987, 2022. 2, 6, 7, 8  
+[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 3  
+[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 4, 5, 6, 7  
+[38] Vishnu Suresh Lokhande, Songwong Tasneeyapant, Abhay Venkatesh, Sathya N Ravi, and Vikas Singh. Generating accurate pseudo-labels in semi-supervised learning and avoiding overconfident predictions via hermite polynomial activations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11435-11443, 2020. 2  
+[39] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April
+
+24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017.5  
+[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5  
+[41] Chao Peng, Tete Xiao, Zeming Li, Yuning Jiang, Xiangyu Zhang, Kai Jia, Gang Yu, and Jian Sun. Megdet: A large minibatch object detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6181-6189, 2018. 5  
+[42] Rui Peng, Ronggang Wang, Yawen Lai, Luyang Tang, and Yangang Cai. Excavating the potential capacity of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15560-15569, 2021. 8  
+[43] Andra Petrovai and Sergiu Nedevschi. Exploiting pseudo labels in a self-supervised learning framework for improved monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1578-1588, 2022. 6, 8  
+[44] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11557-11568, 2021. 2  
+[45] Andrea Pilzer, Dan Xu, Mihai Puscas, Elisa Ricci, and Nicu Sebe. Unsupervised adversarial depth estimation using cycled generative networks. In 2018 international conference on 3D vision (3DV), pages 587-595. IEEE, 2018. 8  
+[46] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, volume 11006, pages 369-386. SPIE, 2019. 5  
+[47] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 6  
+[48] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2  
+[49] Aaron Van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv e-prints, pages arXiv-1807, 2018. 1, 2, 3  
+[50] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3  
+[51] Feng Wang, Huiyu Wang, Chen Wei, Alan Yuille, and Wei Shen. Cp 2: Copy-paste contrastive pretraining for semantic segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXX, pages 499-515. Springer, 2022. 5, 6  
+[52] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 8
+
+[53] Lijun Wang, Jianming Zhang, Oliver Wang, Zhe Lin, and Huchuan Lu. Sdc-depth: Semantic divide-and-conquer network for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 541-550, 2020. 6, 8  
+[54] 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. 3  
+[55] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476, 2021. 6  
+[56] Hao Wu and Saurabh Prasad. Semi-supervised deep learning using pseudo labels for hyperspectral image classification. IEEE Transactions on Image Processing, 27(3):1259-1270, 2017. 2  
+[57] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 1, 2, 3  
+[58] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 5  
+[59] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Padnet: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 675-684, 2018. 8  
+[60] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. Advances in Neural Information Processing Systems, 34:30008-30022, 2021. 6, 7  
+[61] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Learning a discriminative feature network for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1857-1866, 2018. 6  
+[62] Jiaming Zhang, Kailun Yang, Angela Constantinescu, Kunyu Peng, Karin Müller, and Rainer Stiefelhagen. Trans4trans: Efficient transformer for transparent object and semantic scene segmentation in real-world navigation assistance. IEEE Transactions on Intelligent Transportation Systems, 23(10):19173-19186, 2022. 6  
+[63] Qiming Zhang, Yufei Xu, Jing Zhang, and Dacheng Tao. Vsa: learning varied-size window attention in vision transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV, pages 466-483. Springer, 2022. 7  
+[64] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Zequn Jie, Xiang Li, and Jian Yang. Joint task-recursive learning for semantic segmentation and depth estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 235–251, 2018. 8  
+[65] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and
+
+Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. arXiv preprint arXiv:2208.03543, 2022. 8  
+[66] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 6  
+[67] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 3  
+[68] 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. 3  
+[69] Chengxu Zhuang, Alex Lin Zhai, and Daniel Yamins. Local aggregation for unsupervised learning of visual embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6002-6012, 2019. 2  
+[70] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022. 6

2023/Semantic Information in Contrastive Learning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:483d0ebdbf1bba339d203bdc5f5cc8bae61ba0a9230ee118850c8a9e38fdb290
+size 863772

2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/706f36ef-c7c5-4ea7-8481-e3a3bb79a94c_content_list.json

@@ -0,0 +1,2069 @@
+[
+    {
+        "type": "text",
+        "text": "Semantic-Aware Dynamic Parameter for Video Inpainting Transformer",
+        "text_level": 1,
+        "bbox": [
+            122,
+            130,
+            844,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Eunhye Lee $^{1*}$ , Jinsu Yoo $^{2*}$ , Yunjeong Yang $^{2}$ , Sungyong Baik $^{2,3}$ , Tae Hyun Kim $^{1\\dagger}$   \n{dldms1345, jinsuyoo, yunjoeongyang, dsybaik, taehyunkim}@hanyang.ac.kr  \n $^{1}$ Dept. of Computer Science,  $^{2}$ Dept. of Artificial Intelligence,  $^{3}$ Dept. of Data Science, Hanyang University",
+        "bbox": [
+            163,
+            179,
+            803,
+            234
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            268,
+            313,
+            284
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent learning-based video inpainting approaches have achieved considerable progress. However, they still cannot fully utilize semantic information within the video frames and predict improper scene layout, failing to restore clear object boundaries for mixed scenes. To mitigate this problem, we introduce a new transformer-based video inpainting technique that can exploit semantic information within the input and considerably improve reconstruction quality. In this study, we use the mixture-of-experts scheme and train multiple experts to handle mixed scenes, including various semantics. We leverage these multiple experts and produce locally (token-wise) different network parameters to achieve semantic-aware inpainting results. Extensive experiments on YouTube-VOS and DAVIS benchmark datasets demonstrate that, compared with existing conventional video inpainting approaches, the proposed method has superior performance in synthesizing visually pleasing videos with much clearer semantic structures and textures.",
+        "bbox": [
+            75,
+            300,
+            472,
+            571
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            588,
+            209,
+            604
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "A constant increase in demand for video content in our daily lives (e.g., YouTube or TikTok) has led to the development of video inpainting methods, which aim to complete a missing region or erase unwanted areas such as watermarks and captions from a given input video. In particular, recent convolutional neural networks (CNNs) for video inpainting tasks have shown promising results [34, 15, 2, 43, 37, 9]. In addition, transformer-based inpainting networks [39, 26, 23, 29] significantly escalate the quality of inpainting results through their large network capacity and superior local/global connectivity based on attention mechanisms.",
+        "bbox": [
+            75,
+            607,
+            468,
+            773
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, conventional approaches are unable to fully utilize the semantic information within the input video or distinguish the class-specific characteristics of objects with different semantics. As a result, they frequently fail to recover proper object structure, texture, and scene layout. Several studies [32, 24, 19, 25] have suggested that utiliz",
+        "bbox": [
+            75,
+            773,
+            470,
+            864
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ing semantic information can lead to better results and produce visually more plausible images when filling missing regions. However, research on video inpainting has yet to explore the potential of incorporating semantic maps.",
+        "bbox": [
+            500,
+            270,
+            890,
+            330
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Video inpainting is a challenging task that requires to render temporally consistent video frames. To address this issue, optical flow has been extensively explored, although it is computationally intensive [15, 37, 9, 23, 41, 40, 13]. In this study, we demonstrate that semantic information can also be used to enforce temporal consistency, and we introduce a new technique that further facilitates the use of semantic cues in a video by dynamically mixing multiple class-specific experts to handle objects with distinct semantics adaptively.",
+        "bbox": [
+            496,
+            334,
+            893,
+            487
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Although previous semantic-guided inpainting networks used the predicted segmentation map as additional input [32] or developed a dedicated semantic-aware inference module [24] to learn semantic-aware parameters, the restoration performance is limited, given the use of shared network parameters for the different semantics. This problem has been addressed in a recent study [25] by using a semantic-aware attention module that enables the features to attend solely to regions with identical semantic labels. However, the computational cost increases proportionally to the number of classes because the attention mechanism is separately carried out for each category.",
+        "bbox": [
+            496,
+            488,
+            893,
+            671
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To mitigate these problems, we propose a novel semantic-aware dynamic parameter selection approach that effectively utilizes the semantic information within input video frames while retaining the number of operations during the inference phase. Specifically, inspired by Cond-Conv [38], we introduce conditionally parameterized linear operations to learn semantic-aware experts, and produce locally varying (dynamic) parameters by leveraging the given semantic cues based on the notion of mixture-of-experts [31, 30, 38]. Our linear operations with dynamically determined parameters replace the standard feed-forward operation within a conventional transformer block. Notably, we can keep the number of parameters for inference by mixing the expert parameters before performing linear operations rather than aggregating the output features cal",
+        "bbox": [
+            496,
+            674,
+            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,
+            805,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Equal contribution.",
+        "bbox": [
+            101,
+            875,
+            228,
+            886
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "† Corresponding author.",
+        "bbox": [
+            101,
+            887,
+            228,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "12949",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ulated by each expert. Moreover, in contrast to the original CondConv [38] that produces conditional, but locally uniform parameters, ours can generate conditional and locally (token-wise) varying parameters, thus improving each token's representation power.",
+        "bbox": [
+            75,
+            90,
+            468,
+            167
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our extensive experiments witness the outstanding performance of the proposed Semantic-Aware Video Inpainting Transformer (SAVIT) in improving video inpainting results. Specifically, SAVIT elevates quantitative performance on conventional video inpainting benchmark datasets (YouTube-VOS [36] and DAVIS [28]) and produces visually more superior results against state-of-the-art methods. We summarize our contribution as follows:",
+        "bbox": [
+            75,
+            167,
+            468,
+            287
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We tackle leveraging semantic information within the given input video frames for the video inpainting task.",
+            "- We introduce a novel semantic-aware video inpainting transformer based on a mixture-of-experts scheme.",
+            "- We propose a semantic-aware dynamic linear operation to exploit local semantic cues effectively.",
+            "- Extensive experiments demonstrate the superiority and efficacy of our method, especially in recovering semantic structures and textures."
+        ],
+        "bbox": [
+            94,
+            292,
+            467,
+            433
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            450,
+            217,
+            465
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Video inpainting. An active research topic in video inpainting is generating a temporally consistent video. Various approaches, including 3D-CNN-based, flow-based, and transformer-based methods, have been proposed to fill in a missing region inside a video plausibly. Despite their ability to extract spatio-temporal information, 3D-CNN methods [34, 2, 3] fail to fully utilize global correspondence due to limited temporal receptive field. Flow-based methods [15, 37, 9, 23, 41, 40, 13] first estimate motion flow in the missing region and use the flow to predict the missing pixels. However, these methods highly rely on pre-trained inpainting networks and often require manual operations [37, 9]. Recently, transformer-based methods [39, 26, 23, 29] have attracted research attention due to their superiority in capturing information across long-ranged frames. Built upon the vision transformer architecture [8], FuseFormer [26] proposed soft tokenization with spatial overlapping to interact information among neighboring tokens. Followup studies have improved the performance with dedicated modules such as focal attention [23] or discrete latent mapping [29]. In our study, we develop our method on top of transformers to capture long-range dependencies and advocate the effectiveness of leveraging semantic information.",
+        "bbox": [
+            75,
+            470,
+            468,
+            834
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Semantic-guided inpainting. Several works have been proposed to utilize the semantic map for inpainting [32, 24, 25], and its related fields [14, 4, 22, 12]. Among them, SPGNet [32] first fills the missing segmentation map, then",
+        "bbox": [
+            75,
+            839,
+            467,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "synthesizes the inpainted image based on the estimated semantic results. Moreover, SGENet [24] introduced iterative feature refinement using semantic map-based normalization [27] to handle more complex holes in mixed scenes, including objects with different semantics. Recent work [25] proposed to calculate semantic-wise attention to learn the corresponding semantic information. However, most of the previous works have developed upon image inpainting, and leveraging semantic cues within a video has yet to be explored due to the lack of annotated data. Our study mitigates the data acquisition problem by gathering pseudo-annotated videos with pre-trained segmentation network [17], and demonstrates that semantic information can elevate video inpainting performance.",
+        "bbox": [
+            496,
+            90,
+            890,
+            301
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Dynamic filter. As an alternative to fixed filter, scene-adaptive dynamic filter networks [11, 7, 42] adapt their parameters based on a given input image and/or features. These methods allow flexibility in the network in treating various inputs with different characteristics. Notably, CondInst [33] dynamically generates instance-aware parameters for instance segmentation. In this study, we adapt the transformer parameters in reasoning each token conditioned to its features and semantic information.",
+        "bbox": [
+            496,
+            306,
+            890,
+            441
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Mixture-of-experts. By combining the results from multiple expert models, the network can further improve its representation power [30, 10, 31]. Instead of increasing the computational cost during the inference proportional to the number of experts, recent works [38, 5] parameterized the mixed experts to combine the knowledge efficiently. Our work is motivated by CondConv [38]; we mix the experts' parameters before the feature calculation. Specifically, we divide the experts' roles by semantic categories, then combine semantic-aware experts' knowledge at each token.",
+        "bbox": [
+            496,
+            446,
+            890,
+            597
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3. Proposed Method",
+        "text_level": 1,
+        "bbox": [
+            500,
+            604,
+            671,
+            619
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we first introduce the annotation procedure to acquire the semantic maps for videos, followed by a detailed description of the proposed framework SAVIT, which leverages semantic information for video inpainting.",
+        "bbox": [
+            496,
+            623,
+            890,
+            684
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1. Semantic data acquisition",
+        "text_level": 1,
+        "bbox": [
+            500,
+            686,
+            733,
+            702
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Unlike previous semantic-aware approaches [32, 25, 24], which simultaneously estimate the segmentation maps for inpainting, we assume that semantic segmentation maps are predictable from the given clean video frames, which include unmasked regions to remove, using off-the-shelf segmentation methods. However, currently available video segmentation datasets, such as DAVIS [28] or CityScape [6], are limited to specific categories and/or particular objects, leading acquiring ground-truth segmentation maps for general video frames difficult. Alternatively, we utilize a fully pre-trained panoptic segmentation network [17] to predict the foreground and background semantics in more general scenes.",
+        "bbox": [
+            496,
+            704,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "12950",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/19747dbd49d052ab8d77d77bccd7236ed356a66ac7a4be1bae95457e866bafa9.jpg",
+        "image_caption": [
+            "(a) Initial segmentation map"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            125,
+            97,
+            354,
+            299
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/586640cd863417072ec09c23f2d6109ec25d0599f18f8ed8b3b60c0160b7e41b.jpg",
+        "image_caption": [
+            "(b) Super-categorize",
+            "Figure 1: Overview of the semantic data preparation: (a) Temporally inconsistent rough prediction using a pre-trained panoptic segmentation model; (b) Grouping raw predictions into corresponding super-categories; (c) Super-categorized segmentation map produces temporally consistent labels."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            393,
+            95,
+            653,
+            294
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/51008e07925610910e6c24fb9b9ab66fcc56ead2769c1b726f357d65046c93b0.jpg",
+        "image_caption": [
+            "temporally consistent",
+            "(c) Super-categorized segmentation map"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            725,
+            137,
+            828,
+            273
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Fig. 1 illustrates the overall semantic data acquisition process used for training and inference. For training, we first apply the panoptic segmentation network [17] to collect estimated segmentation maps for video frames in conventional video inpainting datasets (e.g., YouTube-VOS [36]). The initial segmentation results consist of 53 background and 80 foreground classes [1]. However, these initial segmentation maps with a large number of classes are erroneous and temporally inconsistent, as shown in Fig. 1 (a). Thus, we enforce the temporal consistency of semantics by grouping the predicted labels to the corresponding supercategory of the class. For example, dog, horse, and sheep regions are grouped into a super-category of animal (Fig. 1 (b)). Although we can lose semantic details to some extent by the super-categorization, such constraint helps the training of our semantic-aware inpainting network through the enforced temporal semantic consistency (Fig. 1 (c)).",
+        "bbox": [
+            75,
+            388,
+            468,
+            645
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In our work, we define eight different super-categories based on the texture similarities and the frequencies of categories, which are measured on the entire training dataset. Specifically, we have five super-categories for foreground objects: 'animal,' 'person,' 'vehicle,' 'plants,' and 'sports.' Moreover, we have two super-categories for background according to their texture; 'plain' and 'patterned.' Finally, we group the rest of the remaining sub-categories into 'etc.' During the test phase, we also obtain the semantic maps in the same fashion.",
+        "bbox": [
+            75,
+            647,
+            468,
+            797
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Semantic-aware dynamic transformer",
+        "text_level": 1,
+        "bbox": [
+            76,
+            806,
+            405,
+            821
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To handle diverse scenes with various objects across different classes in videos, we propose a semantic-aware dynamic transformer. In particular, our transformer employs a mixture-of-experts scheme, in which each expert is a single fully connected layer with its parameters reserved for each",
+        "bbox": [
+            75,
+            824,
+            468,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "corresponding semantic class (i.e., super-category) obtained from Section 3.1. Conditioned on the semantic features of each token, we create a token-wise fully connected layer, where its parameters are determined by weighted summing the class-specific parameters, and the weights for the mixture are produced by our semantic router as in the following.",
+        "bbox": [
+            496,
+            388,
+            890,
+            479
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Learning semantic router. To produce locally (i.e., token-wise) varying parameters conditioned on the semantic information, the network requires a routing function that predicts blending weights to mix the class-specific expert parameters. Our router computes routing weights (i.e., mixture coefficients) for each token by:",
+        "bbox": [
+            496,
+            482,
+            890,
+            573
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nr _ {i} = M L P \\left(f _ {i}\\right) + h a r d m a x \\left(s _ {i}\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            583,
+            584,
+            890,
+            599
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $r_i \\in \\mathbb{R}^C$  is a produced mixture weight vector;  $f_i$  is the feature of an input token;  $s_i \\in \\mathbb{R}^C$  is a histogram of class distribution on the segmentation map for the given  $i$ -th token;  $C$  is the number of semantic experts ( $C = 8$  in this work unless stated otherwise); and MLP is a standard multi-layer perceptron (MLP) composed of two fully-connected layers with Leaky ReLU activation and sigmoid output. Our router uses the most dominant semantic label from the input histogram  $s_i$  by using a hard selection strategy (hardmax operation) rather than a softmax result. In addition, we model our router to be lightweight (i.e., two layers) to minimize the overhead, keeping our framework to have similar computational costs to the standard transformers during inference.",
+        "bbox": [
+            496,
+            609,
+            890,
+            821
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Semantic-aware dynamic parameterization. By using the routing weights  $r_i$  generated by the router for the  $i$ -th token with features  $f_i$ , we produce semantic-aware dynamic parameters  $W_i$  for linear (fully-connected) layers of a transformer block. The formulation of semantic-aware dynamic",
+        "bbox": [
+            496,
+            825,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "12951",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/4b3569247df550c85411f6c17597e8cda4bb180c0b91273b3fecaaf60d7fad56.jpg",
+        "image_caption": [
+            "Figure 2: Illustration of SAFFN. With the histogram of semantic categories (e.g., hardmax) and additional linear layers, the semantic router outputs a routing weight for a specific token. Using the routing weight, SAL first acquires token-specific parameters by the weighted sum of experts, then operates the linear transformation of the token with the parameters. Here, we illustrate the operational flow of a single token.  $\\oplus$  indicates element-wise summation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            112,
+            93,
+            416,
+            455
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "layer (SAL) is given by:",
+        "bbox": [
+            76,
+            585,
+            240,
+            599
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {S A L} \\left(f _ {i}\\right) = f _ {i} \\cdot W _ {i}, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            204,
+            609,
+            468,
+            626
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where the mixed expert parameter  $W_{i}$  for token  $f_{i}$  yields:",
+        "bbox": [
+            76,
+            635,
+            455,
+            650
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nW _ {i} = \\sum_ {c = 1} ^ {C} r _ {i} (c) \\cdot w ^ {c}. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            200,
+            657,
+            468,
+            698
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Here,  $w^{c}$  indicates the parameters of each class-specific expert, while  $r_{i}(c)$  is the  $c$ -th element of routing weight vector  $r_{i}$  and means the mixture coefficient for the corresponding expert  $w^{c}$ . Consequently, our semantic-aware dynamic parameterization produces locally varying (token-wise) parameters. For example, a token corresponding to a goat in the input frame can be handled by parameters specialized for animal super-category, and another token on the rock can be dealt with specific parameters for the plain background. Notably, our framework can also handle a token of complex scenes where semantically different objects are coexisting. This is because our router is designed to produce locally varying weights and give high routing weights",
+        "bbox": [
+            75,
+            703,
+            468,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "for experts corresponding to coexisting objects, rather than choosing a single expert.",
+        "bbox": [
+            498,
+            90,
+            890,
+            119
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Finally, we devise a semantic-aware dynamic feedforward network (SAFFN) as depicted in Fig. 2. SAFFN includes two SAL using the dynamically generated semantic-aware parameters as:",
+        "bbox": [
+            498,
+            121,
+            890,
+            181
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\operatorname {S A F F N} \\left(f _ {i}\\right) = \\operatorname {S A L} \\left(\\operatorname {R e L U} \\left(\\operatorname {S A L} \\left(f _ {i}; W _ {i} ^ {l}\\right)\\right); W _ {i} ^ {l + 1}\\right) \\tag {4} \\\\ = \\max  \\left(\\mathbf {0}, f _ {i} \\cdot W _ {i} ^ {l}\\right) \\cdot W _ {i} ^ {l + 1}. \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            517,
+            186,
+            890,
+            227
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In our SAFFN, each SAL is parameterized by our dynamic parameters  $W_{i}^{l}$ , where  $l$  denotes the layer index of the linear transformation, and ReLU activation is included between two layers. Bias parameters are omitted for brevity.",
+        "bbox": [
+            498,
+            233,
+            890,
+            294
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Overall architecture",
+        "text_level": 1,
+        "bbox": [
+            500,
+            296,
+            691,
+            309
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Fig. 3 illustrates the overall pipeline of the proposed video inpainting network SAVIT. The network is composed of encoder, decoder, and multiple transformer blocks between the encoder and the decoder. First, the encoder produces output features by taking  $N$  consecutive video frames with masks and the corresponding semantic segmentation maps as input. We employ the encoder in FuseFormer [26] and additionally concatenate corresponding semantic maps to input. Next, we carry out the tokenization procedure [26], then forward to several transformer blocks to extract powerful features. Finally, the tokenized features are rearranged into the image-like shape, which is then fed into the decoder that predicts  $N$  inpainted video frames as previous works [26, 23].",
+        "bbox": [
+            498,
+            311,
+            890,
+            523
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Note that our semantic-aware mixture-of-experts schemes can be plugged into any conventional transformer-based approach (e.g., FuseFormer [26], E2FGVI [23]), where a conventional transformer block is just replaced with our semantic-aware transformer block. In our experiments, we embed our dynamic transformer within FuseFormer [26]. Specifically, SAVIT contains eight FuseFormer blocks, each of which consists of a multi-head self-attention (MHSA) operation and fusion feed-forward network (F3N). We replace one of the eight FuseFormer blocks with our dynamic transformer block. Precisely, within the F3N in the fourth FuseFormer block, we modify the two standard linear layers with our dynamic ones and add a semantic router.",
+        "bbox": [
+            496,
+            523,
+            890,
+            734
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Semantic-aware dynamic discriminator. To facilitate each expert's learning of the corresponding category's representation and texture details, we employ a semantic-aware discriminator (SAD) that predicts the real/fake score for each token. Fig. 4 illustrates the architecture of SAD, which has similar structures to the encoder of SAVIT, except for the last layer, where a single SAL is added. Notably, we utilize the same mixture coefficients for dynamic parameters for both SAVIT (generator) and SAD (discriminator), encouraging each expert to learn detailed and discriminative category-specific representation.",
+        "bbox": [
+            496,
+            738,
+            890,
+            904
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "12952",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/c92922e1b79655c66e6835ca0662a9d7faa306d7f5c9e69c77aa9099a5715e58.jpg",
+        "image_caption": [
+            "Figure 3: Overall flow of SAVIT. Our method utilizes additional semantic cues with a mixture-of-experts scheme. The transformer parameters are dynamically adapted to the local semantic information for each token region. With these token-wise dynamic parameters, our network produces more precise inpainting results, generating an object's boundary or synthesizing detailed textures."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            106,
+            93,
+            872,
+            296
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/44b6beda0f9f12c1ea8d3d53d535b9efeb1a69caa4c0efb15182ecd802aba7a1.jpg",
+        "image_caption": [
+            "Figure 4: Schematic illustration of SAD. With the specific mixture-of-experts parameters produced by using the same routing weight from the generator, SAD discriminates whether the inpainted results are realistic or not."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            373,
+            455,
+            505
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4. Loss function",
+        "text_level": 1,
+        "bbox": [
+            76,
+            579,
+            217,
+            593
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Auxiliary token classification loss. In the mixture-of-experts scheme, it is crucial that each expert is assigned with distinct role [18, 30]. If our semantic experts have been assigned properly to each corresponding category, the output tokens from the semantic-aware dynamic block should be helpful for classification. We drive collaborative role division across our semantic experts. Specifically, we regularize the learning of experts during training by performing auxiliary classification tasks on the output tokens from the semantic-aware dynamic block. Such token-wise classification is performed by an auxiliary single linear layer classifier. The classifier predicts the semantic labels of the output tokens, as illustrated with the green box in Fig. 3. As a result, each expert can effectively learn the feature representation corresponding to a specific semantic label.",
+        "bbox": [
+            75,
+            597,
+            468,
+            823
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Overall objective function. Our overall objective function  $\\mathcal{L}_{total}$  is the weighted sum of the following four different loss terms:",
+        "bbox": [
+            75,
+            825,
+            468,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p i x}} + \\lambda_ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\lambda_ {\\text {s a d}} \\mathcal {L} _ {\\text {s a d}}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            91,
+            882,
+            468,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $L_{pix}$  is a pixel-wise reconstruction loss,  $\\mathcal{L}_{adv}$  is an adversarial loss,  $\\mathcal{L}_{cls}$  is token classification loss, and  $\\mathcal{L}_{sad}$  is semantic adversarial loss, while  $\\{\\lambda_{adv},\\lambda_{cls},\\lambda_{sad}\\}$  control the weight of associated loss term. Given the predicted video frame  $\\hat{Y}$  and its ground truth counterpart  $Y$ , we employ the L1 pixel-wise reconstruction loss as:",
+        "bbox": [
+            498,
+            369,
+            890,
+            460
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {p i x} = \\left| \\left| \\hat {Y} - Y \\right| \\right| _ {1}. \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            627,
+            468,
+            890,
+            487
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As for the adversarial loss, we train our inpainting network to fool a discriminator  $D$  as follows:",
+        "bbox": [
+            498,
+            497,
+            890,
+            526
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {a d v} = - E _ {\\hat {Y}} [ D (\\hat {Y}) ]. \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            620,
+            536,
+            890,
+            555
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Based on a temporal patch discriminator [26, 23], the discriminator  $D$  is trained to optimize",
+        "bbox": [
+            498,
+            564,
+            890,
+            594
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {D} = E _ {Y} \\left[ \\operatorname {R e L U} (1 - D (Y)) \\right] + E _ {\\hat {Y}} \\left[ \\operatorname {R e L U} (1 + D (\\hat {Y})) \\right]. \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            504,
+            603,
+            890,
+            635
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The token classification loss term is given as",
+        "bbox": [
+            500,
+            636,
+            795,
+            650
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {c l s} = - \\sum_ {i} \\sum_ {c = 1} ^ {C} y _ {i} ^ {c} \\log \\left(\\text {s o f t m a x} \\left(\\hat {y} _ {i}\\right)\\right), \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            560,
+            660,
+            890,
+            700
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\hat{y}_i\\in \\mathbb{R}^C$  is the predicted score from the single-layer classifier conditioned on the output tokens from the semantic-aware dynamic block, while  $y_{i}^{c}$  is the target probability for class  $c$  at the  $i$ -th token. For the ground truth target  $y_{i}^{c}$ , we use the input semantic segmentation map (i.e.,  $y_{i}^{c} = s_{i}(c)$ ). The semantic-aware adversarial loss  $\\mathcal{L}_{sad}$  is similar to  $\\mathcal{L}_{adv}$  in that (8) and (7) are used, but with the semantic-aware discriminator (SAD) instead of  $D$ .",
+        "bbox": [
+            498,
+            710,
+            890,
+            833
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            840,
+            632,
+            856
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we quantitatively and qualitatively demonstrate the effectiveness of SAVIT. Please refer to the supplementary material for more experimental results.",
+        "bbox": [
+            498,
+            858,
+            890,
+            904
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "12953",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/4870b981a1c5e6971b8ce26899f6f126eff11ef3b2c1467013882087aeb7be56.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td colspan=\"4\">YouTube-VOS [36]</td><td colspan=\"4\">DAVIS [28]</td></tr><tr><td>Method</td><td>PSNR↑</td><td>SSIM↑</td><td>VFID↓</td><td>E_warp(%)↓</td><td>PSNR↑</td><td>SSIM↑</td><td>VFID↓</td><td>E_warp(%)↓</td></tr><tr><td>VINet [15]</td><td>29.20</td><td>0.9434</td><td>0.072</td><td>0.1490</td><td>28.96</td><td>0.9411</td><td>0.199</td><td>0.1785</td></tr><tr><td>LGTSM [3]</td><td>29.74</td><td>0.9504</td><td>0.070</td><td>0.1859</td><td>28.57</td><td>0.9409</td><td>0.170</td><td>0.1640</td></tr><tr><td>CAP [21]</td><td>31.58</td><td>0.9607</td><td>0.071</td><td>0.1470</td><td>30.28</td><td>0.9521</td><td>0.182</td><td>0.1533</td></tr><tr><td>STTN [39]</td><td>32.34</td><td>0.9655</td><td>0.053</td><td>0.0907</td><td>30.67</td><td>0.9560</td><td>0.149</td><td>0.1449</td></tr><tr><td>E2FGVI [23]</td><td>33.71</td><td>0.9700</td><td>0.046</td><td>0.0864</td><td>33.01</td><td>0.9721</td><td>0.116</td><td>0.1315</td></tr><tr><td>ISVI [41]</td><td>31.19</td><td>0.9569</td><td>0.053</td><td>0.1173</td><td>32.28</td><td>0.9669</td><td>0.129</td><td>0.1589</td></tr><tr><td>FuseFormer [26]</td><td>33.16</td><td>0.9673</td><td>0.051</td><td>0.0900</td><td>32.54</td><td>0.9700</td><td>0.138</td><td>0.1362</td></tr><tr><td>SAVIT (Ours)</td><td>33.97</td><td>0.9727</td><td>0.043</td><td>0.0436</td><td>33.14</td><td>0.9748</td><td>0.107</td><td>0.0673</td></tr></table>",
+        "bbox": [
+            135,
+            88,
+            831,
+            237
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/47b4c628aef37a12370a5b7acd28a8bcd2ca46da086d404859a1073a7a4fa63b.jpg",
+        "table_caption": [
+            "Table 1: Quantitative comparison of SAVIT against state-of-the-art methods. SAVIT outperforms the existing approaches in terms of various evaluation metrics. Notably, SAVIT significantly elevates the inpainting performance of FuseFormer [26], which is our baseline architecture, witnessing the superiority of leveraging semantic cues within the input video."
+        ],
+        "table_footnote": [
+            "Table 2: Ablation experiments on various network configurations under DAVIS [28] dataset. Starting from the performance of the baseline architecture [26], the inpainting performance gradually increases as the configurations are added since the network can benefit from the semantic cues within a video."
+        ],
+        "table_body": "<table><tr><td></td><td>Configuration</td><td>PSNR↑</td><td>VFID↓</td></tr><tr><td>A</td><td>FuseFormersmall [26]</td><td>30.51</td><td>0.182</td></tr><tr><td>B</td><td>+ super-categorized segmentation map</td><td>30.90</td><td>0.161</td></tr><tr><td>C</td><td>+ SAL</td><td>30.92</td><td>0.164</td></tr><tr><td>D</td><td>+ token classification loss (λcls)</td><td>31.19</td><td>0.165</td></tr><tr><td>E</td><td>+ SAD</td><td>31.01</td><td>0.157</td></tr></table>",
+        "bbox": [
+            94,
+            295,
+            447,
+            381
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Implementation details",
+        "text_level": 1,
+        "bbox": [
+            76,
+            486,
+            290,
+            502
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Datasets and evaluation metrics. We use two conventional video inpainting datasets, YouTube-VOS [36] and DAVIS [28], for our experiments. YouTube-VOS dataset consists of 3471 training, 474 validation, and 508 test videos from mixed scenes. Meanwhile, DAVIS dataset comprises 150 video clips, 60 of which are used for training and the remaining 90 clips are used for evaluation. Following previous works [26, 23], we use only the training videos in the YouTube-VOS for training, while the evaluation is conducted on YouTube-VOS testset and 50 selected clips of DAVIS testset. During training, we create random mask shapes for input masked frames as in [39, 26, 23] and use the method discussed in Section 3.1 to obtain a segmentation map for the entire dataset.",
+        "bbox": [
+            75,
+            503,
+            468,
+            715
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The inpainting performance is evaluated under two standard scenarios: fixed region inpainting and foreground object removal. For the fixed region inpainting, we use free-form fixed masks for each video provided by FuseFormer [26]. As for the foreground object removal, we remove the main objects using frame-by-frame annotations provided by the stakeholders. Notably, we employ a simple k-nearest neighbors algorithm to complete the missing object region in the segmentation maps for the object removal task (please refer to supplementary material for detail).",
+        "bbox": [
+            75,
+            717,
+            468,
+            868
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We compare the results with respect to four evaluation metrics: PSNR, SSIM, temporal warping error  $E_{\\text{warp}}$  [20],",
+        "bbox": [
+            76,
+            869,
+            470,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "and Video-based Fréchet Inception Distance (VFID) [35]. PSNR and SSIM are traditional metrics used to measure the average quality of produced frames. VFID and  $E_{\\text{warp}}$  measure the video's perceptual quality and temporal stability (consistency), respectively, and low values indicate better quality.",
+        "bbox": [
+            496,
+            297,
+            890,
+            388
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Training settings. We follow the training settings of our baseline network FuseFormer [26]. Specifically, we randomly sample five frames from a video, resize the frames to  $240 \\times 432$ , and utilize random horizontal flipping augmentation. The network is trained for  $500k$  iterations with Adam optimizer [16] and a batch size of eight. The initial learning rate is 1e-4 and decreases to 1e-5 after  $400k$  iterations. The coefficients for our loss functions,  $\\lambda_{adv}$ ,  $\\lambda_{cls}$ , and  $\\lambda_{sad}$  are 0.01, 0.01, and 0.001, respectively. We implement our network on the PyTorch framework and use eight NVIDIA RTX A6000 GPUs for training.",
+        "bbox": [
+            496,
+            393,
+            892,
+            559
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Video inpainting results",
+        "text_level": 1,
+        "bbox": [
+            498,
+            561,
+            718,
+            578
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Quantitative results. Table 1 quantitatively compares our video inpainting results on fixed region inpainting setting against seven state-of-the-art video inpainting networks: VINet [15], LGTSM [3], CAP [21], STTN [39], E2FVGI [23], ISVI [41], and FuseFormer [26]. The result shows that our method remarkably elevates the performance of FuseFormer [26], on which we base our architecture, and achieves superior performances on both YouTube-VOS and DAVIS datasets compared to the state-of-the-art methods. Notably, the warping error has also significantly improved, indicating that semantic information benefits in generating temporally consistent video. These results show the effectiveness of employing mixture-of-experts to dynamically leverage the semantic cues within given frames for video inpainting.",
+        "bbox": [
+            496,
+            579,
+            892,
+            806
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Qualitative results. Fig. 5 shows that SAVIT has visually pleasing inpainting results, especially in synthesizing semantic structure, boundary, and texture of objects. Notably, unlike the baseline methods that rely solely on inherent video clues and cannot accurately restore the shape of the human body, SAVIT effectively completes the region",
+        "bbox": [
+            496,
+            809,
+            893,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "12954",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/8f9ef754c0e2f525e324a8415e1ff9fad5cd58bd9f064040d0f75fe35a052a96.jpg",
+        "image_caption": [
+            "Masked Frames"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            138,
+            74,
+            274,
+            133
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2e12573d17e4ed4bbdef05d514d4cfcb748be6c60ec3ea4406cc4f14a89dcc42.jpg",
+        "image_caption": [
+            "STTN"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            276,
+            74,
+            418,
+            133
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/83c9e68628e4dbb172d29abcd1c981678252f40549c178078f8137a4c8933100.jpg",
+        "image_caption": [
+            "FuseFormer"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            418,
+            74,
+            555,
+            136
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/db5af5a1e67cd1cd0acaf227123157af4be6685c972685b42ce5c94e51797cd3.jpg",
+        "image_caption": [
+            "E2FGVI"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            555,
+            74,
+            689,
+            136
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/61ddfec66c64ef64f1e9d49cf480898400a41ac45fb5c6bfe0be325c09473e8b.jpg",
+        "image_caption": [
+            "SAVIT (Ours)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            689,
+            74,
+            828,
+            136
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7c4147a3778a7f13a666cc6763133e5c893c7817b42f7aa996fffa7b08e93034.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            138,
+            135,
+            274,
+            191
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4317c7295e4383c139139d008c49d540460ab19648ef61ed0cf72cbaef892e5e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            276,
+            135,
+            418,
+            194
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/ecaeab738521227e51db4e6f4d7f990f33ba22d70d047337a3000c2d4e66acca.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            418,
+            135,
+            553,
+            194
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/596c988000b2e45b2463901bdaf47f6f65d7aa85adf85a2b3a5baf6a4af84b05.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            555,
+            135,
+            689,
+            194
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d38d9b0ac28599f1352c93ab68bfc66c176a0638a94010869bb1a98aa065e1e3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            689,
+            135,
+            828,
+            195
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/a1a24f9ceac5f5e6855c058b28efb731e8dea5b271b002eeb283b9b5c66c6125.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            138,
+            194,
+            274,
+            252
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7ffc4d195e1f7d7db840bb62f8f3a22634d9d0814194e2b3aaa1e8ca4a05858b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            276,
+            194,
+            418,
+            253
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/37a9502c5f70e0f52997fd5faf314adf0d73e5c8aa12dffbaa726d0e6a2ffca3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            418,
+            194,
+            553,
+            253
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/650781ecd42b506da60b43ac943a79287cd01d68a99b4e566ebc8a09b2b0a2d4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            555,
+            194,
+            689,
+            253
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7df933a9d73d7f4f7bd3ad244b9e0a1deeef2701e8cd1a7628d7a6f83b3c36db.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            689,
+            194,
+            828,
+            253
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/5e557ee2e7efeac777aa2741c8494cb4f211df93eb81c60d5d6a36303eff8b3f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            138,
+            253,
+            274,
+            313
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/dcf1dd3bbcea01e073c1197ba4cbebcd62a0a184871f1a5ca58fdcabbe11c58f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            276,
+            253,
+            418,
+            313
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/42fd4d9da2aab0f05f9d2a04a54016784b64d742ba53a31b6a55936660320420.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            418,
+            253,
+            553,
+            313
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d93db6a3e4bbe625e5f75f2b0655a892517bbc3be5b6146d13c362c99a0414bf.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            555,
+            253,
+            689,
+            313
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/553275ed3a56e3b5667a7bdab89edf9dd46a4fcc0d412a38112262cb2cfdd065.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            689,
+            253,
+            828,
+            314
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/c828aff8235dcc0408eb3b55ccdd947e47ef8ef765be8108fdff4cf5d7ec5bed.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            138,
+            314,
+            274,
+            373
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4c0f95f58d30ef09e84edbffc1a1b192cdbe5f452721b1cf40217a51f2e623af.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            276,
+            314,
+            418,
+            373
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e3c27885d3d128e97fba23afbbce14e479568f22a10023d3c1bfe49bf2c8b4a6.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            418,
+            314,
+            553,
+            373
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/17c81c33b45fe4596f76f1f2309547b48631b4081b48ed8a112cb15b1a55cd37.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            555,
+            314,
+            689,
+            373
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/24999c0908660f75dffaefda4aacc2d52fc0ebbcdfd15e9cf96a9c276f4ac988.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            689,
+            314,
+            828,
+            373
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f0f4cb76128310bec266cfe7d02c9a6186e03a5ac2bcba8fdb10ef43e18e283c.jpg",
+        "image_caption": [
+            "Figure 5: Qualitative comparison of SAVIT with baseline video inpainting networks in fixed region inpainting and object removal. SAVIT more accurately recovers the object's boundaries and appropriate textures."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            138,
+            375,
+            274,
+            433
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/bca9e385492a285e57d66712941e82a4338c0d4129693e942c32479790841a37.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            276,
+            375,
+            418,
+            434
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/9f42f31a2daef94f9b2b42752ca633f7b518305740b6117c0885179112dcfbe2.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            418,
+            375,
+            553,
+            434
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/cb457ca45898752bc9c94b4e07c1966d9b8b20b1cb5d2f9cbd80f4781d8335d0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            555,
+            375,
+            689,
+            434
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/13229ab8c7ddaeef1b4fb360903aaa689533aff721ecdfc4730ce24fae886665.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            689,
+            375,
+            828,
+            434
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e1640d5d2232cd6f429e96210ddf9ca508d1ce50b726c4733789844f3c55c9aa.jpg",
+        "image_caption": [
+            "Figure 6: User study result."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            484,
+            174,
+            556
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/c307646ea73a9204afc13553f891a750ba243675718221088d8c5b4d33b39756.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Configuration</td><td>PSNR†</td></tr><tr><td>hardmax (SAVIT)</td><td>31.01</td></tr><tr><td>softmax</td><td>30.91</td></tr><tr><td>random routing weight</td><td>19.86</td></tr></table>",
+        "bbox": [
+            232,
+            488,
+            436,
+            554
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Effect of the routing weight on the DAVIS [28] dataset.",
+        "bbox": [
+            207,
+            558,
+            464,
+            587
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "by leveraging additional semantic information (third and fourth rows). Furthermore, SAVIT demonstrates the capability of generating more semantically plausible textures, as observed from the detailed textures of synthesized grass and wall (second and fifth rows).",
+        "bbox": [
+            75,
+            603,
+            468,
+            679
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "User study. We conduct a user study on the foreground object removal task, where ground truth target video frames for evaluation are not available, to investigate the real-world applicability of our method using DAVIS [28] testset. We ask 24 raters to choose a better video between two randomly ordered videos, including the inpainted results by FuseFormer [26] and ours, fifteen times per rater. The result presented in Fig. 6 shows that raters prefer ours to the baseline, indicating the impact of exploiting semantic cues in completing missing object regions.",
+        "bbox": [
+            75,
+            686,
+            468,
+            838
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5. Ablation studies",
+        "text_level": 1,
+        "bbox": [
+            76,
+            847,
+            236,
+            863
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We perform ablation experiments based on a smaller FuseFormer network (FuseFormer<sub>small</sub>) by halving the dimension of the hidden layer.",
+        "bbox": [
+            75,
+            867,
+            468,
+            912
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7d79bd840aa5d68ad9ebdacbe7ef6254649b43d37d35a6105892e0055d143ebc.jpg",
+        "image_caption": [
+            "SAVIT w/o SAD",
+            "Figure 7: Effect of semantic-aware discriminator (SAD). The texture of the black swan's body is blurry in the case without SAD."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            489,
+            692,
+            571
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7b4fe8511bf1a6cea21b03c4df82bc973f361a247cdf9aa04169684d0c5aae11.jpg",
+        "image_caption": [
+            "SAVIT w/ SAD"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            488,
+            890,
+            571
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Network configuration. Table 2 compares the inpainting performance for various configurations of SAVIT. First, we observe that the baseline network's performance improves with additional segmentation map inputs (rows A and B). However, a subtle additional performance improvement (0.02dB PSNR gain, rows B and C) is achieved when we naively add our proposed dynamic linear layer to the network. Rather, applying the proposed token-wise classification loss to the objective boosts the performance (0.2dB PSNR gain, row D). Overall performance improvement by applying mixture-of-experts (rows C and D) highlights the importance and effectiveness of employing mixture-of-experts to handle semantics dynamically. Moreover, using a semantic-aware dynamic discriminator (row E) degrades PSNR while improving VFID. However, as shown in Fig. 7, utilizing the semantic-aware discriminator produces more visually plausible results, which solidifies the advantage of",
+        "bbox": [
+            496,
+            643,
+            890,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "12955",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7e6a537c1d41f40c7c2e54af2f8543267c4121a3c6aa7101e40a7c65fc67c471.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            85,
+            181,
+            138
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/683f84253a3d4df672db5955b8591fbc2ca835d10e89c5ec21dfbeccd40c1a70.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            183,
+            85,
+            284,
+            138
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/11ccf9fa0c90e978f58ca19e20a7cac7e10d8d28e023726e7caa73f05c919c7a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            285,
+            85,
+            383,
+            138
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a8d45a3c5ca0fdbd138b0ba804b9beae8318c856b29c33689fde41ac4da28eca.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            385,
+            85,
+            486,
+            138
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/fb3a4e1974f878c4a2d0956f15bd8ac4a1cef61e231d6fe7729a0c353fa4e9b7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            486,
+            85,
+            586,
+            138
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/77a8f6095b22a6baff8c284663afe9ad4981f940925da0c6a9cecb1f2b7d1b2d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            586,
+            85,
+            686,
+            138
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b022bd318f8cd0d92db791a753f0df51156f333fe89c87dbfe32607c873c0515.jpg",
+        "image_caption": [
+            "Figure 8: Visualization of the semantic router results. (1) to (8) are the visualization of routing weights of the corresponding super-category. The routing weights are rearranged to the image-like features. Brighter pixel values indicate higher routing weights."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            145,
+            181,
+            196
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/caf30627110a1a0f1af152c9b326d99547050e2cc717445342f9dec54715d7e5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            183,
+            145,
+            284,
+            196
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e7f834c23e09fd6ebd51fa147a7a1c3a815dd1dd386c76e6d98730eb5e259917.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            285,
+            145,
+            383,
+            196
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/86c02db76eb995e3c627518ee35bf74f1185d927e568fc96194ffdf13e483d9c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            385,
+            145,
+            486,
+            196
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/36285cb7505347e38e657601e9874d0e418bb7a51053a0d0ba0493839ba9fe31.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            486,
+            145,
+            586,
+            196
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ef1a5e631bc5ee88923b3a65cc4799f2042c7af9ec19d28047201a5ca510acca.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            586,
+            145,
+            686,
+            196
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/eb94e0b8cacb880f280f963286a163d8a1646f2f6bcde629380f53f81d7c2ec5.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td># Experts</td><td>PSNR↑</td><td>VFID↓</td></tr><tr><td>4</td><td>30.83</td><td>0.165</td></tr><tr><td>6</td><td>30.90</td><td>0.168</td></tr><tr><td>8</td><td>31.01</td><td>0.157</td></tr></table>",
+        "bbox": [
+            697,
+            88,
+            885,
+            152
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a34b416150bc72a0c10c74a5618c4197e585ae0fd65437606b1f410293f19073.jpg",
+        "image_caption": [
+            "Figure 9: Qualitative comparison on different routing weight strategies."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            253,
+            274,
+            299
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/563d625ad537ac3e52b57ea182197f70af9719ba06347e067b5f16e05b581f56.jpg",
+        "table_caption": [
+            "Table 4: Inpainting results by changing the numbers of experts under the DAVIS [28] dataset. More experts render better inpainting results."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td># blocks</td><td># params.</td><td>PSNR↑</td></tr><tr><td>1</td><td>26.0M</td><td>30.40</td></tr><tr><td>2</td><td>33.1M</td><td>30.45</td></tr></table>",
+        "bbox": [
+            285,
+            255,
+            464,
+            303
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the discriminator in improving perceptual quality.",
+        "bbox": [
+            76,
+            369,
+            405,
+            385
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Visualization of routing weights. Fig. 8 visualizes the routing weights of each expert from the trained semantic router. We rearrange the tokens to image-like features for visualization, and brighter values indicate higher routing weights in the corresponding region. The bright regions among different experts indicate that the experts collaborate during the inpainting procedure rather than simply utilizing a specific expert. We see that routing weights tend to change locally smoothly while the input segmentation map changes abruptly.",
+        "bbox": [
+            75,
+            388,
+            468,
+            541
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Designing choice on semantic router. Table 3 compares different design choices of our semantic router. The performance drop by replacing the trained routing weights (i.e., mixture coefficients) with random weights indicates the discriminative role of trained experts. Moreover, the results of softmax routing strategy show a noticeable performance degradation against hardmax strategy quantitatively. Fig. 9 also shows that softmax results tend to have blurry texture, especially on the edge area. Therefore, Hardmax was chosen for our final model to facilitate role division between experts better, as it enables the learning of distinct parameters by dividing the input for each expert before the router blends the token based on its features.",
+        "bbox": [
+            75,
+            544,
+            468,
+            739
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Impact of the number of experts. Table 4 compares the inpainting performance by changing the number of experts. Notably, as the number of experts (i.e., semantic categories) increases, the network can benefit from more diverse semantic clues in terms of PSNR and VFID values, and we set the number of experts for our final model to eight.",
+        "bbox": [
+            75,
+            744,
+            468,
+            835
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Impact of the number of semantic blocks. Table 5 compares the inpainting performance by applying our semantic module on a different number of blocks. The result indicates that using more semantic blocks produces a slightly",
+        "bbox": [
+            75,
+            839,
+            468,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/65f8596b3ea6a40f58f013b1318484f4198424a7be21fd7b00fc04542280202c.jpg",
+        "table_caption": [
+            "Table 5: Ablation on different number of semantic blocks."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Mask ratio</td><td>PSNR†</td><td>VFID↓</td></tr><tr><td rowspan=\"2\">FuseFormersmall</td><td>30%</td><td>22.57</td><td>0.554</td></tr><tr><td>60%</td><td>18.10</td><td>1.059</td></tr><tr><td rowspan=\"2\">FuseFormersmall + Ours</td><td>30%</td><td>23.57</td><td>0.412</td></tr><tr><td>60%</td><td>19.21</td><td>0.750</td></tr></table>",
+        "bbox": [
+            537,
+            255,
+            849,
+            330
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Comparison on varying mask ratio.",
+        "bbox": [
+            549,
+            333,
+            841,
+            347
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "better inpainting performance. Although we apply our semantic module one time due to hardware limits for our final model, it is expected that the results will be naturally improved by increasing the number of semantic blocks.",
+        "bbox": [
+            498,
+            364,
+            890,
+            425
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Input video with large corruption. Table 6 compares our method against FuseFormer<sub>small</sub> under different corrupted region ratio settings. It is shown that, despite larger masks, our method still outperforms the baseline by a large margin by exploiting the semantic information.",
+        "bbox": [
+            498,
+            439,
+            890,
+            515
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion and Future Works",
+        "text_level": 1,
+        "bbox": [
+            500,
+            529,
+            776,
+            545
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this study, we have explored leveraging semantic information within a given video for transformer-based video inpainting. Accordingly, we propose an effective semantic-aware dynamic transformer with the notion of mixture-of-experts. Our dynamic transformer contains class-specific expert parameters where experts are efficiently combined to perform token-wise dynamic linear operations based on the token feature and its semantic information. The experiments demonstrate the effectiveness of our method, especially in recovering semantic structures. Although we have utilized pseudo labels for our study, we believe more accurate and denser annotated datasets can improve the inpainting performance further since our method relies on the quality of the segmentation label.",
+        "bbox": [
+            496,
+            550,
+            890,
+            762
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            500,
+            782,
+            666,
+            799
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2023-00222776) and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06.",
+        "bbox": [
+            498,
+            810,
+            890,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "12956",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            173,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3",
+            "[2] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2",
+            "[3] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Learnable gated temporal shift module for deep video inpainting. In Proceedings of the British Machine Vision Conference (BMVC), 2019. 2, 6",
+            "[4] Cheng Chen, Jiayin Cai, Yao Hu, Xu Tang, Xinggang Wang, Chun Yuan, Xiang Bai, and Song Bai. Deep interactive video inpainting: An invisibility cloak for harry potter. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2021. 2",
+            "[5] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2",
+            "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2",
+            "[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2",
+            "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2",
+            "[9] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2",
+            "[10] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 1991. 2",
+            "[11] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. In Advances in Neural Information Processing Systems (NeurIPS), 2016. 2",
+            "[12] Jun-Gyu Jin, Jaehyun Bae, Han-gyul Baek, and Sang-hyo Park. Object-ratio-preserving video retargeting framework based on segmentation and inpainting. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 2",
+            "[13] Jaeyeon Kang, Seoung Wug Oh, and Seon Joo Kim. Error compensation framework for flow-guided video inpainting."
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            902
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2",
+            "[14] Bholeshwar Khurana, Soumya Ranjan Dash, Abhishek Bhatia, Aniruddha Mahapatra, Hrituraj Singh, and Kuldeep Kulkarni. Semie: Semantically-aware image extrapolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2",
+            "[15] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6",
+            "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6",
+            "[17] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3",
+            "[18] Xiangtao Kong, Hengyuan Zhao, Yu Qiao, and Chao Dong. Classsr: A general framework to accelerate super-resolution networks by data characteristic. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5",
+            "[19] Avisek Lahiri, Arnav Kumar Jain, Sanskar Agrawal, Pabitra Mitra, and Prabir Kumar Biswas. Prior guided gan based semantic inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1",
+            "[20] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6",
+            "[21] Sungho Lee, Seoul Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 6",
+            "[22] Jiacheng Li, Chang Chen, and Zhiwei Xiong. Contextual outpainting with object-level contrastive learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2",
+            "[23] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6",
+            "[24] Liang Liao, Jing Xiao, Zheng Wang, Chia-Wen Lin, and Shin'ichi Satoh. Guidance and evaluation: Semantic-aware image inpainting for mixed scenes. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2",
+            "[25] Liang Liao, Jing Xiao, Zheng Wang, Chia-Wen Lin, and Shin'ichi Satoh. Image inpainting guided by coherence priors of semantics and textures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2",
+            "[26] Rui Liu, Hanming Deng, Yangyi Huang, Xiaoyu Shi, Lewei Lu, Wenxiu Sun, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fuseformer: Fusing fine-grained information in"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "12957",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "transformers for video inpainting. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2, 4, 5, 6, 7",
+            "[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2",
+            "[28] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7, 8",
+            "[29] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2",
+            "[30] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 2, 5",
+            "[31] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In Proceedings of the International Conference on Learning Representations (ICLR), 2017. 1, 2",
+            "[32] Yuhang Song, Chao Yang, Yeji Shen, Peng Wang, Qin Huang, and C-C Jay Kuo. Spg-net: Segmentation prediction and guidance network for image inpainting. In Proceedings of the British Machine Vision Conference (BMVC), 2018. 1, 2",
+            "[33] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2",
+            "[34] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In Association for the Advancement of Artificial Intelligence (AAAI), 2019. 1, 2",
+            "[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 6",
+            "[36] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2, 3, 6",
+            "[37] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2",
+            "[38] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 2",
+            "[39] Yanhong Zeng, Jianlong Fu, and Hongyang Chao. Learning joint spatial-temporal transformations for video inpainting."
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            902
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 6",
+            "[40] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2",
+            "[41] Kaidong Zhang, Jingjing Fu, and Dong Liu. Inertia-guided flow completion and style fusion for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6",
+            "[42] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2",
+            "[43] Xueyan Zou, Linjie Yang, Ding Liu, and Yong Jae Lee. Progressive temporal feature alignment network for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            345
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "12958",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/706f36ef-c7c5-4ea7-8481-e3a3bb79a94c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f98da5dd607f7198d8257b265b39db0b86beaad76cf43d71e327fff047225a35
+size 6988144

2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/full.md

@@ -0,0 +1,394 @@
+# Semantic-Aware Dynamic Parameter for Video Inpainting Transformer
+
+Eunhye Lee $^{1*}$ , Jinsu Yoo $^{2*}$ , Yunjeong Yang $^{2}$ , Sungyong Baik $^{2,3}$ , Tae Hyun Kim $^{1\dagger}$   
+{dldms1345, jinsuyoo, yunjoeongyang, dsybaik, taehyunkim}@hanyang.ac.kr  
+ $^{1}$ Dept. of Computer Science,  $^{2}$ Dept. of Artificial Intelligence,  $^{3}$ Dept. of Data Science, Hanyang University
+
+# Abstract
+
+Recent learning-based video inpainting approaches have achieved considerable progress. However, they still cannot fully utilize semantic information within the video frames and predict improper scene layout, failing to restore clear object boundaries for mixed scenes. To mitigate this problem, we introduce a new transformer-based video inpainting technique that can exploit semantic information within the input and considerably improve reconstruction quality. In this study, we use the mixture-of-experts scheme and train multiple experts to handle mixed scenes, including various semantics. We leverage these multiple experts and produce locally (token-wise) different network parameters to achieve semantic-aware inpainting results. Extensive experiments on YouTube-VOS and DAVIS benchmark datasets demonstrate that, compared with existing conventional video inpainting approaches, the proposed method has superior performance in synthesizing visually pleasing videos with much clearer semantic structures and textures.
+
+# 1. Introduction
+
+A constant increase in demand for video content in our daily lives (e.g., YouTube or TikTok) has led to the development of video inpainting methods, which aim to complete a missing region or erase unwanted areas such as watermarks and captions from a given input video. In particular, recent convolutional neural networks (CNNs) for video inpainting tasks have shown promising results [34, 15, 2, 43, 37, 9]. In addition, transformer-based inpainting networks [39, 26, 23, 29] significantly escalate the quality of inpainting results through their large network capacity and superior local/global connectivity based on attention mechanisms.
+
+However, conventional approaches are unable to fully utilize the semantic information within the input video or distinguish the class-specific characteristics of objects with different semantics. As a result, they frequently fail to recover proper object structure, texture, and scene layout. Several studies [32, 24, 19, 25] have suggested that utiliz
+
+ing semantic information can lead to better results and produce visually more plausible images when filling missing regions. However, research on video inpainting has yet to explore the potential of incorporating semantic maps.
+
+Video inpainting is a challenging task that requires to render temporally consistent video frames. To address this issue, optical flow has been extensively explored, although it is computationally intensive [15, 37, 9, 23, 41, 40, 13]. In this study, we demonstrate that semantic information can also be used to enforce temporal consistency, and we introduce a new technique that further facilitates the use of semantic cues in a video by dynamically mixing multiple class-specific experts to handle objects with distinct semantics adaptively.
+
+Although previous semantic-guided inpainting networks used the predicted segmentation map as additional input [32] or developed a dedicated semantic-aware inference module [24] to learn semantic-aware parameters, the restoration performance is limited, given the use of shared network parameters for the different semantics. This problem has been addressed in a recent study [25] by using a semantic-aware attention module that enables the features to attend solely to regions with identical semantic labels. However, the computational cost increases proportionally to the number of classes because the attention mechanism is separately carried out for each category.
+
+To mitigate these problems, we propose a novel semantic-aware dynamic parameter selection approach that effectively utilizes the semantic information within input video frames while retaining the number of operations during the inference phase. Specifically, inspired by Cond-Conv [38], we introduce conditionally parameterized linear operations to learn semantic-aware experts, and produce locally varying (dynamic) parameters by leveraging the given semantic cues based on the notion of mixture-of-experts [31, 30, 38]. Our linear operations with dynamically determined parameters replace the standard feed-forward operation within a conventional transformer block. Notably, we can keep the number of parameters for inference by mixing the expert parameters before performing linear operations rather than aggregating the output features cal
+
+ulated by each expert. Moreover, in contrast to the original CondConv [38] that produces conditional, but locally uniform parameters, ours can generate conditional and locally (token-wise) varying parameters, thus improving each token's representation power.
+
+Our extensive experiments witness the outstanding performance of the proposed Semantic-Aware Video Inpainting Transformer (SAVIT) in improving video inpainting results. Specifically, SAVIT elevates quantitative performance on conventional video inpainting benchmark datasets (YouTube-VOS [36] and DAVIS [28]) and produces visually more superior results against state-of-the-art methods. We summarize our contribution as follows:
+
+- We tackle leveraging semantic information within the given input video frames for the video inpainting task.  
+- We introduce a novel semantic-aware video inpainting transformer based on a mixture-of-experts scheme.  
+- We propose a semantic-aware dynamic linear operation to exploit local semantic cues effectively.  
+- Extensive experiments demonstrate the superiority and efficacy of our method, especially in recovering semantic structures and textures.
+
+# 2. Related Work
+
+Video inpainting. An active research topic in video inpainting is generating a temporally consistent video. Various approaches, including 3D-CNN-based, flow-based, and transformer-based methods, have been proposed to fill in a missing region inside a video plausibly. Despite their ability to extract spatio-temporal information, 3D-CNN methods [34, 2, 3] fail to fully utilize global correspondence due to limited temporal receptive field. Flow-based methods [15, 37, 9, 23, 41, 40, 13] first estimate motion flow in the missing region and use the flow to predict the missing pixels. However, these methods highly rely on pre-trained inpainting networks and often require manual operations [37, 9]. Recently, transformer-based methods [39, 26, 23, 29] have attracted research attention due to their superiority in capturing information across long-ranged frames. Built upon the vision transformer architecture [8], FuseFormer [26] proposed soft tokenization with spatial overlapping to interact information among neighboring tokens. Followup studies have improved the performance with dedicated modules such as focal attention [23] or discrete latent mapping [29]. In our study, we develop our method on top of transformers to capture long-range dependencies and advocate the effectiveness of leveraging semantic information.
+
+Semantic-guided inpainting. Several works have been proposed to utilize the semantic map for inpainting [32, 24, 25], and its related fields [14, 4, 22, 12]. Among them, SPGNet [32] first fills the missing segmentation map, then
+
+synthesizes the inpainted image based on the estimated semantic results. Moreover, SGENet [24] introduced iterative feature refinement using semantic map-based normalization [27] to handle more complex holes in mixed scenes, including objects with different semantics. Recent work [25] proposed to calculate semantic-wise attention to learn the corresponding semantic information. However, most of the previous works have developed upon image inpainting, and leveraging semantic cues within a video has yet to be explored due to the lack of annotated data. Our study mitigates the data acquisition problem by gathering pseudo-annotated videos with pre-trained segmentation network [17], and demonstrates that semantic information can elevate video inpainting performance.
+
+Dynamic filter. As an alternative to fixed filter, scene-adaptive dynamic filter networks [11, 7, 42] adapt their parameters based on a given input image and/or features. These methods allow flexibility in the network in treating various inputs with different characteristics. Notably, CondInst [33] dynamically generates instance-aware parameters for instance segmentation. In this study, we adapt the transformer parameters in reasoning each token conditioned to its features and semantic information.
+
+Mixture-of-experts. By combining the results from multiple expert models, the network can further improve its representation power [30, 10, 31]. Instead of increasing the computational cost during the inference proportional to the number of experts, recent works [38, 5] parameterized the mixed experts to combine the knowledge efficiently. Our work is motivated by CondConv [38]; we mix the experts' parameters before the feature calculation. Specifically, we divide the experts' roles by semantic categories, then combine semantic-aware experts' knowledge at each token.
+
+# 3. Proposed Method
+
+In this section, we first introduce the annotation procedure to acquire the semantic maps for videos, followed by a detailed description of the proposed framework SAVIT, which leverages semantic information for video inpainting.
+
+# 3.1. Semantic data acquisition
+
+Unlike previous semantic-aware approaches [32, 25, 24], which simultaneously estimate the segmentation maps for inpainting, we assume that semantic segmentation maps are predictable from the given clean video frames, which include unmasked regions to remove, using off-the-shelf segmentation methods. However, currently available video segmentation datasets, such as DAVIS [28] or CityScape [6], are limited to specific categories and/or particular objects, leading acquiring ground-truth segmentation maps for general video frames difficult. Alternatively, we utilize a fully pre-trained panoptic segmentation network [17] to predict the foreground and background semantics in more general scenes.
+
+![](images/19747dbd49d052ab8d77d77bccd7236ed356a66ac7a4be1bae95457e866bafa9.jpg)  
+(a) Initial segmentation map
+
+![](images/586640cd863417072ec09c23f2d6109ec25d0599f18f8ed8b3b60c0160b7e41b.jpg)  
+(b) Super-categorize  
+Figure 1: Overview of the semantic data preparation: (a) Temporally inconsistent rough prediction using a pre-trained panoptic segmentation model; (b) Grouping raw predictions into corresponding super-categories; (c) Super-categorized segmentation map produces temporally consistent labels.
+
+![](images/51008e07925610910e6c24fb9b9ab66fcc56ead2769c1b726f357d65046c93b0.jpg)  
+temporally consistent  
+(c) Super-categorized segmentation map
+
+Fig. 1 illustrates the overall semantic data acquisition process used for training and inference. For training, we first apply the panoptic segmentation network [17] to collect estimated segmentation maps for video frames in conventional video inpainting datasets (e.g., YouTube-VOS [36]). The initial segmentation results consist of 53 background and 80 foreground classes [1]. However, these initial segmentation maps with a large number of classes are erroneous and temporally inconsistent, as shown in Fig. 1 (a). Thus, we enforce the temporal consistency of semantics by grouping the predicted labels to the corresponding supercategory of the class. For example, dog, horse, and sheep regions are grouped into a super-category of animal (Fig. 1 (b)). Although we can lose semantic details to some extent by the super-categorization, such constraint helps the training of our semantic-aware inpainting network through the enforced temporal semantic consistency (Fig. 1 (c)).
+
+In our work, we define eight different super-categories based on the texture similarities and the frequencies of categories, which are measured on the entire training dataset. Specifically, we have five super-categories for foreground objects: 'animal,' 'person,' 'vehicle,' 'plants,' and 'sports.' Moreover, we have two super-categories for background according to their texture; 'plain' and 'patterned.' Finally, we group the rest of the remaining sub-categories into 'etc.' During the test phase, we also obtain the semantic maps in the same fashion.
+
+# 3.2. Semantic-aware dynamic transformer
+
+To handle diverse scenes with various objects across different classes in videos, we propose a semantic-aware dynamic transformer. In particular, our transformer employs a mixture-of-experts scheme, in which each expert is a single fully connected layer with its parameters reserved for each
+
+corresponding semantic class (i.e., super-category) obtained from Section 3.1. Conditioned on the semantic features of each token, we create a token-wise fully connected layer, where its parameters are determined by weighted summing the class-specific parameters, and the weights for the mixture are produced by our semantic router as in the following.
+
+Learning semantic router. To produce locally (i.e., token-wise) varying parameters conditioned on the semantic information, the network requires a routing function that predicts blending weights to mix the class-specific expert parameters. Our router computes routing weights (i.e., mixture coefficients) for each token by:
+
+$$
+r _ {i} = M L P \left(f _ {i}\right) + h a r d m a x \left(s _ {i}\right), \tag {1}
+$$
+
+where  $r_i \in \mathbb{R}^C$  is a produced mixture weight vector;  $f_i$  is the feature of an input token;  $s_i \in \mathbb{R}^C$  is a histogram of class distribution on the segmentation map for the given  $i$ -th token;  $C$  is the number of semantic experts ( $C = 8$  in this work unless stated otherwise); and MLP is a standard multi-layer perceptron (MLP) composed of two fully-connected layers with Leaky ReLU activation and sigmoid output. Our router uses the most dominant semantic label from the input histogram  $s_i$  by using a hard selection strategy (hardmax operation) rather than a softmax result. In addition, we model our router to be lightweight (i.e., two layers) to minimize the overhead, keeping our framework to have similar computational costs to the standard transformers during inference.
+
+Semantic-aware dynamic parameterization. By using the routing weights  $r_i$  generated by the router for the  $i$ -th token with features  $f_i$ , we produce semantic-aware dynamic parameters  $W_i$  for linear (fully-connected) layers of a transformer block. The formulation of semantic-aware dynamic
+
+![](images/4b3569247df550c85411f6c17597e8cda4bb180c0b91273b3fecaaf60d7fad56.jpg)  
+Figure 2: Illustration of SAFFN. With the histogram of semantic categories (e.g., hardmax) and additional linear layers, the semantic router outputs a routing weight for a specific token. Using the routing weight, SAL first acquires token-specific parameters by the weighted sum of experts, then operates the linear transformation of the token with the parameters. Here, we illustrate the operational flow of a single token.  $\oplus$  indicates element-wise summation.
+
+layer (SAL) is given by:
+
+$$
+\operatorname {S A L} \left(f _ {i}\right) = f _ {i} \cdot W _ {i}, \tag {2}
+$$
+
+where the mixed expert parameter  $W_{i}$  for token  $f_{i}$  yields:
+
+$$
+W _ {i} = \sum_ {c = 1} ^ {C} r _ {i} (c) \cdot w ^ {c}. \tag {3}
+$$
+
+Here,  $w^{c}$  indicates the parameters of each class-specific expert, while  $r_{i}(c)$  is the  $c$ -th element of routing weight vector  $r_{i}$  and means the mixture coefficient for the corresponding expert  $w^{c}$ . Consequently, our semantic-aware dynamic parameterization produces locally varying (token-wise) parameters. For example, a token corresponding to a goat in the input frame can be handled by parameters specialized for animal super-category, and another token on the rock can be dealt with specific parameters for the plain background. Notably, our framework can also handle a token of complex scenes where semantically different objects are coexisting. This is because our router is designed to produce locally varying weights and give high routing weights
+
+for experts corresponding to coexisting objects, rather than choosing a single expert.
+
+Finally, we devise a semantic-aware dynamic feedforward network (SAFFN) as depicted in Fig. 2. SAFFN includes two SAL using the dynamically generated semantic-aware parameters as:
+
+$$
+\begin{array}{l} \operatorname {S A F F N} \left(f _ {i}\right) = \operatorname {S A L} \left(\operatorname {R e L U} \left(\operatorname {S A L} \left(f _ {i}; W _ {i} ^ {l}\right)\right); W _ {i} ^ {l + 1}\right) \tag {4} \\ = \max  \left(\mathbf {0}, f _ {i} \cdot W _ {i} ^ {l}\right) \cdot W _ {i} ^ {l + 1}. \\ \end{array}
+$$
+
+In our SAFFN, each SAL is parameterized by our dynamic parameters  $W_{i}^{l}$ , where  $l$  denotes the layer index of the linear transformation, and ReLU activation is included between two layers. Bias parameters are omitted for brevity.
+
+# 3.3. Overall architecture
+
+Fig. 3 illustrates the overall pipeline of the proposed video inpainting network SAVIT. The network is composed of encoder, decoder, and multiple transformer blocks between the encoder and the decoder. First, the encoder produces output features by taking  $N$  consecutive video frames with masks and the corresponding semantic segmentation maps as input. We employ the encoder in FuseFormer [26] and additionally concatenate corresponding semantic maps to input. Next, we carry out the tokenization procedure [26], then forward to several transformer blocks to extract powerful features. Finally, the tokenized features are rearranged into the image-like shape, which is then fed into the decoder that predicts  $N$  inpainted video frames as previous works [26, 23].
+
+Note that our semantic-aware mixture-of-experts schemes can be plugged into any conventional transformer-based approach (e.g., FuseFormer [26], E2FGVI [23]), where a conventional transformer block is just replaced with our semantic-aware transformer block. In our experiments, we embed our dynamic transformer within FuseFormer [26]. Specifically, SAVIT contains eight FuseFormer blocks, each of which consists of a multi-head self-attention (MHSA) operation and fusion feed-forward network (F3N). We replace one of the eight FuseFormer blocks with our dynamic transformer block. Precisely, within the F3N in the fourth FuseFormer block, we modify the two standard linear layers with our dynamic ones and add a semantic router.
+
+Semantic-aware dynamic discriminator. To facilitate each expert's learning of the corresponding category's representation and texture details, we employ a semantic-aware discriminator (SAD) that predicts the real/fake score for each token. Fig. 4 illustrates the architecture of SAD, which has similar structures to the encoder of SAVIT, except for the last layer, where a single SAL is added. Notably, we utilize the same mixture coefficients for dynamic parameters for both SAVIT (generator) and SAD (discriminator), encouraging each expert to learn detailed and discriminative category-specific representation.
+
+![](images/c92922e1b79655c66e6835ca0662a9d7faa306d7f5c9e69c77aa9099a5715e58.jpg)  
+Figure 3: Overall flow of SAVIT. Our method utilizes additional semantic cues with a mixture-of-experts scheme. The transformer parameters are dynamically adapted to the local semantic information for each token region. With these token-wise dynamic parameters, our network produces more precise inpainting results, generating an object's boundary or synthesizing detailed textures.
+
+![](images/44b6beda0f9f12c1ea8d3d53d535b9efeb1a69caa4c0efb15182ecd802aba7a1.jpg)  
+Figure 4: Schematic illustration of SAD. With the specific mixture-of-experts parameters produced by using the same routing weight from the generator, SAD discriminates whether the inpainted results are realistic or not.
+
+# 3.4. Loss function
+
+Auxiliary token classification loss. In the mixture-of-experts scheme, it is crucial that each expert is assigned with distinct role [18, 30]. If our semantic experts have been assigned properly to each corresponding category, the output tokens from the semantic-aware dynamic block should be helpful for classification. We drive collaborative role division across our semantic experts. Specifically, we regularize the learning of experts during training by performing auxiliary classification tasks on the output tokens from the semantic-aware dynamic block. Such token-wise classification is performed by an auxiliary single linear layer classifier. The classifier predicts the semantic labels of the output tokens, as illustrated with the green box in Fig. 3. As a result, each expert can effectively learn the feature representation corresponding to a specific semantic label.
+
+Overall objective function. Our overall objective function  $\mathcal{L}_{total}$  is the weighted sum of the following four different loss terms:
+
+$$
+\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {p i x}} + \lambda_ {\text {a d v}} \mathcal {L} _ {\text {a d v}} + \lambda_ {\text {c l s}} \mathcal {L} _ {\text {c l s}} + \lambda_ {\text {s a d}} \mathcal {L} _ {\text {s a d}}, \tag {5}
+$$
+
+where  $L_{pix}$  is a pixel-wise reconstruction loss,  $\mathcal{L}_{adv}$  is an adversarial loss,  $\mathcal{L}_{cls}$  is token classification loss, and  $\mathcal{L}_{sad}$  is semantic adversarial loss, while  $\{\lambda_{adv},\lambda_{cls},\lambda_{sad}\}$  control the weight of associated loss term. Given the predicted video frame  $\hat{Y}$  and its ground truth counterpart  $Y$ , we employ the L1 pixel-wise reconstruction loss as:
+
+$$
+\mathcal {L} _ {p i x} = \left| \left| \hat {Y} - Y \right| \right| _ {1}. \tag {6}
+$$
+
+As for the adversarial loss, we train our inpainting network to fool a discriminator  $D$  as follows:
+
+$$
+\mathcal {L} _ {a d v} = - E _ {\hat {Y}} [ D (\hat {Y}) ]. \tag {7}
+$$
+
+Based on a temporal patch discriminator [26, 23], the discriminator  $D$  is trained to optimize
+
+$$
+\mathcal {L} _ {D} = E _ {Y} \left[ \operatorname {R e L U} (1 - D (Y)) \right] + E _ {\hat {Y}} \left[ \operatorname {R e L U} (1 + D (\hat {Y})) \right]. \tag {8}
+$$
+
+The token classification loss term is given as
+
+$$
+\mathcal {L} _ {c l s} = - \sum_ {i} \sum_ {c = 1} ^ {C} y _ {i} ^ {c} \log \left(\text {s o f t m a x} \left(\hat {y} _ {i}\right)\right), \tag {9}
+$$
+
+where  $\hat{y}_i\in \mathbb{R}^C$  is the predicted score from the single-layer classifier conditioned on the output tokens from the semantic-aware dynamic block, while  $y_{i}^{c}$  is the target probability for class  $c$  at the  $i$ -th token. For the ground truth target  $y_{i}^{c}$ , we use the input semantic segmentation map (i.e.,  $y_{i}^{c} = s_{i}(c)$ ). The semantic-aware adversarial loss  $\mathcal{L}_{sad}$  is similar to  $\mathcal{L}_{adv}$  in that (8) and (7) are used, but with the semantic-aware discriminator (SAD) instead of  $D$ .
+
+# 4. Experiments
+
+In this section, we quantitatively and qualitatively demonstrate the effectiveness of SAVIT. Please refer to the supplementary material for more experimental results.
+
+<table><tr><td></td><td colspan="4">YouTube-VOS [36]</td><td colspan="4">DAVIS [28]</td></tr><tr><td>Method</td><td>PSNR↑</td><td>SSIM↑</td><td>VFID↓</td><td>E_warp(%)↓</td><td>PSNR↑</td><td>SSIM↑</td><td>VFID↓</td><td>E_warp(%)↓</td></tr><tr><td>VINet [15]</td><td>29.20</td><td>0.9434</td><td>0.072</td><td>0.1490</td><td>28.96</td><td>0.9411</td><td>0.199</td><td>0.1785</td></tr><tr><td>LGTSM [3]</td><td>29.74</td><td>0.9504</td><td>0.070</td><td>0.1859</td><td>28.57</td><td>0.9409</td><td>0.170</td><td>0.1640</td></tr><tr><td>CAP [21]</td><td>31.58</td><td>0.9607</td><td>0.071</td><td>0.1470</td><td>30.28</td><td>0.9521</td><td>0.182</td><td>0.1533</td></tr><tr><td>STTN [39]</td><td>32.34</td><td>0.9655</td><td>0.053</td><td>0.0907</td><td>30.67</td><td>0.9560</td><td>0.149</td><td>0.1449</td></tr><tr><td>E2FGVI [23]</td><td>33.71</td><td>0.9700</td><td>0.046</td><td>0.0864</td><td>33.01</td><td>0.9721</td><td>0.116</td><td>0.1315</td></tr><tr><td>ISVI [41]</td><td>31.19</td><td>0.9569</td><td>0.053</td><td>0.1173</td><td>32.28</td><td>0.9669</td><td>0.129</td><td>0.1589</td></tr><tr><td>FuseFormer [26]</td><td>33.16</td><td>0.9673</td><td>0.051</td><td>0.0900</td><td>32.54</td><td>0.9700</td><td>0.138</td><td>0.1362</td></tr><tr><td>SAVIT (Ours)</td><td>33.97</td><td>0.9727</td><td>0.043</td><td>0.0436</td><td>33.14</td><td>0.9748</td><td>0.107</td><td>0.0673</td></tr></table>
+
+Table 1: Quantitative comparison of SAVIT against state-of-the-art methods. SAVIT outperforms the existing approaches in terms of various evaluation metrics. Notably, SAVIT significantly elevates the inpainting performance of FuseFormer [26], which is our baseline architecture, witnessing the superiority of leveraging semantic cues within the input video.  
+
+<table><tr><td></td><td>Configuration</td><td>PSNR↑</td><td>VFID↓</td></tr><tr><td>A</td><td>FuseFormersmall [26]</td><td>30.51</td><td>0.182</td></tr><tr><td>B</td><td>+ super-categorized segmentation map</td><td>30.90</td><td>0.161</td></tr><tr><td>C</td><td>+ SAL</td><td>30.92</td><td>0.164</td></tr><tr><td>D</td><td>+ token classification loss (λcls)</td><td>31.19</td><td>0.165</td></tr><tr><td>E</td><td>+ SAD</td><td>31.01</td><td>0.157</td></tr></table>
+
+Table 2: Ablation experiments on various network configurations under DAVIS [28] dataset. Starting from the performance of the baseline architecture [26], the inpainting performance gradually increases as the configurations are added since the network can benefit from the semantic cues within a video.
+
+# 4.1. Implementation details
+
+Datasets and evaluation metrics. We use two conventional video inpainting datasets, YouTube-VOS [36] and DAVIS [28], for our experiments. YouTube-VOS dataset consists of 3471 training, 474 validation, and 508 test videos from mixed scenes. Meanwhile, DAVIS dataset comprises 150 video clips, 60 of which are used for training and the remaining 90 clips are used for evaluation. Following previous works [26, 23], we use only the training videos in the YouTube-VOS for training, while the evaluation is conducted on YouTube-VOS testset and 50 selected clips of DAVIS testset. During training, we create random mask shapes for input masked frames as in [39, 26, 23] and use the method discussed in Section 3.1 to obtain a segmentation map for the entire dataset.
+
+The inpainting performance is evaluated under two standard scenarios: fixed region inpainting and foreground object removal. For the fixed region inpainting, we use free-form fixed masks for each video provided by FuseFormer [26]. As for the foreground object removal, we remove the main objects using frame-by-frame annotations provided by the stakeholders. Notably, we employ a simple k-nearest neighbors algorithm to complete the missing object region in the segmentation maps for the object removal task (please refer to supplementary material for detail).
+
+We compare the results with respect to four evaluation metrics: PSNR, SSIM, temporal warping error  $E_{\text{warp}}$  [20],
+
+and Video-based Fréchet Inception Distance (VFID) [35]. PSNR and SSIM are traditional metrics used to measure the average quality of produced frames. VFID and  $E_{\text{warp}}$  measure the video's perceptual quality and temporal stability (consistency), respectively, and low values indicate better quality.
+
+Training settings. We follow the training settings of our baseline network FuseFormer [26]. Specifically, we randomly sample five frames from a video, resize the frames to  $240 \times 432$ , and utilize random horizontal flipping augmentation. The network is trained for  $500k$  iterations with Adam optimizer [16] and a batch size of eight. The initial learning rate is 1e-4 and decreases to 1e-5 after  $400k$  iterations. The coefficients for our loss functions,  $\lambda_{adv}$ ,  $\lambda_{cls}$ , and  $\lambda_{sad}$  are 0.01, 0.01, and 0.001, respectively. We implement our network on the PyTorch framework and use eight NVIDIA RTX A6000 GPUs for training.
+
+# 4.2. Video inpainting results
+
+Quantitative results. Table 1 quantitatively compares our video inpainting results on fixed region inpainting setting against seven state-of-the-art video inpainting networks: VINet [15], LGTSM [3], CAP [21], STTN [39], E2FVGI [23], ISVI [41], and FuseFormer [26]. The result shows that our method remarkably elevates the performance of FuseFormer [26], on which we base our architecture, and achieves superior performances on both YouTube-VOS and DAVIS datasets compared to the state-of-the-art methods. Notably, the warping error has also significantly improved, indicating that semantic information benefits in generating temporally consistent video. These results show the effectiveness of employing mixture-of-experts to dynamically leverage the semantic cues within given frames for video inpainting.
+
+Qualitative results. Fig. 5 shows that SAVIT has visually pleasing inpainting results, especially in synthesizing semantic structure, boundary, and texture of objects. Notably, unlike the baseline methods that rely solely on inherent video clues and cannot accurately restore the shape of the human body, SAVIT effectively completes the region
+
+![](images/8f9ef754c0e2f525e324a8415e1ff9fad5cd58bd9f064040d0f75fe35a052a96.jpg)  
+Masked Frames
+
+![](images/2e12573d17e4ed4bbdef05d514d4cfcb748be6c60ec3ea4406cc4f14a89dcc42.jpg)  
+STTN
+
+![](images/83c9e68628e4dbb172d29abcd1c981678252f40549c178078f8137a4c8933100.jpg)  
+FuseFormer
+
+![](images/db5af5a1e67cd1cd0acaf227123157af4be6685c972685b42ce5c94e51797cd3.jpg)  
+E2FGVI
+
+![](images/61ddfec66c64ef64f1e9d49cf480898400a41ac45fb5c6bfe0be325c09473e8b.jpg)  
+SAVIT (Ours)
+
+![](images/7c4147a3778a7f13a666cc6763133e5c893c7817b42f7aa996fffa7b08e93034.jpg)
+
+![](images/4317c7295e4383c139139d008c49d540460ab19648ef61ed0cf72cbaef892e5e.jpg)
+
+![](images/ecaeab738521227e51db4e6f4d7f990f33ba22d70d047337a3000c2d4e66acca.jpg)
+
+![](images/596c988000b2e45b2463901bdaf47f6f65d7aa85adf85a2b3a5baf6a4af84b05.jpg)
+
+![](images/d38d9b0ac28599f1352c93ab68bfc66c176a0638a94010869bb1a98aa065e1e3.jpg)
+
+![](images/a1a24f9ceac5f5e6855c058b28efb731e8dea5b271b002eeb283b9b5c66c6125.jpg)
+
+![](images/7ffc4d195e1f7d7db840bb62f8f3a22634d9d0814194e2b3aaa1e8ca4a05858b.jpg)
+
+![](images/37a9502c5f70e0f52997fd5faf314adf0d73e5c8aa12dffbaa726d0e6a2ffca3.jpg)
+
+![](images/650781ecd42b506da60b43ac943a79287cd01d68a99b4e566ebc8a09b2b0a2d4.jpg)
+
+![](images/7df933a9d73d7f4f7bd3ad244b9e0a1deeef2701e8cd1a7628d7a6f83b3c36db.jpg)
+
+![](images/5e557ee2e7efeac777aa2741c8494cb4f211df93eb81c60d5d6a36303eff8b3f.jpg)
+
+![](images/dcf1dd3bbcea01e073c1197ba4cbebcd62a0a184871f1a5ca58fdcabbe11c58f.jpg)
+
+![](images/42fd4d9da2aab0f05f9d2a04a54016784b64d742ba53a31b6a55936660320420.jpg)
+
+![](images/d93db6a3e4bbe625e5f75f2b0655a892517bbc3be5b6146d13c362c99a0414bf.jpg)
+
+![](images/553275ed3a56e3b5667a7bdab89edf9dd46a4fcc0d412a38112262cb2cfdd065.jpg)
+
+![](images/c828aff8235dcc0408eb3b55ccdd947e47ef8ef765be8108fdff4cf5d7ec5bed.jpg)
+
+![](images/4c0f95f58d30ef09e84edbffc1a1b192cdbe5f452721b1cf40217a51f2e623af.jpg)
+
+![](images/e3c27885d3d128e97fba23afbbce14e479568f22a10023d3c1bfe49bf2c8b4a6.jpg)
+
+![](images/17c81c33b45fe4596f76f1f2309547b48631b4081b48ed8a112cb15b1a55cd37.jpg)
+
+![](images/24999c0908660f75dffaefda4aacc2d52fc0ebbcdfd15e9cf96a9c276f4ac988.jpg)
+
+![](images/f0f4cb76128310bec266cfe7d02c9a6186e03a5ac2bcba8fdb10ef43e18e283c.jpg)  
+Figure 5: Qualitative comparison of SAVIT with baseline video inpainting networks in fixed region inpainting and object removal. SAVIT more accurately recovers the object's boundaries and appropriate textures.
+
+![](images/bca9e385492a285e57d66712941e82a4338c0d4129693e942c32479790841a37.jpg)
+
+![](images/9f42f31a2daef94f9b2b42752ca633f7b518305740b6117c0885179112dcfbe2.jpg)
+
+![](images/cb457ca45898752bc9c94b4e07c1966d9b8b20b1cb5d2f9cbd80f4781d8335d0.jpg)
+
+![](images/13229ab8c7ddaeef1b4fb360903aaa689533aff721ecdfc4730ce24fae886665.jpg)
+
+![](images/e1640d5d2232cd6f429e96210ddf9ca508d1ce50b726c4733789844f3c55c9aa.jpg)  
+Figure 6: User study result.
+
+<table><tr><td>Configuration</td><td>PSNR†</td></tr><tr><td>hardmax (SAVIT)</td><td>31.01</td></tr><tr><td>softmax</td><td>30.91</td></tr><tr><td>random routing weight</td><td>19.86</td></tr></table>
+
+Table 3: Effect of the routing weight on the DAVIS [28] dataset.
+
+by leveraging additional semantic information (third and fourth rows). Furthermore, SAVIT demonstrates the capability of generating more semantically plausible textures, as observed from the detailed textures of synthesized grass and wall (second and fifth rows).
+
+User study. We conduct a user study on the foreground object removal task, where ground truth target video frames for evaluation are not available, to investigate the real-world applicability of our method using DAVIS [28] testset. We ask 24 raters to choose a better video between two randomly ordered videos, including the inpainted results by FuseFormer [26] and ours, fifteen times per rater. The result presented in Fig. 6 shows that raters prefer ours to the baseline, indicating the impact of exploiting semantic cues in completing missing object regions.
+
+# 5. Ablation studies
+
+We perform ablation experiments based on a smaller FuseFormer network (FuseFormer<sub>small</sub>) by halving the dimension of the hidden layer.
+
+![](images/7d79bd840aa5d68ad9ebdacbe7ef6254649b43d37d35a6105892e0055d143ebc.jpg)  
+SAVIT w/o SAD  
+Figure 7: Effect of semantic-aware discriminator (SAD). The texture of the black swan's body is blurry in the case without SAD.
+
+![](images/7b4fe8511bf1a6cea21b03c4df82bc973f361a247cdf9aa04169684d0c5aae11.jpg)  
+SAVIT w/ SAD
+
+Network configuration. Table 2 compares the inpainting performance for various configurations of SAVIT. First, we observe that the baseline network's performance improves with additional segmentation map inputs (rows A and B). However, a subtle additional performance improvement (0.02dB PSNR gain, rows B and C) is achieved when we naively add our proposed dynamic linear layer to the network. Rather, applying the proposed token-wise classification loss to the objective boosts the performance (0.2dB PSNR gain, row D). Overall performance improvement by applying mixture-of-experts (rows C and D) highlights the importance and effectiveness of employing mixture-of-experts to handle semantics dynamically. Moreover, using a semantic-aware dynamic discriminator (row E) degrades PSNR while improving VFID. However, as shown in Fig. 7, utilizing the semantic-aware discriminator produces more visually plausible results, which solidifies the advantage of
+
+![](images/7e6a537c1d41f40c7c2e54af2f8543267c4121a3c6aa7101e40a7c65fc67c471.jpg)
+
+![](images/683f84253a3d4df672db5955b8591fbc2ca835d10e89c5ec21dfbeccd40c1a70.jpg)
+
+![](images/11ccf9fa0c90e978f58ca19e20a7cac7e10d8d28e023726e7caa73f05c919c7a.jpg)
+
+![](images/a8d45a3c5ca0fdbd138b0ba804b9beae8318c856b29c33689fde41ac4da28eca.jpg)
+
+![](images/fb3a4e1974f878c4a2d0956f15bd8ac4a1cef61e231d6fe7729a0c353fa4e9b7.jpg)
+
+![](images/77a8f6095b22a6baff8c284663afe9ad4981f940925da0c6a9cecb1f2b7d1b2d.jpg)
+
+![](images/b022bd318f8cd0d92db791a753f0df51156f333fe89c87dbfe32607c873c0515.jpg)  
+Figure 8: Visualization of the semantic router results. (1) to (8) are the visualization of routing weights of the corresponding super-category. The routing weights are rearranged to the image-like features. Brighter pixel values indicate higher routing weights.
+
+![](images/caf30627110a1a0f1af152c9b326d99547050e2cc717445342f9dec54715d7e5.jpg)
+
+![](images/e7f834c23e09fd6ebd51fa147a7a1c3a815dd1dd386c76e6d98730eb5e259917.jpg)
+
+![](images/86c02db76eb995e3c627518ee35bf74f1185d927e568fc96194ffdf13e483d9c.jpg)
+
+![](images/36285cb7505347e38e657601e9874d0e418bb7a51053a0d0ba0493839ba9fe31.jpg)
+
+![](images/ef1a5e631bc5ee88923b3a65cc4799f2042c7af9ec19d28047201a5ca510acca.jpg)
+
+<table><tr><td># Experts</td><td>PSNR↑</td><td>VFID↓</td></tr><tr><td>4</td><td>30.83</td><td>0.165</td></tr><tr><td>6</td><td>30.90</td><td>0.168</td></tr><tr><td>8</td><td>31.01</td><td>0.157</td></tr></table>
+
+![](images/a34b416150bc72a0c10c74a5618c4197e585ae0fd65437606b1f410293f19073.jpg)  
+Figure 9: Qualitative comparison on different routing weight strategies.
+
+Table 4: Inpainting results by changing the numbers of experts under the DAVIS [28] dataset. More experts render better inpainting results.  
+
+<table><tr><td># blocks</td><td># params.</td><td>PSNR↑</td></tr><tr><td>1</td><td>26.0M</td><td>30.40</td></tr><tr><td>2</td><td>33.1M</td><td>30.45</td></tr></table>
+
+the discriminator in improving perceptual quality.
+
+Visualization of routing weights. Fig. 8 visualizes the routing weights of each expert from the trained semantic router. We rearrange the tokens to image-like features for visualization, and brighter values indicate higher routing weights in the corresponding region. The bright regions among different experts indicate that the experts collaborate during the inpainting procedure rather than simply utilizing a specific expert. We see that routing weights tend to change locally smoothly while the input segmentation map changes abruptly.
+
+Designing choice on semantic router. Table 3 compares different design choices of our semantic router. The performance drop by replacing the trained routing weights (i.e., mixture coefficients) with random weights indicates the discriminative role of trained experts. Moreover, the results of softmax routing strategy show a noticeable performance degradation against hardmax strategy quantitatively. Fig. 9 also shows that softmax results tend to have blurry texture, especially on the edge area. Therefore, Hardmax was chosen for our final model to facilitate role division between experts better, as it enables the learning of distinct parameters by dividing the input for each expert before the router blends the token based on its features.
+
+Impact of the number of experts. Table 4 compares the inpainting performance by changing the number of experts. Notably, as the number of experts (i.e., semantic categories) increases, the network can benefit from more diverse semantic clues in terms of PSNR and VFID values, and we set the number of experts for our final model to eight.
+
+Impact of the number of semantic blocks. Table 5 compares the inpainting performance by applying our semantic module on a different number of blocks. The result indicates that using more semantic blocks produces a slightly
+
+Table 5: Ablation on different number of semantic blocks.  
+
+<table><tr><td>Method</td><td>Mask ratio</td><td>PSNR†</td><td>VFID↓</td></tr><tr><td rowspan="2">FuseFormersmall</td><td>30%</td><td>22.57</td><td>0.554</td></tr><tr><td>60%</td><td>18.10</td><td>1.059</td></tr><tr><td rowspan="2">FuseFormersmall + Ours</td><td>30%</td><td>23.57</td><td>0.412</td></tr><tr><td>60%</td><td>19.21</td><td>0.750</td></tr></table>
+
+Table 6: Comparison on varying mask ratio.
+
+better inpainting performance. Although we apply our semantic module one time due to hardware limits for our final model, it is expected that the results will be naturally improved by increasing the number of semantic blocks.
+
+Input video with large corruption. Table 6 compares our method against FuseFormer<sub>small</sub> under different corrupted region ratio settings. It is shown that, despite larger masks, our method still outperforms the baseline by a large margin by exploiting the semantic information.
+
+# 6. Conclusion and Future Works
+
+In this study, we have explored leveraging semantic information within a given video for transformer-based video inpainting. Accordingly, we propose an effective semantic-aware dynamic transformer with the notion of mixture-of-experts. Our dynamic transformer contains class-specific expert parameters where experts are efficiently combined to perform token-wise dynamic linear operations based on the token feature and its semantic information. The experiments demonstrate the effectiveness of our method, especially in recovering semantic structures. Although we have utilized pseudo labels for our study, we believe more accurate and denser annotated datasets can improve the inpainting performance further since our method relies on the quality of the segmentation label.
+
+# Acknowledgements
+
+This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2023-00222776) and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06.
+
+# References
+
+[1] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3  
+[2] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 2  
+[3] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Learnable gated temporal shift module for deep video inpainting. In Proceedings of the British Machine Vision Conference (BMVC), 2019. 2, 6  
+[4] Cheng Chen, Jiayin Cai, Yao Hu, Xu Tang, Xinggang Wang, Chun Yuan, Xiang Bai, and Song Bai. Deep interactive video inpainting: An invisibility cloak for harry potter. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2021. 2  
+[5] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2  
+[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2  
+[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2  
+[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2  
+[9] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2  
+[10] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 1991. 2  
+[11] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. In Advances in Neural Information Processing Systems (NeurIPS), 2016. 2  
+[12] Jun-Gyu Jin, Jaehyun Bae, Han-gyul Baek, and Sang-hyo Park. Object-ratio-preserving video retargeting framework based on segmentation and inpainting. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 2  
+[13] Jaeyeon Kang, Seoung Wug Oh, and Seon Joo Kim. Error compensation framework for flow-guided video inpainting.
+
+In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2  
+[14] Bholeshwar Khurana, Soumya Ranjan Dash, Abhishek Bhatia, Aniruddha Mahapatra, Hrituraj Singh, and Kuldeep Kulkarni. Semie: Semantically-aware image extrapolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2  
+[15] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6  
+[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6  
+[17] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3  
+[18] Xiangtao Kong, Hengyuan Zhao, Yu Qiao, and Chao Dong. Classsr: A general framework to accelerate super-resolution networks by data characteristic. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5  
+[19] Avisek Lahiri, Arnav Kumar Jain, Sanskar Agrawal, Pabitra Mitra, and Prabir Kumar Biswas. Prior guided gan based semantic inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1  
+[20] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 6  
+[21] Sungho Lee, Seoul Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 6  
+[22] Jiacheng Li, Chang Chen, and Zhiwei Xiong. Contextual outpainting with object-level contrastive learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2  
+[23] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 4, 5, 6  
+[24] Liang Liao, Jing Xiao, Zheng Wang, Chia-Wen Lin, and Shin'ichi Satoh. Guidance and evaluation: Semantic-aware image inpainting for mixed scenes. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2  
+[25] Liang Liao, Jing Xiao, Zheng Wang, Chia-Wen Lin, and Shin'ichi Satoh. Image inpainting guided by coherence priors of semantics and textures. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2  
+[26] Rui Liu, Hanming Deng, Yangyi Huang, Xiaoyu Shi, Lewei Lu, Wenxiu Sun, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fuseformer: Fusing fine-grained information in
+
+transformers for video inpainting. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2, 4, 5, 6, 7  
+[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2  
+[28] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7, 8  
+[29] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2  
+[30] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 2, 5  
+[31] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In Proceedings of the International Conference on Learning Representations (ICLR), 2017. 1, 2  
+[32] Yuhang Song, Chao Yang, Yeji Shen, Peng Wang, Qin Huang, and C-C Jay Kuo. Spg-net: Segmentation prediction and guidance network for image inpainting. In Proceedings of the British Machine Vision Conference (BMVC), 2018. 1, 2  
+[33] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2  
+[34] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In Association for the Advancement of Artificial Intelligence (AAAI), 2019. 1, 2  
+[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 6  
+[36] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2, 3, 6  
+[37] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2  
+[38] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 2  
+[39] Yanhong Zeng, Jianlong Fu, and Hongyang Chao. Learning joint spatial-temporal transformations for video inpainting.
+
+In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 6  
+[40] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2  
+[41] Kaidong Zhang, Jingjing Fu, and Dong Liu. Inertia-guided flow completion and style fusion for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6  
+[42] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2  
+[43] Xueyan Zou, Linjie Yang, Ding Liu, and Yong Jae Lee. Progressive temporal feature alignment network for video inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1

2023/Semantic-Aware Dynamic Parameter for Video Inpainting Transformer/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da700231c69591bc0d4d0041aa45734362a3306cf6e95a0576d5362558a1a057
+size 702409

2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/840dd92d-8817-48f0-ac2c-586a3d27151d_content_list.json

@@ -0,0 +1,1773 @@
+[
+    {
+        "type": "text",
+        "text": "Semantic-Aware Implicit Template Learning via Part Deformation Consistency",
+        "text_level": 1,
+        "bbox": [
+            84,
+            130,
+            883,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sihyeon Kim Minseok Joo Jaewon Lee  \nJuyeon Ko Juhan Cha Hyunwoo J. Kim*",
+        "bbox": [
+            287,
+            180,
+            676,
+            217
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Department of Computer Science and Engineering, Korea University",
+        "bbox": [
+            210,
+            227,
+            756,
+            244
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{sh_bs15, wlgkcjf87, 2jlejyu, juyon98, hanchaa, hyunwoojkim}@korea.ac.kr",
+        "bbox": [
+            168,
+            247,
+            800,
+            262
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            297,
+            313,
+            313
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Learning implicit templates as neural fields has recently shown impressive performance in unsupervised shape correspondence. Despite the success, we observe current approaches, which solely rely on geometric information, often learn suboptimal deformation across generic object shapes, which have high structural variability. In this paper, we highlight the importance of part deformation consistency and propose a semantic-aware implicit template learning framework to enable semantically plausible deformation. By leveraging semantic prior from a self-supervised feature extractor, we suggest local conditioning with novel semantic-aware deformation code and deformation consistency regularizations regarding part deformation, global deformation, and global scaling. Our extensive experiments demonstrate the superiority of the proposed method over baselines in various tasks: keypoint transfer, part label transfer, and texture transfer. More interestingly, our framework shows a larger performance gain under more challenging settings. We also provide qualitative analyses to validate the effectiveness of semantic-aware deformation. The code is available at https://github.com/mlvlab/PDC.",
+        "bbox": [
+            73,
+            328,
+            473,
+            648
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            656,
+            209,
+            671
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Template learning is essential for shape analysis since the template, which is a compact representation of arbitrary shapes, provides a good prior for shape fitting so that we can establish dense correspondence and transfer key attributes such as texture or keypoints between shapes. Traditionally, handcrafted templates such as meshes, part elements with primitive shapes (e.g., sphere, cylinder, cone), or parametric models have been widely used for human/animal body [3, 4, 5, 6, 7, 8, 9], human face [10, 11], human hand [12, 13, 14], and generic objects [15, 16, 17]. Defining templates for less complex shapes with consistent topologies and similar structures is relatively straightforward. However, for shapes like generic object shapes with high structural variability or",
+        "bbox": [
+            75,
+            681,
+            470,
+            878
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/b8c7719d3d700afb1d874be71a8a7714a1fb5c826244ecbfeb71ba0467281d7f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            504,
+            297,
+            893,
+            512
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/068252e044fa8ddddc802a3c0d2b98ea36759befa0f5a17a5bcffdd3449bdcb3.jpg",
+        "image_caption": [
+            "Figure 1. Importance of part deformation consistency. Given input shapes with high structural variability, only considering low-level geometry during template learning [1, 2] can lead to suboptimal deformation (e.g., cannot distinguish between seat and arm where two parts are geometrically close while semantically distant: low seat-low arm, high seat-high arm). Ours succeeds in transferring shape attributes by encouraging part deformation consistency with semantic information."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            503,
+            513,
+            893,
+            625
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "non-rigid shapes with large-scale deformations, manually defining suitable templates becomes challenging.",
+        "bbox": [
+            496,
+            750,
+            890,
+            777
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address this challenge, recent works [1, 2, 18, 19] have focused on learning implicit templates based on neural fields [20]. Neural fields, which are known for strong representation power, efficiency, and flexibility, have been successful in multiple 3D vision tasks, e.g., shape/scene reconstruction [21, 22, 23, 24], neural rendering [25, 26, 27], and human digitization [28, 29, 30]. Likewise, neural fields-based template learning have shown superior perfor",
+        "bbox": [
+            496,
+            779,
+            892,
+            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": "*Corresponding author.",
+        "bbox": [
+            94,
+            887,
+            220,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "593",
+        "bbox": [
+            485,
+            945,
+            511,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "mance on dense correspondence between complex shapes. Both [1] and [2] suggest decomposing the implicit representation of shapes into a deformation field and a template field. Here, the template is learned as a continuous function, which is expected to capture shared structures among the given shapes within a category (e.g., airplane, chair), while the deformation depends solely on the conditioned geometry of each shape (i.e., SDF value and surface normal). We observe current approaches often learn suboptimal deformation especially when a shape category has high structural variability like diverse local deformation scales and unseen/missing part structures; see Figure 1. That is, incorporating extra knowledge of part semantics is necessary to handle such cases.",
+        "bbox": [
+            75,
+            90,
+            472,
+            303
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a novel framework to learn semantically plausible deformation by distilling the knowledge of self-supervised feature extractor [31] and employing it as a semantic prior to satisfy part deformation consistency during the template learning process. Part deformation consistency is our inductive bias that the deformation result within the same part should be consistent. By leveraging semantic prior to understand part arrangements of shapes, we suggest local conditioning with semantic-aware deformation code and carefully designed regularizations to encourage semantically corresponding deformation. Semantic-aware deformation code is a point-wise soft assignment of part deformation priors, where part deformation priors are multiple latent codes to encode information on corresponding part deformation. We calculate assigned weights considering the part semantics of each point. In addition, we propose input space regularization and latent space regularization to encourage part-level distances to be close, providing an explicit way to control the deformation field so that the model can learn flexible deformation and successfully learn common structures under diverse shapes. Lastly, we suggest global scale consistency regularization to preserve the volume of the template against large-scale deformations. To validate the effectiveness of our framework, we conduct extensive surrogate experiments reflecting different levels of correspondence such as keypoint transfer, part label transfer, and texture transfer, and demonstrate competitive performance over baselines with qualitative analyses.",
+        "bbox": [
+            75,
+            306,
+            472,
+            744
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions are summarized as follows: ① We propose a new framework that learns global implicit template fields founded on part deformation consistency while understanding part semantics to enable semantically plausible deformation, ② We impose hybrid conditioning by combining global conditioning of geometry and local conditioning of part semantics for flexible deformation, ③ We design novel regularizations to successfully manipulate the template fields, ④ We show the effectiveness of our framework with both qualitative and quantitative experiments.",
+        "bbox": [
+            75,
+            750,
+            472,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Works",
+        "text_level": 1,
+        "bbox": [
+            500,
+            89,
+            650,
+            106
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Template learning for shape correspondence. Defining a template for a set of arbitrary shapes has been studied for a long time. A high-quality template provides a good initialization for shape alignment and various applications in vision and graphics. For shapes with consistent topologies and structures such as faces or bodies of humans and animals, handcrafted templates have been popular [3, 5, 7, 8, 9] and still widely used today. On the other hand, it was more challenging for generic object shapes with inconsistent topology and diverse part structures. Prior works suggest to predefine or learn local templates from data as in primitive parts [15] or part elements [32, 16, 33, 17, 34] to handle the inconsistent structure. Recent works have tried to learn a canonical shape of generic objects through implicit template field [1, 2, 35, 36], which are proven to be effective for representing shapes. However, existing approaches mostly lack explicit consideration of semantic knowledge, which enables semantically plausible deformation so that the learned template captures a common structure even with the various structured shapes.",
+        "bbox": [
+            496,
+            130,
+            893,
+            434
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Neural fields for 3D representations. Recent studies have shown that representing 3D shapes or scenes as implicit representations with a neural network has several advantages: powerful representation, flexibility, and memory efficiency. These are called neural fields [20]. For instance, Occupancy networks [22] proposed a method to represent 3D geometry as a continuous occupancy function. DeepSDF [23] learns 3D complex surfaces with a neural network of auto-decoder framework for 3D modeling. Further, implicit representation has been introduced to various applications [37, 38]. To leverage the advantages, many applications based on neural fields have emerged such as neural rendering [25, 26, 27], human digitization [28, 29, 30], and generative modeling [39, 40, 41, 42]. They have also been applied to learn the template of complex shapes as mentioned above.",
+        "bbox": [
+            496,
+            446,
+            893,
+            676
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Utilizing self-supervised segmentation models. Recently, there has been a growing body of research on analyzing complex shapes in a self-supervised manner due to the high cost of annotation for 3D data. Unsupervised semantic segmentation/co-segmentation models like BAE-Net [31], RIM-Net [43], AdaCoSeg [44], and PartSLIP [45] have demonstrated the ability to comprehend part semantics without ground truth labels. For 2D and multi-modal tasks, using large-scale self-supervised segmentation models like DINO [46] as a feature extractor has been widely adopted, including pixelNeRF [47], Feature Field [48], and ClipN-erF [49]. Inspired by these recent works, we utilize pretrained unsupervised part segmentation models to harness part semantic information for template learning.",
+        "bbox": [
+            496,
+            688,
+            893,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "594",
+        "bbox": [
+            485,
+            945,
+            511,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/2c8c62fcf1b166e69f5abf8d598351ffd928a7e37ec0a0d5fc4bebd571ae0dd2.jpg",
+        "image_caption": [
+            "Figure 2. Framework of semantic-aware implicit template learning. To learn semantic-aware implicit template field, (a) we extract part semantics from pretrained feature extractor  $E$  and incorporate knowledge into semantic-aware deformation code  $(\\alpha)$  and (b) part deformation consistency regularizations. With optimizing with the reconstruction loss  $\\mathcal{L}_{\\mathrm{rec}}$  and several regularizers, our framework succeeds to learn semantically plausible deformation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            87,
+            279,
+            242
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/ee7a919c08e58d123d46598646c37fef17ae02e62c65416039063fcee5aef637.jpg",
+        "image_caption": [
+            "(a)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            282,
+            88,
+            555,
+            242
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/f523e7f3c92f502460306e545568713d7bd79a9809d81f4cd9455b4a4c8abcb5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            558,
+            85,
+            684,
+            241
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/801ade73fef454d3add5f9bef7712229db9601be40b7f48707a00ebd1f041aa5.jpg",
+        "image_caption": [
+            "(b)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            696,
+            85,
+            888,
+            242
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            76,
+            337,
+            215,
+            353
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Implicit templates have been studied with neural fields. Given a 3D shape  $P$ , a neural field, which is a neural network  $\\mathcal{F}_{\\theta}$  parameterized by  $\\theta$ , approximates continuous functions such as an occupancy function [21, 22] or a Signed Distance Function (SDF) [23] as",
+        "bbox": [
+            75,
+            362,
+            468,
+            438
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}) = s. \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            217,
+            450,
+            468,
+            465
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Herein,  $\\pmb{x} \\in \\mathbb{R}^3$  is a query point and  $s$  is an occupancy/SDF function value. To represent multiple 3D shapes, for each shape, a shape latent code  $\\pmb{z}_P \\in \\mathbb{R}^d$  is learned. In general, a surface is defined as the zero-level set of a SDF, i.e.,  $\\{x | \\mathcal{F}_{\\theta}(x; z_P) = 0\\}$  as in DeepSDF [23].",
+        "bbox": [
+            76,
+            476,
+            468,
+            551
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The neural field can be further decomposed into an implicit template field  $\\mathcal{T}$  and a deformation field  $\\mathcal{D}[1,2]$ :",
+        "bbox": [
+            75,
+            551,
+            468,
+            582
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}) = \\mathcal {T} _ {\\theta_ {2}} \\left(\\mathcal {D} _ {\\theta_ {1}} (\\boldsymbol {x}; \\boldsymbol {z} _ {P})\\right). \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            165,
+            593,
+            468,
+            609
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The implicit template field  $\\mathcal{T}_{\\theta_2}$  captures a common structure and outputs an occupancy/SDF value  $s$  in the template space.  $\\mathcal{D}_{\\theta_1}$  maps a point  $x$  from a shape-specific space to the common template space by adding a deformation vector  $\\Delta \\pmb{x} \\in \\mathbb{R}^3$ , i.e.,  $\\pmb{x} + \\Delta \\pmb{x}$ . In DIF-Net [2],  $\\mathcal{D}_{\\theta_1}$  additionally predicts a correction value  $\\Delta s \\in \\mathbb{R}$  to capture instance-specific local structures. In this case, the neural field  $\\mathcal{F}_{\\theta}$  is defined as",
+        "bbox": [
+            75,
+            619,
+            468,
+            739
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}) = \\mathcal {T} _ {\\theta_ {2}} (\\boldsymbol {x} + \\Delta \\boldsymbol {x}) + \\Delta s \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            156,
+            750,
+            468,
+            773
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\Delta s, \\Delta \\boldsymbol {x} = \\mathcal {D} _ {\\theta_ {1}} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}).\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            178,
+            768,
+            336,
+            784
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The existing approaches [1, 2] learn the implicit template field by minimizing the distance between the template and the deformed individual shapes. The distance is measured by SDF values and surface normals. Also, several regularizers are imposed such as deformation smoothness (e.g., penalizing the spatial gradient of  $\\mathcal{D}_{\\theta_1}$  and restricting the deformation scale between points). However, the loss functions",
+        "bbox": [
+            75,
+            794,
+            468,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "and regularizers are computed by low-level geometric information. They are not effective enough to handle generic objects with large structural variability. Hence, in this paper, we take part semantics into account for template learning to achieve 'semantically plausible' deformation and dense correspondence.",
+        "bbox": [
+            496,
+            338,
+            890,
+            429
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4. Method",
+        "text_level": 1,
+        "bbox": [
+            500,
+            441,
+            589,
+            457
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We present a semantic-aware implicit template learning framework based on part deformation consistency. In this section, we first delineate the overall architecture of our method equipped with part deformation priors and semantic-aware deformation code. Then, we introduce our deformation consistency regularizers including part deformation consistency regularizations. Lastly, we discuss the training procedure with several additional loss functions.",
+        "bbox": [
+            496,
+            468,
+            890,
+            589
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1. Semantic-aware implicit template field",
+        "text_level": 1,
+        "bbox": [
+            498,
+            597,
+            831,
+            614
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We here present a novel semantic-aware implicit template learning framework. The overall pipeline of the proposed method is shown in Figure 2(a).",
+        "bbox": [
+            496,
+            621,
+            890,
+            666
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Template field and deformation field. We extend the implicit template learning by incorporating semantic information into the deformation field. Specifically, given a query point  $\\pmb{x}$  of a shape  $P$ , the deformation field  $\\mathcal{D}_{\\theta_1}$  predicts a deformation vector  $\\Delta x$  and a correction value  $\\Delta s$  conditioned on a point-wise semantic-aware deformation code (SDC)  $\\alpha(x; P)$  in addition to a shape latent code  $z_P$ . Then, our framework is defined as",
+        "bbox": [
+            496,
+            666,
+            890,
+            785
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {F} _ {\\theta} (\\boldsymbol {x}; P, \\mathcal {E}) := \\mathcal {T} _ {\\theta_ {2}} (\\boldsymbol {x} + \\Delta \\boldsymbol {x}) + \\Delta s, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            568,
+            794,
+            890,
+            818
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\Delta \\boldsymbol {x}, \\Delta s = \\mathcal {D} _ {\\theta_ {1}} (\\boldsymbol {x}; \\boldsymbol {z} _ {P}, \\boldsymbol {\\alpha} (\\boldsymbol {x}; P)),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            598,
+            814,
+            818,
+            830
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{E}$  is a pretrained feature extractor. We assign  $\\alpha (\\pmb {x};P)$  based on a point-wise semantic feature, which is extracted from  $\\mathcal{E}$ . For simplicity, we denote the neural field  $\\mathcal{F}_{\\theta}(\\pmb {x};P,\\mathcal{E})$  by  $\\mathcal{F}_{\\theta}(\\pmb {x})$  and SDC  $\\alpha (\\pmb {x};P)$  by  $\\alpha (\\pmb {x})$ .",
+        "bbox": [
+            500,
+            839,
+            890,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "595",
+        "bbox": [
+            485,
+            945,
+            509,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Feature extractor. In this paper, for feature extraction, we mainly use BAE-Net [31] as an encoder  $\\mathcal{E}$ , which is proven effective in capturing part semantics. Note that we pretrained BAE-Net on selected categories of ShapeNetV2 [50] by self-supervision without any extra labels or data for a fair comparison. The features from BAE-Net are  $k$  dimensional vectors, i.e.,  $\\boldsymbol{o}^P(\\boldsymbol{x}) = \\mathcal{E}(\\boldsymbol{x}, P) \\in \\mathbb{R}^k$ .  $k$  is a hyperparameter, which is closely related to the number of semantic parts. Since the number of semantic parts is unknown in self-supervised learning, we simply chose sufficiently large  $k$ .",
+        "bbox": [
+            75,
+            90,
+            472,
+            256
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Semantic-aware deformation code. Our framework assumes that different semantic parts necessitate different deformations. To achieve this, the proposed method learns a deformation prior for each semantic part. The deformation priors are encoded as latent codes  $e_1, \\dots, e_k \\in \\mathbb{R}^{d'}$ . A point  $x$  is softly assigned to  $k$  semantic parts using the softmax values of semantic features  $o^P(x)$ . Combining the soft assignment with deformation priors, semantic-aware deformation code (SDC) denoted as  $\\alpha(x) \\in \\mathbb{R}^{d'}$  is defined as",
+        "bbox": [
+            75,
+            257,
+            472,
+            393
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol {\\alpha} (\\boldsymbol {x}) = \\sum_ {i = 1} ^ {k} \\frac {\\exp \\left(\\boldsymbol {o} _ {i} ^ {P} (\\boldsymbol {x})\\right) \\boldsymbol {e} _ {i}}{\\sum_ {j = 1} ^ {k} \\exp \\left(\\boldsymbol {o} _ {j} ^ {P} (\\boldsymbol {x})\\right)}, \\forall i \\in \\{1, \\dots , k \\}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            96,
+            416,
+            468,
+            459
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\pmb{o}_i^P(\\pmb{x})$  denotes the  $i$ -th dimension value of the semantic features. The SDC at  $\\pmb{x}$  is computed by a linear combination of deformation priors. We learn the point's probability of belonging to each of the  $k$  parts, which is served as part semantics.",
+        "bbox": [
+            75,
+            468,
+            468,
+            545
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Remarks. For comparison, we also applied hard assignment of part deformation priors for SDC, i.e.,  $\\alpha(\\pmb{x}) = e_{i^*}$ , where  $i^* = \\operatorname{argmax}_i (\\pmb{o}_i^P(\\pmb{x}))$ . However, as points are completely separated with local conditioning of the corresponding part deformation prior  $e_{i^*}$ , the final deformation result tends to break the shape geometry (see Figure 3), leading to performance degradation as shown in Table 1. Thereby, we choose the soft assignment to enable flexible deformation while preserving the shape geometry.",
+        "bbox": [
+            75,
+            545,
+            468,
+            681
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We train the deformation field  $\\mathcal{D}_{\\theta_1}$  and  $\\mathcal{T}_{\\theta_2}$  in an end-to-end fashion. In our experiments,  $\\mathcal{D}_{\\theta_1}$  is a 6-layer MLP conditioned on both global geometry  $z_P$  via hypernetwork  $\\Psi$  [51, 52] and local part semantics  $\\alpha(x)$  via concatenation with input coordinates. The shape latent code  $z_P$  and part deformation priors  $\\{\\pmb{e}_i\\}$  are randomly initialized and jointly optimized with neural fields.  $\\mathcal{T}_{\\theta_2}$  is a 5-layer MLP.",
+        "bbox": [
+            75,
+            681,
+            468,
+            787
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/ed6d6b55ebcc8e221a658f50958afd732547feaaea2845337de00adced3979fd.jpg",
+        "table_caption": [
+            "Table 1. Ablations of different assignment strategies for SDC. We conduct a quantitative comparison between hard and soft assignments with 5-shot part label transfer on ShapeNet chair."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Assignment strategies</td><td>Hard</td><td>Soft</td></tr><tr><td>chair (5-shot PLT)</td><td>77.0</td><td>84.2</td></tr></table>",
+        "bbox": [
+            143,
+            840,
+            403,
+            883
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/3ffcd3cae66c1d892adb41f0722cb7d9eaec6987c5d6eee85e1dd68309f33035.jpg",
+        "image_caption": [
+            "Figure 3. Effects of different assignment strategies for SDC. Hard assignment breaks the shape geometry, while soft assignment preserves the geometry while learning flexible deformations."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            550,
+            102,
+            846,
+            196
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2. Part deformation consistency regularizations",
+        "text_level": 1,
+        "bbox": [
+            498,
+            257,
+            879,
+            273
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In addition to conditioning neural fields, leveraging semantic information, we propose novel part deformation consistency regularizations to encourage semantically plausible deformation. Semantically similar points should be close after deformation in both an input space and a latent space. That is, the deformation of points from the same part should be consistent. Figure 2(b) illustrates the concept of our regularizations. Given a pair of deformed shapes  $P', Q'$ , part deformation consistency regularizations are defined as",
+        "bbox": [
+            496,
+            280,
+            892,
+            417
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p d c - g e o}} = \\sum_ {i = 1} ^ {k} \\left(\\frac {1}{| \\mathcal {P} _ {i} ^ {\\prime} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {P} _ {i} ^ {\\prime}} \\| \\boldsymbol {x} - C (\\boldsymbol {x}, \\mathcal {Q} _ {i} ^ {\\prime}) \\| _ {2} ^ {2} \\right. \\\\ \\left. + \\frac {1}{\\left| \\mathcal {Q} _ {i} ^ {\\prime} \\right|} \\sum_ {\\boldsymbol {y} \\in \\mathcal {Q} _ {i} ^ {\\prime}} \\| \\boldsymbol {y} - C (\\boldsymbol {y}, \\mathcal {P} _ {i} ^ {\\prime}) \\| _ {2} ^ {2}\\right), \\text {a n d} \\tag {6} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            526,
+            440,
+            890,
+            540
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p d c} \\text {- s e m}} = \\sum_ {i = 1} ^ {k} \\left(\\frac {1}{| \\mathcal {P} _ {i} ^ {\\prime} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {P} _ {i} ^ {\\prime}} \\| \\boldsymbol {o} ^ {P ^ {\\prime}} (\\boldsymbol {x}) - \\boldsymbol {o} ^ {Q ^ {\\prime}} \\left(C (\\boldsymbol {x}, \\mathcal {Q} _ {i} ^ {\\prime})\\right) \\| _ {2} ^ {2} \\right. \\\\ \\left. + \\frac {1}{| \\mathcal {Q} _ {i} ^ {\\prime} |} \\sum_ {\\boldsymbol {y} \\in \\mathcal {Q} _ {i} ^ {\\prime}} \\| \\boldsymbol {o} ^ {Q ^ {\\prime}} (\\boldsymbol {y}) - \\boldsymbol {o} ^ {P ^ {\\prime}} \\left(C (\\boldsymbol {y}, \\mathcal {P} _ {i} ^ {\\prime})\\right) \\| _ {2} ^ {2}\\right), \\tag {7} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            506,
+            566,
+            890,
+            680
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $C$  is a correspondence matching function for a point  $\\pmb{x}$  and its nearest point in a set  $\\mathcal{Q}$ , i.e.,  $C(\\pmb{x},\\mathcal{Q}) = \\operatorname*{argmin}_{\\pmb{y}\\in \\mathcal{Q}}\\| \\pmb{x} - \\pmb{y}\\|_2^2$ .  $\\mathcal{P}_i'$  and  $\\mathcal{Q}_i'$  denote sampled point sets of",
+        "bbox": [
+            496,
+            691,
+            893,
+            748
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "$i^{\\mathrm{th}}$  deformed part shapes of  $P^{\\prime}$  and  $Q^{\\prime}$ . They are defined as  $\\mathcal{P}_i^\\prime = \\{\\pmb {x} + \\mathcal{D}_{\\theta_1}(\\pmb {x};\\pmb {z}_P,\\pmb {\\alpha}(\\pmb {x};P))\\mid i = \\operatorname *{argmax}_{k}(\\pmb {o}_k^P (\\pmb {x})),\\pmb {x}\\in$",
+        "bbox": [
+            496,
+            748,
+            890,
+            785
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "$\\Omega_P\\}$  and  $\\mathcal{Q}_i^\\prime = \\{\\pmb {x} + \\mathcal{D}_{\\theta_1}(\\pmb {x};\\pmb {z}_Q,\\pmb {\\alpha}(\\pmb {x};Q))|i =$  argmax  $(\\pmb{o}_k^Q (\\pmb {x})),\\pmb {x}\\in \\Omega_Q\\}$  , where  $\\Omega_P$  and  $\\Omega_{Q}$  are sam",
+        "bbox": [
+            496,
+            785,
+            890,
+            824
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "pled query point sets for shape  $P$  and  $Q$ . Note that each deformed part shape is extracted based on part semantics from original shapes  $P$  and  $Q$ . Here, we explicitly guide the part-level deformation among various shapes to be consistent. Moreover, since we understand part arrangement and",
+        "bbox": [
+            496,
+            824,
+            890,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "596",
+        "bbox": [
+            485,
+            945,
+            511,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/c11deb1e2f34ad9c566244693f1e65d187963d0795697e23acc5908163a4b361.jpg",
+        "image_caption": [
+            "Figure 4. Keypoint transfer result on ShapeNet. We transfer keypoint labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations including results on ScanObjectNN [53], see the supplement."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            90,
+            888,
+            234
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/cafeb9d58e02ba8740aa2249583b33b699f5429902f293ffbc09e7ad539f23d0.jpg",
+        "image_caption": [
+            "(a)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            314,
+            171,
+            359
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/4b36f80cd31eee68e6a5b448f1efa22671b904462eb9211f55d754c17fb51573.jpg",
+        "image_caption": [
+            "(b)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            192,
+            316,
+            302,
+            353
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/f5876dceecdf1c4231267053c9f0d994cbe0c8858bedf5f786589aa659e3d8c1.jpg",
+        "image_caption": [
+            "(c)",
+            "Figure 5. Efficacy of global scaling consistency  $\\mathcal{L}_{\\mathrm{scale}}$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            321,
+            306,
+            455,
+            358
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "only penalize matched parts, our method does not forcibly deform \"unseen/missing match\" cases.",
+        "bbox": [
+            75,
+            402,
+            467,
+            431
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We partly observed spatial distortion when part deformation consistency was too strongly imposed. This is because using external semantics for deformation could break the continuity of geometric fields. In addition, we propose global scale consistency and global deformation consistency to address the concern. For the former one, we disentangle a scaling operation from the entire non-rigid deformation  $\\mathcal{D}_{\\theta_1}$  to calculate the global scale of the learned implicit template field. That is, we want to approximate  $\\mathcal{D}$  with the global scaling of the scaling factor of  $r$ , i.e.,  $\\mathcal{D} \\approx rI$ . Then the problem becomes finding  $r$  from",
+        "bbox": [
+            75,
+            433,
+            468,
+            599
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\min  _ {r} \\sum_ {s} \\sum_ {i} \\| \\boldsymbol {x} _ {i} ^ {s} + \\Delta \\boldsymbol {x} _ {i} ^ {s} - r \\boldsymbol {x} _ {i} ^ {s} \\| _ {2} ^ {2}, \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            156,
+            606,
+            468,
+            637
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $s$  and  $i$  are an index for shape and point, respectively. A closed-form solution of (8) is  $r = \\sum_{s}\\sum_{i}\\frac{\\pmb{x}_i^{s\\top}(\\pmb{x}_i^s + \\Delta\\pmb{x}_i^s)}{\\pmb{x}_i^{s\\top}\\pmb{x}_i^s}$ . See the supplement for proof. Since we want the learned  $\\mathcal{D}$  to be insensitive to global scale operation, i.e., the implicit template to keep its scale by the average scale of input shapes, we penalize the expected  $\\mathbb{E}[r]$  over input shapes  $P_{1}\\dots P_{N}$  in a single batch as:",
+        "bbox": [
+            75,
+            643,
+            468,
+            762
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {s c a l e}} = | \\mathbb {E} [ r ] - 1 | \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            207,
+            770,
+            468,
+            787
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Figure 5 illuminates the effect of  $\\mathcal{L}_{\\mathrm{scale}}$ . For the same input shapes, Figure 5(a) is a car template learned without  $\\mathcal{L}_{\\mathrm{scale}}$ , Figure 5(b) is a car template learned with  $\\mathcal{L}_{\\mathrm{scale}}$ , and Figure 5(c) is the result of placing two shapes in the same space. It shows  $\\mathcal{L}_{\\mathrm{scale}}$  helps the model to learn a more desirable template by penalizing the change in the global scale of the learned template fields.",
+        "bbox": [
+            75,
+            794,
+            468,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For the latter one, we add global deformation consistency for a pair of deformed shapes  $P', Q'$ :",
+        "bbox": [
+            498,
+            305,
+            890,
+            335
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g e o}} = \\mathbf {C D} \\left(P ^ {\\prime}, Q ^ {\\prime}\\right), \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            625,
+            340,
+            890,
+            358
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{CD}$  denotes a conventional Chamfer loss. Thus, the deformation consistency regularization can be written as:",
+        "bbox": [
+            498,
+            364,
+            890,
+            393
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d c}} = \\gamma_ {1} \\mathcal {L} _ {\\mathrm {p d c - g e o}} + \\gamma_ {2} \\mathcal {L} _ {\\mathrm {p d c - s e m}} + \\gamma_ {3} \\mathcal {L} _ {\\text {s c a l e}} + \\gamma_ {4} \\mathcal {L} _ {\\text {g e o}}. \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            511,
+            400,
+            890,
+            417
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3. Optimization",
+        "text_level": 1,
+        "bbox": [
+            500,
+            445,
+            638,
+            462
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We train our deformation and template fields with a reconstruction loss and several regularizers including our deformation consistency regularization. Following [2], the reconstruction loss  $\\mathcal{L}_{\\mathrm{rec}}$  for shape  $P$  is",
+        "bbox": [
+            496,
+            468,
+            890,
+            529
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e c}} = \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} | \\mathcal {F} _ {\\theta} (\\boldsymbol {x}) - \\bar {s} | + \\sum_ {\\boldsymbol {x} \\in P} (1 - \\langle \\nabla \\mathcal {F} _ {\\theta} (\\boldsymbol {x}), \\bar {n} \\rangle) \\\\ + \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} | \\| \\nabla \\mathcal {F} _ {\\theta} (\\boldsymbol {x}) \\| _ {2} - 1 | + \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P} \\backslash P} \\exp (- \\delta \\cdot | \\mathcal {F} _ {\\theta} (\\boldsymbol {x}) |), \\tag {12} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            534,
+            890,
+            606
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\bar{s}$  denotes the ground-truth SDF value of each query point,  $\\bar{n}$  is the ground truth surface normal vector of each surface point and  $\\nabla \\mathcal{F}_{\\theta}$  the spatial derivative of a 3D field. The second term of (12) is derived from the property of the SDF function that the gradient of an SDF is identical to the surface normal for surface points and the third is implicit geometric regularization [54] derived from the Eikonal equation. The last term pertains to penalizing off-surface points to be away from the surface points, which is controlled by  $\\delta \\gg 1$ .",
+        "bbox": [
+            496,
+            608,
+            890,
+            757
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Since we do not know ground truth mapping between the template and input shapes, we manipulate the deformation field through multiple regularizers. First, we employ our proposed deformation consistency regularization (11). For further regularizers, we adopt [2] such as deformation smoothness  $\\mathcal{L}_{\\mathrm{smooth}}$ , normal consistency  $\\mathcal{L}_{\\mathrm{normal}}$  and minimal correction  $\\mathcal{L}_{\\mathrm{c}}$ , which are defined as",
+        "bbox": [
+            496,
+            758,
+            890,
+            864
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {s m o o t h}} = \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} \\sum_ {d \\in \\{X, Y, Z \\}} \\| \\nabla \\mathcal {D} _ {\\theta_ {1}} ^ {v} (\\cdot , \\boldsymbol {z} _ {P}) | _ {d} (\\boldsymbol {x}) \\| _ {2} ^ {2}, \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            514,
+            869,
+            890,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "597",
+        "bbox": [
+            485,
+            945,
+            509,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/da82899a2c9ce4b1dc7fc4cccec1d8fceb2404da3d183bb1d68551ee615a29cb.jpg",
+        "image_caption": [
+            "Figure 6. Part label transfer result on ShapeNet. We transfer part segment labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations, please refer to the supplement."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            92,
+            883,
+            232
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {n o r m a l}} = \\sum_ {i} \\sum_ {\\boldsymbol {x} \\in \\mathcal {S} _ {i}} \\left(1 - \\left\\langle \\nabla \\mathcal {T} _ {\\theta_ {2}} \\left(\\boldsymbol {x} + D _ {\\theta_ {1}} ^ {v} (\\boldsymbol {x})\\right), \\bar {n} \\right\\rangle\\right), \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            96,
+            310,
+            468,
+            357
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c}} = \\sum_ {i} \\sum_ {\\boldsymbol {x} \\in \\Omega_ {P}} | \\mathcal {D} _ {\\theta_ {1}} ^ {\\Delta s} (\\boldsymbol {x}) |. \\tag {15}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            184,
+            368,
+            468,
+            401
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We also apply  $\\ell_2$  regularization on shape latent codes and part deformation priors",
+        "bbox": [
+            76,
+            407,
+            468,
+            438
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {e m b}} = \\sum_ {i} ^ {N} \\| z _ {i} \\| _ {2} ^ {2} + \\sum_ {j} ^ {k} \\| e _ {j} \\| _ {2} ^ {2}. \\tag {16}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            165,
+            448,
+            468,
+            491
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The regularizers are collectively written as",
+        "bbox": [
+            76,
+            501,
+            359,
+            517
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} = \\gamma_ {5} \\mathcal {L} _ {\\text {s m o o t h}} + \\gamma_ {6} \\mathcal {L} _ {\\text {n o r m a l}} + \\gamma_ {7} \\mathcal {L} _ {\\mathrm {c}} + \\gamma_ {8} \\mathcal {L} _ {\\text {e m b}}. \\tag {17}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            99,
+            527,
+            468,
+            545
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Hence, the final loss becomes",
+        "bbox": [
+            76,
+            556,
+            274,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\mathrm {d c}} + \\mathcal {L} _ {\\text {r e g}}. \\tag {18}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            192,
+            583,
+            468,
+            599
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluating dense correspondence. Given two pairs of seen or unseen shapes  $P, Q$ , we can evaluate the dense correspondence between  $P$  and  $Q$  by deforming each shape to the common template space. Moreover, the uncertainty of dense correspondence performance between two points  $p \\in P$  and  $q \\in Q$  can be measured as below:",
+        "bbox": [
+            75,
+            611,
+            468,
+            702
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nu (p, q) = 1 - \\exp (- \\gamma \\| (p + \\Delta p) - (q + \\Delta q) \\| _ {2} ^ {2}). \\tag {19}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            94,
+            712,
+            468,
+            729
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "$\\gamma$  is a hyperparameter for modulation and  $\\Delta p, \\Delta q$  are output from  $\\mathcal{D}_{\\theta_1}$ . Note that since each part deformation prior  $e$  is shared based on part semantics across entire training shapes, we do not need a test-time optimization on  $e$  for an unseen shape at inference time.",
+        "bbox": [
+            75,
+            741,
+            468,
+            816
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            829,
+            209,
+            847
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We demonstrate the effectiveness of our proposed framework and provide both quantitative and qualitative results. In Section 5.1, we conduct multiple label transfer",
+        "bbox": [
+            75,
+            854,
+            468,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/14e884a9b2c350111152cd9b325302420efe9ad71d8c1301a84b84ea6fe4b7b3.jpg",
+        "table_caption": [
+            "Table 2. Keypoint correspondence accuracy in KeypointNet [55]. PCK scores with threshold of  $0.05 / 0.1$  are reported."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>PCK</td><td>airplane</td><td>car</td><td>chair</td><td>table</td></tr><tr><td>AtlasNetV2 [34]</td><td>23.9 / 45.3</td><td>33.5 / 64.0</td><td>25.6 / 50.2</td><td>25.3 / 48.0</td></tr><tr><td>DIT [2]</td><td>13.6 / 34.2</td><td>58.9 / 87.7</td><td>33.3 / 66.7</td><td>26.2 / 55.5</td></tr><tr><td>DIF [1]</td><td>16.5 / 40.5</td><td>67.8 / 90.4</td><td>34.3 / 57.6</td><td>37.9 / 62.9</td></tr><tr><td>Ours</td><td>24.2 / 52.3</td><td>69.3 / 91.8</td><td>61.2 / 83.2</td><td>50.8/77.9</td></tr></table>",
+        "bbox": [
+            501,
+            311,
+            893,
+            388
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/2d58bb777779ba8ee363b11e57efc839f42847be5442e9977970070eaba1d71c.jpg",
+        "table_caption": [
+            "Table 3. 5-shot part label transfer in four categories of ShapeNet. The reported numbers are overall mIoU."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>model</td><td>airplane</td><td>car</td><td>chair</td><td>table</td></tr><tr><td>AtlasNetV2 [34]</td><td>57.8</td><td>59.3</td><td>66.9</td><td>70.6</td></tr><tr><td>DIT [1]</td><td>64.2</td><td>66.1</td><td>80.4</td><td>86.0</td></tr><tr><td>DIF [2]</td><td>73.9</td><td>69.6</td><td>82.1</td><td>86.6</td></tr><tr><td>Ours</td><td>75.1</td><td>70.4</td><td>84.2</td><td>86.9</td></tr></table>",
+        "bbox": [
+            517,
+            428,
+            875,
+            525
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "tasks for dense correspondence evaluation. In Section 5.2, we measure shape reconstruction accuracy to evaluate the representational power for unknown shapes. In Section 5.3, we present advanced dense correspondence analysis under more challenging settings. Lastly, in Section 5.4, we perform ablation studies to analyze the proposed regularization and the sensitivity over feature extractor [31].",
+        "bbox": [
+            496,
+            549,
+            890,
+            655
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Datasets and Baselines. We train and validate our model and baselines on four categories (e.g., chair, table, airplane and car) of ShapeNetV2 [50]. We follow the training split of DIF [2] for each category. For label transfer tasks, we use the labels from ShapeNet-Part [56] and KeypointNet [55]. DIT [1] and DIF [2] are the most relevant baseline models as they both globally learn implicit template fields, and we select AtlasNetV2 [34] as an additional baseline. Dense correspondence is measured with local patch composition in [34]. For more implementation details, analyses, and visualizations, see the supplement.",
+        "bbox": [
+            496,
+            657,
+            890,
+            823
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Dense correspondence",
+        "text_level": 1,
+        "bbox": [
+            500,
+            832,
+            707,
+            848
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the quality of learned dense correspondence, we conduct various surrogate tasks including keypoint transfer, part label transfer, and texture transfer due to",
+        "bbox": [
+            496,
+            854,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "598",
+        "bbox": [
+            486,
+            945,
+            511,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d992a9a2429a7e8c487f80fb2c12c05698fa643a016b4a70d35d2eca80ee5f3e.jpg",
+        "image_caption": [
+            "Figure 7. Texture transfer result on ShapeNet. We transfer part segment labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations, please refer to the supplement."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            89,
+            92,
+            877,
+            258
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/90e6e2b99689eff2b04221393f9d7235b1b3c75beda2f97bc90fbbe1218e35c2.jpg",
+        "table_caption": [
+            "Table 4. Reconstruction accuracy for unseen shapes in four categories of ShapeNet. The reported numbers are chamfer distance (CD) between ground truth and reconstructed meshes at a resolution of  $256^{3}$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>CD (×103)↓</td><td>airplane</td><td>car</td><td>chair</td><td>table</td></tr><tr><td>DIT [1]</td><td>0.2261</td><td>0.4082</td><td>0.3063</td><td>0.6187</td></tr><tr><td>DIF [2]</td><td>0.2093</td><td>0.3157</td><td>0.4229</td><td>0.2640</td></tr><tr><td>Ours</td><td>0.1859</td><td>0.3150</td><td>0.2265</td><td>0.2796</td></tr></table>",
+        "bbox": [
+            84,
+            356,
+            460,
+            439
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "the lack of labeled datasets for 3D dense correspondence. For the following tasks, we transfer respective labels from source shapes to target shapes in a common template space.",
+        "bbox": [
+            75,
+            470,
+            468,
+            517
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Keypoint label transfer. Table 2 presents the quantitative result in a 5-shot keypoint transfer task. We measure the performance by a percentage of correct keypoints (PCK) [57]. We compute the distance between the transferred keypoints and the ground truth points to record the PCK score with the threshold of  $0.05 / 0.1$ . Our method achieves the best performance in all four categories with a significant improvement. Especially for categories with high structural variability (e.g., chair and table), a large performance gain is observed. For instance, even compared to the best baseline, on the chair category with the threshold of 0.1, our method (83.2) outperforms DIT [1] (66.7) by  $16.5\\%$ , and on the table category with the threshold of 0.1, our method (77.9) outperforms DIF [2] (62.9) by  $15\\%$ . We additionally provide the visualization of keypoint transfer in Figure 4, which also support the importance of understanding part semantics for shape correspondence. In particular, two keypoints of pea-green and orange from the source chair are geometrically close but semantically different since one is for the arm part and the other is for the seat part. Among several methods, only our framework succeeds in disentangling two keypoints in the target chair. We provide more qualitative results regarding keypoint transfer results in real-world data [53] to validate the robustness of our method in the supplement.",
+        "bbox": [
+            75,
+            523,
+            472,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Part label transfer. For few-shot part label transfer tasks, we transfer part segment labels from source shapes to target shapes. In detail, for each (target) point of the deformed target shape, we find the  $n$  nearest points of the deformed source shapes and perform weighted voting using their labels and distances to the target point. In our experiments, we used  $n = 10$  nearest points. For a 5-shot part label transfer task in Table 3, five source shapes were used just the same as [2] for every category. Table 3 shows our framework achieves the best performance over all four categories. Moreover, Figure 6 illustrates the qualitative results of one-shot part label transfer on ShapeNet [50]. Similar to keypoint transfer, Figure 6 also proves the effectiveness of semantic knowledge in dense correspondence.",
+        "bbox": [
+            496,
+            306,
+            893,
+            518
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Texture transfer. Figure 7 shows texture transfer results on ShapeNetV2 [50] to compare established dense correspondence between baselines and ours with higher-frequency label transfer. We observe notable texture transfer results. For instance, our framework successfully transfers the source car's texture of the windshield (gray) to the corresponding part of the target cars, whereas baselines transfer the same textures to the roof part of the target cars.",
+        "bbox": [
+            496,
+            518,
+            893,
+            640
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2. Shape reconstruction",
+        "text_level": 1,
+        "bbox": [
+            500,
+            650,
+            700,
+            666
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We report the reconstruction quality of unseen shapes with Chamfer Distance (CD) to validate the representational power of our framework in Table 4. A smaller CD means more accurate reconstruction. As in previous work, we report the  $\\mathrm{CD} \\times 10^{3}$  for readability. Also, we consistently discuss the CD and performance gain below at this scale. We optimize the corresponding shape latent code for a given unseen shape. We measure CD as in DIF [2], where reconstructed meshes are extracted at a resolution of  $256^{3}$ , and CD is calculated with  $10K$  sampled points. Our framework on average achieved the best reconstruction results surpassing DIF by a margin of 0.0512. On average, our method and DIF exhibit 0.2518 and 0.3030 reconstruction errors, respectively. In three out of four categories, our method achieved the best reconstruction accuracy.",
+        "bbox": [
+            496,
+            674,
+            893,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "599",
+        "bbox": [
+            486,
+            945,
+            509,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3. More challenging setups",
+        "text_level": 1,
+        "bbox": [
+            76,
+            90,
+            302,
+            107
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To further validate the effectiveness of our method, we perform additional experiments and compare the results with DIF [2]. Overall, the results below imply current methods that lack part semantics and do not consider part deformation consistency may fail to learn semantically plausible deformation in challenging cases.",
+        "bbox": [
+            75,
+            113,
+            468,
+            204
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "2-shot part label transfer within different subcategories. Our objective is to assess dense correspondence performance under difficult settings. We conduct 2-shot part label transfer within two different subcategories belonging to the same category, where shapes from different subcategories tend to have different geometry for same semantics. Since ShapeNet [50] includes incorrect subcategory labels [58], we manually cleaned subcategory labels for four categories (e.g., airplane: airliner/jet, car: sedan/jeep, chair: straight chair/sofa, table: pedestal table/short table). Figure 8 displays examples of shapes from each subcategory. For robust evaluations, we randomly select the source shape 5 times for each subcategory. Table 5 shows our framework achieves a significant performance gain in all cases by over  $7.2\\%$  on average compared to DIF [2].",
+        "bbox": [
+            75,
+            205,
+            470,
+            430
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Unsupervised dense correspondence in DFAUST [59]. Furthermore, we evaluate unsupervised dense correspondence with non-rigid shapes [59]. For more challenging settings, we train both DIF [2] and our model for two different subjects with different sizes and heights (e.g., subject ID of 50002, 50026) and four distinct actions (e.g., shake arms, chicken wings, running on spot, and jumping jacks). Colormap in Figure 9 demonstrates the correspondence quality, where we transfer color over the up-axis. In the source shape, the arms are labeled in red. After transferring these color labels to the target shapes, our method preserves the red/orange labeling of the arms across different motions, whereas the baseline fails to capture the same labeling (i.e., yellow hands) since it lacks semantic knowledge during training. Additionally, we conduct keypoint transfer tasks and measure the PCK score as identical as in Section 5.1. Our model achieves 74.9/85.2 with a threshold of 0.05/0.1, whereas DIF [2] achieves 67.0/80.7 with the same threshold, showing our superior performance in the challenging settings (see the supplement for the full keypoint transfer result). Note that we can also learn implicit templates for non-rigid shapes with supervision as in [19]. We conduct experiments in unsupervised settings since our focus is to validate the effectiveness of utilizing additional semantic information for template learning.",
+        "bbox": [
+            75,
+            431,
+            470,
+            809
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4. Ablations",
+        "text_level": 1,
+        "bbox": [
+            76,
+            816,
+            187,
+            830
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Ablation on regularizations. We conduct ablation studies regarding the proposed regularizations including part deformation consistency regularizations, global scale consistency regularization, and global deformation consistency",
+        "bbox": [
+            75,
+            839,
+            468,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6cd10adf61d83a66ca80750a8786e2ac526e6ea83495786caa65bca3c3b03ef5.jpg",
+        "image_caption": [
+            "Figure 8. Types of source shapes for 2-shot part label transfer."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            90,
+            905,
+            152
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/dc1a2082addac86b30bcfc88ccc5cdc6b77036672b43248d1c28ef43db747ff9.jpg",
+        "table_caption": [
+            "Table 5. 2-shot part label transfer in subcategories of ShapeNet. The reported numbers are overall mIoU over 5 runs."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>subcategories</td><td>DIF [2]</td><td>Ours</td></tr><tr><td>airplane(airliner/jet)</td><td>64.5±1.61</td><td>72.8±0.99 (+8.3↑)</td></tr><tr><td>car(sedan/jeep)</td><td>65.1±2.34</td><td>69.1±1.46 (+4.0↑)</td></tr><tr><td>chair(straight/sofa)</td><td>66.3±6.06</td><td>77.5±5.81 (+11.2↑)</td></tr><tr><td>table(pedestal/short)</td><td>74.4±10.1</td><td>79.7±6.26 (+5.3↑)</td></tr></table>",
+        "bbox": [
+            509,
+            210,
+            885,
+            301
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/87a877925235bd64d7f0802809ba8f9a944ff6a557b07ab22f5af790ef6063d1.jpg",
+        "image_caption": [
+            "Figure 9. Color transfer on DFAUST [59]. We transfer different color labels over the up-axis from source shapes (colored in ivory) to target shapes and compare ours with DIF [2]."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            316,
+            893,
+            436
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/4a4aa9ec1574d04903a1e52c1ee7901ed039391446da6f32798f612695cd34be.jpg",
+        "table_caption": [
+            "Table 6. Ablations for every proposed regularization. We report mIoU for 5-shot part label transfer on ShapeNet chair. Note that  $\\mathcal{L}_{\\mathrm{pdc}}$  denotes  $\\mathcal{L}_{\\mathrm{pdc\\_geo}} + \\mathcal{L}_{\\mathrm{pdc\\_sem}}$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Lpdc</td><td>Lgeo</td><td>Lscale</td><td>chair</td></tr><tr><td>(a)</td><td>✓</td><td></td><td></td><td>83.4</td></tr><tr><td>(b)</td><td>✓</td><td>✓</td><td></td><td>83.9</td></tr><tr><td>(c)</td><td>✓</td><td>✓</td><td>✓</td><td>84.2</td></tr></table>",
+        "bbox": [
+            573,
+            537,
+            820,
+            616
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/16617a34d40e30f85270118a5840eaf7ff9adf29a7f500f6fcf9c99f2b8a6a47.jpg",
+        "table_caption": [
+            "Table 7. BAE-Net  $k$  ablation. We report mIoU for 5-shot part label transfer on ShapeNet chair over 3 runs."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>k=2</td><td>k=4</td><td>k=6</td><td>k=8</td><td>k=10</td></tr><tr><td>chair</td><td>83.8±0.51</td><td>83.5±1.48</td><td>82.8±1.27</td><td>83.6±0.85</td><td>83.7±1.33</td></tr></table>",
+        "bbox": [
+            506,
+            656,
+            879,
+            689
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "regularization. We conduct 5-shot part label transfer within a chair category, as same as Section 5.1. Table 6 shows that each component of our proposed regularizers contributes to performance improvement for dense correspondence. Note that by only using  $\\mathcal{L}_{\\mathrm{pdc}}$ , i.e., using  $\\mathcal{L}_{\\mathrm{pdc\\_geo}}$  and  $\\mathcal{L}_{\\mathrm{pdc\\_sem}}$ , it still outperforms than baselines (80.4 [1] and 82.1 [2]). The best performance was achieved by incorporating all proposed regularizations.",
+        "bbox": [
+            496,
+            703,
+            890,
+            824
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "BAE-Net [31] with different  $k$ . We evaluate the effect of  $k$  for BAE-Net [31] with mean/standard deviation over 3 runs for ShapeNet [50] chair in We train BAE-Net [31] with a range of  $k = (2,4,6,8,10)$ , where  $k = 4$  is used in the main paper, and report mIoU for 5-shot part label",
+        "bbox": [
+            496,
+            824,
+            893,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "600",
+        "bbox": [
+            486,
+            945,
+            511,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "transfer task. Table 7 shows the efficacy of semantic priors, where our framework achieves the average mIoU of 83.5, which shows higher performance than baselines. More importantly, we observed that the performance is fairly robust over  $k$ .",
+        "bbox": [
+            75,
+            90,
+            472,
+            167
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            76,
+            179,
+            196,
+            194
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We have proposed a new framework that learns a semantic-aware implicit template field. Our framework extracts semantic knowledge from a self-supervised feature extractor given unlabelled point clouds. By leveraging the semantic information, we propose novel conditioning with semantic-aware deformation code and part deformation consistency regularizations to encourage semantically plausible deformation. Furthermore, we suggest global scaling consistency and global deformation consistency regularizations to enhance the learning process. Through quantitative and qualitative evaluation, we demonstrate the superiority of our proposed method and verified the effect of all proposed components in our framework.",
+        "bbox": [
+            75,
+            203,
+            472,
+            400
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments. This research was supported in part by the MSIT (Ministry of Science and ICT), Korea, under the ICT Creative Consilience program (IITP-2023-2020-0-01819) supervised by the IITP (Institute for Information & communications Technology Planning & Evaluation); the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF-2023R1A2C2005373); and Google Cloud Research Credits program with the award (GCLF-67WL-WU6D-MF14, V9P3-99XD-64CL-EH5H).",
+        "bbox": [
+            75,
+            417,
+            472,
+            568
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            579,
+            174,
+            594
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3d shape representation. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8",
+            "[2] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8",
+            "[3] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. TOG, 2015. 1, 2",
+            "[4] Brett Allen, Brian Curless, and Zoran Popovic. Articulated body deformation from range scan data. TOG, 2002. 1",
+            "[5] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In ECCV, 2018. 1, 2",
+            "[6] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 1",
+            "[7] Wang Zeng, Wanli Ouyang, Ping Luo, Wentao Liu, and Xiaogang Wang. 3d human mesh regression with dense correspondence. In CVPR, 2020. 1, 2"
+        ],
+        "bbox": [
+            86,
+            604,
+            468,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[8] Sing Chun Lee and Misha Kazhdan. Dense point-to-point correspondences between genus-zero shapes. In Computer Graphics Forum, 2019. 1, 2",
+            "[9] Silvia Zuffi, Angjoo Kanazawa, David W. Jacobs, and Michael J. Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 1, 2",
+            "[10] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In AVSS, 2009. 1",
+            "[11] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. TOG, 2017. 1",
+            "[12] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 1",
+            "[13] Srinath Sridhar, Antti Oulasvirta, and Christian Theobalt. Interactive markerless articulated hand motion tracking using rgb and depth data. In ICCV, 2013. 1",
+            "[14] Iason Oikonomidis, Manolis IA Lourakis, and Antonis A Argyros. Evolutionary quasi-random search for hand articulations tracking. In CVPR, 2014. 1",
+            "[15] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 1, 2",
+            "[16] Vignesh Ganapathi-Subramanian, Olga Diamanti, Soeren Pirk, Chengcheng Tang, Matthias Niessner, and Leonidas Guibas. Parsing geometry using structure-aware shape templates. In 3DV, 2018. 1, 2",
+            "[17] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation. In CVPR, 2018. 1, 2",
+            "[18] Pablo Palafox, Nikolaos Sarafianos, Tony Tung, and Angela Dai. Spams: Structured implicit parametric models. In CVPR, 2022. 1",
+            "[19] Ramana Sundararaman, Gautam Pai, and Maks Ovsjanikov. Implicit field supervision for robust non-rigid shape matching. In ECCV, 2022. 1, 8",
+            "[20] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1, 2",
+            "[21] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. 2019. 1, 3",
+            "[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 1, 2, 3",
+            "[23] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1, 2, 3"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "601",
+        "bbox": [
+            486,
+            945,
+            509,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[24] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, 2020. 1",
+            "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2",
+            "[26] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 2",
+            "[27] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In CVPR, 2021. 1, 2",
+            "[28] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 1, 2",
+            "[29] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In CVPR, 2021. 1, 2",
+            "[30] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 1, 2",
+            "[31] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. In ICCV, 2019. 2, 4, 6, 8",
+            "[32] Vladimir G Kim, Wilmot Li, Niloy J Mitra, Siddhartha Chaudhuri, Stephen DiVerdi, and Thomas Funkhouser. Learning part-based templates from large collections of 3d shapes. TOG, 2013. 2",
+            "[33] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T. Freeman, and Thomas A. Funkhouser. Learning shape templates with structured implicit functions. In ICCV, 2019. 2",
+            "[34] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. In NeurIPS, 2019. 2, 6",
+            "[35] Jiahui Lei and Kostas Daniilidis. Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In CVPR, 2022. 2",
+            "[36] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural template: Topology-aware reconstruction and disentangled generation of 3d meshes. In CVPR, 2022. 2",
+            "[37] Zekun Hao, Hadar Averbuch-Elor, Noah Snavely, and Serge Belongie. Dualsdf: Semantic shape manipulation using a two-level representation. In CVPR, 2020. 2",
+            "[38] Yueqi Duan, Haidong Zhu, He Wang, Li Yi, Ram Nevatia, and Leonidas J. Guibas. Curriculum deepsdf. In ECCV, 2020. 2"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[39] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2",
+            "[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2",
+            "[41] Michael Niemeyer and Andreas Geiger. Campari: Camera-aware decomposed generative neural radiance fields. In 3DV, 2021. 2",
+            "[42] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, et al. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2",
+            "[43] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. Rimnet: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In CVPR, 2022. 2",
+            "[44] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J Guibas, and Hao Zhang. Adacoseg: Adaptive shape co-segmentation with group consistency loss. In CVPR, 2020. 2",
+            "[45] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Part slip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In CVPR, 2023. 2",
+            "[46] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2",
+            "[47] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 2",
+            "[48] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In NeurIPS, 2022. 2",
+            "[49] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2022. 2",
+            "[50] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 4, 6, 7, 8",
+            "[51] David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. In ICLR, 2017. 4",
+            "[52] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 4",
+            "[53] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 5, 7"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "602",
+        "bbox": [
+            486,
+            945,
+            511,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[54] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 5",
+            "[55] Yang You, Yujing Lou, Chengkun Li, Zhoujun Cheng, Liangwei Li, Lizhuang Ma, Cewu Lu, and Weiming Wang. Keypointnet: A large-scale 3d keypoint dataset aggregated from numerous human annotations. In CVPR, 2020. 6",
+            "[56] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3d shape collections. TOG, 2016. 6",
+            "[57] Li Yi, Hao Su, Xingwen Guo, and Leonidas Guibas. Syncspeccnn: Synchronized spectral cnn for 3d shape segmentation. In CVPR, 2017. 7",
+            "[58] Xinhai Liu, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Transactions on Image Processing, 2021. 8",
+            "[59] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In CVPR, 2017. 8"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            407
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "603",
+        "bbox": [
+            486,
+            945,
+            509,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/840dd92d-8817-48f0-ac2c-586a3d27151d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2588ef90fd0697b8f30d88d98ed503c417918fb95bc4f5c5ad7a745dd7094b2f
+size 3691891

2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/full.md

@@ -0,0 +1,382 @@
+# Semantic-Aware Implicit Template Learning via Part Deformation Consistency
+
+Sihyeon Kim Minseok Joo Jaewon Lee  
+Juyeon Ko Juhan Cha Hyunwoo J. Kim*
+
+Department of Computer Science and Engineering, Korea University
+
+{sh_bs15, wlgkcjf87, 2jlejyu, juyon98, hanchaa, hyunwoojkim}@korea.ac.kr
+
+# Abstract
+
+Learning implicit templates as neural fields has recently shown impressive performance in unsupervised shape correspondence. Despite the success, we observe current approaches, which solely rely on geometric information, often learn suboptimal deformation across generic object shapes, which have high structural variability. In this paper, we highlight the importance of part deformation consistency and propose a semantic-aware implicit template learning framework to enable semantically plausible deformation. By leveraging semantic prior from a self-supervised feature extractor, we suggest local conditioning with novel semantic-aware deformation code and deformation consistency regularizations regarding part deformation, global deformation, and global scaling. Our extensive experiments demonstrate the superiority of the proposed method over baselines in various tasks: keypoint transfer, part label transfer, and texture transfer. More interestingly, our framework shows a larger performance gain under more challenging settings. We also provide qualitative analyses to validate the effectiveness of semantic-aware deformation. The code is available at https://github.com/mlvlab/PDC.
+
+# 1. Introduction
+
+Template learning is essential for shape analysis since the template, which is a compact representation of arbitrary shapes, provides a good prior for shape fitting so that we can establish dense correspondence and transfer key attributes such as texture or keypoints between shapes. Traditionally, handcrafted templates such as meshes, part elements with primitive shapes (e.g., sphere, cylinder, cone), or parametric models have been widely used for human/animal body [3, 4, 5, 6, 7, 8, 9], human face [10, 11], human hand [12, 13, 14], and generic objects [15, 16, 17]. Defining templates for less complex shapes with consistent topologies and similar structures is relatively straightforward. However, for shapes like generic object shapes with high structural variability or
+
+![](images/b8c7719d3d700afb1d874be71a8a7714a1fb5c826244ecbfeb71ba0467281d7f.jpg)
+
+![](images/068252e044fa8ddddc802a3c0d2b98ea36759befa0f5a17a5bcffdd3449bdcb3.jpg)  
+Figure 1. Importance of part deformation consistency. Given input shapes with high structural variability, only considering low-level geometry during template learning [1, 2] can lead to suboptimal deformation (e.g., cannot distinguish between seat and arm where two parts are geometrically close while semantically distant: low seat-low arm, high seat-high arm). Ours succeeds in transferring shape attributes by encouraging part deformation consistency with semantic information.
+
+non-rigid shapes with large-scale deformations, manually defining suitable templates becomes challenging.
+
+To address this challenge, recent works [1, 2, 18, 19] have focused on learning implicit templates based on neural fields [20]. Neural fields, which are known for strong representation power, efficiency, and flexibility, have been successful in multiple 3D vision tasks, e.g., shape/scene reconstruction [21, 22, 23, 24], neural rendering [25, 26, 27], and human digitization [28, 29, 30]. Likewise, neural fields-based template learning have shown superior perfor
+
+mance on dense correspondence between complex shapes. Both [1] and [2] suggest decomposing the implicit representation of shapes into a deformation field and a template field. Here, the template is learned as a continuous function, which is expected to capture shared structures among the given shapes within a category (e.g., airplane, chair), while the deformation depends solely on the conditioned geometry of each shape (i.e., SDF value and surface normal). We observe current approaches often learn suboptimal deformation especially when a shape category has high structural variability like diverse local deformation scales and unseen/missing part structures; see Figure 1. That is, incorporating extra knowledge of part semantics is necessary to handle such cases.
+
+In this paper, we propose a novel framework to learn semantically plausible deformation by distilling the knowledge of self-supervised feature extractor [31] and employing it as a semantic prior to satisfy part deformation consistency during the template learning process. Part deformation consistency is our inductive bias that the deformation result within the same part should be consistent. By leveraging semantic prior to understand part arrangements of shapes, we suggest local conditioning with semantic-aware deformation code and carefully designed regularizations to encourage semantically corresponding deformation. Semantic-aware deformation code is a point-wise soft assignment of part deformation priors, where part deformation priors are multiple latent codes to encode information on corresponding part deformation. We calculate assigned weights considering the part semantics of each point. In addition, we propose input space regularization and latent space regularization to encourage part-level distances to be close, providing an explicit way to control the deformation field so that the model can learn flexible deformation and successfully learn common structures under diverse shapes. Lastly, we suggest global scale consistency regularization to preserve the volume of the template against large-scale deformations. To validate the effectiveness of our framework, we conduct extensive surrogate experiments reflecting different levels of correspondence such as keypoint transfer, part label transfer, and texture transfer, and demonstrate competitive performance over baselines with qualitative analyses.
+
+Our contributions are summarized as follows: ① We propose a new framework that learns global implicit template fields founded on part deformation consistency while understanding part semantics to enable semantically plausible deformation, ② We impose hybrid conditioning by combining global conditioning of geometry and local conditioning of part semantics for flexible deformation, ③ We design novel regularizations to successfully manipulate the template fields, ④ We show the effectiveness of our framework with both qualitative and quantitative experiments.
+
+# 2. Related Works
+
+Template learning for shape correspondence. Defining a template for a set of arbitrary shapes has been studied for a long time. A high-quality template provides a good initialization for shape alignment and various applications in vision and graphics. For shapes with consistent topologies and structures such as faces or bodies of humans and animals, handcrafted templates have been popular [3, 5, 7, 8, 9] and still widely used today. On the other hand, it was more challenging for generic object shapes with inconsistent topology and diverse part structures. Prior works suggest to predefine or learn local templates from data as in primitive parts [15] or part elements [32, 16, 33, 17, 34] to handle the inconsistent structure. Recent works have tried to learn a canonical shape of generic objects through implicit template field [1, 2, 35, 36], which are proven to be effective for representing shapes. However, existing approaches mostly lack explicit consideration of semantic knowledge, which enables semantically plausible deformation so that the learned template captures a common structure even with the various structured shapes.
+
+Neural fields for 3D representations. Recent studies have shown that representing 3D shapes or scenes as implicit representations with a neural network has several advantages: powerful representation, flexibility, and memory efficiency. These are called neural fields [20]. For instance, Occupancy networks [22] proposed a method to represent 3D geometry as a continuous occupancy function. DeepSDF [23] learns 3D complex surfaces with a neural network of auto-decoder framework for 3D modeling. Further, implicit representation has been introduced to various applications [37, 38]. To leverage the advantages, many applications based on neural fields have emerged such as neural rendering [25, 26, 27], human digitization [28, 29, 30], and generative modeling [39, 40, 41, 42]. They have also been applied to learn the template of complex shapes as mentioned above.
+
+Utilizing self-supervised segmentation models. Recently, there has been a growing body of research on analyzing complex shapes in a self-supervised manner due to the high cost of annotation for 3D data. Unsupervised semantic segmentation/co-segmentation models like BAE-Net [31], RIM-Net [43], AdaCoSeg [44], and PartSLIP [45] have demonstrated the ability to comprehend part semantics without ground truth labels. For 2D and multi-modal tasks, using large-scale self-supervised segmentation models like DINO [46] as a feature extractor has been widely adopted, including pixelNeRF [47], Feature Field [48], and ClipN-erF [49]. Inspired by these recent works, we utilize pretrained unsupervised part segmentation models to harness part semantic information for template learning.
+
+![](images/2c8c62fcf1b166e69f5abf8d598351ffd928a7e37ec0a0d5fc4bebd571ae0dd2.jpg)  
+Figure 2. Framework of semantic-aware implicit template learning. To learn semantic-aware implicit template field, (a) we extract part semantics from pretrained feature extractor  $E$  and incorporate knowledge into semantic-aware deformation code  $(\alpha)$  and (b) part deformation consistency regularizations. With optimizing with the reconstruction loss  $\mathcal{L}_{\mathrm{rec}}$  and several regularizers, our framework succeeds to learn semantically plausible deformation.
+
+![](images/ee7a919c08e58d123d46598646c37fef17ae02e62c65416039063fcee5aef637.jpg)  
+(a)
+
+![](images/f523e7f3c92f502460306e545568713d7bd79a9809d81f4cd9455b4a4c8abcb5.jpg)
+
+![](images/801ade73fef454d3add5f9bef7712229db9601be40b7f48707a00ebd1f041aa5.jpg)  
+(b)
+
+# 3. Preliminaries
+
+Implicit templates have been studied with neural fields. Given a 3D shape  $P$ , a neural field, which is a neural network  $\mathcal{F}_{\theta}$  parameterized by  $\theta$ , approximates continuous functions such as an occupancy function [21, 22] or a Signed Distance Function (SDF) [23] as
+
+$$
+\mathcal {F} _ {\theta} (\boldsymbol {x}; \boldsymbol {z} _ {P}) = s. \tag {1}
+$$
+
+Herein,  $\pmb{x} \in \mathbb{R}^3$  is a query point and  $s$  is an occupancy/SDF function value. To represent multiple 3D shapes, for each shape, a shape latent code  $\pmb{z}_P \in \mathbb{R}^d$  is learned. In general, a surface is defined as the zero-level set of a SDF, i.e.,  $\{x | \mathcal{F}_{\theta}(x; z_P) = 0\}$  as in DeepSDF [23].
+
+The neural field can be further decomposed into an implicit template field  $\mathcal{T}$  and a deformation field  $\mathcal{D}[1,2]$ :
+
+$$
+\mathcal {F} _ {\theta} (\boldsymbol {x}; \boldsymbol {z} _ {P}) = \mathcal {T} _ {\theta_ {2}} \left(\mathcal {D} _ {\theta_ {1}} (\boldsymbol {x}; \boldsymbol {z} _ {P})\right). \tag {2}
+$$
+
+The implicit template field  $\mathcal{T}_{\theta_2}$  captures a common structure and outputs an occupancy/SDF value  $s$  in the template space.  $\mathcal{D}_{\theta_1}$  maps a point  $x$  from a shape-specific space to the common template space by adding a deformation vector  $\Delta \pmb{x} \in \mathbb{R}^3$ , i.e.,  $\pmb{x} + \Delta \pmb{x}$ . In DIF-Net [2],  $\mathcal{D}_{\theta_1}$  additionally predicts a correction value  $\Delta s \in \mathbb{R}$  to capture instance-specific local structures. In this case, the neural field  $\mathcal{F}_{\theta}$  is defined as
+
+$$
+\mathcal {F} _ {\theta} (\boldsymbol {x}; \boldsymbol {z} _ {P}) = \mathcal {T} _ {\theta_ {2}} (\boldsymbol {x} + \Delta \boldsymbol {x}) + \Delta s \tag {3}
+$$
+
+$$
+\Delta s, \Delta \boldsymbol {x} = \mathcal {D} _ {\theta_ {1}} (\boldsymbol {x}; \boldsymbol {z} _ {P}).
+$$
+
+The existing approaches [1, 2] learn the implicit template field by minimizing the distance between the template and the deformed individual shapes. The distance is measured by SDF values and surface normals. Also, several regularizers are imposed such as deformation smoothness (e.g., penalizing the spatial gradient of  $\mathcal{D}_{\theta_1}$  and restricting the deformation scale between points). However, the loss functions
+
+and regularizers are computed by low-level geometric information. They are not effective enough to handle generic objects with large structural variability. Hence, in this paper, we take part semantics into account for template learning to achieve 'semantically plausible' deformation and dense correspondence.
+
+# 4. Method
+
+We present a semantic-aware implicit template learning framework based on part deformation consistency. In this section, we first delineate the overall architecture of our method equipped with part deformation priors and semantic-aware deformation code. Then, we introduce our deformation consistency regularizers including part deformation consistency regularizations. Lastly, we discuss the training procedure with several additional loss functions.
+
+# 4.1. Semantic-aware implicit template field
+
+We here present a novel semantic-aware implicit template learning framework. The overall pipeline of the proposed method is shown in Figure 2(a).
+
+Template field and deformation field. We extend the implicit template learning by incorporating semantic information into the deformation field. Specifically, given a query point  $\pmb{x}$  of a shape  $P$ , the deformation field  $\mathcal{D}_{\theta_1}$  predicts a deformation vector  $\Delta x$  and a correction value  $\Delta s$  conditioned on a point-wise semantic-aware deformation code (SDC)  $\alpha(x; P)$  in addition to a shape latent code  $z_P$ . Then, our framework is defined as
+
+$$
+\mathcal {F} _ {\theta} (\boldsymbol {x}; P, \mathcal {E}) := \mathcal {T} _ {\theta_ {2}} (\boldsymbol {x} + \Delta \boldsymbol {x}) + \Delta s, \tag {4}
+$$
+
+$$
+\Delta \boldsymbol {x}, \Delta s = \mathcal {D} _ {\theta_ {1}} (\boldsymbol {x}; \boldsymbol {z} _ {P}, \boldsymbol {\alpha} (\boldsymbol {x}; P)),
+$$
+
+where  $\mathcal{E}$  is a pretrained feature extractor. We assign  $\alpha (\pmb {x};P)$  based on a point-wise semantic feature, which is extracted from  $\mathcal{E}$ . For simplicity, we denote the neural field  $\mathcal{F}_{\theta}(\pmb {x};P,\mathcal{E})$  by  $\mathcal{F}_{\theta}(\pmb {x})$  and SDC  $\alpha (\pmb {x};P)$  by  $\alpha (\pmb {x})$ .
+
+Feature extractor. In this paper, for feature extraction, we mainly use BAE-Net [31] as an encoder  $\mathcal{E}$ , which is proven effective in capturing part semantics. Note that we pretrained BAE-Net on selected categories of ShapeNetV2 [50] by self-supervision without any extra labels or data for a fair comparison. The features from BAE-Net are  $k$  dimensional vectors, i.e.,  $\boldsymbol{o}^P(\boldsymbol{x}) = \mathcal{E}(\boldsymbol{x}, P) \in \mathbb{R}^k$ .  $k$  is a hyperparameter, which is closely related to the number of semantic parts. Since the number of semantic parts is unknown in self-supervised learning, we simply chose sufficiently large  $k$ .
+
+Semantic-aware deformation code. Our framework assumes that different semantic parts necessitate different deformations. To achieve this, the proposed method learns a deformation prior for each semantic part. The deformation priors are encoded as latent codes  $e_1, \dots, e_k \in \mathbb{R}^{d'}$ . A point  $x$  is softly assigned to  $k$  semantic parts using the softmax values of semantic features  $o^P(x)$ . Combining the soft assignment with deformation priors, semantic-aware deformation code (SDC) denoted as  $\alpha(x) \in \mathbb{R}^{d'}$  is defined as
+
+$$
+\boldsymbol {\alpha} (\boldsymbol {x}) = \sum_ {i = 1} ^ {k} \frac {\exp \left(\boldsymbol {o} _ {i} ^ {P} (\boldsymbol {x})\right) \boldsymbol {e} _ {i}}{\sum_ {j = 1} ^ {k} \exp \left(\boldsymbol {o} _ {j} ^ {P} (\boldsymbol {x})\right)}, \forall i \in \{1, \dots , k \}, \tag {5}
+$$
+
+where  $\pmb{o}_i^P(\pmb{x})$  denotes the  $i$ -th dimension value of the semantic features. The SDC at  $\pmb{x}$  is computed by a linear combination of deformation priors. We learn the point's probability of belonging to each of the  $k$  parts, which is served as part semantics.
+
+Remarks. For comparison, we also applied hard assignment of part deformation priors for SDC, i.e.,  $\alpha(\pmb{x}) = e_{i^*}$ , where  $i^* = \operatorname{argmax}_i (\pmb{o}_i^P(\pmb{x}))$ . However, as points are completely separated with local conditioning of the corresponding part deformation prior  $e_{i^*}$ , the final deformation result tends to break the shape geometry (see Figure 3), leading to performance degradation as shown in Table 1. Thereby, we choose the soft assignment to enable flexible deformation while preserving the shape geometry.
+
+We train the deformation field  $\mathcal{D}_{\theta_1}$  and  $\mathcal{T}_{\theta_2}$  in an end-to-end fashion. In our experiments,  $\mathcal{D}_{\theta_1}$  is a 6-layer MLP conditioned on both global geometry  $z_P$  via hypernetwork  $\Psi$  [51, 52] and local part semantics  $\alpha(x)$  via concatenation with input coordinates. The shape latent code  $z_P$  and part deformation priors  $\{\pmb{e}_i\}$  are randomly initialized and jointly optimized with neural fields.  $\mathcal{T}_{\theta_2}$  is a 5-layer MLP.
+
+Table 1. Ablations of different assignment strategies for SDC. We conduct a quantitative comparison between hard and soft assignments with 5-shot part label transfer on ShapeNet chair.  
+
+<table><tr><td>Assignment strategies</td><td>Hard</td><td>Soft</td></tr><tr><td>chair (5-shot PLT)</td><td>77.0</td><td>84.2</td></tr></table>
+
+![](images/3ffcd3cae66c1d892adb41f0722cb7d9eaec6987c5d6eee85e1dd68309f33035.jpg)  
+Figure 3. Effects of different assignment strategies for SDC. Hard assignment breaks the shape geometry, while soft assignment preserves the geometry while learning flexible deformations.
+
+# 4.2. Part deformation consistency regularizations
+
+In addition to conditioning neural fields, leveraging semantic information, we propose novel part deformation consistency regularizations to encourage semantically plausible deformation. Semantically similar points should be close after deformation in both an input space and a latent space. That is, the deformation of points from the same part should be consistent. Figure 2(b) illustrates the concept of our regularizations. Given a pair of deformed shapes  $P', Q'$ , part deformation consistency regularizations are defined as
+
+$$
+\begin{array}{l} \mathcal {L} _ {\mathrm {p d c - g e o}} = \sum_ {i = 1} ^ {k} \left(\frac {1}{| \mathcal {P} _ {i} ^ {\prime} |} \sum_ {\boldsymbol {x} \in \mathcal {P} _ {i} ^ {\prime}} \| \boldsymbol {x} - C (\boldsymbol {x}, \mathcal {Q} _ {i} ^ {\prime}) \| _ {2} ^ {2} \right. \\ \left. + \frac {1}{\left| \mathcal {Q} _ {i} ^ {\prime} \right|} \sum_ {\boldsymbol {y} \in \mathcal {Q} _ {i} ^ {\prime}} \| \boldsymbol {y} - C (\boldsymbol {y}, \mathcal {P} _ {i} ^ {\prime}) \| _ {2} ^ {2}\right), \text {a n d} \tag {6} \\ \end{array}
+$$
+
+$$
+\begin{array}{l} \mathcal {L} _ {\mathrm {p d c} \text {- s e m}} = \sum_ {i = 1} ^ {k} \left(\frac {1}{| \mathcal {P} _ {i} ^ {\prime} |} \sum_ {\boldsymbol {x} \in \mathcal {P} _ {i} ^ {\prime}} \| \boldsymbol {o} ^ {P ^ {\prime}} (\boldsymbol {x}) - \boldsymbol {o} ^ {Q ^ {\prime}} \left(C (\boldsymbol {x}, \mathcal {Q} _ {i} ^ {\prime})\right) \| _ {2} ^ {2} \right. \\ \left. + \frac {1}{| \mathcal {Q} _ {i} ^ {\prime} |} \sum_ {\boldsymbol {y} \in \mathcal {Q} _ {i} ^ {\prime}} \| \boldsymbol {o} ^ {Q ^ {\prime}} (\boldsymbol {y}) - \boldsymbol {o} ^ {P ^ {\prime}} \left(C (\boldsymbol {y}, \mathcal {P} _ {i} ^ {\prime})\right) \| _ {2} ^ {2}\right), \tag {7} \\ \end{array}
+$$
+
+where  $C$  is a correspondence matching function for a point  $\pmb{x}$  and its nearest point in a set  $\mathcal{Q}$ , i.e.,  $C(\pmb{x},\mathcal{Q}) = \operatorname*{argmin}_{\pmb{y}\in \mathcal{Q}}\| \pmb{x} - \pmb{y}\|_2^2$ .  $\mathcal{P}_i'$  and  $\mathcal{Q}_i'$  denote sampled point sets of
+
+$i^{\mathrm{th}}$  deformed part shapes of  $P^{\prime}$  and  $Q^{\prime}$ . They are defined as  $\mathcal{P}_i^\prime = \{\pmb {x} + \mathcal{D}_{\theta_1}(\pmb {x};\pmb {z}_P,\pmb {\alpha}(\pmb {x};P))\mid i = \operatorname *{argmax}_{k}(\pmb {o}_k^P (\pmb {x})),\pmb {x}\in$
+
+$\Omega_P\}$  and  $\mathcal{Q}_i^\prime = \{\pmb {x} + \mathcal{D}_{\theta_1}(\pmb {x};\pmb {z}_Q,\pmb {\alpha}(\pmb {x};Q))|i =$  argmax  $(\pmb{o}_k^Q (\pmb {x})),\pmb {x}\in \Omega_Q\}$  , where  $\Omega_P$  and  $\Omega_{Q}$  are sam
+
+pled query point sets for shape  $P$  and  $Q$ . Note that each deformed part shape is extracted based on part semantics from original shapes  $P$  and  $Q$ . Here, we explicitly guide the part-level deformation among various shapes to be consistent. Moreover, since we understand part arrangement and
+
+![](images/c11deb1e2f34ad9c566244693f1e65d187963d0795697e23acc5908163a4b361.jpg)  
+Figure 4. Keypoint transfer result on ShapeNet. We transfer keypoint labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations including results on ScanObjectNN [53], see the supplement.
+
+![](images/cafeb9d58e02ba8740aa2249583b33b699f5429902f293ffbc09e7ad539f23d0.jpg)  
+(a)
+
+![](images/4b36f80cd31eee68e6a5b448f1efa22671b904462eb9211f55d754c17fb51573.jpg)  
+(b)
+
+![](images/f5876dceecdf1c4231267053c9f0d994cbe0c8858bedf5f786589aa659e3d8c1.jpg)  
+(c)  
+Figure 5. Efficacy of global scaling consistency  $\mathcal{L}_{\mathrm{scale}}$
+
+only penalize matched parts, our method does not forcibly deform "unseen/missing match" cases.
+
+We partly observed spatial distortion when part deformation consistency was too strongly imposed. This is because using external semantics for deformation could break the continuity of geometric fields. In addition, we propose global scale consistency and global deformation consistency to address the concern. For the former one, we disentangle a scaling operation from the entire non-rigid deformation  $\mathcal{D}_{\theta_1}$  to calculate the global scale of the learned implicit template field. That is, we want to approximate  $\mathcal{D}$  with the global scaling of the scaling factor of  $r$ , i.e.,  $\mathcal{D} \approx rI$ . Then the problem becomes finding  $r$  from
+
+$$
+\min  _ {r} \sum_ {s} \sum_ {i} \| \boldsymbol {x} _ {i} ^ {s} + \Delta \boldsymbol {x} _ {i} ^ {s} - r \boldsymbol {x} _ {i} ^ {s} \| _ {2} ^ {2}, \tag {8}
+$$
+
+where  $s$  and  $i$  are an index for shape and point, respectively. A closed-form solution of (8) is  $r = \sum_{s}\sum_{i}\frac{\pmb{x}_i^{s\top}(\pmb{x}_i^s + \Delta\pmb{x}_i^s)}{\pmb{x}_i^{s\top}\pmb{x}_i^s}$ . See the supplement for proof. Since we want the learned  $\mathcal{D}$  to be insensitive to global scale operation, i.e., the implicit template to keep its scale by the average scale of input shapes, we penalize the expected  $\mathbb{E}[r]$  over input shapes  $P_{1}\dots P_{N}$  in a single batch as:
+
+$$
+\mathcal {L} _ {\text {s c a l e}} = | \mathbb {E} [ r ] - 1 | \tag {9}
+$$
+
+Figure 5 illuminates the effect of  $\mathcal{L}_{\mathrm{scale}}$ . For the same input shapes, Figure 5(a) is a car template learned without  $\mathcal{L}_{\mathrm{scale}}$ , Figure 5(b) is a car template learned with  $\mathcal{L}_{\mathrm{scale}}$ , and Figure 5(c) is the result of placing two shapes in the same space. It shows  $\mathcal{L}_{\mathrm{scale}}$  helps the model to learn a more desirable template by penalizing the change in the global scale of the learned template fields.
+
+For the latter one, we add global deformation consistency for a pair of deformed shapes  $P', Q'$ :
+
+$$
+\mathcal {L} _ {\mathrm {g e o}} = \mathbf {C D} \left(P ^ {\prime}, Q ^ {\prime}\right), \tag {10}
+$$
+
+where  $\mathbf{CD}$  denotes a conventional Chamfer loss. Thus, the deformation consistency regularization can be written as:
+
+$$
+\mathcal {L} _ {\mathrm {d c}} = \gamma_ {1} \mathcal {L} _ {\mathrm {p d c - g e o}} + \gamma_ {2} \mathcal {L} _ {\mathrm {p d c - s e m}} + \gamma_ {3} \mathcal {L} _ {\text {s c a l e}} + \gamma_ {4} \mathcal {L} _ {\text {g e o}}. \tag {11}
+$$
+
+# 4.3. Optimization
+
+We train our deformation and template fields with a reconstruction loss and several regularizers including our deformation consistency regularization. Following [2], the reconstruction loss  $\mathcal{L}_{\mathrm{rec}}$  for shape  $P$  is
+
+$$
+\begin{array}{l} \mathcal {L} _ {\text {r e c}} = \sum_ {\boldsymbol {x} \in \Omega_ {P}} | \mathcal {F} _ {\theta} (\boldsymbol {x}) - \bar {s} | + \sum_ {\boldsymbol {x} \in P} (1 - \langle \nabla \mathcal {F} _ {\theta} (\boldsymbol {x}), \bar {n} \rangle) \\ + \sum_ {\boldsymbol {x} \in \Omega_ {P}} | \| \nabla \mathcal {F} _ {\theta} (\boldsymbol {x}) \| _ {2} - 1 | + \sum_ {\boldsymbol {x} \in \Omega_ {P} \backslash P} \exp (- \delta \cdot | \mathcal {F} _ {\theta} (\boldsymbol {x}) |), \tag {12} \\ \end{array}
+$$
+
+where  $\bar{s}$  denotes the ground-truth SDF value of each query point,  $\bar{n}$  is the ground truth surface normal vector of each surface point and  $\nabla \mathcal{F}_{\theta}$  the spatial derivative of a 3D field. The second term of (12) is derived from the property of the SDF function that the gradient of an SDF is identical to the surface normal for surface points and the third is implicit geometric regularization [54] derived from the Eikonal equation. The last term pertains to penalizing off-surface points to be away from the surface points, which is controlled by  $\delta \gg 1$ .
+
+Since we do not know ground truth mapping between the template and input shapes, we manipulate the deformation field through multiple regularizers. First, we employ our proposed deformation consistency regularization (11). For further regularizers, we adopt [2] such as deformation smoothness  $\mathcal{L}_{\mathrm{smooth}}$ , normal consistency  $\mathcal{L}_{\mathrm{normal}}$  and minimal correction  $\mathcal{L}_{\mathrm{c}}$ , which are defined as
+
+$$
+\mathcal {L} _ {\text {s m o o t h}} = \sum_ {\boldsymbol {x} \in \Omega_ {P}} \sum_ {d \in \{X, Y, Z \}} \| \nabla \mathcal {D} _ {\theta_ {1}} ^ {v} (\cdot , \boldsymbol {z} _ {P}) | _ {d} (\boldsymbol {x}) \| _ {2} ^ {2}, \tag {13}
+$$
+
+![](images/da82899a2c9ce4b1dc7fc4cccec1d8fceb2404da3d183bb1d68551ee615a29cb.jpg)  
+Figure 6. Part label transfer result on ShapeNet. We transfer part segment labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations, please refer to the supplement.
+
+$$
+\mathcal {L} _ {\text {n o r m a l}} = \sum_ {i} \sum_ {\boldsymbol {x} \in \mathcal {S} _ {i}} \left(1 - \left\langle \nabla \mathcal {T} _ {\theta_ {2}} \left(\boldsymbol {x} + D _ {\theta_ {1}} ^ {v} (\boldsymbol {x})\right), \bar {n} \right\rangle\right), \tag {14}
+$$
+
+$$
+\mathcal {L} _ {\mathrm {c}} = \sum_ {i} \sum_ {\boldsymbol {x} \in \Omega_ {P}} | \mathcal {D} _ {\theta_ {1}} ^ {\Delta s} (\boldsymbol {x}) |. \tag {15}
+$$
+
+We also apply  $\ell_2$  regularization on shape latent codes and part deformation priors
+
+$$
+\mathcal {L} _ {\mathrm {e m b}} = \sum_ {i} ^ {N} \| z _ {i} \| _ {2} ^ {2} + \sum_ {j} ^ {k} \| e _ {j} \| _ {2} ^ {2}. \tag {16}
+$$
+
+The regularizers are collectively written as
+
+$$
+\mathcal {L} _ {\text {r e g}} = \gamma_ {5} \mathcal {L} _ {\text {s m o o t h}} + \gamma_ {6} \mathcal {L} _ {\text {n o r m a l}} + \gamma_ {7} \mathcal {L} _ {\mathrm {c}} + \gamma_ {8} \mathcal {L} _ {\text {e m b}}. \tag {17}
+$$
+
+Hence, the final loss becomes
+
+$$
+\mathcal {L} = \mathcal {L} _ {\text {r e c}} + \mathcal {L} _ {\mathrm {d c}} + \mathcal {L} _ {\text {r e g}}. \tag {18}
+$$
+
+Evaluating dense correspondence. Given two pairs of seen or unseen shapes  $P, Q$ , we can evaluate the dense correspondence between  $P$  and  $Q$  by deforming each shape to the common template space. Moreover, the uncertainty of dense correspondence performance between two points  $p \in P$  and  $q \in Q$  can be measured as below:
+
+$$
+u (p, q) = 1 - \exp (- \gamma \| (p + \Delta p) - (q + \Delta q) \| _ {2} ^ {2}). \tag {19}
+$$
+
+$\gamma$  is a hyperparameter for modulation and  $\Delta p, \Delta q$  are output from  $\mathcal{D}_{\theta_1}$ . Note that since each part deformation prior  $e$  is shared based on part semantics across entire training shapes, we do not need a test-time optimization on  $e$  for an unseen shape at inference time.
+
+# 5. Experiments
+
+We demonstrate the effectiveness of our proposed framework and provide both quantitative and qualitative results. In Section 5.1, we conduct multiple label transfer
+
+Table 2. Keypoint correspondence accuracy in KeypointNet [55]. PCK scores with threshold of  $0.05 / 0.1$  are reported.  
+
+<table><tr><td>PCK</td><td>airplane</td><td>car</td><td>chair</td><td>table</td></tr><tr><td>AtlasNetV2 [34]</td><td>23.9 / 45.3</td><td>33.5 / 64.0</td><td>25.6 / 50.2</td><td>25.3 / 48.0</td></tr><tr><td>DIT [2]</td><td>13.6 / 34.2</td><td>58.9 / 87.7</td><td>33.3 / 66.7</td><td>26.2 / 55.5</td></tr><tr><td>DIF [1]</td><td>16.5 / 40.5</td><td>67.8 / 90.4</td><td>34.3 / 57.6</td><td>37.9 / 62.9</td></tr><tr><td>Ours</td><td>24.2 / 52.3</td><td>69.3 / 91.8</td><td>61.2 / 83.2</td><td>50.8/77.9</td></tr></table>
+
+Table 3. 5-shot part label transfer in four categories of ShapeNet. The reported numbers are overall mIoU.  
+
+<table><tr><td>model</td><td>airplane</td><td>car</td><td>chair</td><td>table</td></tr><tr><td>AtlasNetV2 [34]</td><td>57.8</td><td>59.3</td><td>66.9</td><td>70.6</td></tr><tr><td>DIT [1]</td><td>64.2</td><td>66.1</td><td>80.4</td><td>86.0</td></tr><tr><td>DIF [2]</td><td>73.9</td><td>69.6</td><td>82.1</td><td>86.6</td></tr><tr><td>Ours</td><td>75.1</td><td>70.4</td><td>84.2</td><td>86.9</td></tr></table>
+
+tasks for dense correspondence evaluation. In Section 5.2, we measure shape reconstruction accuracy to evaluate the representational power for unknown shapes. In Section 5.3, we present advanced dense correspondence analysis under more challenging settings. Lastly, in Section 5.4, we perform ablation studies to analyze the proposed regularization and the sensitivity over feature extractor [31].
+
+Datasets and Baselines. We train and validate our model and baselines on four categories (e.g., chair, table, airplane and car) of ShapeNetV2 [50]. We follow the training split of DIF [2] for each category. For label transfer tasks, we use the labels from ShapeNet-Part [56] and KeypointNet [55]. DIT [1] and DIF [2] are the most relevant baseline models as they both globally learn implicit template fields, and we select AtlasNetV2 [34] as an additional baseline. Dense correspondence is measured with local patch composition in [34]. For more implementation details, analyses, and visualizations, see the supplement.
+
+# 5.1. Dense correspondence
+
+To evaluate the quality of learned dense correspondence, we conduct various surrogate tasks including keypoint transfer, part label transfer, and texture transfer due to
+
+![](images/d992a9a2429a7e8c487f80fb2c12c05698fa643a016b4a70d35d2eca80ee5f3e.jpg)  
+Figure 7. Texture transfer result on ShapeNet. We transfer part segment labels from source shapes (colored in ivory) to target shapes and compare ours with DIT [1] and DIF [2]. For additional visualizations, please refer to the supplement.
+
+Table 4. Reconstruction accuracy for unseen shapes in four categories of ShapeNet. The reported numbers are chamfer distance (CD) between ground truth and reconstructed meshes at a resolution of  $256^{3}$ .  
+
+<table><tr><td>CD (×103)↓</td><td>airplane</td><td>car</td><td>chair</td><td>table</td></tr><tr><td>DIT [1]</td><td>0.2261</td><td>0.4082</td><td>0.3063</td><td>0.6187</td></tr><tr><td>DIF [2]</td><td>0.2093</td><td>0.3157</td><td>0.4229</td><td>0.2640</td></tr><tr><td>Ours</td><td>0.1859</td><td>0.3150</td><td>0.2265</td><td>0.2796</td></tr></table>
+
+the lack of labeled datasets for 3D dense correspondence. For the following tasks, we transfer respective labels from source shapes to target shapes in a common template space.
+
+Keypoint label transfer. Table 2 presents the quantitative result in a 5-shot keypoint transfer task. We measure the performance by a percentage of correct keypoints (PCK) [57]. We compute the distance between the transferred keypoints and the ground truth points to record the PCK score with the threshold of  $0.05 / 0.1$ . Our method achieves the best performance in all four categories with a significant improvement. Especially for categories with high structural variability (e.g., chair and table), a large performance gain is observed. For instance, even compared to the best baseline, on the chair category with the threshold of 0.1, our method (83.2) outperforms DIT [1] (66.7) by  $16.5\%$ , and on the table category with the threshold of 0.1, our method (77.9) outperforms DIF [2] (62.9) by  $15\%$ . We additionally provide the visualization of keypoint transfer in Figure 4, which also support the importance of understanding part semantics for shape correspondence. In particular, two keypoints of pea-green and orange from the source chair are geometrically close but semantically different since one is for the arm part and the other is for the seat part. Among several methods, only our framework succeeds in disentangling two keypoints in the target chair. We provide more qualitative results regarding keypoint transfer results in real-world data [53] to validate the robustness of our method in the supplement.
+
+Part label transfer. For few-shot part label transfer tasks, we transfer part segment labels from source shapes to target shapes. In detail, for each (target) point of the deformed target shape, we find the  $n$  nearest points of the deformed source shapes and perform weighted voting using their labels and distances to the target point. In our experiments, we used  $n = 10$  nearest points. For a 5-shot part label transfer task in Table 3, five source shapes were used just the same as [2] for every category. Table 3 shows our framework achieves the best performance over all four categories. Moreover, Figure 6 illustrates the qualitative results of one-shot part label transfer on ShapeNet [50]. Similar to keypoint transfer, Figure 6 also proves the effectiveness of semantic knowledge in dense correspondence.
+
+Texture transfer. Figure 7 shows texture transfer results on ShapeNetV2 [50] to compare established dense correspondence between baselines and ours with higher-frequency label transfer. We observe notable texture transfer results. For instance, our framework successfully transfers the source car's texture of the windshield (gray) to the corresponding part of the target cars, whereas baselines transfer the same textures to the roof part of the target cars.
+
+# 5.2. Shape reconstruction
+
+We report the reconstruction quality of unseen shapes with Chamfer Distance (CD) to validate the representational power of our framework in Table 4. A smaller CD means more accurate reconstruction. As in previous work, we report the  $\mathrm{CD} \times 10^{3}$  for readability. Also, we consistently discuss the CD and performance gain below at this scale. We optimize the corresponding shape latent code for a given unseen shape. We measure CD as in DIF [2], where reconstructed meshes are extracted at a resolution of  $256^{3}$ , and CD is calculated with  $10K$  sampled points. Our framework on average achieved the best reconstruction results surpassing DIF by a margin of 0.0512. On average, our method and DIF exhibit 0.2518 and 0.3030 reconstruction errors, respectively. In three out of four categories, our method achieved the best reconstruction accuracy.
+
+# 5.3. More challenging setups
+
+To further validate the effectiveness of our method, we perform additional experiments and compare the results with DIF [2]. Overall, the results below imply current methods that lack part semantics and do not consider part deformation consistency may fail to learn semantically plausible deformation in challenging cases.
+
+2-shot part label transfer within different subcategories. Our objective is to assess dense correspondence performance under difficult settings. We conduct 2-shot part label transfer within two different subcategories belonging to the same category, where shapes from different subcategories tend to have different geometry for same semantics. Since ShapeNet [50] includes incorrect subcategory labels [58], we manually cleaned subcategory labels for four categories (e.g., airplane: airliner/jet, car: sedan/jeep, chair: straight chair/sofa, table: pedestal table/short table). Figure 8 displays examples of shapes from each subcategory. For robust evaluations, we randomly select the source shape 5 times for each subcategory. Table 5 shows our framework achieves a significant performance gain in all cases by over  $7.2\%$  on average compared to DIF [2].
+
+Unsupervised dense correspondence in DFAUST [59]. Furthermore, we evaluate unsupervised dense correspondence with non-rigid shapes [59]. For more challenging settings, we train both DIF [2] and our model for two different subjects with different sizes and heights (e.g., subject ID of 50002, 50026) and four distinct actions (e.g., shake arms, chicken wings, running on spot, and jumping jacks). Colormap in Figure 9 demonstrates the correspondence quality, where we transfer color over the up-axis. In the source shape, the arms are labeled in red. After transferring these color labels to the target shapes, our method preserves the red/orange labeling of the arms across different motions, whereas the baseline fails to capture the same labeling (i.e., yellow hands) since it lacks semantic knowledge during training. Additionally, we conduct keypoint transfer tasks and measure the PCK score as identical as in Section 5.1. Our model achieves 74.9/85.2 with a threshold of 0.05/0.1, whereas DIF [2] achieves 67.0/80.7 with the same threshold, showing our superior performance in the challenging settings (see the supplement for the full keypoint transfer result). Note that we can also learn implicit templates for non-rigid shapes with supervision as in [19]. We conduct experiments in unsupervised settings since our focus is to validate the effectiveness of utilizing additional semantic information for template learning.
+
+# 5.4. Ablations
+
+Ablation on regularizations. We conduct ablation studies regarding the proposed regularizations including part deformation consistency regularizations, global scale consistency regularization, and global deformation consistency
+
+![](images/6cd10adf61d83a66ca80750a8786e2ac526e6ea83495786caa65bca3c3b03ef5.jpg)  
+Figure 8. Types of source shapes for 2-shot part label transfer.
+
+Table 5. 2-shot part label transfer in subcategories of ShapeNet. The reported numbers are overall mIoU over 5 runs.  
+
+<table><tr><td>subcategories</td><td>DIF [2]</td><td>Ours</td></tr><tr><td>airplane(airliner/jet)</td><td>64.5±1.61</td><td>72.8±0.99 (+8.3↑)</td></tr><tr><td>car(sedan/jeep)</td><td>65.1±2.34</td><td>69.1±1.46 (+4.0↑)</td></tr><tr><td>chair(straight/sofa)</td><td>66.3±6.06</td><td>77.5±5.81 (+11.2↑)</td></tr><tr><td>table(pedestal/short)</td><td>74.4±10.1</td><td>79.7±6.26 (+5.3↑)</td></tr></table>
+
+![](images/87a877925235bd64d7f0802809ba8f9a944ff6a557b07ab22f5af790ef6063d1.jpg)  
+Figure 9. Color transfer on DFAUST [59]. We transfer different color labels over the up-axis from source shapes (colored in ivory) to target shapes and compare ours with DIF [2].
+
+Table 6. Ablations for every proposed regularization. We report mIoU for 5-shot part label transfer on ShapeNet chair. Note that  $\mathcal{L}_{\mathrm{pdc}}$  denotes  $\mathcal{L}_{\mathrm{pdc\_geo}} + \mathcal{L}_{\mathrm{pdc\_sem}}$ .  
+
+<table><tr><td></td><td>Lpdc</td><td>Lgeo</td><td>Lscale</td><td>chair</td></tr><tr><td>(a)</td><td>✓</td><td></td><td></td><td>83.4</td></tr><tr><td>(b)</td><td>✓</td><td>✓</td><td></td><td>83.9</td></tr><tr><td>(c)</td><td>✓</td><td>✓</td><td>✓</td><td>84.2</td></tr></table>
+
+Table 7. BAE-Net  $k$  ablation. We report mIoU for 5-shot part label transfer on ShapeNet chair over 3 runs.  
+
+<table><tr><td></td><td>k=2</td><td>k=4</td><td>k=6</td><td>k=8</td><td>k=10</td></tr><tr><td>chair</td><td>83.8±0.51</td><td>83.5±1.48</td><td>82.8±1.27</td><td>83.6±0.85</td><td>83.7±1.33</td></tr></table>
+
+regularization. We conduct 5-shot part label transfer within a chair category, as same as Section 5.1. Table 6 shows that each component of our proposed regularizers contributes to performance improvement for dense correspondence. Note that by only using  $\mathcal{L}_{\mathrm{pdc}}$ , i.e., using  $\mathcal{L}_{\mathrm{pdc\_geo}}$  and  $\mathcal{L}_{\mathrm{pdc\_sem}}$ , it still outperforms than baselines (80.4 [1] and 82.1 [2]). The best performance was achieved by incorporating all proposed regularizations.
+
+BAE-Net [31] with different  $k$ . We evaluate the effect of  $k$  for BAE-Net [31] with mean/standard deviation over 3 runs for ShapeNet [50] chair in We train BAE-Net [31] with a range of  $k = (2,4,6,8,10)$ , where  $k = 4$  is used in the main paper, and report mIoU for 5-shot part label
+
+transfer task. Table 7 shows the efficacy of semantic priors, where our framework achieves the average mIoU of 83.5, which shows higher performance than baselines. More importantly, we observed that the performance is fairly robust over  $k$ .
+
+# 6. Conclusion
+
+We have proposed a new framework that learns a semantic-aware implicit template field. Our framework extracts semantic knowledge from a self-supervised feature extractor given unlabelled point clouds. By leveraging the semantic information, we propose novel conditioning with semantic-aware deformation code and part deformation consistency regularizations to encourage semantically plausible deformation. Furthermore, we suggest global scaling consistency and global deformation consistency regularizations to enhance the learning process. Through quantitative and qualitative evaluation, we demonstrate the superiority of our proposed method and verified the effect of all proposed components in our framework.
+
+Acknowledgments. This research was supported in part by the MSIT (Ministry of Science and ICT), Korea, under the ICT Creative Consilience program (IITP-2023-2020-0-01819) supervised by the IITP (Institute for Information & communications Technology Planning & Evaluation); the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF-2023R1A2C2005373); and Google Cloud Research Credits program with the award (GCLF-67WL-WU6D-MF14, V9P3-99XD-64CL-EH5H).
+
+# References
+
+[1] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3d shape representation. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8  
+[2] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8  
+[3] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. TOG, 2015. 1, 2  
+[4] Brett Allen, Brian Curless, and Zoran Popovic. Articulated body deformation from range scan data. TOG, 2002. 1  
+[5] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. In ECCV, 2018. 1, 2  
+[6] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In CVPR, 2018. 1  
+[7] Wang Zeng, Wanli Ouyang, Ping Luo, Wentao Liu, and Xiaogang Wang. 3d human mesh regression with dense correspondence. In CVPR, 2020. 1, 2
+
+[8] Sing Chun Lee and Misha Kazhdan. Dense point-to-point correspondences between genus-zero shapes. In Computer Graphics Forum, 2019. 1, 2  
+[9] Silvia Zuffi, Angjoo Kanazawa, David W. Jacobs, and Michael J. Black. 3d menagerie: Modeling the 3d shape and pose of animals. In CVPR, 2017. 1, 2  
+[10] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In AVSS, 2009. 1  
+[11] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. TOG, 2017. 1  
+[12] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 1  
+[13] Srinath Sridhar, Antti Oulasvirta, and Christian Theobalt. Interactive markerless articulated hand motion tracking using rgb and depth data. In ICCV, 2013. 1  
+[14] Iason Oikonomidis, Manolis IA Lourakis, and Antonis A Argyros. Evolutionary quasi-random search for hand articulations tracking. In CVPR, 2014. 1  
+[15] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 1, 2  
+[16] Vignesh Ganapathi-Subramanian, Olga Diamanti, Soeren Pirk, Chengcheng Tang, Matthias Niessner, and Leonidas Guibas. Parsing geometry using structure-aware shape templates. In 3DV, 2018. 1, 2  
+[17] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation. In CVPR, 2018. 1, 2  
+[18] Pablo Palafox, Nikolaos Sarafianos, Tony Tung, and Angela Dai. Spams: Structured implicit parametric models. In CVPR, 2022. 1  
+[19] Ramana Sundararaman, Gautam Pai, and Maks Ovsjanikov. Implicit field supervision for robust non-rigid shape matching. In ECCV, 2022. 1, 8  
+[20] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1, 2  
+[21] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. 2019. 1, 3  
+[22] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 1, 2, 3  
+[23] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 1, 2, 3
+
+[24] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, 2020. 1  
+[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2  
+[26] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 2  
+[27] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In CVPR, 2021. 1, 2  
+[28] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 1, 2  
+[29] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In CVPR, 2021. 1, 2  
+[30] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 1, 2  
+[31] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. In ICCV, 2019. 2, 4, 6, 8  
+[32] Vladimir G Kim, Wilmot Li, Niloy J Mitra, Siddhartha Chaudhuri, Stephen DiVerdi, and Thomas Funkhouser. Learning part-based templates from large collections of 3d shapes. TOG, 2013. 2  
+[33] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T. Freeman, and Thomas A. Funkhouser. Learning shape templates with structured implicit functions. In ICCV, 2019. 2  
+[34] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3d shape generation and matching. In NeurIPS, 2019. 2, 6  
+[35] Jiahui Lei and Kostas Daniilidis. Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In CVPR, 2022. 2  
+[36] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural template: Topology-aware reconstruction and disentangled generation of 3d meshes. In CVPR, 2022. 2  
+[37] Zekun Hao, Hadar Averbuch-Elor, Noah Snavely, and Serge Belongie. Dualsdf: Semantic shape manipulation using a two-level representation. In CVPR, 2020. 2  
+[38] Yueqi Duan, Haidong Zhu, He Wang, Li Yi, Ram Nevatia, and Leonidas J. Guibas. Curriculum deepsdf. In ECCV, 2020. 2
+
+[39] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2  
+[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2  
+[41] Michael Niemeyer and Andreas Geiger. Campari: Camera-aware decomposed generative neural radiance fields. In 3DV, 2021. 2  
+[42] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, et al. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2  
+[43] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. Rimnet: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In CVPR, 2022. 2  
+[44] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J Guibas, and Hao Zhang. Adacoseg: Adaptive shape co-segmentation with group consistency loss. In CVPR, 2020. 2  
+[45] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Part slip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In CVPR, 2023. 2  
+[46] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2  
+[47] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 2  
+[48] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In NeurIPS, 2022. 2  
+[49] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2022. 2  
+[50] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 4, 6, 7, 8  
+[51] David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. In ICLR, 2017. 4  
+[52] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 4  
+[53] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 5, 7
+
+[54] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 5  
+[55] Yang You, Yujing Lou, Chengkun Li, Zhoujun Cheng, Liangwei Li, Lizhuang Ma, Cewu Lu, and Weiming Wang. Keypointnet: A large-scale 3d keypoint dataset aggregated from numerous human annotations. In CVPR, 2020. 6  
+[56] Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3d shape collections. TOG, 2016. 6  
+[57] Li Yi, Hao Su, Xingwen Guo, and Leonidas Guibas. Syncspeccnn: Synchronized spectral cnn for 3d shape segmentation. In CVPR, 2017. 7  
+[58] Xinhai Liu, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Fine-grained 3d shape classification with hierarchical part-view attention. IEEE Transactions on Image Processing, 2021. 8  
+[59] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In CVPR, 2017. 8

2023/Semantic-Aware Implicit Template Learning via Part Deformation Consistency/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d19d7439e6b01d2d6515880d933d28b0044d57274205d157f166d02f3018aa0e
+size 603754

2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/e3cccf23-d23a-4b30-90e1-8c04059b69dd_content_list.json

@@ -0,0 +1,2001 @@
+[
+    {
+        "type": "text",
+        "text": "Semantically Structured Image Compression via Irregular Group-Based Decoupling",
+        "text_level": 1,
+        "bbox": [
+            238,
+            130,
+            733,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ruoyu Feng $^{1*}$  Yixin Gao $^{1*}$  Xin Jin $^{2}$  Runsen Feng $^{1}$  Zhibo Chen $^{1,\\dagger}$",
+        "bbox": [
+            194,
+            202,
+            774,
+            222
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>University of Science and Technology of China <sup>2</sup>Eastern Institute of Technology, Ningbo",
+        "bbox": [
+            210,
+            223,
+            759,
+            239
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            273,
+            313,
+            289
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Image compression techniques typically focus on compressing rectangular images for human consumption, however, resulting in transmitting redundant content for downstream applications. To overcome this limitation, some previous works propose to semantically structure the bitstream, which can meet specific application requirements by selective transmission and reconstruction. Nevertheless, they divide the input image into multiple rectangular regions according to semantics and ignore avoiding information interaction among them, causing waste of bitrate and distorted reconstruction of region boundaries. In this paper, we propose to decouple an image into multiple groups with irregular shapes based on a customized group mask and compress them independently. Our group mask describes the image at a finer granularity, enabling significant bitrate saving by reducing the transmission of redundant content. Moreover, to ensure the fidelity of selective reconstruction, this paper proposes the concept of group-independent transform that maintain the independence among distinct groups. And we instantiate it by the proposed Group-Independent Swin-Block (GI Swin-Block). Experimental results demonstrate that our framework structures the bitstream with negligible cost, and exhibits superior performance on both visual quality and intelligent task supporting.",
+        "bbox": [
+            76,
+            306,
+            473,
+            684
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            715,
+            209,
+            729
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Image signal serves as a critical information carrier for various applications in modern society. Image compression techniques aim at converting images into compact representations (i.e. bitstreams) to save transmission and storage resources. Lossy image compression is one of the most practical techniques, as it allows for the restoration of important content while discarding a small amount of inessential information. In the past decades, traditional image compres",
+        "bbox": [
+            75,
+            741,
+            470,
+            863
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/82944d9ba34b0a08d6b1cec5006054c78948fdbab5baaf9006f2c024e98a9be7.jpg",
+        "image_caption": [
+            "Figure 1. The input image is decoupled into groups according to distinct semantics. Then the semantically structured bitstream (SSB) is generated by compressing the image based on the partitioned groups. The SSB facilitates downstream applications by selective bitstream transmission and partial reconstruction, depending on the specific task requirements."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            270,
+            890,
+            470
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "sion standards [10,54,61,65,67] have been extensively studied and utilized. With the fast development of deep learning, neural image codes [7,8,16,27,34,40,41,50-52,75] rapidly evolved and achieved promising results. Meanwhile, more and more media contents tend to be handled by machine vision algorithms, such as recognition [17,26,29,32,47], detection [43,44,55-57], and segmentation [5,9,11,12,28,46,48,69,73]. However, most compression methods are mainly developed for compressing regular rectangular images for human consumption, without considering the efficiency and functionality for downstream tasks or human-machine interaction scenarios.",
+        "bbox": [
+            496,
+            623,
+            893,
+            805
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recently, the field of image coding for machines (ICM) has emerged to develop a joint efficient and analytical framework for supporting intelligent analytics. End-to-end optimization of the trade-off between specific task loss and compression rate is a promising way [4, 13, 14, 18, 19, 38, 39, 49, 60], but it lacks generalization for massively diverse",
+        "bbox": [
+            496,
+            809,
+            893,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            238,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* First two authors contributed equally.",
+        "bbox": [
+            76,
+            875,
+            287,
+            887
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "† Corresponding author.",
+        "bbox": [
+            80,
+            888,
+            209,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "17237",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "applications. To overcome this limitation, Feng et al. [21] propose to compress general and compact features learned by self-supervised learning under entropy constraints for supporting downstream tasks. Nevertheless, it requires retraining the task model with the proposed features as inputs, which makes the overall performance heavily depend on the effectiveness of feature extraction. The methods mentioned above are designed specifically for scenarios of compression for machine vision without considering situations in which human involvement is required. Semantically structured image compression (SSIC) [62] proposes to generate a semantically structured bitstream (SSB) by separately compressing rectangular regions of detected objects using a pre-prepared object detection toolbox. Although SSB is efficient in supporting intelligent tasks and human-machine interaction through partial transmission and reconstruction of the bitstream, its division approach based on rectangular regions may encounter problems with overlapped objects. SSIC addresses this issue by replacing the overlapping objects with a larger rectangular region, which can result in a waste of bitrate. In addition, SSIC generates the bitstream of each object by compressing the corresponding latent variables in the latent domain directly, without considering the interaction and dependence of features during the transform process. This can lead to blurry and distorted group boundaries in the partial reconstruction scenario, which in turn affects the reconstruction quality.",
+        "bbox": [
+            76,
+            90,
+            472,
+            500
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Going beyond the rectangular-based division, this paper proposes to decouple the image into multiple groups with irregular shapes based on a customized group mask. Then the SSB is generated by independently compressing these groups and can support various requirements via selective transmission and reconstruction. Notably, the generation of the group mask offers high flexibility in terms of shapes of groups, means of pre-analysis, and partition criteria, enabling customization to suit diverse application scenarios and requirements. Moreover, to avoid potential quality degradation in the partial reconstruction scenario, we propose the concept of the group-independent transform, which ensures the independence among groups in the latent representations and therefore the quality of the selective reconstruction will not be affected by the absence of other groups. More specifically, we instantiate it by carefully designing the Group-Independent Swin-Block (GI Swin-Block), which is an extension of Swin Transformer [47] tailored to our situation and requirements. GI SwinBlocks make use of the hierarchical modeling capability of the Swin-Transformer [47, 75], achieving high coding efficiency under the premise of group independence. By combining the group mask based decoupling and the Group-Independent Swin-Block, our proposed method can efficiently support various downstream applications including human-machine interaction and machine vision tasks with",
+        "bbox": [
+            76,
+            508,
+            472,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "only one bitstream generated.",
+        "bbox": [
+            500,
+            90,
+            697,
+            106
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The main contributions of our approach are summarized as follows:",
+        "bbox": [
+            500,
+            106,
+            890,
+            135
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose to decouple an image into multiple groups with irregular shapes for structuring the bitstream. Our group mask can describe spatial division at a finer granularity manner than a typical rectangle, saving bitrate by reducing redundant content transmission.",
+            "- We propose the group-independent transform and instantiate it by carefully designing the Group-Independent Swin-Block (GI Swin-Block), which maintains powerful transformation capability and ensures the independence among groups in the latent representations.",
+            "- Experimental results demonstrate that our proposed model achieves state-of-the-art compression ability and superior downstream tasks performance, which is a codec with both high compression efficiency and functionality."
+        ],
+        "bbox": [
+            516,
+            143,
+            890,
+            401
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Works",
+        "text_level": 1,
+        "bbox": [
+            500,
+            409,
+            650,
+            425
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Image Compression",
+        "text_level": 1,
+        "bbox": [
+            500,
+            434,
+            689,
+            450
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Traditional Image Compression. Traditional image compression standards, such as JPEG [65], JPEG2000 [54], HEVC [61], and VVC Intra [10], have been extensively used in practice after several decades of development. These standards rely on transform coding [23], which decomposes the lossy image compression task into three parts: transform, quantization, and entropy coding. Each module of these standards is manually designed with multiple modes, and rate-distortion optimization is performed to determine the optimal mode. However, the completely hand-crafted structure of traditional CODECs limits their flexibility and scalability to support various objectives, such as MS-SSIM and classification accuracy, as they cannot be optimized in an end-to-end manner.",
+        "bbox": [
+            496,
+            460,
+            890,
+            671
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Learned Image Compression. In recent years, learned image compression methods based on nonlinear transform coding [6] have achieved rapid progress. Early works in this area concentrated on enabling end-to-end training by developing differential quantization and rate estimation techniques [1, 7, 63]. Subsequently, a considerable amount of work design powerful neural network modules to enhance the transform [16, 75], quantization [25, 70], and entropy model [8, 27, 52, 53]. As a result, some neural image CODECs [16, 24, 27, 52, 53] can achieve comparable or even superior performance to traditional coding standards like HEVC [61] and VVC [10]. Moreover, some works [2, 51, 58] introduce perceptual loss to effectively improve reconstruction quality. However, existing image compression methods focus on compressing the entire image",
+        "bbox": [
+            496,
+            674,
+            890,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "17238",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "without considering selective transmission and reconstruction of the bitstream for arbitrary reconstruction, leading to significant bandwidth waste when addressing downstream applications with different requirements.",
+        "bbox": [
+            76,
+            90,
+            468,
+            151
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2. Image Coding for Machines",
+        "text_level": 1,
+        "bbox": [
+            76,
+            162,
+            326,
+            178
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Image coding for machines aims at compress source images to serve downstream tasks, such as recognition [17, 26, 29, 32, 47], detection [43, 44, 55-57], and segmentation [5, 9, 11, 12, 28, 46, 48, 69, 73]. Joint optimization of task loss and bitrate [3, 31, 38, 42, 66] in an end-to-end manner is a natural and promising way. Another direction is compress the features corresponding to downstream tasks [4, 13, 14, 18, 19, 49, 60]. However, these methods tend to be biased towards specific tasks and may lack generalizability to a wide range of intelligent applications. To address the limitation of generalization, a novel representation learning based method is proposed [21] to learn general and compact features by combining contrastive learning with entropy constraints. The learned features are used to replace the original images as the new source for compression and transmission, resulting in a significant improvement in coding efficiency across various intelligent tasks. Nevertheless, this approach necessitates re-training the task model using the proposed features as input, and its overall performance is heavily reliant on the effectiveness of feature extraction. In this paper, we propose a unified compression framework for machine vision support without task-oriented joint training or additional feature learning. Moreover, it can also conduct full or selective reconstruction for human perception according to requirements.",
+        "bbox": [
+            76,
+            186,
+            470,
+            564
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3. Object-oriented Image Compression",
+        "text_level": 1,
+        "bbox": [
+            76,
+            574,
+            393,
+            590
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Object-oriented image compression is first proposed in MPEG-4 Visual [20, 35, 59] by compressing visual object planes (VOP) with arbitrary shapes, targeting content-based interactivity and scalability for the image compression technique. However, due to the difficulty in annotating the complex and intricate shapes of the target and the lack of actual corresponding downstream applications, MPEG-4 Visual has not been widely used. Recently, with the fast development of deep learning [29, 64], the practical scenarios of image compression for downstream applications have increased rapidly. Sun et al. [33, 62] propose the semantically structured image compression (SSIC) of decoupling the image into objects and compressing the corresponding latent variables independently of the neural image codec to generate the semantically structured bitstream. Nevertheless, the rectangular-based division adopted by SSIC lacks generalization to irregular objects and flexibility for customization. Additionally, selecting the representation of objects directly in the latent space ignores the interaction among elements during the encoding transform process, leading to",
+        "bbox": [
+            76,
+            599,
+            468,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "significant edge distortion of objects in partial reconstruction. This paper proposes to conduct semantically structured image compression by decoupling the image into irregular groups based on the group mask, which is flexible and customizable. Furthermore, we propose the group-independent transform conducted during both the encoding and decoding processes, which enables efficient partial or complete reconstruction of images under various requirements.",
+        "bbox": [
+            496,
+            90,
+            890,
+            226
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            500,
+            241,
+            591,
+            257
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Overview",
+        "text_level": 1,
+        "bbox": [
+            500,
+            267,
+            609,
+            282
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The proposed method is an efficient and flexible version of semantically structured image compression [62]. It can satisfy multiple application requirements via transmitting and reconstructing partial spatial regions guided by a block-wise group mask. Our model takes the group mask as guidance to ensure that interactions of transform only occur within the same group, thereby achieving group independence during the redundancy removal process. The group mask, which is generated based on the pre-analysis, such as object detection, semantic segmentation, and saliency detection, provides high flexibility and customization for structuring the bitstream and is considered as side information. Then the entropy coding process is conducted on the variables of each group distinctly, resulting in a bitstream structured by semantics. The bitstream can be partially or fully transmitted according to the requirements of the decoder side, and then the receiver conducts entropy decoding on the bitstream and reorganizes the spatial-wise arrangement of latent variables according to the group mask and the group indexes.",
+        "bbox": [
+            496,
+            291,
+            890,
+            592
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The overall network structure is illustrated in Fig. 2, which incorporates the group mask and GI Swin-Block for guiding and instantiating the group-independent transform, respectively. Section 3.2 and Section 3.3 introduce the group mask and group independent, respectively. The detailed implementation of our network architecture is presented in Section 3.4.",
+        "bbox": [
+            496,
+            593,
+            890,
+            699
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Group Mask",
+        "text_level": 1,
+        "bbox": [
+            500,
+            710,
+            635,
+            726
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Before compression, the group mask is generated according to the results of pre-analysis techniques such as object detection, instance segmentation, and saliency detection. Figure 3 provides an example of this process. More specifically, for the input image  $x$  with a height of  $H$  and a width of  $W$ , the spatial resolution of the group mask is the same as the input image, and it consists of  $\\frac{H}{B} \\times \\frac{W}{B}$  blocks, where  $B$  is the side length of each block. And the value of it indicates the index of the corresponding group.",
+        "bbox": [
+            496,
+            734,
+            890,
+            869
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our proposed group mask can clearly divide the elements in the latent variable space after the one-pass down",
+        "bbox": [
+            500,
+            869,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "17239",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/d76eb38af4aa73a7eb406dfec3fa55e911ae3f0cb944f421209f8772f5801d39.jpg",
+        "image_caption": [
+            "Figure 2. The network architecture of our proposed model with the channel-wise auto-regressive model (ChARM). ConvT denotes transposed convolution. AE and AD are respectively arithmetic encoding and arithmetic decoding. In Ours-Hyper model, we remove the ChARM component and instead output  $\\mu$  and  $\\sigma$  directly from the hyperdecoder  $h_s$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            88,
+            893,
+            345
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/d018b1156d0090bc1bf3f1b2206388b15125c66e51427de92ecb2ae225e90d97.jpg",
+        "image_caption": [
+            "Figure 3. An example of group mask generation. The input image is pre-analyzed by object detection, then the group mask is generated based on the results of the pre-analysis."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            425,
+            467,
+            520
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "sampling transform. Besides, downsampling the group mask by a factor of  $B$  before conducting compression can greatly reduce the overhead bitrate. The block size and the overhead bitrate have an inverse relationship. Furthermore, as shown in Figure 4, the generation of the group mask can be flexibly customized, in terms of the way of pre-analysis and the criteria of block allocation, etc.",
+        "bbox": [
+            75,
+            606,
+            468,
+            712
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Group-Independent transform",
+        "text_level": 1,
+        "bbox": [
+            76,
+            724,
+            349,
+            742
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Lossy image compression based on transform coding can be divided into three modularized components: transform, quantization, and entropy coding. While quantization and entropy coding do not affect the generation and usage of the semantically structured bitstream (SSB), traditional transform applied to the entire image to remove spatial redundancy inevitably creates inter-group dependencies during compression. In situations of selective transmission and reconstruction, the incomplete representation that results from the reorganization of a subset of all groups can",
+        "bbox": [
+            75,
+            750,
+            470,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "lead to inaccurate reconstruction due to the lack of intergroup dependencies. Therefore, we propose the concept of the group-independent transform for semantically structured bitstream generation. The basic idea is to constrain the transform to be conducted only within each group.",
+        "bbox": [
+            496,
+            428,
+            890,
+            503
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "A straight and natural way is by using transformer [64] with the customized attention map corresponding to the group mask. However, image compression differs significantly from high-level understanding tasks, which the transformer excels at. The presence of numerous complex long-range dependencies can impede convergence and adversely affect the final performance. In order to achieve high coding efficiency while maintaining the group-independent property, and inspired by the strong hierarchical representation modeling capability of Swin-Transformer [47, 75], we extend the Swin-Block of Swin-Transformer to our proposed group-independent Swin-Block (GI Swin-Block). The GI Swin-Block serves as the core component of the transform. To be more specific, as shown in Figure 5a, by merging the window partition and the group partition into the group-independent window partition, allowing the self-attention to be computed within the partitioned local regions, enabling us to achieve high coding efficiency while maintaining the group-independent property. Additionally, the cross-window connections are introduced by merging the shifted window partition and group partition similarly, as depicted in Figure 5b.",
+        "bbox": [
+            496,
+            503,
+            892,
+            837
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4. Network architecture and pipeline",
+        "text_level": 1,
+        "bbox": [
+            500,
+            845,
+            802,
+            864
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Figure 2 illustrate the proposed architecture. The main transform is performed in a staggered manner through",
+        "bbox": [
+            500,
+            869,
+            893,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "17240",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/64e2a7e5d133e8584b985bb8fc4289ce2b243157912d5b5644d5c8bb2a027ef0.jpg",
+        "image_caption": [
+            "Input Image"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            83,
+            97,
+            151,
+            161
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/847ce4cb612360f0315619348f88a9c1ae9f73c5efdf51227898cba8b0507cf1.jpg",
+        "image_caption": [
+            "Pre-Analysis"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            151,
+            98,
+            218,
+            161
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/7da7c82cd4385373807b1a9982a8b99f7f115051fb63e0034447d20768c16861.jpg",
+        "image_caption": [
+            "(a)",
+            "Possible Group Masks"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            218,
+            97,
+            308,
+            161
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/57afce836be7bf66ddd76947463897e79365a5ec9ab2725cade8d1ad37afb6c0.jpg",
+        "image_caption": [
+            "(b)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            316,
+            97,
+            383,
+            161
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/7a899820c4768ae05823d7a91afd1b4c4e267e1306e2d69a2fa71390d363631d.jpg",
+        "image_caption": [
+            "(c)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            390,
+            97,
+            459,
+            161
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "the incorporation of upsample/downsample and group-independent Swin-Block (GI Swin-Block). Upsample contains a pixel shuffle operation and a convolution layer with  $1 \\times 1$  kernel size, whereas downsample contains a pixel unshuffle operation and a convolution layer with  $1 \\times 1$  kernel size. For entropy coding, we adopt the Mean & Scale (M&S) Hyperprior model [52] with the strong channel-wise auto-regressive model (ChARM) [53] to predict the probability distribution of latent variables.",
+        "bbox": [
+            75,
+            309,
+            468,
+            444
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Specifically, given a source image  $\\pmb{x}$ , the encoder  $g_{a}$  converts it to latent representations  $\\pmb{y}$  conditioned on the given group mask  $\\pmb{m} = \\sum_{i=1}^{N} \\pmb{m}_{i}$ , where  $i \\in \\{1,2,3,\\dots,N\\}$ , indicating the index of the corresponding group in all  $N$  groups. In each  $\\pmb{m}_{i}$ , the values of positions corresponding to the  $i_{th}$  group are set as  $i$ , and all values of other positions are set as 0. Similarly, we can define  $\\pmb{x} = \\sum_{i=1}^{N} \\pmb{x}_{i}$  and  $\\pmb{y} = \\sum_{i=1}^{N} \\pmb{y}_{i}$  from the perspective of group, and transform only occurs within each group, which is given by",
+        "bbox": [
+            76,
+            445,
+            468,
+            583
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol {y} _ {i} = g _ {a} \\left(\\boldsymbol {x} _ {i} \\mid \\boldsymbol {m} _ {i}\\right). \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            210,
+            590,
+            468,
+            607
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Thanks to the group-independent transform, we can easily implement the transform of all groups within one forward-pass, thus, the process can also be written as  $\\pmb{y} = g_{a}(\\pmb{x}|\\pmb{m})$ .",
+        "bbox": [
+            76,
+            617,
+            468,
+            674
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Then the latents are quantized to discrete representations  $\\hat{\\pmb{y}}$  and selectively transmitted to the decoder side with predicted probability distribution through the entropy model. Concretely, the side information  $\\pmb{z}$  is extracted by the hyper-encoder  $\\pmb{z} = h_{a}(\\pmb{y})$ . The quantized hyper-latent  $\\hat{\\pmb{z}} = Q(\\pmb{z})$  is modeled and entropy-coded with a learned factorized prior. In ChARM, latent  $\\hat{\\pmb{y}}$  is split into  $S$  (we choose  $S = 10$ ) slices along with channel dimension, and each slice  $\\hat{\\pmb{y}}_s$  is entropy-coded based on the previous slices  $\\hat{\\pmb{y}}_{< s}$ .  $\\hat{\\pmb{y}}$  is modeled by a conditional Gaussian distribution convolving with a unit uniform noise.",
+        "bbox": [
+            76,
+            674,
+            468,
+            839
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After selectively entropy decoding the quantized latents, we have  $\\hat{\\pmb{y}}_p = \\sum_{j\\in \\mathbf{P}}\\hat{\\pmb{y}}_j$  , where  $\\mathbf{P}\\subseteq \\{1,2,\\dots ,N\\}$  is the gather of required groups. The reconstruction image  $\\hat{x}_p = \\sum_{j\\in \\mathbf{P}}\\hat{x}_j$  will be calculated as follows",
+        "bbox": [
+            76,
+            839,
+            468,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/260198688b7f50c623cc5d37cb6e3268728877286a4ba52d8e5c7cfbd0f3b894.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            84,
+            890,
+            172
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/29995e04e180ef31d391f4849e88039ea00f178dd36534d0563f50a847c3de6a.jpg",
+        "image_caption": [
+            "Figure 4. Illustration of flexibility to customize the group mask: (a). Taking overlapping objects as one group with the guidance of bounding boxes. (b). Taking overlapping objects as one group with the guidance of instance masks. (c) Taking overlapping objects as distinct groups with the guidance of instance masks.",
+            "(a) GI window partition under regular window partition.",
+            "(b) GI window partition under shifted window partition.",
+            "Figure 5. Group-independent window partition of GI SwinBlock under regular and shifted window partition. Self-attention is conducted inside each local region."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            202,
+            893,
+            290
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\boldsymbol {x}} _ {j} = g _ {s} \\left(\\hat {\\boldsymbol {y}} _ {j} \\mid \\boldsymbol {m} _ {j}\\right). \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            629,
+            386,
+            890,
+            405
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Similar to Equation (1), Equation (2) can also be written as  $\\hat{\\pmb{x}} = g_s(\\hat{\\pmb{y}} |\\pmb {m})$ . Note that the rounding operation is non-differentiable, thus we apply uniform noise for learning the entropy model and use a rounded one as the input of the decoder  $g_{s}$  as in [53].",
+        "bbox": [
+            498,
+            412,
+            890,
+            488
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            500,
+            632,
+            517
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Experiment Setup",
+        "text_level": 1,
+        "bbox": [
+            500,
+            525,
+            674,
+            541
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Image Codec. We train our high bitrate models  $(>= 0.6\\mathrm{bpp})$  on the Flickr 2K dataset, which has the same setting as [74] and train our low bitrate models  $(< 0.6\\mathrm{bpp})$  on the COCO 2017 training set [45]. Ours-Hyper are trained with 3.2M iterations, while Ours-ChARM models are trained by initializing their weights from higher bitrate hyperprior models for another 1.2M iterations. Each batch contains 8 random  $256 \\times 256$  crops from the training dataset. The learning rate is set as  $5e - 5$  and is decayed by a factor of 10 at  $2.8M$  iterations. Training loss  $L = R + \\beta D$  is the weighted combination of the rate-distortion trade-off, balanced by the Lagrange multiplier  $\\beta$ . We adopt the mean squared error (MSE) in the RGB color space as the distortion metric  $D$ . The bitrate range of our image codec is achieved by applying different  $\\beta$  that  $\\beta \\in \\{512, 1024, 2048, 4096\\}$ . During the training stage, we randomly generate the group masks for the training of the group-independent transform, the detailed operations are shown in the supplementary material.",
+        "bbox": [
+            496,
+            550,
+            890,
+            838
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Datasets and Protocol. For the entire image reconstruction quality, we use the widely-used Kodak dataset [37] to evaluate the coding efficiency of our models. We evaluate the selective reconstruction quality on objects",
+        "bbox": [
+            498,
+            839,
+            890,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "17241",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/9b12c7d4a0e73874705e51a40266a9d7bb47f380cc053c4c9c90bc580acf9213.jpg",
+        "image_caption": [
+            "(a) Entire image reconstruction on Kodak."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            151,
+            80,
+            486,
+            337
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/454acd5af833d1e28fe6c84492629d937cfdd848dcee5678633f682360df2fbb.jpg",
+        "image_caption": [
+            "(b) All categories objects reconstruction on COCO."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            80,
+            839,
+            200
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/05afc7ad8aab7a1a4c1e804b9e235000091a7895845f4c471278342b1924d8c5.jpg",
+        "image_caption": [
+            "(c) Human category objects reconstruction on COCO."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            218,
+            836,
+            337
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/86f7d85cbcfd47b5265e68850ba59a0ba7be47d47417f88613bd42ec786aa812.jpg",
+        "image_caption": [
+            "Figure 6. Rate distortion comparison of compression efficiency on both entire and partial reconstruction scenarios. The PSNR of partial reconstructions is only calculated on the pre-detected bounding boxes."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            400,
+            272,
+            511
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/353c8bda89cf67445284c9a35188dad4bc74df5b7b498bec08ef0bae8c6d4a9e.jpg",
+        "image_caption": [
+            "Figure 7. Performance comparison on instance segmentation (left) and pose estimation (right) on COCO."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            277,
+            400,
+            460,
+            511
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "of all categories and human objects using 40 and 20 images, which are randomly selected from the COCO 2017 validation set. The PSNR of the region of interest serves as the metric for measuring the objective quality. Additionally, the bits per pixel (bpp) are calculated by dividing the total bits of the transmitted bitstream by the number of pixels in the region of interest. To verify the effectiveness of our proposed method in supporting downstream tasks, we conduct experiments on instance segmentation and pose estimation using the COCO 2017 validation set. [45]. This dataset is widely used for dense prediction tasks and comprises 118K training, 5K validation, and 20K test-dev images. We evaluate the performances of instance segmentation and human keypoint detection using Mask R-CNN (X101-FPN backbone) and Keypoint R-CNN (X101-FPN backbone), respectively, implemented in the Detectron2 toolbox [68].",
+        "bbox": [
+            75,
+            555,
+            472,
+            797
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Quantitative Results",
+        "text_level": 1,
+        "bbox": [
+            76,
+            800,
+            272,
+            816
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For comparisons, we choose recent neural image codes [16,27,52,75], powerful classical image codes [10, 61], and the functional codec that supports semantically structured bitstream [62]. Among them, ELIC [27] is reproduced and performs closely as their report. It is crit-",
+        "bbox": [
+            75,
+            825,
+            468,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "ical to emphasize that the encoder does not possess prior knowledge regarding the regions of interest required by the decoder. And we utilize object detection as the pre-analysis by default, which is one of the most fundamental visual analysis tools. The semantically structured bitstreams are generated based on the bounding boxes detected by the detector of Mask R-CNN (X101-FPN backbone), and are selectively decoded and reconstructed into images for evaluation according to the requirements of relative semantics. In all experiments of quantitative results, the block size  $B$  of group masks is set to 32 to balance the trade-off between performance and overhead bitrate, and we take the union of overlapping groups as one group. More experiments compared with other methods of image coding for machines are presented in the supplementary.",
+        "bbox": [
+            496,
+            400,
+            893,
+            628
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Reconstruction Quality. For entire image reconstruction, Figure. 6a demonstrates that our proposed model achieves state-of-the-art rate-distortion performance, surpassing VTM 18.2 in terms of PSNR at all bitrates. Notably, structuring the bitstream according to semantics introduces negligible effects. Therefore, our model has a strong potential to achieve both high coding efficiency and functionality simultaneously. For partial image reconstruction in scenarios of specific interest, as shown in Figure. 6b and Figure. 6c, our model has achieved a significant improvement compared to other CODECs. Specifically, the semantically structured bitstream enabled our model and SSIC [62] to avoid transmitting and decoding bitstreams corresponding to the entire image, which is required of ELIC [27]. Additionally, our model's flexible block-level group mask and group-independent transform spatially decouple latents from the semantic level, resulting in sharper and more re",
+        "bbox": [
+            496,
+            643,
+            893,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "17242",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/9a7d6896638011c7da58bfb5d5a4d44d52557be07a841e5bd38414a2e789d8df.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            79,
+            222,
+            148
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d26c13f5ecbf86b15f0e1465c92066dad84819ca749f521a992c3f2938c59cc1.jpg",
+        "image_caption": [
+            "0.516/29.36"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            227,
+            80,
+            357,
+            148
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/b82832be0cdaefe0fb89afa2c13c282dce936f6ee9ef022d4683c3edd1066af1.jpg",
+        "image_caption": [
+            "0.251/29.40"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            364,
+            80,
+            493,
+            148
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/91153e25e2b2538a795999db486136efc42bc8df64aa3cf632dee8600fcbb228.jpg",
+        "image_caption": [
+            "0.195 / 30.14"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            80,
+            630,
+            148
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2321acaaa637c289a5c0a1f8e3f959ffd0cb018b4c55e85bea0cd98459d59d91.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            656,
+            80,
+            710,
+            148
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2e1c2e28dc631c2fdcd56c02a5221d6e48d989aa79fe96945ba03b26ca87ddfb.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            715,
+            80,
+            766,
+            148
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f357cb43ff7ddff1b85d23af659ca4e98589f68dcb4b8cd5c7ee436d77d2a67d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            80,
+            823,
+            148
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/60b72b7b5778c6a5edad1e354da5f0f2a811a9de8c487dbf196f0444f14857a6.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            828,
+            80,
+            880,
+            148
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/6596592032ef0de60f2972318ae404dc93639bf21f1ce1008c80b23f6288bb76.jpg",
+        "image_caption": [
+            "Original Image"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            91,
+            161,
+            220,
+            226
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/5815dc1ed2654d37d11e6a5ea48fc4b30b7d7b13e331035ac5fda3288bfc736b.jpg",
+        "image_caption": [
+            "0.318/28.61",
+            "ELIC"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            227,
+            161,
+            357,
+            226
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/c9a827a04c52c7686e159b55bdd6a0952adfae61c751cdbf42badde15c87a27c.jpg",
+        "image_caption": [
+            "0.295/28.53",
+            "SSIC"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            364,
+            161,
+            493,
+            226
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/c8c1e82b2c0da3fcbbfaaee39d9ebac3dd039bccb0aceaf89619198a4d7d54f8.jpg",
+        "image_caption": [
+            "0.316 / 29.25",
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            161,
+            630,
+            226
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/5f2953d350535e7801e274afa9f8d1063982131c77f25427a099cd4d84f3be2e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            656,
+            161,
+            707,
+            193
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/adece8258b93916b6efd26abe44ab7f10348c55c69fafce5b5d6b4d2f710cfce.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            656,
+            194,
+            707,
+            226
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/9911aecd9cc119cf8b57d5b021fd2ab6a46f042fb417a95b1bd287df5dc9bbdd.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            714,
+            161,
+            766,
+            193
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2ea4415c651ebfca4edbf6c5af7750d68a43b3fac712ef5bb3eccb1f4c7dfe15.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            715,
+            194,
+            766,
+            226
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/608398d18aaf1880d92b461c09528722f51fa2b0b3bbda7116c4551a529a0d70.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            161,
+            823,
+            193
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2ade7db95446bd023e7632e8331c570c6f4ac7f7e4808aec24940599d2122fba.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            194,
+            823,
+            226
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/c2fc8807996df6a36b79859eeeca1e3317e3151ff8419059fb1432c567730fc3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            828,
+            161,
+            880,
+            193
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/8415f5019a1d8711b9a137210601a8f997b8e81047da306068fcf626e1d7b406.jpg",
+        "image_caption": [
+            "Figure 8. Visualization of reconstructed images for human consumption from the COCO 2017 dataset, in the scenario where foreground objects are the regions of interest. The numbers below the images correspond to bits per pixel and PSNR."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            828,
+            194,
+            880,
+            226
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/a13a0777271ce508fd1754ea8677adf56d6b83d632d1b3afe9937524c4a52099.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            300,
+            236,
+            390
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/aa1c98fa10d31e1c2136f8e7d61b4021f37b7bae999e9f18f5f1ff57e58abf21.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            243,
+            300,
+            398,
+            388
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/861d3d4a4d8256d62b1215bbe6b9b627198ca1d1994661459f2a6fda843512dc.jpg",
+        "image_caption": [
+            "1.850bpp"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            406,
+            300,
+            560,
+            388
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/330b8d1349b7574c9d6833e3b9f55fc3ebf4f4452a1c3e7aa65971fb0e685025.jpg",
+        "image_caption": [
+            "1.592bpp"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            568,
+            300,
+            723,
+            388
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/49fa984a24ca38df917f62282ed49a2bea31b9f456fb45145a7541e5c816f5ff.jpg",
+        "image_caption": [
+            "1.211bpp"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            732,
+            300,
+            883,
+            388
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/acd3064c9b691bceecd9ea5b338e637633cc619e06e0d49549e35befe6077a76.jpg",
+        "image_caption": [
+            "Original Image",
+            "Figure 9. Visualization of reconstructed images for downstream tasks from the COCO 2017 dataset. The first line corresponds to instance segmentation, and the second one corresponds to pose estimation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            402,
+            235,
+            474
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/1506b4316dbd3ba8017e120cfb41ee1f8d2d8f01e5922c859f08579b8a1222cf.jpg",
+        "image_caption": [
+            "Ground Truth"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            243,
+            402,
+            398,
+            474
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f7f548d9eee40c0acab91e4583932cf93071d80c036a465f43df8a5714770586.jpg",
+        "image_caption": [
+            "2.229bpp",
+            "ELIC"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            406,
+            402,
+            558,
+            474
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d811158090ed31fc98c9a7d750aff7fa179417fb1a258abe2e11091460623ccb.jpg",
+        "image_caption": [
+            "1.335bpp",
+            "SSIC"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            566,
+            402,
+            720,
+            473
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/8f696fbfa4d9872515a66aa0bee837dbf4d4afb5e2e216a55c05aae7b68a03eb.jpg",
+        "image_caption": [
+            "1.048bpp",
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            730,
+            402,
+            883,
+            473
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "alistic boundaries compared to SSIC [62]. Moreover, in the case of overlapping regions, our model can significantly save bitrate by grouping them into an irregular group instead of replacing them with a larger bounding box, which would introduce more distortion.",
+        "bbox": [
+            75,
+            541,
+            468,
+            616
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Downstream Task Supporting. Figure. 7 demonstrates our excellent performance on instance segmentation and pose estimation. On instance segmentation, our model achieves significantly improved performance at low bitrates (<0.6bpp) compared with other methods, which can be attributed to the faithful reconstruction of RoI boundaries in our approach. It is worth noting that the superior performance of our method on pose estimation is due to the ability of our model to preserve fine details and boundaries, which are critical for accurately localizing human keypoints. In addition, human objects are a minority in the image and are often sparsely distributed, and compared with SSIC, our method can significantly reduce bitrate by avoiding transmitting the latents corresponding to the rectangular area containing all overlapped objects.",
+        "bbox": [
+            75,
+            619,
+            470,
+            845
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3. Qualitative Results",
+        "text_level": 1,
+        "bbox": [
+            76,
+            849,
+            261,
+            864
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Reconstruction for Human Consumption. When the reconstructed images are mainly intended for human con",
+        "bbox": [
+            75,
+            869,
+            468,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "sumption with specific semantic interests, as shown in Figure 8, bitstreams generated by CODEs designed for compressing regular rectangular images are required to be fully transmitted and serve full reconstruction, regardless of the requirements and the content of images. Although SSIC [62] partially addresses this issue through generating SSB through bounding box based division, it lacks flexibility and may introduce additional irrelevant disturbance when merging overlapping objects into larger bounding boxes. Moreover, SSIC generates the bitstream by directly dividing the corresponding positions in the latent variable space after the transform based on ConvNets, which can lead to lost dependencies during selective reconstruction, resulting in blurred and distorted region boundaries. With group mask based partitioning and group-independent transform, our method decouples images to generate the SSB more efficiently while ensuring that no distortion or blurring is introduced after selective reconstruction, leading to both significant bitrate saving and a better visual experience.",
+        "bbox": [
+            496,
+            541,
+            890,
+            829
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Reconstruction for Intelligent Tasks. When using reconstructed images for downstream intelligent tasks, selectively transmitting and decoding the semantically structured bitstream based on the prior knowledge of relative seman",
+        "bbox": [
+            496,
+            839,
+            892,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "17243",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/799f42b5c5c2dcb10d91b8b35101445a0edbdc4ea3afb0a026a1b970294ae9bf.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            80,
+            171,
+            127
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e7fa75907c6ea135bb0420e8e2dd0678a0a05ebfef73509e511ed710500ac633.jpg",
+        "image_caption": [
+            "Original Image"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            130,
+            171,
+            176
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/37fc29c7ed7d0877f883a4ea8c36c41c4fa9ce64956e6d5946ec7108b3b187eb.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            179,
+            80,
+            267,
+            127
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/d4db6a3afb325cfeeef3f20e5835f794e0dca570673340f8d777d1a6a7c915ac.jpg",
+        "image_caption": [
+            "Pre-Analysis"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            179,
+            130,
+            267,
+            176
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a2799c7805e83a4a04e9f5a0784d05410ae31451129c4fd211fa829470a7017b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            274,
+            80,
+            366,
+            127
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/43a71bd43fd4e03274e00f62bdb929c436d733b3f0c4338dd5c132687033a308.jpg",
+        "image_caption": [
+            "Based on"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            274,
+            130,
+            366,
+            176
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/8769c172d99d89cd87ba721a8bb8fcb780119bb35cd6cf6d0b62192c2879d15d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            372,
+            80,
+            464,
+            127
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/0f6557947609d733be5376ff9746a43819fdc09b4b7e1962eeb2d422f0efc629.jpg",
+        "image_caption": [
+            "Bounding Boxes",
+            "Based on",
+            "Instance Maps"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            372,
+            130,
+            464,
+            176
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6fa09b9e4356de1d3cfc882bdc6df10b92f7f6b79b3643749b9d5c70e87784fe.jpg",
+        "image_caption": [
+            "Figure 10. Selective reconstructions from the semantically structured bitstream. The bitstream can be generated based on different pre-analysis methods.",
+            "Original Image",
+            "Figure 11. Illustration of semantically-aware encryption based on our proposed method."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            84,
+            253,
+            171,
+            301
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b261705210627fc10e0391bf0a3ff55d4f1a540e4577571c0703c93c0f99b029.jpg",
+        "image_caption": [
+            "Pre-Analysis"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            181,
+            253,
+            267,
+            301
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/0c2f30714811b2af2c34764895014333a29fcf7d3308a8cf5052e087eede41ef.jpg",
+        "image_caption": [
+            "Encrypt People"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            276,
+            253,
+            362,
+            301
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b773c9f4052d737dee2ed6239d88df1083153744309efc5282895de23c2fcdcc.jpg",
+        "image_caption": [
+            "Encrypt People",
+            "and Vehicles"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            372,
+            253,
+            460,
+            301
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tics can significantly save bitrate. As illustrated in Figure 9, conventional codecs designed for compressing rectangular images require the entire bitstream to be transmitted and decoded. Downstream models then perform intelligent analytics on the fully reconstructed image, resulting in a waste of bitrate. Similar to image reconstruction for human consumption, SSIC can perform selective transmission to save bitrate. However, its bounding box based partitioning may not be optimal for bitrate savings, and the resulting blurred and distorted region boundaries and irrelevant content can further impede downstream intelligent analytics. (misidentifying the car as a person shown in Figure 9). Our method can more efficiently generate the SSB and support various downstream tasks with specific requirements with both accuracy and coding efficiency.",
+        "bbox": [
+            75,
+            364,
+            467,
+            590
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Customizability and Flexibility. It's flexible to customize the semantically structured bitstream based on different semantical partition criteria. For instance, as shown in Figure 10, the generation and usage of semantically structured bitstreams can be based on either object detection or instance segmentation, depending on the specific requirements. It should be noted that the methods of pre-analysis are not limited to object detection and instance segmentation as shown in Figure 10, but can include other techniques such as saliency detection [22, 30, 72], semantic segmentation [5, 48, 69, 71, 73], panoptic segmentation [15, 36], etc., and even human annotation.",
+        "bbox": [
+            75,
+            595,
+            467,
+            775
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Semantically-Aware Encryption. As presented in Figure 11, our proposed method can be applied to bitstream encryption, enabling selective and even layered encryption with semantic priors based on the user's security level. In some cases, the selectively encrypted SSB can allow for secure transmission and storage of sensitive information while minimizing the impact on visual quality and downstream analytics. Please refer to the supplementary material for",
+        "bbox": [
+            75,
+            780,
+            467,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/456c61c5f8400fe58ecc3baef621b45b8a82c4c31dc7286b67f525c6784f8833.jpg",
+        "image_caption": [
+            "(a) RoI reconstruction.",
+            "Figure 12. Ablation study."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            80,
+            687,
+            190
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/c3bcf2a27219e7eb2317217dfe73097f5e1251e12e9589ea6cdc245eef733f9d.jpg",
+        "image_caption": [
+            "(b) Intelligent task supporting."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            702,
+            80,
+            883,
+            190
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "details of implementation.",
+        "bbox": [
+            500,
+            234,
+            673,
+            250
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            500,
+            266,
+            650,
+            282
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In order to show the effectiveness of our proposed group mask and group-independent transform (GIT) for both RoI-aware selective reconstruction and downstream intelligent analytics, we conducted the ablation study at different bitrates. Specifically, we use people as the region of interest for RoI reconstruction, and the results for taking all foreground objects as RoI are presented in the supplementary material. Additionally, we conducted an ablation study on pose estimation of COCO 2017 for intelligent task support, and the results on instance segmentation are also included in the supplementary material.",
+        "bbox": [
+            496,
+            292,
+            890,
+            458
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Group Mask. Figure 12 shows that group mask based partitioning saves significant bitrate compared with naive bounding box based partitioning. Meanwhile, both the visual quality of RoI reconstruction and downstream task performances would not be impeded.",
+        "bbox": [
+            496,
+            460,
+            890,
+            537
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Group-Independent transform. As shown in Figure 12a, group-independent transform is crucial for the visual quality of RoI reconstruction. And the downstream task performance would be impeded by the blurred and distorted region boundaries without group-independent transform, which is shown in Figure 12b.",
+        "bbox": [
+            496,
+            540,
+            890,
+            631
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            647,
+            617,
+            662
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this work, we propose to generate semantically structured bitstream with strong functionality based on group masks, which are highly flexible, customizable, while maintaining lightweight overhead. Moreover, we first propose the concept of group-independent transform and instantiate it by designing the Group-Independent Swin-Block (GI-Swin Block) to ensure independence among distinct groups, resulting in pleasing reconstructions of RoI regions. Particularly, our proposed method outperforms VTM-18.2 and achieves comparable coding efficiency with the SOTA neural image CODECs, while offering strong functionality. The experimental results have demonstrated the effectiveness of our proposed method across several applications, including human consumption of regions of interest, downstream task support, and semantically-aware encryption.",
+        "bbox": [
+            496,
+            674,
+            890,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "17244",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement. This work was supported in part by NSFC under Grant U1908209, 62021001. This work was also supported in part by ZJNSFC under Grant LQ23F010008.",
+        "bbox": [
+            76,
+            90,
+            472,
+            151
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            165,
+            174,
+            181
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Eirikur Agustsson, Fabian Mentzer, Michael Tschannen, Lukas Cavigelli, Radu Timofte, Luca Benini, and Luc V Gool. Soft-to-hard vector quantization for end-to-end learning compressible representations. *NeurIPS*, 30, 2017. 2",
+            "[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In ICCV, pages 221-231, 2019. 2",
+            "[3] Mohammad Akbari, Jie Liang, and Jingning Han. Dsslic: Deep semantic segmentation-based layered image compression. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2042-2046. IEEE, 2019. 3",
+            "[4] Artem Babenko, Anton Slesarev, Alexandr Chigorin, and Victor Lempitsky. Neural codes for image retrieval. In ECCV, pages 584-599. Springer, 2014. 1, 3",
+            "[5] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI, 39(12):2481-2495, 2017. 1, 3, 8",
+            "[6] Johannes Ballé, Philip A Chou, David Minnen, Saurabh Singh, Nick Johnston, Eirikur Agustsson, Sung Jin Hwang, and George Toderici. Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing, 15(2):339-353, 2020. 2",
+            "[7] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In *ICLR*, 2017. 1, 2",
+            "[8] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In ICLR, 2018. 1, 2",
+            "[9] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, pages 9157-9166, 2019. 1, 3",
+            "[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. TCSVT, 2021. 1, 2, 6",
+            "[11] 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. TPAMI, 40(4):834-848, 2017. 1, 3",
+            "[12] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 1, 3",
+            "[13] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. TIP, 29:2230-2243, 2019. 1, 3"
+        ],
+        "bbox": [
+            78,
+            189,
+            470,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[14] Zhuo Chen, Kui Fan, Shiqi Wang, Ling-Yu Duan, Weisi Lin, and Alex Kot. Lossy intermediate deep learning feature compression and evaluation. In ACM MM, pages 2414-2422, 2019. 1, 3",
+            "[15] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, pages 12475-12485, 2020. 8",
+            "[16] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In CVPR, pages 7939-7948, 2020. 1, 2, 6",
+            "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 3",
+            "[18] Ling-Yu Duan, Vijay Chandrasekhar, Jie Chen, Jie Lin, Zhe Wang, Tiejun Huang, Bernd Girod, and Wen Gao. Overview of thempeg-cdvs standard. TIP, 25(1):179-194, 2015. 1, 3",
+            "[19] Ling-Yu Duan, Yihang Lou, Yan Bai, Tiejun Huang, Wen Gao, Vijay Chandrasekhar, Jie Lin, Shiqi Wang, and Alex Chichung Kot. Compact descriptors for video analysis: The emerging mpg standard. IEEE MultiMedia, 26(2):44-54, 2018. 1, 3",
+            "[20] Touradj Ebrahimi and Caspar Horne. Mpeg-4 natural video coding—an overview. Signal Processing: Image Communication, 15(4-5):365-385, 2000. 3",
+            "[21] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In ECCV, pages 510-528. Springer, 2022. 2, 3",
+            "[22] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. TPAMI, 34(10):1915-1926, 2011. 8",
+            "[23] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 2",
+            "[24] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. TCSVT, 32(4):2329-2341, 2021. 2",
+            "[25] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Soft then hard: Rethinking the quantization in neural image compression. In ICML, pages 3920-3929. PMLR, 2021. 2",
+            "[26] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. NeurIPS, 34:15908-15919, 2021. 1, 3",
+            "[27] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In CVPR, pages 5718-5727, 2022. 1, 2, 6",
+            "[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3",
+            "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 3"
+        ],
+        "bbox": [
+            501,
+            90,
+            893,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "17245",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[30] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In CVPR, pages 1-8. IEEE, 2007. 8",
+            "[31] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3",
+            "[32] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. ECCV, 2022. 1, 3",
+            "[33] Xin Jin, Ruoyu Feng, Simeng Sun, Runsen Feng, Tianyu He, and Zhibo Chen. Semantically video coding: Instill static-dynamic clues into structured bitstream for ai tasks. arXiv preprint arXiv:2201.10162, 2022. 3",
+            "[34] Nick Johnston, Damien Vincent, David Minnen, Michele Covell, Saurabh Singh, Troy Chinen, Sung Jin Hwang, Joel Shor, and George Toderici. Improved lossy image compression with priming and spatially adaptive bit rates for recurrent networks. In CVPR, pages 4385-4393, 2018. 1",
+            "[35] Aggelos K Katsaggelos, Lisimachos P Kondi, Fabian W Meier, Jorn Ostermann, and Guido M Schuster. Mpeg-4 and rate-distortion-based shape-coding techniques. Proceedings of the IEEE, 86(6):1126-1154, 1998. 3",
+            "[36] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, pages 9404-9413, 2019. 8",
+            "[37] Eastman Kodak. Kodak lossless true color image suite (photocd pcd0992). http://r0k.us/graphics/kodak, 1993.5",
+            "[38] Nam Le, Honglei Zhang, Francesco Cricri, Ramin GhaznaviYouvalari, and Esa Rahtu. Image coding for machines: An end-to-end learned approach. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1590-1594. IEEE, 2021. 1, 3",
+            "[39] Nam Le, Honglei Zhang, Francesco Cricri, Ramin Ghaznavi Youvalari, Hamed Rezazadegan Tavakoli, and Esa Rahtu. Learned image coding for machines: A content-adaptive approach. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1",
+            "[40] Mu Li, Wangmeng Zuo, Shuhang Gu, Jane You, and David Zhang. Learning content-weighted deep image compression. TPAMI, 2020. 1",
+            "[41] Mu Li, Wangmeng Zuo, Shuhang Gu, Debin Zhao, and David Zhang. Learning convolutional networks for content-weighted image compression. In CVPR, pages 3214-3223, 2018. 1",
+            "[42] Xin Li, Jun Shi, and Zhibo Chen. Task-driven semantic coding via reinforcement learning. TIP, 2021. 3",
+            "[43] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. ECCV, 2022. 1, 3",
+            "[44] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 1, 3",
+            "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence"
+        ],
+        "bbox": [
+            78,
+            92,
+            467,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 5, 6",
+            "[46] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, pages 8759-8768, 2018. 1, 3",
+            "[47] 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, pages 10012-10022, 2021. 1, 2, 3, 4",
+            "[48] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 3, 8",
+            "[49] Siwei Ma, Xiang Zhang, Shiqi Wang, Xinfeng Zhang, Chuanmin Jia, and Shanshe Wang. Joint feature and texture coding: Toward smart video representation via front-end intelligence. TCSVT, 29(10):3095-3105, 2018. 1, 3",
+            "[50] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Conditional probability models for deep image compression. In CVPR, pages 4394-4402, 2018. 1",
+            "[51] Fabian Mentzer, George Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. arXiv preprint arXiv:2006.09965, 2020. 1, 2",
+            "[52] David Minnen, Johannes Balle, and George Toderici. Joint autoregressive and hierarchical priors for learned image compression. In NeurIPS, 2018. 1, 2, 5, 6",
+            "[53] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 5",
+            "[54] Majid Rabbani and Rajan Joshi. An overview of the JPEG 2000 still image compression standard. Signal processing: Image communication, 17(1):3-48, 2002. 1, 2",
+            "[55] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1, 3",
+            "[56] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, pages 7263-7271, 2017. 1, 3",
+            "[57] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28:91-99, 2015. 1, 3",
+            "[58] Oren Rippel and Lubomir Bourdev. Real-time adaptive image compression. In ICML, pages 2922-2930. PMLR, 2017. 2",
+            "[59] Thomas Sikora. Thempeg-4 video standard verification model. TCSVT, 7(1):19-31, 1997. 3",
+            "[60] Saurabh Singh, Sami Abu-El-Haija, Nick Johnston, Johannes Balle, Abhinav Shrivastava, and George Toderici. End-to-end learning of compressible features. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3349–3353. IEEE, 2020. 1, 3",
+            "[61] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. TCSVT, 22(12):1649-1668, 2012. 1, 2, 6",
+            "[62] Simeng Sun, Tianyu He, and Zhibo Chen. Semantic structured image coding framework for multiple intelligent applications. TCSVT, 2020. 2, 3, 6, 7"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "17246",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[63] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. arXiv preprint arXiv:1703.00395, 2017. 2",
+            "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3, 4",
+            "[65] Gregory K Wallace. The JPEG still picture compression standard. IEEE transactions on consumer electronics, 38(1):xviii-xxxiv, 1992. 1, 2",
+            "[66] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. TMM, 2021. 3",
+            "[67] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. TCSVT, 13(7):560-576, 2003. 1",
+            "[68] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6",
+            "[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 34:12077-12090, 2021. 1, 3, 8",
+            "[70] Yibo Yang, Robert Bamler, and Stephan Mandt. Improving inference for neural image compression. NeurIPS, 33:573-584, 2020. 2",
+            "[71] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 2881-2890, 2017. 8",
+            "[72] Rui Zhao, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Saliency detection by multi-context deep learning. In CVPR, pages 1265-1274, 2015. 8",
+            "[73] 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 CVPR, pages 6881-6890, 2021. 1, 3, 8",
+            "[74] Lei Zhou, Zhenhong Sun, Xiangji Wu, and Junmin Wu. End-to-end optimized image compression with attention mechanism. In CVPR workshops, page 0, 2019. 5",
+            "[75] Yinhao Zhu, Yang Yang, and Taco Cohen. Transformer-based transform coding. In ICLR, 2022. 1, 2, 4, 6"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            702
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "17247",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/e3cccf23-d23a-4b30-90e1-8c04059b69dd_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4251bfefca76eacf3a242926a64e9843bea1312e77831c613554e6e02de7a1de
+size 6743904

2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/full.md

@@ -0,0 +1,407 @@
+# Semantically Structured Image Compression via Irregular Group-Based Decoupling
+
+Ruoyu Feng $^{1*}$  Yixin Gao $^{1*}$  Xin Jin $^{2}$  Runsen Feng $^{1}$  Zhibo Chen $^{1,\dagger}$
+
+<sup>1</sup>University of Science and Technology of China <sup>2</sup>Eastern Institute of Technology, Ningbo
+
+# Abstract
+
+Image compression techniques typically focus on compressing rectangular images for human consumption, however, resulting in transmitting redundant content for downstream applications. To overcome this limitation, some previous works propose to semantically structure the bitstream, which can meet specific application requirements by selective transmission and reconstruction. Nevertheless, they divide the input image into multiple rectangular regions according to semantics and ignore avoiding information interaction among them, causing waste of bitrate and distorted reconstruction of region boundaries. In this paper, we propose to decouple an image into multiple groups with irregular shapes based on a customized group mask and compress them independently. Our group mask describes the image at a finer granularity, enabling significant bitrate saving by reducing the transmission of redundant content. Moreover, to ensure the fidelity of selective reconstruction, this paper proposes the concept of group-independent transform that maintain the independence among distinct groups. And we instantiate it by the proposed Group-Independent Swin-Block (GI Swin-Block). Experimental results demonstrate that our framework structures the bitstream with negligible cost, and exhibits superior performance on both visual quality and intelligent task supporting.
+
+# 1. Introduction
+
+Image signal serves as a critical information carrier for various applications in modern society. Image compression techniques aim at converting images into compact representations (i.e. bitstreams) to save transmission and storage resources. Lossy image compression is one of the most practical techniques, as it allows for the restoration of important content while discarding a small amount of inessential information. In the past decades, traditional image compres
+
+![](images/82944d9ba34b0a08d6b1cec5006054c78948fdbab5baaf9006f2c024e98a9be7.jpg)  
+Figure 1. The input image is decoupled into groups according to distinct semantics. Then the semantically structured bitstream (SSB) is generated by compressing the image based on the partitioned groups. The SSB facilitates downstream applications by selective bitstream transmission and partial reconstruction, depending on the specific task requirements.
+
+sion standards [10,54,61,65,67] have been extensively studied and utilized. With the fast development of deep learning, neural image codes [7,8,16,27,34,40,41,50-52,75] rapidly evolved and achieved promising results. Meanwhile, more and more media contents tend to be handled by machine vision algorithms, such as recognition [17,26,29,32,47], detection [43,44,55-57], and segmentation [5,9,11,12,28,46,48,69,73]. However, most compression methods are mainly developed for compressing regular rectangular images for human consumption, without considering the efficiency and functionality for downstream tasks or human-machine interaction scenarios.
+
+Recently, the field of image coding for machines (ICM) has emerged to develop a joint efficient and analytical framework for supporting intelligent analytics. End-to-end optimization of the trade-off between specific task loss and compression rate is a promising way [4, 13, 14, 18, 19, 38, 39, 49, 60], but it lacks generalization for massively diverse
+
+applications. To overcome this limitation, Feng et al. [21] propose to compress general and compact features learned by self-supervised learning under entropy constraints for supporting downstream tasks. Nevertheless, it requires retraining the task model with the proposed features as inputs, which makes the overall performance heavily depend on the effectiveness of feature extraction. The methods mentioned above are designed specifically for scenarios of compression for machine vision without considering situations in which human involvement is required. Semantically structured image compression (SSIC) [62] proposes to generate a semantically structured bitstream (SSB) by separately compressing rectangular regions of detected objects using a pre-prepared object detection toolbox. Although SSB is efficient in supporting intelligent tasks and human-machine interaction through partial transmission and reconstruction of the bitstream, its division approach based on rectangular regions may encounter problems with overlapped objects. SSIC addresses this issue by replacing the overlapping objects with a larger rectangular region, which can result in a waste of bitrate. In addition, SSIC generates the bitstream of each object by compressing the corresponding latent variables in the latent domain directly, without considering the interaction and dependence of features during the transform process. This can lead to blurry and distorted group boundaries in the partial reconstruction scenario, which in turn affects the reconstruction quality.
+
+Going beyond the rectangular-based division, this paper proposes to decouple the image into multiple groups with irregular shapes based on a customized group mask. Then the SSB is generated by independently compressing these groups and can support various requirements via selective transmission and reconstruction. Notably, the generation of the group mask offers high flexibility in terms of shapes of groups, means of pre-analysis, and partition criteria, enabling customization to suit diverse application scenarios and requirements. Moreover, to avoid potential quality degradation in the partial reconstruction scenario, we propose the concept of the group-independent transform, which ensures the independence among groups in the latent representations and therefore the quality of the selective reconstruction will not be affected by the absence of other groups. More specifically, we instantiate it by carefully designing the Group-Independent Swin-Block (GI Swin-Block), which is an extension of Swin Transformer [47] tailored to our situation and requirements. GI SwinBlocks make use of the hierarchical modeling capability of the Swin-Transformer [47, 75], achieving high coding efficiency under the premise of group independence. By combining the group mask based decoupling and the Group-Independent Swin-Block, our proposed method can efficiently support various downstream applications including human-machine interaction and machine vision tasks with
+
+only one bitstream generated.
+
+The main contributions of our approach are summarized as follows:
+
+- We propose to decouple an image into multiple groups with irregular shapes for structuring the bitstream. Our group mask can describe spatial division at a finer granularity manner than a typical rectangle, saving bitrate by reducing redundant content transmission.  
+- We propose the group-independent transform and instantiate it by carefully designing the Group-Independent Swin-Block (GI Swin-Block), which maintains powerful transformation capability and ensures the independence among groups in the latent representations.  
+- Experimental results demonstrate that our proposed model achieves state-of-the-art compression ability and superior downstream tasks performance, which is a codec with both high compression efficiency and functionality.
+
+# 2. Related Works
+
+# 2.1. Image Compression
+
+Traditional Image Compression. Traditional image compression standards, such as JPEG [65], JPEG2000 [54], HEVC [61], and VVC Intra [10], have been extensively used in practice after several decades of development. These standards rely on transform coding [23], which decomposes the lossy image compression task into three parts: transform, quantization, and entropy coding. Each module of these standards is manually designed with multiple modes, and rate-distortion optimization is performed to determine the optimal mode. However, the completely hand-crafted structure of traditional CODECs limits their flexibility and scalability to support various objectives, such as MS-SSIM and classification accuracy, as they cannot be optimized in an end-to-end manner.
+
+Learned Image Compression. In recent years, learned image compression methods based on nonlinear transform coding [6] have achieved rapid progress. Early works in this area concentrated on enabling end-to-end training by developing differential quantization and rate estimation techniques [1, 7, 63]. Subsequently, a considerable amount of work design powerful neural network modules to enhance the transform [16, 75], quantization [25, 70], and entropy model [8, 27, 52, 53]. As a result, some neural image CODECs [16, 24, 27, 52, 53] can achieve comparable or even superior performance to traditional coding standards like HEVC [61] and VVC [10]. Moreover, some works [2, 51, 58] introduce perceptual loss to effectively improve reconstruction quality. However, existing image compression methods focus on compressing the entire image
+
+without considering selective transmission and reconstruction of the bitstream for arbitrary reconstruction, leading to significant bandwidth waste when addressing downstream applications with different requirements.
+
+# 2.2. Image Coding for Machines
+
+Image coding for machines aims at compress source images to serve downstream tasks, such as recognition [17, 26, 29, 32, 47], detection [43, 44, 55-57], and segmentation [5, 9, 11, 12, 28, 46, 48, 69, 73]. Joint optimization of task loss and bitrate [3, 31, 38, 42, 66] in an end-to-end manner is a natural and promising way. Another direction is compress the features corresponding to downstream tasks [4, 13, 14, 18, 19, 49, 60]. However, these methods tend to be biased towards specific tasks and may lack generalizability to a wide range of intelligent applications. To address the limitation of generalization, a novel representation learning based method is proposed [21] to learn general and compact features by combining contrastive learning with entropy constraints. The learned features are used to replace the original images as the new source for compression and transmission, resulting in a significant improvement in coding efficiency across various intelligent tasks. Nevertheless, this approach necessitates re-training the task model using the proposed features as input, and its overall performance is heavily reliant on the effectiveness of feature extraction. In this paper, we propose a unified compression framework for machine vision support without task-oriented joint training or additional feature learning. Moreover, it can also conduct full or selective reconstruction for human perception according to requirements.
+
+# 2.3. Object-oriented Image Compression
+
+Object-oriented image compression is first proposed in MPEG-4 Visual [20, 35, 59] by compressing visual object planes (VOP) with arbitrary shapes, targeting content-based interactivity and scalability for the image compression technique. However, due to the difficulty in annotating the complex and intricate shapes of the target and the lack of actual corresponding downstream applications, MPEG-4 Visual has not been widely used. Recently, with the fast development of deep learning [29, 64], the practical scenarios of image compression for downstream applications have increased rapidly. Sun et al. [33, 62] propose the semantically structured image compression (SSIC) of decoupling the image into objects and compressing the corresponding latent variables independently of the neural image codec to generate the semantically structured bitstream. Nevertheless, the rectangular-based division adopted by SSIC lacks generalization to irregular objects and flexibility for customization. Additionally, selecting the representation of objects directly in the latent space ignores the interaction among elements during the encoding transform process, leading to
+
+significant edge distortion of objects in partial reconstruction. This paper proposes to conduct semantically structured image compression by decoupling the image into irregular groups based on the group mask, which is flexible and customizable. Furthermore, we propose the group-independent transform conducted during both the encoding and decoding processes, which enables efficient partial or complete reconstruction of images under various requirements.
+
+# 3. Method
+
+# 3.1. Overview
+
+The proposed method is an efficient and flexible version of semantically structured image compression [62]. It can satisfy multiple application requirements via transmitting and reconstructing partial spatial regions guided by a block-wise group mask. Our model takes the group mask as guidance to ensure that interactions of transform only occur within the same group, thereby achieving group independence during the redundancy removal process. The group mask, which is generated based on the pre-analysis, such as object detection, semantic segmentation, and saliency detection, provides high flexibility and customization for structuring the bitstream and is considered as side information. Then the entropy coding process is conducted on the variables of each group distinctly, resulting in a bitstream structured by semantics. The bitstream can be partially or fully transmitted according to the requirements of the decoder side, and then the receiver conducts entropy decoding on the bitstream and reorganizes the spatial-wise arrangement of latent variables according to the group mask and the group indexes.
+
+The overall network structure is illustrated in Fig. 2, which incorporates the group mask and GI Swin-Block for guiding and instantiating the group-independent transform, respectively. Section 3.2 and Section 3.3 introduce the group mask and group independent, respectively. The detailed implementation of our network architecture is presented in Section 3.4.
+
+# 3.2. Group Mask
+
+Before compression, the group mask is generated according to the results of pre-analysis techniques such as object detection, instance segmentation, and saliency detection. Figure 3 provides an example of this process. More specifically, for the input image  $x$  with a height of  $H$  and a width of  $W$ , the spatial resolution of the group mask is the same as the input image, and it consists of  $\frac{H}{B} \times \frac{W}{B}$  blocks, where  $B$  is the side length of each block. And the value of it indicates the index of the corresponding group.
+
+Our proposed group mask can clearly divide the elements in the latent variable space after the one-pass down
+
+![](images/d76eb38af4aa73a7eb406dfec3fa55e911ae3f0cb944f421209f8772f5801d39.jpg)  
+Figure 2. The network architecture of our proposed model with the channel-wise auto-regressive model (ChARM). ConvT denotes transposed convolution. AE and AD are respectively arithmetic encoding and arithmetic decoding. In Ours-Hyper model, we remove the ChARM component and instead output  $\mu$  and  $\sigma$  directly from the hyperdecoder  $h_s$ .
+
+![](images/d018b1156d0090bc1bf3f1b2206388b15125c66e51427de92ecb2ae225e90d97.jpg)  
+Figure 3. An example of group mask generation. The input image is pre-analyzed by object detection, then the group mask is generated based on the results of the pre-analysis.
+
+sampling transform. Besides, downsampling the group mask by a factor of  $B$  before conducting compression can greatly reduce the overhead bitrate. The block size and the overhead bitrate have an inverse relationship. Furthermore, as shown in Figure 4, the generation of the group mask can be flexibly customized, in terms of the way of pre-analysis and the criteria of block allocation, etc.
+
+# 3.3. Group-Independent transform
+
+Lossy image compression based on transform coding can be divided into three modularized components: transform, quantization, and entropy coding. While quantization and entropy coding do not affect the generation and usage of the semantically structured bitstream (SSB), traditional transform applied to the entire image to remove spatial redundancy inevitably creates inter-group dependencies during compression. In situations of selective transmission and reconstruction, the incomplete representation that results from the reorganization of a subset of all groups can
+
+lead to inaccurate reconstruction due to the lack of intergroup dependencies. Therefore, we propose the concept of the group-independent transform for semantically structured bitstream generation. The basic idea is to constrain the transform to be conducted only within each group.
+
+A straight and natural way is by using transformer [64] with the customized attention map corresponding to the group mask. However, image compression differs significantly from high-level understanding tasks, which the transformer excels at. The presence of numerous complex long-range dependencies can impede convergence and adversely affect the final performance. In order to achieve high coding efficiency while maintaining the group-independent property, and inspired by the strong hierarchical representation modeling capability of Swin-Transformer [47, 75], we extend the Swin-Block of Swin-Transformer to our proposed group-independent Swin-Block (GI Swin-Block). The GI Swin-Block serves as the core component of the transform. To be more specific, as shown in Figure 5a, by merging the window partition and the group partition into the group-independent window partition, allowing the self-attention to be computed within the partitioned local regions, enabling us to achieve high coding efficiency while maintaining the group-independent property. Additionally, the cross-window connections are introduced by merging the shifted window partition and group partition similarly, as depicted in Figure 5b.
+
+# 3.4. Network architecture and pipeline
+
+Figure 2 illustrate the proposed architecture. The main transform is performed in a staggered manner through
+
+![](images/64e2a7e5d133e8584b985bb8fc4289ce2b243157912d5b5644d5c8bb2a027ef0.jpg)  
+Input Image
+
+![](images/847ce4cb612360f0315619348f88a9c1ae9f73c5efdf51227898cba8b0507cf1.jpg)  
+Pre-Analysis
+
+![](images/7da7c82cd4385373807b1a9982a8b99f7f115051fb63e0034447d20768c16861.jpg)  
+(a)  
+Possible Group Masks
+
+![](images/57afce836be7bf66ddd76947463897e79365a5ec9ab2725cade8d1ad37afb6c0.jpg)  
+(b)
+
+![](images/7a899820c4768ae05823d7a91afd1b4c4e267e1306e2d69a2fa71390d363631d.jpg)  
+(c)
+
+the incorporation of upsample/downsample and group-independent Swin-Block (GI Swin-Block). Upsample contains a pixel shuffle operation and a convolution layer with  $1 \times 1$  kernel size, whereas downsample contains a pixel unshuffle operation and a convolution layer with  $1 \times 1$  kernel size. For entropy coding, we adopt the Mean & Scale (M&S) Hyperprior model [52] with the strong channel-wise auto-regressive model (ChARM) [53] to predict the probability distribution of latent variables.
+
+Specifically, given a source image  $\pmb{x}$ , the encoder  $g_{a}$  converts it to latent representations  $\pmb{y}$  conditioned on the given group mask  $\pmb{m} = \sum_{i=1}^{N} \pmb{m}_{i}$ , where  $i \in \{1,2,3,\dots,N\}$ , indicating the index of the corresponding group in all  $N$  groups. In each  $\pmb{m}_{i}$ , the values of positions corresponding to the  $i_{th}$  group are set as  $i$ , and all values of other positions are set as 0. Similarly, we can define  $\pmb{x} = \sum_{i=1}^{N} \pmb{x}_{i}$  and  $\pmb{y} = \sum_{i=1}^{N} \pmb{y}_{i}$  from the perspective of group, and transform only occurs within each group, which is given by
+
+$$
+\boldsymbol {y} _ {i} = g _ {a} \left(\boldsymbol {x} _ {i} \mid \boldsymbol {m} _ {i}\right). \tag {1}
+$$
+
+Thanks to the group-independent transform, we can easily implement the transform of all groups within one forward-pass, thus, the process can also be written as  $\pmb{y} = g_{a}(\pmb{x}|\pmb{m})$ .
+
+Then the latents are quantized to discrete representations  $\hat{\pmb{y}}$  and selectively transmitted to the decoder side with predicted probability distribution through the entropy model. Concretely, the side information  $\pmb{z}$  is extracted by the hyper-encoder  $\pmb{z} = h_{a}(\pmb{y})$ . The quantized hyper-latent  $\hat{\pmb{z}} = Q(\pmb{z})$  is modeled and entropy-coded with a learned factorized prior. In ChARM, latent  $\hat{\pmb{y}}$  is split into  $S$  (we choose  $S = 10$ ) slices along with channel dimension, and each slice  $\hat{\pmb{y}}_s$  is entropy-coded based on the previous slices  $\hat{\pmb{y}}_{< s}$ .  $\hat{\pmb{y}}$  is modeled by a conditional Gaussian distribution convolving with a unit uniform noise.
+
+After selectively entropy decoding the quantized latents, we have  $\hat{\pmb{y}}_p = \sum_{j\in \mathbf{P}}\hat{\pmb{y}}_j$  , where  $\mathbf{P}\subseteq \{1,2,\dots ,N\}$  is the gather of required groups. The reconstruction image  $\hat{x}_p = \sum_{j\in \mathbf{P}}\hat{x}_j$  will be calculated as follows
+
+![](images/260198688b7f50c623cc5d37cb6e3268728877286a4ba52d8e5c7cfbd0f3b894.jpg)
+
+![](images/29995e04e180ef31d391f4849e88039ea00f178dd36534d0563f50a847c3de6a.jpg)  
+Figure 4. Illustration of flexibility to customize the group mask: (a). Taking overlapping objects as one group with the guidance of bounding boxes. (b). Taking overlapping objects as one group with the guidance of instance masks. (c) Taking overlapping objects as distinct groups with the guidance of instance masks.  
+(a) GI window partition under regular window partition.  
+(b) GI window partition under shifted window partition.  
+Figure 5. Group-independent window partition of GI SwinBlock under regular and shifted window partition. Self-attention is conducted inside each local region.
+
+$$
+\hat {\boldsymbol {x}} _ {j} = g _ {s} \left(\hat {\boldsymbol {y}} _ {j} \mid \boldsymbol {m} _ {j}\right). \tag {2}
+$$
+
+Similar to Equation (1), Equation (2) can also be written as  $\hat{\pmb{x}} = g_s(\hat{\pmb{y}} |\pmb {m})$ . Note that the rounding operation is non-differentiable, thus we apply uniform noise for learning the entropy model and use a rounded one as the input of the decoder  $g_{s}$  as in [53].
+
+# 4. Experiments
+
+# 4.1. Experiment Setup
+
+Image Codec. We train our high bitrate models  $(>= 0.6\mathrm{bpp})$  on the Flickr 2K dataset, which has the same setting as [74] and train our low bitrate models  $(< 0.6\mathrm{bpp})$  on the COCO 2017 training set [45]. Ours-Hyper are trained with 3.2M iterations, while Ours-ChARM models are trained by initializing their weights from higher bitrate hyperprior models for another 1.2M iterations. Each batch contains 8 random  $256 \times 256$  crops from the training dataset. The learning rate is set as  $5e - 5$  and is decayed by a factor of 10 at  $2.8M$  iterations. Training loss  $L = R + \beta D$  is the weighted combination of the rate-distortion trade-off, balanced by the Lagrange multiplier  $\beta$ . We adopt the mean squared error (MSE) in the RGB color space as the distortion metric  $D$ . The bitrate range of our image codec is achieved by applying different  $\beta$  that  $\beta \in \{512, 1024, 2048, 4096\}$ . During the training stage, we randomly generate the group masks for the training of the group-independent transform, the detailed operations are shown in the supplementary material.
+
+Evaluation Datasets and Protocol. For the entire image reconstruction quality, we use the widely-used Kodak dataset [37] to evaluate the coding efficiency of our models. We evaluate the selective reconstruction quality on objects
+
+![](images/9b12c7d4a0e73874705e51a40266a9d7bb47f380cc053c4c9c90bc580acf9213.jpg)  
+(a) Entire image reconstruction on Kodak.
+
+![](images/454acd5af833d1e28fe6c84492629d937cfdd848dcee5678633f682360df2fbb.jpg)  
+(b) All categories objects reconstruction on COCO.
+
+![](images/05afc7ad8aab7a1a4c1e804b9e235000091a7895845f4c471278342b1924d8c5.jpg)  
+(c) Human category objects reconstruction on COCO.
+
+![](images/86f7d85cbcfd47b5265e68850ba59a0ba7be47d47417f88613bd42ec786aa812.jpg)  
+Figure 6. Rate distortion comparison of compression efficiency on both entire and partial reconstruction scenarios. The PSNR of partial reconstructions is only calculated on the pre-detected bounding boxes.
+
+![](images/353c8bda89cf67445284c9a35188dad4bc74df5b7b498bec08ef0bae8c6d4a9e.jpg)  
+Figure 7. Performance comparison on instance segmentation (left) and pose estimation (right) on COCO.
+
+of all categories and human objects using 40 and 20 images, which are randomly selected from the COCO 2017 validation set. The PSNR of the region of interest serves as the metric for measuring the objective quality. Additionally, the bits per pixel (bpp) are calculated by dividing the total bits of the transmitted bitstream by the number of pixels in the region of interest. To verify the effectiveness of our proposed method in supporting downstream tasks, we conduct experiments on instance segmentation and pose estimation using the COCO 2017 validation set. [45]. This dataset is widely used for dense prediction tasks and comprises 118K training, 5K validation, and 20K test-dev images. We evaluate the performances of instance segmentation and human keypoint detection using Mask R-CNN (X101-FPN backbone) and Keypoint R-CNN (X101-FPN backbone), respectively, implemented in the Detectron2 toolbox [68].
+
+# 4.2. Quantitative Results
+
+For comparisons, we choose recent neural image codes [16,27,52,75], powerful classical image codes [10, 61], and the functional codec that supports semantically structured bitstream [62]. Among them, ELIC [27] is reproduced and performs closely as their report. It is crit-
+
+ical to emphasize that the encoder does not possess prior knowledge regarding the regions of interest required by the decoder. And we utilize object detection as the pre-analysis by default, which is one of the most fundamental visual analysis tools. The semantically structured bitstreams are generated based on the bounding boxes detected by the detector of Mask R-CNN (X101-FPN backbone), and are selectively decoded and reconstructed into images for evaluation according to the requirements of relative semantics. In all experiments of quantitative results, the block size  $B$  of group masks is set to 32 to balance the trade-off between performance and overhead bitrate, and we take the union of overlapping groups as one group. More experiments compared with other methods of image coding for machines are presented in the supplementary.
+
+Reconstruction Quality. For entire image reconstruction, Figure. 6a demonstrates that our proposed model achieves state-of-the-art rate-distortion performance, surpassing VTM 18.2 in terms of PSNR at all bitrates. Notably, structuring the bitstream according to semantics introduces negligible effects. Therefore, our model has a strong potential to achieve both high coding efficiency and functionality simultaneously. For partial image reconstruction in scenarios of specific interest, as shown in Figure. 6b and Figure. 6c, our model has achieved a significant improvement compared to other CODECs. Specifically, the semantically structured bitstream enabled our model and SSIC [62] to avoid transmitting and decoding bitstreams corresponding to the entire image, which is required of ELIC [27]. Additionally, our model's flexible block-level group mask and group-independent transform spatially decouple latents from the semantic level, resulting in sharper and more re
+
+![](images/9a7d6896638011c7da58bfb5d5a4d44d52557be07a841e5bd38414a2e789d8df.jpg)
+
+![](images/d26c13f5ecbf86b15f0e1465c92066dad84819ca749f521a992c3f2938c59cc1.jpg)  
+0.516/29.36
+
+![](images/b82832be0cdaefe0fb89afa2c13c282dce936f6ee9ef022d4683c3edd1066af1.jpg)  
+0.251/29.40
+
+![](images/91153e25e2b2538a795999db486136efc42bc8df64aa3cf632dee8600fcbb228.jpg)  
+0.195 / 30.14
+
+![](images/2321acaaa637c289a5c0a1f8e3f959ffd0cb018b4c55e85bea0cd98459d59d91.jpg)
+
+![](images/2e1c2e28dc631c2fdcd56c02a5221d6e48d989aa79fe96945ba03b26ca87ddfb.jpg)
+
+![](images/f357cb43ff7ddff1b85d23af659ca4e98589f68dcb4b8cd5c7ee436d77d2a67d.jpg)
+
+![](images/60b72b7b5778c6a5edad1e354da5f0f2a811a9de8c487dbf196f0444f14857a6.jpg)
+
+![](images/6596592032ef0de60f2972318ae404dc93639bf21f1ce1008c80b23f6288bb76.jpg)  
+Original Image
+
+![](images/5815dc1ed2654d37d11e6a5ea48fc4b30b7d7b13e331035ac5fda3288bfc736b.jpg)  
+0.318/28.61  
+ELIC
+
+![](images/c9a827a04c52c7686e159b55bdd6a0952adfae61c751cdbf42badde15c87a27c.jpg)  
+0.295/28.53  
+SSIC
+
+![](images/c8c1e82b2c0da3fcbbfaaee39d9ebac3dd039bccb0aceaf89619198a4d7d54f8.jpg)  
+0.316 / 29.25  
+Ours
+
+![](images/5f2953d350535e7801e274afa9f8d1063982131c77f25427a099cd4d84f3be2e.jpg)
+
+![](images/adece8258b93916b6efd26abe44ab7f10348c55c69fafce5b5d6b4d2f710cfce.jpg)
+
+![](images/9911aecd9cc119cf8b57d5b021fd2ab6a46f042fb417a95b1bd287df5dc9bbdd.jpg)
+
+![](images/2ea4415c651ebfca4edbf6c5af7750d68a43b3fac712ef5bb3eccb1f4c7dfe15.jpg)
+
+![](images/608398d18aaf1880d92b461c09528722f51fa2b0b3bbda7116c4551a529a0d70.jpg)
+
+![](images/2ade7db95446bd023e7632e8331c570c6f4ac7f7e4808aec24940599d2122fba.jpg)
+
+![](images/c2fc8807996df6a36b79859eeeca1e3317e3151ff8419059fb1432c567730fc3.jpg)
+
+![](images/8415f5019a1d8711b9a137210601a8f997b8e81047da306068fcf626e1d7b406.jpg)  
+Figure 8. Visualization of reconstructed images for human consumption from the COCO 2017 dataset, in the scenario where foreground objects are the regions of interest. The numbers below the images correspond to bits per pixel and PSNR.
+
+![](images/a13a0777271ce508fd1754ea8677adf56d6b83d632d1b3afe9937524c4a52099.jpg)
+
+![](images/aa1c98fa10d31e1c2136f8e7d61b4021f37b7bae999e9f18f5f1ff57e58abf21.jpg)
+
+![](images/861d3d4a4d8256d62b1215bbe6b9b627198ca1d1994661459f2a6fda843512dc.jpg)  
+1.850bpp
+
+![](images/330b8d1349b7574c9d6833e3b9f55fc3ebf4f4452a1c3e7aa65971fb0e685025.jpg)  
+1.592bpp
+
+![](images/49fa984a24ca38df917f62282ed49a2bea31b9f456fb45145a7541e5c816f5ff.jpg)  
+1.211bpp
+
+![](images/acd3064c9b691bceecd9ea5b338e637633cc619e06e0d49549e35befe6077a76.jpg)  
+Original Image  
+Figure 9. Visualization of reconstructed images for downstream tasks from the COCO 2017 dataset. The first line corresponds to instance segmentation, and the second one corresponds to pose estimation.
+
+![](images/1506b4316dbd3ba8017e120cfb41ee1f8d2d8f01e5922c859f08579b8a1222cf.jpg)  
+Ground Truth
+
+![](images/f7f548d9eee40c0acab91e4583932cf93071d80c036a465f43df8a5714770586.jpg)  
+2.229bpp  
+ELIC
+
+![](images/d811158090ed31fc98c9a7d750aff7fa179417fb1a258abe2e11091460623ccb.jpg)  
+1.335bpp  
+SSIC
+
+![](images/8f696fbfa4d9872515a66aa0bee837dbf4d4afb5e2e216a55c05aae7b68a03eb.jpg)  
+1.048bpp  
+Ours
+
+alistic boundaries compared to SSIC [62]. Moreover, in the case of overlapping regions, our model can significantly save bitrate by grouping them into an irregular group instead of replacing them with a larger bounding box, which would introduce more distortion.
+
+Downstream Task Supporting. Figure. 7 demonstrates our excellent performance on instance segmentation and pose estimation. On instance segmentation, our model achieves significantly improved performance at low bitrates (<0.6bpp) compared with other methods, which can be attributed to the faithful reconstruction of RoI boundaries in our approach. It is worth noting that the superior performance of our method on pose estimation is due to the ability of our model to preserve fine details and boundaries, which are critical for accurately localizing human keypoints. In addition, human objects are a minority in the image and are often sparsely distributed, and compared with SSIC, our method can significantly reduce bitrate by avoiding transmitting the latents corresponding to the rectangular area containing all overlapped objects.
+
+# 4.3. Qualitative Results
+
+Reconstruction for Human Consumption. When the reconstructed images are mainly intended for human con
+
+sumption with specific semantic interests, as shown in Figure 8, bitstreams generated by CODEs designed for compressing regular rectangular images are required to be fully transmitted and serve full reconstruction, regardless of the requirements and the content of images. Although SSIC [62] partially addresses this issue through generating SSB through bounding box based division, it lacks flexibility and may introduce additional irrelevant disturbance when merging overlapping objects into larger bounding boxes. Moreover, SSIC generates the bitstream by directly dividing the corresponding positions in the latent variable space after the transform based on ConvNets, which can lead to lost dependencies during selective reconstruction, resulting in blurred and distorted region boundaries. With group mask based partitioning and group-independent transform, our method decouples images to generate the SSB more efficiently while ensuring that no distortion or blurring is introduced after selective reconstruction, leading to both significant bitrate saving and a better visual experience.
+
+Reconstruction for Intelligent Tasks. When using reconstructed images for downstream intelligent tasks, selectively transmitting and decoding the semantically structured bitstream based on the prior knowledge of relative seman
+
+![](images/799f42b5c5c2dcb10d91b8b35101445a0edbdc4ea3afb0a026a1b970294ae9bf.jpg)
+
+![](images/e7fa75907c6ea135bb0420e8e2dd0678a0a05ebfef73509e511ed710500ac633.jpg)  
+Original Image
+
+![](images/37fc29c7ed7d0877f883a4ea8c36c41c4fa9ce64956e6d5946ec7108b3b187eb.jpg)
+
+![](images/d4db6a3afb325cfeeef3f20e5835f794e0dca570673340f8d777d1a6a7c915ac.jpg)  
+Pre-Analysis
+
+![](images/a2799c7805e83a4a04e9f5a0784d05410ae31451129c4fd211fa829470a7017b.jpg)
+
+![](images/43a71bd43fd4e03274e00f62bdb929c436d733b3f0c4338dd5c132687033a308.jpg)  
+Based on
+
+![](images/8769c172d99d89cd87ba721a8bb8fcb780119bb35cd6cf6d0b62192c2879d15d.jpg)
+
+![](images/0f6557947609d733be5376ff9746a43819fdc09b4b7e1962eeb2d422f0efc629.jpg)  
+Bounding Boxes  
+Based on  
+Instance Maps
+
+![](images/6fa09b9e4356de1d3cfc882bdc6df10b92f7f6b79b3643749b9d5c70e87784fe.jpg)  
+Figure 10. Selective reconstructions from the semantically structured bitstream. The bitstream can be generated based on different pre-analysis methods.  
+Original Image  
+Figure 11. Illustration of semantically-aware encryption based on our proposed method.
+
+![](images/b261705210627fc10e0391bf0a3ff55d4f1a540e4577571c0703c93c0f99b029.jpg)  
+Pre-Analysis
+
+![](images/0c2f30714811b2af2c34764895014333a29fcf7d3308a8cf5052e087eede41ef.jpg)  
+Encrypt People
+
+![](images/b773c9f4052d737dee2ed6239d88df1083153744309efc5282895de23c2fcdcc.jpg)  
+Encrypt People  
+and Vehicles
+
+tics can significantly save bitrate. As illustrated in Figure 9, conventional codecs designed for compressing rectangular images require the entire bitstream to be transmitted and decoded. Downstream models then perform intelligent analytics on the fully reconstructed image, resulting in a waste of bitrate. Similar to image reconstruction for human consumption, SSIC can perform selective transmission to save bitrate. However, its bounding box based partitioning may not be optimal for bitrate savings, and the resulting blurred and distorted region boundaries and irrelevant content can further impede downstream intelligent analytics. (misidentifying the car as a person shown in Figure 9). Our method can more efficiently generate the SSB and support various downstream tasks with specific requirements with both accuracy and coding efficiency.
+
+Customizability and Flexibility. It's flexible to customize the semantically structured bitstream based on different semantical partition criteria. For instance, as shown in Figure 10, the generation and usage of semantically structured bitstreams can be based on either object detection or instance segmentation, depending on the specific requirements. It should be noted that the methods of pre-analysis are not limited to object detection and instance segmentation as shown in Figure 10, but can include other techniques such as saliency detection [22, 30, 72], semantic segmentation [5, 48, 69, 71, 73], panoptic segmentation [15, 36], etc., and even human annotation.
+
+Semantically-Aware Encryption. As presented in Figure 11, our proposed method can be applied to bitstream encryption, enabling selective and even layered encryption with semantic priors based on the user's security level. In some cases, the selectively encrypted SSB can allow for secure transmission and storage of sensitive information while minimizing the impact on visual quality and downstream analytics. Please refer to the supplementary material for
+
+![](images/456c61c5f8400fe58ecc3baef621b45b8a82c4c31dc7286b67f525c6784f8833.jpg)  
+(a) RoI reconstruction.  
+Figure 12. Ablation study.
+
+![](images/c3bcf2a27219e7eb2317217dfe73097f5e1251e12e9589ea6cdc245eef733f9d.jpg)  
+(b) Intelligent task supporting.
+
+details of implementation.
+
+# 5. Ablation Study
+
+In order to show the effectiveness of our proposed group mask and group-independent transform (GIT) for both RoI-aware selective reconstruction and downstream intelligent analytics, we conducted the ablation study at different bitrates. Specifically, we use people as the region of interest for RoI reconstruction, and the results for taking all foreground objects as RoI are presented in the supplementary material. Additionally, we conducted an ablation study on pose estimation of COCO 2017 for intelligent task support, and the results on instance segmentation are also included in the supplementary material.
+
+Group Mask. Figure 12 shows that group mask based partitioning saves significant bitrate compared with naive bounding box based partitioning. Meanwhile, both the visual quality of RoI reconstruction and downstream task performances would not be impeded.
+
+Group-Independent transform. As shown in Figure 12a, group-independent transform is crucial for the visual quality of RoI reconstruction. And the downstream task performance would be impeded by the blurred and distorted region boundaries without group-independent transform, which is shown in Figure 12b.
+
+# 6. Conclusion
+
+In this work, we propose to generate semantically structured bitstream with strong functionality based on group masks, which are highly flexible, customizable, while maintaining lightweight overhead. Moreover, we first propose the concept of group-independent transform and instantiate it by designing the Group-Independent Swin-Block (GI-Swin Block) to ensure independence among distinct groups, resulting in pleasing reconstructions of RoI regions. Particularly, our proposed method outperforms VTM-18.2 and achieves comparable coding efficiency with the SOTA neural image CODECs, while offering strong functionality. The experimental results have demonstrated the effectiveness of our proposed method across several applications, including human consumption of regions of interest, downstream task support, and semantically-aware encryption.
+
+Acknowledgement. This work was supported in part by NSFC under Grant U1908209, 62021001. This work was also supported in part by ZJNSFC under Grant LQ23F010008.
+
+# References
+
+[1] Eirikur Agustsson, Fabian Mentzer, Michael Tschannen, Lukas Cavigelli, Radu Timofte, Luca Benini, and Luc V Gool. Soft-to-hard vector quantization for end-to-end learning compressible representations. *NeurIPS*, 30, 2017. 2  
+[2] Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool. Generative adversarial networks for extreme learned image compression. In ICCV, pages 221-231, 2019. 2  
+[3] Mohammad Akbari, Jie Liang, and Jingning Han. Dsslic: Deep semantic segmentation-based layered image compression. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2042-2046. IEEE, 2019. 3  
+[4] Artem Babenko, Anton Slesarev, Alexandr Chigorin, and Victor Lempitsky. Neural codes for image retrieval. In ECCV, pages 584-599. Springer, 2014. 1, 3  
+[5] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI, 39(12):2481-2495, 2017. 1, 3, 8  
+[6] Johannes Ballé, Philip A Chou, David Minnen, Saurabh Singh, Nick Johnston, Eirikur Agustsson, Sung Jin Hwang, and George Toderici. Nonlinear transform coding. IEEE Journal of Selected Topics in Signal Processing, 15(2):339-353, 2020. 2  
+[7] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In *ICLR*, 2017. 1, 2  
+[8] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In ICLR, 2018. 1, 2  
+[9] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, pages 9157-9166, 2019. 1, 3  
+[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. TCSVT, 2021. 1, 2, 6  
+[11] 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. TPAMI, 40(4):834-848, 2017. 1, 3  
+[12] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 1, 3  
+[13] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. TIP, 29:2230-2243, 2019. 1, 3
+
+[14] Zhuo Chen, Kui Fan, Shiqi Wang, Ling-Yu Duan, Weisi Lin, and Alex Kot. Lossy intermediate deep learning feature compression and evaluation. In ACM MM, pages 2414-2422, 2019. 1, 3  
+[15] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, pages 12475-12485, 2020. 8  
+[16] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In CVPR, pages 7939-7948, 2020. 1, 2, 6  
+[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 3  
+[18] Ling-Yu Duan, Vijay Chandrasekhar, Jie Chen, Jie Lin, Zhe Wang, Tiejun Huang, Bernd Girod, and Wen Gao. Overview of thempeg-cdvs standard. TIP, 25(1):179-194, 2015. 1, 3  
+[19] Ling-Yu Duan, Yihang Lou, Yan Bai, Tiejun Huang, Wen Gao, Vijay Chandrasekhar, Jie Lin, Shiqi Wang, and Alex Chichung Kot. Compact descriptors for video analysis: The emerging mpg standard. IEEE MultiMedia, 26(2):44-54, 2018. 1, 3  
+[20] Touradj Ebrahimi and Caspar Horne. Mpeg-4 natural video coding—an overview. Signal Processing: Image Communication, 15(4-5):365-385, 2000. 3  
+[21] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In ECCV, pages 510-528. Springer, 2022. 2, 3  
+[22] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. TPAMI, 34(10):1915-1926, 2011. 8  
+[23] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 2  
+[24] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. TCSVT, 32(4):2329-2341, 2021. 2  
+[25] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Soft then hard: Rethinking the quantization in neural image compression. In ICML, pages 3920-3929. PMLR, 2021. 2  
+[26] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. NeurIPS, 34:15908-15919, 2021. 1, 3  
+[27] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In CVPR, pages 5718-5727, 2022. 1, 2, 6  
+[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3  
+[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 3
+
+[30] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In CVPR, pages 1-8. IEEE, 2007. 8  
+[31] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3  
+[32] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. ECCV, 2022. 1, 3  
+[33] Xin Jin, Ruoyu Feng, Simeng Sun, Runsen Feng, Tianyu He, and Zhibo Chen. Semantically video coding: Instill static-dynamic clues into structured bitstream for ai tasks. arXiv preprint arXiv:2201.10162, 2022. 3  
+[34] Nick Johnston, Damien Vincent, David Minnen, Michele Covell, Saurabh Singh, Troy Chinen, Sung Jin Hwang, Joel Shor, and George Toderici. Improved lossy image compression with priming and spatially adaptive bit rates for recurrent networks. In CVPR, pages 4385-4393, 2018. 1  
+[35] Aggelos K Katsaggelos, Lisimachos P Kondi, Fabian W Meier, Jorn Ostermann, and Guido M Schuster. Mpeg-4 and rate-distortion-based shape-coding techniques. Proceedings of the IEEE, 86(6):1126-1154, 1998. 3  
+[36] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, pages 9404-9413, 2019. 8  
+[37] Eastman Kodak. Kodak lossless true color image suite (photocd pcd0992). http://r0k.us/graphics/kodak, 1993.5  
+[38] Nam Le, Honglei Zhang, Francesco Cricri, Ramin GhaznaviYouvalari, and Esa Rahtu. Image coding for machines: An end-to-end learned approach. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1590-1594. IEEE, 2021. 1, 3  
+[39] Nam Le, Honglei Zhang, Francesco Cricri, Ramin Ghaznavi Youvalari, Hamed Rezazadegan Tavakoli, and Esa Rahtu. Learned image coding for machines: A content-adaptive approach. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1  
+[40] Mu Li, Wangmeng Zuo, Shuhang Gu, Jane You, and David Zhang. Learning content-weighted deep image compression. TPAMI, 2020. 1  
+[41] Mu Li, Wangmeng Zuo, Shuhang Gu, Debin Zhao, and David Zhang. Learning convolutional networks for content-weighted image compression. In CVPR, pages 3214-3223, 2018. 1  
+[42] Xin Li, Jun Shi, and Zhibo Chen. Task-driven semantic coding via reinforcement learning. TIP, 2021. 3  
+[43] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. ECCV, 2022. 1, 3  
+[44] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 1, 3  
+[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence
+
+Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 5, 6  
+[46] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, pages 8759-8768, 2018. 1, 3  
+[47] 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, pages 10012-10022, 2021. 1, 2, 3, 4  
+[48] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 1, 3, 8  
+[49] Siwei Ma, Xiang Zhang, Shiqi Wang, Xinfeng Zhang, Chuanmin Jia, and Shanshe Wang. Joint feature and texture coding: Toward smart video representation via front-end intelligence. TCSVT, 29(10):3095-3105, 2018. 1, 3  
+[50] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Conditional probability models for deep image compression. In CVPR, pages 4394-4402, 2018. 1  
+[51] Fabian Mentzer, George Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. arXiv preprint arXiv:2006.09965, 2020. 1, 2  
+[52] David Minnen, Johannes Balle, and George Toderici. Joint autoregressive and hierarchical priors for learned image compression. In NeurIPS, 2018. 1, 2, 5, 6  
+[53] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 5  
+[54] Majid Rabbani and Rajan Joshi. An overview of the JPEG 2000 still image compression standard. Signal processing: Image communication, 17(1):3-48, 2002. 1, 2  
+[55] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1, 3  
+[56] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, pages 7263-7271, 2017. 1, 3  
+[57] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28:91-99, 2015. 1, 3  
+[58] Oren Rippel and Lubomir Bourdev. Real-time adaptive image compression. In ICML, pages 2922-2930. PMLR, 2017. 2  
+[59] Thomas Sikora. Thempeg-4 video standard verification model. TCSVT, 7(1):19-31, 1997. 3  
+[60] Saurabh Singh, Sami Abu-El-Haija, Nick Johnston, Johannes Balle, Abhinav Shrivastava, and George Toderici. End-to-end learning of compressible features. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3349–3353. IEEE, 2020. 1, 3  
+[61] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. TCSVT, 22(12):1649-1668, 2012. 1, 2, 6  
+[62] Simeng Sun, Tianyu He, and Zhibo Chen. Semantic structured image coding framework for multiple intelligent applications. TCSVT, 2020. 2, 3, 6, 7
+
+[63] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. arXiv preprint arXiv:1703.00395, 2017. 2  
+[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3, 4  
+[65] Gregory K Wallace. The JPEG still picture compression standard. IEEE transactions on consumer electronics, 38(1):xviii-xxxiv, 1992. 1, 2  
+[66] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. TMM, 2021. 3  
+[67] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. TCSVT, 13(7):560-576, 2003. 1  
+[68] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6  
+[69] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS, 34:12077-12090, 2021. 1, 3, 8  
+[70] Yibo Yang, Robert Bamler, and Stephan Mandt. Improving inference for neural image compression. NeurIPS, 33:573-584, 2020. 2  
+[71] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, pages 2881-2890, 2017. 8  
+[72] Rui Zhao, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Saliency detection by multi-context deep learning. In CVPR, pages 1265-1274, 2015. 8  
+[73] 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 CVPR, pages 6881-6890, 2021. 1, 3, 8  
+[74] Lei Zhou, Zhenhong Sun, Xiangji Wu, and Junmin Wu. End-to-end optimized image compression with attention mechanism. In CVPR workshops, page 0, 2019. 5  
+[75] Yinhao Zhu, Yang Yang, and Taco Cohen. Transformer-based transform coding. In ICLR, 2022. 1, 2, 4, 6

2023/Semantically Structured Image Compression via Irregular Group-Based Decoupling/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d2b0b40ab3c1a7892d32c706c19675285b8b8532f66d0e623351a278c32fda9
+size 661304

2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/0da0877f-f78d-4f97-9a49-55c50b2292de_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d88715185fb924acb4113735c2a10b31bef8875837bc7837f221e9e9d6a23f8
+size 3914233

2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/full.md

@@ -0,0 +1,381 @@
+# Semantics Meets Temporal Correspondence: Self-supervised Object-centric Learning in Videos
+
+Rui Qian<sup>1</sup> Shuangrui Ding<sup>1</sup> Xian Liu<sup>1</sup> Dahua Lin<sup>1,2*</sup>
+
+<sup>1</sup>The Chinese University of Hong Kong <sup>2</sup>Shanghai Artificial Intelligence Laboratory
+
+{qr021,ds023,lx021,dhlin}@ie.cuhk.edu.hk
+
+# Abstract
+
+Self-supervised methods have shown remarkable progress in learning high-level semantics and low-level temporal correspondence. Building on these results, we take one step further and explore the possibility of integrating these two features to enhance object-centric representations. Our preliminary experiments indicate that query slot attention can extract different semantic components from the RGB feature map, while random sampling based slot attention can exploit temporal correspondence cues between frames to assist instance identification. Motivated by this, we propose a novel semantic-aware masked slot attention on top of the fused semantic features and correspondence maps. It comprises two slot attention stages with a set of shared learnable Gaussian distributions. In the first stage, we use the mean vectors as slot initialization to decompose potential semantics and generate semantic segmentation masks through iterative attention. In the second stage, for each semantics, we randomly sample slots from the corresponding Gaussian distribution and perform masked feature aggregation within the semantic area to exploit temporal correspondence patterns for instance identification. We adopt semantic- and instance-level temporal consistency as self-supervision to encourage temporally coherent object-centric representations. Our model effectively identifies multiple object instances with semantic structure, reaching promising results on unsupervised video object discovery. Furthermore, we achieve state-of-the-art performance on dense label propagation tasks, demonstrating the potential for object-centric analysis. The code is released at https://github.com/shvdiwnkozbw/SMTC.
+
+# 1. Introduction
+
+As one of the fundamental cognitive capabilities, human beings easily distinguish different objects, establish visual
+
+![](images/06fe90c013fc77671267ef4b2e01a9a4c5c5ec86500384249bf7b59b4041c790.jpg)  
+(a) Query slot attention on semantic feature.
+
+![](images/95713c39e96e743b2ba94bccd45808cf70187045eaa76c8e70f3830f7ba15b8e.jpg)
+
+![](images/f54196344cb58166b76125737f0d095a1fff030f384f1f1f8b67aba06fea8ca8.jpg)  
+(b) Random sampling slot attention on correspondence map.
+
+![](images/a9fc3ad0a33a1ce8c341f7a5026395f04b9f055999083bf9257420247028d229.jpg)  
+Figure 1. Fig. 1(a) presents the results of query slot attention on top of the RGB feature map. It successfully decomposes different semantics, e.g., camels and fence. Fig. 1(b) visualizes the correspondence map after PCA dimensionality reduction, showing that different instances have different correspondence patterns. And the slot attention with random sampling coarsely distinguishes two camels with some redundant borders. Best viewed in color.
+
+correspondence and perform object-centric analysis from temporally continuous observations. This ability can be attributed to two indispensable visual mechanisms: high-level semantic discrimination as well as low-level temporal correspondence, which enable humans to effectively understand and interact with the world.
+
+Motivated by this, computer vision researchers equip machines with these capacities to enhance object-centric perception [79, 60, 34]. To achieve this goal, early works rely on human annotations or weak supervision to perceive object semantics [60, 63, 13], identify geometric positions [39, 71, 31, 91], and establish temporal correspondence [41, 79, 75, 77], but their generalization ability is limited. Recently, there emerge a host of fully unsupervised methods to learn robust representations for semantic discrimination [14, 38, 33, 70, 29, 37, 12] or spatiotemporal correspondence [44, 81, 86, 55, 52, 43], which achieve promising performance. Given this encouraging re
+
+sult, we naturally come up with a question: Is it possible to jointly leverage the semantics and correspondence to discover object instances and distill object-centric representations without human annotations?
+
+Regarding this problem, our intuition is that the high-level semantics delineates meaningful foreground areas in a top-down manner, while when looking into more frames, the low-level correspondence temporally associates coherent objects and separates individual instances in a bottom-up fashion. For instance, in a football scene, the semantic cue differentiates the foreground that includes several players, while the temporal correspondence links distinct players through dynamic movements and geometric relationships. These two aspects collectively contribute to object-centric representations. Unfortunately, most of the existing works only concentrate on one of these features. [14, 38, 33, 70, 82] succeed in developing high-level semantics, but this abstract semantics alone is insufficient to distinguish instances. Whereas, [44, 81, 78, 55] excel in detailed correspondence, but lack semantic structure and result in redundancy and ambiguity.
+
+In this paper, we propose a new architecture, Semantics Meets Temporal Correspondence (SMTC), to jointly leverage semantics and temporal correspondence to distill object-centric representations from RGB sequences. Specifically, we first extract frame-wise visual features as the semantic representation. Then we calculate dense feature correlations between adjacent frames as the correspondence map which encodes temporal relationships. To mine the object-centric knowledge, we take inspiration from [59, 88, 50, 25], and investigate using different formulations of slot attention on them. The preliminary experiments show that the original slot attention with random sampling on RGB feature map suffers from complex scene components in real-world videos [59, 72], but the revised query slot attention [88, 46] can decompose different semantic components as shown in Fig. 1(a). As for the correspondence map, different objects present diverse temporal correspondence patterns. Comparing to semantic features, these patterns reveal low-level geometric relationships, which are comparatively simple but vary with specific scenes. Hence, query slot attention fails but the random sampling based formulation performs surprisingly well as shown in Fig. 1(b), coarsely separating different object areas with some redundant borders.
+
+Motivated by this, we develop semantic-aware masked slot attention, which comprises a set of Gaussian distributions with learnable mean and standard deviation vectors, on top of the fused semantic and correspondence representations. The intuition is that the mean vectors could represent potential semantic centers, which act similarly to the query slot attention to separate semantic components. While the deviation vectors introduce perturbations around
+
+the semantic centers to capture distinct temporal correspondence patterns of different instances. Technically, we formulate two slot attention stages to achieve this goal. Firstly, we use the mean vectors as slot initialization to generate semantic segmentation masks. Secondly, for each semantics, we randomly sample slot vectors from the Gaussian distribution, then perform iterative attention and masked aggregation within the corresponding semantic mask area to distinguish instances. We enforce temporal consistency on the semantic masks as well as object instance slots to enhance temporal coherency and refine object-centric representations. Comparing with existing works on object-centric learning in videos [25, 50, 88, 84], our model is free of pre-computed motion or depth prior, and explicitly identifies multiple objects with semantic structure.
+
+In summary, our contributions are: (1) We propose a novel self-supervised architecture that unifies semantic discrimination and temporal correspondence to distill object-centric representations in videos. (2) We demonstrate that simple feature correlation can effectively represent temporal correspondence cues when used in conjunction with semantic features. Building on this observation, we develop semantic-aware masked slot attention, which operates on fused visual features and correspondence maps, to distinguish multiple object instances with semantic structure without relying on motion or depth priors. (3) We achieve promising results on unsupervised object discovery in both single and multiple object scenarios, and reach state-of-the-art performance on label propagation tasks, demonstrating that we learn discriminative and temporally consistent object-centric representations.
+
+# 2. Related Work
+
+Unsupervised Object Discovery is an essential process to formulate object-centric representations, which aims to identify objects without human annotations. There exist a series of works focusing on this problem [59, 32, 26, 27, 28, 7]. Typically, [59] develops slot attention to iteratively update latent object representations from random initializations but it has difficulty scaling to complex real-world scenes. To tackle this challenge, [72] employs feature reconstruction as objective to reduce redundancy, [46] develops query slot initialization to encode visual concepts. Further, a line of works extend object discovery to video domain [50, 48, 17, 5, 88]. Most of them build on slot attention architecture, and adopt extra pre-computed priors, e.g., optical flow [88, 84, 16, 21], depth [25], geometric positions [50], as input or supervision to assist object discovery. In our work, we find that the simple feature correlation provides informative temporal cues. And we develop semantic-aware masked slot attention to identify object instances with semantic structure without resorting to extra priors.
+
+Self-supervised Representation Learning aims to learn
+
+![](images/fe93decb9a5c6db629d51dcb51e2632af2e45a5db995b6fabaad9a7b39399dd8.jpg)  
+Figure 2. An overview of our framework. We first extract frame-wise features and calculate dense feature correlation, then fuse them to pass through semantic-aware masked slot attention, which comprises two slot attention stages with  $N$  shared learnable Gaussian distributions. In the first semantic slot attention stage, the mean vectors serve as slot initialization to generate a set of segmentation masks for semantic decomposition. In the second masked slot attention stage, which runs on  $N$  semantics in parallel, we randomly sample slots from the Gaussian distribution of each semantics and perform masked feature aggregation within the semantic area to identify distinct instances. We enforce semantic and instance temporal consistency to train the architecture in a teacher-student manner, with the teacher marked in gray.
+
+robust representation without human annotations. Early works design various pretext tasks to generate pseudo labels as self-supervision [49, 30, 47, 10, 1, 62]. Later, contrastive learning with instance discrimination revolutionized the field [38, 14, 70, 29, 35, 83, 65, 42, 11, 20, 3, 68, 35, 36, 69, 19, 22]. These works generally contribute to general semantics but lack the ability to distill object-centric representations. To delve into this problem, [82, 85] preserve spatial sensitivity and perform contrastive learning on dense features. [86, 12] point out that semantic pre-training implicitly encodes object correspondence. While in our work, we learn to explicitly identify different objects with instance-level consistency to progressively refine object-centric representations.
+
+Self-supervised Correspondence is a fundamental problem in computer vision [55, 78, 81, 80]. Typical works rely on low-level statistics like color to generate the dense correspondence flow [51, 52, 78], or establish temporal cycles to track image regions [81, 44, 6]. Further, [43, 86, 12] incorporate correspondence learning with semantic discrimination, but they either regard correspondence as a side effect of semantic discrimination [86, 12] or independently learn two pathway features with late fusion [43]. Among these works, high-level semantics is not well utilized, and there exists high redundancy due to large background areas. In contrast, our method explicitly separates semantic components to alleviate redundancy and formulate instance-level correspondence from an object-centric perspective.
+
+# 3. Method
+
+Our framework is shown in Fig. 2, which adopts a teacher-student structure similar to recent self-supervised methods [38, 12, 33]. Specifically, we first extract framewise features from an RGB sequence, and calculate dense feature correlation as temporal correspondence representations. After that, we fuse them to pass through semantic-aware masked slot attention to separate multiple object instances with semantic structure, and enforce temporal consistency to refine the object-centric representations.
+
+# 3.1. Feature Encoding
+
+Given a RGB sequence  $v = \{I_1, I_2, \dots, I_T\}$ , we employ a shared visual encoder  $f$  to extract frame-wise features:
+
+$$
+F _ {t} = f \left(I _ {t}\right) \in \mathbb {R} ^ {H W \times D}, \tag {1}
+$$
+
+where  $H, W, D$  respectively denotes height, width and channel dimension,  $t$  indicates frame index. For simplicity, we omit the subscript indicating student or teacher pathway, and use  $\hat{F}$  to denote teacher outputs. Intuitively,  $F_{t}$  contains appearance and semantic cues in the input sequence, and the next step is to formulate an effective representation for the temporal correspondence cues. To do this, inspired by the cost volume in motion estimation [24, 73, 74], we calculate dense feature correlation to generate the temporal
+
+<table><tr><td>Feature</td><td>Query</td><td>Random</td></tr><tr><td>RGB</td><td>58.1</td><td>38.4</td></tr><tr><td>Correlation</td><td>21.5</td><td>56.3</td></tr></table>
+
+Table 1. Preliminary results on RGB feature map and feature correlation. 'Query' ('Random') denotes slot attention with query (random sampling) initialization. We report IoU on DAVIS-2016.
+
+correspondence map for each timestamp:
+
+$$
+C _ {t j} = F _ {t} F _ {j} ^ {T} \in \mathbb {R} ^ {H W \times H W}, \tag {2}
+$$
+
+where we randomly sample one frame index  $j \neq t$  to compute feature correlation. The channel activations of  $C_{tj}$  encode low-level geometric correlations between frame  $t$  and  $j$ , which indicates potential partitions of different objects as illustrated in Fig. 1(b).
+
+To jointly leverage the semantic and correspondence cues, we employ two linear transformation heads,  $h_f$  and  $h_c$ , to respectively project  $F_t$  and  $C_{tj}$  to a shared  $D$ -dimensional embedding space, which are then fused with element-wise summation:
+
+$$
+R _ {t} = h _ {f} \left(F _ {t}\right) + h _ {c} \left(C _ {t j}\right) \in \mathbb {R} ^ {H W \times D}. \tag {3}
+$$
+
+In this way,  $R_{t}$  simultaneously encodes semantic and temporal correspondence information, serving as an intermediate feature for object-centric analysis.
+
+# 3.2. Preliminary Experiments with Slot Attention
+
+Given these representations, the next step is to formulate effective ways to identify individual objects. To do this, we take inspiration from slot attention [59, 50, 88, 46], which employs a set of slot vectors to iteratively attend to specific objects. Generally, there are two alternatives for slot initialization, random sampling from a Gaussian distribution with learnable parameters  $\mu, \sigma \in \mathbb{R}^D$  [59, 50] or directly inheriting from a set of learnable queries [46, 88].
+
+To investigate the effectiveness of these two formulations, we conduct preliminary experiments on the semantic feature  $F_{t}$  and correspondence map  $C_{tj}$  respectively. We follow the settings in [72], using a frozen DINO pre-trained ViT to extract visual feature and generate correspondence map, on top of which we perform two kinds of slot attention. We train and evaluate them on DAVIS-2016 dataset [66]. As indicated by Fig. 1 and Table. 1, we find that query slot attention decomposes distinct semantic components on  $F_{t}$ , while the original slot attention with random sampling performs much worse and struggles in complex real-world scenes. Conversely, query slot attention does not work on the correspondence map, and random sampling initialization enables the model to effectively utilize the temporal correspondence patterns in  $C_{tj}$  to distinguish instances. An intuitive explanation into these observations is the RGB
+
+features contain rich semantics and there exist consistent patterns for similar objects. Slot attention with learnable queries can capture inter-sample semantic patterns to identify objects. While random initialization fails to memorize semantic patterns, thus resulting in worse performance. In contrast, correspondence features reveal inter-frame geometric correlations, shown in Fig. 1(b), which vary with specific scenes. There are no consistent patterns for similar objects, so the learnable queries fail but random initialization works well to capture correspondence cues. The complementarity between these two conditions motivates two-stage slot attention design to combine the attributes of query and random sampling initialization, thus better exploiting the semantic and correspondence cues in  $R_{t}$ .
+
+# 3.3. Semantic-aware Masked Slot Attention
+
+We propose semantic-aware masked slot attention, which comprises  $N$  Gaussian distributions with learnable means  $\pmb{\mu} = \{\mu_1,\mu_2,\dots,\mu_N\} \in \mathbb{R}^{N\times D}$  and standard deviations  $\sigma = \{\sigma_{1},\sigma_{2},\dots,\sigma_{N}\} \in \mathbb{R}^{N\times D}$ . With this design, the mean vectors act similarly to the learnable queries, each representing a potential semantic center. And the deviations introduce perturbations around semantic centers to model temporal correspondence patterns. To achieve this goal, we employ two stages of iterative attention with shared learnable parameters. First, we use the mean vectors as slot initialization to decompose different semantics in the form of segmentation. Second, we randomly sample slots from each Gaussian distribution  $(\mu_n,\mathrm{diag}(\sigma_n))$ , and perform masked aggregation within the corresponding semantic area to identify instances. The process is detailed as follows.
+
+Semantic Decomposition. In the first stage, we directly use  $\pmb{\mu}$  without randomness as slot initialization  $S = \pmb{\mu} \in \mathbb{R}^{N \times D}$ . Then following the standard slot attention procedure [59], we employ three linear transformation heads to map  $S$  and  $R_{t}$  into  $Q \in \mathbb{R}^{N \times D}, K, V \in \mathbb{R}^{HW \times D}$ , and iteratively calculate attention score and update slot representations. Mathematically, we formulate each iteration as:
+
+$$
+M := \operatorname {s o f t m a x} \left(\frac {1}{\sqrt {D}} Q K ^ {T}, 0\right) \in \mathbb {R} ^ {N \times H W}, \tag {4}
+$$
+
+$$
+A [ n, i ] := \frac {M [ n , i ]}{\sum_ {l} M [ n , l ] + \epsilon}, A \in \mathbb {R} ^ {N \times H W}, \tag {5}
+$$
+
+$$
+S := \operatorname {G R U} (\text {i n p u t} = A V, \text {s t a t e s} = S). \tag {6}
+$$
+
+Note that the slot attention weight  $M$  is normalized along the query axis, and the weighted mean coefficient  $A$  is employed to aggregate the value to update the slots. This mechanism induces competition among slots and enforce different slots to take over different semantic features. We iterate this process for three times, and take the slot attention weight (slot vectors) of the final iteration as the segmentation mask (representation) for different semantic compo
+
+nents. To avoid ambiguity in presentation, we denote them as  $\mathcal{M} \in \mathbb{R}^{N \times HW}$  and  $S \in \mathbb{R}^{N \times D}$ , where  $S[n]$  presents  $n$ -th semantic center representation,  $\mathcal{M}[n]$  indicates the probability that each pixel belongs to  $n$ -th semantics.
+
+Instance Identification. Within each semantics, there could exist multiple object instances. The Gaussian distributions we introduce help to distinguish them with temporal correspondence cues. Specifically, in the second stage, for each semantics, we randomly sample slots from the Gaussian distribution, using the perturbations around the semantic center to capture distinct temporal correspondence patterns. For simplicity, we take the  $n$ -th semantics as an example. First, we sample  $P$  vectors,  $S \sim \mathcal{N}(\mu_n, \mathrm{diag}(\sigma_n)) \in \mathbb{R}^{P \times D}$ , as slot initialization to represent  $P$  potential instances. Next, we follow the first stage formulation to respectively project  $S$  and  $R_t$  into  $Q \in \mathbb{R}^{P \times D}$  and  $K, V \in \mathbb{R}^{HW \times D}$ , and obtain the attention score  $M \in \mathbb{R}^{P \times HW}$ . To preserve the semantic attributes and discriminate correspondence cues, we perform masked feature aggregation within the semantic mask areas. The only mathematical difference in this process is the weighted mean coefficient computation:
+
+$$
+\begin{array}{l} A [ p, i ] := \frac {M [ p , i ] \widetilde {\mathcal {M}} [ n , i ]}{\sum_ {l} M [ p , l ] \widetilde {\mathcal {M}} [ n , l ] + \epsilon}, A \in \mathbb {R} ^ {P \times H W}, \tag {7} \\ \widetilde {\mathcal {M}} [ n, i ] = \operatorname {b i n a r i z e} (\mathcal {M} [ n, i ], \tau), \mathcal {M} \in \mathbb {R} ^ {N \times H W}, \\ \end{array}
+$$
+
+where we use the  $n$ -th semantic mask  $\widetilde{\mathcal{M}}[n] \in \{0,1\}^{HW}$  binarized with threshold  $\tau$  to constrain the model to only aggregate the visual features related to  $n$ -th semantics. And we empirically find a comparatively large threshold  $\tau = 0.5$  helps to preserve object regions with clear semantics and filter out background. The masked slot attention for all semantics follows the same rule and is calculated in parallel. Similarly, we iterate the process for three times, use the final slot attention weight as the instance segmentation masks, and take the final slot vectors as the object-centric representations  $\mathcal{O} \in \mathbb{R}^{N \times P \times D}$ , where  $\mathcal{O}[n,p] \in \mathbb{R}^D$  denotes the  $p$ -th potential object instance of  $n$ -th semantics. In this way, our semantic-aware masked slot attention explicitly segments multiple instances with semantic structure.
+
+# 3.4. Training
+
+We use temporal consistency as self-supervision to optimize semantic masks  $\mathcal{M}$  and instance representations  $\mathcal{O}$ .
+
+Dense Semantic Alignment. Given the semantic mask of each timestamp  $\mathcal{M}_t\in \mathbb{R}^{N\times HW}$ , as well as  $\hat{\mathcal{M}}_t$  from teacher pathway, we aim to align the spatially dense semantic distributions across time. To achieve this, the first step is to determine the corresponding patches between different frames. Utilizing the feature correlation  $C_{tj}\in \mathbb{R}^{HW\times HW}$  which indicates dense feature similarity between timestamps  $t,j$ , we can infer patch-level correspondence from
+
+this cue with optimal transport [11, 2, 58, 18]. Formally, we take  $-\hat{C}_{tj}$  from teacher pathway as cost matrix, and solve the optimal transport strategy  $\pi_{tj}^{*} \in \mathbb{R}^{HW \times HW}$  between two uniform distributions to indicate patch correspondence:
+
+$$
+\begin{array}{l} \min  _ {\pi_ {t j}} \sum_ {u = 1} ^ {H W} \sum_ {v = 1} ^ {H W} - \hat {C} _ {t j} [ u, v ] \pi_ {t j} [ u, v ] \\ \text {s . t .} \sum_ {v = 1} ^ {H W} \pi_ {t j} [ \cdot , v ] = \frac {1}{H W} \mathbf {1} ^ {H W}, \tag {8} \\ \sum_ {u = 1} ^ {H W} \pi_ {t j} [ u, \cdot ] = \frac {1}{H W} \mathbf {1} ^ {H W}, \\ \pi_ {t j} [ u, v ] \geq 0 \quad u, v \in \{1, 2, \dots , H W \}, \\ \end{array}
+$$
+
+where  $u, v$  denotes spatial index,  $\mathbf{1}^{HW}$  is a  $HW$ -dimensional vector of all ones. Note that it is feasible to adopt other marginal distribution formulations with prior knowledge, e.g., class-agnostic activation map [4], to facilitate training, with details discussed in Supplementary Material. We employ Sinkhorn-Knopp algorithm [18] to obtain the optimal transport matrix  $\pi_{t,j}^*$ , and formulate the dense semantic alignment objective in the form of cross entropy:
+
+$$
+\mathcal {L} _ {s e m} = - \sum_ {\substack {t, j = 1 \\ t \neq j}} ^ {T} \sum_ {u, v = 1} ^ {H W} \sum_ {n = 1} ^ {N} \pi_ {t j} ^ {*} [ u, v ] \hat {\mathcal {M}} _ {j} [ n, v ] \log \mathcal {M} _ {t} [ n, u ]. \tag{9}
+$$
+
+By minimizing the weighted sum of cross entropy, we achieve temporally consistent semantic distributions among corresponding spatial areas.
+
+Semantic Mask Regularization. To encourage the semantic centers to emphasize different visual contents and cover diverse semantics, we apply a simple regularization to  $\mathcal{M}_t$ :
+
+$$
+\mathcal {L} _ {r e g} = \sum_ {t = 1} ^ {T} \sum_ {\substack {n, j = 1 \\ j \neq n}} ^ {N} \frac {\mathcal {M} _ {t} [ n ] ^ {T} \mathcal {M} _ {t} [ j ]}{\| \mathcal {M} _ {t} [ n ] \| _ {2} \| \mathcal {M} _ {t} [ j ] \| _ {2}}. \tag{10}
+$$
+
+This simple regularization suppresses the cosine similarity between different semantic masks to avoid collapse.
+
+Instance Representation Consistency. Apart from the semantic distributions, it is necessary to ensure that the representations of each object instance are temporally consistent. Given the instance representations  $\mathcal{O}_t\in \mathbb{R}^{N\times P\times HW}$ , as well as  $\hat{\mathcal{O}}_t$  from teacher pathway, we need to first match corresponding instances between different timestamps. For illustration, considering  $n$ -th semantics of time  $t,j$ , we adopt bipartite matching [9, 15] based on the cosine similarity between  $\mathcal{O}_t[n]$  and  $\hat{\mathcal{O}}_j[n]$  to generate one-to-one instance correspondence, with the matching function denoted as  $\varepsilon (\cdot)$ .
+
+In this way,  $\mathcal{O}_t[n,p]$  and  $\hat{\mathcal{O}}_j[n,\varepsilon(p)]$  are considered as a positive pair to be aligned. However, for each video, there exists absent semantics and object occlusion such that the number of visible object instances varies. Thus, there could be redundant slots not attending to objects, and it is crucial to filter out these invalid instance representations to reduce distractions. Mathematically, for timestamp  $t$ , we introduce  $\mathcal{I}_t \in \{0,1\}^{N \times P}$  as the valid instance indicator. For the  $p$ -th instance of  $n$ -th semantics to be valid, two criteria must be met: (1) The ratio of the  $n$ -th semantic area is above threshold  $\tau_1 = 0.2$  to filter out non-existing semantics; (2) The instance representation is close to the semantic center representation with cosine similarity larger than  $\tau_2 = 0.5$  to exclude redundant slots. The conditions are formulated as:
+
+$$
+\mathcal {I} _ {t} [ n, p ] = 1 \Leftrightarrow \left\{ \begin{array}{l} \frac {1}{H W} \sum_ {u = 1} ^ {H W} \widetilde {\mathcal {M}} _ {t} [ n, u ] \geq \tau_ {1}, \\ \cos \left(\mathcal {S} _ {t} [ n ], \mathcal {O} _ {t} [ n, p ]\right) \geq \tau_ {2}. \end{array} \right. \tag {11}
+$$
+
+And  $\hat{\mathcal{L}}_j\in \{0,1\}^{N\times P}$  is defined in the same manner. We use a margin loss to encourage object detail consistency over valid instance representations:
+
+$$
+\begin{array}{l} \mathcal{L}_{obj} = \sum_{\substack{t,j = 1\\ t\neq j}}^{T}\sum_{n = 1}^{N}\sum_{p = 1}^{P}\mathcal{I}_{t}[n,p]\big\{ \\ \hat {\mathcal {I}} _ {j} [ n, \varepsilon (p) ] \| \mathcal {O} _ {t} [ n, p ] - \hat {\mathcal {O}} _ {j} [ n, \varepsilon (p) ] \| _ {2} \tag {12} \\ + \sum_{\substack{q = 1\\ q\neq \varepsilon (p)}}^{P}\mathrm{relu}(\lambda -\| \mathcal{O}_{t}[n,p] - \hat{\mathcal{O}}_{j}[n,q]\|_{2})\} \\ \end{array}
+$$
+
+where  $\lambda$  is a margin hyper-parameter. Note that with the  $\mathcal{I}_t$  and  $\hat{\mathcal{I}}_j$  constraints, our formulation can handle object occlusions which lead to varying number of visible instances in different frames. By minimizing the margin loss, we encourage the model to distill discriminative and temporally consistent object-centric representations.
+
+Overall, we take the summation of three objectives for training:
+
+$$
+\mathcal {L} = \mathcal {L} _ {\text {s e m}} + \mathcal {L} _ {\text {o b j}} + \mathcal {L} _ {\text {r e g}}. \tag {13}
+$$
+
+The student pathway  $\theta$  is updated with gradient descent, and the teacher parameters  $\hat{\theta}$  are updated with momentum as:
+
+$$
+\hat {\theta} \leftarrow m \hat {\theta} + (1 - m) \theta , \tag {14}
+$$
+
+where  $m$  is momentum coefficient set to 0.999 in default. This momentum update mechanism results in a slowly evolving teacher network, which progressively distills object knowledge, provides reliable self-supervision signals and enables us to train semantic-aware masked slot attention without relying on widely adopted reconstruction objective in existing works [59, 50, 88, 84, 72, 46].
+
+# 4. Experiments
+
+# 4.1. Dataset
+
+We train our model on YouTube-VOS [87], a challenging video dataset that contains multiple object instances of distinct semantics in each video. We evaluate our method on two lines of tasks: (1) Unsupervised object segmentation on DAVIS-2016 [66], SegTrack-v2 [54], FMBS-59 [64] and challenging multiple object segmentation on DAVIS-2017-Unsupervised [8]. We respectively calculate the mean per frame intersection over union (IoU) and  $\mathcal{J}\&\mathcal{F}$  on single and multiple object discovery benchmarks. (2) Lable propagation tasks including semi-supervised video object segmentation on DAVIS-2017 [67], human pose tracking on JHMDB [45], human part tracking on VIP [90]. We adopt the same settings as [44, 43] and report standard  $\mathcal{J}$  and  $\mathcal{F}$  score on DAVIS, probability of a correct pose (PCK) on JHMDB and mean intersection over union (mIoU) on VIP.
+
+# 4.2. Implementation Details
+
+We sample  $T = 4$  frames with frame rate FPS  $= 4$  as the input RGB sequence, where each frame is augmented with random crop, horizontal flip and color jitter, and finally resized to  $256 \times 256$ . We adopt ViT-Small/16 [23] and ResNet-50 [40] as the visual encoder, specified in each experiment, to extract frame-wise features. Then for semantic-aware masked slot attention, we set the number of learnable Gaussian distributions to  $N = 16$ , the number of instances of each semantics to  $P = 4$  in default and use 3 iterations to update slot representations and attention maps.
+
+In training, the visual encoder is initialized with self-supervised pre-trained weights from DINO ViT-Small/16 [12] or MoCo ResNet-50 [38]. We use AdamW optimizer [61] with learning rate  $2 \times 10^{-4}$  for batch size 128 to update student parameters, and the teacher pathway is updated in momentum. In inference, we employ the same evaluation protocol as [44] to evaluate the temporal object correspondence performance on label propagation tasks. And for object discovery, we maintain the valid instances filtered with Eq. 11 as candidate objects. For single object evaluation, we merge all candidate objects as foreground areas to calculate IoU. For multiple object benchmark, we follow [8] to match ground-truth and our predictions and report  $\mathcal{J} \& \mathcal{F}$  score.
+
+# 4.3. Comparison with State-of-the-art
+
+Single Object Discovery. To measure our model's ability to decompose different objects, we first present the quantitative results on single object discovery without post processing (e.g., CRF, spectral clustering) in Table 2. Many existing works resort to optical flow [89, 53, 88, 84] or synthetic data [84] as weak supervision to learn temporal dynamics and generate semantic-agnostic segmentation masks for
+
+<table><tr><td>Model</td><td>RGB</td><td>Flow</td><td>DAVIS</td><td>ST-v2</td><td>FBMS</td></tr><tr><td>CIS [89]</td><td>✓</td><td>✓</td><td>71.5</td><td>62.5</td><td>63.5</td></tr><tr><td>AMD [57]</td><td>✓</td><td>X</td><td>57.8</td><td>57.0</td><td>47.5</td></tr><tr><td>DINO [12]</td><td>✓</td><td>X</td><td>52.3</td><td>46.5</td><td>50.3</td></tr><tr><td>SIMO [53]</td><td>X</td><td>✓</td><td>67.8</td><td>62.0</td><td>-</td></tr><tr><td>MG [88]</td><td>X</td><td>✓</td><td>68.3</td><td>58.6</td><td>53.1</td></tr><tr><td>OCLR [84]</td><td>X</td><td>✓</td><td>72.1</td><td>67.6</td><td>65.4</td></tr><tr><td>GWM [16]</td><td>✓</td><td>✓</td><td>71.2</td><td>69.0</td><td>66.9</td></tr><tr><td>SMTC</td><td>✓</td><td>X</td><td>71.8</td><td>69.3</td><td>68.4</td></tr><tr><td>\( SMTC^† \)</td><td>✓</td><td>X</td><td>70.8</td><td>68.4</td><td>66.5</td></tr></table>
+
+Table 2. Quantitative results on single object discovery. We compare per frame mean IoU on DAVIS-2016, SegTrack-v2 and FBMS-59 without any post-processing.  $\mathsf{SMTC}^{\dagger}$  denotes only with first slot attention stage for semantic decomposition in inference.  
+
+<table><tr><td>Model</td><td>Backbone</td><td>J&amp;F</td><td>J</td><td>F</td></tr><tr><td>DINOSAUR [72]</td><td>ViT-S/16</td><td>21.4</td><td>19.2</td><td>23.7</td></tr><tr><td>SMTC</td><td>ViT-S/16</td><td>40.5</td><td>36.4</td><td>44.6</td></tr><tr><td>SMTC</td><td>ResNet-50</td><td>39.0</td><td>35.5</td><td>42.6</td></tr><tr><td>RVOS [76]</td><td>ResNet-101</td><td>41.2</td><td>36.8</td><td>45.7</td></tr><tr><td>ProReduce [56]</td><td>ResNet-101</td><td>68.3</td><td>65.0</td><td>71.6</td></tr></table>
+
+Table 3. Quantitative results on multiple object discovery on DAVIS-2017-Unsupervised. Gray denotes supervised training.
+
+moving objects. DINO baseline [12] uses the self-attention map from the last layer as an object segmentation mask. While our method only relies on RGB frames and explicitly discriminates object semantics in a fully self-supervised manner. From the comparison, we observe that our SMTC with only semantic decomposition largely exceeds RGB-only methods [12, 57] and comparable to optical flow-based methods, and our formulation with instance identification further improves the performance. This phenomenon indicates that the learned semantic decomposition with temporal correspondence cues have the potential to substitute precomputed optical flow to guide single object segmentation. Multiple Object Discovery. Most of the existing video segmentation methods with fully unsupervised training cannot explicitly distinguish multiple object instances. OCLR [84] adopts layered flow representations to identify multiple objects on re-annotated DAVIS-2017-motion dataset, but it cannot distinguish instances with common motion. On the contrary, our method jointly utilizes semantics and temporal correspondence to separate arbitrary object instances on DAVIS-2017-Unsupervised benchmark. For fair comparison, we re-run DINOSAUR [72] on DAVIS-2017 with 5 slots as an unsupervised baseline. As presented in Table 3, our method significantly outperforms DINOSAUR. This is because our method explicitly decomposes different semantics, exploits temporal correspondence among multiple frames to distinguish instances and encourages instance
+
+level temporal consistency. While DINOSAUR is only supervised with DINO feature reconstruction without considering temporal relationships and semantic discrimination. Besides, our method is comparable with supervised baseline RVOS [76], and the gap to supervised state-of-the-art [56] is majorly due to human annotated masks as supervision leading to much more precise segmentation boundaries. This conjecture can be demonstrated by Fig. 3, which also reveals that our model is able to learn discriminative object-centric representations to separate different instances.
+
+Label Propagation. Finally, we compare the performance on label propagation tasks in Table 4 to validate the temporal consistency of our learned features. Though not specifically designed for these tasks, our model achieves state-of-the-art results on three benchmarks. The most related work to our SMTC is SFC [43], which also utilizes both high-level semantic and low-level correspondence. The difference is that SFC only employs late fusion to incorporate these two cues in inference, while our model jointly exploits semantics and temporal correspondence in training to identify object instances, perform instance-level alignment and refine temporally consistent object-centric representations.
+
+# 4.4. Ablation Study
+
+We report IoU on DAVIS-2016, and  $\mathcal{J}\&\mathcal{F}$  on DAVIS-2017-Unsupervised for single and multiple object discovery. Refer to Supplementary Material for more ablations.
+
+Slot Attention Formulation. We compare the different formulations of semantic-aware masked slot attention in Table 5. With only the first slot attention stage, i.e., the 'Semantic' part, the model cannot distinguish object instances of the same semantics, thus the performance on multiple object discovery drops significantly as indicated by Ours-B. While with only the second stage, i.e., the 'Instance' part, the performance on both single and multiple object segmentation drops as indicated by Ours-C. This reveals that the pre-computed semantic masks play an important role in alleviating interference in instance slot update and filtering valid sample for instance-level alignment.
+
+Number of Learnable Gaussian Distributions. Comparing Ours-A, Ours-E and Ours-F in Table 5, we observe that the performance improves when  $N$  increases since larger  $N$  guides our model to distill more fine-grained semantics. And we present an extreme setting with  $N = 1$  in Ours-D, where our formulation degenerates into the original slot attention with random sampling from one Gaussian distribution [59]. Note that although Ours-C also has no access to semantic masks, it still maintains the potential to discriminate different semantics due to multiple learnable means. Its superiority to Ours-D demonstrates the effectiveness of semantic discrimination in object discovery.
+
+Feature Usage. We explore using different feature input to our semantic-aware masked slot attention in Table 6. We
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Backbone</td><td colspan="3">DAVIS</td><td colspan="2">JHMDB</td><td>VIP</td></tr><tr><td>J&amp;F</td><td>J</td><td>F</td><td>PCK@0.1</td><td>PCK@0.2</td><td>mIoU</td></tr><tr><td>Supervised [40]</td><td>ResNet-50</td><td>66.0</td><td>63.7</td><td>68.4</td><td>59.2</td><td>78.3</td><td>39.5</td></tr><tr><td>MoCo [38]</td><td>ResNet-50</td><td>65.4</td><td>63.2</td><td>67.6</td><td>60.4</td><td>79.3</td><td>36.1</td></tr><tr><td>VFS [86]</td><td>ResNet-50</td><td>68.9</td><td>66.5</td><td>71.3</td><td>60.9</td><td>80.7</td><td>43.2</td></tr><tr><td>DINO [12]</td><td>ViT-S/16</td><td>61.8</td><td>60.2</td><td>63.4</td><td>45.4</td><td>75.2</td><td>37.9</td></tr><tr><td>TimeCycle [81]</td><td>ResNet-50</td><td>40.7</td><td>41.9</td><td>39.4</td><td>57.7</td><td>78.5</td><td>28.9</td></tr><tr><td>UVC [55]</td><td>ResNet-50</td><td>56.3</td><td>54.5</td><td>58.1</td><td>56.5</td><td>76.6</td><td>34.2</td></tr><tr><td>MAST [51]</td><td>ResNet-18</td><td>65.5</td><td>63.3</td><td>67.6</td><td>-</td><td>-</td><td>-</td></tr><tr><td>CRW [44]</td><td>ResNet-18</td><td>67.6</td><td>64.8</td><td>70.2</td><td>58.8</td><td>80.3</td><td>37.6</td></tr><tr><td>SFC [43]</td><td>ResNet-18+ResNet-50</td><td>71.2</td><td>68.3</td><td>74.0</td><td>61.9</td><td>83.0</td><td>38.4</td></tr><tr><td>SMTC</td><td>ViT-S/16</td><td>67.6</td><td>64.1</td><td>71.2</td><td>53.2</td><td>79.6</td><td>39.2</td></tr><tr><td>SMTC</td><td>ResNet-50</td><td>73.0</td><td>69.4</td><td>76.6</td><td>62.5</td><td>84.1</td><td>38.8</td></tr></table>
+
+Table 4. Quantitative results on label propagation tasks: semi-supervised video object segmentation on DAVIS-2017, pose tracking on JHMDB, human part tracking on VIP. We present the backbone for comprehensive comparison, and report comparing results from [86, 43].  
+
+<table><tr><td>Model</td><td>N</td><td>Semantic</td><td>Masked</td><td>IoU</td><td>J&amp;F</td></tr><tr><td>Ours-A</td><td>16</td><td>✓</td><td>✓</td><td>71.8</td><td>40.5</td></tr><tr><td>Ours-B</td><td>16</td><td>✓</td><td>X</td><td>66.4</td><td>24.5</td></tr><tr><td>Ours-C</td><td>16</td><td>X</td><td>✓</td><td>52.1</td><td>30.3</td></tr><tr><td>Ours-D</td><td>1</td><td>-</td><td>✓</td><td>47.4</td><td>22.1</td></tr><tr><td>Ours-E</td><td>8</td><td>✓</td><td>✓</td><td>68.4</td><td>37.9</td></tr><tr><td>Ours-F</td><td>32</td><td>✓</td><td>✓</td><td>71.9</td><td>40.7</td></tr></table>
+
+Table 5. Ablation studies on the number of learnable Gaussian distributions, and the formulation of semantic-aware masked slot attention. 'Semantic' ('Instance') denotes the first (second) attention stage for semantic decomposition (instance identification).
+
+<table><tr><td>Feature</td><td>Semantic</td><td>Instance</td><td>IoU</td><td>J&amp;F</td></tr><tr><td>Fused</td><td>✓</td><td>✓</td><td>71.8</td><td>40.5</td></tr><tr><td rowspan="3">RGB-only</td><td>✓</td><td>✓</td><td>67.7</td><td>20.1</td></tr><tr><td>✓</td><td>×</td><td>67.5</td><td>18.8</td></tr><tr><td>×</td><td>✓</td><td>41.5</td><td>15.3</td></tr><tr><td rowspan="3">Correlation-only</td><td>✓</td><td>✓</td><td>38.8</td><td>14.2</td></tr><tr><td>✓</td><td>×</td><td>21.1</td><td>10.7</td></tr><tr><td>×</td><td>✓</td><td>59.4</td><td>27.5</td></tr></table>
+
+Table 6. Ablation studies on the feature usage. We compare different slot attention formulations on different feature inputs.
+
+observe that with RGB-only input, the 'Semantic' part guarantees satisfactory performance on single object discovery, but performs poor on multiple object benchmarks even with the 'Instance' part. This coincides with our intuition that it is difficult to distinguish object instances with solely semantic cues, instead we need to look into more frames and resort to temporal correspondence. While for the correlation-only input, the 'Instance' part coarsely segments objects and
+
+<table><tr><td>Lsem</td><td>Lreg</td><td>Lobj</td><td>IoU</td><td>J&amp;F</td></tr><tr><td>✓</td><td>X</td><td>X</td><td>61.3</td><td>21.6</td></tr><tr><td>✓</td><td>✓</td><td>X</td><td>66.4</td><td>24.5</td></tr><tr><td>X</td><td>X</td><td>✓</td><td>50.6</td><td>30.4</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>71.7</td><td>40.5</td></tr></table>
+
+Table 7. Ablation studies on the learning objectives. We report the results on DAVIS-2016 and DAVIS-2017-Unsupervised.  
+
+<table><tr><td>T</td><td>FPS</td><td>IoU</td><td>J&amp;F</td></tr><tr><td>2</td><td>4</td><td>65.4</td><td>36.9</td></tr><tr><td>4</td><td>2</td><td>66.3</td><td>37.7</td></tr><tr><td>4</td><td>4</td><td>71.8</td><td>40.5</td></tr><tr><td>8</td><td>4</td><td>71.9</td><td>40.8</td></tr></table>
+
+Table 8. Ablation studies on the hyper-parameters of frame sampling. We compare different number of frames per clip and different FPS.
+
+the 'Semantic' part exerts negative impact. This is because the correspondence map contains little semantic information, the semantic decomposition results in severe ambiguity, thus impairing the performance.
+
+Learning Objectives. We also present a comprehensive ablation on the learning objectives in Table 7, where we report single object discovery on DAVIS-2016 and multiple object discovery results on DAVIS-2017-Unsupervised. From the comparison, we observe that  $\mathcal{L}_{sem}$  is fundamental to decompose object semantics, which demonstrates the necessity of semantic decomposition in the first stage.  $\mathcal{L}_{reg}$  also leads to improvement by encouraging the learnable queries to cover more diverse semantics. As for  $\mathcal{L}_{obj}$ , this objective introduces instance-level alignment and significantly enhances multiple object discovery performance on DAVIS-2017-Unsupervised.
+
+Frame Sampling. Another important thing in training is the frame sampling procedure. The number of frames per
+
+![](images/321920c59f87079ae3354195f90e07dba2a40c7ad71b4348a4a4c2136517e82f.jpg)  
+Figure 3. Visualization of semantic and instance segmentation map. The red boxes outline the ambiguous areas.
+
+![](images/5c852b66a77f8f4dba7488c70f5cf5273325a7f16a637a4829c96daa9cffb1c0.jpg)  
+Figure 4. Visualization of instance alignment. The arrows point out the matched instances across time.
+
+clip and sampling frame rate determines the temporal reception field of our model. We compare different sampling hyper-parameters in Table 8. The results indicate that it is necessary to use large FPS to provide rich temporal dynamics. And by setting comparatively large FPS, our model can have access to more temporal information without introducing more frames, reaching a trade-off between performance and efficiency.
+
+# 4.5. Further Discussion
+
+Visualization of Object Discovery. We visualize the our semantic decomposition map and instance identification map in Fig. 3, where the same color denotes the segmentation mask of the same semantics. The visualization reveals that our method is able to distinguish multiple object instances with semantic structure. For example, different people belong to the same semantic center, and quadrupeds such as dogs and pigs belong to another semantics. And our model also separates different instances of the same semantics, such as multiple persons and different pigs, with minor ambiguity on small object part, e.g., pig legs, mobile phones. Besides, we also visualize our instance-level alignment during training in Fig. 4. Our bipartite matching well aligns the corresponding instances in different frames, and the valid sample constraint of Eq. 11 effectively handles object occlusion in videos. For example, when encountering an occluded person instance, our method filters out the redundant slots not attending to objects, only correlating valid slots for robust instance-level alignment.
+
+Limitation. Due to the lack of the human annotated segmentation masks, our model faces challenges in generating precise boundaries for each instance, particularly for small objects. One potential solution is to incorporate multi-scale feature pyramid to further improve dense perception. We leave it in the future work. Despite this limitation, our work effectively demonstrates the benefits of leveraging both semantics and temporal correspondence to discover object instances with semantic structure and distill discriminative
+
+and temporally consistent object-centric representations.
+
+# 5. Conclusion
+
+In this work, we propose a novel self-supervised framework jointly exploiting high-level semantics and low-level temporal correspondence to enhance object-centric perception. Specifically, we represent semantic and temporal correspondence cues using the RGB feature map and dense feature correlation, respectively. These cues are fused and fed into semantic-aware masked slot attention which comprises a set of learnable Gaussian distributions. This design allows us to leverage the mean vectors as potential semantic centers for semantic decomposition, and use the perturbations introduced by the standard deviation vectors around the semantic centers to make use of the temporal correspondence cues for instance identification. To distill discriminative and temporally consistent object-centric representations, we devise semantic- and instance-level alignment that is robust to object occlusion as self-supervision. We demonstrate the effectiveness of our model for object-centric analysis through the state-of-the-art performance on label propagation tasks, as well as the promising results on unsupervised object discovery in both single and multiple object scenarios.
+
+# Acknowledgement
+
+We thank Tianfan Xue for constructive feedback. This project is funded in part by Shanghai AI Laboratory (P23KS00020, 2022ZD0160201), CUHK Interdisciplinary AI Research Institute, and the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK.
+
+# References
+
+[1] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. Advances in Neural Information Processing Systems, 33:9758-9770, 2020.  
+[2] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. arXiv preprint arXiv:1911.05371, 2019.  
+[3] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019.  
+[4] Kyungjune Baek, Minhyun Lee, and Hyunjung Shim. Psynet: Self-supervised approach to object localization using point symmetric transformation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10451-10459, 2020.  
+[5] Beril Besbinar and Pascal Frossard. Self-supervision by prediction for object discovery in videos. In 2021 IEEE Interna
+
+tional Conference on Image Processing (ICIP), pages 1509-1513. IEEE, 2021.  
+[6] Zhangxing Bian, Allan Jabri, Alexei A Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6508-6519, 2022.  
+[7] Christopher P Burgess, Loic Matthey, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019.  
+[8] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv preprint arXiv:1905.00737, 2019.  
+[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020.  
+[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018.  
+[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020.  
+[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021.  
+[13] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017.  
+[14] 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. PMLR, 2020.  
+[15] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021.  
+[16] Subhabrata Choudhury, Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Guess what moves: unsupervised video and image segmentation by anticipating motion. arXiv preprint arXiv:2205.07844, 2022.  
+[17] Eric Crawford and Joelle Pineau. Exploiting spatial invariance for scalable unsupervised object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3684-3692, 2020.  
+[18] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013.
+
+[19] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9716-9726, 2022.  
+[20] Shuangrui Ding, Rui Qian, and Hongkai Xiong. Dual contrastive learning for spatio-temporal representation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5649-5658, 2022.  
+[21] Shuangrui Ding, Weidi Xie, Yabo Chen, Rui Qian, Xiaopeng Zhang, Hongkai Xiong, and Qi Tian. Motion-inductive self-supervised object discovery in videos. arXiv preprint arXiv:2210.00221, 2022.  
+[22] Shuangrui Ding, Peisen Zhao, Xiaopeng Zhang, Rui Qian, Hongkai Xiong, and Qi Tian. Prune spatio-temporal tokens by semantic-aware temporal accumulation. arXiv preprint arXiv:2308.04549, 2023.  
+[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.  
+[24] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 2758-2766, 2015.  
+[25] Gamaleldin F Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C Mozer, and Thomas Kipf. Savi++: Towards end-to-end object-centric learning from real-world videos. arXiv preprint arXiv:2206.07764, 2022.  
+[26] Patrick Emami, Pan He, Sanjay Ranka, and Anand Rangarajan. Efficient iterative amortized inference for learning symmetric and disentangled multi-object representations. In International Conference on Machine Learning, pages 2970-2981. PMLR, 2021.  
+[27] Martin Engelcke, Adam R Kosiorek, Oiwi Parker Jones, and Ingmar Posner. Genesis: Generative scene inference and sampling with object-centric latent representations. arXiv preprint arXiv:1907.13052, 2019.  
+[28] Martin Engelcke, Oiwi Parker Jones, and Ingmar Posner. Genesis-v2: Inferring unordered object representations without iterative refinement. Advances in Neural Information Processing Systems, 34:8085-8094, 2021.  
+[29] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3299-3309, 2021.  
+[30] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018.
+
+[31] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015.  
+[32] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In International Conference on Machine Learning, pages 2424-2433. PMLR, 2019.  
+[33] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020.  
+[34] Kai Han, Rafael S Rezende, Bumsub Ham, Kwan-Yee K Wong, Minsu Cho, Cordelia Schmid, and Jean Ponce. Sc net: Learning semantic correspondence. In Proceedings of the IEEE international conference on computer vision, pages 1831-1840, 2017.  
+[35] Tengda Han, Weidi Xie, and Andrew Zisserman. Video representation learning by dense predictive coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019.  
+[36] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in Neural Information Processing Systems, 33:5679-5690, 2020.  
+[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022.  
+[38] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020.  
+[39] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017.  
+[40] 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.  
+[41] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 fps with deep regression networks. In European conference on computer vision, pages 749-765. Springer, 2016.  
+[42] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International conference on machine learning, pages 4182-4192. PMLR, 2020.  
+[43] Yingdong Hu, Renhao Wang, Kaifeng Zhang, and Yang Gao. Semantic-aware fine-grained correspondence. arXiv preprint arXiv:2207.10456, 2022.  
+[44] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances
+
+in neural information processing systems, 33:19545-19560, 2020.  
+[45] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013.  
+[46] Baoxiong Jia, Yu Liu, and Siyuan Huang. Unsupervised object-centric learning with bi-level optimized query slot attention. arXiv preprint arXiv:2210.08990, 2022.  
+[47] Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction. arXiv preprint arXiv:1811.11387, 2018.  
+[48] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthew, Antonia Creswell, Matt Botvinick, Alexander Lerchner, and Chris Burgess. Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. Advances in Neural Information Processing Systems, 34:20146-20159, 2021.  
+[49] Dahun Kim, Donghyeon Cho, Donggeun Yoo, and In So Kweon. Learning image representations by completing damaged jigsaw puzzles. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 793-802. IEEE, 2018.  
+[50] Thomas Kipf, Gamaleldin F Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. arXiv preprint arXiv:2111.12594, 2021.  
+[51] Zihang Lai, Erika Lu, and Weidi Xie. Mast: A memory-augmented self-supervised tracker. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6479-6488, 2020.  
+[52] Zihang Lai and Weidi Xie. Self-supervised learning for video correspondence flow. arXiv preprint arXiv:1905.00875, 2019.  
+[53] Hala Lamdouar, Weidi Xie, and Andrew Zisserman. Segmenting invisible moving objects. 2021.  
+[54] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In Proceedings of the IEEE international conference on computer vision, pages 2192-2199, 2013.  
+[55] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019.  
+[56] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1739-1748, 2021.  
+[57] Runtao Liu, Zhirong Wu, Stella Yu, and Stephen Lin. The emergence of objectness: Learning zero-shot segmentation from videos. Advances in Neural Information Processing Systems, 34:13137-13152, 2021.  
+[58] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection withoutImagenet. arXiv preprint arXiv:2011.13677, 2020.
+
+[59] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020.  
+[60] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015.  
+[61] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018.  
+[62] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European conference on computer vision, pages 527-544. Springer, 2016.  
+[63] Joe Yue-Hei Ng, Jonghyun Choi, Jan Neumann, and Larry S Davis. Actionflownet: Learning motion representation for action recognition. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1616–1624. IEEE, 2018.  
+[64] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. IEEE transactions on pattern analysis and machine intelligence, 36(6):1187-1200, 2013.  
+[65] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018.  
+[66] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016.  
+[67] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017.  
+[68] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Static and dynamic concepts for self-supervised video representation learning. In European Conference on Computer Vision, pages 145-164. Springer, 2022.  
+[69] Rui Qian, Yuxi Li, Huabin Liu, John See, Shuangrui Ding, Xian Liu, Dian Li, and Weiyao Lin. Enhancing self-supervised video representation learning via multi-level feature optimization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7990-8001, 2021.  
+[70] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6964-6974, 2021.  
+[71] 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.
+
+[72] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860, 2022.  
+[73] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018.  
+[74] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020.  
+[75] Jack Valmadre, Luca Bertinetto, Joao Henriques, Andrea Vedaldi, and Philip HS Torr. End-to-end representation learning for correlation filter based tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2805-2813, 2017.  
+[76] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019.  
+[77] Paul Voigtlaender and Bastian Leibe. Online adaptation of convolutional neural networks for video object segmentation. arXiv preprint arXiv:1706.09364, 2017.  
+[78] Carl Vondrick, Abhinav Shrivastava, Alireza Fathi, Sergio Guadarrama, and Kevin Murphy. Tracking emerges by colorizing videos. In Proceedings of the European conference on computer vision (ECCV), pages 391-408, 2018.  
+[79] Naiyan Wang and Dit-Yan Yeung. Learning a deep compact image representation for visual tracking. Advances in neural information processing systems, 26, 2013.  
+[80] Xiaolong Wang, Kaiming He, and Abhinav Gupta. Transitive invariance for self-supervised visual representation learning. In Proceedings of the IEEE international conference on computer vision, pages 1329-1338, 2017.  
+[81] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2566-2576, 2019.  
+[82] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021.  
+[83] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018.  
+[84] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. arXiv preprint arXiv:2207.02206, 2022.  
+[85] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level
+
+consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021.  
+[86] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10075-10085, 2021.  
+[87] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 585-601, 2018.  
+[88] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7177-7188, 2021.  
+[89] Yanchao Yang, Antonio Loquercio, Davide Scaramuzza, and Stefano Soatto. Unsupervised moving object detection via contextual information separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 879-888, 2019.  
+[90] Qixian Zhou, Xiaodan Liang, Ke Gong, and Liang Lin. Adaptive temporal encoding network for video instance-level human parsing. In Proceedings of the 26th ACM international conference on Multimedia, pages 1527-1535, 2018.  
+[91] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. In Proceedings of the IEEE international conference on computer vision, pages 408-417, 2017.

2023/Semantics Meets Temporal Correspondence_ Self-supervised Object-centric Learning in Videos/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af4d1bc599c9881bd894a8fa9b37216f3935288dde9f46a753a9e14148b794d9
+size 499855

2023/Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning/fd85c784-c77b-49cb-91b7-f3a6f94e50c3_content_list.json

@@ -0,0 +1,1566 @@
+[
+    {
+        "type": "text",
+        "text": "Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning",
+        "text_level": 1,
+        "bbox": [
+            241,
+            130,
+            728,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Kaiyou Song Shan Zhang Zimeng Luo Tong Wang Jin Xie Megvii Technology",
+        "bbox": [
+            217,
+            203,
+            746,
+            244
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{songkaiyou, zhangshan, luozimeng, wangtong, xiejin}@megvii.com",
+        "bbox": [
+            205,
+            250,
+            764,
+            263
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            297,
+            313,
+            314
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In contrastive self-supervised learning, the common way to learn discriminative representation is to pull different augmented \"views\" of the same image closer while pushing all other images further apart, which has been proven to be effective. However, it is unavoidable to construct undesirable views containing different semantic concepts during the augmentation procedure. It would damage the semantic consistency of representation to pull these augmentations closer in the feature space indiscriminately. In this study, we introduce feature-level augmentation and propose a novel semantics-consistent feature search (SCFS) method to mitigate this negative effect. The main idea of SCFS is to adaptively search semantics-consistent features to enhance the contrast between semantics-consistent regions in different augmentations. Thus, the trained model can learn to focus on meaningful object regions, improving the semantic representation ability. Extensive experiments conducted on different datasets and tasks demonstrate that SCFS effectively improves the performance of self-supervised learning and achieves state-of-the-art performance on different downstream tasks.",
+        "bbox": [
+            76,
+            332,
+            473,
+            648
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            680,
+            209,
+            696
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Due to the tremendous potential in learning discriminative feature representation without using data annotations, self-supervised learning, including contrastive learning [16, 5] and masked image modeling (MIM) [1, 15, 38] has received much attention in the representation learning field. Contrastive learning, as a type of discriminative self-supervised learning method, is heavily studied and has shown remarkable progress in the computer vision field in recent years. It aims at pulling different augmented \"views\" of the same image (positive pairs) closer while pushing diverse images (negative pairs) far from each other. To this end, a contrastive",
+        "bbox": [
+            76,
+            707,
+            470,
+            875
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/4d603971121f051b93cdb76af88d0e95709712c0cdf30fedf59da5215805e844.jpg",
+        "image_caption": [
+            "Figure 1. Semantic inconsistency of over-augmentation. (a) shows three augmentations of two images, in which the third augmentation is over-augmented and contains only background. (b) shows category probability distributions of the corresponding images in (a), which are obtained from a supervised pre-trained ResNet50 [18] model. (c)(d)(e) show three different samples of an image (the data-augmented image, the original image, and the semantics-consistent feature-augmented sample generated by Eq. (7) in this study) and their corresponding probability distributions, which point out that the over-augmented image generates different category with the original image, while the feature-augmented sample gets a balanced category probability."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            511,
+            297,
+            880,
+            532
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "loss between the features of different views extracted from an encoder network is employed to train the encoder network end-to-end. According to whether the negative pairs are used, current contrastive learning can be generally divided into two categories.",
+        "bbox": [
+            496,
+            731,
+            893,
+            808
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The first category [16, 5] utilizes both positive pairs and negative pairs for contrast. MoCo [16, 7] uses a momentum update mechanism to maintain a memory bank of negative examples. SimCLR [5, 6] directly trains a single encoder network with a large batch size to ensure sufficient positive and negative samples for learning. Based on MoCo and Sim",
+        "bbox": [
+            496,
+            809,
+            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": "1Code: https://github.com/skyoux/scfs",
+        "bbox": [
+            94,
+            886,
+            377,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "16099",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "CLR, some methods [24, 20, 36, 10, 49, 39, 19, 9, 48, 14] are proposed to improve the performance. For example, MSF [20], ISD [36] and NNCLR [10] aim to search semantics-consistent samples for contrast, solving the false negative problem. While some studies, such as Momentum2Teacher [24] and DCL [49], aim to solve the limitation that large batch size is necessary for satisfactory performance.",
+        "bbox": [
+            75,
+            90,
+            470,
+            196
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The second category of contrastive learning methods [13, 8, 3, 50, 2, 12, 4, 31, 35] only constructs positive pairs for contrast. Based on MoCo [16] and SimCLR [5], respectively, BYOL [13] and SimSiam [8] abandon the negative samples and use an asymmetric architecture to avoid model collapse. SwAV [3] uses online clustering to cluster samples and forces the consistency among cluster assignments of different augmentations. After that, some studies [4, 39, 10, 19] point out that enriching the augmented samples can improve the performance of contrastive learning. In addition, the study in [31] shows that improving the quality of positive augmented samples is important for self-supervised learning.",
+        "bbox": [
+            75,
+            199,
+            470,
+            381
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "However, it is unavoidable to construct data augmentations containing different semantic concepts. Fig. 1(a) shows three augmentations of two images, in which the third augmentation is over-augmented and contains only the background. Fig. 1(b) shows category probability distributions of the corresponding images in (a), which are obtained from a supervised pre-trained ResNet50 [18]. We observed that the probability distribution of over-augmented images changes greatly compared with the first two augmentations, which indicates that the semantic information of the over-augmented images deviates from the normally-augmented images. Similar observation can be found in Fig. 1(c)(d). The original image in (d) shows a max probability for \"ambulance\", while the over-augmented image in (c) represents the different category \"telescope\". Due to such semantic inconsistency, conducting contrastive learning on these overaugmentations is harmful to representation learning. In this study, we found that semantics-consistent feature augmentation (Fig. 1(e), generated by Eq. (7)) can balance the original semantics \"ambulance\" and the over-augmented semantics \"telescope\", which can alleviate the influence of semantic inconsistency.",
+        "bbox": [
+            75,
+            383,
+            470,
+            715
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Motivated by this observation, we propose a novel semantics-consistent feature search (SCFS) method to alleviate the negative influence of semantic inconsistency in contrastive learning. SCFS utilizes the global feature of a view to adaptively search the semantics-consistent features of another view for contrast according to their similarity. It constructs informative feature augmentations and conducts contrast learning between feature augmentations and data augmentations. Thus, the pre-trained model can learn to focus on meaningful object regions to alleviate the negative influence of unmatched semantic alignment in current contrastive learning for better representation learning. In",
+        "bbox": [
+            75,
+            719,
+            470,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "addition, the feature search is conducted on multiple layers of the backbone network, further enhancing the semantic alignment at different scales of features. Extensive experiments conducted on different datasets and tasks demonstrate that SCFS effectively improves the performance of self-supervised learning and achieves state-of-the-art performance on different downstream tasks. For example, it achieves state-of-the-art  $75.7\\%$  ImageNet top-1 accuracy under the pre-training setting of 1024 batch size and 800 epochs for ResNet50.",
+        "bbox": [
+            496,
+            90,
+            893,
+            241
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The main contributions of this study are threefold:",
+        "bbox": [
+            517,
+            242,
+            851,
+            257
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- A novel contrastive learning method, i.e., SCFS, is proposed, and it can enhance semantic alignment in contrastive learning. To our knowledge, this is the first work that defines a feature search task in contrastive learning.",
+            "- We expand contrastive learning from a data-to-data manner to a feature-to-data manner, which enriches the diversity of augmentations.",
+            "- The proposed SCFS achieves state-of-the-art performance on different downstream tasks."
+        ],
+        "bbox": [
+            517,
+            267,
+            890,
+            436
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Works",
+        "text_level": 1,
+        "bbox": [
+            500,
+            452,
+            648,
+            468
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recently, some studies [33, 44, 42, 45, 46, 25, 47, 40, 23, 41] pointed out that the problem of semantic inconsistency is more serious for downstream dense prediction tasks, such as object detection and instance segmentation. Therefore, these methods utilize region-level and pixel-level features for contrast. In this study, the proposed SCFS construct feature-level augmentations using dense feature maps. Therefore, this section introduces related studies that conduct contrastive learning using region-level and pixel-level features.",
+        "bbox": [
+            496,
+            477,
+            893,
+            613
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Region-level contrastive learning. SCRL [33] minimizes the distance between two local features, which are cropped from two corresponding feature maps of two views. ReSim [44] aligns regional representations by sliding a fixed-sized window across the overlapping area between two views to improve the performance for localization-based tasks. SoCo [42], ORL [46], and UniVIP [25] extract object region proposals and use them to construct region-level features for contrastive learning. They achieve good performance for downstream dense prediction tasks.",
+        "bbox": [
+            496,
+            614,
+            893,
+            763
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Pixel-level contrastive learning. To obtain a more fine-grained representation, several studies [47, 40, 41] design pixel-level contrastive learning task, which assumes that features extracted from the same pixel of different views should be treated as positive pairs while pixels from others must be distinguished. PixPro [47] utilizes a pixel propagation module to select similar pixel features for contrast and encourages consistency between positive pixel pairs. DenseCL [40] proposes a dense projection head to generate dense",
+        "bbox": [
+            496,
+            765,
+            893,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "16100",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "feature vectors for pixel-level contrastive learning. SetSim [41] is designed to realize pixel-wise similarity learning by filtering out noisy backgrounds.",
+        "bbox": [
+            75,
+            90,
+            468,
+            136
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As summarized above, data augmentations bring rich information while increasing uncertainty in contrastive learning. While methods that utilize region-level features expand the granularity of feature representation by alleviating the influence of noises. Recently, to solve the problem of semantic inconsistency in video contrastive learning, VITO [30] uses an MLP to learn attention masks for each view independently and fuses the attention pooling features from multi-layers for contrastive learning. Unlike VITO [30], SCFS uses a cross-view search to calculate attention and conducts contrastive learning at multi-layers independently. In addition, our method bridges the correlation between data and feature augmentations and extends the contrastive learning task to a semantics-consistent feature search task.",
+        "bbox": [
+            75,
+            137,
+            470,
+            347
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Methods",
+        "text_level": 1,
+        "bbox": [
+            76,
+            362,
+            174,
+            378
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The overall architecture of SCFS is shown in Fig. 2. It consists of an encoder and a momentum encoder. The momentum encoder is an exponential-moving-average version of the encoder. SCFS consists of two contrastive learning tasks: the contrast between data augmentations  $(\\mathcal{L}_d)$ , which is designed based on the baseline DINO [4], and the contrast between data augmentations and feature augmentations  $(\\mathcal{L}_{fs})$ . In this section, we first introduce the contrast between data augmentations in Sec. 3.1. Then, the key feature search module of SCFS is introduced in detail in Sec. 3.2. Next, we will introduce the overall loss of SCFS in Sec. 3.3. Finally, the implementation details are presented in Sec. 3.4.",
+        "bbox": [
+            75,
+            387,
+            470,
+            569
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Contrast Between Data Augmentations",
+        "text_level": 1,
+        "bbox": [
+            76,
+            579,
+            413,
+            595
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given a pair of global augmentations  $(I_1$  and  $I_{2}$  ) of an input image, the feature representations of the two augmentations are used to calculate the global contrastive loss. Specifically,  $\\pmb {f}_1 = E_\\pmb {\\theta}(\\pmb {I}_1)$  and  $\\pmb{f^{\\prime}}_2 = E_{\\pmb{\\theta}^{\\prime}}(\\pmb {I}_2)$  , where  $\\pmb{\\theta}$  and  $\\pmb{\\theta}^{\\prime}$  are parameters of the encoder and the momentum encoder, respectively.  $\\pmb {f}_1,\\pmb{f}^{\\prime}_2\\in R^K$ $K$  is the output dimension.  $\\pmb{f}_{1}$  is normalized with a softmax function:",
+        "bbox": [
+            75,
+            603,
+            468,
+            708
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP _ {1} ^ {i} = \\frac {\\exp \\left(f _ {1} ^ {i} / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(f _ {1} ^ {k} / \\tau\\right)} \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            186,
+            717,
+            468,
+            756
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\tau > 0$  is a temperature parameter that controls the sharpness of the output distribution. Note that  $P_{2}^{\\prime}$  is obtained by normalizing  $f_{2}^{\\prime}$  with a similar softmax function with temperature  $\\tau^{\\prime}$ .  $I_{1}$  and  $I_{2}$  are fed to the momentum encoder and encoder symmetrically, and  $P_{1}^{\\prime}$  and  $P_{2}$  are obtained respectively. Following DINO [4], the cross-entropy loss is employed as the contrastive loss between two global views:",
+        "bbox": [
+            75,
+            767,
+            468,
+            872
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {g} = - \\left(P _ {2} ^ {\\prime} \\log \\left(P _ {1}\\right) + P _ {1} ^ {\\prime} \\log \\left(P _ {2}\\right)\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            150,
+            883,
+            468,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To enrich augmentations, the multi-crop strategy [3] is employed. Multiple local augmentations  $\\mathbf{I}_l$  is also constructed and fed to the encoder:  $\\mathbf{f}_l = E_{\\boldsymbol{\\theta}}(\\mathbf{I}_l)$ .  $P_l$  is obtained by normalizing  $\\mathbf{f}_l$  with the softmax function with temperature  $\\tau$ . The contrast between local views and global views can be calculated:",
+        "bbox": [
+            496,
+            90,
+            893,
+            181
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {l} = \\sum_ {n = 1} ^ {N} - \\left(P _ {1} ^ {\\prime} \\log \\left(P _ {l} ^ {n}\\right) + P _ {2} ^ {\\prime} \\log \\left(P _ {l} ^ {n}\\right)\\right) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            550,
+            189,
+            892,
+            231
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $N$  denotes the number of local views. Thus, the overall loss is the sum of global loss and local loss:",
+        "bbox": [
+            496,
+            239,
+            890,
+            268
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {d} = \\mathcal {L} _ {g} + \\mathcal {L} _ {l} \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            643,
+            281,
+            892,
+            297
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Semantics-Consistent Feature Search",
+        "text_level": 1,
+        "bbox": [
+            500,
+            306,
+            821,
+            321
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As introduced in Sec. 1, it is unavoidable to construct augmentations that contain different semantic concepts during the augmentation procedure. It's harmful to pull these augmentations close indiscriminately in the feature space. Therefore, we propose the SCFS method to enhance the contrast between semantics-consistent regions in different augmentations.",
+        "bbox": [
+            496,
+            329,
+            893,
+            434
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The architecture of SCFS is shown in Fig. 2. By feeding the local augmentations  $I_{l}$  to the encoder, feature maps from different stages of the backbone ResNet50 [18] are extracted. Specifically, the output features from different stages, i.e.,  $Res2$ ,  $Res3$  and  $Res4$ , are utilized to conduct SCFS, ensuring that each stage of the backbone produces discriminative features:  $\\{^{2}F_{l},^{3}F_{l},^{4}F_{l}\\} = E_{\\theta}(I_{l})$ , where  $^i F_l\\in R^{W_l^i\\times H_l^i\\times C^i}$ ,  $W_{l}^{i},H_{l}^{i},C^{i}$  denote the width, height and channel dimension, respectively. Next, the global average pooling operation is conducted on each  $^i F_{l}$  in the spatial dimensions:",
+        "bbox": [
+            496,
+            435,
+            893,
+            603
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n{ } ^ { i } \\boldsymbol { f } _ { l } = \\frac { 1 } { W _ { l } ^ { i } \\times H _ { l } ^ { i } } \\sum _ { x = 1 } ^ { W _ { l } ^ { i } } \\sum _ { y = 1 } ^ { H _ { l } ^ { i } } { } ^ { i } \\boldsymbol { F } _ { l } ( x , y , z ) \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            568,
+            613,
+            892,
+            657
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  ${}^i f_l \\in R^{C^i}$ . Meanwhile, the global augmentations  $I_g$  ( $g = 1,2$ ) are fed to the momentum encoder to extract feature maps from different stages:  $\\{^2 F_g',^3 F_g',^4 F_g'\\} = E_{\\pmb{\\theta}'}(\\pmb {I}_g),{}^i F_g' \\in R^{W_g^i\\times H_g^i\\times C^i}, W_g^i,H_g^i,C^i$  denote width, height and channel dimension, respectively.",
+        "bbox": [
+            496,
+            670,
+            890,
+            750
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Then, based on  ${}^i\\pmb {f}_l$  and  ${}^i\\pmb{F}_{g}^{\\prime}$ , SCFS aims to adaptively search the most semantics-consistent features in  ${}^i\\pmb{F}_{g}^{\\prime}$  for contrast, while suppressing irrelevant features. In SCFS, each feature  ${}^i\\pmb {f}_l$  of the local data augmentations is treated as query, and the features  ${}^i\\pmb{F}_{g}^{\\prime}$  of the global augmentations are treated as keys. The similarity between  ${}^i\\pmb {f}_l$  and  ${}^i\\pmb{F}_{g}^{\\prime}$  is calculated:",
+        "bbox": [
+            496,
+            750,
+            893,
+            857
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol {A} (x, y) = \\frac {{} ^ {i} \\boldsymbol {f} _ {l} \\cdot {} ^ {i} \\boldsymbol {F} _ {g} ^ {\\prime} (x , y)}{\\| ^ {i} \\boldsymbol {f} _ {l} \\| _ {2} \\| ^ {i} \\boldsymbol {F} _ {g} ^ {\\prime} (x , y) \\| _ {2}} \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            583,
+            866,
+            892,
+            905
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "16101",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/2f47ebc92a19132eefcdbaca58e53be431c38d99e09fc3c383ded5eae72b6d7d.jpg",
+        "image_caption": [
+            "Figure 2. Overall architecture of the proposed semantics-consistent feature search (SCFS). It consists of an encoder and a momentum encoder. There are two contrastive learning tasks: the contrast between data augmentations  $(\\mathcal{L}_d)$  in the final feature space and the feature search task conducted on multiple layers  $(\\mathcal{L}_{fs})$ . The details of the feature search procedure is shown on the right."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            158,
+            88,
+            558,
+            268
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/e11192e29bee3c44aa0413a27f29c9325f5cabf6877fd1fcae5ca6defae4acc1.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            563,
+            89,
+            810,
+            268
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{A} \\in R^{W_g^i \\times H_g^i}$  is the attention map, and  $x = 1, \\ldots, W_g^i$ ,  $y = 1, \\ldots, H_g^i$ ,  $\\| \\cdot \\|_2$  is the L2 norm. The attention map  $\\mathbf{A}$  activates the semantics-consistent regions of the local augmentation on the global augmentation. Thus, the higher portion of local regions can be searched. To select semantic features and suppress irrelevant local features, we directly multiply the attention map  $\\mathbf{A}$  with  ${}^i\\mathbf{F}_g'$  to obtain the semantics-consistent feature augmentations:",
+        "bbox": [
+            75,
+            337,
+            473,
+            460
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n{ } ^ { i } \\boldsymbol { F } _ { l g } ^ { \\prime } = \\boldsymbol { A } \\cdot { } ^ { i } \\boldsymbol { F } _ { g } ^ { \\prime }  ( 7 )\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            217,
+            465,
+            468,
+            487
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "This operation can be regarded as attention-weighted average pooling. Through feature search,  $N$  local data augmentations  $\\pmb{I}_l$  can search  $N$  corresponding semantics-consistent features  ${}^i\\pmb{F}_{lg}'$  from a global data augmentation  $\\pmb{I}_g$ . That is, in terms of the global data augmentation,  $N$  different features are constructed in the feature space through the feature search procedure. Therefore, we term  ${}^i\\pmb{F}_{lg}'$  as feature-level augmentations.",
+        "bbox": [
+            75,
+            491,
+            468,
+            611
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Next,  ${}^i\\mathbf{F}_l$  and  ${}^i\\mathbf{F}_{lg}'$  are fed to corresponding projection heads to obtain their final representations for contrast:",
+        "bbox": [
+            75,
+            611,
+            468,
+            643
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left\\{ \\begin{array}{l} ^ {i} \\boldsymbol {f} _ {l} = \\mathrm {H} _ {i} \\left(^ {i} \\boldsymbol {F} _ {l}\\right) \\\\ ^ {i} \\boldsymbol {f} _ {l g} ^ {\\prime} = H _ {i} ^ {\\prime} \\left(^ {i} \\boldsymbol {F} _ {l g} ^ {\\prime}\\right) \\end{array} \\right. \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            192,
+            648,
+            468,
+            683
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathrm{H}_i$  and  $\\mathrm{H}'_i$  denote the projection heads on the  $i$ -th layer of the encoder and the momentum encoder, respectively.  ${}^i\\pmb {f}_l$  and  ${}^i\\pmb {f}_{lg}'$  are normalized with softmax function with temperature  $\\tau$  and  $\\tau'$ , respectively, as the same formulation in Eq. (1). The corresponding output probability  ${}^iP_l$  and  ${}^iP_{lg}'$  are employed to calculate the contrast loss between local data augmentations and feature augmentations:",
+        "bbox": [
+            75,
+            688,
+            472,
+            794
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n{ } ^ { i } \\mathcal { L } _ { f s } = \\sum _ { g = 1 } ^ { 2 } \\sum _ { n = 1 } ^ { N } - \\left( ^ { i } P _ { l g } ^ { \\prime } \\log \\left( ^ { i } P _ { l } ^ { n } \\right) \\right) \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            148,
+            800,
+            468,
+            843
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The overall feature search loss is the sum of all layers:",
+        "bbox": [
+            96,
+            849,
+            457,
+            864
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {f s} = \\sum_ {i \\in V _ {L}} ^ {i} \\mathcal {L} _ {f s} \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            214,
+            869,
+            468,
+            904
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $V_{L}$  denotes the set of layers to conduct SCFS.",
+        "bbox": [
+            498,
+            339,
+            846,
+            354
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Overall Loss",
+        "text_level": 1,
+        "bbox": [
+            500,
+            373,
+            633,
+            388
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The overall loss is the sum of the contrastive loss between data augmentations and the feature search loss:",
+        "bbox": [
+            498,
+            398,
+            890,
+            430
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {d} + \\mathcal {L} _ {f s} \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            643,
+            449,
+            890,
+            465
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Through SCFS, the contrast between feature augmentations and data augmentations is bridged. The model can adaptively search the semantics-consistent features for contrast. Therefore, it can enhance the importance of semantics-consistent regions in different augmentations, alleviating the uncertainty in contrastive learning introduced by data augmentations that contain different semantic concepts.",
+        "bbox": [
+            498,
+            486,
+            893,
+            592
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4. Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            500,
+            611,
+            718,
+            628
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "SCFS is based on DINO [4] and we follow the most hyper-parameter settings of DINO. For a fair comparison, the standard ResNet50 [18] is employed as the default backbone network.",
+        "bbox": [
+            498,
+            637,
+            890,
+            696
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For data augmentation, the global augmentations consist of random cropping, resizing to  $224 \\times 224$ , random horizontal flip, gaussian blur, and color jittering. And the local augmentations consist of random cropping, resizing to  $96 \\times 96$ , random horizontal flip, gaussian blur, and color jittering. For feature augmentations in SCFS, the Res2, Res3, and Res4 layers are used. Two global views with  $N = 8$  local views are the default setting of augmentation.",
+        "bbox": [
+            496,
+            700,
+            893,
+            821
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The projection head for the contrast between data augmentations consists of a four-layer multi-layer-perceptron (MLP) with the same architecture as DINO [4]. The projection head for feature search consists of three convolutional layers and two FC layers.",
+        "bbox": [
+            496,
+            825,
+            893,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "16102",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/a261d5d96710eba8221daecb261b08400865a1aabfe57128aaf815ebbacea0bf.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 1. Linear probing and  $k$ -NN accuracy (\\%) on ImageNet. The result with “*” is reproduced for fair comparison. LP denotes linear probing. Bold font and underline indicate the best results under the setting of 256 batch size and 200 epochs and the setting of 1024 batch size and 800 epochs, respectively."
+        ],
+        "table_body": "<table><tr><td>Method</td><td>Batch Size</td><td>Epochs</td><td>LP</td><td>k-NN</td></tr><tr><td>Supervised</td><td>256</td><td>100</td><td>76.2</td><td>74.8</td></tr><tr><td>SimCLR [5]</td><td>4096</td><td>1000</td><td>69.3</td><td>-</td></tr><tr><td>BYOL [13]</td><td>4096</td><td>1000</td><td>74.3</td><td>66.9</td></tr><tr><td>BYOL [13]</td><td>4096</td><td>200</td><td>70.6</td><td>-</td></tr><tr><td>SwAV [3]</td><td>4096</td><td>800</td><td>75.3</td><td>-</td></tr><tr><td>SwAV [3]</td><td>256</td><td>200</td><td>72.7</td><td>-</td></tr><tr><td>MoCo-v2 [16]</td><td>256</td><td>200</td><td>67.5</td><td>54.3</td></tr><tr><td>SimSiam [8]</td><td>256</td><td>200</td><td>70.0</td><td>-</td></tr><tr><td>ISD [36]</td><td>256</td><td>200</td><td>69.8</td><td>62.0</td></tr><tr><td>MSF [20]</td><td>256</td><td>200</td><td>71.4</td><td>64.0</td></tr><tr><td>NNCLR [10]</td><td>4096</td><td>200</td><td>70.7</td><td>-</td></tr><tr><td>Barlow Twins [50]</td><td>2048</td><td>1000</td><td>73.2</td><td>-</td></tr><tr><td>VICReg [2]</td><td>2048</td><td>1000</td><td>73.2</td><td>-</td></tr><tr><td>OBoW [12]</td><td>256</td><td>200</td><td>73.8</td><td>-</td></tr><tr><td>DCL [49]</td><td>256</td><td>200</td><td>66.9</td><td>-</td></tr><tr><td>CLSA [39]</td><td>256</td><td>200</td><td>73.3</td><td>-</td></tr><tr><td>AdCo [19]</td><td>256</td><td>200</td><td>73.2</td><td>-</td></tr><tr><td>DetCo [45]</td><td>256</td><td>200</td><td>68.6</td><td>-</td></tr><tr><td>UniVIP [25]</td><td>4096</td><td>200</td><td>73.1</td><td>-</td></tr><tr><td>HCSC [14]</td><td>256</td><td>200</td><td>73.3</td><td>-</td></tr><tr><td>MoCo-v3 [9]</td><td>4096</td><td>300</td><td>72.8</td><td>-</td></tr><tr><td>MoCo-v3 [9]</td><td>4096</td><td>1000</td><td>74.6</td><td>-</td></tr><tr><td>DINO* [4]</td><td>256</td><td>200</td><td>73.0</td><td>64.0</td></tr><tr><td>DINO [4]</td><td>4080</td><td>800</td><td>75.3</td><td>67.5</td></tr><tr><td>SCFS</td><td>256</td><td>200</td><td>73.9</td><td>65.5</td></tr><tr><td>SCFS</td><td>1024</td><td>800</td><td>75.7</td><td>68.5</td></tr></table>",
+        "bbox": [
+            84,
+            88,
+            460,
+            518
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            625,
+            209,
+            642
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, comprehensive experiments are conducted to demonstrate the effectiveness of SCFS. We evaluate the performance on different downstream tasks, including ImageNet classification, object detection, instance segmentation, and other classification task on small datasets. In addition, we conduct ablation experiments to analyze the influence of each component in SCFS.",
+        "bbox": [
+            75,
+            651,
+            470,
+            756
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Comparing with SSL methods on ImageNet",
+        "text_level": 1,
+        "bbox": [
+            76,
+            768,
+            447,
+            787
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "$k$ -NN and Linear Probing Accuracy on ImageNet. After pre-training on the ImageNet ILSVRC-2012 [34] training set, the pre-trained models are evaluated on the ImageNet ILSVRC-2012 validation set. For  $k$ -NN, it is evaluated as in study [43]. For linear probing, we train a linear classifier from scratch based on the feature extracted by a fixed backbone with 100 epochs [16]. The top-1 accuracy is adopted",
+        "bbox": [
+            75,
+            794,
+            470,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/557e7ef00384ae78338e786dfa9c39cb02142d5ddddc4717c6cc15d21f3d672e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">Batch Size Epochs</td><td colspan=\"2\">Top-1</td><td colspan=\"2\">Top-5</td></tr><tr><td></td><td></td><td>1%</td><td>10%</td><td>1%</td><td>10%</td></tr><tr><td>Supervised [51]</td><td>256</td><td>90</td><td>25.4</td><td>56.4</td><td>48.4</td><td>80.4</td></tr><tr><td>SimCLR [5]</td><td>4096</td><td>1000</td><td>48.3</td><td>65.6</td><td>75.5</td><td>87.8</td></tr><tr><td>BYOL [13]</td><td>4096</td><td>1000</td><td>53.2</td><td>68.8</td><td>78.4</td><td>89.0</td></tr><tr><td>SwAV [3]</td><td>4096</td><td>800</td><td>53.9</td><td>70.2</td><td>78.5</td><td>89.9</td></tr><tr><td>DINO [4]</td><td>4080</td><td>800</td><td>50.2</td><td>69.3</td><td>74.0</td><td>89.1</td></tr><tr><td>SCFS</td><td>1024</td><td>800</td><td>54.3</td><td>70.5</td><td>78.6</td><td>90.2</td></tr></table>",
+        "bbox": [
+            509,
+            88,
+            883,
+            232
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/2588c583ac30b4edffc275433e0569e3afdec46d580cedca47138db6e2a5ed9d.jpg",
+        "table_caption": [
+            "Table 2. Evaluation on small labeled ImageNet. Bold font indicates the best result."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Epochs</td><td>APb</td><td>APb50</td><td>APb75</td></tr><tr><td>Scratch</td><td>-</td><td>33.8</td><td>60.2</td><td>33.1</td></tr><tr><td>Supervised</td><td>90</td><td>53.5</td><td>81.3</td><td>58.8</td></tr><tr><td>SimCLR [5]</td><td>1000</td><td>56.3</td><td>81.9</td><td>62.5</td></tr><tr><td>BYOL [13]</td><td>300</td><td>51.9</td><td>81.0</td><td>56.5</td></tr><tr><td>SwAV [3]</td><td>400</td><td>45.1</td><td>77.4</td><td>46.5</td></tr><tr><td>DINO [4]</td><td>800</td><td>55.9</td><td>82.1</td><td>62.3</td></tr><tr><td>SCFS</td><td>800</td><td>57.4</td><td>83.0</td><td>63.6</td></tr></table>",
+        "bbox": [
+            511,
+            279,
+            880,
+            412
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Results for PASCAL VOC object detection using Faster R-CNN [32] with ResNet50-C4. Bold font indicates the best result.",
+        "bbox": [
+            500,
+            422,
+            893,
+            450
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "as the evaluation metric.",
+        "bbox": [
+            500,
+            474,
+            663,
+            488
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The results are reported in Tab. 1. With the standard ResNet50 [18] architecture and pre-trained with 256 batch size for 200 epoch, the proposed SCFS achieves the best  $k$ -NN top-1 accuracy  $65.5\\%$  and the best linear probing top-1 accuracy  $73.9\\%$ , outperforming its baseline DINO [4] by  $1.5\\%$  and  $0.9\\%$ , respectively. In addition, with 1024 batch size and 800 epoch, SCFS achieves the best  $k$ -NN accuracy  $(68.5\\%)$  and linear probing accuracy  $(75.7\\%)$ , outperforming the accuracy of DINO [4] trained with 4080 batch size for 800 epoch. This result demonstrates that SCFS can improve the representation learning performance by searching semantics-consistent features for contrast.",
+        "bbox": [
+            496,
+            489,
+            893,
+            671
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Semi-Supervised Learning on ImageNet. In this part, we evaluate the performance of SCFS under the semi-supervised setting. Specifically, we use  $1\\%$  and  $10\\%$  of the labeled training data from ImageNet [34] for finetuning, which follows the semi-supervised protocol in SimCLR [5]. The same splits of  $1\\%$  and  $10\\%$  of ImageNet labeled training data in SimCLRv2 [6] are used. The results are reported in Tab. 2. After finetuning using  $1\\%$  and  $10\\%$  training data, SCFS outperforms all the compared methods. The results demonstrate that SCFS achieves the best feature representation quality.",
+        "bbox": [
+            496,
+            671,
+            895,
+            823
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Transfer Learning on Downstream Tasks",
+        "text_level": 1,
+        "bbox": [
+            500,
+            832,
+            851,
+            848
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Object Detection and Instance Segmentation. In this part, we evaluate the representations of SCFS on dense prediction tasks, i.e., object detection and instance segmentation,",
+        "bbox": [
+            496,
+            854,
+            893,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "16103",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/9805798297ae8665ee55cf52316dd8d4159f48c59879cfe4b813a0bac8a47fab.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">Epochs</td><td colspan=\"6\">1×schedule</td><td colspan=\"6\">2×schedule</td></tr><tr><td>\\(AP^b\\)</td><td>\\(AP^b_{50}\\)</td><td>\\(AP^b_{75}\\)</td><td>\\(AP^s\\)</td><td>\\(AP^s_{50}\\)</td><td>\\(AP^s_{75}\\)</td><td>\\(AP^b\\)</td><td>\\(AP^b_{50}\\)</td><td>\\(AP^b_{75}\\)</td><td>\\(AP^s\\)</td><td>\\(AP^s_{50}\\)</td><td>\\(AP^s_{75}\\)</td></tr><tr><td>Scratch</td><td>-</td><td>31.0</td><td>49.5</td><td>33.2</td><td>28.5</td><td>46.8</td><td>30.4</td><td>38.4</td><td>57.5</td><td>42.0</td><td>34.7</td><td>54.8</td><td>37.2</td></tr><tr><td>Supervised</td><td>90</td><td>38.9</td><td>59.6</td><td>42.7</td><td>35.4</td><td>56.5</td><td>38.1</td><td>41.3</td><td>61.3</td><td>45.0</td><td>37.3</td><td>58.3</td><td>40.3</td></tr><tr><td>MoCo [16]</td><td>200</td><td>38.5</td><td>58.9</td><td>42.0</td><td>35.1</td><td>55.9</td><td>37.7</td><td>40.8</td><td>61.6</td><td>44.7</td><td>36.9</td><td>58.4</td><td>39.7</td></tr><tr><td>MoCo v2 [7]</td><td>200</td><td>40.4</td><td>60.2</td><td>44.2</td><td>36.4</td><td>57.2</td><td>38.9</td><td>41.7</td><td>61.6</td><td>45.6</td><td>37.6</td><td>58.7</td><td>40.5</td></tr><tr><td>BYOL [13]</td><td>300</td><td>40.4</td><td>61.6</td><td>44.1</td><td>37.2</td><td>58.8</td><td>39.8</td><td>42.3</td><td>62.6</td><td>46.2</td><td>38.3</td><td>59.6</td><td>41.1</td></tr><tr><td>SwAV [3]</td><td>400</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>42.3</td><td>62.8</td><td>46.3</td><td>38.2</td><td>60.0</td><td>41.0</td></tr><tr><td>\\(ReSim-FPN^T\\) [44]</td><td>200</td><td>39.8</td><td>60.2</td><td>43.5</td><td>36.0</td><td>57.1</td><td>38.6</td><td>41.4</td><td>61.9</td><td>45.4</td><td>37.5</td><td>59.1</td><td>40.3</td></tr><tr><td>SetSim [41]</td><td>200</td><td>40.2</td><td>60.7</td><td>43.9</td><td>36.4</td><td>57.7</td><td>39.0</td><td>41.6</td><td>62.4</td><td>45.9</td><td>37.7</td><td>59.4</td><td>40.6</td></tr><tr><td>DenseCL [40]</td><td>200</td><td>40.3</td><td>59.9</td><td>44.3</td><td>36.4</td><td>57.0</td><td>39.2</td><td>41.2</td><td>61.9</td><td>45.1</td><td>37.3</td><td>58.9</td><td>40.1</td></tr><tr><td>DSC [23]</td><td>200</td><td>39.4</td><td>58.9</td><td>43.2</td><td>35.7</td><td>56.1</td><td>38.3</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>HSA [48]</td><td>800</td><td>40.2</td><td>60.9</td><td>43.9</td><td>36.5</td><td>57.9</td><td>39.1</td><td>42.2</td><td>63.0</td><td>46.1</td><td>38.1</td><td>59.9</td><td>40.9</td></tr><tr><td>DetCo [45]</td><td>800</td><td>40.1</td><td>61.0</td><td>43.9</td><td>36.4</td><td>58.0</td><td>38.9</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>ORL* [46]</td><td>800</td><td>40.3</td><td>60.2</td><td>44.4</td><td>36.3</td><td>57.3</td><td>38.9</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DINO [4]</td><td>800</td><td>40.0</td><td>61.6</td><td>43.4</td><td>36.5</td><td>58.6</td><td>39.1</td><td>41.9</td><td>62.6</td><td>46.0</td><td>37.8</td><td>59.7</td><td>40.6</td></tr><tr><td>SCFS</td><td>800</td><td>40.5</td><td>61.8</td><td>44.0</td><td>36.7</td><td>58.8</td><td>39.2</td><td>42.1</td><td>63.4</td><td>46.1</td><td>38.1</td><td>60.2</td><td>41.0</td></tr></table>",
+        "bbox": [
+            127,
+            88,
+            841,
+            363
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/3ecca5e5e90135872ecaeddb8722b14257db43a983d5059353969ddff9008647.jpg",
+        "table_caption": [
+            "Table 4. Object detection and instance segmentation on COCO using Mask R-CNN [17] with ResNet50-FPN. Bold font indicates the best result. * indicates pre-training on COCO [26]."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>CIFAR-10</td><td>CIFAR-100</td><td>CUB-Bird</td><td>Stanford-Cars</td><td>Aircraft</td><td>Oxford-Pets</td></tr><tr><td>Supervised</td><td>97.5</td><td>86.4</td><td>81.3</td><td>92.1</td><td>86.0</td><td>92.1</td></tr><tr><td>SimCLR[5]</td><td>97.7</td><td>85.9</td><td>-</td><td>91.3</td><td>88.1</td><td>89.2</td></tr><tr><td>BYOL[13]</td><td>97.8</td><td>86.1</td><td>-</td><td>91.6</td><td>88.1</td><td>91.7</td></tr><tr><td>DINO[4]*</td><td>97.7</td><td>86.6</td><td>81.0</td><td>91.1</td><td>87.4</td><td>91.5</td></tr><tr><td>SCFS</td><td>97.8</td><td>86.7</td><td>82.7</td><td>91.6</td><td>88.5</td><td>91.9</td></tr></table>",
+        "bbox": [
+            187,
+            411,
+            784,
+            513
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 5. Transfer learning results from ImageNet with the standard ResNet50 [18]. * denotes the results are reproduced in this study. Bold font indicates the best result.",
+        "bbox": [
+            75,
+            523,
+            893,
+            551
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "on mainstream datasets PASCAL VOC [11] and MS COCO [26] datasets. On the PASCAL VOC dataset [11], the trainval07+12 set is used as the training set, and the test2007 set is used as the test set. Following [42], Faster R-CNN detector [32] with the ResNet50-C4 backbone initialized by the self-supervised pre-trained model is trained end-to-end. On the COCO dataset, the train2017 set is used for training and the val2017 set is used for evaluation. The Mask R-CNN [17] with R50-FPN is used. The  $\\mathrm{AP^{b}}$ ,  $\\mathrm{AP_{50}^{b}}$  and  $\\mathrm{AP_{75}^{b}}$  metrics are used for object detection. While the  $\\mathrm{AP^{s}}$ ,  $\\mathrm{AP_{50}^{s}}$  and  $\\mathrm{AP_{75}^{s}}$  metrics are used for instance segmentation.",
+        "bbox": [
+            75,
+            573,
+            472,
+            742
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The experimental results are shown in Tab. 3 and Tab. 4. SCFS achieves best performance on the two datasets. For example, on VOC, SCFS achieves  $57.4\\% \\mathrm{AP}^{\\mathrm{b}}$ ,  $83.0\\% \\mathrm{AP}_{50}^{\\mathrm{b}}$  and  $63.6\\% \\mathrm{AP}_{75}^{\\mathrm{b}}$ . The  $\\mathrm{AP}^{\\mathrm{b}}$  of SCFS outperforms its baseline DINO by  $1.5\\%$ . These results show that SCFS also has good transfer ability on dense prediction tasks.",
+        "bbox": [
+            75,
+            744,
+            470,
+            835
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Other Classification Tasks. In this part, we focus on the performance of self-supervised models when they are finetuned on small datasets, including CIFAR [22] and fine grained datasets [21, 37, 28, 29]. The results are shown in Tab. 5.",
+        "bbox": [
+            75,
+            839,
+            472,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The proposed SCFS shows the best performance on all the small datasets, which demonstrates that SCFS has good generalization ability.",
+        "bbox": [
+            496,
+            574,
+            893,
+            621
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Pre-training on Uncurated Dataset",
+        "text_level": 1,
+        "bbox": [
+            498,
+            633,
+            803,
+            650
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The proposed SCFS can solve the problem of semantic inconsistency during pre-training, which is important when pre-training on uncurated datasets since this problem is more serious. To verify this, we pre-train SCFS and DINO on COCO [26], which is much more Uncurated than ImageNet. The same hyper-parameters used on ImageNet are applied to train the models with 512 batch size for 500 epochs. After pre-training, we fine-tune the pre-trained models on COCO for object detection and instance segmentation. The Mask R-CNN [17] with R50-FPN is used. As shown in Tab. 6, SCFS improves the performance significantly compared to its baseline DINO. In addition, compared to other dense pixel-level and region-level methods, such as DenseCL [40] and ORL [46], SCFS also achieves the best performance. This experiment verifies that SCFS can effectively solve the problem of semantic inconsistency during pre-training.",
+        "bbox": [
+            496,
+            657,
+            893,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "16104",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/1478d593ae9846acc2ccbcbeeebf2af6822945c5c83eacdd8dacbe41198dc827.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 6. Pre-training and than Fine-tuning on COCO [26] using Mask R-CNN [17] with ResNet50-FPN and  $1 \\times$  schedule. All models pre-trained on COCO are pre-trained with 512 batch size for 800 epochs. Bold font indicates the best result. * indicates pre-training on ImageNet [34]"
+        ],
+        "table_body": "<table><tr><td>Method</td><td>k-NN</td><td>APb</td><td>APb50</td><td>APb75</td><td>APs</td><td>APs50</td><td>APs75</td></tr><tr><td>Scratch</td><td>-</td><td>31.0</td><td>49.5</td><td>33.2</td><td>28.5</td><td>46.8</td><td>30.4</td></tr><tr><td>Supervised*</td><td>-</td><td>38.9</td><td>59.6</td><td>42.7</td><td>35.4</td><td>56.5</td><td>38.1</td></tr><tr><td>SimCLR [5]</td><td>35.3</td><td>37.0</td><td>56.8</td><td>40.3</td><td>33.7</td><td>53.8</td><td>36.1</td></tr><tr><td>MoCov2 [7]</td><td>39.4</td><td>38.5</td><td>58.1</td><td>42.1</td><td>34.8</td><td>55.3</td><td>37.3</td></tr><tr><td>BYOL [13]</td><td>42.6</td><td>39.5</td><td>59.3</td><td>43.2</td><td>35.6</td><td>56.5</td><td>38.2</td></tr><tr><td>DenseCL [40]</td><td>32.5</td><td>39.6</td><td>59.3</td><td>43.3</td><td>35.7</td><td>56.5</td><td>38.4</td></tr><tr><td>ORL [46]</td><td>43.6</td><td>40.3</td><td>60.2</td><td>44.4</td><td>36.3</td><td>57.3</td><td>38.9</td></tr><tr><td>UniVIP [25]</td><td>-</td><td>40.8</td><td>-</td><td>-</td><td>36.8</td><td>-</td><td>-</td></tr><tr><td>DINO [4]</td><td>42.6</td><td>39.0</td><td>59.6</td><td>42.9</td><td>35.6</td><td>56.8</td><td>38.0</td></tr><tr><td>SCFS</td><td>44.6</td><td>40.9</td><td>61.6</td><td>44.4</td><td>36.9</td><td>58.4</td><td>39.5</td></tr></table>",
+        "bbox": [
+            81,
+            88,
+            460,
+            266
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4. Ablation Studies",
+        "text_level": 1,
+        "bbox": [
+            76,
+            371,
+            240,
+            386
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We analyze the influence of each component in SCFS. To speed up the training time, the ImageNet100 dataset, which contains 100 randomly selected categories from ImageNet [34], is adopted. All the models are pre-trained on the ImageNet100 training set with 256 batch size for 200 epoch, and tested on the validation set. The  $k$ -NN and linear probing top-1 accuracy are used as the evaluation metrics.",
+        "bbox": [
+            75,
+            396,
+            468,
+            501
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Influence of Different Contrast Modes. The contrast mode can be divided into three types: contrast between two global data augmentations used in all contrastive learning methods  $(\\mathrm{G_d}2\\mathrm{G_d})$ ; contrast between local data augmentations and global data augmentations used in multi-crop strategy  $(\\mathrm{L_d}2\\mathrm{G_d})$ ; and contrast between local data augmentations and local feature augmentations used in SCFS  $(\\mathrm{L_d}2\\mathrm{L_f})$ .",
+        "bbox": [
+            75,
+            503,
+            468,
+            609
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The results are shown in Tab. 7. With multi-crop, DINO [4] (81.1%) improves accuracy by 3.0% compared to DINO without multi-crop baseline. SCFS (84.8%) further improves accuracy by 3.7% by introducing a contrast between local data augmentation and local feature augmentation. Some attention maps of SCFS and DINO are shown in Fig. 3. SCFS can more accurately focus on semantics-consistent regions between global view and local views, while DINO is easily influenced by background.",
+        "bbox": [
+            75,
+            609,
+            470,
+            747
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We also add multi-layer feature contrastive learning on DINO. The result in Tab. 7 (the \"DINO w ML\" row) verifies the improvements of SCFS are not totally owed to multi-layer contrast.",
+        "bbox": [
+            75,
+            748,
+            470,
+            808
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In addition, we directly crop the corresponding region of local augmentation on the feature map of global augmentation for contrastive learning. As shown in Tab. 7, this variant (ROIAlign) of SCFS also outperforms the DINO baseline, which shows that the directly cropped features are also beneficial for contrastive learning. And the ROIAlign variant",
+        "bbox": [
+            75,
+            810,
+            470,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/287e3074a6911c2fc81ac5c04e8a4c21eaf394efbd1df87d966a1160e6257505.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Contrast Mode</td><td>Gd2Gd</td><td>Ld2Gd</td><td>Ld2Lf</td><td>Gd2Gf</td><td>k-NN</td><td>LP</td></tr><tr><td>DINO w/o MC</td><td>✓</td><td></td><td></td><td></td><td>78.1</td><td>83.7</td></tr><tr><td>DINO</td><td>✓</td><td>✓</td><td></td><td></td><td>81.1</td><td>87.0</td></tr><tr><td>DINO w ML</td><td>✓</td><td>✓</td><td></td><td></td><td>82.2</td><td>87.4</td></tr><tr><td>SCFS</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>84.8</td><td>89.2</td></tr><tr><td>ROI Crop</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>83.9</td><td>88.1</td></tr><tr><td>SCFS w/o MC</td><td></td><td></td><td></td><td>✓</td><td>79.7</td><td>86.3</td></tr></table>",
+        "bbox": [
+            504,
+            89,
+            888,
+            205
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 7. Influence of different contrast modes. MC, ML, and LP denote multi-crop, multi-layer, and linear probing, respectively.",
+        "bbox": [
+            498,
+            214,
+            890,
+            243
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7f8aa03f2395644f4897a8537d84e58c10244e784f755028b6dc16c7b3747956.jpg",
+        "image_caption": [
+            "Figure 3. Attention maps of SCFS (the third row) compared with DINO [4] (the second row). In each example, (a) shows a global image, and its four local images in (b) are constructed by  $2 \\times 2$  jigsaw. (d) and (f) show the attention maps that highlight the semantics-consistent regions between the local images in (b) and the global image in (a). They are obtained by multiplying the globally average pooled feature maps from the encoder (Res4) of the local images in (b) with the feature map (Res4) of the global image in (a). And the encoder is the trained DINO ResNet50 model and SCFS ResNet50 model in (d) and (f), respectively. (c) and (e) show the mean attention maps of DINO and SCFS respectively, which are obtained by multiplying the mean globally average pooled feature map of the four local images in (b) with feature map of the global image in (a)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            511,
+            261,
+            875,
+            430
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "of SCFS achieves lower accuracy than SCFS, demonstrating that the soft feature search in SCFS is better than the hard ROIAlign since ROIAlign may damage the continuous semantic context of the feature map.",
+        "bbox": [
+            496,
+            640,
+            893,
+            700
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Further, we also test the performance of SCFS under the setting without multi-crop. That is, the feature search is conducted between two global data augmentations. We term this contrast mode as  $\\mathrm{G_d}2\\mathrm{G_f}$ . As shown in the \"SCFS w/o MC\" row, SCFS also improves the performance compared to its baseline (the \"DINO w/o MC\" row), which proves that SCFS is also helpful in solving the semantic inconsistency caused by other augmentations, not only the multi-crop augmentation strategy.",
+        "bbox": [
+            496,
+            702,
+            893,
+            839
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Influence of Multi-Layer Contrast. The influence of the feature layer that is used for feature search is analyzed. The Res2, Res3 and Res4 in the ResNet50 [18] backbone are evaluated. As shown in Tab. 8, the performance improves",
+        "bbox": [
+            496,
+            839,
+            893,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "16105",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/2e8294c96dda461c7703e22846aa5179dd74bab00a7f7c92563d475e60abd0eb.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Res2</td><td>Res3</td><td>Res4</td><td>k-NN</td><td>Linear Probing</td></tr><tr><td>✓</td><td></td><td></td><td>82.0</td><td>86.1</td></tr><tr><td>✓</td><td>✓</td><td></td><td>84.3</td><td>87.5</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>84.8</td><td>89.2</td></tr></table>",
+        "bbox": [
+            106,
+            88,
+            442,
+            160
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/32cc7bcf55c389eba83a68856410513d9704b827fa327950f9ffe264b5128b8b.jpg",
+        "image_caption": [
+            "Figure 4. The  $k$ -NN accuracy of features from different layers."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            196,
+            428,
+            345
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b5fcca0d8ad18777935c3d077f1178bacd21da419d0ae5694771f3ec4b1e8ecc.jpg",
+        "image_caption": [
+            "Figure 5. Influence of local augmentation number."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            377,
+            428,
+            523
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "with the increase of feature layer numbers, which demonstrates that conducting feature search on more layers is helpful for representation learning.",
+        "bbox": [
+            75,
+            566,
+            470,
+            612
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Further, we evaluate the  $k$ -NN accuracy using feature maps from different layers to observe the influence of feature search on the representation of middle layers. We also choose the features extracted by the Res2, Res3 and Res4 layer of ResNet50. The results are shown in Fig. 4. Compared with DINO [4], SCFS achieves better performance with features from all middle layers on ImageNet100, which verifies that enhancing the semantic consistency can improve the semantic representation of shallow layers. Compared with supervised learning, SCFS has higher performance on res2 and res3 layer, which shows that SCFS is more advantageous in the shallow layer feature representation.",
+        "bbox": [
+            75,
+            613,
+            468,
+            792
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Influence of Local Augmentation Number. In this part, we analyze the performance difference with the change of local augmentation numbers. The results are shown in Fig. 5. The performance of DINO and SCFS is steadily improved when adding more local augmentations for contrast. In addition, SCFS improves the performance under different local augmentation numbers, which demonstrates that semantics-",
+        "bbox": [
+            75,
+            795,
+            470,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/e9c309b5021e66bfb4a1297ecb79190f00e75c004ec45b7e6280ea2c0d9d1573.jpg",
+        "table_caption": [
+            "Table 8. Influence of different feature augmentation layer."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Cropping Scale</td><td>0.05-0.14</td><td>0.14-0.25</td><td>0.25-0.4</td><td>0.4-0.6</td></tr><tr><td>DINO [4]</td><td>87.0</td><td>87.2</td><td>85.6</td><td>82.6</td></tr><tr><td>SCFS</td><td>89.2</td><td>88.7</td><td>87.0</td><td>83.1</td></tr></table>",
+        "bbox": [
+            504,
+            88,
+            890,
+            137
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/c9da0c2713e13b0ec73f1c0914867db92bcfadc816e901347292fb8bf7915f3f.jpg",
+        "table_caption": [
+            "Table 9. Influence of random cropping scale. We report the linear probing top-1 accuracy."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Backbone</td><td>Batch Size</td><td>Epochs</td><td>k-NN</td><td>LP</td></tr><tr><td>DINO [4]</td><td>R101</td><td>256</td><td>200</td><td>81.0</td><td>86.3</td></tr><tr><td>SCFS</td><td>R101</td><td>256</td><td>200</td><td>85.1</td><td>88.3</td></tr><tr><td>DINO [4]</td><td>ViT-S-16</td><td>256</td><td>200</td><td>75.0</td><td>80.4</td></tr><tr><td>SCFS</td><td>ViT-S-16</td><td>256</td><td>200</td><td>76.3</td><td>81.0</td></tr><tr><td>DINO [4]</td><td>ViT-B-16</td><td>256</td><td>200</td><td>76.2</td><td>80.7</td></tr><tr><td>SCFS</td><td>ViT-B-16</td><td>256</td><td>200</td><td>77.2</td><td>82.3</td></tr></table>",
+        "bbox": [
+            504,
+            189,
+            883,
+            308
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 10. Experiments on other backbones. LP denotes linear probing.",
+        "bbox": [
+            498,
+            316,
+            890,
+            344
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "consistent feature search is helpful to alleviate the influence of semantics inconsistent data augmentations.",
+        "bbox": [
+            498,
+            371,
+            890,
+            401
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Influence of Random Cropping Scale. The hyperparameter of the random cropping scale influences the degree of semantic inconsistency. In this experiment, we analyze the performance of SCFS and DINO [4] under different random cropping scales. As shown in Tab. 9, owing to the feature search, SCFS is more adaptive to different cropping scales than the baseline DINO [4].",
+        "bbox": [
+            498,
+            402,
+            892,
+            508
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Experiments on Other Backbones. In this part, we conduct experiments on other backbones to further evaluate the effectiveness of SCFS. Apart for the default Reset50 used in other experiments, ResNet101 and Vision Transformer (ViT-S and ViT-B) are tested. The results are shown in Tab. 10. SCFS achieves significant improvement on different backbones compared to its baseline DINO, which demonstrates that SCFS is applicable to different backbones.",
+        "bbox": [
+            498,
+            510,
+            893,
+            630
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusions",
+        "text_level": 1,
+        "bbox": [
+            500,
+            647,
+            625,
+            662
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this study, we aim to alleviate the problem of unmatched semantic alignment in current contrastive learning by expanding the augmentations from data space to feature space. The proposed semantics-consistent feature search (SCFS) adaptively searches semantics-consistent local features between different views for contrast, while suppressing irrelevant local features during pre-training. It conducts contrast learning between feature augmentation and data augmentation. The experimental results demonstrate that SCFS can learn to focus on meaningful object regions and effectively improve the performance of self-supervised learning. The feature search procedure in SCFS is learnable parameter-free. We will utilize the self-attention mechanism in Transformer to perform the feature search procedure to further boost its performance in future work.",
+        "bbox": [
+            496,
+            674,
+            893,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "16106",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            173,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1",
+            "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. 2, 5",
+            "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. NeurIPS, 33:9912-9924, 2020. 2, 3, 5, 6, 12",
+            "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 2, 3, 4, 5, 6, 7, 8, 11, 12",
+            "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 1, 2, 5, 6, 7, 12",
+            "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. NeurIPS, 33:22243-22255, 2020. 1, 5",
+            "[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 6, 7",
+            "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. 2, 5",
+            "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 2, 5, 12",
+            "[10] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In ICCV, pages 9588–9597, 2021. 2, 5",
+            "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88(2):303–338, 2010. 6",
+            "[12] Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, and Patrick Perez. Obow: Online bag-of-visual-words generation for self-supervised learning. In CVPR, pages 6830–6840, 2021. 2, 5",
+            "[13] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 33:21271-21284, 2020. 2, 5, 6, 7, 12",
+            "[14] Yuanfan Guo, Minghao Xu, Jiawen Li, Bingbing Ni, Xuanyu Zhu, Zhenbang Sun, and Yi Xu. Hcsc: Hierarchical contrastive selective coding. In CVPR, pages 9706-9715, 2022. 2, 5",
+            "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 1, 2, 5, 6",
+            "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 6, 7",
+            "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 2, 3, 4, 5, 6, 7",
+            "[19] Qianjiang Hu, Xiao Wang, Wei Hu, and Guo-Jun Qi. Adco: Adversarial contrast for efficient learning of unsupervised representations from self-trained negative adversaries. In CVPR, pages 1074-1083, 2021. 2, 5",
+            "[20] Soroush Abbasi Koohpayegani, Ajinkya Tejankar, and Hamed Piri siavash. Mean shift for self-supervised learning. In ICCV, pages 10326-10335, 2021. 2, 5",
+            "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6",
+            "[22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6",
+            "[23] Xiaoni Li, Yu Zhou, Yifei Zhang, Aoting Zhang, Wei Wang, Ning Jiang, Haiying Wu, and Weiping Wang. Dense semantic contrast for self-supervised visual representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1368-1376, 2021. 2, 6",
+            "[24] Zeming Li, Songtao Liu, and Jian Sun. Momentum^2 teacher: Momentum teacher with momentum statistics for self-supervised learning. arXiv preprint arXiv:2101.07525, 2021. 2",
+            "[25] Zhaowen Li, Yousong Zhu, Fan Yang, Wei Li, Chaoyang Zhao, Yingying Chen, Zhiyang Chen, Jiahao Xie, Liwei Wu, Rui Zhao, et al. Univip: A unified framework for self-supervised visual pre-training. In CVPR, pages 14627-14636, 2022. 2, 5, 7",
+            "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 6, 7",
+            "[27] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 11",
+            "[28] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6",
+            "[29] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In  $CVPR$ , pages 3498-3505. IEEE, 2012. 6",
+            "[30] Nikhil Parthasarathy, SM Eslami, João Carreira, and Olivier J Henaff. Self-supervised video pretraining yields strong image representations. arXiv preprint arXiv:2210.06433, 2022. 3, 12",
+            "[31] Xiangyu Peng, Kai Wang, Zheng Zhu, Mang Wang, and Yang You. Crafting better contrastive views for siamese representation learning. In CVPR, pages 16031-16040, 2022. 2"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "16107",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[32] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 5, 6",
+            "[33] Byungseok Roh, Wuhyun Shin, Ildoo Kim, and Sungwooong Kim. Spatially consistent representation learning. In CVPR, pages 1144–1153, 2021. 2",
+            "[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 5, 7",
+            "[35] Kaiyou Song, Jin Xie, Shan Zhang, and Zimeng Luo. Multi-mode online knowledge distillation for self-supervised visual representation learning. In CVPR, 2023. 2",
+            "[36] Ajinkya Tejankar, Soroush Abbasi Koohpayegani, Vipin Pillai, Paolo Favaro, and Hamed Pirsiavash. Isd: Self-supervised learning by iterative similarity distillation. In ICCV, pages 9609-9618, 2021. 2, 5",
+            "[37] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6",
+            "[38] Haochen Wang, Kaiyou Song, Junsong Fan, Yuxi Wang, Jin Xie, and Zhaoxiang Zhang. Hard patches mining for masked image modeling. In CVPR, 2023. 1",
+            "[39] Xiao Wang and Guo-Jun Qi. Contrastive learning with stronger augmentations. IEEE TPAMI, 2022. 2, 5",
+            "[40] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In CVPR, pages 3024-3033, 2021. 2, 6, 7",
+            "[41] Zhaoqing Wang, Qiang Li, Guoxin Zhang, Pengfei Wan, Wen Zheng, Nannan Wang, Mingming Gong, and Tongliang Liu. Exploring set similarity for dense self-supervised representation learning. In CVPR, pages 16590-16599, 2022. 2, 3, 6",
+            "[42] Fangyun Wei, Yue Gao, Zhirong Wu, Han Hu, and Stephen Lin. Aligning pretraining for detection via object-level contrastive learning. NeurIPS, 34, 2021. 2, 6",
+            "[43] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742, 2018. 5",
+            "[44] Tete Xiao, Colorado J Reed, Xiaolong Wang, Kurt Keutzer, and Trevor Darrell. Region similarity representation learning. In ICCV, pages 10539–10548, 2021. 2, 6",
+            "[45] Enze Xie, Jian Ding, Wenhai Wang, Xiaohang Zhan, Hang Xu, Peize Sun, Zhenguo Li, and Ping Luo. Detco: Unsupervised contrastive learning for object detection. In ICCV, pages 8392-8401, 2021. 2, 5, 6",
+            "[46] Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Unsupervised object-level representation learning from scene images. NeurIPS, 34:28864-28876, 2021. 2, 6, 7",
+            "[47] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In CVPR, pages 16684-16693, 2021. 2",
+            "[48] Haohang Xu, Xiaopeng Zhang, Hao Li, Lingxi Xie, Wenrui Dai, Hongkai Xiong, and Qi Tian. Seed the views: Hierarchi"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "cal semantic alignment for contrastive representation learning. IEEE TPAMI, 2022. 2, 6",
+            "[49] Chun-Hsiao Yeh, Cheng-Yao Hong, Yen-Chi Hsu, Tyng-Luh Liu, Yubei Chen, and Yann LeCun. Decoupled contrastive learning. In ECCV, pages 668-684. Springer, 2022. 2, 5",
+            "[50] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, pages 12310-12320. PMLR, 2021. 2, 5",
+            "[51] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In ICCV, pages 1476-1485, 2019. 5"
+        ],
+        "bbox": [
+            503,
+            92,
+            892,
+            247
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "16108",
+        "bbox": [
+            480,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 9
+    }
+]