Add Batch fb309d1d-5e57-43f9-bcd2-ec761930d450 data

.gitattributes

@@ -3301,3 +3301,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2024/EASE-DETR_[[:space:]]Easing[[:space:]]the[[:space:]]Competition[[:space:]]among[[:space:]]Object[[:space:]]Queries/9ff0d4e0-f729-41a1-9284-a3462c62983a_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/ECLIPSE_[[:space:]]A[[:space:]]Resource-Efficient[[:space:]]Text-to-Image[[:space:]]Prior[[:space:]]for[[:space:]]Image[[:space:]]Generations/83a8980c-9503-4c5f-99c0-a6eb3b04e979_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/ECLIPSE_[[:space:]]Efficient[[:space:]]Continual[[:space:]]Learning[[:space:]]in[[:space:]]Panoptic[[:space:]]Segmentation[[:space:]]with[[:space:]]Visual[[:space:]]Prompt[[:space:]]Tuning/9f39aafa-e0ef-4ad0-a765-ad142df3abb7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/ECoDepth_[[:space:]]Effective[[:space:]]Conditioning[[:space:]]of[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]Monocular[[:space:]]Depth[[:space:]]Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EFHQ_[[:space:]]Multi-purpose[[:space:]]ExtremePose-Face-HQ[[:space:]]dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EFormer_[[:space:]]Enhanced[[:space:]]Transformer[[:space:]]towards[[:space:]]Semantic-Contour[[:space:]]Features[[:space:]]of[[:space:]]Foreground[[:space:]]for[[:space:]]Portraits[[:space:]]Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EGTR_[[:space:]]Extracting[[:space:]]Graph[[:space:]]from[[:space:]]Transformer[[:space:]]for[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation/174c2199-8aa8-4e45-8d2e-205eb2e2e099_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EMAGE_[[:space:]]Towards[[:space:]]Unified[[:space:]]Holistic[[:space:]]Co-Speech[[:space:]]Gesture[[:space:]]Generation[[:space:]]via[[:space:]]Expressive[[:space:]]Masked[[:space:]]Audio[[:space:]]Gesture[[:space:]]Modeling/f8a85da7-9221-4a37-baf0-e5a3b46705e3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EMCAD_[[:space:]]Efficient[[:space:]]Multi-scale[[:space:]]Convolutional[[:space:]]Attention[[:space:]]Decoding[[:space:]]for[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation/32047b53-ccf5-41e9-bc53-4e712089f958_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EMOPortraits_[[:space:]]Emotion-enhanced[[:space:]]Multimodal[[:space:]]One-shot[[:space:]]Head[[:space:]]Avatars/ab68d1cd-73a3-4a4c-9896-716aeaaff436_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/ERMVP_[[:space:]]Communication-Efficient[[:space:]]and[[:space:]]Collaboration-Robust[[:space:]]Multi-Vehicle[[:space:]]Perception[[:space:]]in[[:space:]]Challenging[[:space:]]Environments/197e9226-8724-4a9f-ac9a-1556459fa452_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/ES3_[[:space:]]Evolving[[:space:]]Self-Supervised[[:space:]]Learning[[:space:]]of[[:space:]]Robust[[:space:]]Audio-Visual[[:space:]]Speech[[:space:]]Representations/09b7366b-807f-4dfb-a9c8-45edf9472928_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/ESCAPE_[[:space:]]Encoding[[:space:]]Super-keypoints[[:space:]]for[[:space:]]Category-Agnostic[[:space:]]Pose[[:space:]]Estimation/64b3bbb9-9580-410a-adfe-c8a2bd5b8bca_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/ESR-NeRF_[[:space:]]Emissive[[:space:]]Source[[:space:]]Reconstruction[[:space:]]Using[[:space:]]LDR[[:space:]]Multi-view[[:space:]]Images/2c6d481f-7a23-410c-a62a-93255bbc656e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EVCap_[[:space:]]Retrieval-Augmented[[:space:]]Image[[:space:]]Captioning[[:space:]]with[[:space:]]External[[:space:]]Visual-Name[[:space:]]Memory[[:space:]]for[[:space:]]Open-World[[:space:]]Comprehension/003cc44f-e449-40f1-aec0-baf1c6235010_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EVS-assisted[[:space:]]Joint[[:space:]]Deblurring[[:space:]]Rolling-Shutter[[:space:]]Correction[[:space:]]and[[:space:]]Video[[:space:]]Frame[[:space:]]Interpolation[[:space:]]through[[:space:]]Sensor[[:space:]]Inverse[[:space:]]Modeling/2babbe27-3090-44dc-8f9d-2d16b50cc5ab_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Each[[:space:]]Test[[:space:]]Image[[:space:]]Deserves[[:space:]]A[[:space:]]Specific[[:space:]]Prompt_[[:space:]]Continual[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]for[[:space:]]2D[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation/b547c627-f0bf-4b29-9f5d-10851e261fb1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EarthLoc_[[:space:]]Astronaut[[:space:]]Photography[[:space:]]Localization[[:space:]]by[[:space:]]Indexing[[:space:]]Earth[[:space:]]from[[:space:]]Space/09848c3c-ff6b-4e53-9236-ab0f624d8495_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EasyDrag_[[:space:]]Efficient[[:space:]]Point-based[[:space:]]Manipulation[[:space:]]on[[:space:]]Diffusion[[:space:]]Models/3c97514b-8a67-4bdc-8795-3e7cfc02ceca_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Eclipse_[[:space:]]Disambiguating[[:space:]]Illumination[[:space:]]and[[:space:]]Materials[[:space:]]using[[:space:]]Unintended[[:space:]]Shadows/a9a99e7b-5bee-4200-8036-13e74c55e36e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Edge-Aware[[:space:]]3D[[:space:]]Instance[[:space:]]Segmentation[[:space:]]Network[[:space:]]with[[:space:]]Intelligent[[:space:]]Semantic[[:space:]]Prior/ef85b90b-3f1b-4322-ac5a-c5d6ae60802c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Edit[[:space:]]One[[:space:]]for[[:space:]]All_[[:space:]]Interactive[[:space:]]Batch[[:space:]]Image[[:space:]]Editing/d0962876-52e1-4fef-ac92-9ddbf4fa5230_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EditGuard_[[:space:]]Versatile[[:space:]]Image[[:space:]]Watermarking[[:space:]]for[[:space:]]Tamper[[:space:]]Localization[[:space:]]and[[:space:]]Copyright[[:space:]]Protection/43cb9c2b-dd1a-4670-9150-d4345483687e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Editable[[:space:]]Scene[[:space:]]Simulation[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving[[:space:]]via[[:space:]]Collaborative[[:space:]]LLM-Agents/49eb43d9-92f1-4849-a7c5-3da61380ffc6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Effective[[:space:]]Video[[:space:]]Mirror[[:space:]]Detection[[:space:]]with[[:space:]]Inconsistent[[:space:]]Motion[[:space:]]Cues/423d0156-aefb-4c1d-9839-9825d419ff81_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]3D[[:space:]]Implicit[[:space:]]Head[[:space:]]Avatar[[:space:]]with[[:space:]]Mesh-anchored[[:space:]]Hash[[:space:]]Table[[:space:]]Blendshapes/afaaa8c5-698e-4fcb-9bde-21aa2469de9a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Dataset[[:space:]]Distillation[[:space:]]via[[:space:]]Minimax[[:space:]]Diffusion/75146b44-b76a-42b0-8456-4c37446317b7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Deformable[[:space:]]ConvNets_[[:space:]]Rethinking[[:space:]]Dynamic[[:space:]]and[[:space:]]Sparse[[:space:]]Operator[[:space:]]for[[:space:]]Vision[[:space:]]Applications/d839f1b7-f6aa-4faa-bb5a-7c661a9aa4e5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Detection[[:space:]]of[[:space:]]Long[[:space:]]Consistent[[:space:]]Cycles[[:space:]]and[[:space:]]its[[:space:]]Application[[:space:]]to[[:space:]]Distributed[[:space:]]Synchronization/c0dd6b15-09a8-4eba-bcfb-24b8083dd375_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Hyperparameter[[:space:]]Optimization[[:space:]]with[[:space:]]Adaptive[[:space:]]Fidelity[[:space:]]Identification/fa155004-89c9-4599-9ab7-9a5d6edc5fbc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]LoFTR_[[:space:]]Semi-Dense[[:space:]]Local[[:space:]]Feature[[:space:]]Matching[[:space:]]with[[:space:]]Sparse-Like[[:space:]]Speed/14a55345-a1a3-4ac4-a5a2-7c153c54b832_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Meshflow[[:space:]]and[[:space:]]Optical[[:space:]]Flow[[:space:]]Estimation[[:space:]]from[[:space:]]Event[[:space:]]Cameras/141a2ee1-d1ef-416d-8c61-bc462b1074b9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Model[[:space:]]Stealing[[:space:]]Defense[[:space:]]with[[:space:]]Noise[[:space:]]Transition[[:space:]]Matrix/5181a0c8-f9af-47c9-b829-4e394817b1e7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Multi-scale[[:space:]]Network[[:space:]]with[[:space:]]Learnable[[:space:]]Discrete[[:space:]]Wavelet[[:space:]]Transform[[:space:]]for[[:space:]]Blind[[:space:]]Motion[[:space:]]Deblurring/4af0ea1e-f85d-4e0b-ab29-04d268e0386c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Multitask[[:space:]]Dense[[:space:]]Predictor[[:space:]]via[[:space:]]Binarization/b91bc222-326c-4762-aabe-433ec8965e13_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Privacy-Preserving[[:space:]]Visual[[:space:]]Localization[[:space:]]Using[[:space:]]3D[[:space:]]Ray[[:space:]]Clouds/61345f4c-613d-4e27-91d1-09fcd6a52a24_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Scene[[:space:]]Recovery[[:space:]]Using[[:space:]]Luminous[[:space:]]Flux[[:space:]]Prior/0327d2f4-6c3a-4e74-a5a8-041807da336a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Solution[[:space:]]of[[:space:]]Point-Line[[:space:]]Absolute[[:space:]]Pose/8561d28f-e735-46cc-9618-c6c206bdafc2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Stitchable[[:space:]]Task[[:space:]]Adaptation/03126c3f-71ba-4532-921c-1c95d792e515_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models/f86a8c24-7089-4003-88ca-fa1a5c70fe2d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]Vision-Language[[:space:]]Pre-training[[:space:]]by[[:space:]]Cluster[[:space:]]Masking/24f9f648-410c-4f99-8615-dc6d0380efad_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficient[[:space:]]and[[:space:]]Effective[[:space:]]Weakly-Supervised[[:space:]]Action[[:space:]]Segmentation[[:space:]]via[[:space:]]Action-Transition-Aware[[:space:]]Boundary[[:space:]]Alignment/1ae3fdc0-77c3-4dc9-87e2-e3b0568d2efa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EfficientDreamer_[[:space:]]High-Fidelity[[:space:]]and[[:space:]]Robust[[:space:]]3D[[:space:]]Creation[[:space:]]via[[:space:]]Orthogonal-view[[:space:]]Diffusion[[:space:]]Priors/6a261e68-4855-488d-9a23-298bfd3be3c9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EfficientSAM_[[:space:]]Leveraged[[:space:]]Masked[[:space:]]Image[[:space:]]Pretraining[[:space:]]for[[:space:]]Efficient[[:space:]]Segment[[:space:]]Anything/f1e649e9-e5d4-48b3-8018-af1f58065ff2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Efficiently[[:space:]]Assemble[[:space:]]Normalization[[:space:]]Layers[[:space:]]and[[:space:]]Regularization[[:space:]]for[[:space:]]Federated[[:space:]]Domain[[:space:]]Generalization/16ce0297-575e-48b3-acb7-4a9aa0af4fe6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Ego-Exo4D_[[:space:]]Understanding[[:space:]]Skilled[[:space:]]Human[[:space:]]Activity[[:space:]]from[[:space:]]First-[[:space:]]and[[:space:]]Third-Person[[:space:]]Perspectives/742959dc-4498-434e-af9a-b1cc7f6e5bd2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EgoExoLearn_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Bridging[[:space:]]Asynchronous[[:space:]]Ego-[[:space:]]and[[:space:]]Exo-centric[[:space:]]View[[:space:]]of[[:space:]]Procedural[[:space:]]Activities[[:space:]]in[[:space:]]Real[[:space:]]World/134e2089-370f-4ad6-9686-84a1d7af0ed4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EgoGen_[[:space:]]An[[:space:]]Egocentric[[:space:]]Synthetic[[:space:]]Data[[:space:]]Generator/717ad579-fa9d-4ce5-996d-77c046c2cdbf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EgoThink_[[:space:]]Evaluating[[:space:]]First-Person[[:space:]]Perspective[[:space:]]Thinking[[:space:]]Capability[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models/2ff0d147-b907-475e-8aa3-5584b71f3014_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Egocentric[[:space:]]Whole-Body[[:space:]]Motion[[:space:]]Capture[[:space:]]with[[:space:]]FisheyeViT[[:space:]]and[[:space:]]Diffusion-Based[[:space:]]Motion[[:space:]]Refinement/90ece1d3-5ea4-4b56-9fe8-5a7d9a524eef_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/ElasticDiffusion_[[:space:]]Training-free[[:space:]]Arbitrary[[:space:]]Size[[:space:]]Image[[:space:]]Generation[[:space:]]through[[:space:]]Global-Local[[:space:]]Content[[:space:]]Separation/de40da54-b1b7-40fa-9aec-0f4f1bc4f406_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Elite360D_[[:space:]]Towards[[:space:]]Efficient[[:space:]]360[[:space:]]Depth[[:space:]]Estimation[[:space:]]via[[:space:]]Semantic-[[:space:]]and[[:space:]]Distance-Aware[[:space:]]Bi-Projection[[:space:]]Fusion/b933bca2-9ff4-42fc-9bd7-be64c39e1936_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Embodied[[:space:]]Multi-Modal[[:space:]]Agent[[:space:]]trained[[:space:]]by[[:space:]]an[[:space:]]LLM[[:space:]]from[[:space:]]a[[:space:]]Parallel[[:space:]]TextWorld/911fda92-445d-4ff5-8ef4-3ff0a7daf4bf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EmbodiedScan_[[:space:]]A[[:space:]]Holistic[[:space:]]Multi-Modal[[:space:]]3D[[:space:]]Perception[[:space:]]Suite[[:space:]]Towards[[:space:]]Embodied[[:space:]]AI/ab5e0a2a-2a5f-41d4-a08d-01fd4564b83b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Embracing[[:space:]]Unimodal[[:space:]]Aleatoric[[:space:]]Uncertainty[[:space:]]for[[:space:]]Robust[[:space:]]Multimodal[[:space:]]Fusion/b864c4f0-5e7f-441f-a2a0-028202223b38_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Emergent[[:space:]]Open-Vocabulary[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]from[[:space:]]Off-the-shelf[[:space:]]Vision-Language[[:space:]]Models/bf254b56-9ba7-4039-9f52-56365463bb00_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EmoGen_[[:space:]]Emotional[[:space:]]Image[[:space:]]Content[[:space:]]Generation[[:space:]]with[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models/a0b8035c-47d5-4f15-8dd7-134ed07d94b6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/EmoVIT_[[:space:]]Revolutionizing[[:space:]]Emotion[[:space:]]Insights[[:space:]]with[[:space:]]Visual[[:space:]]Instruction[[:space:]]Tuning/337f29a8-74df-4960-89a4-6eff2797f1cc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Emotional[[:space:]]Speech-driven[[:space:]]3D[[:space:]]Body[[:space:]]Animation[[:space:]]via[[:space:]]Disentangled[[:space:]]Latent[[:space:]]Diffusion/8fe0dd5b-5118-4e0a-ac09-f70da9816cdc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Empowering[[:space:]]Resampling[[:space:]]Operation[[:space:]]for[[:space:]]Ultra-High-Definition[[:space:]]Image[[:space:]]Enhancement[[:space:]]with[[:space:]]Model-Aware[[:space:]]Guidance/5e31bf7f-5e48-4030-90cc-b526f724e932_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Emu[[:space:]]Edit_[[:space:]]Precise[[:space:]]Image[[:space:]]Editing[[:space:]]via[[:space:]]Recognition[[:space:]]and[[:space:]]Generation[[:space:]]Tasks/b6371d78-25ca-4b11-84b1-a566ed9d62f2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/En3D_[[:space:]]An[[:space:]]Enhanced[[:space:]]Generative[[:space:]]Model[[:space:]]for[[:space:]]Sculpting[[:space:]]3D[[:space:]]Humans[[:space:]]from[[:space:]]2D[[:space:]]Synthetic[[:space:]]Data/cb7c8896-3bc3-4bc4-a3cb-b2aedebff579_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/End-to-End[[:space:]]Spatio-Temporal[[:space:]]Action[[:space:]]Localisation[[:space:]]with[[:space:]]Video[[:space:]]Transformers/62e0dda5-0752-47db-aad5-cb1b83945b43_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/End-to-End[[:space:]]Temporal[[:space:]]Action[[:space:]]Detection[[:space:]]with[[:space:]]1B[[:space:]]Parameters[[:space:]]Across[[:space:]]1000[[:space:]]Frames/aed9af76-c260-41b9-be21-a2e11ea6ab0b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Endow[[:space:]]SAM[[:space:]]with[[:space:]]Keen[[:space:]]Eyes_[[:space:]]Temporal-spatial[[:space:]]Prompt[[:space:]]Learning[[:space:]]for[[:space:]]Video[[:space:]]Camouflaged[[:space:]]Object[[:space:]]Detection/e8cb1a37-360c-44d3-92de-d83afed15e31_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Enhance[[:space:]]Image[[:space:]]Classification[[:space:]]via[[:space:]]Inter-Class[[:space:]]Image[[:space:]]Mixup[[:space:]]with[[:space:]]Diffusion[[:space:]]Model/d24eac86-8eb2-4146-81d8-2b9455b8e75f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Enhanced[[:space:]]Motion-Text[[:space:]]Alignment[[:space:]]for[[:space:]]Image-to-Video[[:space:]]Transfer[[:space:]]Learning/1b431374-7f10-4d10-9e33-68850035e408_origin.pdf filter=lfs diff=lfs merge=lfs -text

2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_content_list.json

@@ -0,0 +1,1329 @@
+[
+    {
+        "type": "text",
+        "text": "ECoDepth: Effective Conditioning of Diffusion Models for Monocular Depth Estimation",
+        "text_level": 1,
+        "bbox": [
+            207,
+            130,
+            759,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Suraj Patni*",
+        "bbox": [
+            222,
+            203,
+            320,
+            220
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Aradhye Agarwal*",
+        "bbox": [
+            408,
+            204,
+            557,
+            220
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Chetan Arora",
+        "bbox": [
+            635,
+            204,
+            745,
+            219
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Indian Institute of Technology Delhi",
+        "bbox": [
+            339,
+            222,
+            629,
+            239
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "https://ecodepth-iitd.github.io",
+        "bbox": [
+            297,
+            239,
+            666,
+            256
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            291,
+            312,
+            306
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In the absence of parallax cues, a learning based single image depth estimation (SIDE) model relies heavily on shading and contextual cues in the image. While this simplicity is attractive, it is necessary to train such models on large and varied datasets, which are difficult to capture. It has been shown that using embeddings from pretrained foundational models, such as CLIP, improves zero shot transfer in several applications. Taking inspiration from this, in our paper we explore the use of global image priors generated from a pre-trained ViT model to provide more detailed contextual information. We argue that the embedding vector from a ViT model, pre-trained on a large dataset, captures greater relevant information for SIDE than the usual route of generating pseudo image captions, followed by CLIP based text embeddings. Based on this idea, we propose a new SIDE model using a diffusion backbone which is conditioned on ViT embeddings. Our proposed design establishes a new state-of-the-art (SOTA) for SIDE on NYU Depth v2 dataset, achieving Abs Rel error of 0.059(14% improvement) compared to 0.069 by the current SOTA (VPD). And on KITTI dataset, achieving Sq Rel error of 0.139 (2% improvement) compared to 0.142 by the current SOTA (GED). For zero shot transfer with a model trained on NYU Depth v2, we report mean relative improvement of (20%, 23%, 81%, 25%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, compared to (16%, 18%, 45%, 9%) by ZoEDepth. The code is available in our project page.",
+        "bbox": [
+            75,
+            323,
+            473,
+            748
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            775,
+            207,
+            790
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Single Image Depth Estimation (SIDE) is the task of predicting per pixel depth using a single RGB image from a monocular camera. It is a fundamental problem in computer vision with applications in several domains, viz robotics, autonomous driving, and augmented reality. The problem is",
+        "bbox": [
+            75,
+            801,
+            468,
+            878
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/f7274f44c4d22bc9eae8a19f69aa92c1c20646d54843a02397296650e2bfc12a.jpg",
+        "image_caption": [
+            "(a) Sun-RGBD [42]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            290,
+            696,
+            330
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/214cfea1c3468673bbbfb8b1f65db15a952fafc13969355b377513969847d17f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            290,
+            890,
+            330
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/ff8d25236e5638817716f97d36604908ba6a66250a0304d48025d379b65c2095.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            344,
+            694,
+            383
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/f56327bab79fcd3f27c08fbef298e0b0e23fd81917afe234efe190f0d7e73e38.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            344,
+            890,
+            383
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/acfee1ae7e238a4d60c7914c77d3468b68bc4fb10c24754766404c5d2a5dd4a9.jpg",
+        "image_caption": [
+            "(b) iBims1 [19]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            398,
+            696,
+            439
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/b6c9dc2b7d14f841f908bf571798f99779b924830b846d5fe98305823fd3225f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            398,
+            890,
+            439
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/5a9b6186ef62db411c91c5843145d756b45f98d60f13a0e2aa97b4f3d446142e.jpg",
+        "image_caption": [
+            "(c) DIODE [49]",
+            "(d) HyperSim [35]",
+            "Figure 1. Qualitative results across four different datasets, demonstrating the zero-shot performance of our model trained only on the NYU Depth v2 dataset. Corresponding quantitative results are presented in Table 3. The first column displays RGB images, the second column depicts ground truth depth, and the third column showcases our model's predicted depths. Additional images for each dataset are available in the Supplementary Material."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            452,
+            696,
+            493
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/8d5070d4cbc6c4ebcfef4196245823ef4f87ec26324f3274b07d2d886dd1ddc2.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            452,
+            890,
+            493
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "typically formulated in two flavors: metric depth estimation (MDE), and relative depth estimation (RDE). As the names suggest, MDE deals with estimation in physical units such as meters, whereas, RDE techniques focus on relative depth only, and require a per-image affine transformation as post-processing to convert to a physical unit.",
+        "bbox": [
+            496,
+            642,
+            890,
+            734
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Conventional geometric techniques for depth estimation typically rely on feature correspondence, parallax, and triangulation from two or more views. However, the problem becomes ill-posed for estimation from a single view. Intuitively, depth map is a 3D representation of the scene, whereas an RGB image is a 2D projection of the scene. Hence, it cannot be uniquely determined from a single RGB image. Therefore, learning-based SIDE models rely on visual cues like 'shape from shading' and other contextual priors for per-pixel depth prediction.",
+        "bbox": [
+            496,
+            734,
+            890,
+            885
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "A data-driven approach makes the SIDE pipeline sim",
+        "bbox": [
+            517,
+            885,
+            890,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Equal contribution.",
+        "bbox": [
+            94,
+            886,
+            205,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "28285",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "pler, but also makes the learnt model dependent on the quality of training data. It has been observed that such models overfit on a particular training distribution/domain and fail to generalize on unseen data. This is especially true for MDE models when the range of depth in the training dataset is limited. Hence, training on multiple datasets, with wide variations in depth ranges has been proposed [4, 34].",
+        "bbox": [
+            75,
+            90,
+            470,
+            196
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "On the other hand, development of large foundational models (LFMs) in recent years has altered the preferred design approach for many computer vision problems. These huge models are trained using extensive datasets of unlabeled images and learning objectives that are agnostic to specific tasks. The learnt embeddings from such pre-trained models have been shown to help generalization and zero-shot transfer in many applications. We are aware of at least two works for MDE in SIDE problem that have appeared in the last few months and make use of such foundational models. VPD [54] uses a text-to-image diffusion model pretrained on LAION-400M [39] dataset having large-scale image-text pairs as the backbone. The model prompts the denoising UNet with textual inputs to make the visual contents interact with the text prompts. Since the problem formulation doesn't include text description as the input, the model generates simple descriptions such as 'A photo of a {scene name}' based on the scene label given in the NYU Depth v2 dataset. Another work, TADP [20] improves upon VPD. Instead of simple image descriptions, TADP uses BLIP-2 [23] to generate image captions, and then uses CLIP [31] embeddings of the pseudo-caption to condition the diffusion model.",
+        "bbox": [
+            75,
+            200,
+            473,
+            547
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We view the above works as providing robust semantic context to LFMs for the actual task at hand, which helps in visual recognition in general, as well as SIDE. While we do agree with the broad motivations of these works, the question that we ask is if pseudo-captions are the most effective way to provide the semantic context. Textual descriptions of an image typically focus on large salient objects and emphasize on their relationships. On the other hand, large vision models for image classification, typically contain representation for even smaller objects present in the scene. Even when a single object is present in the scene, the representations typically capture uncertainties and ambiguities inherent in the scene. Hence, we posit that using embeddings from a transformer model, pretrained on a large image dataset unrelated to the SIDE task, captures more relevant information, and is a better alternative to using pseudocaptions' CLIP embedding. Since diffusion-based models have shown their superiority for the dense prediction tasks in recent works [8, 16, 20, 38, 54]. Hence, we propose a diffusion backbone for our model, along with a novel CIDE module which employs ViT [7] to extract semantic context embeddings. These embeddings are subsequently utilized to condition the diffusion backbone.",
+        "bbox": [
+            75,
+            553,
+            472,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Contributions. The key contributions of this work are:",
+        "bbox": [
+            500,
+            90,
+            864,
+            104
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(1) We propose a new model for MDE in a SIDE task. The proposed model uses a conditional diffusion architecture with the semantic context being supplied through embeddings generated using a ViT model. Achieving a new state-of-the-art (SOTA) performance, our method outperforms existing approaches on benchmark datasets, including the NYU Depth v2 indoor and KITTI outdoor datasets. Notably, we report a significant improvement of  $14\\%$  in absolute relative error, achieving 0.059 compared to the current SOTA (VPD) performance of 0.069 on NYU Depth v2. And we report an improvement of  $2\\%$  in square relative error, achieving 0.139 compared to the current SOTA (GED) performance of 0.142 on KITTI.",
+            "(2) We show, qualitatively as well as quantitatively, that using ViT embeddings to provide semantic context is a better alternative to generating pseudo captions and then using its CLIP embeddings to condition a SIDE model. In contrast to TADP [20], which uses pseudo captions, but only achieves a RMSE of 0.225 on NYU Depth v2, we report a lower, and a new SOTA, error of 0.218 (VPD [54] achieves 0.254).",
+            "(3) We show that providing ViT conditioning, helps our model perform better in a zero shot transfer task. ZoEDepth [4], the current SOTA for zero-shot transfer, reports an improvement of (16%, 18%, 45%, 9%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, after training their model on 12 other datasets and NYU Depth v2. In contrast, we only train on NYU Depth v2, and report a much larger improvement of (21%, 23%, 81%, 25%)."
+        ],
+        "bbox": [
+            500,
+            106,
+            893,
+            574
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            585,
+            640,
+            601
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Traditional Methods. Earlier techniques for SIDE have used Markov Random Fields [47], non-parametric depth sampling [18], and structural similarity with prior depth map [13] to predict pixel-wise depth.",
+        "bbox": [
+            498,
+            611,
+            890,
+            672
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Deep Learning Techniques for SIDE. Modern techniques have approached the problem as a dense regression problem. CNNs have been the dominant architecture for the SIDE in the last decade, with global-local network stack [9], and multi-scale [21], or encoder-decoder architecture [11] as some of the popular solution strategies. Recently, PixelFormer [1] used a transformer-based encoder-decoder architecture with skip connections from encoders to decoders. MIM [51] proposed masked image modeling as a general-purpose pre-training for geometric and motion tasks such as SIDE and pose estimation. Similarly, AiT [28] used mask augmentation and proposed soft tokens to generalize visual prediction tasks. While earlier works either focused on MDE for a specific dataset or RDE for generalization on multiple datasets, ZoEDepth [4] has proposed",
+        "bbox": [
+            498,
+            674,
+            893,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "28286",
+        "bbox": [
+            478,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/c7e530d0b4f2d38341472733a1549e75c90b00840b1be1a8bba2c60c1793cff7.jpg",
+        "image_caption": [
+            "Figure 2. An overview of our proposed model: The latent representation of the input image undergoes a diffusion process, which is conditioned by our proposed CIDe module. Within the CIDe module, the input image is fed through the frozen ViT model. From this, a linear combination of the learnt embeddings is computed, which is transformed to generate a 768-dimensional contextual embedding. This embedding is utilized to condition the diffusion backbone. Subsequently, hierarchical feature maps are extracted from the UNet's decoder which are concatenated and processed through a depth regressor to generate the depth map."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            92,
+            880,
+            300
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "a generalized method for MDE that performs well in zero-shot transfer. We outperform [4] by a large margin even when training on a single dataset - NYU Depth v2 for indoor scenes or KITTI for outdoor scenes. Whereas, [4] uses 12 datasets for pre-training and then fine-tunes on NYU Depth v2 or KITTI for zero shot transfer.",
+        "bbox": [
+            75,
+            410,
+            467,
+            500
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Diffusion-based Methods with Pretraining on Large Datasets. Recently, many techniques for SIDE has used diffusion architectures. These techniques [8, 16, 20, 38, 54], exploit prior knowledge acquired by pretraining on large datasets like LAION-400M [39], which consists of 400 million image-text pairs. In contrast, depth datasets, such as NYU Depth v2 and KITTI, contain around 20-30 thousand image-depth pairs. DepthGen [38] and DDP [16] work on a noise-to-depthmap paradigm and use images for conditional guidance of the diffusion process. DepthGen employs self-supervised pretraining on tasks like colorization, inpainting, and JPEG artifact removal, followed by supervised training on indoor and outdoor datasets [6, 12, 26, 27, 48]. DDP [16] decouples the image encoder and map decoder, allowing the image encoder to run just once, while the lightweight map decoder is run multiple times. VPD [54] and TADP [20] use denoising UNet [37] as a backbone to extract the rich features at multiple scales. Also, they utilize text instead of image for conditioning the diffusion backbone.",
+        "bbox": [
+            75,
+            503,
+            467,
+            805
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Vision Transformer for Scene Understanding. The transformer architecture was initially proposed for NLP tasks [50], but introduced to the computer vision community as Vision Transformer (ViT) in [7]. Prior to this, CNNs dominated computer vision problems due to their ability to capture spatial hierarchies in image data. However, such archi",
+        "bbox": [
+            75,
+            809,
+            467,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "tectures were constrained due to their ability to learn spatially localized features only. Transformer architecture has a weaker inductive bias and allows ViT to learn long-range dependencies, and robust, generalizable features. ViT architectures have replaced CNNs for SOTA performance on most computer vision tasks in recent years. We use a pretrained ViT model for providing semantic information to the diffusion backbone in our model.",
+        "bbox": [
+            498,
+            410,
+            890,
+            530
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Proposed Methodology",
+        "text_level": 1,
+        "bbox": [
+            500,
+            561,
+            718,
+            579
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            500,
+            590,
+            640,
+            606
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Problem Formulation. The objective of single image depth prediction task is to predict continuous values, denoted as  $\\mathbf{y} \\in \\mathbb{R}^{H \\times W}$ , for every pixel present in the input RGB image,  $\\mathbf{x} \\in [0,255]^{3 \\times H \\times W}$ . Here  $H$  and  $W$  represent the height and width respectively of the input image.",
+        "bbox": [
+            498,
+            619,
+            890,
+            696
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Diffusion Model. Diffusion models are a class of generative models that progressively inject noise into the input data (forward pass) and then learn to reconstruct the original data in a reverse denoising process (reverse pass). There are three formulations of diffusion models: denoising diffusion probabilistic models (DDPMs) [14], score-based generative models [43, 44], and those based on stochastic differential equations [45, 46]. DDPMs are of relevance to our paper, and are described below. The architecture of DDPMs consists of two Markov chains: a forward chain that adds noise to the data, and a reverse chain that converts noise back to data by learning transition kernels parameterized by deep neural networks. Formally, the forward pass is modeled as",
+        "bbox": [
+            496,
+            703,
+            892,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "28287",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "a Markov process:",
+        "bbox": [
+            76,
+            90,
+            202,
+            104
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbb {P} \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}\\right). \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            113,
+            468,
+            140
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Here  $\\mathbf{z}_t$  denotes the random variable at the  $t^{\\mathrm{th}}$  time step,  $\\mathcal{N}(\\mathbf{z};\\boldsymbol {\\mu},\\boldsymbol {\\sigma})$  denotes Gaussian probability distribution, and  $\\beta_{t}$  is the noise schedule. The above equation leads to the analytic form of  $\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{z}_0),\\forall t\\in \\{0,1,\\dots ,T\\}$ :",
+        "bbox": [
+            76,
+            147,
+            468,
+            208
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {z} _ {t} = \\sqrt {\\bar {\\beta} _ {t}} \\mathbf {z} _ {0} + \\sqrt {1 - \\bar {\\beta} _ {t}} \\boldsymbol {\\epsilon}, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            181,
+            215,
+            468,
+            242
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\bar{\\beta}_t = \\prod_{s=1}^t \\beta_s$  and  $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . The model then gradually removes noise by executing a learnable Markov chain in the reverse time direction, parameterized by a normal prior distribution  $\\mathbb{P}(\\mathbf{z}_T) = \\mathcal{N}(\\mathbf{z}_T;0,\\mathbf{I})$  and a learnable transition kernel  $\\mathbb{P}_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t)$  given by:",
+        "bbox": [
+            76,
+            250,
+            468,
+            327
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbb {P} _ {\\theta} \\left(\\mathbf {z} _ {t - 1} \\mid \\mathbf {z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {z} _ {t}, t\\right), \\Sigma_ {\\theta} \\left(\\mathbf {z} _ {t}, t\\right)\\right). \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            99,
+            334,
+            468,
+            351
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The goal of the training process is to approximately match the reverse Markov chain with the actual time reversal of the forward Markov chain. Mathematically, parameter  $\\theta$  is adjusted so that the joint distribution of the reverse Markov chain  $\\mathbb{P}_{\\theta}(\\mathbf{z}_0,\\mathbf{z}_1,\\dots ,\\mathbf{z}_T)\\coloneqq \\mathbb{P}(\\mathbf{z}_T)\\prod_{t = 1}^T\\mathbb{P}_{\\theta}(\\mathbf{z}_{t - 1}\\mid \\mathbf{z}_t)$  closely approximates that of the forward Markov chain  $\\mathbb{P}(\\mathbf{z}_0,\\mathbf{z}_1,\\dots ,\\mathbf{z}_T)\\coloneqq \\mathbb{P}(\\mathbf{z}_0)\\prod_{t = 1}^T\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{z}_{t - 1})$ . This is achieved by minimizing the following loss:",
+        "bbox": [
+            76,
+            359,
+            468,
+            479
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbb {E} _ {t \\sim \\mathcal {U} [ 1, T ], \\mathbf {z} _ {0} \\sim \\mathbb {P} (\\mathbf {z} _ {0}), \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t) \\| ^ {2} \\right], \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            102,
+            487,
+            468,
+            513
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{z}_t$  is computed using Eq. (2), and  $\\epsilon_{\\theta}$  is predicted using a neural network, typically a UNet architecture [37]. In a conditional diffusion model  $\\epsilon_{\\theta}(\\mathbf{z}_t,t)$  gets replaced by  $\\epsilon_{\\theta}(\\mathbf{z}_t,t,\\mathcal{C})$ , where  $\\mathcal{C}$  is a conditioning variable.",
+        "bbox": [
+            76,
+            518,
+            468,
+            580
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Our Architecture",
+        "text_level": 1,
+        "bbox": [
+            76,
+            587,
+            246,
+            602
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Image Encoder and Latent Diffusion. Diffusion models typically take large number of time steps to train, and are difficult to converge. Recently, [36] proposed a new diffusion with improved convergence properties and is called \"stable-diffusion\". The key idea is to perform the diffusion in latent space, with latent embedding learnt separately through a variational autoencoder (VAE). The Encoder of VAE first transforms the input image  $\\mathbf{x}$  of size  $(H,W)$  to latent space, then we follow latent diffusion formulation and utilize the UNet used in Stable Diffusion[36]. Utilizing latent diffusion formulation enables our architecture to capture multi-resolution features. Hence, we aggregate the feature maps from different layers of the UNet module (implementing conditional diffusion) by bringing them all to 1/4th resolution of the latent space, resulting in a feature map of size  $8e\\times H / 32\\times W / 32$ .",
+        "bbox": [
+            76,
+            611,
+            468,
+            851
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Exploiting Semantic Context with Conditional Diffusion. Recall that we formulated the single image depth estimation as a dense regression problem, predicting  $\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{x})$",
+        "bbox": [
+            76,
+            854,
+            468,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/c884c50e858aad2ed14417c86c14d50488599881dbd4715cd3e63339ec6572b7.jpg",
+        "image_caption": [
+            "(a)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            598,
+            88,
+            629,
+            215
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/57991546efa28fa9bd31fcb0cf70cb7c634b4e25f7be459feddfdac4e6a8ced9.jpg",
+        "image_caption": [
+            "(b)",
+            "Figure 3. (a) Probabilistic graphical model corresponding to VPD. (b) The same corresponding to our formulation. Here,  $\\mathcal{C}$  represents the semantic embedding derived from our CIDE module. This embedding is internally generated by passing  $\\mathbf{x}$  through the ViT, resulting in  $\\mathcal{E}$ . Subsequently,  $\\mathcal{E}$  undergoes further processing to yield  $\\mathcal{C}$ , which is then utilized in the conditional diffusion module implementing  $\\mathbb{P}(\\mathbf{z}_0\\mid \\mathbf{z}_t,\\mathcal{C})$ . The output of the conditional diffusion module is fed into the Depth Regressor module within our architecture, implementing  $\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{z}_0)$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            746,
+            88,
+            808,
+            215
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{y}$  and  $\\mathbf{x}$  denote output depth, and input image respectively. In our diffusion formulation, we predict  $\\mathbf{z}_0$  which is then used to predict pixel-wise depth,  $\\mathbf{y}$ . It has been shown that noise prediction in a diffusion model,  $\\epsilon_{\\theta}$ , can be seen as predicting the gradient of the density function,  $\\nabla_{\\mathbf{z}_t} \\log \\mathbb{P}(\\mathbf{z}_t)$ . Hence, overall architecture for SIDE using a diffusion architecture can be seen as factorizing the conditional probability, as shown in the probabilistic graph model in Fig. 3a. To utilize additional semantic context generated from a ViT model, we condition it on the ViT embeddings as shown in Fig. 3b. Mathematically, we model:",
+        "bbox": [
+            496,
+            392,
+            890,
+            559
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbb {P} (\\mathbf {y} \\mid \\mathbf {x}, \\mathcal {E}) = \\mathbb {P} (\\mathbf {y} \\mid \\mathbf {z} _ {0}) \\mathbb {P} (\\mathbf {z} _ {0} \\mid \\mathbf {z} _ {t}, \\mathcal {C}) \\mathbb {P} (\\mathbf {z} _ {t} \\mid \\mathbf {x}) \\mathbb {P} (\\mathcal {C} \\mid \\mathbf {x}),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            513,
+            587,
+            875,
+            604
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where",
+        "bbox": [
+            500,
+            608,
+            544,
+            619
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbb {P} (\\mathcal {C} \\mid \\mathbf {x}) = \\mathbb {P} (\\mathcal {C} \\mid \\mathcal {E}) \\mathbb {P} (\\mathcal {E} \\mid \\mathbf {x}). \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            532,
+            626,
+            890,
+            643
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Here, the first term,  $\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{z}_0)$ , is implemented through the Depth Regressor module explained earlier. Similarly,  $\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{x})$  is implemented using the VAE's Encoder as described earlier. We generate conditional information,  $\\mathcal{C}$  using our Comprehensive Image Detail Embedding module (hereafter CIDE, and explained below). The CIDE module takes  $\\mathbf{x}$  as input and generates embedding vector  $\\mathcal{C}$  (of dimension 768 in our design) as the output, thus, implementing  $\\mathbb{P}(\\mathcal{C}\\mid \\mathbf{x})$  given in Eq. (5). We use  $\\mathcal{E}$  to denote the embedding vector of a ViT module, and  $\\mathbb{P}(\\mathcal{E}\\mid \\mathbf{x})$  is implemented through the ViT.  $\\mathbb{P}(\\mathcal{C}\\mid \\mathcal{E})$  is implemented using downstream modules in CIDE consisting of learnable embeddings. The second term,  $\\mathbb{P}(\\mathbf{z}_0\\mid \\mathbf{z}_t,\\mathcal{C})$ , is implemented using conditional diffusion,",
+        "bbox": [
+            496,
+            656,
+            890,
+            866
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Comprehensive Image Detail Embedding (CIDE) Module. As described earlier, we believe using pseudo-captions",
+        "bbox": [
+            496,
+            869,
+            890,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "28288",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/85cecef06f97900f28fcbdaa57d7bdd2247ae7c37b92b6324ec104478fa24f85.jpg",
+        "table_caption": [
+            "Table 1. Results on Indoor NYU Depth v2 [27] Dataset. Results that are bold perform best.  $\\uparrow$  means the metric should be higher,  $\\downarrow$  indicate lower is better. The evaluation uses an upper bound of 10 meters on the ground truth depth map. All the numbers for other works have been taken from the corresponding papers. For MIM, and ZoEDepth we have used SwinV2-L 1K, and ZoeDepth-M12-N versions respectively. We see an overall improvement against SOTA on all the metrics used for evaluation."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Venue</td><td>Abs Rel↓</td><td>RMSE↓</td><td>log10↓</td><td>Sq Rel↓</td><td>δ1↑</td><td>δ2↑</td><td>δ3↑</td></tr><tr><td>Eigen et al.[9]</td><td>NIPS&#x27;14</td><td>0.158</td><td>0.641</td><td>-</td><td>-</td><td>0.769</td><td>0.950</td><td>0.988</td></tr><tr><td>DORN[10]</td><td>CVPR&#x27;18</td><td>0.115</td><td>0.509</td><td>0.051</td><td>-</td><td>0.828</td><td>0.965</td><td>0.992</td></tr><tr><td>SharpNet[32]</td><td>ICCV&#x27;19</td><td>0.139</td><td>0.502</td><td>0.047</td><td>-</td><td>0.836</td><td>0.966</td><td>0.993</td></tr><tr><td>Chen et al.[5]</td><td>IJCAI-19</td><td>0.111</td><td>0.514</td><td>0.048</td><td>-</td><td>0.878</td><td>0.977</td><td>0.994</td></tr><tr><td>BTS[22]</td><td>Arxiv&#x27;19</td><td>0.110</td><td>0.392</td><td>0.047</td><td>0.066</td><td>0.885</td><td>0.978</td><td>0.994</td></tr><tr><td>AdaBins[2]</td><td>CVPR&#x27;21</td><td>0.103</td><td>0.364</td><td>0.044</td><td>-</td><td>0.903</td><td>0.984</td><td>0.997</td></tr><tr><td>DPT[33]</td><td>ICCV&#x27;21</td><td>0.110</td><td>0.357</td><td>0.045</td><td>-</td><td>0.904</td><td>0.988</td><td>0.998</td></tr><tr><td>P3Depth[30]</td><td>CVPR&#x27;22</td><td>0.104</td><td>0.356</td><td>0.043</td><td>-</td><td>0.898</td><td>0.981</td><td>0.996</td></tr><tr><td>NeWCRFs[53]</td><td>CVPR&#x27;22</td><td>0.095</td><td>0.334</td><td>0.041</td><td>0.045</td><td>0.922</td><td>0.992</td><td>0.998</td></tr><tr><td>SwinV2-B[24]</td><td>CVPR&#x27;22</td><td>0.133</td><td>0.462</td><td>0.059</td><td>-</td><td>0.819</td><td>0.975</td><td>0.995</td></tr><tr><td>SwinV2-L[24]</td><td>CVPR&#x27;22</td><td>0.112</td><td>0.381</td><td>0.051</td><td>-</td><td>0.886</td><td>0.984</td><td>0.997</td></tr><tr><td>Localbins[3]</td><td>ECCV&#x27;22</td><td>0.099</td><td>0.357</td><td>0.042</td><td>-</td><td>0.907</td><td>0.987</td><td>0.998</td></tr><tr><td>Jun et al.[17]</td><td>ECCV&#x27;22</td><td>0.098</td><td>0.355</td><td>0.042</td><td>-</td><td>0.913</td><td>0.987</td><td>0.998</td></tr><tr><td>PixelFormer[1]</td><td>WACV&#x27;23</td><td>0.090</td><td>0.322</td><td>0.039</td><td>0.043</td><td>0.929</td><td>0.991</td><td>0.998</td></tr><tr><td>DDP[16]</td><td>ICCV&#x27;23</td><td>0.094</td><td>0.329</td><td>0.040</td><td>-</td><td>0.921</td><td>0.990</td><td>0.998</td></tr><tr><td>MIM [51]</td><td>CVPR&#x27;23</td><td>0.083</td><td>0.287</td><td>0.035</td><td>-</td><td>0.949</td><td>0.994</td><td>0.999</td></tr><tr><td>AiT[28]</td><td>ICCV&#x27;23</td><td>0.076</td><td>0.275</td><td>0.033</td><td>-</td><td>0.954</td><td>0.994</td><td>0.999</td></tr><tr><td>ZoeDepth [4]</td><td>Arxiv&#x27;23</td><td>0.075</td><td>0.270</td><td>0.032</td><td>0.030</td><td>0.955</td><td>0.995</td><td>0.999</td></tr><tr><td>VPD[54]</td><td>ICCV&#x27;23</td><td>0.069</td><td>0.254</td><td>0.030</td><td>0.027</td><td>0.964</td><td>0.995</td><td>0.999</td></tr><tr><td>Ours</td><td>CVPR&#x27;24</td><td>0.059</td><td>0.218</td><td>0.026</td><td>0.013</td><td>0.978</td><td>0.997</td><td>0.999</td></tr></table>",
+        "bbox": [
+            106,
+            155,
+            867,
+            496
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/a76cc74dd06135df0021484b881dda524bb0070b423a6e292f03cff9e997573e.jpg",
+        "image_caption": [
+            "Figure 4. Visual Comparison on NYU Depth v2 Indoor Dataset. Note, our method's ability to delineate objects in terms of their depth, such as the table lamp in Row 5, even when such information is absent from the ground truth depth map."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            158,
+            507,
+            812,
+            825
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "28289",
+        "bbox": [
+            478,
+            945,
+            521,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/7e3acb0e160e30d314f36226aad83e643f08cfb3b7d455b1bbed0204d4726d20.jpg",
+        "table_caption": [
+            "Table 2. Performance on the Outdoor KITTI [12] Dataset. Please refer to the caption of Tab. 1 for notation details. For ZoEDepth results, we use the ZoeDepth-M12-K version following the authors' recommendation. In instances where results for certain methods were not reported in the respective works, denoted by “-”, and the code is unavailable, we were unable to generate the missing numbers. Despite the saturation of results on the outdoor KITTI dataset, our method consistently achieves comparable or superior performance to the state-of-the-art (SOTA) across all metrics. VPD[54] cannot be trained on KITTI dataset as it required a per image text label which is not present in KITTI. The symbol  $\\dagger$  indicates that the method utilizes additional information beyond RGB."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Venue</td><td>Abs Rel↓</td><td>Sq Rel↓</td><td>RMSElog ↓</td><td>RMSE↓</td><td>δ1 ↑</td><td>δ2 ↑</td><td>δ3 ↑</td></tr><tr><td>Eigen et al.[9]</td><td>NIPS&#x27;14</td><td>0.203</td><td>1.517</td><td>0.282</td><td>6.307</td><td>0.702</td><td>0.898</td><td>0.967</td></tr><tr><td>DORN[10]</td><td>CVPR&#x27;18</td><td>0.072</td><td>0.307</td><td>0.120</td><td>2.727</td><td>0.932</td><td>0.984</td><td>0.994</td></tr><tr><td>BTS[22]</td><td>Arxiv&#x27;19</td><td>0.059</td><td>0.241</td><td>0.096</td><td>2.756</td><td>0.956</td><td>0.993</td><td>0.998</td></tr><tr><td>AdaBins[2]</td><td>CVPR&#x27;21</td><td>0.067</td><td>0.190</td><td>0.088</td><td>2.960</td><td>0.949</td><td>0.992</td><td>0.998</td></tr><tr><td>DPT[33]</td><td>ICCV&#x27;21</td><td>0.060</td><td>-</td><td>0.092</td><td>2.573</td><td>0.959</td><td>0.995</td><td>0.996</td></tr><tr><td>P3Depth[30]</td><td>CVPR&#x27;22</td><td>0.071</td><td>0.270</td><td>0.103</td><td>2.842</td><td>0.953</td><td>0.993</td><td>0.998</td></tr><tr><td>NeWCRFs[53]</td><td>CVPR&#x27;22</td><td>0.052</td><td>0.155</td><td>0.079</td><td>2.129</td><td>0.974</td><td>0.997</td><td>0.999</td></tr><tr><td>PixelFormer[1]</td><td>WACV&#x27;23</td><td>0.051</td><td>0.149</td><td>0.077</td><td>2.081</td><td>0.976</td><td>0.997</td><td>0.999</td></tr><tr><td>ZoeDepth [4]</td><td>Arxiv&#x27;23</td><td>0.054</td><td>0.189</td><td>0.083</td><td>2.440</td><td>0.97</td><td>0.996</td><td>0.999</td></tr><tr><td>DDP [16]</td><td>ICCV&#x27;23</td><td>0.050</td><td>0.148</td><td>0.076</td><td>2.072</td><td>0.975</td><td>0.997</td><td>0.999</td></tr><tr><td>URCDC [40]</td><td>ToM&#x27;23</td><td>0.050</td><td>0.142</td><td>0.076</td><td>2.032</td><td>0.977</td><td>0.997</td><td>0.999</td></tr><tr><td>IEBins [41]</td><td>NeurIPS&#x27;23</td><td>0.050</td><td>0.142</td><td>0.075</td><td>2.011</td><td>0.978</td><td>0.998</td><td>0.999</td></tr><tr><td>MIM [51]</td><td>CVPR&#x27;23</td><td>0.050</td><td>0.139</td><td>0.075</td><td>1.966</td><td>0.977</td><td>0.998</td><td>1.000</td></tr><tr><td>GEDepth[52]†</td><td>ICCV&#x27;23</td><td>0.048</td><td>0.142</td><td>0.076</td><td>2.044</td><td>0.976</td><td>0.997</td><td>0.999</td></tr><tr><td>Ours</td><td>CVPR&#x27;24</td><td>0.048</td><td>0.139</td><td>0.074</td><td>2.039</td><td>0.979</td><td>0.998</td><td>1.000</td></tr></table>",
+        "bbox": [
+            84,
+            183,
+            888,
+            450
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/fd5626577cb0a6d7c11ad79b7f023d80f1eaefeb0fdc9d0368fc06a025dcee80.jpg",
+        "image_caption": [
+            "Figure 5. Visual Comparison on KITTI Outdoor Dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            460,
+            890,
+            705
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "to generate the semantic context has limited utility, as the textual descriptions typically focus on large salient objects only. Instead we propose our CIDE module which use embeddings from a pre-trained ViT, and extract detailed semantic context from these embeddings. For this we take 1000 dimensional logit vector from ViT, and pass it through a two layer MLP which converts it to a 100 dimensional vector. Subsequently, we employ this vector to compute the linear combination of 100 learnable embeddings.",
+        "bbox": [
+            75,
+            756,
+            472,
+            891
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "This resulting embedding undergoes a linear transformation to yield a semantic context vector of dimension 768, which is then passed to conditional diffusion module.",
+        "bbox": [
+            496,
+            756,
+            890,
+            800
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Depth Regressor. The output feature map undergoes through an Upsampling Decoder, comprised of deconvolution layers, followed by a Depth Regressor. The Depth Regressor is essentially a two-layer convolutional neural network (CNN), with the initial layer having dimensions  $\\mathrm{Conv}(3\\times 3)$ , 192, and the subsequent layer  $\\mathrm{Conv}(3\\times 3)$ , 1.",
+        "bbox": [
+            496,
+            809,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "28290",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/2e0b5f3c63f0cc8f262938f75cd81372b094e58cf9940af92c8e9463c4040b78.jpg",
+        "table_caption": [
+            "Table 3. Quantitative results for zero-shot transfer to four unseen indoor datasets. mRI denotes the mean relative improvement with respect to NeWCRFs across all metrics (δ1, REL, RMSE). Evaluation depth is capped at 8m for SUN RGB-D, 10m for iBims and DIODE Indoor, and 80m for HyperSim. Best results are in bold, second best are underlined. Our mRI outperforms all methods across all datasets by a large margin. † denotes that ZoeD is trained on 12 datasets and our method is trained only on NYUv2."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"4\">SUN RGB-D</td><td colspan=\"4\">iBims-1 Benchmark</td><td colspan=\"4\">DIODE Indoor</td><td colspan=\"4\">HyperSim</td></tr><tr><td>δ1↑</td><td>REL↓</td><td>RMSE↓</td><td>mRIθ↑</td><td>δ1↑</td><td>REL↓</td><td>RMSE↓</td><td>mRIθ↑</td><td>δ1↑</td><td>REL↓</td><td>RMSE↓</td><td>mRIθ↑</td><td>δ1↑</td><td>REL↓</td><td>RMSE↓</td><td>mRIθ↑</td></tr><tr><td>BTS [22]</td><td>0.740</td><td>0.172</td><td>0.515</td><td>-14.2%</td><td>0.538</td><td>0.231</td><td>0.919</td><td>-6.9%</td><td>0.210</td><td>0.418</td><td>1.905</td><td>2.3%</td><td>0.225</td><td>0.476</td><td>6.404</td><td>-8.6%</td></tr><tr><td>AdaBins [2]</td><td>0.771</td><td>0.159</td><td>0.476</td><td>-7.0%</td><td>0.555</td><td>0.212</td><td>0.901</td><td>-2.1%</td><td>0.174</td><td>0.443</td><td>1.963</td><td>-7.2%</td><td>0.221</td><td>0.483</td><td>6.546</td><td>-10.5%</td></tr><tr><td>LocalBins [3]</td><td>0.777</td><td>0.156</td><td>0.470</td><td>-5.6%</td><td>0.558</td><td>0.211</td><td>0.880</td><td>-0.7%</td><td>0.229</td><td>0.412</td><td>1.853</td><td>7.1%</td><td>0.234</td><td>0.468</td><td>6.362</td><td>-6.6%</td></tr><tr><td>NeWCRFs [53]</td><td>0.798</td><td>0.151</td><td>0.424</td><td>0.0%</td><td>0.548</td><td>0.206</td><td>0.861</td><td>0.0%</td><td>0.187</td><td>0.404</td><td>1.867</td><td>0.0%</td><td>0.255</td><td>0.442</td><td>6.017</td><td>0.0%</td></tr><tr><td>VPD [54]</td><td>0.861</td><td>0.121</td><td>0.355</td><td>14.7%</td><td>0.627</td><td>0.187</td><td>0.767</td><td>11.5%</td><td>0.480</td><td>0.392</td><td>1.295</td><td>63.4%</td><td>0.333</td><td>0.531</td><td>5.111</td><td>8.5%</td></tr><tr><td>ZoeD-M12-N† [4]</td><td>0.864</td><td>0.119</td><td>0.346</td><td>16.0%</td><td>0.658</td><td>0.169</td><td>0.711</td><td>18.5%</td><td>0.376</td><td>0.327</td><td>1.588</td><td>45.0%</td><td>0.292</td><td>0.410</td><td>5.771</td><td>8.6%</td></tr><tr><td>Ours</td><td>0.885</td><td>0.112</td><td>0.319</td><td>20.5%</td><td>0.688</td><td>0.163</td><td>0.664</td><td>23.1%</td><td>0.545</td><td>0.344</td><td>1.164</td><td>81.3%</td><td>0.394</td><td>0.442</td><td>4.739</td><td>25.2%</td></tr></table>",
+        "bbox": [
+            76,
+            157,
+            913,
+            287
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/10a7e4cdf5394b729bfc2c2fda0f2f9fb4531171367e28fe6cc3a38ac46faaed.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            135,
+            305,
+            846,
+            460
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f31058aa86b9b19c5376f9c5dabda1e36277891d933e7679fad2e97499e3bd57.jpg",
+        "image_caption": [
+            "Figure 6. Visualization of improvements over VPD [54] in our model due to ViT embeddings passed down as conditional vectors (in blue and green). In the above images, ViT detects the desktop computer (first image) and table lamp (second image) with high probability, and they are thus better detected by our model. Additional visualizations are provided in the supplementary material."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            135,
+            478,
+            844,
+            638
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments and Results",
+        "text_level": 1,
+        "bbox": [
+            76,
+            718,
+            310,
+            734
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Datasets and Evaluation. We use NYU Depth v2 [27] and KITTI [12] as the primary datasets for training. The NYU Depth v2 dataset is a widely used indoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images in the train set and 654 in the test set. The dataset covers a wide range of indoor scenes and includes challenging scenarios such as reflective surfaces, transparent objects, and occlusions. The ground truth depth maps are obtained using a structured light sensor and are provided at a resolution of  $640 \\times 480$ .",
+        "bbox": [
+            75,
+            750,
+            470,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "KITTI dataset is a widely used outdoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images. The dataset covers outdoor driving scenarios and includes varying lighting conditions, weather, and occlusions. The ground truth depth maps are obtained using a Velodyne LiDAR sensor and are provided at a resolution of  $1242 \\times 375$ . To demonstrate generalizability, we evaluate zero-shot performance on the following datasets: Sun-RGBD [42], iBims1 [19], DIODE [49], and HyperSim [35]. Whereas, the main paper contains mostly quantitative, and only some representative",
+        "bbox": [
+            496,
+            719,
+            890,
+            887
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "28291",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/5fdba6a9d50ed2fb62d14f37e682b07bbb0186f63efe75d3ebcd71721d942fd5.jpg",
+        "table_caption": [
+            "Table 4. Effectiveness of different embeddings for guiding the diffusion process for depth estimation on NYU Depth v2 dataset. In the third row, rather than using pseudo caption embeddings generated from the scene label (as implemented by VPD [54]), we provide the one-hot vector representing the scene label as a condition to the diffusion model. We observe a slight improvement in the metrics, highlighting that information content in the caption is similar to that in one-hot label."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Embeddings</td><td>RMSE↓</td><td>Abs Rel↓</td><td>δ1↑</td></tr><tr><td>Scene label emb. [54]</td><td>0.254</td><td>0.069</td><td>0.964</td></tr><tr><td>Text caption emb. [20]</td><td>0.225</td><td>0.062</td><td>0.976</td></tr><tr><td>One-hot vector emb.</td><td>0.244</td><td>0.067</td><td>0.968</td></tr><tr><td>Proposed scene emb.</td><td>0.218</td><td>0.059</td><td>0.978</td></tr></table>",
+        "bbox": [
+            83,
+            210,
+            464,
+            304
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tive visual results, detailed visual results on each dataset are included in the supplementary.",
+        "bbox": [
+            76,
+            329,
+            468,
+            359
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. Our model is implemented using PyTorch [29]. For optimization, we have used AdamW optimizer [25] with  $\\beta_0$  values of 0.9 and 0.999, a batch size of 32, and a weight decay of 0.1. We train our model for 25 epochs for both KITTI and NYU Depth v2 datasets, with an initial learning rate of  $3 \\times 10^{-5}$ . We first linearly increase learning rate to  $5 \\times 10^{-4}$ , and then linearly decrease across training iterations. We use usual data augmentation techniques, including random hue addition, horizontal flipping, changing the image brightness, and Cut Depth [15]. Our model takes approximately 21 minutes per epoch to train using 8 NVIDIA A100 GPUs.",
+        "bbox": [
+            75,
+            362,
+            468,
+            544
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.1. Comparison on Benchmark Datasets",
+        "text_level": 1,
+        "bbox": [
+            76,
+            554,
+            395,
+            569
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Comparison on NYU Depth v2. Tab. 1 shows the comparison of our proposed method with SOTA methods on the indoor NYU Depth v2 dataset [27]. We achieve a new state of the art on this dataset. Our methods perform better than the previous SOTA ([54])) by a large margin of  $14\\%$  in terms of RMSE. Fig. 4 provides a qualitative comparison on the dataset.",
+        "bbox": [
+            75,
+            577,
+            468,
+            681
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Comparison on KITTI. Tab. 2 shows the comparison with various methods on the KITTI dataset [12]. Fig. 5 shows the qualitative results. Unlike NYU Depth v2 dataset, KITTI is an outdoor dataset. On this dataset also, we achieve similar or better performance than all existing state-of-the-art techniques.",
+        "bbox": [
+            75,
+            685,
+            468,
+            777
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.2. Generalization and Zero Shot Transfer",
+        "text_level": 1,
+        "bbox": [
+            76,
+            786,
+            411,
+            801
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Unlike state of the art on zero short transfer (ZoEDepth [4]), which requiring training on many datasets (12 in their case) for effective zero shot transfer, we show that our model generalizes well to other unseen dataset even when trained on a single NYU Depth v2 dataset. Tab. 3 shows quantitative results to back our findings.",
+        "bbox": [
+            75,
+            809,
+            468,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.3. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            500,
+            90,
+            653,
+            107
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of Contextual Information. As highlighted in the motivation, a key observation of this study is the richness of information contained within the output probability vector of the ViT, surpassing the textual embeddings employed in current state-of-the-art (e.g., VPD [54]). To compare with the utilization of scene label information (as implemented by VPD), we construct the conditioning embedding as a one-hot vector using the scene label, and subsequently transform it using an MLP. As illustrated in Tab. 4, we observe a slight improvement over VPD, indicating that the information content in pseudo-captions resembles that of one-hot labels. Furthermore, our proposed method surpasses both the approaches.",
+        "bbox": [
+            496,
+            113,
+            890,
+            311
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Qualitative Results. Perhaps the most important task would be to verify that use of rich probability vector instead of text embeddings actually results in an improvement in depth as a direct consequence. We do this by considering the top few objects predicted by the ViT and correlating this with the depth predicted at these objects. We hypothesise when a particular class say dog is predicted with a high probability (hence there must be a dog in the image), then the corresponding depth must also be more accurate. We show this in Fig. 6.",
+        "bbox": [
+            496,
+            311,
+            890,
+            465
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            472,
+            617,
+            487
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We presented a new architecture block Comprehensive Image Detail Embedding (CIDE) module for robust monocular depth estimation in this paper. Our key idea is to highlight the limitations of using pseudo-captions to provide contextual information, and instead propose to use richer class-wise probability generated by a classification model, such as ViT. The motivation is that, while textual embedding typically highlights salient objects, class-wise probability vector preserves more details, including smaller objects in the background also. We implement the idea using proposed CIDE module cascaded with a conditional diffusion pipeline for monocular depth estimation. We demonstrated the effectiveness of our approach on several benchmark datasets and showed that it outperforms SOTA methods by a significant margin.",
+        "bbox": [
+            496,
+            497,
+            890,
+            724
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement. We thank Kartik Anand for assistance with some of the experiments. This work has been supported by funding through Staqu Technologies Pvt. Ltd., and Department of Science and Technology (DST), Government of India. The travel for Suraj Patni is supported through Research Acceleration funds of Department of Computer Science and Engineering, IIT Delhi.",
+        "bbox": [
+            496,
+            727,
+            890,
+            834
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            500,
+            847,
+            596,
+            863
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "ref_text",
+        "text": "[1] Ashutosh Agarwal and Chetan Arora. Attention attention everywhere: Monocular depth prediction with skip attention.",
+        "bbox": [
+            506,
+            872,
+            890,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "28292",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5861-5870, 2023. 2, 5, 6",
+            "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4009-4018, 2021. 5, 6, 7",
+            "[3] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 5, 7",
+            "[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 2, 3, 5, 6, 7, 8",
+            "[5] Xiaotian Chen, Xuejin Chen, and Zheng-Jun Zha. Structure-aware residual pyramid network for monocular depth estimation. arXiv preprint arXiv:1907.06023, 2019. 5",
+            "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3",
+            "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3",
+            "[8] Yiqun Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation, 2023. 2, 3",
+            "[9] 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. 2, 5, 6",
+            "[10] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 5, 6",
+            "[11] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 2",
+            "[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013. 3, 6, 7, 8",
+            "[13] José L. Herrera, Carlos R. del Blanco, and Narciso García. Automatic depth extraction from 2d images using a cluster-based learning framework. IEEE Transactions on Image Processing, 27(7):3288-3299, 2018. 2",
+            "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[15] Yasunori Ishii and Takayoshi Yamashita. Cutdepth: Edge-aware data augmentation in depth estimation. arXiv preprint arXiv:2107.07684, 2021. 8",
+            "[16] Yuanfeng Ji, Zhe Chen, Enze Xie, Lanqing Hong, Xihui Liu, Zhaoqiang Liu, Tong Lu, Zhenguo Li, and Ping Luo. Ddp: Diffusion model for dense visual prediction. arXiv preprint arXiv:2303.17559, 2023. 2, 3, 5, 6",
+            "[17] Jinyoung Jun, Jae-Han Lee, Chul Lee, and Chang-Su Kim. Depth map decomposition for monocular depth estimation. In European Conference on Computer Vision, pages 18-34. Springer, 2022. 5",
+            "[18] Kevin Karsch, Ce Liu, and Sing Bing Kang. Depth transfer: Depth extraction from video using non-parametric sampling. IEEE transactions on pattern analysis and machine intelligence, 36(11):2144-2158, 2014. 2",
+            "[19] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 7",
+            "[20] Neehar Kondapaneni, Markus Marks, Manuel Knott, Rogério Guimarães, and Pietro Perona. Text-image alignment for diffusion-based perception. arXiv preprint arXiv:2310.00031, 2023. 2, 3, 8",
+            "[21] 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. 2",
+            "[22] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 5, 6, 7",
+            "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2",
+            "[24] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 5",
+            "[25] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICML 2019. 8",
+            "[26] John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J Davison. Scelenet rgb-d: 5m photorealistic images of synthetic indoor trajectories with ground truth. arXiv preprint arXiv:1612.05079, 2016. 3",
+            "[27] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 3, 5, 7, 8",
+            "[28] Jia Ning, Chen Li, Zheng Zhang, Chunyu Wang, Zigang Geng, Qi Dai, Kun He, and Han Hu. All in tokens: Unifying output space of visual tasks via soft token. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19900-19910, 2023. 2, 5"
+        ],
+        "bbox": [
+            501,
+            92,
+            892,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "28293",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[29] 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. Curran Associates, Inc., 2019. 8",
+            "[30] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1610–1621, 2022. 5, 6",
+            "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2",
+            "[32] Michael Ramamonjisoa and Vincent Lepetit. Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 5",
+            "[33] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 5, 6",
+            "[34] 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",
+            "[35] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021. 1, 7",
+            "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4",
+            "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3, 4",
+            "[38] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 2, 3",
+            "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m:"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 3",
+            "[40] Shuwei Shao, Zhongcai Pei, Weihai Chen, Ran Li, Zhong Liu, and Zhengguo Li. Urcdc-depth: Uncertainty rectified cross-distillation with cutflip for monocular depth estimation. arXiv preprint arXiv:2302.08149, 2023. 6",
+            "[41] Shuwei Shao, Zhongcai Pei, Xingming Wu, Zhong Liu, Weihai Chen, and Zhengguo Li. Iebins: Iterative elastic bins for monocular depth estimation. arXiv preprint arXiv:2309.14137, 2023. 6",
+            "[42] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 567-576, 2015. 1, 7",
+            "[43] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3",
+            "[44] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 3",
+            "[45] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3",
+            "[46] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. Advances in Neural Information Processing Systems, 34: 1415-1428, 2021. 3",
+            "[47] M. Sun, A. Y. Ng, and A. Saxena. Make3d: Learning 3d scene structure from a single still image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(05):824-840, 2009. 2",
+            "[48] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3",
+            "[49] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 1, 7",
+            "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3",
+            "[51] Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14475-14485, 2023. 2, 5, 6",
+            "[52] Xiaodong Yang, Zhuang Ma, Zhiyu Ji, and Zhe Ren. Gedgepth: Ground embedding for monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12719-12727, 2023. 6",
+            "[53] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Newcrfs: Neural window fully-connected crfs for"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "28294",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "monocular depth estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 5, 6, 7",
+            "[54] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. ICCV, 2023. 2, 3, 5, 6, 7, 8"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            186
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "28295",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_model.json

@@ -0,0 +1,2015 @@
+[
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.107,
+                0.003,
+                0.182,
+                0.043
+            ],
+            "angle": 0,
+            "content": "CVF"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.237,
+                0.001,
+                0.808,
+                0.047
+            ],
+            "angle": 0,
+            "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.209,
+                0.131,
+                0.761,
+                0.176
+            ],
+            "angle": 0,
+            "content": "ECoDepth: Effective Conditioning of Diffusion Models for Monocular Depth Estimation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.223,
+                0.204,
+                0.321,
+                0.222
+            ],
+            "angle": 0,
+            "content": "Suraj Patni*"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.409,
+                0.205,
+                0.558,
+                0.222
+            ],
+            "angle": 0,
+            "content": "Aradhye Agarwal*"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.636,
+                0.205,
+                0.746,
+                0.22
+            ],
+            "angle": 0,
+            "content": "Chetan Arora"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.34,
+                0.223,
+                0.63,
+                0.24
+            ],
+            "angle": 0,
+            "content": "Indian Institute of Technology Delhi"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.299,
+                0.241,
+                0.667,
+                0.257
+            ],
+            "angle": 0,
+            "content": "https://ecodepth-iitd.github.io"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.235,
+                0.292,
+                0.313,
+                0.307
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.324,
+                0.474,
+                0.749
+            ],
+            "angle": 0,
+            "content": "In the absence of parallax cues, a learning based single image depth estimation (SIDE) model relies heavily on shading and contextual cues in the image. While this simplicity is attractive, it is necessary to train such models on large and varied datasets, which are difficult to capture. It has been shown that using embeddings from pretrained foundational models, such as CLIP, improves zero shot transfer in several applications. Taking inspiration from this, in our paper we explore the use of global image priors generated from a pre-trained ViT model to provide more detailed contextual information. We argue that the embedding vector from a ViT model, pre-trained on a large dataset, captures greater relevant information for SIDE than the usual route of generating pseudo image captions, followed by CLIP based text embeddings. Based on this idea, we propose a new SIDE model using a diffusion backbone which is conditioned on ViT embeddings. Our proposed design establishes a new state-of-the-art (SOTA) for SIDE on NYU Depth v2 dataset, achieving Abs Rel error of 0.059(14% improvement) compared to 0.069 by the current SOTA (VPD). And on KITTI dataset, achieving Sq Rel error of 0.139 (2% improvement) compared to 0.142 by the current SOTA (GED). For zero shot transfer with a model trained on NYU Depth v2, we report mean relative improvement of (20%, 23%, 81%, 25%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, compared to (16%, 18%, 45%, 9%) by ZoEDepth. The code is available in our project page."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.776,
+                0.208,
+                0.791
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.802,
+                0.47,
+                0.879
+            ],
+            "angle": 0,
+            "content": "Single Image Depth Estimation (SIDE) is the task of predicting per pixel depth using a single RGB image from a monocular camera. It is a fundamental problem in computer vision with applications in several domains, viz robotics, autonomous driving, and augmented reality. The problem is"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.501,
+                0.291,
+                0.697,
+                0.331
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.698,
+                0.291,
+                0.891,
+                0.331
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.65,
+                0.334,
+                0.745,
+                0.344
+            ],
+            "angle": 0,
+            "content": "(a) Sun-RGBD [42]"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.502,
+                0.345,
+                0.696,
+                0.385
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.698,
+                0.345,
+                0.892,
+                0.385
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.656,
+                0.388,
+                0.737,
+                0.399
+            ],
+            "angle": 0,
+            "content": "(b) iBims1 [19]"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.502,
+                0.399,
+                0.697,
+                0.44
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.698,
+                0.399,
+                0.891,
+                0.44
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.659,
+                0.442,
+                0.734,
+                0.453
+            ],
+            "angle": 0,
+            "content": "(c) DIODE [49]"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.502,
+                0.453,
+                0.697,
+                0.494
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.698,
+                0.453,
+                0.891,
+                0.494
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.649,
+                0.496,
+                0.745,
+                0.508
+            ],
+            "angle": 0,
+            "content": "(d) HyperSim [35]"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.498,
+                0.52,
+                0.892,
+                0.617
+            ],
+            "angle": 0,
+            "content": "Figure 1. Qualitative results across four different datasets, demonstrating the zero-shot performance of our model trained only on the NYU Depth v2 dataset. Corresponding quantitative results are presented in Table 3. The first column displays RGB images, the second column depicts ground truth depth, and the third column showcases our model's predicted depths. Additional images for each dataset are available in the Supplementary Material."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.643,
+                0.892,
+                0.735
+            ],
+            "angle": 0,
+            "content": "typically formulated in two flavors: metric depth estimation (MDE), and relative depth estimation (RDE). As the names suggest, MDE deals with estimation in physical units such as meters, whereas, RDE techniques focus on relative depth only, and require a per-image affine transformation as post-processing to convert to a physical unit."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.735,
+                0.892,
+                0.886
+            ],
+            "angle": 0,
+            "content": "Conventional geometric techniques for depth estimation typically rely on feature correspondence, parallax, and triangulation from two or more views. However, the problem becomes ill-posed for estimation from a single view. Intuitively, depth map is a 3D representation of the scene, whereas an RGB image is a 2D projection of the scene. Hence, it cannot be uniquely determined from a single RGB image. Therefore, learning-based SIDE models rely on visual cues like 'shape from shading' and other contextual priors for per-pixel depth prediction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.519,
+                0.886,
+                0.892,
+                0.902
+            ],
+            "angle": 0,
+            "content": "A data-driven approach makes the SIDE pipeline sim"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.096,
+                0.887,
+                0.206,
+                0.9
+            ],
+            "angle": 0,
+            "content": "*Equal contribution."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "28285"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.092,
+                0.471,
+                0.198
+            ],
+            "angle": 0,
+            "content": "pler, but also makes the learnt model dependent on the quality of training data. It has been observed that such models overfit on a particular training distribution/domain and fail to generalize on unseen data. This is especially true for MDE models when the range of depth in the training dataset is limited. Hence, training on multiple datasets, with wide variations in depth ranges has been proposed [4, 34]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.202,
+                0.474,
+                0.549
+            ],
+            "angle": 0,
+            "content": "On the other hand, development of large foundational models (LFMs) in recent years has altered the preferred design approach for many computer vision problems. These huge models are trained using extensive datasets of unlabeled images and learning objectives that are agnostic to specific tasks. The learnt embeddings from such pre-trained models have been shown to help generalization and zero-shot transfer in many applications. We are aware of at least two works for MDE in SIDE problem that have appeared in the last few months and make use of such foundational models. VPD [54] uses a text-to-image diffusion model pretrained on LAION-400M [39] dataset having large-scale image-text pairs as the backbone. The model prompts the denoising UNet with textual inputs to make the visual contents interact with the text prompts. Since the problem formulation doesn't include text description as the input, the model generates simple descriptions such as 'A photo of a {scene name}' based on the scene label given in the NYU Depth v2 dataset. Another work, TADP [20] improves upon VPD. Instead of simple image descriptions, TADP uses BLIP-2 [23] to generate image captions, and then uses CLIP [31] embeddings of the pseudo-caption to condition the diffusion model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.554,
+                0.473,
+                0.901
+            ],
+            "angle": 0,
+            "content": "We view the above works as providing robust semantic context to LFMs for the actual task at hand, which helps in visual recognition in general, as well as SIDE. While we do agree with the broad motivations of these works, the question that we ask is if pseudo-captions are the most effective way to provide the semantic context. Textual descriptions of an image typically focus on large salient objects and emphasize on their relationships. On the other hand, large vision models for image classification, typically contain representation for even smaller objects present in the scene. Even when a single object is present in the scene, the representations typically capture uncertainties and ambiguities inherent in the scene. Hence, we posit that using embeddings from a transformer model, pretrained on a large image dataset unrelated to the SIDE task, captures more relevant information, and is a better alternative to using pseudocaptions' CLIP embedding. Since diffusion-based models have shown their superiority for the dense prediction tasks in recent works [8, 16, 20, 38, 54]. Hence, we propose a diffusion backbone for our model, along with a novel CIDE module which employs ViT [7] to extract semantic context embeddings. These embeddings are subsequently utilized to condition the diffusion backbone."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.092,
+                0.866,
+                0.106
+            ],
+            "angle": 0,
+            "content": "Contributions. The key contributions of this work are:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.501,
+                0.107,
+                0.894,
+                0.318
+            ],
+            "angle": 0,
+            "content": "(1) We propose a new model for MDE in a SIDE task. The proposed model uses a conditional diffusion architecture with the semantic context being supplied through embeddings generated using a ViT model. Achieving a new state-of-the-art (SOTA) performance, our method outperforms existing approaches on benchmark datasets, including the NYU Depth v2 indoor and KITTI outdoor datasets. Notably, we report a significant improvement of \\(14\\%\\) in absolute relative error, achieving 0.059 compared to the current SOTA (VPD) performance of 0.069 on NYU Depth v2. And we report an improvement of \\(2\\%\\) in square relative error, achieving 0.139 compared to the current SOTA (GED) performance of 0.142 on KITTI."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.501,
+                0.318,
+                0.894,
+                0.438
+            ],
+            "angle": 0,
+            "content": "(2) We show, qualitatively as well as quantitatively, that using ViT embeddings to provide semantic context is a better alternative to generating pseudo captions and then using its CLIP embeddings to condition a SIDE model. In contrast to TADP [20], which uses pseudo captions, but only achieves a RMSE of 0.225 on NYU Depth v2, we report a lower, and a new SOTA, error of 0.218 (VPD [54] achieves 0.254)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.501,
+                0.439,
+                0.894,
+                0.575
+            ],
+            "angle": 0,
+            "content": "(3) We show that providing ViT conditioning, helps our model perform better in a zero shot transfer task. ZoEDepth [4], the current SOTA for zero-shot transfer, reports an improvement of (16%, 18%, 45%, 9%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, after training their model on 12 other datasets and NYU Depth v2. In contrast, we only train on NYU Depth v2, and report a much larger improvement of (21%, 23%, 81%, 25%)."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.501,
+                0.107,
+                0.894,
+                0.575
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.587,
+                0.642,
+                0.602
+            ],
+            "angle": 0,
+            "content": "2. Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.612,
+                0.892,
+                0.673
+            ],
+            "angle": 0,
+            "content": "Traditional Methods. Earlier techniques for SIDE have used Markov Random Fields [47], non-parametric depth sampling [18], and structural similarity with prior depth map [13] to predict pixel-wise depth."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.675,
+                0.895,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Deep Learning Techniques for SIDE. Modern techniques have approached the problem as a dense regression problem. CNNs have been the dominant architecture for the SIDE in the last decade, with global-local network stack [9], and multi-scale [21], or encoder-decoder architecture [11] as some of the popular solution strategies. Recently, PixelFormer [1] used a transformer-based encoder-decoder architecture with skip connections from encoders to decoders. MIM [51] proposed masked image modeling as a general-purpose pre-training for geometric and motion tasks such as SIDE and pose estimation. Similarly, AiT [28] used mask augmentation and proposed soft tokens to generalize visual prediction tasks. While earlier works either focused on MDE for a specific dataset or RDE for generalization on multiple datasets, ZoEDepth [4] has proposed"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.521,
+                0.957
+            ],
+            "angle": 0,
+            "content": "28286"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.081,
+                0.093,
+                0.882,
+                0.301
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.315,
+                0.893,
+                0.384
+            ],
+            "angle": 0,
+            "content": "Figure 2. An overview of our proposed model: The latent representation of the input image undergoes a diffusion process, which is conditioned by our proposed CIDe module. Within the CIDe module, the input image is fed through the frozen ViT model. From this, a linear combination of the learnt embeddings is computed, which is transformed to generate a 768-dimensional contextual embedding. This embedding is utilized to condition the diffusion backbone. Subsequently, hierarchical feature maps are extracted from the UNet's decoder which are concatenated and processed through a depth regressor to generate the depth map."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.411,
+                0.468,
+                0.501
+            ],
+            "angle": 0,
+            "content": "a generalized method for MDE that performs well in zero-shot transfer. We outperform [4] by a large margin even when training on a single dataset - NYU Depth v2 for indoor scenes or KITTI for outdoor scenes. Whereas, [4] uses 12 datasets for pre-training and then fine-tunes on NYU Depth v2 or KITTI for zero shot transfer."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.505,
+                0.468,
+                0.806
+            ],
+            "angle": 0,
+            "content": "Diffusion-based Methods with Pretraining on Large Datasets. Recently, many techniques for SIDE has used diffusion architectures. These techniques [8, 16, 20, 38, 54], exploit prior knowledge acquired by pretraining on large datasets like LAION-400M [39], which consists of 400 million image-text pairs. In contrast, depth datasets, such as NYU Depth v2 and KITTI, contain around 20-30 thousand image-depth pairs. DepthGen [38] and DDP [16] work on a noise-to-depthmap paradigm and use images for conditional guidance of the diffusion process. DepthGen employs self-supervised pretraining on tasks like colorization, inpainting, and JPEG artifact removal, followed by supervised training on indoor and outdoor datasets [6, 12, 26, 27, 48]. DDP [16] decouples the image encoder and map decoder, allowing the image encoder to run just once, while the lightweight map decoder is run multiple times. VPD [54] and TADP [20] use denoising UNet [37] as a backbone to extract the rich features at multiple scales. Also, they utilize text instead of image for conditioning the diffusion backbone."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.81,
+                0.468,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Vision Transformer for Scene Understanding. The transformer architecture was initially proposed for NLP tasks [50], but introduced to the computer vision community as Vision Transformer (ViT) in [7]. Prior to this, CNNs dominated computer vision problems due to their ability to capture spatial hierarchies in image data. However, such archi"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.411,
+                0.892,
+                0.531
+            ],
+            "angle": 0,
+            "content": "tectures were constrained due to their ability to learn spatially localized features only. Transformer architecture has a weaker inductive bias and allows ViT to learn long-range dependencies, and robust, generalizable features. ViT architectures have replaced CNNs for SOTA performance on most computer vision tasks in recent years. We use a pretrained ViT model for providing semantic information to the diffusion backbone in our model."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.562,
+                0.719,
+                0.58
+            ],
+            "angle": 0,
+            "content": "3. Proposed Methodology"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.592,
+                0.642,
+                0.607
+            ],
+            "angle": 0,
+            "content": "3.1. Preliminaries"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.621,
+                0.892,
+                0.698
+            ],
+            "angle": 0,
+            "content": "Problem Formulation. The objective of single image depth prediction task is to predict continuous values, denoted as \\(\\mathbf{y} \\in \\mathbb{R}^{H \\times W}\\), for every pixel present in the input RGB image, \\(\\mathbf{x} \\in [0,255]^{3 \\times H \\times W}\\). Here \\(H\\) and \\(W\\) represent the height and width respectively of the input image."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.704,
+                0.893,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Diffusion Model. Diffusion models are a class of generative models that progressively inject noise into the input data (forward pass) and then learn to reconstruct the original data in a reverse denoising process (reverse pass). There are three formulations of diffusion models: denoising diffusion probabilistic models (DDPMs) [14], score-based generative models [43, 44], and those based on stochastic differential equations [45, 46]. DDPMs are of relevance to our paper, and are described below. The architecture of DDPMs consists of two Markov chains: a forward chain that adds noise to the data, and a reverse chain that converts noise back to data by learning transition kernels parameterized by deep neural networks. Formally, the forward pass is modeled as"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.957
+            ],
+            "angle": 0,
+            "content": "28287"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.092,
+                0.204,
+                0.106
+            ],
+            "angle": 0,
+            "content": "a Markov process:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.123,
+                0.114,
+                0.47,
+                0.141
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbb {P} \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}\\right). \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.148,
+                0.47,
+                0.209
+            ],
+            "angle": 0,
+            "content": "Here \\(\\mathbf{z}_t\\) denotes the random variable at the \\(t^{\\mathrm{th}}\\) time step, \\(\\mathcal{N}(\\mathbf{z};\\boldsymbol {\\mu},\\boldsymbol {\\sigma})\\) denotes Gaussian probability distribution, and \\(\\beta_{t}\\) is the noise schedule. The above equation leads to the analytic form of \\(\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{z}_0),\\forall t\\in \\{0,1,\\dots ,T\\}\\):"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.182,
+                0.217,
+                0.47,
+                0.243
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbf {z} _ {t} = \\sqrt {\\bar {\\beta} _ {t}} \\mathbf {z} _ {0} + \\sqrt {1 - \\bar {\\beta} _ {t}} \\boldsymbol {\\epsilon}, \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.25,
+                0.469,
+                0.328
+            ],
+            "angle": 0,
+            "content": "where \\(\\bar{\\beta}_t = \\prod_{s=1}^t \\beta_s\\) and \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). The model then gradually removes noise by executing a learnable Markov chain in the reverse time direction, parameterized by a normal prior distribution \\(\\mathbb{P}(\\mathbf{z}_T) = \\mathcal{N}(\\mathbf{z}_T;0,\\mathbf{I})\\) and a learnable transition kernel \\(\\mathbb{P}_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t)\\) given by:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.1,
+                0.335,
+                0.469,
+                0.352
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbb {P} _ {\\theta} \\left(\\mathbf {z} _ {t - 1} \\mid \\mathbf {z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {z} _ {t}, t\\right), \\Sigma_ {\\theta} \\left(\\mathbf {z} _ {t}, t\\right)\\right). \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.36,
+                0.469,
+                0.481
+            ],
+            "angle": 0,
+            "content": "The goal of the training process is to approximately match the reverse Markov chain with the actual time reversal of the forward Markov chain. Mathematically, parameter \\(\\theta\\) is adjusted so that the joint distribution of the reverse Markov chain \\(\\mathbb{P}_{\\theta}(\\mathbf{z}_0,\\mathbf{z}_1,\\dots ,\\mathbf{z}_T)\\coloneqq \\mathbb{P}(\\mathbf{z}_T)\\prod_{t = 1}^T\\mathbb{P}_{\\theta}(\\mathbf{z}_{t - 1}\\mid \\mathbf{z}_t)\\) closely approximates that of the forward Markov chain \\(\\mathbb{P}(\\mathbf{z}_0,\\mathbf{z}_1,\\dots ,\\mathbf{z}_T)\\coloneqq \\mathbb{P}(\\mathbf{z}_0)\\prod_{t = 1}^T\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{z}_{t - 1})\\). This is achieved by minimizing the following loss:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.103,
+                0.488,
+                0.469,
+                0.514
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbb {E} _ {t \\sim \\mathcal {U} [ 1, T ], \\mathbf {z} _ {0} \\sim \\mathbb {P} (\\mathbf {z} _ {0}), \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t) \\| ^ {2} \\right], \\tag {4}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.52,
+                0.469,
+                0.581
+            ],
+            "angle": 0,
+            "content": "where \\(\\mathbf{z}_t\\) is computed using Eq. (2), and \\(\\epsilon_{\\theta}\\) is predicted using a neural network, typically a UNet architecture [37]. In a conditional diffusion model \\(\\epsilon_{\\theta}(\\mathbf{z}_t,t)\\) gets replaced by \\(\\epsilon_{\\theta}(\\mathbf{z}_t,t,\\mathcal{C})\\), where \\(\\mathcal{C}\\) is a conditioning variable."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.588,
+                0.248,
+                0.603
+            ],
+            "angle": 0,
+            "content": "3.2. Our Architecture"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.612,
+                0.47,
+                0.852
+            ],
+            "angle": 0,
+            "content": "Image Encoder and Latent Diffusion. Diffusion models typically take large number of time steps to train, and are difficult to converge. Recently, [36] proposed a new diffusion with improved convergence properties and is called \"stable-diffusion\". The key idea is to perform the diffusion in latent space, with latent embedding learnt separately through a variational autoencoder (VAE). The Encoder of VAE first transforms the input image \\(\\mathbf{x}\\) of size \\((H,W)\\) to latent space, then we follow latent diffusion formulation and utilize the UNet used in Stable Diffusion[36]. Utilizing latent diffusion formulation enables our architecture to capture multi-resolution features. Hence, we aggregate the feature maps from different layers of the UNet module (implementing conditional diffusion) by bringing them all to 1/4th resolution of the latent space, resulting in a feature map of size \\(8e\\times H / 32\\times W / 32\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.856,
+                0.469,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Exploiting Semantic Context with Conditional Diffusion. Recall that we formulated the single image depth estimation as a dense regression problem, predicting \\(\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{x})\\)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.599,
+                0.089,
+                0.63,
+                0.216
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.607,
+                0.219,
+                0.621,
+                0.228
+            ],
+            "angle": 0,
+            "content": "(a)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.747,
+                0.089,
+                0.81,
+                0.216
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.773,
+                0.219,
+                0.787,
+                0.228
+            ],
+            "angle": 0,
+            "content": "(b)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.498,
+                0.241,
+                0.892,
+                0.366
+            ],
+            "angle": 0,
+            "content": "Figure 3. (a) Probabilistic graphical model corresponding to VPD. (b) The same corresponding to our formulation. Here, \\(\\mathcal{C}\\) represents the semantic embedding derived from our CIDE module. This embedding is internally generated by passing \\(\\mathbf{x}\\) through the ViT, resulting in \\(\\mathcal{E}\\). Subsequently, \\(\\mathcal{E}\\) undergoes further processing to yield \\(\\mathcal{C}\\), which is then utilized in the conditional diffusion module implementing \\(\\mathbb{P}(\\mathbf{z}_0\\mid \\mathbf{z}_t,\\mathcal{C})\\). The output of the conditional diffusion module is fed into the Depth Regressor module within our architecture, implementing \\(\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{z}_0)\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.393,
+                0.892,
+                0.56
+            ],
+            "angle": 0,
+            "content": "where \\(\\mathbf{y}\\) and \\(\\mathbf{x}\\) denote output depth, and input image respectively. In our diffusion formulation, we predict \\(\\mathbf{z}_0\\) which is then used to predict pixel-wise depth, \\(\\mathbf{y}\\). It has been shown that noise prediction in a diffusion model, \\(\\epsilon_{\\theta}\\), can be seen as predicting the gradient of the density function, \\(\\nabla_{\\mathbf{z}_t} \\log \\mathbb{P}(\\mathbf{z}_t)\\). Hence, overall architecture for SIDE using a diffusion architecture can be seen as factorizing the conditional probability, as shown in the probabilistic graph model in Fig. 3a. To utilize additional semantic context generated from a ViT model, we condition it on the ViT embeddings as shown in Fig. 3b. Mathematically, we model:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.514,
+                0.588,
+                0.876,
+                0.605
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbb {P} (\\mathbf {y} \\mid \\mathbf {x}, \\mathcal {E}) = \\mathbb {P} (\\mathbf {y} \\mid \\mathbf {z} _ {0}) \\mathbb {P} (\\mathbf {z} _ {0} \\mid \\mathbf {z} _ {t}, \\mathcal {C}) \\mathbb {P} (\\mathbf {z} _ {t} \\mid \\mathbf {x}) \\mathbb {P} (\\mathcal {C} \\mid \\mathbf {x}),\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.609,
+                0.545,
+                0.621
+            ],
+            "angle": 0,
+            "content": "where"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.533,
+                0.627,
+                0.891,
+                0.644
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbb {P} (\\mathcal {C} \\mid \\mathbf {x}) = \\mathbb {P} (\\mathcal {C} \\mid \\mathcal {E}) \\mathbb {P} (\\mathcal {E} \\mid \\mathbf {x}). \\tag {5}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.657,
+                0.892,
+                0.867
+            ],
+            "angle": 0,
+            "content": "Here, the first term, \\(\\mathbb{P}(\\mathbf{y}\\mid \\mathbf{z}_0)\\), is implemented through the Depth Regressor module explained earlier. Similarly, \\(\\mathbb{P}(\\mathbf{z}_t\\mid \\mathbf{x})\\) is implemented using the VAE's Encoder as described earlier. We generate conditional information, \\(\\mathcal{C}\\) using our Comprehensive Image Detail Embedding module (hereafter CIDE, and explained below). The CIDE module takes \\(\\mathbf{x}\\) as input and generates embedding vector \\(\\mathcal{C}\\) (of dimension 768 in our design) as the output, thus, implementing \\(\\mathbb{P}(\\mathcal{C}\\mid \\mathbf{x})\\) given in Eq. (5). We use \\(\\mathcal{E}\\) to denote the embedding vector of a ViT module, and \\(\\mathbb{P}(\\mathcal{E}\\mid \\mathbf{x})\\) is implemented through the ViT. \\(\\mathbb{P}(\\mathcal{C}\\mid \\mathcal{E})\\) is implemented using downstream modules in CIDE consisting of learnable embeddings. The second term, \\(\\mathbb{P}(\\mathbf{z}_0\\mid \\mathbf{z}_t,\\mathcal{C})\\), is implemented using conditional diffusion,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.871,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Comprehensive Image Detail Embedding (CIDE) Module. As described earlier, we believe using pseudo-captions"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "28288"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.076,
+                0.089,
+                0.895,
+                0.149
+            ],
+            "angle": 0,
+            "content": "Table 1. Results on Indoor NYU Depth v2 [27] Dataset. Results that are bold perform best. \\(\\uparrow\\) means the metric should be higher, \\(\\downarrow\\) indicate lower is better. The evaluation uses an upper bound of 10 meters on the ground truth depth map. All the numbers for other works have been taken from the corresponding papers. For MIM, and ZoEDepth we have used SwinV2-L 1K, and ZoeDepth-M12-N versions respectively. We see an overall improvement against SOTA on all the metrics used for evaluation."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.107,
+                0.156,
+                0.869,
+                0.497
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Method</td><td>Venue</td><td>Abs Rel↓</td><td>RMSE↓</td><td>log10↓</td><td>Sq Rel↓</td><td>δ1↑</td><td>δ2↑</td><td>δ3↑</td></tr><tr><td>Eigen et al.[9]</td><td>NIPS&#x27;14</td><td>0.158</td><td>0.641</td><td>-</td><td>-</td><td>0.769</td><td>0.950</td><td>0.988</td></tr><tr><td>DORN[10]</td><td>CVPR&#x27;18</td><td>0.115</td><td>0.509</td><td>0.051</td><td>-</td><td>0.828</td><td>0.965</td><td>0.992</td></tr><tr><td>SharpNet[32]</td><td>ICCV&#x27;19</td><td>0.139</td><td>0.502</td><td>0.047</td><td>-</td><td>0.836</td><td>0.966</td><td>0.993</td></tr><tr><td>Chen et al.[5]</td><td>IJCAI-19</td><td>0.111</td><td>0.514</td><td>0.048</td><td>-</td><td>0.878</td><td>0.977</td><td>0.994</td></tr><tr><td>BTS[22]</td><td>Arxiv&#x27;19</td><td>0.110</td><td>0.392</td><td>0.047</td><td>0.066</td><td>0.885</td><td>0.978</td><td>0.994</td></tr><tr><td>AdaBins[2]</td><td>CVPR&#x27;21</td><td>0.103</td><td>0.364</td><td>0.044</td><td>-</td><td>0.903</td><td>0.984</td><td>0.997</td></tr><tr><td>DPT[33]</td><td>ICCV&#x27;21</td><td>0.110</td><td>0.357</td><td>0.045</td><td>-</td><td>0.904</td><td>0.988</td><td>0.998</td></tr><tr><td>P3Depth[30]</td><td>CVPR&#x27;22</td><td>0.104</td><td>0.356</td><td>0.043</td><td>-</td><td>0.898</td><td>0.981</td><td>0.996</td></tr><tr><td>NeWCRFs[53]</td><td>CVPR&#x27;22</td><td>0.095</td><td>0.334</td><td>0.041</td><td>0.045</td><td>0.922</td><td>0.992</td><td>0.998</td></tr><tr><td>SwinV2-B[24]</td><td>CVPR&#x27;22</td><td>0.133</td><td>0.462</td><td>0.059</td><td>-</td><td>0.819</td><td>0.975</td><td>0.995</td></tr><tr><td>SwinV2-L[24]</td><td>CVPR&#x27;22</td><td>0.112</td><td>0.381</td><td>0.051</td><td>-</td><td>0.886</td><td>0.984</td><td>0.997</td></tr><tr><td>Localbins[3]</td><td>ECCV&#x27;22</td><td>0.099</td><td>0.357</td><td>0.042</td><td>-</td><td>0.907</td><td>0.987</td><td>0.998</td></tr><tr><td>Jun et al.[17]</td><td>ECCV&#x27;22</td><td>0.098</td><td>0.355</td><td>0.042</td><td>-</td><td>0.913</td><td>0.987</td><td>0.998</td></tr><tr><td>PixelFormer[1]</td><td>WACV&#x27;23</td><td>0.090</td><td>0.322</td><td>0.039</td><td>0.043</td><td>0.929</td><td>0.991</td><td>0.998</td></tr><tr><td>DDP[16]</td><td>ICCV&#x27;23</td><td>0.094</td><td>0.329</td><td>0.040</td><td>-</td><td>0.921</td><td>0.990</td><td>0.998</td></tr><tr><td>MIM [51]</td><td>CVPR&#x27;23</td><td>0.083</td><td>0.287</td><td>0.035</td><td>-</td><td>0.949</td><td>0.994</td><td>0.999</td></tr><tr><td>AiT[28]</td><td>ICCV&#x27;23</td><td>0.076</td><td>0.275</td><td>0.033</td><td>-</td><td>0.954</td><td>0.994</td><td>0.999</td></tr><tr><td>ZoeDepth [4]</td><td>Arxiv&#x27;23</td><td>0.075</td><td>0.270</td><td>0.032</td><td>0.030</td><td>0.955</td><td>0.995</td><td>0.999</td></tr><tr><td>VPD[54]</td><td>ICCV&#x27;23</td><td>0.069</td><td>0.254</td><td>0.030</td><td>0.027</td><td>0.964</td><td>0.995</td><td>0.999</td></tr><tr><td>Ours</td><td>CVPR&#x27;24</td><td>0.059</td><td>0.218</td><td>0.026</td><td>0.013</td><td>0.978</td><td>0.997</td><td>0.999</td></tr></table>"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.16,
+                0.508,
+                0.813,
+                0.827
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.837,
+                0.897,
+                0.867
+            ],
+            "angle": 0,
+            "content": "Figure 4. Visual Comparison on NYU Depth v2 Indoor Dataset. Note, our method's ability to delineate objects in terms of their depth, such as the table lamp in Row 5, even when such information is absent from the ground truth depth map."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.946,
+                0.522,
+                0.957
+            ],
+            "angle": 0,
+            "content": "28289"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.076,
+                0.089,
+                0.895,
+                0.175
+            ],
+            "angle": 0,
+            "content": "Table 2. Performance on the Outdoor KITTI [12] Dataset. Please refer to the caption of Tab. 1 for notation details. For ZoEDepth results, we use the ZoeDepth-M12-K version following the authors' recommendation. In instances where results for certain methods were not reported in the respective works, denoted by “-”, and the code is unavailable, we were unable to generate the missing numbers. Despite the saturation of results on the outdoor KITTI dataset, our method consistently achieves comparable or superior performance to the state-of-the-art (SOTA) across all metrics. VPD[54] cannot be trained on KITTI dataset as it required a per image text label which is not present in KITTI. The symbol \\(\\dagger\\) indicates that the method utilizes additional information beyond RGB."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.085,
+                0.184,
+                0.889,
+                0.451
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Method</td><td>Venue</td><td>Abs Rel↓</td><td>Sq Rel↓</td><td>RMSElog ↓</td><td>RMSE↓</td><td>δ1 ↑</td><td>δ2 ↑</td><td>δ3 ↑</td></tr><tr><td>Eigen et al.[9]</td><td>NIPS&#x27;14</td><td>0.203</td><td>1.517</td><td>0.282</td><td>6.307</td><td>0.702</td><td>0.898</td><td>0.967</td></tr><tr><td>DORN[10]</td><td>CVPR&#x27;18</td><td>0.072</td><td>0.307</td><td>0.120</td><td>2.727</td><td>0.932</td><td>0.984</td><td>0.994</td></tr><tr><td>BTS[22]</td><td>Arxiv&#x27;19</td><td>0.059</td><td>0.241</td><td>0.096</td><td>2.756</td><td>0.956</td><td>0.993</td><td>0.998</td></tr><tr><td>AdaBins[2]</td><td>CVPR&#x27;21</td><td>0.067</td><td>0.190</td><td>0.088</td><td>2.960</td><td>0.949</td><td>0.992</td><td>0.998</td></tr><tr><td>DPT[33]</td><td>ICCV&#x27;21</td><td>0.060</td><td>-</td><td>0.092</td><td>2.573</td><td>0.959</td><td>0.995</td><td>0.996</td></tr><tr><td>P3Depth[30]</td><td>CVPR&#x27;22</td><td>0.071</td><td>0.270</td><td>0.103</td><td>2.842</td><td>0.953</td><td>0.993</td><td>0.998</td></tr><tr><td>NeWCRFs[53]</td><td>CVPR&#x27;22</td><td>0.052</td><td>0.155</td><td>0.079</td><td>2.129</td><td>0.974</td><td>0.997</td><td>0.999</td></tr><tr><td>PixelFormer[1]</td><td>WACV&#x27;23</td><td>0.051</td><td>0.149</td><td>0.077</td><td>2.081</td><td>0.976</td><td>0.997</td><td>0.999</td></tr><tr><td>ZoeDepth [4]</td><td>Arxiv&#x27;23</td><td>0.054</td><td>0.189</td><td>0.083</td><td>2.440</td><td>0.97</td><td>0.996</td><td>0.999</td></tr><tr><td>DDP [16]</td><td>ICCV&#x27;23</td><td>0.050</td><td>0.148</td><td>0.076</td><td>2.072</td><td>0.975</td><td>0.997</td><td>0.999</td></tr><tr><td>URCDC [40]</td><td>ToM&#x27;23</td><td>0.050</td><td>0.142</td><td>0.076</td><td>2.032</td><td>0.977</td><td>0.997</td><td>0.999</td></tr><tr><td>IEBins [41]</td><td>NeurIPS&#x27;23</td><td>0.050</td><td>0.142</td><td>0.075</td><td>2.011</td><td>0.978</td><td>0.998</td><td>0.999</td></tr><tr><td>MIM [51]</td><td>CVPR&#x27;23</td><td>0.050</td><td>0.139</td><td>0.075</td><td>1.966</td><td>0.977</td><td>0.998</td><td>1.000</td></tr><tr><td>GEDepth[52]†</td><td>ICCV&#x27;23</td><td>0.048</td><td>0.142</td><td>0.076</td><td>2.044</td><td>0.976</td><td>0.997</td><td>0.999</td></tr><tr><td>Ours</td><td>CVPR&#x27;24</td><td>0.048</td><td>0.139</td><td>0.074</td><td>2.039</td><td>0.979</td><td>0.998</td><td>1.000</td></tr></table>"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.08,
+                0.462,
+                0.891,
+                0.707
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.303,
+                0.715,
+                0.666,
+                0.73
+            ],
+            "angle": 0,
+            "content": "Figure 5. Visual Comparison on KITTI Outdoor Dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.757,
+                0.473,
+                0.892
+            ],
+            "angle": 0,
+            "content": "to generate the semantic context has limited utility, as the textual descriptions typically focus on large salient objects only. Instead we propose our CIDE module which use embeddings from a pre-trained ViT, and extract detailed semantic context from these embeddings. For this we take 1000 dimensional logit vector from ViT, and pass it through a two layer MLP which converts it to a 100 dimensional vector. Subsequently, we employ this vector to compute the linear combination of 100 learnable embeddings."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.757,
+                0.892,
+                0.801
+            ],
+            "angle": 0,
+            "content": "This resulting embedding undergoes a linear transformation to yield a semantic context vector of dimension 768, which is then passed to conditional diffusion module."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.81,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Depth Regressor. The output feature map undergoes through an Upsampling Decoder, comprised of deconvolution layers, followed by a Depth Regressor. The Depth Regressor is essentially a two-layer convolutional neural network (CNN), with the initial layer having dimensions \\(\\mathrm{Conv}(3\\times 3)\\), 192, and the subsequent layer \\(\\mathrm{Conv}(3\\times 3)\\), 1."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "28290"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.076,
+                0.089,
+                0.892,
+                0.146
+            ],
+            "angle": 0,
+            "content": "Table 3. Quantitative results for zero-shot transfer to four unseen indoor datasets. mRI denotes the mean relative improvement with respect to NeWCRFs across all metrics (δ1, REL, RMSE). Evaluation depth is capped at 8m for SUN RGB-D, 10m for iBims and DIODE Indoor, and 80m for HyperSim. Best results are in bold, second best are underlined. Our mRI outperforms all methods across all datasets by a large margin. † denotes that ZoeD is trained on 12 datasets and our method is trained only on NYUv2."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.078,
+                0.158,
+                0.915,
+                0.288
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"4\">SUN RGB-D</td><td colspan=\"4\">iBims-1 Benchmark</td><td colspan=\"4\">DIODE Indoor</td><td colspan=\"4\">HyperSim</td></tr><tr><td>δ1↑</td><td>REL↓</td><td>RMSE↓</td><td>mRIθ↑</td><td>δ1↑</td><td>REL↓</td><td>RMSE↓</td><td>mRIθ↑</td><td>δ1↑</td><td>REL↓</td><td>RMSE↓</td><td>mRIθ↑</td><td>δ1↑</td><td>REL↓</td><td>RMSE↓</td><td>mRIθ↑</td></tr><tr><td>BTS [22]</td><td>0.740</td><td>0.172</td><td>0.515</td><td>-14.2%</td><td>0.538</td><td>0.231</td><td>0.919</td><td>-6.9%</td><td>0.210</td><td>0.418</td><td>1.905</td><td>2.3%</td><td>0.225</td><td>0.476</td><td>6.404</td><td>-8.6%</td></tr><tr><td>AdaBins [2]</td><td>0.771</td><td>0.159</td><td>0.476</td><td>-7.0%</td><td>0.555</td><td>0.212</td><td>0.901</td><td>-2.1%</td><td>0.174</td><td>0.443</td><td>1.963</td><td>-7.2%</td><td>0.221</td><td>0.483</td><td>6.546</td><td>-10.5%</td></tr><tr><td>LocalBins [3]</td><td>0.777</td><td>0.156</td><td>0.470</td><td>-5.6%</td><td>0.558</td><td>0.211</td><td>0.880</td><td>-0.7%</td><td>0.229</td><td>0.412</td><td>1.853</td><td>7.1%</td><td>0.234</td><td>0.468</td><td>6.362</td><td>-6.6%</td></tr><tr><td>NeWCRFs [53]</td><td>0.798</td><td>0.151</td><td>0.424</td><td>0.0%</td><td>0.548</td><td>0.206</td><td>0.861</td><td>0.0%</td><td>0.187</td><td>0.404</td><td>1.867</td><td>0.0%</td><td>0.255</td><td>0.442</td><td>6.017</td><td>0.0%</td></tr><tr><td>VPD [54]</td><td>0.861</td><td>0.121</td><td>0.355</td><td>14.7%</td><td>0.627</td><td>0.187</td><td>0.767</td><td>11.5%</td><td>0.480</td><td>0.392</td><td>1.295</td><td>63.4%</td><td>0.333</td><td>0.531</td><td>5.111</td><td>8.5%</td></tr><tr><td>ZoeD-M12-N† [4]</td><td>0.864</td><td>0.119</td><td>0.346</td><td>16.0%</td><td>0.658</td><td>0.169</td><td>0.711</td><td>18.5%</td><td>0.376</td><td>0.327</td><td>1.588</td><td>45.0%</td><td>0.292</td><td>0.410</td><td>5.771</td><td>8.6%</td></tr><tr><td>Ours</td><td>0.885</td><td>0.112</td><td>0.319</td><td>20.5%</td><td>0.688</td><td>0.163</td><td>0.664</td><td>23.1%</td><td>0.545</td><td>0.344</td><td>1.164</td><td>81.3%</td><td>0.394</td><td>0.442</td><td>4.739</td><td>25.2%</td></tr></table>"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.137,
+                0.306,
+                0.847,
+                0.462
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.137,
+                0.479,
+                0.846,
+                0.639
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.652,
+                0.89,
+                0.694
+            ],
+            "angle": 0,
+            "content": "Figure 6. Visualization of improvements over VPD [54] in our model due to ViT embeddings passed down as conditional vectors (in blue and green). In the above images, ViT detects the desktop computer (first image) and table lamp (second image) with high probability, and they are thus better detected by our model. Additional visualizations are provided in the supplementary material."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.719,
+                0.312,
+                0.736
+            ],
+            "angle": 0,
+            "content": "4. Experiments and Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.75,
+                0.471,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Datasets and Evaluation. We use NYU Depth v2 [27] and KITTI [12] as the primary datasets for training. The NYU Depth v2 dataset is a widely used indoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images in the train set and 654 in the test set. The dataset covers a wide range of indoor scenes and includes challenging scenarios such as reflective surfaces, transparent objects, and occlusions. The ground truth depth maps are obtained using a structured light sensor and are provided at a resolution of \\(640 \\times 480\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.72,
+                0.892,
+                0.888
+            ],
+            "angle": 0,
+            "content": "KITTI dataset is a widely used outdoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images. The dataset covers outdoor driving scenarios and includes varying lighting conditions, weather, and occlusions. The ground truth depth maps are obtained using a Velodyne LiDAR sensor and are provided at a resolution of \\(1242 \\times 375\\). To demonstrate generalizability, we evaluate zero-shot performance on the following datasets: Sun-RGBD [42], iBims1 [19], DIODE [49], and HyperSim [35]. Whereas, the main paper contains mostly quantitative, and only some representative"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.946,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "28291"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.077,
+                0.089,
+                0.472,
+                0.201
+            ],
+            "angle": 0,
+            "content": "Table 4. Effectiveness of different embeddings for guiding the diffusion process for depth estimation on NYU Depth v2 dataset. In the third row, rather than using pseudo caption embeddings generated from the scene label (as implemented by VPD [54]), we provide the one-hot vector representing the scene label as a condition to the diffusion model. We observe a slight improvement in the metrics, highlighting that information content in the caption is similar to that in one-hot label."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.084,
+                0.212,
+                0.465,
+                0.305
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Embeddings</td><td>RMSE↓</td><td>Abs Rel↓</td><td>δ1↑</td></tr><tr><td>Scene label emb. [54]</td><td>0.254</td><td>0.069</td><td>0.964</td></tr><tr><td>Text caption emb. [20]</td><td>0.225</td><td>0.062</td><td>0.976</td></tr><tr><td>One-hot vector emb.</td><td>0.244</td><td>0.067</td><td>0.968</td></tr><tr><td>Proposed scene emb.</td><td>0.218</td><td>0.059</td><td>0.978</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.33,
+                0.47,
+                0.361
+            ],
+            "angle": 0,
+            "content": "tive visual results, detailed visual results on each dataset are included in the supplementary."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.363,
+                0.47,
+                0.545
+            ],
+            "angle": 0,
+            "content": "Implementation Details. Our model is implemented using PyTorch [29]. For optimization, we have used AdamW optimizer [25] with \\(\\beta_0\\) values of 0.9 and 0.999, a batch size of 32, and a weight decay of 0.1. We train our model for 25 epochs for both KITTI and NYU Depth v2 datasets, with an initial learning rate of \\(3 \\times 10^{-5}\\). We first linearly increase learning rate to \\(5 \\times 10^{-4}\\), and then linearly decrease across training iterations. We use usual data augmentation techniques, including random hue addition, horizontal flipping, changing the image brightness, and Cut Depth [15]. Our model takes approximately 21 minutes per epoch to train using 8 NVIDIA A100 GPUs."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.555,
+                0.396,
+                0.57
+            ],
+            "angle": 0,
+            "content": "4.1. Comparison on Benchmark Datasets"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.578,
+                0.469,
+                0.683
+            ],
+            "angle": 0,
+            "content": "Comparison on NYU Depth v2. Tab. 1 shows the comparison of our proposed method with SOTA methods on the indoor NYU Depth v2 dataset [27]. We achieve a new state of the art on this dataset. Our methods perform better than the previous SOTA ([54])) by a large margin of \\(14\\%\\) in terms of RMSE. Fig. 4 provides a qualitative comparison on the dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.686,
+                0.469,
+                0.779
+            ],
+            "angle": 0,
+            "content": "Comparison on KITTI. Tab. 2 shows the comparison with various methods on the KITTI dataset [12]. Fig. 5 shows the qualitative results. Unlike NYU Depth v2 dataset, KITTI is an outdoor dataset. On this dataset also, we achieve similar or better performance than all existing state-of-the-art techniques."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.787,
+                0.412,
+                0.803
+            ],
+            "angle": 0,
+            "content": "4.2. Generalization and Zero Shot Transfer"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.81,
+                0.469,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Unlike state of the art on zero short transfer (ZoEDepth [4]), which requiring training on many datasets (12 in their case) for effective zero shot transfer, we show that our model generalizes well to other unseen dataset even when trained on a single NYU Depth v2 dataset. Tab. 3 shows quantitative results to back our findings."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.091,
+                0.655,
+                0.108
+            ],
+            "angle": 0,
+            "content": "4.3. Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.114,
+                0.892,
+                0.312
+            ],
+            "angle": 0,
+            "content": "Effect of Contextual Information. As highlighted in the motivation, a key observation of this study is the richness of information contained within the output probability vector of the ViT, surpassing the textual embeddings employed in current state-of-the-art (e.g., VPD [54]). To compare with the utilization of scene label information (as implemented by VPD), we construct the conditioning embedding as a one-hot vector using the scene label, and subsequently transform it using an MLP. As illustrated in Tab. 4, we observe a slight improvement over VPD, indicating that the information content in pseudo-captions resembles that of one-hot labels. Furthermore, our proposed method surpasses both the approaches."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.313,
+                0.892,
+                0.466
+            ],
+            "angle": 0,
+            "content": "Qualitative Results. Perhaps the most important task would be to verify that use of rich probability vector instead of text embeddings actually results in an improvement in depth as a direct consequence. We do this by considering the top few objects predicted by the ViT and correlating this with the depth predicted at these objects. We hypothesise when a particular class say dog is predicted with a high probability (hence there must be a dog in the image), then the corresponding depth must also be more accurate. We show this in Fig. 6."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.473,
+                0.619,
+                0.488
+            ],
+            "angle": 0,
+            "content": "5. Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.498,
+                0.892,
+                0.725
+            ],
+            "angle": 0,
+            "content": "We presented a new architecture block Comprehensive Image Detail Embedding (CIDE) module for robust monocular depth estimation in this paper. Our key idea is to highlight the limitations of using pseudo-captions to provide contextual information, and instead propose to use richer class-wise probability generated by a classification model, such as ViT. The motivation is that, while textual embedding typically highlights salient objects, class-wise probability vector preserves more details, including smaller objects in the background also. We implement the idea using proposed CIDE module cascaded with a conditional diffusion pipeline for monocular depth estimation. We demonstrated the effectiveness of our approach on several benchmark datasets and showed that it outperforms SOTA methods by a significant margin."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.728,
+                0.892,
+                0.835
+            ],
+            "angle": 0,
+            "content": "Acknowledgement. We thank Kartik Anand for assistance with some of the experiments. This work has been supported by funding through Staqu Technologies Pvt. Ltd., and Department of Science and Technology (DST), Government of India. The travel for Suraj Patni is supported through Research Acceleration funds of Department of Computer Science and Engineering, IIT Delhi."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.501,
+                0.848,
+                0.597,
+                0.864
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.507,
+                0.873,
+                0.892,
+                0.903
+            ],
+            "angle": 0,
+            "content": "[1] Ashutosh Agarwal and Chetan Arora. Attention attention everywhere: Monocular depth prediction with skip attention."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "28292"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.108,
+                0.092,
+                0.47,
+                0.134
+            ],
+            "angle": 0,
+            "content": "In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5861-5870, 2023. 2, 5, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.136,
+                0.472,
+                0.205
+            ],
+            "angle": 0,
+            "content": "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4009-4018, 2021. 5, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.207,
+                0.472,
+                0.262
+            ],
+            "angle": 0,
+            "content": "[3] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 5, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.264,
+                0.472,
+                0.305
+            ],
+            "angle": 0,
+            "content": "[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 2, 3, 5, 6, 7, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.307,
+                0.472,
+                0.348
+            ],
+            "angle": 0,
+            "content": "[5] Xiaotian Chen, Xuejin Chen, and Zheng-Jun Zha. Structure-aware residual pyramid network for monocular depth estimation. arXiv preprint arXiv:1907.06023, 2019. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.35,
+                0.472,
+                0.419
+            ],
+            "angle": 0,
+            "content": "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.421,
+                0.472,
+                0.502
+            ],
+            "angle": 0,
+            "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.504,
+                0.472,
+                0.545
+            ],
+            "angle": 0,
+            "content": "[8] Yiqun Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation, 2023. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.548,
+                0.472,
+                0.602
+            ],
+            "angle": 0,
+            "content": "[9] 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. 2, 5, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.079,
+                0.604,
+                0.472,
+                0.673
+            ],
+            "angle": 0,
+            "content": "[10] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 5, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.079,
+                0.676,
+                0.472,
+                0.744
+            ],
+            "angle": 0,
+            "content": "[11] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.079,
+                0.746,
+                0.472,
+                0.8
+            ],
+            "angle": 0,
+            "content": "[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013. 3, 6, 7, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.079,
+                0.803,
+                0.472,
+                0.857
+            ],
+            "angle": 0,
+            "content": "[13] José L. Herrera, Carlos R. del Blanco, and Narciso García. Automatic depth extraction from 2d images using a cluster-based learning framework. IEEE Transactions on Image Processing, 27(7):3288-3299, 2018. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.079,
+                0.859,
+                0.472,
+                0.902
+            ],
+            "angle": 0,
+            "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.079,
+                0.092,
+                0.472,
+                0.902
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.093,
+                0.893,
+                0.134
+            ],
+            "angle": 0,
+            "content": "[15] Yasunori Ishii and Takayoshi Yamashita. Cutdepth: Edge-aware data augmentation in depth estimation. arXiv preprint arXiv:2107.07684, 2021. 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.136,
+                0.893,
+                0.19
+            ],
+            "angle": 0,
+            "content": "[16] Yuanfeng Ji, Zhe Chen, Enze Xie, Lanqing Hong, Xihui Liu, Zhaoqiang Liu, Tong Lu, Zhenguo Li, and Ping Luo. Ddp: Diffusion model for dense visual prediction. arXiv preprint arXiv:2303.17559, 2023. 2, 3, 5, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.193,
+                0.893,
+                0.248
+            ],
+            "angle": 0,
+            "content": "[17] Jinyoung Jun, Jae-Han Lee, Chul Lee, and Chang-Su Kim. Depth map decomposition for monocular depth estimation. In European Conference on Computer Vision, pages 18-34. Springer, 2022. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.25,
+                0.893,
+                0.305
+            ],
+            "angle": 0,
+            "content": "[18] Kevin Karsch, Ce Liu, and Sing Bing Kang. Depth transfer: Depth extraction from video using non-parametric sampling. IEEE transactions on pattern analysis and machine intelligence, 36(11):2144-2158, 2014. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.307,
+                0.893,
+                0.374
+            ],
+            "angle": 0,
+            "content": "[19] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.377,
+                0.893,
+                0.432
+            ],
+            "angle": 0,
+            "content": "[20] Neehar Kondapaneni, Markus Marks, Manuel Knott, Rogério Guimarães, and Pietro Perona. Text-image alignment for diffusion-based perception. arXiv preprint arXiv:2310.00031, 2023. 2, 3, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.434,
+                0.893,
+                0.502
+            ],
+            "angle": 0,
+            "content": "[21] 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. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.504,
+                0.893,
+                0.559
+            ],
+            "angle": 0,
+            "content": "[22] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 5, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.561,
+                0.893,
+                0.63
+            ],
+            "angle": 0,
+            "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.632,
+                0.893,
+                0.702
+            ],
+            "angle": 0,
+            "content": "[24] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.704,
+                0.893,
+                0.731
+            ],
+            "angle": 0,
+            "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICML 2019. 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.733,
+                0.893,
+                0.788
+            ],
+            "angle": 0,
+            "content": "[26] John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J Davison. Scelenet rgb-d: 5m photorealistic images of synthetic indoor trajectories with ground truth. arXiv preprint arXiv:1612.05079, 2016. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.789,
+                0.893,
+                0.83
+            ],
+            "angle": 0,
+            "content": "[27] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 3, 5, 7, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.832,
+                0.893,
+                0.902
+            ],
+            "angle": 0,
+            "content": "[28] Jia Ning, Chen Li, Zheng Zhang, Chunyu Wang, Zigang Geng, Qi Dai, Kun He, and Han Hu. All in tokens: Unifying output space of visual tasks via soft token. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19900-19910, 2023. 2, 5"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.503,
+                0.093,
+                0.893,
+                0.902
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.945,
+                0.52,
+                0.957
+            ],
+            "angle": 0,
+            "content": "28293"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.092,
+                0.47,
+                0.217
+            ],
+            "angle": 0,
+            "content": "[29] 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. Curran Associates, Inc., 2019. 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.219,
+                0.47,
+                0.288
+            ],
+            "angle": 0,
+            "content": "[30] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1610–1621, 2022. 5, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.29,
+                0.47,
+                0.373
+            ],
+            "angle": 0,
+            "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.375,
+                0.469,
+                0.444
+            ],
+            "angle": 0,
+            "content": "[32] Michael Ramamonjisoa and Vincent Lepetit. Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.446,
+                0.471,
+                0.502
+            ],
+            "angle": 0,
+            "content": "[33] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 5, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.504,
+                0.47,
+                0.574
+            ],
+            "angle": 0,
+            "content": "[34] 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"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.575,
+                0.469,
+                0.657
+            ],
+            "angle": 0,
+            "content": "[35] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021. 1, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.66,
+                0.47,
+                0.73
+            ],
+            "angle": 0,
+            "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.731,
+                0.47,
+                0.815
+            ],
+            "angle": 0,
+            "content": "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3, 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.817,
+                0.469,
+                0.857
+            ],
+            "angle": 0,
+            "content": "[38] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.859,
+                0.469,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m:"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.092,
+                0.471,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.533,
+                0.093,
+                0.89,
+                0.12
+            ],
+            "angle": 0,
+            "content": "Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.121,
+                0.892,
+                0.176
+            ],
+            "angle": 0,
+            "content": "[40] Shuwei Shao, Zhongcai Pei, Weihai Chen, Ran Li, Zhong Liu, and Zhengguo Li. Urcdc-depth: Uncertainty rectified cross-distillation with cutflip for monocular depth estimation. arXiv preprint arXiv:2302.08149, 2023. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.177,
+                0.892,
+                0.232
+            ],
+            "angle": 0,
+            "content": "[41] Shuwei Shao, Zhongcai Pei, Xingming Wu, Zhong Liu, Weihai Chen, and Zhengguo Li. Iebins: Iterative elastic bins for monocular depth estimation. arXiv preprint arXiv:2309.14137, 2023. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.233,
+                0.892,
+                0.288
+            ],
+            "angle": 0,
+            "content": "[42] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 567-576, 2015. 1, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.289,
+                0.892,
+                0.329
+            ],
+            "angle": 0,
+            "content": "[43] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.33,
+                0.892,
+                0.372
+            ],
+            "angle": 0,
+            "content": "[44] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.372,
+                0.892,
+                0.427
+            ],
+            "angle": 0,
+            "content": "[45] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.428,
+                0.892,
+                0.482
+            ],
+            "angle": 0,
+            "content": "[46] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. Advances in Neural Information Processing Systems, 34: 1415-1428, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.484,
+                0.892,
+                0.537
+            ],
+            "angle": 0,
+            "content": "[47] M. Sun, A. Y. Ng, and A. Saxena. Make3d: Learning 3d scene structure from a single still image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(05):824-840, 2009. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.539,
+                0.892,
+                0.622
+            ],
+            "angle": 0,
+            "content": "[48] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.623,
+                0.892,
+                0.692
+            ],
+            "angle": 0,
+            "content": "[49] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 1, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.692,
+                0.892,
+                0.748
+            ],
+            "angle": 0,
+            "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.749,
+                0.892,
+                0.816
+            ],
+            "angle": 0,
+            "content": "[51] Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14475-14485, 2023. 2, 5, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.818,
+                0.892,
+                0.872
+            ],
+            "angle": 0,
+            "content": "[52] Xiaodong Yang, Zhuang Ma, Zhiyu Ji, and Zhe Ren. Gedgepth: Ground embedding for monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12719-12727, 2023. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.873,
+                0.892,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[53] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Newcrfs: Neural window fully-connected crfs for"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.503,
+                0.093,
+                0.892,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "28294"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.11,
+                0.092,
+                0.47,
+                0.134
+            ],
+            "angle": 0,
+            "content": "monocular depth estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 5, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.136,
+                0.47,
+                0.188
+            ],
+            "angle": 0,
+            "content": "[54] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. ICCV, 2023. 2, 3, 5, 6, 7, 8"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.092,
+                0.47,
+                0.188
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "28295"
+        }
+    ]
+]

2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/7e3c5935-f184-4aa8-a48b-682e7477329f_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82db315fa3070134234835799ec765d2db6701c5a0886c590e7f9dd7788b348d
+size 3204672

2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/full.md

@@ -0,0 +1,268 @@
+# ECoDepth: Effective Conditioning of Diffusion Models for Monocular Depth Estimation
+
+Suraj Patni*
+
+Aradhye Agarwal*
+
+Chetan Arora
+
+Indian Institute of Technology Delhi
+
+https://ecodepth-iitd.github.io
+
+# Abstract
+
+In the absence of parallax cues, a learning based single image depth estimation (SIDE) model relies heavily on shading and contextual cues in the image. While this simplicity is attractive, it is necessary to train such models on large and varied datasets, which are difficult to capture. It has been shown that using embeddings from pretrained foundational models, such as CLIP, improves zero shot transfer in several applications. Taking inspiration from this, in our paper we explore the use of global image priors generated from a pre-trained ViT model to provide more detailed contextual information. We argue that the embedding vector from a ViT model, pre-trained on a large dataset, captures greater relevant information for SIDE than the usual route of generating pseudo image captions, followed by CLIP based text embeddings. Based on this idea, we propose a new SIDE model using a diffusion backbone which is conditioned on ViT embeddings. Our proposed design establishes a new state-of-the-art (SOTA) for SIDE on NYU Depth v2 dataset, achieving Abs Rel error of 0.059(14% improvement) compared to 0.069 by the current SOTA (VPD). And on KITTI dataset, achieving Sq Rel error of 0.139 (2% improvement) compared to 0.142 by the current SOTA (GED). For zero shot transfer with a model trained on NYU Depth v2, we report mean relative improvement of (20%, 23%, 81%, 25%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, compared to (16%, 18%, 45%, 9%) by ZoEDepth. The code is available in our project page.
+
+# 1. Introduction
+
+Single Image Depth Estimation (SIDE) is the task of predicting per pixel depth using a single RGB image from a monocular camera. It is a fundamental problem in computer vision with applications in several domains, viz robotics, autonomous driving, and augmented reality. The problem is
+
+![](images/f7274f44c4d22bc9eae8a19f69aa92c1c20646d54843a02397296650e2bfc12a.jpg)  
+(a) Sun-RGBD [42]
+
+![](images/214cfea1c3468673bbbfb8b1f65db15a952fafc13969355b377513969847d17f.jpg)
+
+![](images/ff8d25236e5638817716f97d36604908ba6a66250a0304d48025d379b65c2095.jpg)
+
+![](images/f56327bab79fcd3f27c08fbef298e0b0e23fd81917afe234efe190f0d7e73e38.jpg)
+
+![](images/acfee1ae7e238a4d60c7914c77d3468b68bc4fb10c24754766404c5d2a5dd4a9.jpg)  
+(b) iBims1 [19]
+
+![](images/b6c9dc2b7d14f841f908bf571798f99779b924830b846d5fe98305823fd3225f.jpg)
+
+![](images/5a9b6186ef62db411c91c5843145d756b45f98d60f13a0e2aa97b4f3d446142e.jpg)  
+(c) DIODE [49]  
+(d) HyperSim [35]  
+Figure 1. Qualitative results across four different datasets, demonstrating the zero-shot performance of our model trained only on the NYU Depth v2 dataset. Corresponding quantitative results are presented in Table 3. The first column displays RGB images, the second column depicts ground truth depth, and the third column showcases our model's predicted depths. Additional images for each dataset are available in the Supplementary Material.
+
+![](images/8d5070d4cbc6c4ebcfef4196245823ef4f87ec26324f3274b07d2d886dd1ddc2.jpg)
+
+typically formulated in two flavors: metric depth estimation (MDE), and relative depth estimation (RDE). As the names suggest, MDE deals with estimation in physical units such as meters, whereas, RDE techniques focus on relative depth only, and require a per-image affine transformation as post-processing to convert to a physical unit.
+
+Conventional geometric techniques for depth estimation typically rely on feature correspondence, parallax, and triangulation from two or more views. However, the problem becomes ill-posed for estimation from a single view. Intuitively, depth map is a 3D representation of the scene, whereas an RGB image is a 2D projection of the scene. Hence, it cannot be uniquely determined from a single RGB image. Therefore, learning-based SIDE models rely on visual cues like 'shape from shading' and other contextual priors for per-pixel depth prediction.
+
+A data-driven approach makes the SIDE pipeline sim
+
+pler, but also makes the learnt model dependent on the quality of training data. It has been observed that such models overfit on a particular training distribution/domain and fail to generalize on unseen data. This is especially true for MDE models when the range of depth in the training dataset is limited. Hence, training on multiple datasets, with wide variations in depth ranges has been proposed [4, 34].
+
+On the other hand, development of large foundational models (LFMs) in recent years has altered the preferred design approach for many computer vision problems. These huge models are trained using extensive datasets of unlabeled images and learning objectives that are agnostic to specific tasks. The learnt embeddings from such pre-trained models have been shown to help generalization and zero-shot transfer in many applications. We are aware of at least two works for MDE in SIDE problem that have appeared in the last few months and make use of such foundational models. VPD [54] uses a text-to-image diffusion model pretrained on LAION-400M [39] dataset having large-scale image-text pairs as the backbone. The model prompts the denoising UNet with textual inputs to make the visual contents interact with the text prompts. Since the problem formulation doesn't include text description as the input, the model generates simple descriptions such as 'A photo of a {scene name}' based on the scene label given in the NYU Depth v2 dataset. Another work, TADP [20] improves upon VPD. Instead of simple image descriptions, TADP uses BLIP-2 [23] to generate image captions, and then uses CLIP [31] embeddings of the pseudo-caption to condition the diffusion model.
+
+We view the above works as providing robust semantic context to LFMs for the actual task at hand, which helps in visual recognition in general, as well as SIDE. While we do agree with the broad motivations of these works, the question that we ask is if pseudo-captions are the most effective way to provide the semantic context. Textual descriptions of an image typically focus on large salient objects and emphasize on their relationships. On the other hand, large vision models for image classification, typically contain representation for even smaller objects present in the scene. Even when a single object is present in the scene, the representations typically capture uncertainties and ambiguities inherent in the scene. Hence, we posit that using embeddings from a transformer model, pretrained on a large image dataset unrelated to the SIDE task, captures more relevant information, and is a better alternative to using pseudocaptions' CLIP embedding. Since diffusion-based models have shown their superiority for the dense prediction tasks in recent works [8, 16, 20, 38, 54]. Hence, we propose a diffusion backbone for our model, along with a novel CIDE module which employs ViT [7] to extract semantic context embeddings. These embeddings are subsequently utilized to condition the diffusion backbone.
+
+Contributions. The key contributions of this work are:
+
+(1) We propose a new model for MDE in a SIDE task. The proposed model uses a conditional diffusion architecture with the semantic context being supplied through embeddings generated using a ViT model. Achieving a new state-of-the-art (SOTA) performance, our method outperforms existing approaches on benchmark datasets, including the NYU Depth v2 indoor and KITTI outdoor datasets. Notably, we report a significant improvement of  $14\%$  in absolute relative error, achieving 0.059 compared to the current SOTA (VPD) performance of 0.069 on NYU Depth v2. And we report an improvement of  $2\%$  in square relative error, achieving 0.139 compared to the current SOTA (GED) performance of 0.142 on KITTI.  
+(2) We show, qualitatively as well as quantitatively, that using ViT embeddings to provide semantic context is a better alternative to generating pseudo captions and then using its CLIP embeddings to condition a SIDE model. In contrast to TADP [20], which uses pseudo captions, but only achieves a RMSE of 0.225 on NYU Depth v2, we report a lower, and a new SOTA, error of 0.218 (VPD [54] achieves 0.254).  
+(3) We show that providing ViT conditioning, helps our model perform better in a zero shot transfer task. ZoEDepth [4], the current SOTA for zero-shot transfer, reports an improvement of (16%, 18%, 45%, 9%) over NeWCRF on (Sun-RGBD, iBims1, DIODE, HyperSim) datasets, after training their model on 12 other datasets and NYU Depth v2. In contrast, we only train on NYU Depth v2, and report a much larger improvement of (21%, 23%, 81%, 25%).
+
+# 2. Related Work
+
+Traditional Methods. Earlier techniques for SIDE have used Markov Random Fields [47], non-parametric depth sampling [18], and structural similarity with prior depth map [13] to predict pixel-wise depth.
+
+Deep Learning Techniques for SIDE. Modern techniques have approached the problem as a dense regression problem. CNNs have been the dominant architecture for the SIDE in the last decade, with global-local network stack [9], and multi-scale [21], or encoder-decoder architecture [11] as some of the popular solution strategies. Recently, PixelFormer [1] used a transformer-based encoder-decoder architecture with skip connections from encoders to decoders. MIM [51] proposed masked image modeling as a general-purpose pre-training for geometric and motion tasks such as SIDE and pose estimation. Similarly, AiT [28] used mask augmentation and proposed soft tokens to generalize visual prediction tasks. While earlier works either focused on MDE for a specific dataset or RDE for generalization on multiple datasets, ZoEDepth [4] has proposed
+
+![](images/c7e530d0b4f2d38341472733a1549e75c90b00840b1be1a8bba2c60c1793cff7.jpg)  
+Figure 2. An overview of our proposed model: The latent representation of the input image undergoes a diffusion process, which is conditioned by our proposed CIDe module. Within the CIDe module, the input image is fed through the frozen ViT model. From this, a linear combination of the learnt embeddings is computed, which is transformed to generate a 768-dimensional contextual embedding. This embedding is utilized to condition the diffusion backbone. Subsequently, hierarchical feature maps are extracted from the UNet's decoder which are concatenated and processed through a depth regressor to generate the depth map.
+
+a generalized method for MDE that performs well in zero-shot transfer. We outperform [4] by a large margin even when training on a single dataset - NYU Depth v2 for indoor scenes or KITTI for outdoor scenes. Whereas, [4] uses 12 datasets for pre-training and then fine-tunes on NYU Depth v2 or KITTI for zero shot transfer.
+
+Diffusion-based Methods with Pretraining on Large Datasets. Recently, many techniques for SIDE has used diffusion architectures. These techniques [8, 16, 20, 38, 54], exploit prior knowledge acquired by pretraining on large datasets like LAION-400M [39], which consists of 400 million image-text pairs. In contrast, depth datasets, such as NYU Depth v2 and KITTI, contain around 20-30 thousand image-depth pairs. DepthGen [38] and DDP [16] work on a noise-to-depthmap paradigm and use images for conditional guidance of the diffusion process. DepthGen employs self-supervised pretraining on tasks like colorization, inpainting, and JPEG artifact removal, followed by supervised training on indoor and outdoor datasets [6, 12, 26, 27, 48]. DDP [16] decouples the image encoder and map decoder, allowing the image encoder to run just once, while the lightweight map decoder is run multiple times. VPD [54] and TADP [20] use denoising UNet [37] as a backbone to extract the rich features at multiple scales. Also, they utilize text instead of image for conditioning the diffusion backbone.
+
+Vision Transformer for Scene Understanding. The transformer architecture was initially proposed for NLP tasks [50], but introduced to the computer vision community as Vision Transformer (ViT) in [7]. Prior to this, CNNs dominated computer vision problems due to their ability to capture spatial hierarchies in image data. However, such archi
+
+tectures were constrained due to their ability to learn spatially localized features only. Transformer architecture has a weaker inductive bias and allows ViT to learn long-range dependencies, and robust, generalizable features. ViT architectures have replaced CNNs for SOTA performance on most computer vision tasks in recent years. We use a pretrained ViT model for providing semantic information to the diffusion backbone in our model.
+
+# 3. Proposed Methodology
+
+# 3.1. Preliminaries
+
+Problem Formulation. The objective of single image depth prediction task is to predict continuous values, denoted as  $\mathbf{y} \in \mathbb{R}^{H \times W}$ , for every pixel present in the input RGB image,  $\mathbf{x} \in [0,255]^{3 \times H \times W}$ . Here  $H$  and  $W$  represent the height and width respectively of the input image.
+
+Diffusion Model. Diffusion models are a class of generative models that progressively inject noise into the input data (forward pass) and then learn to reconstruct the original data in a reverse denoising process (reverse pass). There are three formulations of diffusion models: denoising diffusion probabilistic models (DDPMs) [14], score-based generative models [43, 44], and those based on stochastic differential equations [45, 46]. DDPMs are of relevance to our paper, and are described below. The architecture of DDPMs consists of two Markov chains: a forward chain that adds noise to the data, and a reverse chain that converts noise back to data by learning transition kernels parameterized by deep neural networks. Formally, the forward pass is modeled as
+
+a Markov process:
+
+$$
+\mathbb {P} \left(\mathbf {z} _ {t} \mid \mathbf {z} _ {t - 1}\right) = \mathcal {N} \left(\mathbf {z} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {z} _ {t - 1}, \beta_ {t} \boldsymbol {I}\right). \tag {1}
+$$
+
+Here  $\mathbf{z}_t$  denotes the random variable at the  $t^{\mathrm{th}}$  time step,  $\mathcal{N}(\mathbf{z};\boldsymbol {\mu},\boldsymbol {\sigma})$  denotes Gaussian probability distribution, and  $\beta_{t}$  is the noise schedule. The above equation leads to the analytic form of  $\mathbb{P}(\mathbf{z}_t\mid \mathbf{z}_0),\forall t\in \{0,1,\dots ,T\}$ :
+
+$$
+\mathbf {z} _ {t} = \sqrt {\bar {\beta} _ {t}} \mathbf {z} _ {0} + \sqrt {1 - \bar {\beta} _ {t}} \boldsymbol {\epsilon}, \tag {2}
+$$
+
+where  $\bar{\beta}_t = \prod_{s=1}^t \beta_s$  and  $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . The model then gradually removes noise by executing a learnable Markov chain in the reverse time direction, parameterized by a normal prior distribution  $\mathbb{P}(\mathbf{z}_T) = \mathcal{N}(\mathbf{z}_T;0,\mathbf{I})$  and a learnable transition kernel  $\mathbb{P}_{\theta}(\mathbf{z}_{t-1}|\mathbf{z}_t)$  given by:
+
+$$
+\mathbb {P} _ {\theta} \left(\mathbf {z} _ {t - 1} \mid \mathbf {z} _ {t}\right) = \mathcal {N} \left(\mathbf {z} _ {t - 1}; \mu_ {\theta} \left(\mathbf {z} _ {t}, t\right), \Sigma_ {\theta} \left(\mathbf {z} _ {t}, t\right)\right). \tag {3}
+$$
+
+The goal of the training process is to approximately match the reverse Markov chain with the actual time reversal of the forward Markov chain. Mathematically, parameter  $\theta$  is adjusted so that the joint distribution of the reverse Markov chain  $\mathbb{P}_{\theta}(\mathbf{z}_0,\mathbf{z}_1,\dots ,\mathbf{z}_T)\coloneqq \mathbb{P}(\mathbf{z}_T)\prod_{t = 1}^T\mathbb{P}_{\theta}(\mathbf{z}_{t - 1}\mid \mathbf{z}_t)$  closely approximates that of the forward Markov chain  $\mathbb{P}(\mathbf{z}_0,\mathbf{z}_1,\dots ,\mathbf{z}_T)\coloneqq \mathbb{P}(\mathbf{z}_0)\prod_{t = 1}^T\mathbb{P}(\mathbf{z}_t\mid \mathbf{z}_{t - 1})$ . This is achieved by minimizing the following loss:
+
+$$
+\mathbb {E} _ {t \sim \mathcal {U} [ 1, T ], \mathbf {z} _ {0} \sim \mathbb {P} (\mathbf {z} _ {0}), \epsilon \sim \mathcal {N} (0, I)} \left[ \| \epsilon - \epsilon_ {\theta} (\mathbf {z} _ {t}, t) \| ^ {2} \right], \tag {4}
+$$
+
+where  $\mathbf{z}_t$  is computed using Eq. (2), and  $\epsilon_{\theta}$  is predicted using a neural network, typically a UNet architecture [37]. In a conditional diffusion model  $\epsilon_{\theta}(\mathbf{z}_t,t)$  gets replaced by  $\epsilon_{\theta}(\mathbf{z}_t,t,\mathcal{C})$ , where  $\mathcal{C}$  is a conditioning variable.
+
+# 3.2. Our Architecture
+
+Image Encoder and Latent Diffusion. Diffusion models typically take large number of time steps to train, and are difficult to converge. Recently, [36] proposed a new diffusion with improved convergence properties and is called "stable-diffusion". The key idea is to perform the diffusion in latent space, with latent embedding learnt separately through a variational autoencoder (VAE). The Encoder of VAE first transforms the input image  $\mathbf{x}$  of size  $(H,W)$  to latent space, then we follow latent diffusion formulation and utilize the UNet used in Stable Diffusion[36]. Utilizing latent diffusion formulation enables our architecture to capture multi-resolution features. Hence, we aggregate the feature maps from different layers of the UNet module (implementing conditional diffusion) by bringing them all to 1/4th resolution of the latent space, resulting in a feature map of size  $8e\times H / 32\times W / 32$ .
+
+Exploiting Semantic Context with Conditional Diffusion. Recall that we formulated the single image depth estimation as a dense regression problem, predicting  $\mathbb{P}(\mathbf{y}\mid \mathbf{x})$
+
+![](images/c884c50e858aad2ed14417c86c14d50488599881dbd4715cd3e63339ec6572b7.jpg)  
+(a)
+
+![](images/57991546efa28fa9bd31fcb0cf70cb7c634b4e25f7be459feddfdac4e6a8ced9.jpg)  
+(b)  
+Figure 3. (a) Probabilistic graphical model corresponding to VPD. (b) The same corresponding to our formulation. Here,  $\mathcal{C}$  represents the semantic embedding derived from our CIDE module. This embedding is internally generated by passing  $\mathbf{x}$  through the ViT, resulting in  $\mathcal{E}$ . Subsequently,  $\mathcal{E}$  undergoes further processing to yield  $\mathcal{C}$ , which is then utilized in the conditional diffusion module implementing  $\mathbb{P}(\mathbf{z}_0\mid \mathbf{z}_t,\mathcal{C})$ . The output of the conditional diffusion module is fed into the Depth Regressor module within our architecture, implementing  $\mathbb{P}(\mathbf{y}\mid \mathbf{z}_0)$ .
+
+where  $\mathbf{y}$  and  $\mathbf{x}$  denote output depth, and input image respectively. In our diffusion formulation, we predict  $\mathbf{z}_0$  which is then used to predict pixel-wise depth,  $\mathbf{y}$ . It has been shown that noise prediction in a diffusion model,  $\epsilon_{\theta}$ , can be seen as predicting the gradient of the density function,  $\nabla_{\mathbf{z}_t} \log \mathbb{P}(\mathbf{z}_t)$ . Hence, overall architecture for SIDE using a diffusion architecture can be seen as factorizing the conditional probability, as shown in the probabilistic graph model in Fig. 3a. To utilize additional semantic context generated from a ViT model, we condition it on the ViT embeddings as shown in Fig. 3b. Mathematically, we model:
+
+$$
+\mathbb {P} (\mathbf {y} \mid \mathbf {x}, \mathcal {E}) = \mathbb {P} (\mathbf {y} \mid \mathbf {z} _ {0}) \mathbb {P} (\mathbf {z} _ {0} \mid \mathbf {z} _ {t}, \mathcal {C}) \mathbb {P} (\mathbf {z} _ {t} \mid \mathbf {x}) \mathbb {P} (\mathcal {C} \mid \mathbf {x}),
+$$
+
+where
+
+$$
+\mathbb {P} (\mathcal {C} \mid \mathbf {x}) = \mathbb {P} (\mathcal {C} \mid \mathcal {E}) \mathbb {P} (\mathcal {E} \mid \mathbf {x}). \tag {5}
+$$
+
+Here, the first term,  $\mathbb{P}(\mathbf{y}\mid \mathbf{z}_0)$ , is implemented through the Depth Regressor module explained earlier. Similarly,  $\mathbb{P}(\mathbf{z}_t\mid \mathbf{x})$  is implemented using the VAE's Encoder as described earlier. We generate conditional information,  $\mathcal{C}$  using our Comprehensive Image Detail Embedding module (hereafter CIDE, and explained below). The CIDE module takes  $\mathbf{x}$  as input and generates embedding vector  $\mathcal{C}$  (of dimension 768 in our design) as the output, thus, implementing  $\mathbb{P}(\mathcal{C}\mid \mathbf{x})$  given in Eq. (5). We use  $\mathcal{E}$  to denote the embedding vector of a ViT module, and  $\mathbb{P}(\mathcal{E}\mid \mathbf{x})$  is implemented through the ViT.  $\mathbb{P}(\mathcal{C}\mid \mathcal{E})$  is implemented using downstream modules in CIDE consisting of learnable embeddings. The second term,  $\mathbb{P}(\mathbf{z}_0\mid \mathbf{z}_t,\mathcal{C})$ , is implemented using conditional diffusion,
+
+Comprehensive Image Detail Embedding (CIDE) Module. As described earlier, we believe using pseudo-captions
+
+Table 1. Results on Indoor NYU Depth v2 [27] Dataset. Results that are bold perform best.  $\uparrow$  means the metric should be higher,  $\downarrow$  indicate lower is better. The evaluation uses an upper bound of 10 meters on the ground truth depth map. All the numbers for other works have been taken from the corresponding papers. For MIM, and ZoEDepth we have used SwinV2-L 1K, and ZoeDepth-M12-N versions respectively. We see an overall improvement against SOTA on all the metrics used for evaluation.  
+
+<table><tr><td>Method</td><td>Venue</td><td>Abs Rel↓</td><td>RMSE↓</td><td>log10↓</td><td>Sq Rel↓</td><td>δ1↑</td><td>δ2↑</td><td>δ3↑</td></tr><tr><td>Eigen et al.[9]</td><td>NIPS&#x27;14</td><td>0.158</td><td>0.641</td><td>-</td><td>-</td><td>0.769</td><td>0.950</td><td>0.988</td></tr><tr><td>DORN[10]</td><td>CVPR&#x27;18</td><td>0.115</td><td>0.509</td><td>0.051</td><td>-</td><td>0.828</td><td>0.965</td><td>0.992</td></tr><tr><td>SharpNet[32]</td><td>ICCV&#x27;19</td><td>0.139</td><td>0.502</td><td>0.047</td><td>-</td><td>0.836</td><td>0.966</td><td>0.993</td></tr><tr><td>Chen et al.[5]</td><td>IJCAI-19</td><td>0.111</td><td>0.514</td><td>0.048</td><td>-</td><td>0.878</td><td>0.977</td><td>0.994</td></tr><tr><td>BTS[22]</td><td>Arxiv&#x27;19</td><td>0.110</td><td>0.392</td><td>0.047</td><td>0.066</td><td>0.885</td><td>0.978</td><td>0.994</td></tr><tr><td>AdaBins[2]</td><td>CVPR&#x27;21</td><td>0.103</td><td>0.364</td><td>0.044</td><td>-</td><td>0.903</td><td>0.984</td><td>0.997</td></tr><tr><td>DPT[33]</td><td>ICCV&#x27;21</td><td>0.110</td><td>0.357</td><td>0.045</td><td>-</td><td>0.904</td><td>0.988</td><td>0.998</td></tr><tr><td>P3Depth[30]</td><td>CVPR&#x27;22</td><td>0.104</td><td>0.356</td><td>0.043</td><td>-</td><td>0.898</td><td>0.981</td><td>0.996</td></tr><tr><td>NeWCRFs[53]</td><td>CVPR&#x27;22</td><td>0.095</td><td>0.334</td><td>0.041</td><td>0.045</td><td>0.922</td><td>0.992</td><td>0.998</td></tr><tr><td>SwinV2-B[24]</td><td>CVPR&#x27;22</td><td>0.133</td><td>0.462</td><td>0.059</td><td>-</td><td>0.819</td><td>0.975</td><td>0.995</td></tr><tr><td>SwinV2-L[24]</td><td>CVPR&#x27;22</td><td>0.112</td><td>0.381</td><td>0.051</td><td>-</td><td>0.886</td><td>0.984</td><td>0.997</td></tr><tr><td>Localbins[3]</td><td>ECCV&#x27;22</td><td>0.099</td><td>0.357</td><td>0.042</td><td>-</td><td>0.907</td><td>0.987</td><td>0.998</td></tr><tr><td>Jun et al.[17]</td><td>ECCV&#x27;22</td><td>0.098</td><td>0.355</td><td>0.042</td><td>-</td><td>0.913</td><td>0.987</td><td>0.998</td></tr><tr><td>PixelFormer[1]</td><td>WACV&#x27;23</td><td>0.090</td><td>0.322</td><td>0.039</td><td>0.043</td><td>0.929</td><td>0.991</td><td>0.998</td></tr><tr><td>DDP[16]</td><td>ICCV&#x27;23</td><td>0.094</td><td>0.329</td><td>0.040</td><td>-</td><td>0.921</td><td>0.990</td><td>0.998</td></tr><tr><td>MIM [51]</td><td>CVPR&#x27;23</td><td>0.083</td><td>0.287</td><td>0.035</td><td>-</td><td>0.949</td><td>0.994</td><td>0.999</td></tr><tr><td>AiT[28]</td><td>ICCV&#x27;23</td><td>0.076</td><td>0.275</td><td>0.033</td><td>-</td><td>0.954</td><td>0.994</td><td>0.999</td></tr><tr><td>ZoeDepth [4]</td><td>Arxiv&#x27;23</td><td>0.075</td><td>0.270</td><td>0.032</td><td>0.030</td><td>0.955</td><td>0.995</td><td>0.999</td></tr><tr><td>VPD[54]</td><td>ICCV&#x27;23</td><td>0.069</td><td>0.254</td><td>0.030</td><td>0.027</td><td>0.964</td><td>0.995</td><td>0.999</td></tr><tr><td>Ours</td><td>CVPR&#x27;24</td><td>0.059</td><td>0.218</td><td>0.026</td><td>0.013</td><td>0.978</td><td>0.997</td><td>0.999</td></tr></table>
+
+![](images/a76cc74dd06135df0021484b881dda524bb0070b423a6e292f03cff9e997573e.jpg)  
+Figure 4. Visual Comparison on NYU Depth v2 Indoor Dataset. Note, our method's ability to delineate objects in terms of their depth, such as the table lamp in Row 5, even when such information is absent from the ground truth depth map.
+
+Table 2. Performance on the Outdoor KITTI [12] Dataset. Please refer to the caption of Tab. 1 for notation details. For ZoEDepth results, we use the ZoeDepth-M12-K version following the authors' recommendation. In instances where results for certain methods were not reported in the respective works, denoted by “-”, and the code is unavailable, we were unable to generate the missing numbers. Despite the saturation of results on the outdoor KITTI dataset, our method consistently achieves comparable or superior performance to the state-of-the-art (SOTA) across all metrics. VPD[54] cannot be trained on KITTI dataset as it required a per image text label which is not present in KITTI. The symbol  $\dagger$  indicates that the method utilizes additional information beyond RGB.  
+
+<table><tr><td>Method</td><td>Venue</td><td>Abs Rel↓</td><td>Sq Rel↓</td><td>RMSElog ↓</td><td>RMSE↓</td><td>δ1 ↑</td><td>δ2 ↑</td><td>δ3 ↑</td></tr><tr><td>Eigen et al.[9]</td><td>NIPS&#x27;14</td><td>0.203</td><td>1.517</td><td>0.282</td><td>6.307</td><td>0.702</td><td>0.898</td><td>0.967</td></tr><tr><td>DORN[10]</td><td>CVPR&#x27;18</td><td>0.072</td><td>0.307</td><td>0.120</td><td>2.727</td><td>0.932</td><td>0.984</td><td>0.994</td></tr><tr><td>BTS[22]</td><td>Arxiv&#x27;19</td><td>0.059</td><td>0.241</td><td>0.096</td><td>2.756</td><td>0.956</td><td>0.993</td><td>0.998</td></tr><tr><td>AdaBins[2]</td><td>CVPR&#x27;21</td><td>0.067</td><td>0.190</td><td>0.088</td><td>2.960</td><td>0.949</td><td>0.992</td><td>0.998</td></tr><tr><td>DPT[33]</td><td>ICCV&#x27;21</td><td>0.060</td><td>-</td><td>0.092</td><td>2.573</td><td>0.959</td><td>0.995</td><td>0.996</td></tr><tr><td>P3Depth[30]</td><td>CVPR&#x27;22</td><td>0.071</td><td>0.270</td><td>0.103</td><td>2.842</td><td>0.953</td><td>0.993</td><td>0.998</td></tr><tr><td>NeWCRFs[53]</td><td>CVPR&#x27;22</td><td>0.052</td><td>0.155</td><td>0.079</td><td>2.129</td><td>0.974</td><td>0.997</td><td>0.999</td></tr><tr><td>PixelFormer[1]</td><td>WACV&#x27;23</td><td>0.051</td><td>0.149</td><td>0.077</td><td>2.081</td><td>0.976</td><td>0.997</td><td>0.999</td></tr><tr><td>ZoeDepth [4]</td><td>Arxiv&#x27;23</td><td>0.054</td><td>0.189</td><td>0.083</td><td>2.440</td><td>0.97</td><td>0.996</td><td>0.999</td></tr><tr><td>DDP [16]</td><td>ICCV&#x27;23</td><td>0.050</td><td>0.148</td><td>0.076</td><td>2.072</td><td>0.975</td><td>0.997</td><td>0.999</td></tr><tr><td>URCDC [40]</td><td>ToM&#x27;23</td><td>0.050</td><td>0.142</td><td>0.076</td><td>2.032</td><td>0.977</td><td>0.997</td><td>0.999</td></tr><tr><td>IEBins [41]</td><td>NeurIPS&#x27;23</td><td>0.050</td><td>0.142</td><td>0.075</td><td>2.011</td><td>0.978</td><td>0.998</td><td>0.999</td></tr><tr><td>MIM [51]</td><td>CVPR&#x27;23</td><td>0.050</td><td>0.139</td><td>0.075</td><td>1.966</td><td>0.977</td><td>0.998</td><td>1.000</td></tr><tr><td>GEDepth[52]†</td><td>ICCV&#x27;23</td><td>0.048</td><td>0.142</td><td>0.076</td><td>2.044</td><td>0.976</td><td>0.997</td><td>0.999</td></tr><tr><td>Ours</td><td>CVPR&#x27;24</td><td>0.048</td><td>0.139</td><td>0.074</td><td>2.039</td><td>0.979</td><td>0.998</td><td>1.000</td></tr></table>
+
+![](images/fd5626577cb0a6d7c11ad79b7f023d80f1eaefeb0fdc9d0368fc06a025dcee80.jpg)  
+Figure 5. Visual Comparison on KITTI Outdoor Dataset.
+
+to generate the semantic context has limited utility, as the textual descriptions typically focus on large salient objects only. Instead we propose our CIDE module which use embeddings from a pre-trained ViT, and extract detailed semantic context from these embeddings. For this we take 1000 dimensional logit vector from ViT, and pass it through a two layer MLP which converts it to a 100 dimensional vector. Subsequently, we employ this vector to compute the linear combination of 100 learnable embeddings.
+
+This resulting embedding undergoes a linear transformation to yield a semantic context vector of dimension 768, which is then passed to conditional diffusion module.
+
+Depth Regressor. The output feature map undergoes through an Upsampling Decoder, comprised of deconvolution layers, followed by a Depth Regressor. The Depth Regressor is essentially a two-layer convolutional neural network (CNN), with the initial layer having dimensions  $\mathrm{Conv}(3\times 3)$ , 192, and the subsequent layer  $\mathrm{Conv}(3\times 3)$ , 1.
+
+Table 3. Quantitative results for zero-shot transfer to four unseen indoor datasets. mRI denotes the mean relative improvement with respect to NeWCRFs across all metrics (δ1, REL, RMSE). Evaluation depth is capped at 8m for SUN RGB-D, 10m for iBims and DIODE Indoor, and 80m for HyperSim. Best results are in bold, second best are underlined. Our mRI outperforms all methods across all datasets by a large margin. † denotes that ZoeD is trained on 12 datasets and our method is trained only on NYUv2.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">SUN RGB-D</td><td colspan="4">iBims-1 Benchmark</td><td colspan="4">DIODE Indoor</td><td colspan="4">HyperSim</td></tr><tr><td>δ1↑</td><td>REL↓</td><td>RMSE↓</td><td>mRIθ↑</td><td>δ1↑</td><td>REL↓</td><td>RMSE↓</td><td>mRIθ↑</td><td>δ1↑</td><td>REL↓</td><td>RMSE↓</td><td>mRIθ↑</td><td>δ1↑</td><td>REL↓</td><td>RMSE↓</td><td>mRIθ↑</td></tr><tr><td>BTS [22]</td><td>0.740</td><td>0.172</td><td>0.515</td><td>-14.2%</td><td>0.538</td><td>0.231</td><td>0.919</td><td>-6.9%</td><td>0.210</td><td>0.418</td><td>1.905</td><td>2.3%</td><td>0.225</td><td>0.476</td><td>6.404</td><td>-8.6%</td></tr><tr><td>AdaBins [2]</td><td>0.771</td><td>0.159</td><td>0.476</td><td>-7.0%</td><td>0.555</td><td>0.212</td><td>0.901</td><td>-2.1%</td><td>0.174</td><td>0.443</td><td>1.963</td><td>-7.2%</td><td>0.221</td><td>0.483</td><td>6.546</td><td>-10.5%</td></tr><tr><td>LocalBins [3]</td><td>0.777</td><td>0.156</td><td>0.470</td><td>-5.6%</td><td>0.558</td><td>0.211</td><td>0.880</td><td>-0.7%</td><td>0.229</td><td>0.412</td><td>1.853</td><td>7.1%</td><td>0.234</td><td>0.468</td><td>6.362</td><td>-6.6%</td></tr><tr><td>NeWCRFs [53]</td><td>0.798</td><td>0.151</td><td>0.424</td><td>0.0%</td><td>0.548</td><td>0.206</td><td>0.861</td><td>0.0%</td><td>0.187</td><td>0.404</td><td>1.867</td><td>0.0%</td><td>0.255</td><td>0.442</td><td>6.017</td><td>0.0%</td></tr><tr><td>VPD [54]</td><td>0.861</td><td>0.121</td><td>0.355</td><td>14.7%</td><td>0.627</td><td>0.187</td><td>0.767</td><td>11.5%</td><td>0.480</td><td>0.392</td><td>1.295</td><td>63.4%</td><td>0.333</td><td>0.531</td><td>5.111</td><td>8.5%</td></tr><tr><td>ZoeD-M12-N† [4]</td><td>0.864</td><td>0.119</td><td>0.346</td><td>16.0%</td><td>0.658</td><td>0.169</td><td>0.711</td><td>18.5%</td><td>0.376</td><td>0.327</td><td>1.588</td><td>45.0%</td><td>0.292</td><td>0.410</td><td>5.771</td><td>8.6%</td></tr><tr><td>Ours</td><td>0.885</td><td>0.112</td><td>0.319</td><td>20.5%</td><td>0.688</td><td>0.163</td><td>0.664</td><td>23.1%</td><td>0.545</td><td>0.344</td><td>1.164</td><td>81.3%</td><td>0.394</td><td>0.442</td><td>4.739</td><td>25.2%</td></tr></table>
+
+![](images/10a7e4cdf5394b729bfc2c2fda0f2f9fb4531171367e28fe6cc3a38ac46faaed.jpg)
+
+![](images/f31058aa86b9b19c5376f9c5dabda1e36277891d933e7679fad2e97499e3bd57.jpg)  
+Figure 6. Visualization of improvements over VPD [54] in our model due to ViT embeddings passed down as conditional vectors (in blue and green). In the above images, ViT detects the desktop computer (first image) and table lamp (second image) with high probability, and they are thus better detected by our model. Additional visualizations are provided in the supplementary material.
+
+# 4. Experiments and Results
+
+Datasets and Evaluation. We use NYU Depth v2 [27] and KITTI [12] as the primary datasets for training. The NYU Depth v2 dataset is a widely used indoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images in the train set and 654 in the test set. The dataset covers a wide range of indoor scenes and includes challenging scenarios such as reflective surfaces, transparent objects, and occlusions. The ground truth depth maps are obtained using a structured light sensor and are provided at a resolution of  $640 \times 480$ .
+
+KITTI dataset is a widely used outdoor benchmark for monocular depth estimation, containing over 24k densely labeled pairs of RGB and depth images. The dataset covers outdoor driving scenarios and includes varying lighting conditions, weather, and occlusions. The ground truth depth maps are obtained using a Velodyne LiDAR sensor and are provided at a resolution of  $1242 \times 375$ . To demonstrate generalizability, we evaluate zero-shot performance on the following datasets: Sun-RGBD [42], iBims1 [19], DIODE [49], and HyperSim [35]. Whereas, the main paper contains mostly quantitative, and only some representative
+
+Table 4. Effectiveness of different embeddings for guiding the diffusion process for depth estimation on NYU Depth v2 dataset. In the third row, rather than using pseudo caption embeddings generated from the scene label (as implemented by VPD [54]), we provide the one-hot vector representing the scene label as a condition to the diffusion model. We observe a slight improvement in the metrics, highlighting that information content in the caption is similar to that in one-hot label.  
+
+<table><tr><td>Embeddings</td><td>RMSE↓</td><td>Abs Rel↓</td><td>δ1↑</td></tr><tr><td>Scene label emb. [54]</td><td>0.254</td><td>0.069</td><td>0.964</td></tr><tr><td>Text caption emb. [20]</td><td>0.225</td><td>0.062</td><td>0.976</td></tr><tr><td>One-hot vector emb.</td><td>0.244</td><td>0.067</td><td>0.968</td></tr><tr><td>Proposed scene emb.</td><td>0.218</td><td>0.059</td><td>0.978</td></tr></table>
+
+tive visual results, detailed visual results on each dataset are included in the supplementary.
+
+Implementation Details. Our model is implemented using PyTorch [29]. For optimization, we have used AdamW optimizer [25] with  $\beta_0$  values of 0.9 and 0.999, a batch size of 32, and a weight decay of 0.1. We train our model for 25 epochs for both KITTI and NYU Depth v2 datasets, with an initial learning rate of  $3 \times 10^{-5}$ . We first linearly increase learning rate to  $5 \times 10^{-4}$ , and then linearly decrease across training iterations. We use usual data augmentation techniques, including random hue addition, horizontal flipping, changing the image brightness, and Cut Depth [15]. Our model takes approximately 21 minutes per epoch to train using 8 NVIDIA A100 GPUs.
+
+# 4.1. Comparison on Benchmark Datasets
+
+Comparison on NYU Depth v2. Tab. 1 shows the comparison of our proposed method with SOTA methods on the indoor NYU Depth v2 dataset [27]. We achieve a new state of the art on this dataset. Our methods perform better than the previous SOTA ([54])) by a large margin of  $14\%$  in terms of RMSE. Fig. 4 provides a qualitative comparison on the dataset.
+
+Comparison on KITTI. Tab. 2 shows the comparison with various methods on the KITTI dataset [12]. Fig. 5 shows the qualitative results. Unlike NYU Depth v2 dataset, KITTI is an outdoor dataset. On this dataset also, we achieve similar or better performance than all existing state-of-the-art techniques.
+
+# 4.2. Generalization and Zero Shot Transfer
+
+Unlike state of the art on zero short transfer (ZoEDepth [4]), which requiring training on many datasets (12 in their case) for effective zero shot transfer, we show that our model generalizes well to other unseen dataset even when trained on a single NYU Depth v2 dataset. Tab. 3 shows quantitative results to back our findings.
+
+# 4.3. Ablation Study
+
+Effect of Contextual Information. As highlighted in the motivation, a key observation of this study is the richness of information contained within the output probability vector of the ViT, surpassing the textual embeddings employed in current state-of-the-art (e.g., VPD [54]). To compare with the utilization of scene label information (as implemented by VPD), we construct the conditioning embedding as a one-hot vector using the scene label, and subsequently transform it using an MLP. As illustrated in Tab. 4, we observe a slight improvement over VPD, indicating that the information content in pseudo-captions resembles that of one-hot labels. Furthermore, our proposed method surpasses both the approaches.
+
+Qualitative Results. Perhaps the most important task would be to verify that use of rich probability vector instead of text embeddings actually results in an improvement in depth as a direct consequence. We do this by considering the top few objects predicted by the ViT and correlating this with the depth predicted at these objects. We hypothesise when a particular class say dog is predicted with a high probability (hence there must be a dog in the image), then the corresponding depth must also be more accurate. We show this in Fig. 6.
+
+# 5. Conclusion
+
+We presented a new architecture block Comprehensive Image Detail Embedding (CIDE) module for robust monocular depth estimation in this paper. Our key idea is to highlight the limitations of using pseudo-captions to provide contextual information, and instead propose to use richer class-wise probability generated by a classification model, such as ViT. The motivation is that, while textual embedding typically highlights salient objects, class-wise probability vector preserves more details, including smaller objects in the background also. We implement the idea using proposed CIDE module cascaded with a conditional diffusion pipeline for monocular depth estimation. We demonstrated the effectiveness of our approach on several benchmark datasets and showed that it outperforms SOTA methods by a significant margin.
+
+Acknowledgement. We thank Kartik Anand for assistance with some of the experiments. This work has been supported by funding through Staqu Technologies Pvt. Ltd., and Department of Science and Technology (DST), Government of India. The travel for Suraj Patni is supported through Research Acceleration funds of Department of Computer Science and Engineering, IIT Delhi.
+
+# References
+
+[1] Ashutosh Agarwal and Chetan Arora. Attention attention everywhere: Monocular depth prediction with skip attention.
+
+In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5861-5870, 2023. 2, 5, 6  
+[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4009-4018, 2021. 5, 6, 7  
+[3] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 5, 7  
+[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 2, 3, 5, 6, 7, 8  
+[5] Xiaotian Chen, Xuejin Chen, and Zheng-Jun Zha. Structure-aware residual pyramid network for monocular depth estimation. arXiv preprint arXiv:1907.06023, 2019. 5  
+[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3  
+[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3  
+[8] Yiqun Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estimation, 2023. 2, 3  
+[9] 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. 2, 5, 6  
+[10] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 5, 6  
+[11] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 2  
+[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013. 3, 6, 7, 8  
+[13] José L. Herrera, Carlos R. del Blanco, and Narciso García. Automatic depth extraction from 2d images using a cluster-based learning framework. IEEE Transactions on Image Processing, 27(7):3288-3299, 2018. 2  
+[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3
+
+[15] Yasunori Ishii and Takayoshi Yamashita. Cutdepth: Edge-aware data augmentation in depth estimation. arXiv preprint arXiv:2107.07684, 2021. 8  
+[16] Yuanfeng Ji, Zhe Chen, Enze Xie, Lanqing Hong, Xihui Liu, Zhaoqiang Liu, Tong Lu, Zhenguo Li, and Ping Luo. Ddp: Diffusion model for dense visual prediction. arXiv preprint arXiv:2303.17559, 2023. 2, 3, 5, 6  
+[17] Jinyoung Jun, Jae-Han Lee, Chul Lee, and Chang-Su Kim. Depth map decomposition for monocular depth estimation. In European Conference on Computer Vision, pages 18-34. Springer, 2022. 5  
+[18] Kevin Karsch, Ce Liu, and Sing Bing Kang. Depth transfer: Depth extraction from video using non-parametric sampling. IEEE transactions on pattern analysis and machine intelligence, 36(11):2144-2158, 2014. 2  
+[19] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 7  
+[20] Neehar Kondapaneni, Markus Marks, Manuel Knott, Rogério Guimarães, and Pietro Perona. Text-image alignment for diffusion-based perception. arXiv preprint arXiv:2310.00031, 2023. 2, 3, 8  
+[21] 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. 2  
+[22] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 5, 6, 7  
+[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2  
+[24] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 5  
+[25] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICML 2019. 8  
+[26] John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J Davison. Scelenet rgb-d: 5m photorealistic images of synthetic indoor trajectories with ground truth. arXiv preprint arXiv:1612.05079, 2016. 3  
+[27] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 3, 5, 7, 8  
+[28] Jia Ning, Chen Li, Zheng Zhang, Chunyu Wang, Zigang Geng, Qi Dai, Kun He, and Han Hu. All in tokens: Unifying output space of visual tasks via soft token. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19900-19910, 2023. 2, 5
+
+[29] 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. Curran Associates, Inc., 2019. 8  
+[30] Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1610–1621, 2022. 5, 6  
+[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2  
+[32] Michael Ramamonjisoa and Vincent Lepetit. Sharpnet: Fast and accurate recovery of occluding contours in monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 5  
+[33] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 5, 6  
+[34] 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  
+[35] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021. 1, 7  
+[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4  
+[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3, 4  
+[38] Saurabh Saxena, Abhishek Kar, Mohammad Norouzi, and David J Fleet. Monocular depth estimation using diffusion models. arXiv preprint arXiv:2302.14816, 2023. 2, 3  
+[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m:
+
+Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 3  
+[40] Shuwei Shao, Zhongcai Pei, Weihai Chen, Ran Li, Zhong Liu, and Zhengguo Li. Urcdc-depth: Uncertainty rectified cross-distillation with cutflip for monocular depth estimation. arXiv preprint arXiv:2302.08149, 2023. 6  
+[41] Shuwei Shao, Zhongcai Pei, Xingming Wu, Zhong Liu, Weihai Chen, and Zhengguo Li. Iebins: Iterative elastic bins for monocular depth estimation. arXiv preprint arXiv:2309.14137, 2023. 6  
+[42] Shuran Song, Samuel P. Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 567-576, 2015. 1, 7  
+[43] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3  
+[44] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 3  
+[45] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3  
+[46] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. Advances in Neural Information Processing Systems, 34: 1415-1428, 2021. 3  
+[47] M. Sun, A. Y. Ng, and A. Saxena. Make3d: Learning 3d scene structure from a single still image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(05):824-840, 2009. 2  
+[48] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3  
+[49] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 1, 7  
+[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3  
+[51] Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14475-14485, 2023. 2, 5, 6  
+[52] Xiaodong Yang, Zhuang Ma, Zhiyu Ji, and Zhe Ren. Gedgepth: Ground embedding for monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12719-12727, 2023. 6  
+[53] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Newcrfs: Neural window fully-connected crfs for
+
+monocular depth estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 5, 6, 7  
+[54] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. ICCV, 2023. 2, 3, 5, 6, 7, 8

2024/ECoDepth_ Effective Conditioning of Diffusion Models for Monocular Depth Estimation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:863e04a34f1a7e85775c197f2d584e373498273cd760c211dff122e8b0aab040
+size 900243

2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_content_list.json

@@ -0,0 +1,1586 @@
+[
+    {
+        "type": "text",
+        "text": "EFHQ: Multi-purpose ExtremePose-Face-HQ dataset",
+        "text_level": 1,
+        "bbox": [
+            212,
+            130,
+            756,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Trung Tuan Dao $^{1,*}$  Duc Hong Vu $^{1,2,*}$  Cuong Pham $^{1,3}$  Anh Tran $^{1}$",
+        "bbox": [
+            200,
+            189,
+            766,
+            208
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ VinAI Research, Vietnam  $^{2}$ Miami University  $^{3}$ Posts & Telecommunications Inst. of Tech., Vietnam",
+        "bbox": [
+            174,
+            210,
+            795,
+            224
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{v.trungdt21,v.ducvh5,v.cuongpv11,v.anhtt152}@vinai.io vudh3@miamioh.edu cuongpv@ptit.edu.vn",
+        "bbox": [
+            140,
+            229,
+            828,
+            243
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "*Equal contribution",
+        "bbox": [
+            423,
+            244,
+            542,
+            260
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/cbf16de748627893751e10fb88a1c1970cd6edd74202d56555bea8261d2edfef.jpg",
+        "image_caption": [
+            "Figure 1. Benefits of our proposed dataset (EFHQ). Standard large-scale facial datasets have most images at near frontal views, causing inferior performance of trained models on downstream tasks when dealing with extreme head poses. For instance, the trained 2D image generators and text-to-image ones often produce only near frontal faces, while the 3D face generators and face reenactment methods often show distorted outputs at profile views. The recently proposed dataset LPFF [47] partially handles that issue by providing complementary images at extreme head poses for only 2D and 3D image generation tasks. Our proposed dataset EFHQ provides high-quality extreme-pose images to complement a wide range of face-related tasks. It supports 2D and 3D image generation, with generally better diversity than LPFF. EFHQ also helps correct the outputs of text-to-image generation and face reenactment at extreme views. Finally, EFHQ provides a more challenging pose-based face verification benchmark to better assess the quality of face recognition networks."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            143,
+            292,
+            831,
+            465
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            597,
+            313,
+            612
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The existing facial datasets, while having plentiful images at near frontal views, lack images with extreme head poses, leading to the downgraded performance of deep learning models when dealing with profile or pitched faces. This work aims to address this gap by introducing a novel dataset named Extreme Pose Face High-Quality Dataset (EFHQ), which includes a maximum of 450k high-quality images of faces at extreme poses. To produce such a massive dataset, we utilize a novel and meticulous dataset processing pipeline to curate two publicly available datasets, VFHQ and CelebV-HQ, which contain many high-resolution face videos captured in various settings. Our dataset can complement existing datasets on various facial-related tasks, such as facial synthesis with 2D/3D-aware GAN, diffusion-based text-to-image face generation, and face reenactment. Specifically, training with EFHQ helps models generalize well across diverse poses, significantly improving performance in scenarios involving ex",
+        "bbox": [
+            73,
+            628,
+            473,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "treme views, confirmed by extensive experiments. Additionally, we utilize EFHQ to define a challenging cross-view face verification benchmark, in which the performance of SOTA face recognition models drops  $5 - 37\\%$  compared to frontal-to-frontal scenarios, aiming to stimulate studies on face recognition under severe pose conditions in the wild.",
+        "bbox": [
+            496,
+            599,
+            895,
+            689
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            500,
+            722,
+            632,
+            738
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The human face is a central focus in computer vision, with extensive research dedicated to tasks like detection, recognition, generation, and manipulation [39, 42]. Numerous large-scale facial datasets[9, 12, 14, 24, 28, 30, 45, 48, 54, 55] have facilitated the face-related studies. Along with deep learning techniques, many tasks have made a considerable leap in performance in recent years. For instance, face recognition systems can recognize near-frontal faces in the wild with near-perfect accuracy [13, 18]. Facial generative models, such as [5, 15, 49], can synthesize realistic",
+        "bbox": [
+            496,
+            750,
+            893,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "22605",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "images indistinguishable from real faces. Face reenactment systems, including [41, 43, 51], can animate an input face based on poses and expressions from any driver video.",
+        "bbox": [
+            75,
+            90,
+            468,
+            136
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Despite impressive results, the mentioned models produce inferior results when dealing with faces at extreme head poses. As illustrated in Fig. 1, faces synthesized by 3D face generation or face reenactment models are often distorted at profile views, due to the scarcity of extreme pose images in training datasets. Current extensive datasets for face generation leverage portrait images from Google or stock photo repositories like Flickr, which feature images with upfront views and minimal facial rotations. Face reenactment methods leverage video datasets such as Vox-Celeb [29], primarily featuring frames with frontal faces. The instability of pre-processing tools, particularly face alignment modules when handling extreme pose images, further reduces the proportion of those images in the existing datasets. This issue is critical since faces with large head rotations are common in real life, and dealing with them is unavoidable in practical systems.",
+        "bbox": [
+            75,
+            138,
+            470,
+            396
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "This paper aims to fill that void by proposing a novel large-scale dataset of extreme-pose and high-quality facial images called ExtremePose-Face-HQ, or EFHQ for short. Our ambition is to build an \"one-for-all\" dataset that can supplement existing datasets on all mentioned face-related tasks. To achieve that goal, the dataset needs all the following properties: (1) extreme-pose, (2) large-scale, (3) high-quality, (4) in-the-wild, (5) having multiple images for each subject and having subject identity annotated. Given those requirements, we utilize a novel and meticulous dataset processing pipeline to filter and sample frames from two publicly available datasets, including VFHQ [48] and CelebV-HQ [54], which contain high-resolution face videos captured in various settings. This results in a dataset with 450k high-quality images of extreme poses, which are combined with frontal images of the same subject to use as image pairs. These image pairs can support face reenactment and face recognition tasks, while the extreme-pose images alone can complement the existing datasets of face generation training. Note that there is a recent attempt [47] to build a large-pose facial dataset similar to ours. However, that work relies on images crawled from Flickr, thus having limited size (19k images) and missing identity information. Hence, its applications are restricted, covering only several face-generation tasks, unlike our large-scale and multi-purpose dataset.",
+        "bbox": [
+            75,
+            398,
+            470,
+            790
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We perform a comprehensive set of experiments to validate the effectiveness of EFHQ across three distinct face generation sub-tasks and face reenactment scenarios. Our dataset, EFHQ, when combined with standard training sets, significantly enhances the quality of synthesized images in highly profile or pitched views while maintaining synthesis quality in frontal views. EFHQ also provides a more",
+        "bbox": [
+            75,
+            795,
+            468,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "challenging pose-centric face verification benchmark. It reveals the vulnerability of many state-of-the-art face recognition networks when dealing with extreme-pose images, with TAR@FAR=1e-3 scores dropping 5-37%.",
+        "bbox": [
+            496,
+            90,
+            890,
+            151
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In summary, our contributions include (1) a large-scale, high-quality, in-the-wild, extreme-posed dataset (EFHQ) designed to complement various face-related tasks, (2) a novel and meticulous dataset processing pipeline ensuring the creation and quality of the EFHQ dataset, and (3) a series of experiments, with plan-to-release models, demonstrating the usefulness of EFHQ in enhancing performance on a wide range of tasks when dealing with faces at extreme rotations. The final dataset will be publicly available as metadata files to not violate the two original dataset's licenses, alongside the pretrained networks.",
+        "bbox": [
+            496,
+            152,
+            892,
+            316
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            332,
+            640,
+            347
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Synthetic Face Generation. Existing 2D-based models like StyleGAN [14, 15, 17] and 3D-aware models like EG3D [5] demonstrate impressive fidelity, especially for near-frontal views. Nevertheless, their performance significantly degrades when faced with extreme views, primarily due to insufficient training data for such scenarios. Datasets like FFHQ [14] exhibit a notable bias toward near-frontal poses, heavily limiting the range of poses synthesized by current 2D and 3D models despite the 3D-aware generators' ability to handle arbitrary poses.",
+        "bbox": [
+            496,
+            364,
+            890,
+            515
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Diffusion-based text-to-image models [32, 33] can synthesize high-quality images from simple text inputs. However, these models lack mechanisms to control the generated pose. Consequently, the generated facial images remain in a frontal view despite terms like \"profile\" or \"looking-up\" in the input prompts. ControlNet[49] proposes a finetuning approach allowing these models to take additional context as conditional input. Namely, one can use landmarks to guide the generated face to the desired pose. However, the scarcity of profile-view data in training affects the model's performance, particularly in extreme angles. Consequently, effectively managing the long-tailed distribution of these extreme views becomes necessary to address this issue.",
+        "bbox": [
+            496,
+            516,
+            892,
+            712
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Face Reenactment. Face reenactment involves transferring facial expressions and movements from a driving image to a source image. Recent advancements claim high-quality results, especially for near-frontal poses [31, 41, 43, 51]. However, pose collapse persists, particularly in keypoint-based models [41, 51]. Common datasets for this task are VoxCeleb1 [29] and HDTF [50], which primarily feature frontal talking face videos. Despite recent potentially high-quality datasets for reenactment such as VFHQ [48] and CelebV-HQ [54], the frontal-view bias persists. As a result, facial reenactment methods often produce artifacts and lower-quality outputs when faced with significant rotations.",
+        "bbox": [
+            496,
+            719,
+            892,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "22606",
+        "bbox": [
+            478,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/beb48d3203d2a90d8ab0890319d86d2fe69993915e96900069cc7e6ecfa0a320.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td></td><td colspan=\"3\">Attributes</td><td colspan=\"3\">Ready-to-use tasks</td></tr><tr><td>#images</td><td>Multiple images per ID</td><td>High Quality</td><td>Profile View</td><td>Face Verification</td><td>Face Synthesis</td><td>Face Reenactment</td></tr><tr><td>AgeDB</td><td>12K</td><td>✓</td><td></td><td></td><td>✓</td><td></td><td></td></tr><tr><td>CPLFW</td><td>12K</td><td>✓</td><td></td><td>✓</td><td>✓</td><td></td><td></td></tr><tr><td>CFP-FP</td><td>14K</td><td>✓</td><td></td><td>✓</td><td>✓</td><td></td><td></td></tr><tr><td>IJB-x</td><td>138K</td><td>✓</td><td></td><td>✓</td><td>✓</td><td></td><td></td></tr><tr><td>CelebA-HQ</td><td>30K</td><td></td><td>✓</td><td></td><td></td><td>✓</td><td></td></tr><tr><td>FFHQ</td><td>70K</td><td></td><td>✓</td><td></td><td></td><td>✓</td><td></td></tr><tr><td>LPFF</td><td>19K</td><td></td><td>✓</td><td>✓</td><td></td><td>✓</td><td></td></tr><tr><td>CelebV-HQ</td><td>6M</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>✓</td><td></td></tr><tr><td>VFHQ</td><td>3M</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>✓</td><td>✓</td></tr><tr><td>VoxCeleb1</td><td>4M</td><td>✓</td><td></td><td></td><td></td><td>✓</td><td>✓</td></tr><tr><td>Ours</td><td>450,538</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr><tr><td>FFHQ+Ours*</td><td>112,671</td><td></td><td>✓</td><td>✓</td><td></td><td>✓</td><td></td></tr><tr><td>Vox+Ours*</td><td>4.5M</td><td>✓</td><td></td><td>✓</td><td></td><td></td><td>✓</td></tr></table>",
+        "bbox": [
+            158,
+            88,
+            803,
+            296
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Representative comparison of key attributes and supported tasks across existing face datasets and our proposed EFHQ dataset. EFHQ provides large-scale, high-quality images with multiple views per identity, enabling diverse face-related tasks, including synthesis, reenactment, and verification. Combining EFHQ with existing datasets like FFHQ or VoxCeleb can augment available data for selected tasks. Asterisk (*) indicates a combined dataset with a subsampled version of EFHQ.",
+        "bbox": [
+            75,
+            301,
+            890,
+            357
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Face Verification. Despite near-perfect accuracy on benchmark datasets[12, 20, 27, 28, 44], state-of-the-art face verification models[7, 13, 18] fall short in unconstrained environments due to a lack of pose diversity. While CPLFW[52] introduces 6000 cross-pose pairs, its small size and mix of pose samples hinder analyzing performance on specific pose scenarios such as frontal-frontal, frontal-profile, and profile-profile. More granular benchmarking on diverse pose pairs is crucial to identify failure modes and drive progress in robust face verification across poses.",
+        "bbox": [
+            75,
+            369,
+            468,
+            521
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Extreme Pose Face Dataset. Tab. 1 summarizes existing datasets related to face verification, synthesis, and re-enactment, showing that existing face datasets have limitations for multi-task use. Recent efforts like LPFF[47] made progress by crawling from Flickr and processing for diverse views, resulting in a dataset of 19,590 images. The dataset could complement high-quality datasets like FFHQ[14] for generative models. However, due to its limited scale comprising only  $20\\%$  of the combined dataset, the unbalanced distribution between frontal and profile view still needs further improvements. Additionally, since the images are from image hosting services, the dataset fails to provide the identity information necessary for other facial applications.",
+        "bbox": [
+            75,
+            527,
+            468,
+            724
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Dataset Process",
+        "text_level": 1,
+        "bbox": [
+            76,
+            739,
+            232,
+            756
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we discuss the process of producing our novel dataset EFHQ. To fulfill our goals, we carefully curated frames from the two recent facial video datasets, VFHQ[48] and CelebV-HQ[54], as seen in Fig. 2.",
+        "bbox": [
+            75,
+            766,
+            468,
+            828
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Attributes Preparation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            838,
+            287,
+            854
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Attributes Extraction. To curate accurate facial poses and high-quality frames from existing datasets, we first ex",
+        "bbox": [
+            75,
+            869,
+            468,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "tract (1) face bounding boxes, (2) facial landmarks (5/68 keypoints), (3) image quality score, and (4) face identity. For bounding boxes and landmarks, we employ the popular RetinaFace[36] and SynergyNet[46]. We use HyperIQA [40] to grade image quality. For VFHQ, we only generate 68 landmarks and image quality scores since bounding boxes and 5 keypoints are provided. After extracting these attributes, we match them with existing annotations using an IoU-based Hungarian matching algorithm [21]. Regarding identity, VFHQ-wise, we source labels from the annotations. As CelebV-HQ lacks individual identity labels within the video, we use[35] to identify the identities associated with each bounding box.",
+        "bbox": [
+            496,
+            369,
+            890,
+            566
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Due to the instability of head pose estimators when handling extreme-pose images, we strategically apply multiple pose estimators relying on various methodologies: (1) SynergyNet[46], a 3DMM-based model, (2) DirectMHP[53], a landmark-free joint head detector and pose estimator, and (3) an in-house FacePoseNet[6] improved through extensive data augmentation focused on extreme poses and extra training data. We then categorize each estimated pose into a hierarchical binning scheme illustrated in Fig. 2. Next, we gather bin predictions and perform majority voting to arrive at consensus labels. When no agreement emerges, we put it into the  $5^{th}$  bin for \"confusing\" cases. As depicted in Fig. 3, this ensemble approach provides robustness in cases where a single estimator incorrectly categorizes a sample. Please refer to the supplementary for statistics of the extracted bins and the hyperparameters for each model used for attribute extraction.",
+        "bbox": [
+            496,
+            570,
+            890,
+            828
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Annotation Review. Next, we manually review the results to ensure high-quality labels. Given the large dataset, we developed a graphical user interface tool to streamline the review process. Our focus is verifying the binning annota",
+        "bbox": [
+            496,
+            839,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "22607",
+        "bbox": [
+            478,
+            944,
+            517,
+            957
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/f254cce2dcbe27b3334c49712a16370a9280f177f1bfe4f41e39c028a6f522d5.jpg",
+        "image_caption": [
+            "Figure 2. The pipeline of EFHQ dataset creation. Starting with high-quality videos from the VFHQ[48] and CelebV-HQ[54] datasets, single-frame attributes are extracted then manually reviews. Task-specific preprocessing is then applied to generate specialized versions of the dataset for tasks such as face generation, reenactment, and verification."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            158,
+            79,
+            807,
+            256
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/38c77dbb4c806449587b476c0b22ce3f7a86ea80b9d0b108ab1e63be1859d8d5.jpg",
+        "image_caption": [
+            "Figure 3. Example cases where a pose estimator fails to categorize the sample to the correct bin."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            321,
+            447,
+            448
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "tions, which most directly impact label accuracy. Concurrently, we randomly subsample the data based on pose angle and image quality to validate facial landmark quality further and discard cases of landmark prediction failure.",
+        "bbox": [
+            75,
+            496,
+            468,
+            556
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. EFHQ dataset",
+        "text_level": 1,
+        "bbox": [
+            76,
+            571,
+            225,
+            587
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our final EFHQ dataset comprises up to 450k frames with extreme poses extracted from approximately 5,000 clips. Most of the clips in our dataset include at least one frame with a frontal face and multiple frames with extreme pose angles. We also include extreme pose-only clips, representing profile-to-profile pose transfer cases.",
+        "bbox": [
+            75,
+            595,
+            468,
+            688
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/caec04bb3476c53e27cfcf754fc7da82c83f7eb861fc3f48a721d325f51ab89b.jpg",
+        "image_caption": [
+            "Figure 4. Pose distribution comparison between our sampled dataset and other datasets, including FFHQ[14], LPFF[47], CPLFW[52] and 40K random samples from VoxCeleb1 [29]. Our sampled dataset demonstrates greater pose diversity, with increased sample counts across high angle bins."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            705,
+            455,
+            825
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Dataset for subtasks",
+        "text_level": 1,
+        "bbox": [
+            500,
+            321,
+            691,
+            335
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Supplementary datasets for face generation. To address FFHQ's pose distribution gap, we compile a dataset encompassing diverse poses, varied identities, and FFHQ-comparable image quality from EFHQ. Additionally, we also embed an image brightness filter[3] in the sampling process, given that HyperIQA doesn't explicitly assess this aspect. The sampled images are then processed using the same pipeline in [5, 16] and manually reviewed. The aim is to improve coverage of under-represented poses without degrading frontal performance of trained generation models. This yields a dataset of 42,671 images with equitably distributed poses as visualized in Fig. 4. Compared to the original FFHQ distribution, yaw and pitch distributions are significantly improved.",
+        "bbox": [
+            496,
+            358,
+            890,
+            570
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Additionally, for 3D-aware GAN subtasks, we extract camera parameters to serve as conditional pose information, following [5]. We convert these parameters into yaw, pitch angles to validate against our existing annotations. When disagreements emerge, we re-examine the images and filter if needed, to ensure accurate, quality labels.",
+        "bbox": [
+            496,
+            575,
+            890,
+            667
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For diffusion-based text-to-image generation, a specialized dataset is curated to refine Stable Diffusion models[33] by accommodating landmark conditional input through ControlNet[49]. Based on our 3D-aware face generation dataset, we integrate landmark-based conditional images and tailored text prompts for each image. To capture fine facial nuances, we extract 478 facial landmarks and draw the condition image with connected edges between matched keypoints, following Mediapipe framework[26], better-conveying face expression and pose. For text prompts, detailed facial attributes like gender, race, and emotion are derived from existing labels or inferred via BLIP-2 pretrained captioning model [23] and a face attribute estimator [37]. The resulting prompt adheres to a structured format: \"A profile portrait image of a [emotion] [race] [gender].\"",
+        "bbox": [
+            496,
+            674,
+            892,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "22608",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/867f1367cf0d97dea13112a5cf0db97b2e1330fb65f9182a4c2e564667591df6.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Reference Dataset</td><td>FID↓</td><td>Recall↑</td></tr><tr><td>GFFHQ</td><td>FFHQ</td><td>2.84</td><td>0.49</td></tr><tr><td>GFFHQ+LPFF</td><td>FFHQ+LPFF</td><td>3.43</td><td>0.44</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ+EFHQ</td><td>3.44</td><td>0.46</td></tr><tr><td>GFFHQ+LPFF</td><td>FFHQ+EFHQ</td><td>8.26</td><td>0.43</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ+LPFF</td><td>7.04</td><td>0.44</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ+EFHQ</td><td>3.33</td><td>0.46</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ</td><td>3.12</td><td>0.44</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ+LPFF</td><td>4.47</td><td>0.42</td></tr></table>",
+        "bbox": [
+            78,
+            88,
+            470,
+            258
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 2. StyleGAN2-ADA models trained on different datasets. Lower FIDs indicate better fidelity, while higher Recalls indicate better diversity. The second and third blocks show comparisons between unconditional models trained on FFHQ+LPFF and FFHQ+EFHQ when evaluating on the same and cross-dataset settings, with the better metrics in bold.",
+        "bbox": [
+            75,
+            263,
+            468,
+            348
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Supplementary dataset for face reenactment. As EFHQ is crafted, we can employ the entire dataset to complement established training datasets, e.g., VoxCeleb1 [29]. We also built an extra evaluation set with 1200 EFHQ clips.",
+        "bbox": [
+            75,
+            354,
+            468,
+            417
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Benchmarking dataset for face verification. We curate a dataset covering three distinct scenarios: frontal-to-frontal, frontal-to-profile, and profile-to-profile. To enable rigorous benchmarking, we sample 10,000 pairs each for both negative and positive cases per scenario, resulting in a balanced benchmark dataset containing 60,000 image pairs. Images follow the same established preprocessing pipelines in [7]. Next, we filter misaligned images with RetinaFace[36], randomly review, and replace equivalent samples from our diverse corpus, if needed. Finally, to simulate varying image quality, we randomly applied downscaling and compression to samples in the dataset.",
+        "bbox": [
+            75,
+            424,
+            468,
+            604
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Face Generation Subtask",
+        "text_level": 1,
+        "bbox": [
+            76,
+            623,
+            313,
+            638
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. StyleGAN",
+        "text_level": 1,
+        "bbox": [
+            76,
+            648,
+            191,
+            666
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training details. We train StyleGAN2-ADA[15] models from scratch after combining FFHQ with our own curated dataset (denoted FFHQ+EFHQ) at  $1024 \\times 1024$  resolution. We also train a conditional version, allowing direct control over frontal versus profile synthesis and enhancing model performance. Specifically, we categorize the dataset into two classes: \"frontal\" and \"profile\" based on pose angles. All models are trained for 6 days on 8 Nvidia A100 GPUs.",
+        "bbox": [
+            75,
+            680,
+            467,
+            800
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Evaluation protocol. We utilize the Frechet Inception Distance[11] (FID) metric and improved Recall[22] to quantify the generation fidelity and diversity. Regarding the two unconditional models trained with the supplemented FFHQ+LPFF and FFHQ+EFHQ, we evaluate each model with both of these datasets. For conditional mod",
+        "bbox": [
+            75,
+            809,
+            467,
+            898
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/3dcea56d5362a57eb679bd30469ce22064230f4b8d123e344efe25aaf2dc37ea.jpg",
+        "image_caption": [
+            "(a) Samples from the model trained with FFHQ+EFHQ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            532,
+            88,
+            859,
+            174
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/3c0d5c9941056f9827add44aae19a2e757bb2b54eaf548fa151058eafdd04f75.jpg",
+        "image_caption": [
+            "(b) Profile-view comparison of LPFF (left) and EFHQ (right).",
+            "Figure 5. Qualitative result and comparison of generated samples, with truncation  $\\psi = 0.7$ , from StyleGAN2-ADA training with other dataset and ours."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            531,
+            190,
+            857,
+            349
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "els, we evaluate performance on FFHQ, FFHQ+LPFF, and FFHQ+EFHQ to demonstrate the broader coverage of the model trained with our proposed dataset. For FFHQ, we generate 50K images with the \"frontal\" label, whereas for the other two datasets, we generate 50K following their label distributions. This comprehensive quantitative analysis examines the fidelity and variety of the generated images.",
+        "bbox": [
+            496,
+            431,
+            890,
+            537
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Experimental results. We compared the performance of our proposed unconditional models  $(G^{\\mathrm{FFHQ + EFHQ}})$ , against models trained on the FFHQ and FFHQ+LPFF datasets  $(G^{\\mathrm{FFHQ}}$  and  $G^{\\mathrm{FFHQ + LPFF}})$ . Models trained conditionally are specifically subscripted, e.g.,  $G_{c}^{\\mathrm{FFHQ + EFHQ}}$ . The qualitative results in Fig. 5 illustrate that our models generate high-quality frontal faces comparable to the FFHQ model while producing realistic and varied profile faces. Additionally, our method has less noise in extreme poses than the FFHQ+LPFF model. Specifically, Fig. 5b shows LPFF's generated face contains more noise and less photorealistic details than our model, such as the pixelated eye and nose patches. Meanwhile, ours achieves greater realism with smoother skin, well-defined features, noise reduction, and natural lighting. The quantitative results in Tab. 2 show that our unconditional model achieves competitive FID scores and improved Recall to other models, indicating similar fidelity but greater sample pose diversity. In the cross-examination (third block of Tab. 2),  $G^{\\mathrm{FFHQ + EFHQ}}$  has both better FID and improved Recall on the opposite dataset than  $G^{\\mathrm{FFHQ + LPFF}}$ . Finally, our conditional model demonstrates substantially improved FID scores over multiple reference datasets, highlighting its versatility.",
+        "bbox": [
+            496,
+            553,
+            890,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "22609",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/44f3155bf285746e0ba7966b8c0a6c4ed50d0cc4776b7446c00c1152e1796d93.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Reference Dataset</td><td>FID↓</td><td>Pose↓</td><td>Depth↓</td><td>ID↑</td></tr><tr><td>EFHQ</td><td>FFHQ/Trainset</td><td>4.8</td><td>5e-4</td><td>0.234</td><td>0.74</td></tr><tr><td>EFFHQ</td><td>FFHQ/Trainset</td><td>3.7</td><td>5e-4</td><td>0.288</td><td>0.74</td></tr><tr><td>EFFHQ+LPFF</td><td>FFHQ</td><td>6.2</td><td>6e-4</td><td>0.316</td><td>0.77</td></tr><tr><td>EFFHQ+EFHQ</td><td>FFHQ</td><td>3.9</td><td>4e-4</td><td>0.224</td><td>0.77</td></tr><tr><td>EFFHQ+LPFF</td><td>Trainset</td><td>7.0</td><td>6e-4</td><td>0.316</td><td>0.65</td></tr><tr><td>EFFHQ+EFHQ</td><td>Trainset</td><td>4.6</td><td>4e-4</td><td>0.223</td><td>0.65</td></tr></table>",
+        "bbox": [
+            80,
+            88,
+            467,
+            205
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/8e70a13b656bbde69089723c284a0af97e6cca6c1c51f5d55a61b3fe1f63ec4d.jpg",
+        "image_caption": [
+            "Figure 6. Comparison between multiview generated samples, with truncation  $\\psi = 0.8$ , of EG3D model trained with various datasets."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            88,
+            294,
+            455,
+            457
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. EG3D",
+        "text_level": 1,
+        "bbox": [
+            76,
+            507,
+            161,
+            522
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Training details. We followed the same pipeline and hyper-parameters as the original EG3D model [5]. However, based on empirical results, we adjusted  $\\gamma = 5$ . We also fixed the triplane implementation bug (XY, XZ, ZX) [1] to replicate the original idea in the papers (XY, XZ, ZY). We trained the model on 8 Nvidia A100 40GB GPUs for approximately 7 days until convergence.",
+        "bbox": [
+            75,
+            537,
+            467,
+            643
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Protocol. To evaluate the models, mirroring [5], we measure image quality with FID; multi-view consistency and identity preservation with mean cosine similarity of pre-trained face recognition model[7]; pose accuracy and geometry by calculating MSE against pseudo labels from [8]. We evaluate all models on 1,024 generated images, except for FID, which we compute on 50k images.",
+        "bbox": [
+            75,
+            651,
+            467,
+            758
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Experimental results. To evaluate the benefits of our dataset, we trained the same backbone architecture on different datasets: FFHQ  $(E^{\\mathrm{FFHQ}})$ , FFHQ+LPFF  $(E^{\\mathrm{FFHQ} + \\mathrm{LPFF}})$ , and FFHQ+EFHQ (our model,  $E^{\\mathrm{FFHQ} + \\mathrm{EFHQ}}$ ). The quantitative results are presented in Table 3, and qualitative examples are shown in Figure 6. Qualitatively, we observe that  $E^{\\mathrm{FFHQ} + \\mathrm{EFHQ}}$  generates improved face shapes in extreme poses compared to  $E^{\\mathrm{FFHQ}}$ , similar to the improvements seen with  $E^{\\mathrm{FFHQ} + \\mathrm{LPFF}}$ . Quantitatively, the FID",
+        "bbox": [
+            75,
+            763,
+            467,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/7cff13cf3cc12be33b6c195c7d3f103525cc0601828cf4e62e89e6bc85fa9fff.jpg",
+        "table_caption": [
+            "Table 3. EG3D models trained on different datasets. In the second and third blocks, we compare models trained on FFHQ+LPFF and FFHQ+EFHQ with the testset is either FFHQ or their training datasets; the better metrics are in bold. The best metrics on the FFHQ testset are in red, while the best metrics when evaluating on training dataset are underlined."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Metrics</td><td>Dataset</td><td>Pretrained ControlNet</td><td>Ours</td></tr><tr><td>NME↓</td><td>LPFF</td><td>0.085</td><td>0.084</td></tr><tr><td>FID↓</td><td>LPFF</td><td>119.74</td><td>83.39</td></tr><tr><td>FIDeyepatch ↓</td><td>LPFF</td><td>103.77</td><td>49.45</td></tr><tr><td>NME↓</td><td>EFHQ</td><td>0.106</td><td>0.081</td></tr><tr><td>FID↓</td><td>EFHQ</td><td>111.23</td><td>81.90</td></tr><tr><td>FIDeyepatch ↓</td><td>EFHQ</td><td>137.43</td><td>58.60</td></tr></table>",
+        "bbox": [
+            500,
+            88,
+            888,
+            217
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Quantitative results of fine-tuning StableDiffusion v1.5 with ControlNet trained on various datasets, using Normalized Mean Error (NME), FID, and FID on eye region.",
+        "bbox": [
+            498,
+            220,
+            890,
+            262
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "of  $E^{\\mathrm{FFHQ + EFHQ}}$  is comparable to  $E^{\\mathrm{FFHQ}}$  when evaluated on FFHQ, indicating no degradation in frontal view performance. Meanwhile, the competitive FID achieved by  $E^{\\mathrm{FFHQ + EFHQ}}$  on the combined FFHQ+EFHQ dataset also suggests proficient modeling of profile views. Regarding identity consistency, we see a significant drop when the reference dataset contains profile-view images, due to the weak confidence of the pretrained face recognition model when handling profile-view images, as confirmed by its low average cosine similarity of around 0.6 on the simple frontal-to-profile verification set CFP-FP[34]. Furthermore, our model also outperforms in pose accuracy and geometry quality with both frontal and cross-pose dataset.",
+        "bbox": [
+            498,
+            270,
+            890,
+            469
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/91c1c39d34e1ff469ab19168be92e10d78b69610e53bee1af6b1b4b5daca6041.jpg",
+        "image_caption": [
+            "(a) Generated samples; condition image is in between samples."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            513,
+            479,
+            694,
+            578
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/e2849aa6187df629ee39998db75c92f3603e153a68b749d5b56f72afd4fcceeb.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            696,
+            479,
+            880,
+            577
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/ea5572592c76c179b7a23caa798b0464f0d77a81bf0014c7a81b2a1c2f7a141b.jpg",
+        "image_caption": [
+            "(b) Generated eye regions."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            592,
+            697,
+            625
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/42ea89334a93b892c5276bce82ddb8d2426f524d79dcfa74a222ced094510ff8.jpg",
+        "image_caption": [
+            "Figure 7. Comparison profile-view generated samples of pretrained ControlNet (left) and our fine-tuned ControlNet (right) with the prompt: \"A profile portrait image of a person\"."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            699,
+            592,
+            893,
+            625
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. ControlNet",
+        "text_level": 1,
+        "bbox": [
+            500,
+            696,
+            622,
+            710
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Training details. We finetuned Stable Diffusion v1.5 [33] using ControlNet [49] on the EFHQ dataset to enhance its ability to generate extreme facial poses. We use AdamW optimizer [25] with learning rate  $1 \\times e^{-5}$ ,  $\\beta_0 = 0.9$ ,  $\\beta_1 = 0.999$ , and  $\\epsilon = 0.01$ . We train with batch size 24 on 1 Nvidia A100 40GB GPU in 2 days.",
+        "bbox": [
+            498,
+            726,
+            890,
+            816
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Protocol. To assess image qualities across diverse poses, we measure the Normalized Mean Error (NME) of 478-point facial landmarks between the generated and source image. Specifically, we evaluate performance across the full pose distribution in the reference",
+        "bbox": [
+            496,
+            824,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "22610",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/0100f099836e506a48e568b5e3766bbb91f0c71b7470a5ec112ea03488b75695.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"2\">Test Dataset (L1 ↓/ AKD ↓/ AED ↓)</td></tr><tr><td>VoxCeleb1</td><td>EFHQ (Ours)</td></tr><tr><td>TPS(VoxCeleb1</td><td>0.041/1.214/0.132</td><td>0.061/3.814/0.275</td></tr><tr><td>TPS(VoxCeleb1+EFHQ</td><td>0.041/1.250/0.133</td><td>0.051/1.963/0.214</td></tr><tr><td>LIA(VoxCeleb1</td><td>0.044/1.342/0.138</td><td>0.059/3.807/0.266</td></tr><tr><td>LIA(VoxCeleb1+EFHQ</td><td>0.044/1.398/0.142</td><td>0.054/2.291/0.233</td></tr></table>",
+        "bbox": [
+            81,
+            88,
+            470,
+            175
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "dataset. For the prompt, we use: \"A profile portrait image of a person.\" with extra positive and negative keywords, further detailed in supplementary. Utilizing LPFF and EFHQ datasets as the reference datasets, we compute FID to assess synthesized face quality and extend evaluation to cropped eye regions due to unrealistic gazes observed from the original ControlNet model. Note that the models receive condition images created akin to its training data.",
+        "bbox": [
+            75,
+            241,
+            468,
+            363
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Experimental results. Fig. 7a visually compares the baseline ControlNet OpenPose face and our fine-tuned model. Our fine-tuned model adeptly conditions diverse extreme poses, generating high-quality results with precise details while maintaining fidelity to the condition, such as mouth shape, and exhibiting minimal artifacts. These patterns indicate the superiority of the fine-tuned model over the baseline model in handling extreme viewing angles. Additionally, Fig. 7b contrasts the eye regions of both models. Our fine-tuned model accurately captures natural eye gaze and shape from condition images, a notable improvement over the baseline. Quantitative results in Tab. 4 consistently favor our fine-tuned model across all metrics. FID scores on full and eye patches exhibit significant enhancement, and NME performance remains stable on training and reference datasets (EFHQ and LPFF). These findings showcase the fine-tuned model's capability to generate high-quality, geometrically consistent faces across diverse poses.",
+        "bbox": [
+            75,
+            369,
+            470,
+            642
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5. Face Reenactment Subtask",
+        "text_level": 1,
+        "bbox": [
+            76,
+            656,
+            328,
+            672
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1. Evaluation Protocol",
+        "text_level": 1,
+        "bbox": [
+            76,
+            680,
+            266,
+            695
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We follow the same preprocessing and evaluation protocol as FOMM [38]. Specifically, all the frames are cropped to  $256 \\times 256$ , and we use the first frame of a clip as the source image to reconstruct the remaining frames. In the case of the EFHQ test set, most source images have a frontal pose, while the target frames have an extreme pose. Some test cases only contain extreme pose to evaluate self-reconstruction on these cases. We run the experiments using two recent methods, Thin-Plate Spline Motion Model (TPS) [51] and Latent Image Animator (LIA) [43]. For training details, please refer to the supplementary. We select the following three metrics to compare the generated frames and the driving ground truth.  $\\mathcal{L}_1$  measures the av",
+        "bbox": [
+            75,
+            703,
+            470,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/6d045a3c5d3768b18d0d1ca71db35c934026da6a519eeaea4da966eb56060a00.jpg",
+        "table_caption": [
+            "Table 5. Quantitative results of two state-of-the-art reenactment methods (TPS and LIA) when training on VoxCeleb1 and the combined set VoxCeleb1+EFHQ. We present results evaluated either on VoxCeleb1 (left) or EFHQ (right) test sets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td rowspan=\"2\">Trainset</td><td rowspan=\"2\">CPLFW</td><td colspan=\"3\">EFHQ</td></tr><tr><td>f2f</td><td>f2p</td><td>p2p</td></tr><tr><td>R18arface</td><td>Glint360k</td><td>0.80</td><td>0.94</td><td>0.570.37</td><td>0.880.06</td></tr><tr><td>R18arface</td><td>MS1MV3</td><td>0.75</td><td>0.93</td><td>0.570.36</td><td>0.880.05</td></tr><tr><td>R50arface</td><td>Glint360k</td><td>0.91</td><td>0.97</td><td>0.770.20</td><td>0.920.05</td></tr><tr><td>R100arface</td><td>Glint360k</td><td>0.93</td><td>0.98</td><td>0.830.15</td><td>0.820.16</td></tr><tr><td>R100adaface</td><td>WebFace4M</td><td>0.94</td><td>0.99</td><td>0.940.05</td><td>0.970.02</td></tr></table>",
+        "bbox": [
+            503,
+            88,
+            888,
+            195
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 6. Quantitative evaluation of state-of-the-art face recognition models on the cleansed CPLFW and the proposed benchmark dataset. We use the metric TAR@FAR=1e-3 (higher is better). Our EFHQ benchmark includes frontal-frontal (f2f), frontal-profile (f2p), and profile-profile (p2p) subsets. For f2p and p2p, we report the score drop when comparing to f2f in the red subscription. The lowest benchmark score for each model is in bold.",
+        "bbox": [
+            496,
+            199,
+            890,
+            297
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "erage distance among pixel values, indicating reconstruction faithfulness. Average Keypoint Distance (AKD) calculates the mean distance between keypoints, thus assessing the pose accuracy. Average Euclidean Distance (AED) measures identity preservation by computing the average  $\\mathcal{L}_2$  distance between identity embeddings.",
+        "bbox": [
+            496,
+            306,
+            890,
+            398
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2. Experimental results",
+        "text_level": 1,
+        "bbox": [
+            498,
+            412,
+            696,
+            428
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We compare TPS and LIA models trained with our supplementary EFHQ dataset  $(TPS^{\\mathrm{VoxCeleb1 + EFHQ}}$  and  $LIA^{\\mathrm{VoxCeleb1 + EFHQ}})$  against models trained solely on VoxCeleb1  $(TPS^{\\mathrm{VoxCeleb1}}$  and  $LIA^{\\mathrm{VoxCeleb1}})$ . Tab. 5 displays quantitative results, revealing that models trained with EFHQ maintain comparable performance when evaluated on VoxCeleb1, emphasizing high-quality performance in frontal-to-frontal reenactment scenarios. Notably, on the EFHQ test set, both EFHQ-trained models show significant improvement in all metrics, indicating enhanced extreme pose and identity reenactment performance. In Fig. 8, qualitative results prove that EFHQ-trained models excel in transferring facial expressions and motion across diverse poses. Models trained on EFHQ showcase improved image quality, reduced artifacts, and well-preserved shapes of driving faces. These results highlight our models' effectiveness in capturing and reproducing facial motions across varying poses while enhancing fidelity to the original facial features.",
+        "bbox": [
+            496,
+            436,
+            892,
+            724
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6. Face Recognition Subtask",
+        "text_level": 1,
+        "bbox": [
+            498,
+            742,
+            741,
+            760
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.1. Evaluation Protocol",
+        "text_level": 1,
+        "bbox": [
+            500,
+            768,
+            689,
+            784
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To evaluate face recognition performance on extreme poses, we selected two state-of-the-art methods: ArcFace[7] and AdaFace[18]. We further explored various backbone sizes (ResNet18, ResNet50, ResNet101)[10] and training datasets (MS1MV2[9], Glint360K[2], WebFace4M[55]) of increasing size. For metrics, we used True Acceptance Rate (TAR) at False Acceptance Rate (FAR) of  $1 \\times e^{-3}$ .",
+        "bbox": [
+            496,
+            795,
+            890,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "22611",
+        "bbox": [
+            478,
+            944,
+            517,
+            957
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2c8117e4d7e86d126609d5accefe90974ed636e97faaf9d4022551304474fbb7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            168,
+            99,
+            264,
+            172
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/f69d710555708f6551cef3b5ce48de1fde2a208e53247211237ab7c6fbae5fcb.jpg",
+        "image_caption": [
+            "Source"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            168,
+            174,
+            267,
+            247
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/7d2912606c5b61da5c606f88306fd6e060536c5ef9c314d5d5374ab99e85b6c3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            277,
+            99,
+            374,
+            172
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/61c559132cc4890d912246b645d7129a8dff94b65997700f60ef27bdef7545eb.jpg",
+        "image_caption": [
+            "Figure 8. Comparison of performance for frontal-profile face reenactment on EFHQ. From left to right: source image, TPS on VoxCeleb1, TPS with EFHQ, LIA on VoxCeleb1, LIA with EFHQ, driving image. Additional EFHQ training data improves synthesis of extreme poses."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            277,
+            174,
+            375,
+            250
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/21f7cb68aa95c4ae50e04e75b16ffcc4a864c7551ef38fd4b4973bd55631b879.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            379,
+            99,
+            473,
+            172
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/74afe71aafafe1944aae621c88642f991e2738f3e52b03082731ebdf1d978f1f.jpg",
+        "image_caption": [
+            "TPS (VoxCeleb1)",
+            "TPS (Ours)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            379,
+            174,
+            473,
+            250
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/2eb2f40499fafbf4fcb78e429ecb6a3f128bea37332e4a23ac5950ae3a760e86.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            488,
+            99,
+            586,
+            172
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/13c7f831840d7ea36cbc2037bba4722099980fd3cb1850d9ce250d64185c67f0.jpg",
+        "image_caption": [
+            "LIA (VoxCeleb1)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            488,
+            174,
+            586,
+            247
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6d928c801997e6e85bd8d4dd1dab3f7ab8b88397fe154b822f5f89eac285434f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            591,
+            99,
+            687,
+            172
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a37e832c4e7938ac2972ac1b7a72c3542f56a1a39355e19829977df27ba063d8.jpg",
+        "image_caption": [
+            "LIA (Ours)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            591,
+            174,
+            687,
+            247
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/cf4b1d5c92581dc96cd7ada2ad8e39c5f6b911a5e50861f172bbd3c22fc7a3d0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            702,
+            99,
+            797,
+            172
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/227ecfb2eb75f91aa3aa305d5cb140c2676776284e15ef4c1a063540403f72ea.jpg",
+        "image_caption": [
+            "Driving"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            700,
+            174,
+            795,
+            247
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.2. Experimental results",
+        "text_level": 1,
+        "bbox": [
+            76,
+            309,
+            264,
+            325
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Before assessing face recognition networks on EFHQ, we first clean the commonly-used version of CPLFW, processed by[7], by removing misaligned images (please refer to the Supplementary), and evaluate the models on it. Models performed well on this cleansed set, with minimal gaps between ArcFace and AdaFace.",
+        "bbox": [
+            75,
+            332,
+            468,
+            422
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our proposed evaluation benchmark involves three pose-based scenarios. Results in Tab. 6 reveal performance drops of  $5 - 37\\%$  on the frontal-profile subset compared to frontal-frontal, highlighting the challenge of matching faces in different poses. Larger ResNet models outperformed ResNet18 across subsets, showcasing the advantages of increased capacity in handling pose variation. However, the ResNet100 model trained on Glint360K performed worse than smaller models on the profile-profile subset, indicating the need for sufficient profile data to prevent overfitting. Our results also indicate AdaFace's suitability for handling pose variations. In summary, pose mismatch brings significant challenges for face recognition, necessitating larger models trained on diverse data covering profile faces.",
+        "bbox": [
+            75,
+            422,
+            468,
+            635
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7. Discussion",
+        "text_level": 1,
+        "bbox": [
+            76,
+            646,
+            191,
+            660
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e17c16d7e1823307bbc687e505298164fa1f4d5649b691e9a0108114e17bb041.jpg",
+        "image_caption": [
+            "Figure 9. User study results. We reported the percentage of times testers rank whose output is best or tied for each criterion."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            667,
+            446,
+            813
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "User survey. We conduct user surveys to evaluate the face generation improvement of our method through human per",
+        "bbox": [
+            75,
+            869,
+            468,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "ceptual assessment. In this survey, we compare the outputs of the models mentioned in Secs. 4 and 5. The setup of this user survey is included in the supplementary. In all human evaluation results (Fig. 9), our models outperform previous ones, notably on identity and motion test of face reenactment models (74.80% and 64.22%, respectively).",
+        "bbox": [
+            496,
+            310,
+            890,
+            401
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5406b417968dda5059ac39836df61475119ae9f04474919d0df72f427ee2ccb5.jpg",
+        "image_caption": [
+            "Figure 10. Representative examples of the raw LPFF dataset that got excluded from the final dataset, either from misdetections or pose filtering process. Red keypoints represent landmarks prediction of selected samples from LPFF, green keypoints represent landmarks prediction from our pipeline."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            527,
+            419,
+            859,
+            506
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Robustness of data processing pipeline. As shown in Fig. 10, our data processing pipeline can detect additional profile-view faces filtered out by the LPFF pipeline, likely due to misdetections or incorrect pose filtering. This confirms the robustness of our pipeline, and we hope it will enable the community to develop larger and higher-quality in-the-wild facial datasets.",
+        "bbox": [
+            496,
+            594,
+            890,
+            699
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "8. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            720,
+            617,
+            734
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This work has introduced a large-scale, diverse facial dataset to address performance gaps between frontal and profile faces, powered by a novel and robust data processing pipeline. Crucially, we provided tailored sub-datasets for advancing essential tasks like face synthesis and reenactment. Moreover, our new face verification benchmark reveals gaps between techniques, granting valuable insights. Ultimately, we hope this high-quality, diverse facial dataset opens up new and exciting opportunities to push forward cross-pose tasks.",
+        "bbox": [
+            496,
+            747,
+            890,
+            887
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "22612",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            173,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Fix for triplane projection  $\\cdot$  Issue #67  $\\cdot$  NVlabs/eg3d. https://github.com/NVlabs/eg3d/issues/67, 2023.6",
+            "[2] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 7",
+            "[3] Sergey Bezryadin, Pavel Bourov, and Dmitry Ilinih. Brightness calculation in digital image processing. International Symposium on Technologies for Digital Photo Fulfillment, 2007:10-15, 2007. 4, 1",
+            "[4] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4",
+            "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 4, 6",
+            "[6] Feng-Ju Chang, Anh Tuan Tran, Tal Hassner, Iacopo Masi, Ram Nevatia, and Gerard Medioni. Faceposenet: Making a case for landmark-free face alignment. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1599-1608, 2017. 3",
+            "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 3, 5, 6, 7, 8, 1, 2",
+            "[8] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 6, 3",
+            "[9] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 1, 7",
+            "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778. IEEE, 2016. 7",
+            "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 5",
+            "[12] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 1, 3"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[13] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 1, 3",
+            "[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3, 4",
+            "[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Proc. NeurIPS, 2020. 1, 2, 5",
+            "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 4",
+            "[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 2",
+            "[18] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 7",
+            "[19] Davis E. King. Dlib-ml: A machine learning toolkit. J. Mach. Learn. Res., 10:1755-1758, 2009. 3",
+            "[20] Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1931-1939, 2015. 3",
+            "[21] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3",
+            "[22] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Proceedings of the 33rd International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2019. Curran Associates Inc. 5",
+            "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 4",
+            "[24] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 1",
+            "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 6",
+            "[26] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Wan-Teh"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "22613",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for building perception pipelines. CoRR, abs/1906.08172, 2019. 4",
+            "[27] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 3",
+            "[28] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 1, 3",
+            "[29] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. VoxCeleb: A large-scale speaker identification dataset. In Interspeech 2017. ISCA, 2017. 2, 4, 5",
+            "[30] Aaron Nech and Ira Kemelmacher-Shlizerman. Level playing field for million scale face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1",
+            "[31] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-Ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 427-436, 2023. 2",
+            "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv e-prints, pages arXiv-2204, 2022. 2",
+            "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 6",
+            "[34] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 6",
+            "[35] Sefik Ilkin Serengil and Alper Ozpinar. Lightface: A hybrid deep face recognition framework. In 2020 Innovations in Intelligent Systems and Applications Conference (ASYU), pages 23-27. IEEE, 2020. 3",
+            "[36] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 3, 5, 1, 2",
+            "[37] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 4",
+            "[38] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 7, 2"
+        ],
+        "bbox": [
+            78,
+            90,
+            467,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[39] L Sirovich and M Kirby. Low-dimensional procedure for the characterization of human faces. Journal of the Optical Society of America A, 4(3):519-524, 1987. 1",
+            "[40] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3",
+            "[41] Jiale Tao, Biao Wang, Borun Xu, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Structure-aware motion transfer with deformable anchor model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3637-3646, 2022. 2",
+            "[42] Matthew A Turk and Alex P Pentland. Face recognition using eigenfaces. In Proceedings. 1991 IEEE computer society conference on computer vision and pattern recognition, pages 586-587. IEEE Computer Society, 1991. 1",
+            "[43] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In ICLR 2022-The International Conference on Learning Representations, 2022. 2, 7, 3",
+            "[44] Cameron Whitelam, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K Jain, James A Duncan, Kristen Allen, et al. Iarpa janus benchmark-b face dataset. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 90-98, 2017. 3",
+            "[45] L Wolf, T Hassner, and I Maoz. Face recognition in unconstrained videos with matched background similarity. In Proceedings of the 2011 IEEE Conference on Computer Vision and Pattern Recognition, pages 529-534, 2011. 1",
+            "[46] Cho-Ying Wu, Qiangeng Xu, and Ulrich Neumann. Synergy between 3dmm and 3d landmarks for accurate 3d facial geometry. In 2021 International Conference on 3D Vision (3DV), 2021. 3",
+            "[47] Yiqian Wu, Jing Zhang, Hongbo Fu, and Xiaogang Jin. Lpff: A portrait dataset for face generators across large poses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20327-20337, 2023. 1, 2, 3, 4",
+            "[48] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2022. 1, 2, 3, 4",
+            "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 4, 6",
+            "[50] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 2",
+            "[51] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Con"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "22614",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "ference on Computer Vision and Pattern Recognition, pages 3657-3666, 2022. 2, 7, 3",
+            "[52] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7), 2018. 3, 4, 2",
+            "[53] Huayi Zhou, Fei Jiang, and Hongtao Lu. Directmhp: Direct 2d multi-person head pose estimation with full-range angles. arXiv preprint arXiv:2302.01110, 2023. 3",
+            "[54] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 1, 2, 3, 4",
+            "[55] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 7"
+        ],
+        "bbox": [
+            78,
+            92,
+            468,
+            359
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "22615",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_model.json

@@ -0,0 +1,2279 @@
+[
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.107,
+                0.003,
+                0.182,
+                0.043
+            ],
+            "angle": 0,
+            "content": "CVF"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.237,
+                0.001,
+                0.812,
+                0.047
+            ],
+            "angle": 0,
+            "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.214,
+                0.131,
+                0.758,
+                0.154
+            ],
+            "angle": 0,
+            "content": "EFHQ: Multi-purpose ExtremePose-Face-HQ dataset"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.202,
+                0.19,
+                0.767,
+                0.209
+            ],
+            "angle": 0,
+            "content": "Trung Tuan Dao\\(^{1,*}\\) Duc Hong Vu\\(^{1,2,*}\\) Cuong Pham\\(^{1,3}\\) Anh Tran\\(^{1}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.175,
+                0.211,
+                0.797,
+                0.225
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)VinAI Research, Vietnam \\(^{2}\\)Miami University \\(^{3}\\)Posts & Telecommunications Inst. of Tech., Vietnam"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.23,
+                0.83,
+                0.244
+            ],
+            "angle": 0,
+            "content": "{v.trungdt21,v.ducvh5,v.cuongpv11,v.anhtt152}@vinai.io vudh3@miamioh.edu cuongpv@ptit.edu.vn"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.424,
+                0.246,
+                0.543,
+                0.261
+            ],
+            "angle": 0,
+            "content": "*Equal contribution"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.145,
+                0.293,
+                0.832,
+                0.466
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.473,
+                0.896,
+                0.586
+            ],
+            "angle": 0,
+            "content": "Figure 1. Benefits of our proposed dataset (EFHQ). Standard large-scale facial datasets have most images at near frontal views, causing inferior performance of trained models on downstream tasks when dealing with extreme head poses. For instance, the trained 2D image generators and text-to-image ones often produce only near frontal faces, while the 3D face generators and face reenactment methods often show distorted outputs at profile views. The recently proposed dataset LPFF [47] partially handles that issue by providing complementary images at extreme head poses for only 2D and 3D image generation tasks. Our proposed dataset EFHQ provides high-quality extreme-pose images to complement a wide range of face-related tasks. It supports 2D and 3D image generation, with generally better diversity than LPFF. EFHQ also helps correct the outputs of text-to-image generation and face reenactment at extreme views. Finally, EFHQ provides a more challenging pose-based face verification benchmark to better assess the quality of face recognition networks."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.235,
+                0.598,
+                0.314,
+                0.613
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.075,
+                0.629,
+                0.474,
+                0.903
+            ],
+            "angle": 0,
+            "content": "The existing facial datasets, while having plentiful images at near frontal views, lack images with extreme head poses, leading to the downgraded performance of deep learning models when dealing with profile or pitched faces. This work aims to address this gap by introducing a novel dataset named Extreme Pose Face High-Quality Dataset (EFHQ), which includes a maximum of 450k high-quality images of faces at extreme poses. To produce such a massive dataset, we utilize a novel and meticulous dataset processing pipeline to curate two publicly available datasets, VFHQ and CelebV-HQ, which contain many high-resolution face videos captured in various settings. Our dataset can complement existing datasets on various facial-related tasks, such as facial synthesis with 2D/3D-aware GAN, diffusion-based text-to-image face generation, and face reenactment. Specifically, training with EFHQ helps models generalize well across diverse poses, significantly improving performance in scenarios involving ex"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.6,
+                0.896,
+                0.69
+            ],
+            "angle": 0,
+            "content": "treme views, confirmed by extensive experiments. Additionally, we utilize EFHQ to define a challenging cross-view face verification benchmark, in which the performance of SOTA face recognition models drops \\(5 - 37\\%\\) compared to frontal-to-frontal scenarios, aiming to stimulate studies on face recognition under severe pose conditions in the wild."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.501,
+                0.723,
+                0.633,
+                0.739
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.75,
+                0.895,
+                0.903
+            ],
+            "angle": 0,
+            "content": "The human face is a central focus in computer vision, with extensive research dedicated to tasks like detection, recognition, generation, and manipulation [39, 42]. Numerous large-scale facial datasets[9, 12, 14, 24, 28, 30, 45, 48, 54, 55] have facilitated the face-related studies. Along with deep learning techniques, many tasks have made a considerable leap in performance in recent years. For instance, face recognition systems can recognize near-frontal faces in the wild with near-perfect accuracy [13, 18]. Facial generative models, such as [5, 15, 49], can synthesize realistic"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.52,
+                0.958
+            ],
+            "angle": 0,
+            "content": "22605"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.092,
+                0.47,
+                0.137
+            ],
+            "angle": 0,
+            "content": "images indistinguishable from real faces. Face reenactment systems, including [41, 43, 51], can animate an input face based on poses and expressions from any driver video."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.14,
+                0.471,
+                0.397
+            ],
+            "angle": 0,
+            "content": "Despite impressive results, the mentioned models produce inferior results when dealing with faces at extreme head poses. As illustrated in Fig. 1, faces synthesized by 3D face generation or face reenactment models are often distorted at profile views, due to the scarcity of extreme pose images in training datasets. Current extensive datasets for face generation leverage portrait images from Google or stock photo repositories like Flickr, which feature images with upfront views and minimal facial rotations. Face reenactment methods leverage video datasets such as Vox-Celeb [29], primarily featuring frames with frontal faces. The instability of pre-processing tools, particularly face alignment modules when handling extreme pose images, further reduces the proportion of those images in the existing datasets. This issue is critical since faces with large head rotations are common in real life, and dealing with them is unavoidable in practical systems."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.399,
+                0.471,
+                0.791
+            ],
+            "angle": 0,
+            "content": "This paper aims to fill that void by proposing a novel large-scale dataset of extreme-pose and high-quality facial images called ExtremePose-Face-HQ, or EFHQ for short. Our ambition is to build an \"one-for-all\" dataset that can supplement existing datasets on all mentioned face-related tasks. To achieve that goal, the dataset needs all the following properties: (1) extreme-pose, (2) large-scale, (3) high-quality, (4) in-the-wild, (5) having multiple images for each subject and having subject identity annotated. Given those requirements, we utilize a novel and meticulous dataset processing pipeline to filter and sample frames from two publicly available datasets, including VFHQ [48] and CelebV-HQ [54], which contain high-resolution face videos captured in various settings. This results in a dataset with 450k high-quality images of extreme poses, which are combined with frontal images of the same subject to use as image pairs. These image pairs can support face reenactment and face recognition tasks, while the extreme-pose images alone can complement the existing datasets of face generation training. Note that there is a recent attempt [47] to build a large-pose facial dataset similar to ours. However, that work relies on images crawled from Flickr, thus having limited size (19k images) and missing identity information. Hence, its applications are restricted, covering only several face-generation tasks, unlike our large-scale and multi-purpose dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.796,
+                0.47,
+                0.903
+            ],
+            "angle": 0,
+            "content": "We perform a comprehensive set of experiments to validate the effectiveness of EFHQ across three distinct face generation sub-tasks and face reenactment scenarios. Our dataset, EFHQ, when combined with standard training sets, significantly enhances the quality of synthesized images in highly profile or pitched views while maintaining synthesis quality in frontal views. EFHQ also provides a more"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.092,
+                0.892,
+                0.152
+            ],
+            "angle": 0,
+            "content": "challenging pose-centric face verification benchmark. It reveals the vulnerability of many state-of-the-art face recognition networks when dealing with extreme-pose images, with TAR@FAR=1e-3 scores dropping 5-37%."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.153,
+                0.893,
+                0.318
+            ],
+            "angle": 0,
+            "content": "In summary, our contributions include (1) a large-scale, high-quality, in-the-wild, extreme-posed dataset (EFHQ) designed to complement various face-related tasks, (2) a novel and meticulous dataset processing pipeline ensuring the creation and quality of the EFHQ dataset, and (3) a series of experiments, with plan-to-release models, demonstrating the usefulness of EFHQ in enhancing performance on a wide range of tasks when dealing with faces at extreme rotations. The final dataset will be publicly available as metadata files to not violate the two original dataset's licenses, alongside the pretrained networks."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.333,
+                0.642,
+                0.348
+            ],
+            "angle": 0,
+            "content": "2. Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.365,
+                0.892,
+                0.516
+            ],
+            "angle": 0,
+            "content": "Synthetic Face Generation. Existing 2D-based models like StyleGAN [14, 15, 17] and 3D-aware models like EG3D [5] demonstrate impressive fidelity, especially for near-frontal views. Nevertheless, their performance significantly degrades when faced with extreme views, primarily due to insufficient training data for such scenarios. Datasets like FFHQ [14] exhibit a notable bias toward near-frontal poses, heavily limiting the range of poses synthesized by current 2D and 3D models despite the 3D-aware generators' ability to handle arbitrary poses."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.517,
+                0.893,
+                0.713
+            ],
+            "angle": 0,
+            "content": "Diffusion-based text-to-image models [32, 33] can synthesize high-quality images from simple text inputs. However, these models lack mechanisms to control the generated pose. Consequently, the generated facial images remain in a frontal view despite terms like \"profile\" or \"looking-up\" in the input prompts. ControlNet[49] proposes a finetuning approach allowing these models to take additional context as conditional input. Namely, one can use landmarks to guide the generated face to the desired pose. However, the scarcity of profile-view data in training affects the model's performance, particularly in extreme angles. Consequently, effectively managing the long-tailed distribution of these extreme views becomes necessary to address this issue."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.72,
+                0.893,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Face Reenactment. Face reenactment involves transferring facial expressions and movements from a driving image to a source image. Recent advancements claim high-quality results, especially for near-frontal poses [31, 41, 43, 51]. However, pose collapse persists, particularly in keypoint-based models [41, 51]. Common datasets for this task are VoxCeleb1 [29] and HDTF [50], which primarily feature frontal talking face videos. Despite recent potentially high-quality datasets for reenactment such as VFHQ [48] and CelebV-HQ [54], the frontal-view bias persists. As a result, facial reenactment methods often produce artifacts and lower-quality outputs when faced with significant rotations."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.521,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22606"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.16,
+                0.089,
+                0.805,
+                0.297
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Dataset</td><td></td><td colspan=\"3\">Attributes</td><td colspan=\"3\">Ready-to-use tasks</td></tr><tr><td>#images</td><td>Multiple images per ID</td><td>High Quality</td><td>Profile View</td><td>Face Verification</td><td>Face Synthesis</td><td>Face Reenactment</td></tr><tr><td>AgeDB</td><td>12K</td><td>✓</td><td></td><td></td><td>✓</td><td></td><td></td></tr><tr><td>CPLFW</td><td>12K</td><td>✓</td><td></td><td>✓</td><td>✓</td><td></td><td></td></tr><tr><td>CFP-FP</td><td>14K</td><td>✓</td><td></td><td>✓</td><td>✓</td><td></td><td></td></tr><tr><td>IJB-x</td><td>138K</td><td>✓</td><td></td><td>✓</td><td>✓</td><td></td><td></td></tr><tr><td>CelebA-HQ</td><td>30K</td><td></td><td>✓</td><td></td><td></td><td>✓</td><td></td></tr><tr><td>FFHQ</td><td>70K</td><td></td><td>✓</td><td></td><td></td><td>✓</td><td></td></tr><tr><td>LPFF</td><td>19K</td><td></td><td>✓</td><td>✓</td><td></td><td>✓</td><td></td></tr><tr><td>CelebV-HQ</td><td>6M</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>✓</td><td></td></tr><tr><td>VFHQ</td><td>3M</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>✓</td><td>✓</td></tr><tr><td>VoxCeleb1</td><td>4M</td><td>✓</td><td></td><td></td><td></td><td>✓</td><td>✓</td></tr><tr><td>Ours</td><td>450,538</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr><tr><td>FFHQ+Ours*</td><td>112,671</td><td></td><td>✓</td><td>✓</td><td></td><td>✓</td><td></td></tr><tr><td>Vox+Ours*</td><td>4.5M</td><td>✓</td><td></td><td>✓</td><td></td><td></td><td>✓</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.076,
+                0.302,
+                0.892,
+                0.358
+            ],
+            "angle": 0,
+            "content": "Table 1. Representative comparison of key attributes and supported tasks across existing face datasets and our proposed EFHQ dataset. EFHQ provides large-scale, high-quality images with multiple views per identity, enabling diverse face-related tasks, including synthesis, reenactment, and verification. Combining EFHQ with existing datasets like FFHQ or VoxCeleb can augment available data for selected tasks. Asterisk (*) indicates a combined dataset with a subsampled version of EFHQ."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.37,
+                0.47,
+                0.522
+            ],
+            "angle": 0,
+            "content": "Face Verification. Despite near-perfect accuracy on benchmark datasets[12, 20, 27, 28, 44], state-of-the-art face verification models[7, 13, 18] fall short in unconstrained environments due to a lack of pose diversity. While CPLFW[52] introduces 6000 cross-pose pairs, its small size and mix of pose samples hinder analyzing performance on specific pose scenarios such as frontal-frontal, frontal-profile, and profile-profile. More granular benchmarking on diverse pose pairs is crucial to identify failure modes and drive progress in robust face verification across poses."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.529,
+                0.47,
+                0.726
+            ],
+            "angle": 0,
+            "content": "Extreme Pose Face Dataset. Tab. 1 summarizes existing datasets related to face verification, synthesis, and re-enactment, showing that existing face datasets have limitations for multi-task use. Recent efforts like LPFF[47] made progress by crawling from Flickr and processing for diverse views, resulting in a dataset of 19,590 images. The dataset could complement high-quality datasets like FFHQ[14] for generative models. However, due to its limited scale comprising only \\(20\\%\\) of the combined dataset, the unbalanced distribution between frontal and profile view still needs further improvements. Additionally, since the images are from image hosting services, the dataset fails to provide the identity information necessary for other facial applications."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.741,
+                0.233,
+                0.757
+            ],
+            "angle": 0,
+            "content": "3. Dataset Process"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.767,
+                0.469,
+                0.829
+            ],
+            "angle": 0,
+            "content": "In this section, we discuss the process of producing our novel dataset EFHQ. To fulfill our goals, we carefully curated frames from the two recent facial video datasets, VFHQ[48] and CelebV-HQ[54], as seen in Fig. 2."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.839,
+                0.289,
+                0.856
+            ],
+            "angle": 0,
+            "content": "3.1. Attributes Preparation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.871,
+                0.469,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Attributes Extraction. To curate accurate facial poses and high-quality frames from existing datasets, we first ex"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.37,
+                0.892,
+                0.567
+            ],
+            "angle": 0,
+            "content": "tract (1) face bounding boxes, (2) facial landmarks (5/68 keypoints), (3) image quality score, and (4) face identity. For bounding boxes and landmarks, we employ the popular RetinaFace[36] and SynergyNet[46]. We use HyperIQA [40] to grade image quality. For VFHQ, we only generate 68 landmarks and image quality scores since bounding boxes and 5 keypoints are provided. After extracting these attributes, we match them with existing annotations using an IoU-based Hungarian matching algorithm [21]. Regarding identity, VFHQ-wise, we source labels from the annotations. As CelebV-HQ lacks individual identity labels within the video, we use[35] to identify the identities associated with each bounding box."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.571,
+                0.892,
+                0.829
+            ],
+            "angle": 0,
+            "content": "Due to the instability of head pose estimators when handling extreme-pose images, we strategically apply multiple pose estimators relying on various methodologies: (1) SynergyNet[46], a 3DMM-based model, (2) DirectMHP[53], a landmark-free joint head detector and pose estimator, and (3) an in-house FacePoseNet[6] improved through extensive data augmentation focused on extreme poses and extra training data. We then categorize each estimated pose into a hierarchical binning scheme illustrated in Fig. 2. Next, we gather bin predictions and perform majority voting to arrive at consensus labels. When no agreement emerges, we put it into the \\(5^{th}\\) bin for \"confusing\" cases. As depicted in Fig. 3, this ensemble approach provides robustness in cases where a single estimator incorrectly categorizes a sample. Please refer to the supplementary for statistics of the extracted bins and the hyperparameters for each model used for attribute extraction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.84,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Annotation Review. Next, we manually review the results to ensure high-quality labels. Given the large dataset, we developed a graphical user interface tool to streamline the review process. Our focus is verifying the binning annota"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.519,
+                0.958
+            ],
+            "angle": 0,
+            "content": "22607"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.16,
+                0.08,
+                0.808,
+                0.257
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.262,
+                0.895,
+                0.304
+            ],
+            "angle": 0,
+            "content": "Figure 2. The pipeline of EFHQ dataset creation. Starting with high-quality videos from the VFHQ[48] and CelebV-HQ[54] datasets, single-frame attributes are extracted then manually reviews. Task-specific preprocessing is then applied to generate specialized versions of the dataset for tasks such as face generation, reenactment, and verification."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.098,
+                0.322,
+                0.449,
+                0.449
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.077,
+                0.453,
+                0.47,
+                0.48
+            ],
+            "angle": 0,
+            "content": "Figure 3. Example cases where a pose estimator fails to categorize the sample to the correct bin."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.497,
+                0.47,
+                0.558
+            ],
+            "angle": 0,
+            "content": "tions, which most directly impact label accuracy. Concurrently, we randomly subsample the data based on pose angle and image quality to validate facial landmark quality further and discard cases of landmark prediction failure."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.572,
+                0.226,
+                0.588
+            ],
+            "angle": 0,
+            "content": "3.2. EFHQ dataset"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.597,
+                0.47,
+                0.689
+            ],
+            "angle": 0,
+            "content": "Our final EFHQ dataset comprises up to 450k frames with extreme poses extracted from approximately 5,000 clips. Most of the clips in our dataset include at least one frame with a frontal face and multiple frames with extreme pose angles. We also include extreme pose-only clips, representing profile-to-profile pose transfer cases."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.088,
+                0.706,
+                0.456,
+                0.826
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.828,
+                0.47,
+                0.899
+            ],
+            "angle": 0,
+            "content": "Figure 4. Pose distribution comparison between our sampled dataset and other datasets, including FFHQ[14], LPFF[47], CPLFW[52] and 40K random samples from VoxCeleb1 [29]. Our sampled dataset demonstrates greater pose diversity, with increased sample counts across high angle bins."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.323,
+                0.692,
+                0.337
+            ],
+            "angle": 0,
+            "content": "3.3. Dataset for subtasks"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.359,
+                0.892,
+                0.571
+            ],
+            "angle": 0,
+            "content": "Supplementary datasets for face generation. To address FFHQ's pose distribution gap, we compile a dataset encompassing diverse poses, varied identities, and FFHQ-comparable image quality from EFHQ. Additionally, we also embed an image brightness filter[3] in the sampling process, given that HyperIQA doesn't explicitly assess this aspect. The sampled images are then processed using the same pipeline in [5, 16] and manually reviewed. The aim is to improve coverage of under-represented poses without degrading frontal performance of trained generation models. This yields a dataset of 42,671 images with equitably distributed poses as visualized in Fig. 4. Compared to the original FFHQ distribution, yaw and pitch distributions are significantly improved."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.577,
+                0.892,
+                0.669
+            ],
+            "angle": 0,
+            "content": "Additionally, for 3D-aware GAN subtasks, we extract camera parameters to serve as conditional pose information, following [5]. We convert these parameters into yaw, pitch angles to validate against our existing annotations. When disagreements emerge, we re-examine the images and filter if needed, to ensure accurate, quality labels."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.675,
+                0.893,
+                0.903
+            ],
+            "angle": 0,
+            "content": "For diffusion-based text-to-image generation, a specialized dataset is curated to refine Stable Diffusion models[33] by accommodating landmark conditional input through ControlNet[49]. Based on our 3D-aware face generation dataset, we integrate landmark-based conditional images and tailored text prompts for each image. To capture fine facial nuances, we extract 478 facial landmarks and draw the condition image with connected edges between matched keypoints, following Mediapipe framework[26], better-conveying face expression and pose. For text prompts, detailed facial attributes like gender, race, and emotion are derived from existing labels or inferred via BLIP-2 pretrained captioning model [23] and a face attribute estimator [37]. The resulting prompt adheres to a structured format: \"A profile portrait image of a [emotion] [race] [gender].\""
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22608"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.08,
+                0.089,
+                0.472,
+                0.26
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>Reference Dataset</td><td>FID↓</td><td>Recall↑</td></tr><tr><td>GFFHQ</td><td>FFHQ</td><td>2.84</td><td>0.49</td></tr><tr><td>GFFHQ+LPFF</td><td>FFHQ+LPFF</td><td>3.43</td><td>0.44</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ+EFHQ</td><td>3.44</td><td>0.46</td></tr><tr><td>GFFHQ+LPFF</td><td>FFHQ+EFHQ</td><td>8.26</td><td>0.43</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ+LPFF</td><td>7.04</td><td>0.44</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ+EFHQ</td><td>3.33</td><td>0.46</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ</td><td>3.12</td><td>0.44</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ+LPFF</td><td>4.47</td><td>0.42</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.076,
+                0.265,
+                0.47,
+                0.349
+            ],
+            "angle": 0,
+            "content": "Table 2. StyleGAN2-ADA models trained on different datasets. Lower FIDs indicate better fidelity, while higher Recalls indicate better diversity. The second and third blocks show comparisons between unconditional models trained on FFHQ+LPFF and FFHQ+EFHQ when evaluating on the same and cross-dataset settings, with the better metrics in bold."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.356,
+                0.47,
+                0.418
+            ],
+            "angle": 0,
+            "content": "Supplementary dataset for face reenactment. As EFHQ is crafted, we can employ the entire dataset to complement established training datasets, e.g., VoxCeleb1 [29]. We also built an extra evaluation set with 1200 EFHQ clips."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.425,
+                0.47,
+                0.606
+            ],
+            "angle": 0,
+            "content": "Benchmarking dataset for face verification. We curate a dataset covering three distinct scenarios: frontal-to-frontal, frontal-to-profile, and profile-to-profile. To enable rigorous benchmarking, we sample 10,000 pairs each for both negative and positive cases per scenario, resulting in a balanced benchmark dataset containing 60,000 image pairs. Images follow the same established preprocessing pipelines in [7]. Next, we filter misaligned images with RetinaFace[36], randomly review, and replace equivalent samples from our diverse corpus, if needed. Finally, to simulate varying image quality, we randomly applied downscaling and compression to samples in the dataset."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.624,
+                0.314,
+                0.64
+            ],
+            "angle": 0,
+            "content": "4. Face Generation Subtask"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.649,
+                0.192,
+                0.667
+            ],
+            "angle": 0,
+            "content": "4.1. StyleGAN"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.681,
+                0.468,
+                0.801
+            ],
+            "angle": 0,
+            "content": "Training details. We train StyleGAN2-ADA[15] models from scratch after combining FFHQ with our own curated dataset (denoted FFHQ+EFHQ) at \\(1024 \\times 1024\\) resolution. We also train a conditional version, allowing direct control over frontal versus profile synthesis and enhancing model performance. Specifically, we categorize the dataset into two classes: \"frontal\" and \"profile\" based on pose angles. All models are trained for 6 days on 8 Nvidia A100 GPUs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.81,
+                0.468,
+                0.9
+            ],
+            "angle": 0,
+            "content": "Evaluation protocol. We utilize the Frechet Inception Distance[11] (FID) metric and improved Recall[22] to quantify the generation fidelity and diversity. Regarding the two unconditional models trained with the supplemented FFHQ+LPFF and FFHQ+EFHQ, we evaluate each model with both of these datasets. For conditional mod"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.533,
+                0.089,
+                0.86,
+                0.175
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.548,
+                0.178,
+                0.845,
+                0.19
+            ],
+            "angle": 0,
+            "content": "(a) Samples from the model trained with FFHQ+EFHQ."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.532,
+                0.191,
+                0.858,
+                0.35
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.531,
+                0.351,
+                0.858,
+                0.365
+            ],
+            "angle": 0,
+            "content": "(b) Profile-view comparison of LPFF (left) and EFHQ (right)."
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.499,
+                0.369,
+                0.892,
+                0.41
+            ],
+            "angle": 0,
+            "content": "Figure 5. Qualitative result and comparison of generated samples, with truncation \\(\\psi = 0.7\\), from StyleGAN2-ADA training with other dataset and ours."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.433,
+                0.892,
+                0.538
+            ],
+            "angle": 0,
+            "content": "els, we evaluate performance on FFHQ, FFHQ+LPFF, and FFHQ+EFHQ to demonstrate the broader coverage of the model trained with our proposed dataset. For FFHQ, we generate 50K images with the \"frontal\" label, whereas for the other two datasets, we generate 50K following their label distributions. This comprehensive quantitative analysis examines the fidelity and variety of the generated images."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.554,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Experimental results. We compared the performance of our proposed unconditional models \\((G^{\\mathrm{FFHQ + EFHQ}})\\), against models trained on the FFHQ and FFHQ+LPFF datasets \\((G^{\\mathrm{FFHQ}}\\) and \\(G^{\\mathrm{FFHQ + LPFF}})\\). Models trained conditionally are specifically subscripted, e.g., \\(G_{c}^{\\mathrm{FFHQ + EFHQ}}\\). The qualitative results in Fig. 5 illustrate that our models generate high-quality frontal faces comparable to the FFHQ model while producing realistic and varied profile faces. Additionally, our method has less noise in extreme poses than the FFHQ+LPFF model. Specifically, Fig. 5b shows LPFF's generated face contains more noise and less photorealistic details than our model, such as the pixelated eye and nose patches. Meanwhile, ours achieves greater realism with smoother skin, well-defined features, noise reduction, and natural lighting. The quantitative results in Tab. 2 show that our unconditional model achieves competitive FID scores and improved Recall to other models, indicating similar fidelity but greater sample pose diversity. In the cross-examination (third block of Tab. 2), \\(G^{\\mathrm{FFHQ + EFHQ}}\\) has both better FID and improved Recall on the opposite dataset than \\(G^{\\mathrm{FFHQ + LPFF}}\\). Finally, our conditional model demonstrates substantially improved FID scores over multiple reference datasets, highlighting its versatility."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22609"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.081,
+                0.089,
+                0.468,
+                0.206
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>Reference Dataset</td><td>FID↓</td><td>Pose↓</td><td>Depth↓</td><td>ID↑</td></tr><tr><td>EFHQ</td><td>FFHQ/Trainset</td><td>4.8</td><td>5e-4</td><td>0.234</td><td>0.74</td></tr><tr><td>EFFHQ</td><td>FFHQ/Trainset</td><td>3.7</td><td>5e-4</td><td>0.288</td><td>0.74</td></tr><tr><td>EFFHQ+LPFF</td><td>FFHQ</td><td>6.2</td><td>6e-4</td><td>0.316</td><td>0.77</td></tr><tr><td>EFFHQ+EFHQ</td><td>FFHQ</td><td>3.9</td><td>4e-4</td><td>0.224</td><td>0.77</td></tr><tr><td>EFFHQ+LPFF</td><td>Trainset</td><td>7.0</td><td>6e-4</td><td>0.316</td><td>0.65</td></tr><tr><td>EFFHQ+EFHQ</td><td>Trainset</td><td>4.6</td><td>4e-4</td><td>0.223</td><td>0.65</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.077,
+                0.211,
+                0.468,
+                0.294
+            ],
+            "angle": 0,
+            "content": "Table 3. EG3D models trained on different datasets. In the second and third blocks, we compare models trained on FFHQ+LPFF and FFHQ+EFHQ with the testset is either FFHQ or their training datasets; the better metrics are in bold. The best metrics on the FFHQ testset are in red, while the best metrics when evaluating on training dataset are underlined."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.089,
+                0.295,
+                0.457,
+                0.458
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.461,
+                0.468,
+                0.502
+            ],
+            "angle": 0,
+            "content": "Figure 6. Comparison between multiview generated samples, with truncation \\(\\psi = 0.8\\), of EG3D model trained with various datasets."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.508,
+                0.163,
+                0.523
+            ],
+            "angle": 0,
+            "content": "4.2. EG3D"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.539,
+                0.468,
+                0.645
+            ],
+            "angle": 0,
+            "content": "Training details. We followed the same pipeline and hyper-parameters as the original EG3D model [5]. However, based on empirical results, we adjusted \\(\\gamma = 5\\). We also fixed the triplane implementation bug (XY, XZ, ZX) [1] to replicate the original idea in the papers (XY, XZ, ZY). We trained the model on 8 Nvidia A100 40GB GPUs for approximately 7 days until convergence."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.652,
+                0.468,
+                0.759
+            ],
+            "angle": 0,
+            "content": "Evaluation Protocol. To evaluate the models, mirroring [5], we measure image quality with FID; multi-view consistency and identity preservation with mean cosine similarity of pre-trained face recognition model[7]; pose accuracy and geometry by calculating MSE against pseudo labels from [8]. We evaluate all models on 1,024 generated images, except for FID, which we compute on 50k images."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.765,
+                0.468,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Experimental results. To evaluate the benefits of our dataset, we trained the same backbone architecture on different datasets: FFHQ \\((E^{\\mathrm{FFHQ}})\\), FFHQ+LPFF \\((E^{\\mathrm{FFHQ} + \\mathrm{LPFF}})\\), and FFHQ+EFHQ (our model, \\(E^{\\mathrm{FFHQ} + \\mathrm{EFHQ}}\\)). The quantitative results are presented in Table 3, and qualitative examples are shown in Figure 6. Qualitatively, we observe that \\(E^{\\mathrm{FFHQ} + \\mathrm{EFHQ}}\\) generates improved face shapes in extreme poses compared to \\(E^{\\mathrm{FFHQ}}\\), similar to the improvements seen with \\(E^{\\mathrm{FFHQ} + \\mathrm{LPFF}}\\). Quantitatively, the FID"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.501,
+                0.089,
+                0.889,
+                0.218
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Metrics</td><td>Dataset</td><td>Pretrained ControlNet</td><td>Ours</td></tr><tr><td>NME↓</td><td>LPFF</td><td>0.085</td><td>0.084</td></tr><tr><td>FID↓</td><td>LPFF</td><td>119.74</td><td>83.39</td></tr><tr><td>FIDeyepatch ↓</td><td>LPFF</td><td>103.77</td><td>49.45</td></tr><tr><td>NME↓</td><td>EFHQ</td><td>0.106</td><td>0.081</td></tr><tr><td>FID↓</td><td>EFHQ</td><td>111.23</td><td>81.90</td></tr><tr><td>FIDeyepatch ↓</td><td>EFHQ</td><td>137.43</td><td>58.60</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.499,
+                0.221,
+                0.892,
+                0.263
+            ],
+            "angle": 0,
+            "content": "Table 4. Quantitative results of fine-tuning StableDiffusion v1.5 with ControlNet trained on various datasets, using Normalized Mean Error (NME), FID, and FID on eye region."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.271,
+                0.892,
+                0.47
+            ],
+            "angle": 0,
+            "content": "of \\( E^{\\mathrm{FFHQ + EFHQ}} \\) is comparable to \\( E^{\\mathrm{FFHQ}} \\) when evaluated on FFHQ, indicating no degradation in frontal view performance. Meanwhile, the competitive FID achieved by \\( E^{\\mathrm{FFHQ + EFHQ}} \\) on the combined FFHQ+EFHQ dataset also suggests proficient modeling of profile views. Regarding identity consistency, we see a significant drop when the reference dataset contains profile-view images, due to the weak confidence of the pretrained face recognition model when handling profile-view images, as confirmed by its low average cosine similarity of around 0.6 on the simple frontal-to-profile verification set CFP-FP[34]. Furthermore, our model also outperforms in pose accuracy and geometry quality with both frontal and cross-pose dataset."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.514,
+                0.481,
+                0.696,
+                0.579
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.697,
+                0.481,
+                0.882,
+                0.578
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.531,
+                0.58,
+                0.862,
+                0.592
+            ],
+            "angle": 0,
+            "content": "(a) Generated samples; condition image is in between samples."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.502,
+                0.593,
+                0.699,
+                0.626
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.7,
+                0.593,
+                0.894,
+                0.626
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.628,
+                0.628,
+                0.768,
+                0.641
+            ],
+            "angle": 0,
+            "content": "(b) Generated eye regions."
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.499,
+                0.646,
+                0.892,
+                0.689
+            ],
+            "angle": 0,
+            "content": "Figure 7. Comparison profile-view generated samples of pretrained ControlNet (left) and our fine-tuned ControlNet (right) with the prompt: \"A profile portrait image of a person\"."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.698,
+                0.623,
+                0.712
+            ],
+            "angle": 0,
+            "content": "4.3. ControlNet"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.727,
+                0.892,
+                0.818
+            ],
+            "angle": 0,
+            "content": "Training details. We finetuned Stable Diffusion v1.5 [33] using ControlNet [49] on the EFHQ dataset to enhance its ability to generate extreme facial poses. We use AdamW optimizer [25] with learning rate \\( 1 \\times e^{-5} \\), \\( \\beta_0 = 0.9 \\), \\( \\beta_1 = 0.999 \\), and \\( \\epsilon = 0.01 \\). We train with batch size 24 on 1 Nvidia A100 40GB GPU in 2 days."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.825,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Evaluation Protocol. To assess image qualities across diverse poses, we measure the Normalized Mean Error (NME) of 478-point facial landmarks between the generated and source image. Specifically, we evaluate performance across the full pose distribution in the reference"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22610"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.082,
+                0.089,
+                0.472,
+                0.176
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"2\">Test Dataset (L1 ↓/ AKD ↓/ AED ↓)</td></tr><tr><td>VoxCeleb1</td><td>EFHQ (Ours)</td></tr><tr><td>TPS(VoxCeleb1</td><td>0.041/1.214/0.132</td><td>0.061/3.814/0.275</td></tr><tr><td>TPS(VoxCeleb1+EFHQ</td><td>0.041/1.250/0.133</td><td>0.051/1.963/0.214</td></tr><tr><td>LIA(VoxCeleb1</td><td>0.044/1.342/0.138</td><td>0.059/3.807/0.266</td></tr><tr><td>LIA(VoxCeleb1+EFHQ</td><td>0.044/1.398/0.142</td><td>0.054/2.291/0.233</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.076,
+                0.181,
+                0.47,
+                0.237
+            ],
+            "angle": 0,
+            "content": "Table 5. Quantitative results of two state-of-the-art reenactment methods (TPS and LIA) when training on VoxCeleb1 and the combined set VoxCeleb1+EFHQ. We present results evaluated either on VoxCeleb1 (left) or EFHQ (right) test sets."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.242,
+                0.47,
+                0.364
+            ],
+            "angle": 0,
+            "content": "dataset. For the prompt, we use: \"A profile portrait image of a person.\" with extra positive and negative keywords, further detailed in supplementary. Utilizing LPFF and EFHQ datasets as the reference datasets, we compute FID to assess synthesized face quality and extend evaluation to cropped eye regions due to unrealistic gazes observed from the original ControlNet model. Note that the models receive condition images created akin to its training data."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.37,
+                0.471,
+                0.643
+            ],
+            "angle": 0,
+            "content": "Experimental results. Fig. 7a visually compares the baseline ControlNet OpenPose face and our fine-tuned model. Our fine-tuned model adeptly conditions diverse extreme poses, generating high-quality results with precise details while maintaining fidelity to the condition, such as mouth shape, and exhibiting minimal artifacts. These patterns indicate the superiority of the fine-tuned model over the baseline model in handling extreme viewing angles. Additionally, Fig. 7b contrasts the eye regions of both models. Our fine-tuned model accurately captures natural eye gaze and shape from condition images, a notable improvement over the baseline. Quantitative results in Tab. 4 consistently favor our fine-tuned model across all metrics. FID scores on full and eye patches exhibit significant enhancement, and NME performance remains stable on training and reference datasets (EFHQ and LPFF). These findings showcase the fine-tuned model's capability to generate high-quality, geometrically consistent faces across diverse poses."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.657,
+                0.329,
+                0.673
+            ],
+            "angle": 0,
+            "content": "5. Face Reenactment Subtask"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.681,
+                0.267,
+                0.696
+            ],
+            "angle": 0,
+            "content": "5.1. Evaluation Protocol"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.704,
+                0.471,
+                0.903
+            ],
+            "angle": 0,
+            "content": "We follow the same preprocessing and evaluation protocol as FOMM [38]. Specifically, all the frames are cropped to \\(256 \\times 256\\), and we use the first frame of a clip as the source image to reconstruct the remaining frames. In the case of the EFHQ test set, most source images have a frontal pose, while the target frames have an extreme pose. Some test cases only contain extreme pose to evaluate self-reconstruction on these cases. We run the experiments using two recent methods, Thin-Plate Spline Motion Model (TPS) [51] and Latent Image Animator (LIA) [43]. For training details, please refer to the supplementary. We select the following three metrics to compare the generated frames and the driving ground truth. \\(\\mathcal{L}_1\\) measures the av"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.504,
+                0.089,
+                0.89,
+                0.196
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Model</td><td rowspan=\"2\">Trainset</td><td rowspan=\"2\">CPLFW</td><td colspan=\"3\">EFHQ</td></tr><tr><td>f2f</td><td>f2p</td><td>p2p</td></tr><tr><td>R18arface</td><td>Glint360k</td><td>0.80</td><td>0.94</td><td>0.570.37</td><td>0.880.06</td></tr><tr><td>R18arface</td><td>MS1MV3</td><td>0.75</td><td>0.93</td><td>0.570.36</td><td>0.880.05</td></tr><tr><td>R50arface</td><td>Glint360k</td><td>0.91</td><td>0.97</td><td>0.770.20</td><td>0.920.05</td></tr><tr><td>R100arface</td><td>Glint360k</td><td>0.93</td><td>0.98</td><td>0.830.15</td><td>0.820.16</td></tr><tr><td>R100adaface</td><td>WebFace4M</td><td>0.94</td><td>0.99</td><td>0.940.05</td><td>0.970.02</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.498,
+                0.2,
+                0.892,
+                0.299
+            ],
+            "angle": 0,
+            "content": "Table 6. Quantitative evaluation of state-of-the-art face recognition models on the cleansed CPLFW and the proposed benchmark dataset. We use the metric TAR@FAR=1e-3 (higher is better). Our EFHQ benchmark includes frontal-frontal (f2f), frontal-profile (f2p), and profile-profile (p2p) subsets. For f2p and p2p, we report the score drop when comparing to f2f in the red subscription. The lowest benchmark score for each model is in bold."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.307,
+                0.892,
+                0.399
+            ],
+            "angle": 0,
+            "content": "erage distance among pixel values, indicating reconstruction faithfulness. Average Keypoint Distance (AKD) calculates the mean distance between keypoints, thus assessing the pose accuracy. Average Euclidean Distance (AED) measures identity preservation by computing the average \\(\\mathcal{L}_2\\) distance between identity embeddings."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.413,
+                0.697,
+                0.429
+            ],
+            "angle": 0,
+            "content": "5.2. Experimental results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.437,
+                0.893,
+                0.725
+            ],
+            "angle": 0,
+            "content": "We compare TPS and LIA models trained with our supplementary EFHQ dataset \\((TPS^{\\mathrm{VoxCeleb1 + EFHQ}}\\) and \\(LIA^{\\mathrm{VoxCeleb1 + EFHQ}})\\) against models trained solely on VoxCeleb1 \\((TPS^{\\mathrm{VoxCeleb1}}\\) and \\(LIA^{\\mathrm{VoxCeleb1}})\\). Tab. 5 displays quantitative results, revealing that models trained with EFHQ maintain comparable performance when evaluated on VoxCeleb1, emphasizing high-quality performance in frontal-to-frontal reenactment scenarios. Notably, on the EFHQ test set, both EFHQ-trained models show significant improvement in all metrics, indicating enhanced extreme pose and identity reenactment performance. In Fig. 8, qualitative results prove that EFHQ-trained models excel in transferring facial expressions and motion across diverse poses. Models trained on EFHQ showcase improved image quality, reduced artifacts, and well-preserved shapes of driving faces. These results highlight our models' effectiveness in capturing and reproducing facial motions across varying poses while enhancing fidelity to the original facial features."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.743,
+                0.742,
+                0.761
+            ],
+            "angle": 0,
+            "content": "6. Face Recognition Subtask"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.77,
+                0.691,
+                0.785
+            ],
+            "angle": 0,
+            "content": "6.1. Evaluation Protocol"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.796,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "To evaluate face recognition performance on extreme poses, we selected two state-of-the-art methods: ArcFace[7] and AdaFace[18]. We further explored various backbone sizes (ResNet18, ResNet50, ResNet101)[10] and training datasets (MS1MV2[9], Glint360K[2], WebFace4M[55]) of increasing size. For metrics, we used True Acceptance Rate (TAR) at False Acceptance Rate (FAR) of \\( 1 \\times e^{-3} \\)."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.518,
+                0.958
+            ],
+            "angle": 0,
+            "content": "22611"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.169,
+                0.1,
+                0.266,
+                0.173
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.169,
+                0.175,
+                0.268,
+                0.248
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.199,
+                0.255,
+                0.236,
+                0.264
+            ],
+            "angle": 0,
+            "content": "Source"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.279,
+                0.1,
+                0.375,
+                0.173
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.279,
+                0.175,
+                0.376,
+                0.25
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.285,
+                0.254,
+                0.372,
+                0.265
+            ],
+            "angle": 0,
+            "content": "TPS (VoxCeleb1)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.38,
+                0.1,
+                0.475,
+                0.173
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.38,
+                0.175,
+                0.475,
+                0.25
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.398,
+                0.255,
+                0.458,
+                0.266
+            ],
+            "angle": 0,
+            "content": "TPS (Ours)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.49,
+                0.1,
+                0.587,
+                0.173
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.49,
+                0.175,
+                0.587,
+                0.248
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.497,
+                0.254,
+                0.585,
+                0.265
+            ],
+            "angle": 0,
+            "content": "LIA (VoxCeleb1)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.593,
+                0.1,
+                0.688,
+                0.173
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.593,
+                0.175,
+                0.688,
+                0.248
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.61,
+                0.254,
+                0.669,
+                0.265
+            ],
+            "angle": 0,
+            "content": "LIA (Ours)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.703,
+                0.1,
+                0.798,
+                0.173
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.702,
+                0.175,
+                0.797,
+                0.248
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.73,
+                0.254,
+                0.771,
+                0.266
+            ],
+            "angle": 0,
+            "content": "Driving"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.27,
+                0.892,
+                0.3
+            ],
+            "angle": 0,
+            "content": "Figure 8. Comparison of performance for frontal-profile face reenactment on EFHQ. From left to right: source image, TPS on VoxCeleb1, TPS with EFHQ, LIA on VoxCeleb1, LIA with EFHQ, driving image. Additional EFHQ training data improves synthesis of extreme poses."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.31,
+                0.265,
+                0.326
+            ],
+            "angle": 0,
+            "content": "6.2. Experimental results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.333,
+                0.469,
+                0.424
+            ],
+            "angle": 0,
+            "content": "Before assessing face recognition networks on EFHQ, we first clean the commonly-used version of CPLFW, processed by[7], by removing misaligned images (please refer to the Supplementary), and evaluate the models on it. Models performed well on this cleansed set, with minimal gaps between ArcFace and AdaFace."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.424,
+                0.47,
+                0.636
+            ],
+            "angle": 0,
+            "content": "Our proposed evaluation benchmark involves three pose-based scenarios. Results in Tab. 6 reveal performance drops of \\(5 - 37\\%\\) on the frontal-profile subset compared to frontal-frontal, highlighting the challenge of matching faces in different poses. Larger ResNet models outperformed ResNet18 across subsets, showcasing the advantages of increased capacity in handling pose variation. However, the ResNet100 model trained on Glint360K performed worse than smaller models on the profile-profile subset, indicating the need for sufficient profile data to prevent overfitting. Our results also indicate AdaFace's suitability for handling pose variations. In summary, pose mismatch brings significant challenges for face recognition, necessitating larger models trained on diverse data covering profile faces."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.647,
+                0.192,
+                0.661
+            ],
+            "angle": 0,
+            "content": "7. Discussion"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.097,
+                0.669,
+                0.447,
+                0.814
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.819,
+                0.468,
+                0.846
+            ],
+            "angle": 0,
+            "content": "Figure 9. User study results. We reported the percentage of times testers rank whose output is best or tied for each criterion."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.871,
+                0.469,
+                0.902
+            ],
+            "angle": 0,
+            "content": "User survey. We conduct user surveys to evaluate the face generation improvement of our method through human per"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.311,
+                0.892,
+                0.402
+            ],
+            "angle": 0,
+            "content": "ceptual assessment. In this survey, we compare the outputs of the models mentioned in Secs. 4 and 5. The setup of this user survey is included in the supplementary. In all human evaluation results (Fig. 9), our models outperform previous ones, notably on identity and motion test of face reenactment models (74.80% and 64.22%, respectively)."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.529,
+                0.42,
+                0.861,
+                0.507
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.498,
+                0.511,
+                0.892,
+                0.58
+            ],
+            "angle": 0,
+            "content": "Figure 10. Representative examples of the raw LPFF dataset that got excluded from the final dataset, either from misdetections or pose filtering process. Red keypoints represent landmarks prediction of selected samples from LPFF, green keypoints represent landmarks prediction from our pipeline."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.595,
+                0.892,
+                0.7
+            ],
+            "angle": 0,
+            "content": "Robustness of data processing pipeline. As shown in Fig. 10, our data processing pipeline can detect additional profile-view faces filtered out by the LPFF pipeline, likely due to misdetections or incorrect pose filtering. This confirms the robustness of our pipeline, and we hope it will enable the community to develop larger and higher-quality in-the-wild facial datasets."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.721,
+                0.619,
+                0.736
+            ],
+            "angle": 0,
+            "content": "8. Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.748,
+                0.892,
+                0.888
+            ],
+            "angle": 0,
+            "content": "This work has introduced a large-scale, diverse facial dataset to address performance gaps between frontal and profile faces, powered by a novel and robust data processing pipeline. Crucially, we provided tailored sub-datasets for advancing essential tasks like face synthesis and reenactment. Moreover, our new face verification benchmark reveals gaps between techniques, granting valuable insights. Ultimately, we hope this high-quality, diverse facial dataset opens up new and exciting opportunities to push forward cross-pose tasks."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22612"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.08,
+                0.09,
+                0.174,
+                0.106
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.115,
+                0.47,
+                0.156
+            ],
+            "angle": 0,
+            "content": "[1] Fix for triplane projection \\(\\cdot\\) Issue #67 \\(\\cdot\\) NVlabs/eg3d. https://github.com/NVlabs/eg3d/issues/67, 2023.6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.157,
+                0.472,
+                0.227
+            ],
+            "angle": 0,
+            "content": "[2] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.228,
+                0.471,
+                0.282
+            ],
+            "angle": 0,
+            "content": "[3] Sergey Bezryadin, Pavel Bourov, and Dmitry Ilinih. Brightness calculation in digital image processing. International Symposium on Technologies for Digital Photo Fulfillment, 2007:10-15, 2007. 4, 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.284,
+                0.471,
+                0.34
+            ],
+            "angle": 0,
+            "content": "[4] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.341,
+                0.471,
+                0.41
+            ],
+            "angle": 0,
+            "content": "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 4, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.411,
+                0.471,
+                0.48
+            ],
+            "angle": 0,
+            "content": "[6] Feng-Ju Chang, Anh Tuan Tran, Tal Hassner, Iacopo Masi, Ram Nevatia, and Gerard Medioni. Faceposenet: Making a case for landmark-free face alignment. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1599-1608, 2017. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.481,
+                0.471,
+                0.549
+            ],
+            "angle": 0,
+            "content": "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 3, 5, 6, 7, 8, 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.551,
+                0.471,
+                0.62
+            ],
+            "angle": 0,
+            "content": "[8] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 6, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.621,
+                0.471,
+                0.704
+            ],
+            "angle": 0,
+            "content": "[9] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 1, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.705,
+                0.47,
+                0.761
+            ],
+            "angle": 0,
+            "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778. IEEE, 2016. 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.761,
+                0.47,
+                0.83
+            ],
+            "angle": 0,
+            "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.831,
+                0.47,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[12] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 1, 3"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.115,
+                0.472,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.093,
+                0.892,
+                0.175
+            ],
+            "angle": 0,
+            "content": "[13] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.177,
+                0.892,
+                0.245
+            ],
+            "angle": 0,
+            "content": "[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3, 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.246,
+                0.892,
+                0.3
+            ],
+            "angle": 0,
+            "content": "[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Proc. NeurIPS, 2020. 1, 2, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.301,
+                0.892,
+                0.37
+            ],
+            "angle": 0,
+            "content": "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.371,
+                0.892,
+                0.413
+            ],
+            "angle": 0,
+            "content": "[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.414,
+                0.892,
+                0.468
+            ],
+            "angle": 0,
+            "content": "[18] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.469,
+                0.892,
+                0.496
+            ],
+            "angle": 0,
+            "content": "[19] Davis E. King. Dlib-ml: A machine learning toolkit. J. Mach. Learn. Res., 10:1755-1758, 2009. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.497,
+                0.892,
+                0.58
+            ],
+            "angle": 0,
+            "content": "[20] Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1931-1939, 2015. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.581,
+                0.892,
+                0.62
+            ],
+            "angle": 0,
+            "content": "[21] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.622,
+                0.892,
+                0.704
+            ],
+            "angle": 0,
+            "content": "[22] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Proceedings of the 33rd International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2019. Curran Associates Inc. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.705,
+                0.892,
+                0.761
+            ],
+            "angle": 0,
+            "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.761,
+                0.892,
+                0.816
+            ],
+            "angle": 0,
+            "content": "[24] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.817,
+                0.892,
+                0.859
+            ],
+            "angle": 0,
+            "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.86,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[26] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Wan-Teh"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.503,
+                0.093,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22613"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.108,
+                0.092,
+                0.468,
+                0.134
+            ],
+            "angle": 0,
+            "content": "Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for building perception pipelines. CoRR, abs/1906.08172, 2019. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.135,
+                0.468,
+                0.216
+            ],
+            "angle": 0,
+            "content": "[27] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.218,
+                0.468,
+                0.3
+            ],
+            "angle": 0,
+            "content": "[28] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.302,
+                0.468,
+                0.342
+            ],
+            "angle": 0,
+            "content": "[29] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. VoxCeleb: A large-scale speaker identification dataset. In Interspeech 2017. ISCA, 2017. 2, 4, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.344,
+                0.468,
+                0.399
+            ],
+            "angle": 0,
+            "content": "[30] Aaron Nech and Ira Kemelmacher-Shlizerman. Level playing field for million scale face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.4,
+                0.468,
+                0.481
+            ],
+            "angle": 0,
+            "content": "[31] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-Ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 427-436, 2023. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.483,
+                0.468,
+                0.536
+            ],
+            "angle": 0,
+            "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv e-prints, pages arXiv-2204, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.539,
+                0.468,
+                0.607
+            ],
+            "angle": 0,
+            "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.608,
+                0.468,
+                0.677
+            ],
+            "angle": 0,
+            "content": "[34] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.679,
+                0.468,
+                0.733
+            ],
+            "angle": 0,
+            "content": "[35] Sefik Ilkin Serengil and Alper Ozpinar. Lightface: A hybrid deep face recognition framework. In 2020 Innovations in Intelligent Systems and Applications Conference (ASYU), pages 23-27. IEEE, 2020. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.735,
+                0.468,
+                0.789
+            ],
+            "angle": 0,
+            "content": "[36] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 3, 5, 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.79,
+                0.468,
+                0.845
+            ],
+            "angle": 0,
+            "content": "[37] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.846,
+                0.468,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[38] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 7, 2"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.092,
+                0.468,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.093,
+                0.892,
+                0.134
+            ],
+            "angle": 0,
+            "content": "[39] L Sirovich and M Kirby. Low-dimensional procedure for the characterization of human faces. Journal of the Optical Society of America A, 4(3):519-524, 1987. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.137,
+                0.892,
+                0.205
+            ],
+            "angle": 0,
+            "content": "[40] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.207,
+                0.892,
+                0.275
+            ],
+            "angle": 0,
+            "content": "[41] Jiale Tao, Biao Wang, Borun Xu, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Structure-aware motion transfer with deformable anchor model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3637-3646, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.278,
+                0.892,
+                0.332
+            ],
+            "angle": 0,
+            "content": "[42] Matthew A Turk and Alex P Pentland. Face recognition using eigenfaces. In Proceedings. 1991 IEEE computer society conference on computer vision and pattern recognition, pages 586-587. IEEE Computer Society, 1991. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.334,
+                0.892,
+                0.401
+            ],
+            "angle": 0,
+            "content": "[43] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In ICLR 2022-The International Conference on Learning Representations, 2022. 2, 7, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.405,
+                0.892,
+                0.488
+            ],
+            "angle": 0,
+            "content": "[44] Cameron Whitelam, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K Jain, James A Duncan, Kristen Allen, et al. Iarpa janus benchmark-b face dataset. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 90-98, 2017. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.49,
+                0.892,
+                0.545
+            ],
+            "angle": 0,
+            "content": "[45] L Wolf, T Hassner, and I Maoz. Face recognition in unconstrained videos with matched background similarity. In Proceedings of the 2011 IEEE Conference on Computer Vision and Pattern Recognition, pages 529-534, 2011. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.547,
+                0.892,
+                0.602
+            ],
+            "angle": 0,
+            "content": "[46] Cho-Ying Wu, Qiangeng Xu, and Ulrich Neumann. Synergy between 3dmm and 3d landmarks for accurate 3d facial geometry. In 2021 International Conference on 3D Vision (3DV), 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.604,
+                0.892,
+                0.671
+            ],
+            "angle": 0,
+            "content": "[47] Yiqian Wu, Jing Zhang, Hongbo Fu, and Xiaogang Jin. Lpff: A portrait dataset for face generators across large poses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20327-20337, 2023. 1, 2, 3, 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.675,
+                0.892,
+                0.743
+            ],
+            "angle": 0,
+            "content": "[48] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2022. 1, 2, 3, 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.746,
+                0.892,
+                0.801
+            ],
+            "angle": 0,
+            "content": "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 4, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.803,
+                0.892,
+                0.871
+            ],
+            "angle": 0,
+            "content": "[50] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.873,
+                0.892,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[51] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Con"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.503,
+                0.093,
+                0.892,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22614"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.108,
+                0.093,
+                0.468,
+                0.119
+            ],
+            "angle": 0,
+            "content": "ference on Computer Vision and Pattern Recognition, pages 3657-3666, 2022. 2, 7, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.122,
+                0.469,
+                0.177
+            ],
+            "angle": 0,
+            "content": "[52] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7), 2018. 3, 4, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.179,
+                0.469,
+                0.219
+            ],
+            "angle": 0,
+            "content": "[53] Huayi Zhou, Fei Jiang, and Hongtao Lu. Directmhp: Direct 2d multi-person head pose estimation with full-range angles. arXiv preprint arXiv:2302.01110, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.221,
+                0.469,
+                0.275
+            ],
+            "angle": 0,
+            "content": "[54] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 1, 2, 3, 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.278,
+                0.469,
+                0.36
+            ],
+            "angle": 0,
+            "content": "[55] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 7"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.093,
+                0.469,
+                0.36
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22615"
+        }
+    ]
+]

2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/65a10ed9-cef6-42e8-a141-8586b125aea4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be3409fea9659bb691923937f02b2e8710da69be27354a78c9dae72d89611b27
+size 3811407

2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/full.md

@@ -0,0 +1,301 @@
+# EFHQ: Multi-purpose ExtremePose-Face-HQ dataset
+
+Trung Tuan Dao $^{1,*}$  Duc Hong Vu $^{1,2,*}$  Cuong Pham $^{1,3}$  Anh Tran $^{1}$
+
+$^{1}$ VinAI Research, Vietnam  $^{2}$ Miami University  $^{3}$ Posts & Telecommunications Inst. of Tech., Vietnam
+
+{v.trungdt21,v.ducvh5,v.cuongpv11,v.anhtt152}@vinai.io vudh3@miamioh.edu cuongpv@ptit.edu.vn
+
+*Equal contribution
+
+![](images/cbf16de748627893751e10fb88a1c1970cd6edd74202d56555bea8261d2edfef.jpg)  
+Figure 1. Benefits of our proposed dataset (EFHQ). Standard large-scale facial datasets have most images at near frontal views, causing inferior performance of trained models on downstream tasks when dealing with extreme head poses. For instance, the trained 2D image generators and text-to-image ones often produce only near frontal faces, while the 3D face generators and face reenactment methods often show distorted outputs at profile views. The recently proposed dataset LPFF [47] partially handles that issue by providing complementary images at extreme head poses for only 2D and 3D image generation tasks. Our proposed dataset EFHQ provides high-quality extreme-pose images to complement a wide range of face-related tasks. It supports 2D and 3D image generation, with generally better diversity than LPFF. EFHQ also helps correct the outputs of text-to-image generation and face reenactment at extreme views. Finally, EFHQ provides a more challenging pose-based face verification benchmark to better assess the quality of face recognition networks.
+
+# Abstract
+
+The existing facial datasets, while having plentiful images at near frontal views, lack images with extreme head poses, leading to the downgraded performance of deep learning models when dealing with profile or pitched faces. This work aims to address this gap by introducing a novel dataset named Extreme Pose Face High-Quality Dataset (EFHQ), which includes a maximum of 450k high-quality images of faces at extreme poses. To produce such a massive dataset, we utilize a novel and meticulous dataset processing pipeline to curate two publicly available datasets, VFHQ and CelebV-HQ, which contain many high-resolution face videos captured in various settings. Our dataset can complement existing datasets on various facial-related tasks, such as facial synthesis with 2D/3D-aware GAN, diffusion-based text-to-image face generation, and face reenactment. Specifically, training with EFHQ helps models generalize well across diverse poses, significantly improving performance in scenarios involving ex
+
+treme views, confirmed by extensive experiments. Additionally, we utilize EFHQ to define a challenging cross-view face verification benchmark, in which the performance of SOTA face recognition models drops  $5 - 37\%$  compared to frontal-to-frontal scenarios, aiming to stimulate studies on face recognition under severe pose conditions in the wild.
+
+# 1. Introduction
+
+The human face is a central focus in computer vision, with extensive research dedicated to tasks like detection, recognition, generation, and manipulation [39, 42]. Numerous large-scale facial datasets[9, 12, 14, 24, 28, 30, 45, 48, 54, 55] have facilitated the face-related studies. Along with deep learning techniques, many tasks have made a considerable leap in performance in recent years. For instance, face recognition systems can recognize near-frontal faces in the wild with near-perfect accuracy [13, 18]. Facial generative models, such as [5, 15, 49], can synthesize realistic
+
+images indistinguishable from real faces. Face reenactment systems, including [41, 43, 51], can animate an input face based on poses and expressions from any driver video.
+
+Despite impressive results, the mentioned models produce inferior results when dealing with faces at extreme head poses. As illustrated in Fig. 1, faces synthesized by 3D face generation or face reenactment models are often distorted at profile views, due to the scarcity of extreme pose images in training datasets. Current extensive datasets for face generation leverage portrait images from Google or stock photo repositories like Flickr, which feature images with upfront views and minimal facial rotations. Face reenactment methods leverage video datasets such as Vox-Celeb [29], primarily featuring frames with frontal faces. The instability of pre-processing tools, particularly face alignment modules when handling extreme pose images, further reduces the proportion of those images in the existing datasets. This issue is critical since faces with large head rotations are common in real life, and dealing with them is unavoidable in practical systems.
+
+This paper aims to fill that void by proposing a novel large-scale dataset of extreme-pose and high-quality facial images called ExtremePose-Face-HQ, or EFHQ for short. Our ambition is to build an "one-for-all" dataset that can supplement existing datasets on all mentioned face-related tasks. To achieve that goal, the dataset needs all the following properties: (1) extreme-pose, (2) large-scale, (3) high-quality, (4) in-the-wild, (5) having multiple images for each subject and having subject identity annotated. Given those requirements, we utilize a novel and meticulous dataset processing pipeline to filter and sample frames from two publicly available datasets, including VFHQ [48] and CelebV-HQ [54], which contain high-resolution face videos captured in various settings. This results in a dataset with 450k high-quality images of extreme poses, which are combined with frontal images of the same subject to use as image pairs. These image pairs can support face reenactment and face recognition tasks, while the extreme-pose images alone can complement the existing datasets of face generation training. Note that there is a recent attempt [47] to build a large-pose facial dataset similar to ours. However, that work relies on images crawled from Flickr, thus having limited size (19k images) and missing identity information. Hence, its applications are restricted, covering only several face-generation tasks, unlike our large-scale and multi-purpose dataset.
+
+We perform a comprehensive set of experiments to validate the effectiveness of EFHQ across three distinct face generation sub-tasks and face reenactment scenarios. Our dataset, EFHQ, when combined with standard training sets, significantly enhances the quality of synthesized images in highly profile or pitched views while maintaining synthesis quality in frontal views. EFHQ also provides a more
+
+challenging pose-centric face verification benchmark. It reveals the vulnerability of many state-of-the-art face recognition networks when dealing with extreme-pose images, with TAR@FAR=1e-3 scores dropping 5-37%.
+
+In summary, our contributions include (1) a large-scale, high-quality, in-the-wild, extreme-posed dataset (EFHQ) designed to complement various face-related tasks, (2) a novel and meticulous dataset processing pipeline ensuring the creation and quality of the EFHQ dataset, and (3) a series of experiments, with plan-to-release models, demonstrating the usefulness of EFHQ in enhancing performance on a wide range of tasks when dealing with faces at extreme rotations. The final dataset will be publicly available as metadata files to not violate the two original dataset's licenses, alongside the pretrained networks.
+
+# 2. Related Work
+
+Synthetic Face Generation. Existing 2D-based models like StyleGAN [14, 15, 17] and 3D-aware models like EG3D [5] demonstrate impressive fidelity, especially for near-frontal views. Nevertheless, their performance significantly degrades when faced with extreme views, primarily due to insufficient training data for such scenarios. Datasets like FFHQ [14] exhibit a notable bias toward near-frontal poses, heavily limiting the range of poses synthesized by current 2D and 3D models despite the 3D-aware generators' ability to handle arbitrary poses.
+
+Diffusion-based text-to-image models [32, 33] can synthesize high-quality images from simple text inputs. However, these models lack mechanisms to control the generated pose. Consequently, the generated facial images remain in a frontal view despite terms like "profile" or "looking-up" in the input prompts. ControlNet[49] proposes a finetuning approach allowing these models to take additional context as conditional input. Namely, one can use landmarks to guide the generated face to the desired pose. However, the scarcity of profile-view data in training affects the model's performance, particularly in extreme angles. Consequently, effectively managing the long-tailed distribution of these extreme views becomes necessary to address this issue.
+
+Face Reenactment. Face reenactment involves transferring facial expressions and movements from a driving image to a source image. Recent advancements claim high-quality results, especially for near-frontal poses [31, 41, 43, 51]. However, pose collapse persists, particularly in keypoint-based models [41, 51]. Common datasets for this task are VoxCeleb1 [29] and HDTF [50], which primarily feature frontal talking face videos. Despite recent potentially high-quality datasets for reenactment such as VFHQ [48] and CelebV-HQ [54], the frontal-view bias persists. As a result, facial reenactment methods often produce artifacts and lower-quality outputs when faced with significant rotations.
+
+<table><tr><td rowspan="2">Dataset</td><td></td><td colspan="3">Attributes</td><td colspan="3">Ready-to-use tasks</td></tr><tr><td>#images</td><td>Multiple images per ID</td><td>High Quality</td><td>Profile View</td><td>Face Verification</td><td>Face Synthesis</td><td>Face Reenactment</td></tr><tr><td>AgeDB</td><td>12K</td><td>✓</td><td></td><td></td><td>✓</td><td></td><td></td></tr><tr><td>CPLFW</td><td>12K</td><td>✓</td><td></td><td>✓</td><td>✓</td><td></td><td></td></tr><tr><td>CFP-FP</td><td>14K</td><td>✓</td><td></td><td>✓</td><td>✓</td><td></td><td></td></tr><tr><td>IJB-x</td><td>138K</td><td>✓</td><td></td><td>✓</td><td>✓</td><td></td><td></td></tr><tr><td>CelebA-HQ</td><td>30K</td><td></td><td>✓</td><td></td><td></td><td>✓</td><td></td></tr><tr><td>FFHQ</td><td>70K</td><td></td><td>✓</td><td></td><td></td><td>✓</td><td></td></tr><tr><td>LPFF</td><td>19K</td><td></td><td>✓</td><td>✓</td><td></td><td>✓</td><td></td></tr><tr><td>CelebV-HQ</td><td>6M</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>✓</td><td></td></tr><tr><td>VFHQ</td><td>3M</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>✓</td><td>✓</td></tr><tr><td>VoxCeleb1</td><td>4M</td><td>✓</td><td></td><td></td><td></td><td>✓</td><td>✓</td></tr><tr><td>Ours</td><td>450,538</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr><tr><td>FFHQ+Ours*</td><td>112,671</td><td></td><td>✓</td><td>✓</td><td></td><td>✓</td><td></td></tr><tr><td>Vox+Ours*</td><td>4.5M</td><td>✓</td><td></td><td>✓</td><td></td><td></td><td>✓</td></tr></table>
+
+Table 1. Representative comparison of key attributes and supported tasks across existing face datasets and our proposed EFHQ dataset. EFHQ provides large-scale, high-quality images with multiple views per identity, enabling diverse face-related tasks, including synthesis, reenactment, and verification. Combining EFHQ with existing datasets like FFHQ or VoxCeleb can augment available data for selected tasks. Asterisk (*) indicates a combined dataset with a subsampled version of EFHQ.
+
+Face Verification. Despite near-perfect accuracy on benchmark datasets[12, 20, 27, 28, 44], state-of-the-art face verification models[7, 13, 18] fall short in unconstrained environments due to a lack of pose diversity. While CPLFW[52] introduces 6000 cross-pose pairs, its small size and mix of pose samples hinder analyzing performance on specific pose scenarios such as frontal-frontal, frontal-profile, and profile-profile. More granular benchmarking on diverse pose pairs is crucial to identify failure modes and drive progress in robust face verification across poses.
+
+Extreme Pose Face Dataset. Tab. 1 summarizes existing datasets related to face verification, synthesis, and re-enactment, showing that existing face datasets have limitations for multi-task use. Recent efforts like LPFF[47] made progress by crawling from Flickr and processing for diverse views, resulting in a dataset of 19,590 images. The dataset could complement high-quality datasets like FFHQ[14] for generative models. However, due to its limited scale comprising only  $20\%$  of the combined dataset, the unbalanced distribution between frontal and profile view still needs further improvements. Additionally, since the images are from image hosting services, the dataset fails to provide the identity information necessary for other facial applications.
+
+# 3. Dataset Process
+
+In this section, we discuss the process of producing our novel dataset EFHQ. To fulfill our goals, we carefully curated frames from the two recent facial video datasets, VFHQ[48] and CelebV-HQ[54], as seen in Fig. 2.
+
+# 3.1. Attributes Preparation
+
+Attributes Extraction. To curate accurate facial poses and high-quality frames from existing datasets, we first ex
+
+tract (1) face bounding boxes, (2) facial landmarks (5/68 keypoints), (3) image quality score, and (4) face identity. For bounding boxes and landmarks, we employ the popular RetinaFace[36] and SynergyNet[46]. We use HyperIQA [40] to grade image quality. For VFHQ, we only generate 68 landmarks and image quality scores since bounding boxes and 5 keypoints are provided. After extracting these attributes, we match them with existing annotations using an IoU-based Hungarian matching algorithm [21]. Regarding identity, VFHQ-wise, we source labels from the annotations. As CelebV-HQ lacks individual identity labels within the video, we use[35] to identify the identities associated with each bounding box.
+
+Due to the instability of head pose estimators when handling extreme-pose images, we strategically apply multiple pose estimators relying on various methodologies: (1) SynergyNet[46], a 3DMM-based model, (2) DirectMHP[53], a landmark-free joint head detector and pose estimator, and (3) an in-house FacePoseNet[6] improved through extensive data augmentation focused on extreme poses and extra training data. We then categorize each estimated pose into a hierarchical binning scheme illustrated in Fig. 2. Next, we gather bin predictions and perform majority voting to arrive at consensus labels. When no agreement emerges, we put it into the  $5^{th}$  bin for "confusing" cases. As depicted in Fig. 3, this ensemble approach provides robustness in cases where a single estimator incorrectly categorizes a sample. Please refer to the supplementary for statistics of the extracted bins and the hyperparameters for each model used for attribute extraction.
+
+Annotation Review. Next, we manually review the results to ensure high-quality labels. Given the large dataset, we developed a graphical user interface tool to streamline the review process. Our focus is verifying the binning annota
+
+![](images/f254cce2dcbe27b3334c49712a16370a9280f177f1bfe4f41e39c028a6f522d5.jpg)  
+Figure 2. The pipeline of EFHQ dataset creation. Starting with high-quality videos from the VFHQ[48] and CelebV-HQ[54] datasets, single-frame attributes are extracted then manually reviews. Task-specific preprocessing is then applied to generate specialized versions of the dataset for tasks such as face generation, reenactment, and verification.
+
+![](images/38c77dbb4c806449587b476c0b22ce3f7a86ea80b9d0b108ab1e63be1859d8d5.jpg)  
+Figure 3. Example cases where a pose estimator fails to categorize the sample to the correct bin.
+
+tions, which most directly impact label accuracy. Concurrently, we randomly subsample the data based on pose angle and image quality to validate facial landmark quality further and discard cases of landmark prediction failure.
+
+# 3.2. EFHQ dataset
+
+Our final EFHQ dataset comprises up to 450k frames with extreme poses extracted from approximately 5,000 clips. Most of the clips in our dataset include at least one frame with a frontal face and multiple frames with extreme pose angles. We also include extreme pose-only clips, representing profile-to-profile pose transfer cases.
+
+![](images/caec04bb3476c53e27cfcf754fc7da82c83f7eb861fc3f48a721d325f51ab89b.jpg)  
+Figure 4. Pose distribution comparison between our sampled dataset and other datasets, including FFHQ[14], LPFF[47], CPLFW[52] and 40K random samples from VoxCeleb1 [29]. Our sampled dataset demonstrates greater pose diversity, with increased sample counts across high angle bins.
+
+# 3.3. Dataset for subtasks
+
+Supplementary datasets for face generation. To address FFHQ's pose distribution gap, we compile a dataset encompassing diverse poses, varied identities, and FFHQ-comparable image quality from EFHQ. Additionally, we also embed an image brightness filter[3] in the sampling process, given that HyperIQA doesn't explicitly assess this aspect. The sampled images are then processed using the same pipeline in [5, 16] and manually reviewed. The aim is to improve coverage of under-represented poses without degrading frontal performance of trained generation models. This yields a dataset of 42,671 images with equitably distributed poses as visualized in Fig. 4. Compared to the original FFHQ distribution, yaw and pitch distributions are significantly improved.
+
+Additionally, for 3D-aware GAN subtasks, we extract camera parameters to serve as conditional pose information, following [5]. We convert these parameters into yaw, pitch angles to validate against our existing annotations. When disagreements emerge, we re-examine the images and filter if needed, to ensure accurate, quality labels.
+
+For diffusion-based text-to-image generation, a specialized dataset is curated to refine Stable Diffusion models[33] by accommodating landmark conditional input through ControlNet[49]. Based on our 3D-aware face generation dataset, we integrate landmark-based conditional images and tailored text prompts for each image. To capture fine facial nuances, we extract 478 facial landmarks and draw the condition image with connected edges between matched keypoints, following Mediapipe framework[26], better-conveying face expression and pose. For text prompts, detailed facial attributes like gender, race, and emotion are derived from existing labels or inferred via BLIP-2 pretrained captioning model [23] and a face attribute estimator [37]. The resulting prompt adheres to a structured format: "A profile portrait image of a [emotion] [race] [gender]."
+
+<table><tr><td>Model</td><td>Reference Dataset</td><td>FID↓</td><td>Recall↑</td></tr><tr><td>GFFHQ</td><td>FFHQ</td><td>2.84</td><td>0.49</td></tr><tr><td>GFFHQ+LPFF</td><td>FFHQ+LPFF</td><td>3.43</td><td>0.44</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ+EFHQ</td><td>3.44</td><td>0.46</td></tr><tr><td>GFFHQ+LPFF</td><td>FFHQ+EFHQ</td><td>8.26</td><td>0.43</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ+LPFF</td><td>7.04</td><td>0.44</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ+EFHQ</td><td>3.33</td><td>0.46</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ</td><td>3.12</td><td>0.44</td></tr><tr><td>GFFHQ+EFHQ</td><td>FFHQ+LPFF</td><td>4.47</td><td>0.42</td></tr></table>
+
+Table 2. StyleGAN2-ADA models trained on different datasets. Lower FIDs indicate better fidelity, while higher Recalls indicate better diversity. The second and third blocks show comparisons between unconditional models trained on FFHQ+LPFF and FFHQ+EFHQ when evaluating on the same and cross-dataset settings, with the better metrics in bold.
+
+Supplementary dataset for face reenactment. As EFHQ is crafted, we can employ the entire dataset to complement established training datasets, e.g., VoxCeleb1 [29]. We also built an extra evaluation set with 1200 EFHQ clips.
+
+Benchmarking dataset for face verification. We curate a dataset covering three distinct scenarios: frontal-to-frontal, frontal-to-profile, and profile-to-profile. To enable rigorous benchmarking, we sample 10,000 pairs each for both negative and positive cases per scenario, resulting in a balanced benchmark dataset containing 60,000 image pairs. Images follow the same established preprocessing pipelines in [7]. Next, we filter misaligned images with RetinaFace[36], randomly review, and replace equivalent samples from our diverse corpus, if needed. Finally, to simulate varying image quality, we randomly applied downscaling and compression to samples in the dataset.
+
+# 4. Face Generation Subtask
+
+# 4.1. StyleGAN
+
+Training details. We train StyleGAN2-ADA[15] models from scratch after combining FFHQ with our own curated dataset (denoted FFHQ+EFHQ) at  $1024 \times 1024$  resolution. We also train a conditional version, allowing direct control over frontal versus profile synthesis and enhancing model performance. Specifically, we categorize the dataset into two classes: "frontal" and "profile" based on pose angles. All models are trained for 6 days on 8 Nvidia A100 GPUs.
+
+Evaluation protocol. We utilize the Frechet Inception Distance[11] (FID) metric and improved Recall[22] to quantify the generation fidelity and diversity. Regarding the two unconditional models trained with the supplemented FFHQ+LPFF and FFHQ+EFHQ, we evaluate each model with both of these datasets. For conditional mod
+
+![](images/3dcea56d5362a57eb679bd30469ce22064230f4b8d123e344efe25aaf2dc37ea.jpg)  
+(a) Samples from the model trained with FFHQ+EFHQ.
+
+![](images/3c0d5c9941056f9827add44aae19a2e757bb2b54eaf548fa151058eafdd04f75.jpg)  
+(b) Profile-view comparison of LPFF (left) and EFHQ (right).  
+Figure 5. Qualitative result and comparison of generated samples, with truncation  $\psi = 0.7$ , from StyleGAN2-ADA training with other dataset and ours.
+
+els, we evaluate performance on FFHQ, FFHQ+LPFF, and FFHQ+EFHQ to demonstrate the broader coverage of the model trained with our proposed dataset. For FFHQ, we generate 50K images with the "frontal" label, whereas for the other two datasets, we generate 50K following their label distributions. This comprehensive quantitative analysis examines the fidelity and variety of the generated images.
+
+Experimental results. We compared the performance of our proposed unconditional models  $(G^{\mathrm{FFHQ + EFHQ}})$ , against models trained on the FFHQ and FFHQ+LPFF datasets  $(G^{\mathrm{FFHQ}}$  and  $G^{\mathrm{FFHQ + LPFF}})$ . Models trained conditionally are specifically subscripted, e.g.,  $G_{c}^{\mathrm{FFHQ + EFHQ}}$ . The qualitative results in Fig. 5 illustrate that our models generate high-quality frontal faces comparable to the FFHQ model while producing realistic and varied profile faces. Additionally, our method has less noise in extreme poses than the FFHQ+LPFF model. Specifically, Fig. 5b shows LPFF's generated face contains more noise and less photorealistic details than our model, such as the pixelated eye and nose patches. Meanwhile, ours achieves greater realism with smoother skin, well-defined features, noise reduction, and natural lighting. The quantitative results in Tab. 2 show that our unconditional model achieves competitive FID scores and improved Recall to other models, indicating similar fidelity but greater sample pose diversity. In the cross-examination (third block of Tab. 2),  $G^{\mathrm{FFHQ + EFHQ}}$  has both better FID and improved Recall on the opposite dataset than  $G^{\mathrm{FFHQ + LPFF}}$ . Finally, our conditional model demonstrates substantially improved FID scores over multiple reference datasets, highlighting its versatility.
+
+<table><tr><td>Model</td><td>Reference Dataset</td><td>FID↓</td><td>Pose↓</td><td>Depth↓</td><td>ID↑</td></tr><tr><td>EFHQ</td><td>FFHQ/Trainset</td><td>4.8</td><td>5e-4</td><td>0.234</td><td>0.74</td></tr><tr><td>EFFHQ</td><td>FFHQ/Trainset</td><td>3.7</td><td>5e-4</td><td>0.288</td><td>0.74</td></tr><tr><td>EFFHQ+LPFF</td><td>FFHQ</td><td>6.2</td><td>6e-4</td><td>0.316</td><td>0.77</td></tr><tr><td>EFFHQ+EFHQ</td><td>FFHQ</td><td>3.9</td><td>4e-4</td><td>0.224</td><td>0.77</td></tr><tr><td>EFFHQ+LPFF</td><td>Trainset</td><td>7.0</td><td>6e-4</td><td>0.316</td><td>0.65</td></tr><tr><td>EFFHQ+EFHQ</td><td>Trainset</td><td>4.6</td><td>4e-4</td><td>0.223</td><td>0.65</td></tr></table>
+
+![](images/8e70a13b656bbde69089723c284a0af97e6cca6c1c51f5d55a61b3fe1f63ec4d.jpg)  
+Figure 6. Comparison between multiview generated samples, with truncation  $\psi = 0.8$ , of EG3D model trained with various datasets.
+
+# 4.2. EG3D
+
+Training details. We followed the same pipeline and hyper-parameters as the original EG3D model [5]. However, based on empirical results, we adjusted  $\gamma = 5$ . We also fixed the triplane implementation bug (XY, XZ, ZX) [1] to replicate the original idea in the papers (XY, XZ, ZY). We trained the model on 8 Nvidia A100 40GB GPUs for approximately 7 days until convergence.
+
+Evaluation Protocol. To evaluate the models, mirroring [5], we measure image quality with FID; multi-view consistency and identity preservation with mean cosine similarity of pre-trained face recognition model[7]; pose accuracy and geometry by calculating MSE against pseudo labels from [8]. We evaluate all models on 1,024 generated images, except for FID, which we compute on 50k images.
+
+Experimental results. To evaluate the benefits of our dataset, we trained the same backbone architecture on different datasets: FFHQ  $(E^{\mathrm{FFHQ}})$ , FFHQ+LPFF  $(E^{\mathrm{FFHQ} + \mathrm{LPFF}})$ , and FFHQ+EFHQ (our model,  $E^{\mathrm{FFHQ} + \mathrm{EFHQ}}$ ). The quantitative results are presented in Table 3, and qualitative examples are shown in Figure 6. Qualitatively, we observe that  $E^{\mathrm{FFHQ} + \mathrm{EFHQ}}$  generates improved face shapes in extreme poses compared to  $E^{\mathrm{FFHQ}}$ , similar to the improvements seen with  $E^{\mathrm{FFHQ} + \mathrm{LPFF}}$ . Quantitatively, the FID
+
+Table 3. EG3D models trained on different datasets. In the second and third blocks, we compare models trained on FFHQ+LPFF and FFHQ+EFHQ with the testset is either FFHQ or their training datasets; the better metrics are in bold. The best metrics on the FFHQ testset are in red, while the best metrics when evaluating on training dataset are underlined.  
+
+<table><tr><td>Metrics</td><td>Dataset</td><td>Pretrained ControlNet</td><td>Ours</td></tr><tr><td>NME↓</td><td>LPFF</td><td>0.085</td><td>0.084</td></tr><tr><td>FID↓</td><td>LPFF</td><td>119.74</td><td>83.39</td></tr><tr><td>FIDeyepatch ↓</td><td>LPFF</td><td>103.77</td><td>49.45</td></tr><tr><td>NME↓</td><td>EFHQ</td><td>0.106</td><td>0.081</td></tr><tr><td>FID↓</td><td>EFHQ</td><td>111.23</td><td>81.90</td></tr><tr><td>FIDeyepatch ↓</td><td>EFHQ</td><td>137.43</td><td>58.60</td></tr></table>
+
+Table 4. Quantitative results of fine-tuning StableDiffusion v1.5 with ControlNet trained on various datasets, using Normalized Mean Error (NME), FID, and FID on eye region.
+
+of  $E^{\mathrm{FFHQ + EFHQ}}$  is comparable to  $E^{\mathrm{FFHQ}}$  when evaluated on FFHQ, indicating no degradation in frontal view performance. Meanwhile, the competitive FID achieved by  $E^{\mathrm{FFHQ + EFHQ}}$  on the combined FFHQ+EFHQ dataset also suggests proficient modeling of profile views. Regarding identity consistency, we see a significant drop when the reference dataset contains profile-view images, due to the weak confidence of the pretrained face recognition model when handling profile-view images, as confirmed by its low average cosine similarity of around 0.6 on the simple frontal-to-profile verification set CFP-FP[34]. Furthermore, our model also outperforms in pose accuracy and geometry quality with both frontal and cross-pose dataset.
+
+![](images/91c1c39d34e1ff469ab19168be92e10d78b69610e53bee1af6b1b4b5daca6041.jpg)  
+(a) Generated samples; condition image is in between samples.
+
+![](images/e2849aa6187df629ee39998db75c92f3603e153a68b749d5b56f72afd4fcceeb.jpg)
+
+![](images/ea5572592c76c179b7a23caa798b0464f0d77a81bf0014c7a81b2a1c2f7a141b.jpg)  
+(b) Generated eye regions.
+
+![](images/42ea89334a93b892c5276bce82ddb8d2426f524d79dcfa74a222ced094510ff8.jpg)  
+Figure 7. Comparison profile-view generated samples of pretrained ControlNet (left) and our fine-tuned ControlNet (right) with the prompt: "A profile portrait image of a person".
+
+# 4.3. ControlNet
+
+Training details. We finetuned Stable Diffusion v1.5 [33] using ControlNet [49] on the EFHQ dataset to enhance its ability to generate extreme facial poses. We use AdamW optimizer [25] with learning rate  $1 \times e^{-5}$ ,  $\beta_0 = 0.9$ ,  $\beta_1 = 0.999$ , and  $\epsilon = 0.01$ . We train with batch size 24 on 1 Nvidia A100 40GB GPU in 2 days.
+
+Evaluation Protocol. To assess image qualities across diverse poses, we measure the Normalized Mean Error (NME) of 478-point facial landmarks between the generated and source image. Specifically, we evaluate performance across the full pose distribution in the reference
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">Test Dataset (L1 ↓/ AKD ↓/ AED ↓)</td></tr><tr><td>VoxCeleb1</td><td>EFHQ (Ours)</td></tr><tr><td>TPS(VoxCeleb1</td><td>0.041/1.214/0.132</td><td>0.061/3.814/0.275</td></tr><tr><td>TPS(VoxCeleb1+EFHQ</td><td>0.041/1.250/0.133</td><td>0.051/1.963/0.214</td></tr><tr><td>LIA(VoxCeleb1</td><td>0.044/1.342/0.138</td><td>0.059/3.807/0.266</td></tr><tr><td>LIA(VoxCeleb1+EFHQ</td><td>0.044/1.398/0.142</td><td>0.054/2.291/0.233</td></tr></table>
+
+dataset. For the prompt, we use: "A profile portrait image of a person." with extra positive and negative keywords, further detailed in supplementary. Utilizing LPFF and EFHQ datasets as the reference datasets, we compute FID to assess synthesized face quality and extend evaluation to cropped eye regions due to unrealistic gazes observed from the original ControlNet model. Note that the models receive condition images created akin to its training data.
+
+Experimental results. Fig. 7a visually compares the baseline ControlNet OpenPose face and our fine-tuned model. Our fine-tuned model adeptly conditions diverse extreme poses, generating high-quality results with precise details while maintaining fidelity to the condition, such as mouth shape, and exhibiting minimal artifacts. These patterns indicate the superiority of the fine-tuned model over the baseline model in handling extreme viewing angles. Additionally, Fig. 7b contrasts the eye regions of both models. Our fine-tuned model accurately captures natural eye gaze and shape from condition images, a notable improvement over the baseline. Quantitative results in Tab. 4 consistently favor our fine-tuned model across all metrics. FID scores on full and eye patches exhibit significant enhancement, and NME performance remains stable on training and reference datasets (EFHQ and LPFF). These findings showcase the fine-tuned model's capability to generate high-quality, geometrically consistent faces across diverse poses.
+
+# 5. Face Reenactment Subtask
+
+# 5.1. Evaluation Protocol
+
+We follow the same preprocessing and evaluation protocol as FOMM [38]. Specifically, all the frames are cropped to  $256 \times 256$ , and we use the first frame of a clip as the source image to reconstruct the remaining frames. In the case of the EFHQ test set, most source images have a frontal pose, while the target frames have an extreme pose. Some test cases only contain extreme pose to evaluate self-reconstruction on these cases. We run the experiments using two recent methods, Thin-Plate Spline Motion Model (TPS) [51] and Latent Image Animator (LIA) [43]. For training details, please refer to the supplementary. We select the following three metrics to compare the generated frames and the driving ground truth.  $\mathcal{L}_1$  measures the av
+
+Table 5. Quantitative results of two state-of-the-art reenactment methods (TPS and LIA) when training on VoxCeleb1 and the combined set VoxCeleb1+EFHQ. We present results evaluated either on VoxCeleb1 (left) or EFHQ (right) test sets.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Trainset</td><td rowspan="2">CPLFW</td><td colspan="3">EFHQ</td></tr><tr><td>f2f</td><td>f2p</td><td>p2p</td></tr><tr><td>R18arface</td><td>Glint360k</td><td>0.80</td><td>0.94</td><td>0.570.37</td><td>0.880.06</td></tr><tr><td>R18arface</td><td>MS1MV3</td><td>0.75</td><td>0.93</td><td>0.570.36</td><td>0.880.05</td></tr><tr><td>R50arface</td><td>Glint360k</td><td>0.91</td><td>0.97</td><td>0.770.20</td><td>0.920.05</td></tr><tr><td>R100arface</td><td>Glint360k</td><td>0.93</td><td>0.98</td><td>0.830.15</td><td>0.820.16</td></tr><tr><td>R100adaface</td><td>WebFace4M</td><td>0.94</td><td>0.99</td><td>0.940.05</td><td>0.970.02</td></tr></table>
+
+Table 6. Quantitative evaluation of state-of-the-art face recognition models on the cleansed CPLFW and the proposed benchmark dataset. We use the metric TAR@FAR=1e-3 (higher is better). Our EFHQ benchmark includes frontal-frontal (f2f), frontal-profile (f2p), and profile-profile (p2p) subsets. For f2p and p2p, we report the score drop when comparing to f2f in the red subscription. The lowest benchmark score for each model is in bold.
+
+erage distance among pixel values, indicating reconstruction faithfulness. Average Keypoint Distance (AKD) calculates the mean distance between keypoints, thus assessing the pose accuracy. Average Euclidean Distance (AED) measures identity preservation by computing the average  $\mathcal{L}_2$  distance between identity embeddings.
+
+# 5.2. Experimental results
+
+We compare TPS and LIA models trained with our supplementary EFHQ dataset  $(TPS^{\mathrm{VoxCeleb1 + EFHQ}}$  and  $LIA^{\mathrm{VoxCeleb1 + EFHQ}})$  against models trained solely on VoxCeleb1  $(TPS^{\mathrm{VoxCeleb1}}$  and  $LIA^{\mathrm{VoxCeleb1}})$ . Tab. 5 displays quantitative results, revealing that models trained with EFHQ maintain comparable performance when evaluated on VoxCeleb1, emphasizing high-quality performance in frontal-to-frontal reenactment scenarios. Notably, on the EFHQ test set, both EFHQ-trained models show significant improvement in all metrics, indicating enhanced extreme pose and identity reenactment performance. In Fig. 8, qualitative results prove that EFHQ-trained models excel in transferring facial expressions and motion across diverse poses. Models trained on EFHQ showcase improved image quality, reduced artifacts, and well-preserved shapes of driving faces. These results highlight our models' effectiveness in capturing and reproducing facial motions across varying poses while enhancing fidelity to the original facial features.
+
+# 6. Face Recognition Subtask
+
+# 6.1. Evaluation Protocol
+
+To evaluate face recognition performance on extreme poses, we selected two state-of-the-art methods: ArcFace[7] and AdaFace[18]. We further explored various backbone sizes (ResNet18, ResNet50, ResNet101)[10] and training datasets (MS1MV2[9], Glint360K[2], WebFace4M[55]) of increasing size. For metrics, we used True Acceptance Rate (TAR) at False Acceptance Rate (FAR) of  $1 \times e^{-3}$ .
+
+![](images/2c8117e4d7e86d126609d5accefe90974ed636e97faaf9d4022551304474fbb7.jpg)
+
+![](images/f69d710555708f6551cef3b5ce48de1fde2a208e53247211237ab7c6fbae5fcb.jpg)  
+Source
+
+![](images/7d2912606c5b61da5c606f88306fd6e060536c5ef9c314d5d5374ab99e85b6c3.jpg)
+
+![](images/61c559132cc4890d912246b645d7129a8dff94b65997700f60ef27bdef7545eb.jpg)  
+Figure 8. Comparison of performance for frontal-profile face reenactment on EFHQ. From left to right: source image, TPS on VoxCeleb1, TPS with EFHQ, LIA on VoxCeleb1, LIA with EFHQ, driving image. Additional EFHQ training data improves synthesis of extreme poses.
+
+![](images/21f7cb68aa95c4ae50e04e75b16ffcc4a864c7551ef38fd4b4973bd55631b879.jpg)
+
+![](images/74afe71aafafe1944aae621c88642f991e2738f3e52b03082731ebdf1d978f1f.jpg)  
+TPS (VoxCeleb1)  
+TPS (Ours)
+
+![](images/2eb2f40499fafbf4fcb78e429ecb6a3f128bea37332e4a23ac5950ae3a760e86.jpg)
+
+![](images/13c7f831840d7ea36cbc2037bba4722099980fd3cb1850d9ce250d64185c67f0.jpg)  
+LIA (VoxCeleb1)
+
+![](images/6d928c801997e6e85bd8d4dd1dab3f7ab8b88397fe154b822f5f89eac285434f.jpg)
+
+![](images/a37e832c4e7938ac2972ac1b7a72c3542f56a1a39355e19829977df27ba063d8.jpg)  
+LIA (Ours)
+
+![](images/cf4b1d5c92581dc96cd7ada2ad8e39c5f6b911a5e50861f172bbd3c22fc7a3d0.jpg)
+
+![](images/227ecfb2eb75f91aa3aa305d5cb140c2676776284e15ef4c1a063540403f72ea.jpg)  
+Driving
+
+# 6.2. Experimental results
+
+Before assessing face recognition networks on EFHQ, we first clean the commonly-used version of CPLFW, processed by[7], by removing misaligned images (please refer to the Supplementary), and evaluate the models on it. Models performed well on this cleansed set, with minimal gaps between ArcFace and AdaFace.
+
+Our proposed evaluation benchmark involves three pose-based scenarios. Results in Tab. 6 reveal performance drops of  $5 - 37\%$  on the frontal-profile subset compared to frontal-frontal, highlighting the challenge of matching faces in different poses. Larger ResNet models outperformed ResNet18 across subsets, showcasing the advantages of increased capacity in handling pose variation. However, the ResNet100 model trained on Glint360K performed worse than smaller models on the profile-profile subset, indicating the need for sufficient profile data to prevent overfitting. Our results also indicate AdaFace's suitability for handling pose variations. In summary, pose mismatch brings significant challenges for face recognition, necessitating larger models trained on diverse data covering profile faces.
+
+# 7. Discussion
+
+![](images/e17c16d7e1823307bbc687e505298164fa1f4d5649b691e9a0108114e17bb041.jpg)  
+Figure 9. User study results. We reported the percentage of times testers rank whose output is best or tied for each criterion.
+
+User survey. We conduct user surveys to evaluate the face generation improvement of our method through human per
+
+ceptual assessment. In this survey, we compare the outputs of the models mentioned in Secs. 4 and 5. The setup of this user survey is included in the supplementary. In all human evaluation results (Fig. 9), our models outperform previous ones, notably on identity and motion test of face reenactment models (74.80% and 64.22%, respectively).
+
+![](images/5406b417968dda5059ac39836df61475119ae9f04474919d0df72f427ee2ccb5.jpg)  
+Figure 10. Representative examples of the raw LPFF dataset that got excluded from the final dataset, either from misdetections or pose filtering process. Red keypoints represent landmarks prediction of selected samples from LPFF, green keypoints represent landmarks prediction from our pipeline.
+
+Robustness of data processing pipeline. As shown in Fig. 10, our data processing pipeline can detect additional profile-view faces filtered out by the LPFF pipeline, likely due to misdetections or incorrect pose filtering. This confirms the robustness of our pipeline, and we hope it will enable the community to develop larger and higher-quality in-the-wild facial datasets.
+
+# 8. Conclusion
+
+This work has introduced a large-scale, diverse facial dataset to address performance gaps between frontal and profile faces, powered by a novel and robust data processing pipeline. Crucially, we provided tailored sub-datasets for advancing essential tasks like face synthesis and reenactment. Moreover, our new face verification benchmark reveals gaps between techniques, granting valuable insights. Ultimately, we hope this high-quality, diverse facial dataset opens up new and exciting opportunities to push forward cross-pose tasks.
+
+# References
+
+[1] Fix for triplane projection  $\cdot$  Issue #67  $\cdot$  NVlabs/eg3d. https://github.com/NVlabs/eg3d/issues/67, 2023.6  
+[2] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 7  
+[3] Sergey Bezryadin, Pavel Bourov, and Dmitry Ilinih. Brightness calculation in digital image processing. International Symposium on Technologies for Digital Photo Fulfillment, 2007:10-15, 2007. 4, 1  
+[4] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4  
+[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 4, 6  
+[6] Feng-Ju Chang, Anh Tuan Tran, Tal Hassner, Iacopo Masi, Ram Nevatia, and Gerard Medioni. Faceposenet: Making a case for landmark-free face alignment. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1599-1608, 2017. 3  
+[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 3, 5, 6, 7, 8, 1, 2  
+[8] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 6, 3  
+[9] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 1, 7  
+[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778. IEEE, 2016. 7  
+[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 5  
+[12] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 1, 3
+
+[13] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 1, 3  
+[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3, 4  
+[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Proc. NeurIPS, 2020. 1, 2, 5  
+[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 4  
+[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 2  
+[18] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 7  
+[19] Davis E. King. Dlib-ml: A machine learning toolkit. J. Mach. Learn. Res., 10:1755-1758, 2009. 3  
+[20] Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1931-1939, 2015. 3  
+[21] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3  
+[22] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Proceedings of the 33rd International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2019. Curran Associates Inc. 5  
+[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 4  
+[24] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 1  
+[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 6  
+[26] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Wan-Teh
+
+Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for building perception pipelines. CoRR, abs/1906.08172, 2019. 4  
+[27] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 3  
+[28] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 1, 3  
+[29] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. VoxCeleb: A large-scale speaker identification dataset. In Interspeech 2017. ISCA, 2017. 2, 4, 5  
+[30] Aaron Nech and Ira Kemelmacher-Shlizerman. Level playing field for million scale face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1  
+[31] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-Ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 427-436, 2023. 2  
+[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv e-prints, pages arXiv-2204, 2022. 2  
+[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 6  
+[34] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 6  
+[35] Sefik Ilkin Serengil and Alper Ozpinar. Lightface: A hybrid deep face recognition framework. In 2020 Innovations in Intelligent Systems and Applications Conference (ASYU), pages 23-27. IEEE, 2020. 3  
+[36] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 3, 5, 1, 2  
+[37] Sefik Ilkin Serengil and Alper Ozpinar. Hyperextended lightface: A facial attribute analysis framework. In 2021 International Conference on Engineering and Emerging Technologies (ICEET), pages 1-4. IEEE, 2021. 4  
+[38] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 7, 2
+
+[39] L Sirovich and M Kirby. Low-dimensional procedure for the characterization of human faces. Journal of the Optical Society of America A, 4(3):519-524, 1987. 1  
+[40] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3  
+[41] Jiale Tao, Biao Wang, Borun Xu, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Structure-aware motion transfer with deformable anchor model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3637-3646, 2022. 2  
+[42] Matthew A Turk and Alex P Pentland. Face recognition using eigenfaces. In Proceedings. 1991 IEEE computer society conference on computer vision and pattern recognition, pages 586-587. IEEE Computer Society, 1991. 1  
+[43] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In ICLR 2022-The International Conference on Learning Representations, 2022. 2, 7, 3  
+[44] Cameron Whitelam, Emma Taborsky, Austin Blanton, Brianna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K Jain, James A Duncan, Kristen Allen, et al. Iarpa janus benchmark-b face dataset. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 90-98, 2017. 3  
+[45] L Wolf, T Hassner, and I Maoz. Face recognition in unconstrained videos with matched background similarity. In Proceedings of the 2011 IEEE Conference on Computer Vision and Pattern Recognition, pages 529-534, 2011. 1  
+[46] Cho-Ying Wu, Qiangeng Xu, and Ulrich Neumann. Synergy between 3dmm and 3d landmarks for accurate 3d facial geometry. In 2021 International Conference on 3D Vision (3DV), 2021. 3  
+[47] Yiqian Wu, Jing Zhang, Hongbo Fu, and Xiaogang Jin. Lpff: A portrait dataset for face generators across large poses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20327-20337, 2023. 1, 2, 3, 4  
+[48] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2022. 1, 2, 3, 4  
+[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 4, 6  
+[50] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 2  
+[51] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Con
+
+ference on Computer Vision and Pattern Recognition, pages 3657-3666, 2022. 2, 7, 3  
+[52] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7), 2018. 3, 4, 2  
+[53] Huayi Zhou, Fei Jiang, and Hongtao Lu. Directmhp: Direct 2d multi-person head pose estimation with full-range angles. arXiv preprint arXiv:2302.01110, 2023. 3  
+[54] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 1, 2, 3, 4  
+[55] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 7

2024/EFHQ_ Multi-purpose ExtremePose-Face-HQ dataset/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7f02d32324440dae91cf8ef5d64edc18f0dcd5de29b3d16cdd6d700eabe0e64
+size 625399

2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_content_list.json

@@ -0,0 +1,1576 @@
+[
+    {
+        "type": "text",
+        "text": "EFormer: Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting",
+        "text_level": 1,
+        "bbox": [
+            114,
+            128,
+            854,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zitao Wang Qiguang Miao* Yue Xi* Peipei Zhao",
+        "bbox": [
+            259,
+            203,
+            709,
+            222
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Xidian University, China",
+        "bbox": [
+            408,
+            224,
+            558,
+            238
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{21181214174@stu.，qgmiao@，xiyue@，zhpp2023@}xidian.edu.cn",
+        "bbox": [
+            218,
+            241,
+            740,
+            256
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            291,
+            313,
+            306
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The portrait matting task aims to extract an alpha matte with complete semantics and finely detailed contours. In comparison to CNN-based approaches, transformers with self-attention module have a better capacity to capture long-range dependencies and low-frequency semantic information of a portrait. However, recent research shows that the self-attention mechanism struggles with modeling high-frequency contour information and capturing fine contour details, which can lead to bias while predicting the portrait's contours. To deal with this issue, we propose EFormer to enhance the model's attention towards both the low-frequency semantic and high-frequency contour features. For the high-frequency contours, our research demonstrates that cross-attention module between different resolutions can guide our model to allocate attention appropriately to these contour regions. Supported by this, we can successfully extract the high-frequency detail information around the portrait's contours, which were previously ignored by self-attention. Based on the cross-attention module, we further build a semantic and contour detector (SCD) to accurately capture both the low-frequency semantic and high-frequency contour features. And we design a contour-edge extraction branch and semantic extraction branch to extract refined high-frequency contour features and complete low-frequency semantic information, respectively. Finally, we fuse the two kinds of features and leverage the segmentation head to generate a predicted portrait matte. Experiments on VideoMatte240K (JPEG SD Format) and Adobe Image Matting (AIM) datasets demonstrate that EFormer outperforms previous portrait matte methods.",
+        "bbox": [
+            75,
+            323,
+            473,
+            777
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            805,
+            209,
+            821
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Portrait matting aims to extract a precise alpha matte of persons from natural images. It has gathered great momentum in vision applications, including but not limited to photo",
+        "bbox": [
+            76,
+            830,
+            468,
+            878
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/c9937fb43e03a7633aee3ae5e531b87101ba94ad97cd2f40462a3b2430c6fa14.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            290,
+            890,
+            473
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/3b06508ac7a09bba239e870787572c7c30df658ab82ad01b9fb750bf1c4e71f8.jpg",
+        "image_caption": [
+            "Figure 1. (I) Visualization with Grad-CAM of output from each layer within different transformer blocks. (a) The features captured from Self-attention and MLP layers in the original transformer block. (b) In comparison, the features captured from each layer in the proposed transformer block. (II) Comparisons of different transformer blocks. (c) The original transformer block in ViT. (d) The proposed transformer block for portrait matte."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            477,
+            890,
+            656
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "editing and background replacement.",
+        "bbox": [
+            500,
+            794,
+            746,
+            809
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Existing methods[20, 30, 33, 35] attempt to estimate a portrait matte with Convolutional Neural Networks (CNNs) under the guidance of a pre-determined trimap, which is a three-class map that indicates the foreground, background, and unknown region. However, it is time-consuming and labor-intensive to obtain abundant trimaps by manually an-",
+        "bbox": [
+            496,
+            810,
+            893,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding authors.",
+        "bbox": [
+            94,
+            886,
+            227,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "3880",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "notating images. Some methods eliminate the trimap, such as background matting series[21, 28] uses a separate background image instead of the trimap as an auxiliary reference for model prediction. However, in practical cases, it is necessary to provide two aligned images for the model, which include one with only the background and the other one including people. An alternative solution[17, 22, 23] is to rely solely on RGB images as input for predicting portrait matte. Although the previous models have made strides in portrait matte, there is a constraint on the receptive field in CNN-based approaches.",
+        "bbox": [
+            75,
+            90,
+            472,
+            257
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In contrast, Vision Transformer (ViT)[7] and Swin Transformer (Swin) [24] for semantic segmentation employ a self-attention mechanism, which conducts interactions among all pixels in an image to obtain a global receptive field. Previous research [25] shows that ViT and its variants can capture the low-frequency components in images effectively, such as global shapes and structures of a scene or object. However, they usually neglect high-frequency ones in images, such as edges and textures. They work similarly to low-pass filters.",
+        "bbox": [
+            75,
+            258,
+            472,
+            411
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As shown in Fig. 1(a), we can observe the transformer block with self-attention clearly pays little attention to the portrait's high-frequency contour regions. We believe that this phenomenon is attributed to the excessive propagation of global information because global attention is performed in each transformer block during the calculation of multiple blocks. So, the global information is propagated throughout the model in this iterative calculation, ultimately leading it to only focus on low-frequency semantic features with stronger clustering in portraits. Over time, the model gradually ignores the discrete high-frequency detail features around the contours. Therefore, it is desired to capture low- and high-frequency components simultaneously to improve the segmentation performance.",
+        "bbox": [
+            75,
+            412,
+            472,
+            625
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To achieve this goal, Inception Transformer (IFormer)[29] splits all channels into a convolutional path and self-attention path, respectively, to capture high and low frequencies. However, it is required for a channel ratio of each block at different levels in the IFormer. These channel ratios typically require manual adjustment, resulting in significant uncertainty when aiming for optimal model performance. To enhance the model's stability and adaptive capability, we propose a novel architecture EFormer without manually adjusted parameters. In EFormer, we use the cross-attention layer between different resolution features, as shown in Fig. 1(d), which is replaced by convolution to handle high-frequency detail features.",
+        "bbox": [
+            75,
+            626,
+            472,
+            823
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our research further demonstrates that the cross-attention module between different resolutions can formulate a more reasonable attention allocation mechanism, which can accurately capture contour features, as shown in Fig. 1(b). Particularly, EFormer with a cross-attention layer",
+        "bbox": [
+            75,
+            825,
+            472,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "not only can not lose the semantic information within the portrait but also can successfully capture the high-frequency details around the portrait's contours, which were previously ignored by self-attention. It is worth noting that high-frequency components only around the boundary between foreground and background are critical to the improvement of segmentation performance. Therefore, we follow a coarse-to-fine manner. Initially, EFormer needs to locate all high-frequency regions in the image. Then, it filters and extracts the high-frequency regions around the foreground contours.",
+        "bbox": [
+            496,
+            90,
+            890,
+            257
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Based on the cross-attention module, we design a semantic and contour detector (SCD) in EFormer, which cascades the cross-attention layer and self-attention layer to sequentially locate the contour and semantic features. Firstly, the model adjusts its attention to contour features with the guidance of a cross-attention module. Secondly, the self-attention module captures semantic information within a portrait's contours and reverses filtering and correcting the contour information output by the cross-attention layer. This allows the model to gradually capture and match the high-frequency contour information and the low-frequency semantic information. Then, we use multi-layer perceptron (MLP) to build the contour-edge extraction branch (CEEB) and the semantic extraction branch (SEB), respectively. With this support, we independently purify the contour flow and the semantic flow to obtain finer contour information and more comprehensive portrait semantic information, as shown in Fig. 1(b). Finally, we fuse the contour features and semantic features and send them to the segmentation head to estimate the portrait matte. We perform extensive experiments on VideoMatte240K (JPEG SD Format)[21], Adobe Image Matting (AIM)[35], and BG10K. Comparisons with other state-of-the-art models and ablation studies verify that our model performs better than previous works.",
+        "bbox": [
+            496,
+            258,
+            892,
+            619
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To summarize, our contributions are as follows:",
+        "bbox": [
+            517,
+            621,
+            834,
+            636
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose EFormer: an approach that can enhance the transformer's attention towards both semantic and contour features of the foreground.",
+            "- We further build a semantic and contour detector (SCD) to accurately capture the semantic and contour features and design two separate extraction branches to purify the contour flow and the semantic flow.",
+            "- We leverage the cross-attention module between different resolution features in transformer block to autonomously capture and extract high-frequency detail features, which are sparsely distributed along the foreground contour."
+        ],
+        "bbox": [
+            496,
+            638,
+            890,
+            804
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Works",
+        "text_level": 1,
+        "bbox": [
+            500,
+            820,
+            650,
+            835
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Image Matting",
+        "text_level": 1,
+        "bbox": [
+            500,
+            845,
+            651,
+            863
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Image matting is an essential task in the field of computer vision that aims to accurately estimate the foreground in an",
+        "bbox": [
+            500,
+            869,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "3881",
+        "bbox": [
+            482,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "image. Mathematically speaking, an image  $I$  is a combination of an unknown foreground image  $F$  and a background image  $B$  with a probability coefficient alpha matte maps  $\\alpha$ .",
+        "bbox": [
+            75,
+            90,
+            468,
+            137
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nI = \\alpha F + (1 - \\alpha) B \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            199,
+            147,
+            468,
+            164
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Previous image matting solutions have predominantly focused on low-level features, such as color cues[9, 10, 14, 15] or propagation[1, 2, 19], to distinguish the transition areas between foregrounds and backgrounds. However, such traditional matting algorithms commonly struggle to perform properly in complex scenes.",
+        "bbox": [
+            75,
+            174,
+            468,
+            263
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "With the significant advancements making in deep learning, numerous methods based on convolutional neural networks (CNNs) are proposed, leading to notable successes. Some approaches[36, 37, 39] incorporate auxiliary trimap supervisions to enhance matting performance, while other methods[22, 23] leverage trimap-free solutions to estimate alpha matts from image feature maps using an end-to-end segmentation network. In the background matting series[21, 28], an auxiliary input of the background image is used to compute alpha matte. Furthermore, MODNet[17] proposes an end-to-end manner without any auxiliary input. Although prior works based on CNN kernels have improved the accuracy of image matting, the prediction of models is in the limited receptive field.",
+        "bbox": [
+            75,
+            265,
+            468,
+            474
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2. Vision Transformer",
+        "text_level": 1,
+        "bbox": [
+            76,
+            484,
+            264,
+            500
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Different from CNNs, vision transformers with self-attention mechanisms can capture long-term dependencies. Therefore, in the realm of computer vision, vision transformers have gained significant attention in recent times. For pixel-level prediction tasks, such as segmentation, SETR[38], SegFormer[34] and DPT[27] apply transformer as an encoder to attain feature maps of images. Their performances demonstrate the transformer with self-attention is capable of building complete contextual information through global interactions. However, recent studies[25, 29] indicate that transformer \"yet is incompetent in capturing high frequencies that predominantly convey local information\".",
+        "bbox": [
+            75,
+            508,
+            468,
+            702
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Compared to self-attention, cross-attention has more potential to induce customized features. Max-Deeplab[31], MaskFormer[5, 6] and SeMask[13] use query-based methods through cross-attention, which are inspired by DETR[3]. They view segmentation as a set prediction problem. U-Transformer[26] and EUT[8] modify the cross-attention in the original transformer to leverage the information from the encoder, allowing a fine spatial recovery in the decoder. The cross-resolution attention employed by RTFormer[32] enables the gathering of comprehensive contextual information for high-resolution features.",
+        "bbox": [
+            75,
+            704,
+            468,
+            869
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this work, we propose a strategy that using of crossattention between different resolutions guides the model to",
+        "bbox": [
+            76,
+            869,
+            468,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "autonomously locate and capture high-frequency features, particularly those near the portrait's contours. Contrary to IFormer[29], this method eliminates the need for predefined parameters to enhance the model's adaptive capabilities. At the same time, this approach can significantly improve the model's performance in capturing contour details.",
+        "bbox": [
+            496,
+            90,
+            890,
+            181
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3. Refinement for Segmentation",
+        "text_level": 1,
+        "bbox": [
+            498,
+            191,
+            759,
+            207
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the field of portrait segmentation, refining the high-frequency details near the contours is crucial. Many prior studies on segmentation refinement depend on convolutional networks or MLPs that are designed specifically for this purpose.",
+        "bbox": [
+            496,
+            214,
+            890,
+            289
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "PointRend[18], based on the coarse predictions of Mask R-CNN[12], calculates confidence scores as the point-wise uncertainty measure. And it focuses on the points with the higher uncertainty according to set hyperparameters. Finally, PointRend uses a shared MLP to refine the labels of the selected points. Although PointRend has implemented segmentation refinement to some extent, the model still requires manual adjustment of certain hyperparameters to assist in filtering uncertain points. Transfiner[16] processes detected error-prone and incoherent regions with transformer, which include high-frequency regions strewn along object boundaries. However, during this procedure, excessive focus is placed on the local details and incoherent regions, without effectively associating the local features of the object with the overall semantic information.",
+        "bbox": [
+            496,
+            290,
+            892,
+            515
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We believe that locating effective detail distribution areas requires information interaction between local details and global semantics. To achieve this, we design a semantic and contour detector that adopts a cascade structure of cross-attention and self-attention. This approach uses global semantic information garnered from self-attention to autonomously select out local details situated at the borders of the contours. On this base, we further utilize the MLP-based extraction branch to purify contour details, thereby significantly enhancing the effect of refinement.",
+        "bbox": [
+            496,
+            516,
+            890,
+            667
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Model",
+        "text_level": 1,
+        "bbox": [
+            500,
+            681,
+            578,
+            696
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce the process of establishing the complete model architecture of EFormer, which comprises the backbone for encoder, the transformer block for decoder, and the prediction stage, as shown in Fig. 2.",
+        "bbox": [
+            496,
+            705,
+            890,
+            767
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Backbone for Encoder",
+        "text_level": 1,
+        "bbox": [
+            500,
+            776,
+            709,
+            791
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The CNN backbone takes RGB images  $I \\in \\mathbb{R}^{B \\times 3 \\times H \\times W}$  as input, where  $B$  is the number of images, to obtain a pyramid of feature maps  $F = \\left\\{F_{enc\\frac{1}{4}}, F_{enc\\frac{1}{8}}, F_{enc\\frac{1}{16}}\\right\\}$  with dimension  $F_{enc\\frac{1}{i}} \\in \\mathbb{R}^{B \\times C_i \\times \\frac{H}{i} \\times \\frac{W}{i}}$ . As we all know, in the pyramid of feature maps, shallow features have high resolution and rich details, while deep features have low reso",
+        "bbox": [
+            496,
+            797,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "3882",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/19dc18b25a5ce7108cc5db14d5b7e0dec35cccd2f842cb026d2fcf2d6a35cae3.jpg",
+        "image_caption": [
+            "Figure 2. The Architecture of the EFormer. It contains a backbone for encoder, a transformer block for decoder, and the prediction stage, where the transformer block includes a semantic and contour detector, a semantic extraction branch, and a contour-edge extraction branch."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            99,
+            85,
+            874,
+            393
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "lation but more concentrated semantic information. Based on this, our research makes the cross-attention mechanism to filter the high-frequency contour features from high-resolution features, using low-frequency semantic information in low-resolution features. However, since the large resolution of feature maps would bring large computation cost, we use the  $F_{enc\\frac{1}{8}}$  as  $F_{HR}$  (high-resolution features) and use the  $F_{enc\\frac{1}{16}}$  as  $F_{LR}$  (low-resolution features) in our base model. The impact of other choices of  $F_{HR}$  and  $F_{LR}$  on model performance is demonstrated in ablation experiments Sec. 4.4. As shown in Fig. 2, the  $F_{HR}$  serves as the source of contour flow, providing high-frequency contour detail features for the following processes. Correspondingly, the  $F_{LR}$  serves as the source of semantic flow, providing low-frequency semantic information. Then, the  $F_{HR}$  and  $F_{LR}$  are projected to same dimension  $C$  and respectively flattened into high-resolution feature embedding as  $F_{HR}^{em}$  and low-resolution feature embedding as  $F_{LR}^{em}$ , where  $F_{HR}^{em} \\in R^{N\\times B\\times C}$ ,  $F_{LR}^{em} \\in R^{N\\times B\\times C}$ , and  $N = \\frac{H}{8}\\times \\frac{W}{8}$ . Both of them are utilized as inputs for the transformer block.",
+        "bbox": [
+            75,
+            458,
+            472,
+            760
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Transformer Block for Decoder",
+        "text_level": 1,
+        "bbox": [
+            76,
+            771,
+            356,
+            786
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We propose a novel transformer block in the model decoding stage to enhance the ability to capture and extract visual features. It not only does not reduce the focus on low-frequency semantic features but also significantly increases the attention to high-frequency contour features. Its detailed architecture is depicted in Fig. 3. The novel transformer block first uses a Semantic and Contour De",
+        "bbox": [
+            75,
+            794,
+            468,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "tector (SCD) to locate and capture high-frequency contour features and low-frequency semantic features. To achieve this, we cascade a cross-attention layer (CA) and a self-attention layer (SA) in the SCD. Then, as shown in Fig. 3, a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB) are further built to handle high-frequency contour feature flow and low-frequency semantic feature flow respectively. These branches can separately purify more refined high-frequency contour features and extract more complete low-frequency semantic features. This ensures an outcome that is optimal for both types of features.",
+        "bbox": [
+            496,
+            458,
+            890,
+            637
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Semantic and Contour Detector (SCD). There is a fact that low-frequency semantic features far outperform high-frequency contour detail features in terms of quantity and distribution density. So, in the global information propagation process, the low-frequency semantic information easily dominates the feature representations of the original transformer block with self-attention module.",
+        "bbox": [
+            496,
+            641,
+            890,
+            746
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To prevent the loss of high-frequency contour detail information, we first use a cross-attention layer with different resolution features rather than a self-attention layer. The cross-attention layer can capture high-frequency detail features with high semantic correlation in high-resolution features, based on the low-frequency semantic information in low-resolution features. Technically, after getting  $F_{HR}^{em}$  and  $F_{LR}^{em}$ , we use Layer Norm (LN) processing them to maintain consistency. Due to the sensitivity of the attention mechanism towards positional information, we use absolute and",
+        "bbox": [
+            496,
+            750,
+            892,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "3883",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/8b731e62de48786a607962e324bfde9d2bf61d970716c65951847e6993d27c69.jpg",
+        "image_caption": [
+            "Figure 3. Our transformer block. It includes a Semantic and Contour Detector (SCD), a Semantic Extraction Branch (SEB), and a Contour-Edge Extraction Branch (CEEB). The detailed composition of the module is shown in the figure."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            89,
+            88,
+            459,
+            526
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "learnable position encoding (PE) in the SCD. The input  $K$ ,  $Q$ ,  $V$  of the cross-attention layer (CA) is calculated as",
+        "bbox": [
+            75,
+            618,
+            468,
+            648
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nK = L N \\left(F _ {H R} ^ {e m}\\right) + P E \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            657,
+            468,
+            676
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nQ = L N \\left(F _ {L R} ^ {e m}\\right) + P E \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            191,
+            684,
+            468,
+            700
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nV = L N \\left(F _ {L R} ^ {e m} + F _ {H R} ^ {e m}\\right) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            187,
+            705,
+            468,
+            722
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "By performing cross-attention calculations using features of different resolutions, it is possible to automatically locate and capture the discrete contours along the portrait and allocate more attention to corresponding positions. The contour-edge features can be derived from the output of the CA.",
+        "bbox": [
+            75,
+            728,
+            468,
+            816
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {\\text {c o n t o u r - e d g e}} = C A (K, Q, V) \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            161,
+            818,
+            468,
+            835
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Although CA pays extra attention to contour-edge features, it still captures the overall semantics. By the residual connection between CA and  $V$ , both the semantic and contour-edge features of the portrait are enhanced, especially the",
+        "bbox": [
+            75,
+            840,
+            468,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "contour-edge features.",
+        "bbox": [
+            500,
+            92,
+            651,
+            106
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {\\text {e n h a n c e}} = C A (K, Q, V) + V \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            584,
+            116,
+            890,
+            133
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Since semantic information is still present, we use the self-attention layer (SA) after CA. SA can locate the semantic feature within the portrait and contours, thereby emphasizing the semantic attributes at their respective positions. After using LN to normalize  $F_{\\text{enhance}}$ , the input  $K', Q', V'$  of the SA is calculated as",
+        "bbox": [
+            498,
+            143,
+            893,
+            233
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nK ^ {\\prime} = Q ^ {\\prime} = L N \\left(F _ {\\text {e n h a n c e}}\\right) + P E \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            578,
+            243,
+            890,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nV ^ {\\prime} = L N \\left(F _ {\\text {e n h a n c e}}\\right) \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            622,
+            270,
+            890,
+            287
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {\\text {c o n t o u r - s e m a n t i c}} = S A \\left(K ^ {\\prime}, Q ^ {\\prime}, V ^ {\\prime}\\right) + V ^ {\\prime} \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            544,
+            292,
+            890,
+            309
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Through SA filtering the semantic and contour features, the residual connection between  $V'$  and the output of SA, we can produce  $F_{contour - semantic}$  as the output of the semantic and contour detector.",
+        "bbox": [
+            498,
+            315,
+            890,
+            373
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Feature Extraction Branches. After SCD, we use multi-layer perceptron (MLP) to build a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB). To prevent the loss of high-frequency contour detail information, we aggregate  $F_{contour - semantic}$  getting from SCD and  $F_{HR}^{em}$  in the contour flow. After getting the merged feature normalized by LN, we employ CEEB to purify it, in order to further extract and refine high-frequency contour details individually. Similarly, to obtain more comprehensive and consistent portrait semantic features separately, we merge  $F_{contour - semantic}$  and  $F_{LR}^{em}$  in the semantic flow. Then, we normalize the aggregated feature by applying LN and use SEB to filter it.",
+        "bbox": [
+            498,
+            375,
+            892,
+            571
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {\\text {c o n t o u r}} = M L P \\left(L N \\left(F _ {H R} ^ {\\text {e m}} + F _ {\\text {c o n t o u r - s e m a n t i c}}\\right)\\right) \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            514,
+            580,
+            890,
+            612
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {\\text {s e m a n t i c}} = M L P \\left(L N \\left(F _ {L R} ^ {\\text {e m}} + F _ {\\text {c o n t o u r - s e m a n t i c}}\\right)\\right) \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            511,
+            612,
+            890,
+            643
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3. Prediction Stage",
+        "text_level": 1,
+        "bbox": [
+            500,
+            652,
+            663,
+            667
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After getting key characteristics  $F_{semantic} \\in R^{N \\times B \\times C}$  and  $F_{contour} \\in R^{N \\times B \\times C}$  from transformer block, where  $N = \\frac{H}{8} \\times \\frac{W}{8}$ , we transform them into  $F_{semantic}' \\in R^{B \\times C \\times \\frac{H}{8} \\times \\frac{W}{8}}$  and  $F_{contour}' \\in R^{B \\times C \\times \\frac{H}{8} \\times \\frac{W}{8}}$  by converting the vector into the matrix. To achieve improved outcomes in portrait segmentation, we propose fusing the two kinds of features and ultimately acquiring the predicted portrait matte from the segmentation head module.",
+        "bbox": [
+            500,
+            674,
+            890,
+            797
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Fuse. We attain the semantic feature through the residual connection between  $F_{LR}^{\\prime}$  and  $F_{semantic}^{\\prime}$  of semantic flow. Similarly, we get the contour feature from contour flow and subsequently merge the  $F_{semantic}^{\\prime}$  and  $F_{contour}^{\\prime}$  to generate the fused features  $F_{contour - semantic}^{\\prime}$ .",
+        "bbox": [
+            498,
+            797,
+            890,
+            876
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {\\text {s e m a n t i c}} ^ {\\prime} = \\operatorname {C o n v} \\left(F _ {\\text {s e m a n t i c}} ^ {\\prime} + F _ {L R} ^ {\\prime}\\right) \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            562,
+            883,
+            890,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "3884",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {\\text {c o n t o u r}} ^ {\\prime} = \\operatorname {C o n v} \\left(F _ {\\text {c o n t o u r}} ^ {\\prime} + F _ {H R} ^ {\\prime}\\right) \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            147,
+            89,
+            467,
+            108
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {\\text {c o n t o u r - s e m a n t i c}} ^ {\\prime} = \\operatorname {F u s e} \\left(F _ {\\text {s e m a n t i c}} ^ {\\prime} + F _ {\\text {c o n t o u r}} ^ {\\prime}\\right) \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            84,
+            113,
+            467,
+            132
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Segmentation Head. Finally, We gather  $F_{contour - semantic}^{\\prime}$  and  $F_{enc\\frac{1}{4}}$  with higher resolution to predict portrait matte in 1/4 scale of the input image and upscale it to the original scale through bi-linear interpolation as the output of our model.",
+        "bbox": [
+            75,
+            138,
+            467,
+            213
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nm a t t e = H e a d \\left(F _ {c o n t o u r - s e m a n t i c} ^ {\\prime} + F _ {e n c \\frac {1}{4}}\\right) \\tag {15}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            101,
+            222,
+            467,
+            248
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            258,
+            209,
+            276
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Dataset and Evaluation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            284,
+            292,
+            299
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Dataset. We utilize composite training data from two sources: the foreground image dataset VideoMatte240K (JPEG SD Format)[21] and Adobe Image Matting (AIM)[35]. For the background images, we select BG10K following the approach of BGMv2[21], which is a collection of photographs depicting various life scenes without any human portraits. We select foregrounds from VideoMatte240K-JPEG-SD and AIM, and backgrounds from BG10K to composite image datasets. We split VideoMatte240K-JPEG-SD into 234,982/3,007/2720 image sets with image resolutions of  $224 \\times 224$  and  $512 \\times 288$  for training, validating, and testing our model. Similarly, BG10K is split into 9000/1000 image sets, and AIM is split into 214/10 image sets for training and testing, where AIM only includes the portrait images with resolutions at  $512 \\times 512$ . Finally, we use the aforementioned testing datasets to compare and evaluate the performance of EFormer against other models.",
+        "bbox": [
+            75,
+            306,
+            468,
+            578
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation. We mainly consider portrait matting accuracy for evaluation. For portrait matting accuracy, we use Mean Absolute Difference (MAD), Mean Squared Error (MSE), Gradient (Grad), and Connectivity (Conn) as evaluation metrics. We also scale MAD, MSE, Grad, and Conn by  $10^{3}$ ,  $10^{3}$ ,  $10^{-3}$ , and  $10^{-3}$  respectively, for convenience of reference. For all these metrics, the lower number represents better performance.",
+        "bbox": [
+            75,
+            580,
+            467,
+            700
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            76,
+            710,
+            294,
+            726
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Training Settings. For training the network, we use a single RTX 3090 GPU with batch size at 24. The optimizer is AdamW and the initial learning rate is set to  $10^{-4}$ . All models in ablation studies are trained for 25 epochs, with the learning rate decaying by a factor of 0.8 every 5 epochs. Additionally, to augment the data, each input image is subjected to random horizontal flipping.",
+        "bbox": [
+            75,
+            733,
+            467,
+            839
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Backbone. The backbone we use for the network is the pretrained ResNet50[11], since it is the most frequently used backbone in prior works. We use the implementation and weights from torchvision.",
+        "bbox": [
+            75,
+            839,
+            467,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/2b29dd07027dfc431e484454ec4b4fcbcaf97a2dea64be6672f24670cdbd512d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>DeepLabv3[4]</td><td>14.4700</td><td>9.6700</td><td>8.5500</td><td>1.6900</td></tr><tr><td>MODnet[17]</td><td>10.3900</td><td>5.6500</td><td>2.0200</td><td>1.0400</td></tr><tr><td>BGMv2[21]</td><td>4.1858</td><td>1.7934</td><td>1.3011</td><td>0.5326</td></tr><tr><td>RVM[22]</td><td>5.9900</td><td>1.1700</td><td>1.1000</td><td>0.3400</td></tr><tr><td>EFormer</td><td>2.3097</td><td>0.6637</td><td>0.4580</td><td>0.2740</td></tr></table>",
+        "bbox": [
+            513,
+            88,
+            879,
+            195
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/ea643b03dda1a2cfc98ef9e9a8cf8ac9e1fce6990aeb2dc38ccb6b152968a860.jpg",
+        "table_caption": [
+            "Table 1. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at  $512 \\times 288$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>DeepLabv3[4]</td><td>29.64</td><td>23.78</td><td>20.17</td><td>7.71</td></tr><tr><td>MODnet[17]</td><td>21.66</td><td>14.27</td><td>5.37</td><td>5.23</td></tr><tr><td>BGMv2[21]</td><td>44.61</td><td>39.08</td><td>5.54</td><td>11.60</td></tr><tr><td>RVM[22]</td><td>14.84</td><td>8.93</td><td>4.35</td><td>3.83</td></tr><tr><td>EFormer</td><td>7.47</td><td>2.13</td><td>2.83</td><td>1.90</td></tr></table>",
+        "bbox": [
+            519,
+            260,
+            870,
+            367
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Comparisons on the test set of AIM. Bold indicates the best performance among these models under the inputs with the same resolution at  $512 \\times 512$ .",
+        "bbox": [
+            498,
+            378,
+            890,
+            419
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Transformer Block for Decoder. The transformer block is based on the transformer decoder of DETR. We continuously use four transformer blocks in the decoder of the network. We set the channel  $C = 256$  and the number of attention heads  $M = 8$  for each multi-head attention module.",
+        "bbox": [
+            496,
+            443,
+            890,
+            532
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Prediction Stage. After obtaining the semantic and contour feature maps of the portrait from the stacked transformer blocks, we fuse both of them. Then, the alpha mattes predicted by the segmentation head are upsampled to the original size of the input image with bi-linear interpolation.",
+        "bbox": [
+            496,
+            534,
+            890,
+            609
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Loss function. To compute losses during the training stage, we use a Binary CrossEntropy Loss function.",
+        "bbox": [
+            498,
+            609,
+            888,
+            640
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {b c e} = - (g \\times \\log (m) + (1 - g) \\times \\log (1 - m)) \\tag {16}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            516,
+            648,
+            888,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Where,  $m$  is the prediction of the network,  $g$  is the ground truth.",
+        "bbox": [
+            498,
+            672,
+            890,
+            702
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Comparisons to State-of-the-art Methods",
+        "text_level": 1,
+        "bbox": [
+            500,
+            710,
+            851,
+            726
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To demonstrate the progressiveness of EFormer, we compare it against the latest state-of-the-art trimap-free portrait matting solutions. These are including DeepLabV3[4] with ResNet101[11] backbone, BGMv2[21] with ResNet50[11] backbone, MODNet[17], and RVM[22]. Our assessment is based on the testing datasets of VideoMatte240K-JPEG-SD and AIM, respectively. As the training files for MODNet are not available, we use its official weights. Furthermore, since BGMv2, DeepLabV3, and RVM have already been trained on all datasets, our method can be compared with them in a fair manner.",
+        "bbox": [
+            496,
+            734,
+            890,
+            898
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "3885",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/a44cfc197d899a897cdae63917124a1992d807f4b3ea9c1a1cb1af45bf9a7ef0.jpg",
+        "image_caption": [
+            "Figure 4. Visualization of portrait matte predictions from MODnet, BGMv2, and EFormer under challenging image from the test set of VideoMatte240K (JPEG SD Format). Our model shows a better ability to distinguish ambiguous foreground contours, as indicated by the red box. Please zoom in to view the detailed information."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            83,
+            88,
+            893,
+            428
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/d66d74013182b05d955a7fc314b2f295b3303d5a4d703b858ad43c84fa974857.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>DeepLabv3[4]</td><td>18.1586</td><td>14.7623</td><td>6.0015</td><td>2.8824</td></tr><tr><td>MODnet[17]</td><td>13.1400</td><td>8.9656</td><td>2.6028</td><td>2.1543</td></tr><tr><td>BGMv2[21]</td><td>7.0471</td><td>2.3543</td><td>1.5621</td><td>1.0651</td></tr><tr><td>RVM[22]</td><td>6.1357</td><td>2.2647</td><td>0.6523</td><td>0.2866</td></tr><tr><td>EFormer</td><td>4.1357</td><td>1.4280</td><td>0.3760</td><td>0.1753</td></tr></table>",
+        "bbox": [
+            84,
+            501,
+            460,
+            609
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As shown in the Tab. 1, Tab. 2 and Tab. 3, our model exhibits superior performance in portrait matte tasks. It has a stronger capability to distinguish the contour of a portrait. And it can finely segment edge details (e.g. fingers), as exemplified in Fig. 4.",
+        "bbox": [
+            75,
+            686,
+            468,
+            763
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4. Ablation Studies",
+        "text_level": 1,
+        "bbox": [
+            76,
+            771,
+            240,
+            786
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Training Epochs. In the ablation study with an image resolution of  $224 \\times 224$ , we conduct 25 training epochs. Every 5 epochs of training, we evaluate the training effect of the model. At the same time, the learning rate is reduced by 0.8 times. The experimental results indicate that our model achieves optimal performance in the 20 epoch, as shown in the Tab. 4.",
+        "bbox": [
+            75,
+            794,
+            470,
+            898
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/3e01d395210f49963e3779f01492181715d5c54a39bd96f04650def1402a3c47.jpg",
+        "table_caption": [
+            "Table 3. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at  $224 \\times 224$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>5-epoch</td><td>4.3953</td><td>1.4824</td><td>0.4089</td><td>0.1865</td></tr><tr><td>10-epoch</td><td>4.3108</td><td>1.4766</td><td>0.3908</td><td>0.1834</td></tr><tr><td>15-epoch</td><td>4.2867</td><td>1.5238</td><td>0.4087</td><td>0.1833</td></tr><tr><td>20-epoch</td><td>4.1357</td><td>1.4280</td><td>0.3760</td><td>0.1753</td></tr><tr><td>25-epoch</td><td>4.1811</td><td>1.4797</td><td>0.3858</td><td>0.1779</td></tr></table>",
+        "bbox": [
+            532,
+            501,
+            859,
+            609
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Ablation Study on the Training Epochs with the test set of VideoMatte240K-JPEG-SD (resolution at  $224 \\times 224$ ). **Bold** indicates the best performance among these models. We use the result of 20-epoch training as a strong baseline model.",
+        "bbox": [
+            496,
+            619,
+            893,
+            675
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The Functional Role of Each Module. To demonstrate the effectiveness of the model in capturing and extracting contour details and portrait semantics, we analyze the outputs of CA, SA, CEEB, and SEB, and utilize Class Activation Map (CAM) to visualize the attention allocation of each layer's features, as shown in Fig. 5.",
+        "bbox": [
+            496,
+            703,
+            890,
+            792
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "From the first line of Fig. 5, it can be evident that after the cross-attention layer, the model clearly strengthens its attention to the contours of the portrait. Following the contour prior, the self-attention layer captures the portrait's semantics thoroughly. CEEB and SEB models provide sharper contour and more comprehensive semantic information.",
+        "bbox": [
+            496,
+            794,
+            890,
+            883
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Comparing the second and first lines of Fig. 5, it is ob",
+        "bbox": [
+            517,
+            885,
+            890,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "3886",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f4a86092d25db2d4a4c9dd6b16240410a2a59c5e7287d4d91926926071d90d7f.jpg",
+        "image_caption": [
+            "Figure 5. We use Grad-CAM to visualize the attention distribution of various features in CA, SA, CEEB, and SEB. The model in the first row contains four complete modules. For comparison, the SA layer and CA layer were removed from the models in the second and third rows, respectively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            89,
+            467,
+            275
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/9284b5286ab0833f03b0deb33b734389a968a9090ebdffccfb98e06530372592.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>CA</td><td>SA</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>✓</td><td></td><td>4.2762</td><td>1.5344</td><td>0.4092</td><td>0.1823</td></tr><tr><td></td><td>✓</td><td>4.1710</td><td>1.4354</td><td>0.3773</td><td>0.1769</td></tr><tr><td>✓</td><td>✓</td><td>4.1357</td><td>1.4280</td><td>0.3760</td><td>0.1753</td></tr></table>",
+        "bbox": [
+            89,
+            373,
+            455,
+            450
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "served that removing the self-attention layer considerably weakens the semantic information extracted by SEB. In the comparison between the third and first lines of Fig. 5, it can be indicated that removing the cross-attention layer significantly weakens the contour information extracted by CEEB.",
+        "bbox": [
+            75,
+            547,
+            468,
+            625
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "After qualitative analysis, the prediction results of all the above models are evaluated. It is concluded that the model with CA, SA, CEEB, and SEB is both effective and rational, as demonstrated in the Tab. 5.",
+        "bbox": [
+            75,
+            625,
+            468,
+            684
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As is well known, when it comes to extracting features for portrait segmentation tasks, semantic information plays a significant role, in contrast, contour details account for a relatively small proportion. Owing to this, as shown in the Tab. 5 and Fig. 5, the model without SA leads to a significant increase in all metrics due to the lack of amount of semantic information, while the model without CA experiences a slight increase in all indicators due to the loss of contour details. In order to fully harness the potential of the vision transformer, we integrate both CA and SA to ensure semantic completeness and enhance attention towards contour details, thereby achieving more comprehensive and fine-grained segmentation results.",
+        "bbox": [
+            75,
+            686,
+            468,
+            883
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The Selection of  $F_{HR}$  and  $F_{LR}$ . In our base model,",
+        "bbox": [
+            76,
+            885,
+            468,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/e71b74d17dbb4b9a3f2efbc7a347c6138f07d21bd6f697e21c3171dcf0fce96a.jpg",
+        "table_caption": [
+            "Table 5. Ablation Study on the Functional Role of CA and SA with the test set of VideoMatte240K-JPEG-SD (resolution at  $224 \\times 224$ ). **Bold** indicates the best performance among these models."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>FHR</td><td>FLR</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>Fenc1/8</td><td>Fenc1/16</td><td>4.1357</td><td>1.4280</td><td>0.3760</td><td>0.1753</td></tr><tr><td>Fenc1/4</td><td>Fenc1/8</td><td>3.9970</td><td>1.3343</td><td>0.3328</td><td>0.1679</td></tr><tr><td>Fenc1/4</td><td>Fenc1/16</td><td>3.9730</td><td>1.2949</td><td>0.3215</td><td>0.1665</td></tr></table>",
+        "bbox": [
+            500,
+            88,
+            890,
+            170
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6. Ablation Study on the Selection of  $F_{HR}$  and  $F_{LR}$  with the test set of VideoMatte240K-JPEG-SD (resolution at  $224 \\times 224$ ). Bold indicates the best performance among these models.",
+        "bbox": [
+            498,
+            180,
+            890,
+            223
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "considering the large resolution of feature maps would bring large computation cost, we use the  $F_{enc}^{\\frac{1}{8}}$  as  $F_{HR}$  (high-resolution features) and use the  $F_{enc}^{\\frac{1}{16}}$  as  $F_{LR}$  (low-resolution features). Then the  $F_{HR}$  serves as the source of contour flow, providing high-frequency contour detail features. Correspondingly, the  $F_{LR}$  serves as the source of semantic flow, providing low-frequency semantic information. As shown in the Tab. 6, after using  $F_{enc}^{\\frac{1}{4}}$  with higher resolution as  $F_{HR}$ , our model can capture more high-frequency contour detail features and its performance is better. And after using  $F_{enc}^{\\frac{1}{16}}$  with lower resolution as  $F_{LR}$ , our model can use more concentrated low-frequency semantic information, which makes all indicators optimal.",
+        "bbox": [
+            496,
+            258,
+            890,
+            455
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            484,
+            617,
+            500
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a strategy that use of cross-attention module between different resolutions guides the EFormer to autonomously locate and capture high-frequency contour features. Without manual parameter tuning, the proposed method can adaptability capture high-frequency details neglected by previous transformer-based methods. Based on the cross-attention module, we further construct a semantic and contour detector (SCD), which can enhance the transformer's attention to both the high-frequency contour and low-frequency semantic features. Additionally, we build a contour-edge extraction branch (CEEB) and a semantic extraction branch (SEB) to extract finer high-frequency contour details and more comprehensive portrait's low-frequency semantic information. Extensive experiments show that our method outperforms previous portrait matting solutions and exhibits robust performance during testing and inference.",
+        "bbox": [
+            496,
+            515,
+            890,
+            772
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            500,
+            800,
+            660,
+            816
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The work was jointly supported by the National Natural Science Foundations of China under grant No.62272364, the Teaching Reform Project of Shaanxi Higher Continuing Education under Grant No.21XJZ004.",
+        "bbox": [
+            498,
+            829,
+            890,
+            887
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "3887",
+        "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] Yagiz Aksoy, Tunc Ozan Aydin, and Marc Pollefeys. Designing effective inter-pixel information flow for natural image matting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 29-37, 2017. 3",
+            "[2] Yagiz Aksoy, Tae-Hyun Oh, Sylvain Paris, Marc Pollefeys, and Wojciech Matusik. Semantic soft segmentation. ACM Transactions on Graphics (TOG), 37(4):1-13, 2018. 3",
+            "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3",
+            "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6, 7",
+            "[5] 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. 3",
+            "[6] 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. 3",
+            "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2",
+            "[8] Dazhao Du, Bing Su, Yu Li, Zhongang Qi, Lingyu Si, and Ying Shan. Efficient u-transformer with boundary-aware loss for action segmentation. arXiv preprint arXiv:2205.13425, 2022. 3",
+            "[9] Xiaoxue Feng, Xiaohui Liang, and Zili Zhang. A cluster sampling method for image matting via sparse coding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 204-219. Springer, 2016. 3",
+            "[10] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaou Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056. IEEE, 2011. 3",
+            "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6",
+            "[12] 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. 3",
+            "[13] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. 3",
+            "[14] Jubin Johnson, Ehsan Shahrian Varnousfaderani, Hisham Cholakkal, and Deepu Rajan. Sparse coding for alpha mat"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "ting. IEEE Transactions on Image Processing, 25(7):3032-3043, 2016. 3",
+            "[15] Levent Karacan, Aykut Erdem, and Erkut Erdem. Image matting with kl-divergence based sparse sampling. In Proceedings of the IEEE international conference on computer vision, pages 424-432, 2015. 3",
+            "[16] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4412–4421, 2022. 3",
+            "[17] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1140-1147, 2022. 2, 3, 6, 7",
+            "[18] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3",
+            "[19] Anat Levin, Alex Rav-Acha, and Dani Lischinski. Spectral matting. IEEE transactions on pattern analysis and machine intelligence, 30(10):1699-1712, 2008. 3",
+            "[20] Yaoyi Li and Hongtao Lu. Natural image matting via guided contextual attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11450-11457, 2020. 1",
+            "[21] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 2, 3, 6, 7",
+            "[22] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 2, 3, 6, 7",
+            "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2, 3",
+            "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2",
+            "[25] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 2, 3",
+            "[26] Olivier Petit, Nicolas Thome, Clement Rambour, Loic Themyr, Toby Collins, and Luc Soler. U-net transformer: Self and cross attention for medical image segmentation. In Machine Learning in Medical Imaging: 12th International Workshop, MLMI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, September 27, 2021, Proceedings 12, pages 267-276. Springer, 2021. 3",
+            "[27] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "3888",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021.3",
+            "[28] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 2, 3",
+            "[29] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xinchao Wang, and Shuicheng Yan. Inception transformer. Advances in Neural Information Processing Systems, 35:23495-23509, 2022. 2, 3",
+            "[30] Yanan Sun, Guanzhi Wang, Qiao Gu, Chi-Keung Tang, and Yu-Wing Tai. Deep video matting via spatio-temporal alignment and aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6975–6984, 2021. 1",
+            "[31] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5463-5474, 2021. 3",
+            "[32] Jian Wang, Chenhui Gou, Qiman Wu, Haocheng Feng, Junyu Han, Errui Ding, and Jingdong Wang. Rrformer: Efficient design for real-time semantic segmentation with transformer. Advances in Neural Information Processing Systems, 35:7423-7436, 2022. 3",
+            "[33] Yu Wang, Yi Niu, Peiyong Duan, Jianwei Lin, and Yuanjie Zheng. Deep propagation based image matting. In IJCAI, pages 999-1006, 2018. 1",
+            "[34] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3",
+            "[35] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 1, 2, 6",
+            "[36] Haichao Yu, Ning Xu, Zilong Huang, Yuqian Zhou, and Humphrey Shi. High-resolution deep image matting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3217-3224, 2021. 3",
+            "[37] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5128-5137, 2021. 3",
+            "[38] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 3",
+            "[39] Bingke Zhu, Yingying Chen, Jinqiao Wang, Si Liu, Bo Zhang, and Ming Tang. Fast deep matting for portrait animation on mobile phone. In Proceedings of the 25th ACM international conference on Multimedia, pages 297-305, 2017. 3"
+        ],
+        "bbox": [
+            78,
+            92,
+            468,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "3889",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_model.json

@@ -0,0 +1,2035 @@
+[
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.107,
+                0.003,
+                0.182,
+                0.043
+            ],
+            "angle": 0,
+            "content": "CVF"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.237,
+                0.001,
+                0.808,
+                0.047
+            ],
+            "angle": 0,
+            "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.13,
+                0.856,
+                0.177
+            ],
+            "angle": 0,
+            "content": "EFormer: Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.26,
+                0.204,
+                0.71,
+                0.223
+            ],
+            "angle": 0,
+            "content": "Zitao Wang Qiguang Miao* Yue Xi* Peipei Zhao"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.41,
+                0.225,
+                0.56,
+                0.239
+            ],
+            "angle": 0,
+            "content": "Xidian University, China"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.22,
+                0.242,
+                0.741,
+                0.257
+            ],
+            "angle": 0,
+            "content": "{21181214174@stu.，qgmiao@，xiyue@，zhpp2023@}xidian.edu.cn"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.235,
+                0.292,
+                0.314,
+                0.308
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.324,
+                0.474,
+                0.778
+            ],
+            "angle": 0,
+            "content": "The portrait matting task aims to extract an alpha matte with complete semantics and finely detailed contours. In comparison to CNN-based approaches, transformers with self-attention module have a better capacity to capture long-range dependencies and low-frequency semantic information of a portrait. However, recent research shows that the self-attention mechanism struggles with modeling high-frequency contour information and capturing fine contour details, which can lead to bias while predicting the portrait's contours. To deal with this issue, we propose EFormer to enhance the model's attention towards both the low-frequency semantic and high-frequency contour features. For the high-frequency contours, our research demonstrates that cross-attention module between different resolutions can guide our model to allocate attention appropriately to these contour regions. Supported by this, we can successfully extract the high-frequency detail information around the portrait's contours, which were previously ignored by self-attention. Based on the cross-attention module, we further build a semantic and contour detector (SCD) to accurately capture both the low-frequency semantic and high-frequency contour features. And we design a contour-edge extraction branch and semantic extraction branch to extract refined high-frequency contour features and complete low-frequency semantic information, respectively. Finally, we fuse the two kinds of features and leverage the segmentation head to generate a predicted portrait matte. Experiments on VideoMatte240K (JPEG SD Format) and Adobe Image Matting (AIM) datasets demonstrate that EFormer outperforms previous portrait matte methods."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.806,
+                0.21,
+                0.822
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.832,
+                0.47,
+                0.879
+            ],
+            "angle": 0,
+            "content": "Portrait matting aims to extract a precise alpha matte of persons from natural images. It has gathered great momentum in vision applications, including but not limited to photo"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.503,
+                0.291,
+                0.892,
+                0.474
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.519,
+                0.478,
+                0.892,
+                0.657
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.498,
+                0.671,
+                0.894,
+                0.77
+            ],
+            "angle": 0,
+            "content": "Figure 1. (I) Visualization with Grad-CAM of output from each layer within different transformer blocks. (a) The features captured from Self-attention and MLP layers in the original transformer block. (b) In comparison, the features captured from each layer in the proposed transformer block. (II) Comparisons of different transformer blocks. (c) The original transformer block in ViT. (d) The proposed transformer block for portrait matte."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.795,
+                0.747,
+                0.81
+            ],
+            "angle": 0,
+            "content": "editing and background replacement."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.811,
+                0.894,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Existing methods[20, 30, 33, 35] attempt to estimate a portrait matte with Convolutional Neural Networks (CNNs) under the guidance of a pre-determined trimap, which is a three-class map that indicates the foreground, background, and unknown region. However, it is time-consuming and labor-intensive to obtain abundant trimaps by manually an-"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.096,
+                0.887,
+                0.228,
+                0.901
+            ],
+            "angle": 0,
+            "content": "*Corresponding authors."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "3880"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.092,
+                0.473,
+                0.258
+            ],
+            "angle": 0,
+            "content": "notating images. Some methods eliminate the trimap, such as background matting series[21, 28] uses a separate background image instead of the trimap as an auxiliary reference for model prediction. However, in practical cases, it is necessary to provide two aligned images for the model, which include one with only the background and the other one including people. An alternative solution[17, 22, 23] is to rely solely on RGB images as input for predicting portrait matte. Although the previous models have made strides in portrait matte, there is a constraint on the receptive field in CNN-based approaches."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.26,
+                0.473,
+                0.412
+            ],
+            "angle": 0,
+            "content": "In contrast, Vision Transformer (ViT)[7] and Swin Transformer (Swin) [24] for semantic segmentation employ a self-attention mechanism, which conducts interactions among all pixels in an image to obtain a global receptive field. Previous research [25] shows that ViT and its variants can capture the low-frequency components in images effectively, such as global shapes and structures of a scene or object. However, they usually neglect high-frequency ones in images, such as edges and textures. They work similarly to low-pass filters."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.413,
+                0.473,
+                0.625
+            ],
+            "angle": 0,
+            "content": "As shown in Fig. 1(a), we can observe the transformer block with self-attention clearly pays little attention to the portrait's high-frequency contour regions. We believe that this phenomenon is attributed to the excessive propagation of global information because global attention is performed in each transformer block during the calculation of multiple blocks. So, the global information is propagated throughout the model in this iterative calculation, ultimately leading it to only focus on low-frequency semantic features with stronger clustering in portraits. Over time, the model gradually ignores the discrete high-frequency detail features around the contours. Therefore, it is desired to capture low- and high-frequency components simultaneously to improve the segmentation performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.627,
+                0.473,
+                0.824
+            ],
+            "angle": 0,
+            "content": "To achieve this goal, Inception Transformer (IFormer)[29] splits all channels into a convolutional path and self-attention path, respectively, to capture high and low frequencies. However, it is required for a channel ratio of each block at different levels in the IFormer. These channel ratios typically require manual adjustment, resulting in significant uncertainty when aiming for optimal model performance. To enhance the model's stability and adaptive capability, we propose a novel architecture EFormer without manually adjusted parameters. In EFormer, we use the cross-attention layer between different resolution features, as shown in Fig. 1(d), which is replaced by convolution to handle high-frequency detail features."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.826,
+                0.473,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Our research further demonstrates that the cross-attention module between different resolutions can formulate a more reasonable attention allocation mechanism, which can accurately capture contour features, as shown in Fig. 1(b). Particularly, EFormer with a cross-attention layer"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.092,
+                0.892,
+                0.258
+            ],
+            "angle": 0,
+            "content": "not only can not lose the semantic information within the portrait but also can successfully capture the high-frequency details around the portrait's contours, which were previously ignored by self-attention. It is worth noting that high-frequency components only around the boundary between foreground and background are critical to the improvement of segmentation performance. Therefore, we follow a coarse-to-fine manner. Initially, EFormer needs to locate all high-frequency regions in the image. Then, it filters and extracts the high-frequency regions around the foreground contours."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.259,
+                0.893,
+                0.621
+            ],
+            "angle": 0,
+            "content": "Based on the cross-attention module, we design a semantic and contour detector (SCD) in EFormer, which cascades the cross-attention layer and self-attention layer to sequentially locate the contour and semantic features. Firstly, the model adjusts its attention to contour features with the guidance of a cross-attention module. Secondly, the self-attention module captures semantic information within a portrait's contours and reverses filtering and correcting the contour information output by the cross-attention layer. This allows the model to gradually capture and match the high-frequency contour information and the low-frequency semantic information. Then, we use multi-layer perceptron (MLP) to build the contour-edge extraction branch (CEEB) and the semantic extraction branch (SEB), respectively. With this support, we independently purify the contour flow and the semantic flow to obtain finer contour information and more comprehensive portrait semantic information, as shown in Fig. 1(b). Finally, we fuse the contour features and semantic features and send them to the segmentation head to estimate the portrait matte. We perform extensive experiments on VideoMatte240K (JPEG SD Format)[21], Adobe Image Matting (AIM)[35], and BG10K. Comparisons with other state-of-the-art models and ablation studies verify that our model performs better than previous works."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.518,
+                0.622,
+                0.836,
+                0.637
+            ],
+            "angle": 0,
+            "content": "To summarize, our contributions are as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.639,
+                0.892,
+                0.683
+            ],
+            "angle": 0,
+            "content": "- We propose EFormer: an approach that can enhance the transformer's attention towards both semantic and contour features of the foreground."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.684,
+                0.892,
+                0.743
+            ],
+            "angle": 0,
+            "content": "- We further build a semantic and contour detector (SCD) to accurately capture the semantic and contour features and design two separate extraction branches to purify the contour flow and the semantic flow."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.744,
+                0.892,
+                0.805
+            ],
+            "angle": 0,
+            "content": "- We leverage the cross-attention module between different resolution features in transformer block to autonomously capture and extract high-frequency detail features, which are sparsely distributed along the foreground contour."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.498,
+                0.639,
+                0.892,
+                0.805
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.821,
+                0.651,
+                0.836
+            ],
+            "angle": 0,
+            "content": "2. Related Works"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.847,
+                0.652,
+                0.864
+            ],
+            "angle": 0,
+            "content": "2.1. Image Matting"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.871,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Image matting is an essential task in the field of computer vision that aims to accurately estimate the foreground in an"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.514,
+                0.957
+            ],
+            "angle": 0,
+            "content": "3881"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.092,
+                0.47,
+                0.138
+            ],
+            "angle": 0,
+            "content": "image. Mathematically speaking, an image \\(I\\) is a combination of an unknown foreground image \\(F\\) and a background image \\(B\\) with a probability coefficient alpha matte maps \\(\\alpha\\)."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.2,
+                0.148,
+                0.469,
+                0.165
+            ],
+            "angle": 0,
+            "content": "\\[\nI = \\alpha F + (1 - \\alpha) B \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.175,
+                0.469,
+                0.265
+            ],
+            "angle": 0,
+            "content": "Previous image matting solutions have predominantly focused on low-level features, such as color cues[9, 10, 14, 15] or propagation[1, 2, 19], to distinguish the transition areas between foregrounds and backgrounds. However, such traditional matting algorithms commonly struggle to perform properly in complex scenes."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.266,
+                0.47,
+                0.476
+            ],
+            "angle": 0,
+            "content": "With the significant advancements making in deep learning, numerous methods based on convolutional neural networks (CNNs) are proposed, leading to notable successes. Some approaches[36, 37, 39] incorporate auxiliary trimap supervisions to enhance matting performance, while other methods[22, 23] leverage trimap-free solutions to estimate alpha matts from image feature maps using an end-to-end segmentation network. In the background matting series[21, 28], an auxiliary input of the background image is used to compute alpha matte. Furthermore, MODNet[17] proposes an end-to-end manner without any auxiliary input. Although prior works based on CNN kernels have improved the accuracy of image matting, the prediction of models is in the limited receptive field."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.485,
+                0.266,
+                0.5
+            ],
+            "angle": 0,
+            "content": "2.2. Vision Transformer"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.509,
+                0.469,
+                0.703
+            ],
+            "angle": 0,
+            "content": "Different from CNNs, vision transformers with self-attention mechanisms can capture long-term dependencies. Therefore, in the realm of computer vision, vision transformers have gained significant attention in recent times. For pixel-level prediction tasks, such as segmentation, SETR[38], SegFormer[34] and DPT[27] apply transformer as an encoder to attain feature maps of images. Their performances demonstrate the transformer with self-attention is capable of building complete contextual information through global interactions. However, recent studies[25, 29] indicate that transformer \"yet is incompetent in capturing high frequencies that predominantly convey local information\"."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.705,
+                0.469,
+                0.87
+            ],
+            "angle": 0,
+            "content": "Compared to self-attention, cross-attention has more potential to induce customized features. Max-Deeplab[31], MaskFormer[5, 6] and SeMask[13] use query-based methods through cross-attention, which are inspired by DETR[3]. They view segmentation as a set prediction problem. U-Transformer[26] and EUT[8] modify the cross-attention in the original transformer to leverage the information from the encoder, allowing a fine spatial recovery in the decoder. The cross-resolution attention employed by RTFormer[32] enables the gathering of comprehensive contextual information for high-resolution features."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.871,
+                0.469,
+                0.901
+            ],
+            "angle": 0,
+            "content": "In this work, we propose a strategy that using of crossattention between different resolutions guides the model to"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.092,
+                0.892,
+                0.182
+            ],
+            "angle": 0,
+            "content": "autonomously locate and capture high-frequency features, particularly those near the portrait's contours. Contrary to IFormer[29], this method eliminates the need for predefined parameters to enhance the model's adaptive capabilities. At the same time, this approach can significantly improve the model's performance in capturing contour details."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.192,
+                0.761,
+                0.208
+            ],
+            "angle": 0,
+            "content": "2.3. Refinement for Segmentation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.215,
+                0.892,
+                0.29
+            ],
+            "angle": 0,
+            "content": "In the field of portrait segmentation, refining the high-frequency details near the contours is crucial. Many prior studies on segmentation refinement depend on convolutional networks or MLPs that are designed specifically for this purpose."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.291,
+                0.893,
+                0.516
+            ],
+            "angle": 0,
+            "content": "PointRend[18], based on the coarse predictions of Mask R-CNN[12], calculates confidence scores as the point-wise uncertainty measure. And it focuses on the points with the higher uncertainty according to set hyperparameters. Finally, PointRend uses a shared MLP to refine the labels of the selected points. Although PointRend has implemented segmentation refinement to some extent, the model still requires manual adjustment of certain hyperparameters to assist in filtering uncertain points. Transfiner[16] processes detected error-prone and incoherent regions with transformer, which include high-frequency regions strewn along object boundaries. However, during this procedure, excessive focus is placed on the local details and incoherent regions, without effectively associating the local features of the object with the overall semantic information."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.517,
+                0.892,
+                0.669
+            ],
+            "angle": 0,
+            "content": "We believe that locating effective detail distribution areas requires information interaction between local details and global semantics. To achieve this, we design a semantic and contour detector that adopts a cascade structure of cross-attention and self-attention. This approach uses global semantic information garnered from self-attention to autonomously select out local details situated at the borders of the contours. On this base, we further utilize the MLP-based extraction branch to purify contour details, thereby significantly enhancing the effect of refinement."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.682,
+                0.579,
+                0.697
+            ],
+            "angle": 0,
+            "content": "3. Model"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.707,
+                0.892,
+                0.768
+            ],
+            "angle": 0,
+            "content": "In this section, we introduce the process of establishing the complete model architecture of EFormer, which comprises the backbone for encoder, the transformer block for decoder, and the prediction stage, as shown in Fig. 2."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.777,
+                0.71,
+                0.792
+            ],
+            "angle": 0,
+            "content": "3.1. Backbone for Encoder"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.799,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "The CNN backbone takes RGB images \\( I \\in \\mathbb{R}^{B \\times 3 \\times H \\times W} \\) as input, where \\( B \\) is the number of images, to obtain a pyramid of feature maps \\( F = \\left\\{F_{enc\\frac{1}{4}}, F_{enc\\frac{1}{8}}, F_{enc\\frac{1}{16}}\\right\\} \\) with dimension \\( F_{enc\\frac{1}{i}} \\in \\mathbb{R}^{B \\times C_i \\times \\frac{H}{i} \\times \\frac{W}{i}} \\). As we all know, in the pyramid of feature maps, shallow features have high resolution and rich details, while deep features have low reso"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "3882"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.1,
+                0.087,
+                0.875,
+                0.395
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.404,
+                0.893,
+                0.434
+            ],
+            "angle": 0,
+            "content": "Figure 2. The Architecture of the EFormer. It contains a backbone for encoder, a transformer block for decoder, and the prediction stage, where the transformer block includes a semantic and contour detector, a semantic extraction branch, and a contour-edge extraction branch."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.459,
+                0.473,
+                0.761
+            ],
+            "angle": 0,
+            "content": "lation but more concentrated semantic information. Based on this, our research makes the cross-attention mechanism to filter the high-frequency contour features from high-resolution features, using low-frequency semantic information in low-resolution features. However, since the large resolution of feature maps would bring large computation cost, we use the \\( F_{enc\\frac{1}{8}} \\) as \\( F_{HR} \\) (high-resolution features) and use the \\( F_{enc\\frac{1}{16}} \\) as \\( F_{LR} \\) (low-resolution features) in our base model. The impact of other choices of \\( F_{HR} \\) and \\( F_{LR} \\) on model performance is demonstrated in ablation experiments Sec. 4.4. As shown in Fig. 2, the \\( F_{HR} \\) serves as the source of contour flow, providing high-frequency contour detail features for the following processes. Correspondingly, the \\( F_{LR} \\) serves as the source of semantic flow, providing low-frequency semantic information. Then, the \\( F_{HR} \\) and \\( F_{LR} \\) are projected to same dimension \\( C \\) and respectively flattened into high-resolution feature embedding as \\( F_{HR}^{em} \\) and low-resolution feature embedding as \\( F_{LR}^{em} \\), where \\( F_{HR}^{em} \\in R^{N\\times B\\times C} \\), \\( F_{LR}^{em} \\in R^{N\\times B\\times C} \\), and \\( N = \\frac{H}{8}\\times \\frac{W}{8} \\). Both of them are utilized as inputs for the transformer block."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.772,
+                0.357,
+                0.787
+            ],
+            "angle": 0,
+            "content": "3.2. Transformer Block for Decoder"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.795,
+                0.47,
+                0.901
+            ],
+            "angle": 0,
+            "content": "We propose a novel transformer block in the model decoding stage to enhance the ability to capture and extract visual features. It not only does not reduce the focus on low-frequency semantic features but also significantly increases the attention to high-frequency contour features. Its detailed architecture is depicted in Fig. 3. The novel transformer block first uses a Semantic and Contour De"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.459,
+                0.892,
+                0.638
+            ],
+            "angle": 0,
+            "content": "tector (SCD) to locate and capture high-frequency contour features and low-frequency semantic features. To achieve this, we cascade a cross-attention layer (CA) and a self-attention layer (SA) in the SCD. Then, as shown in Fig. 3, a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB) are further built to handle high-frequency contour feature flow and low-frequency semantic feature flow respectively. These branches can separately purify more refined high-frequency contour features and extract more complete low-frequency semantic features. This ensures an outcome that is optimal for both types of features."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.642,
+                0.892,
+                0.747
+            ],
+            "angle": 0,
+            "content": "Semantic and Contour Detector (SCD). There is a fact that low-frequency semantic features far outperform high-frequency contour detail features in terms of quantity and distribution density. So, in the global information propagation process, the low-frequency semantic information easily dominates the feature representations of the original transformer block with self-attention module."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.75,
+                0.893,
+                0.901
+            ],
+            "angle": 0,
+            "content": "To prevent the loss of high-frequency contour detail information, we first use a cross-attention layer with different resolution features rather than a self-attention layer. The cross-attention layer can capture high-frequency detail features with high semantic correlation in high-resolution features, based on the low-frequency semantic information in low-resolution features. Technically, after getting \\( F_{HR}^{em} \\) and \\( F_{LR}^{em} \\), we use Layer Norm (LN) processing them to maintain consistency. Due to the sensitivity of the attention mechanism towards positional information, we use absolute and"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "3883"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.091,
+                0.089,
+                0.46,
+                0.527
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.538,
+                0.47,
+                0.595
+            ],
+            "angle": 0,
+            "content": "Figure 3. Our transformer block. It includes a Semantic and Contour Detector (SCD), a Semantic Extraction Branch (SEB), and a Contour-Edge Extraction Branch (CEEB). The detailed composition of the module is shown in the figure."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.619,
+                0.47,
+                0.65
+            ],
+            "angle": 0,
+            "content": "learnable position encoding (PE) in the SCD. The input \\( K \\), \\( Q \\), \\( V \\) of the cross-attention layer (CA) is calculated as"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.19,
+                0.659,
+                0.469,
+                0.677
+            ],
+            "angle": 0,
+            "content": "\\[\nK = L N \\left(F _ {H R} ^ {e m}\\right) + P E \\tag {2}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.192,
+                0.685,
+                0.469,
+                0.702
+            ],
+            "angle": 0,
+            "content": "\\[\nQ = L N \\left(F _ {L R} ^ {e m}\\right) + P E \\tag {3}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.188,
+                0.707,
+                0.469,
+                0.723
+            ],
+            "angle": 0,
+            "content": "\\[\nV = L N \\left(F _ {L R} ^ {e m} + F _ {H R} ^ {e m}\\right) \\tag {4}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.729,
+                0.47,
+                0.817
+            ],
+            "angle": 0,
+            "content": "By performing cross-attention calculations using features of different resolutions, it is possible to automatically locate and capture the discrete contours along the portrait and allocate more attention to corresponding positions. The contour-edge features can be derived from the output of the CA."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.162,
+                0.819,
+                0.469,
+                0.836
+            ],
+            "angle": 0,
+            "content": "\\[\nF _ {\\text {c o n t o u r - e d g e}} = C A (K, Q, V) \\tag {5}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.841,
+                0.47,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Although CA pays extra attention to contour-edge features, it still captures the overall semantics. By the residual connection between CA and \\(V\\), both the semantic and contour-edge features of the portrait are enhanced, especially the"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.093,
+                0.652,
+                0.107
+            ],
+            "angle": 0,
+            "content": "contour-edge features."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.585,
+                0.117,
+                0.892,
+                0.134
+            ],
+            "angle": 0,
+            "content": "\\[\nF _ {\\text {e n h a n c e}} = C A (K, Q, V) + V \\tag {6}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.144,
+                0.894,
+                0.234
+            ],
+            "angle": 0,
+            "content": "Since semantic information is still present, we use the self-attention layer (SA) after CA. SA can locate the semantic feature within the portrait and contours, thereby emphasizing the semantic attributes at their respective positions. After using LN to normalize \\( F_{\\text{enhance}} \\), the input \\( K', Q', V' \\) of the SA is calculated as"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.58,
+                0.244,
+                0.892,
+                0.262
+            ],
+            "angle": 0,
+            "content": "\\[\nK ^ {\\prime} = Q ^ {\\prime} = L N \\left(F _ {\\text {e n h a n c e}}\\right) + P E \\tag {7}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.624,
+                0.271,
+                0.891,
+                0.288
+            ],
+            "angle": 0,
+            "content": "\\[\nV ^ {\\prime} = L N \\left(F _ {\\text {e n h a n c e}}\\right) \\tag {8}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.545,
+                0.293,
+                0.891,
+                0.31
+            ],
+            "angle": 0,
+            "content": "\\[\nF _ {\\text {c o n t o u r - s e m a n t i c}} = S A \\left(K ^ {\\prime}, Q ^ {\\prime}, V ^ {\\prime}\\right) + V ^ {\\prime} \\tag {9}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.316,
+                0.892,
+                0.374
+            ],
+            "angle": 0,
+            "content": "Through SA filtering the semantic and contour features, the residual connection between \\( V' \\) and the output of SA, we can produce \\( F_{contour - semantic} \\) as the output of the semantic and contour detector."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.376,
+                0.893,
+                0.572
+            ],
+            "angle": 0,
+            "content": "Feature Extraction Branches. After SCD, we use multi-layer perceptron (MLP) to build a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB). To prevent the loss of high-frequency contour detail information, we aggregate \\( F_{contour - semantic} \\) getting from SCD and \\( F_{HR}^{em} \\) in the contour flow. After getting the merged feature normalized by LN, we employ CEEB to purify it, in order to further extract and refine high-frequency contour details individually. Similarly, to obtain more comprehensive and consistent portrait semantic features separately, we merge \\( F_{contour - semantic} \\) and \\( F_{LR}^{em} \\) in the semantic flow. Then, we normalize the aggregated feature by applying LN and use SEB to filter it."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.515,
+                0.582,
+                0.891,
+                0.613
+            ],
+            "angle": 0,
+            "content": "\\[\nF _ {\\text {c o n t o u r}} = M L P \\left(L N \\left(F _ {H R} ^ {\\text {e m}} + F _ {\\text {c o n t o u r - s e m a n t i c}}\\right)\\right) \\tag {10}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.513,
+                0.613,
+                0.891,
+                0.644
+            ],
+            "angle": 0,
+            "content": "\\[\nF _ {\\text {s e m a n t i c}} = M L P \\left(L N \\left(F _ {L R} ^ {\\text {e m}} + F _ {\\text {c o n t o u r - s e m a n t i c}}\\right)\\right) \\tag {11}\n\\]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.653,
+                0.664,
+                0.669
+            ],
+            "angle": 0,
+            "content": "3.3. Prediction Stage"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.675,
+                0.892,
+                0.798
+            ],
+            "angle": 0,
+            "content": "After getting key characteristics \\( F_{semantic} \\in R^{N \\times B \\times C} \\) and \\( F_{contour} \\in R^{N \\times B \\times C} \\) from transformer block, where \\( N = \\frac{H}{8} \\times \\frac{W}{8} \\), we transform them into \\( F_{semantic}' \\in R^{B \\times C \\times \\frac{H}{8} \\times \\frac{W}{8}} \\) and \\( F_{contour}' \\in R^{B \\times C \\times \\frac{H}{8} \\times \\frac{W}{8}} \\) by converting the vector into the matrix. To achieve improved outcomes in portrait segmentation, we propose fusing the two kinds of features and ultimately acquiring the predicted portrait matte from the segmentation head module."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.799,
+                0.892,
+                0.877
+            ],
+            "angle": 0,
+            "content": "Fuse. We attain the semantic feature through the residual connection between \\( F_{LR}^{\\prime} \\) and \\( F_{semantic}^{\\prime} \\) of semantic flow. Similarly, we get the contour feature from contour flow and subsequently merge the \\( F_{semantic}^{\\prime} \\) and \\( F_{contour}^{\\prime} \\) to generate the fused features \\( F_{contour - semantic}^{\\prime} \\)."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.563,
+                0.885,
+                0.892,
+                0.903
+            ],
+            "angle": 0,
+            "content": "\\[\nF _ {\\text {s e m a n t i c}} ^ {\\prime} = \\operatorname {C o n v} \\left(F _ {\\text {s e m a n t i c}} ^ {\\prime} + F _ {L R} ^ {\\prime}\\right) \\tag {12}\n\\]"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "3884"
+        }
+    ],
+    [
+        {
+            "type": "equation",
+            "bbox": [
+                0.148,
+                0.09,
+                0.468,
+                0.109
+            ],
+            "angle": 0,
+            "content": "\\[\nF _ {\\text {c o n t o u r}} ^ {\\prime} = \\operatorname {C o n v} \\left(F _ {\\text {c o n t o u r}} ^ {\\prime} + F _ {H R} ^ {\\prime}\\right) \\tag {13}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.086,
+                0.114,
+                0.468,
+                0.133
+            ],
+            "angle": 0,
+            "content": "\\[\nF _ {\\text {c o n t o u r - s e m a n t i c}} ^ {\\prime} = \\operatorname {F u s e} \\left(F _ {\\text {s e m a n t i c}} ^ {\\prime} + F _ {\\text {c o n t o u r}} ^ {\\prime}\\right) \\tag {14}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.139,
+                0.468,
+                0.214
+            ],
+            "angle": 0,
+            "content": "Segmentation Head. Finally, We gather \\( F_{contour - semantic}^{\\prime} \\) and \\( F_{enc\\frac{1}{4}} \\) with higher resolution to predict portrait matte in 1/4 scale of the input image and upscale it to the original scale through bi-linear interpolation as the output of our model."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.102,
+                0.223,
+                0.468,
+                0.249
+            ],
+            "angle": 0,
+            "content": "\\[\nm a t t e = H e a d \\left(F _ {c o n t o u r - s e m a n t i c} ^ {\\prime} + F _ {e n c \\frac {1}{4}}\\right) \\tag {15}\n\\]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.26,
+                0.21,
+                0.277
+            ],
+            "angle": 0,
+            "content": "4. Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.285,
+                0.294,
+                0.3
+            ],
+            "angle": 0,
+            "content": "4.1. Dataset and Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.308,
+                0.47,
+                0.579
+            ],
+            "angle": 0,
+            "content": "Dataset. We utilize composite training data from two sources: the foreground image dataset VideoMatte240K (JPEG SD Format)[21] and Adobe Image Matting (AIM)[35]. For the background images, we select BG10K following the approach of BGMv2[21], which is a collection of photographs depicting various life scenes without any human portraits. We select foregrounds from VideoMatte240K-JPEG-SD and AIM, and backgrounds from BG10K to composite image datasets. We split VideoMatte240K-JPEG-SD into 234,982/3,007/2720 image sets with image resolutions of \\(224 \\times 224\\) and \\(512 \\times 288\\) for training, validating, and testing our model. Similarly, BG10K is split into 9000/1000 image sets, and AIM is split into 214/10 image sets for training and testing, where AIM only includes the portrait images with resolutions at \\(512 \\times 512\\). Finally, we use the aforementioned testing datasets to compare and evaluate the performance of EFormer against other models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.581,
+                0.468,
+                0.701
+            ],
+            "angle": 0,
+            "content": "Evaluation. We mainly consider portrait matting accuracy for evaluation. For portrait matting accuracy, we use Mean Absolute Difference (MAD), Mean Squared Error (MSE), Gradient (Grad), and Connectivity (Conn) as evaluation metrics. We also scale MAD, MSE, Grad, and Conn by \\(10^{3}\\), \\(10^{3}\\), \\(10^{-3}\\), and \\(10^{-3}\\) respectively, for convenience of reference. For all these metrics, the lower number represents better performance."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.711,
+                0.295,
+                0.727
+            ],
+            "angle": 0,
+            "content": "4.2. Implementation Details"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.734,
+                0.468,
+                0.84
+            ],
+            "angle": 0,
+            "content": "Training Settings. For training the network, we use a single RTX 3090 GPU with batch size at 24. The optimizer is AdamW and the initial learning rate is set to \\(10^{-4}\\). All models in ablation studies are trained for 25 epochs, with the learning rate decaying by a factor of 0.8 every 5 epochs. Additionally, to augment the data, each input image is subjected to random horizontal flipping."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.84,
+                0.468,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Backbone. The backbone we use for the network is the pretrained ResNet50[11], since it is the most frequently used backbone in prior works. We use the implementation and weights from torchvision."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.514,
+                0.089,
+                0.88,
+                0.196
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>DeepLabv3[4]</td><td>14.4700</td><td>9.6700</td><td>8.5500</td><td>1.6900</td></tr><tr><td>MODnet[17]</td><td>10.3900</td><td>5.6500</td><td>2.0200</td><td>1.0400</td></tr><tr><td>BGMv2[21]</td><td>4.1858</td><td>1.7934</td><td>1.3011</td><td>0.5326</td></tr><tr><td>RVM[22]</td><td>5.9900</td><td>1.1700</td><td>1.1000</td><td>0.3400</td></tr><tr><td>EFormer</td><td>2.3097</td><td>0.6637</td><td>0.4580</td><td>0.2740</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.499,
+                0.206,
+                0.892,
+                0.248
+            ],
+            "angle": 0,
+            "content": "Table 1. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at \\(512 \\times 288\\)."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.521,
+                0.261,
+                0.871,
+                0.368
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>DeepLabv3[4]</td><td>29.64</td><td>23.78</td><td>20.17</td><td>7.71</td></tr><tr><td>MODnet[17]</td><td>21.66</td><td>14.27</td><td>5.37</td><td>5.23</td></tr><tr><td>BGMv2[21]</td><td>44.61</td><td>39.08</td><td>5.54</td><td>11.60</td></tr><tr><td>RVM[22]</td><td>14.84</td><td>8.93</td><td>4.35</td><td>3.83</td></tr><tr><td>EFormer</td><td>7.47</td><td>2.13</td><td>2.83</td><td>1.90</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.499,
+                0.379,
+                0.892,
+                0.42
+            ],
+            "angle": 0,
+            "content": "Table 2. Comparisons on the test set of AIM. Bold indicates the best performance among these models under the inputs with the same resolution at \\(512 \\times 512\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.444,
+                0.892,
+                0.533
+            ],
+            "angle": 0,
+            "content": "Transformer Block for Decoder. The transformer block is based on the transformer decoder of DETR. We continuously use four transformer blocks in the decoder of the network. We set the channel \\( C = 256 \\) and the number of attention heads \\( M = 8 \\) for each multi-head attention module."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.535,
+                0.892,
+                0.61
+            ],
+            "angle": 0,
+            "content": "Prediction Stage. After obtaining the semantic and contour feature maps of the portrait from the stacked transformer blocks, we fuse both of them. Then, the alpha mattes predicted by the segmentation head are upsampled to the original size of the input image with bi-linear interpolation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.611,
+                0.89,
+                0.641
+            ],
+            "angle": 0,
+            "content": "Loss function. To compute losses during the training stage, we use a Binary CrossEntropy Loss function."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.517,
+                0.649,
+                0.89,
+                0.666
+            ],
+            "angle": 0,
+            "content": "\\[\nL _ {b c e} = - (g \\times \\log (m) + (1 - g) \\times \\log (1 - m)) \\tag {16}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.673,
+                0.892,
+                0.703
+            ],
+            "angle": 0,
+            "content": "Where, \\( m \\) is the prediction of the network, \\( g \\) is the ground truth."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.712,
+                0.852,
+                0.727
+            ],
+            "angle": 0,
+            "content": "4.3. Comparisons to State-of-the-art Methods"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.735,
+                0.892,
+                0.9
+            ],
+            "angle": 0,
+            "content": "To demonstrate the progressiveness of EFormer, we compare it against the latest state-of-the-art trimap-free portrait matting solutions. These are including DeepLabV3[4] with ResNet101[11] backbone, BGMv2[21] with ResNet50[11] backbone, MODNet[17], and RVM[22]. Our assessment is based on the testing datasets of VideoMatte240K-JPEG-SD and AIM, respectively. As the training files for MODNet are not available, we use its official weights. Furthermore, since BGMv2, DeepLabV3, and RVM have already been trained on all datasets, our method can be compared with them in a fair manner."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "3885"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.084,
+                0.089,
+                0.895,
+                0.429
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.436,
+                0.895,
+                0.478
+            ],
+            "angle": 0,
+            "content": "Figure 4. Visualization of portrait matte predictions from MODnet, BGMv2, and EFormer under challenging image from the test set of VideoMatte240K (JPEG SD Format). Our model shows a better ability to distinguish ambiguous foreground contours, as indicated by the red box. Please zoom in to view the detailed information."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.086,
+                0.502,
+                0.462,
+                0.61
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>DeepLabv3[4]</td><td>18.1586</td><td>14.7623</td><td>6.0015</td><td>2.8824</td></tr><tr><td>MODnet[17]</td><td>13.1400</td><td>8.9656</td><td>2.6028</td><td>2.1543</td></tr><tr><td>BGMv2[21]</td><td>7.0471</td><td>2.3543</td><td>1.5621</td><td>1.0651</td></tr><tr><td>RVM[22]</td><td>6.1357</td><td>2.2647</td><td>0.6523</td><td>0.2866</td></tr><tr><td>EFormer</td><td>4.1357</td><td>1.4280</td><td>0.3760</td><td>0.1753</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.076,
+                0.62,
+                0.47,
+                0.662
+            ],
+            "angle": 0,
+            "content": "Table 3. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at \\(224 \\times 224\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.688,
+                0.47,
+                0.764
+            ],
+            "angle": 0,
+            "content": "As shown in the Tab. 1, Tab. 2 and Tab. 3, our model exhibits superior performance in portrait matte tasks. It has a stronger capability to distinguish the contour of a portrait. And it can finely segment edge details (e.g. fingers), as exemplified in Fig. 4."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.772,
+                0.241,
+                0.787
+            ],
+            "angle": 0,
+            "content": "4.4. Ablation Studies"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.795,
+                0.471,
+                0.9
+            ],
+            "angle": 0,
+            "content": "Training Epochs. In the ablation study with an image resolution of \\(224 \\times 224\\), we conduct 25 training epochs. Every 5 epochs of training, we evaluate the training effect of the model. At the same time, the learning rate is reduced by 0.8 times. The experimental results indicate that our model achieves optimal performance in the 20 epoch, as shown in the Tab. 4."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.534,
+                0.502,
+                0.86,
+                0.61
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>5-epoch</td><td>4.3953</td><td>1.4824</td><td>0.4089</td><td>0.1865</td></tr><tr><td>10-epoch</td><td>4.3108</td><td>1.4766</td><td>0.3908</td><td>0.1834</td></tr><tr><td>15-epoch</td><td>4.2867</td><td>1.5238</td><td>0.4087</td><td>0.1833</td></tr><tr><td>20-epoch</td><td>4.1357</td><td>1.4280</td><td>0.3760</td><td>0.1753</td></tr><tr><td>25-epoch</td><td>4.1811</td><td>1.4797</td><td>0.3858</td><td>0.1779</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.498,
+                0.62,
+                0.894,
+                0.676
+            ],
+            "angle": 0,
+            "content": "Table 4. Ablation Study on the Training Epochs with the test set of VideoMatte240K-JPEG-SD (resolution at \\(224 \\times 224\\)). **Bold** indicates the best performance among these models. We use the result of 20-epoch training as a strong baseline model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.704,
+                0.892,
+                0.794
+            ],
+            "angle": 0,
+            "content": "The Functional Role of Each Module. To demonstrate the effectiveness of the model in capturing and extracting contour details and portrait semantics, we analyze the outputs of CA, SA, CEEB, and SEB, and utilize Class Activation Map (CAM) to visualize the attention allocation of each layer's features, as shown in Fig. 5."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.795,
+                0.892,
+                0.885
+            ],
+            "angle": 0,
+            "content": "From the first line of Fig. 5, it can be evident that after the cross-attention layer, the model clearly strengthens its attention to the contours of the portrait. Following the contour prior, the self-attention layer captures the portrait's semantics thoroughly. CEEB and SEB models provide sharper contour and more comprehensive semantic information."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.519,
+                0.886,
+                0.892,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Comparing the second and first lines of Fig. 5, it is ob"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "3886"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.081,
+                0.09,
+                0.468,
+                0.276
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.287,
+                0.472,
+                0.358
+            ],
+            "angle": 0,
+            "content": "Figure 5. We use Grad-CAM to visualize the attention distribution of various features in CA, SA, CEEB, and SEB. The model in the first row contains four complete modules. For comparison, the SA layer and CA layer were removed from the models in the second and third rows, respectively."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.091,
+                0.374,
+                0.456,
+                0.452
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>CA</td><td>SA</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>✓</td><td></td><td>4.2762</td><td>1.5344</td><td>0.4092</td><td>0.1823</td></tr><tr><td></td><td>✓</td><td>4.1710</td><td>1.4354</td><td>0.3773</td><td>0.1769</td></tr><tr><td>✓</td><td>✓</td><td>4.1357</td><td>1.4280</td><td>0.3760</td><td>0.1753</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.076,
+                0.462,
+                0.47,
+                0.517
+            ],
+            "angle": 0,
+            "content": "Table 5. Ablation Study on the Functional Role of CA and SA with the test set of VideoMatte240K-JPEG-SD (resolution at \\(224 \\times 224\\)). **Bold** indicates the best performance among these models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.548,
+                0.469,
+                0.625
+            ],
+            "angle": 0,
+            "content": "served that removing the self-attention layer considerably weakens the semantic information extracted by SEB. In the comparison between the third and first lines of Fig. 5, it can be indicated that removing the cross-attention layer significantly weakens the contour information extracted by CEEB."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.626,
+                0.469,
+                0.685
+            ],
+            "angle": 0,
+            "content": "After qualitative analysis, the prediction results of all the above models are evaluated. It is concluded that the model with CA, SA, CEEB, and SEB is both effective and rational, as demonstrated in the Tab. 5."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.688,
+                0.47,
+                0.884
+            ],
+            "angle": 0,
+            "content": "As is well known, when it comes to extracting features for portrait segmentation tasks, semantic information plays a significant role, in contrast, contour details account for a relatively small proportion. Owing to this, as shown in the Tab. 5 and Fig. 5, the model without SA leads to a significant increase in all metrics due to the lack of amount of semantic information, while the model without CA experiences a slight increase in all indicators due to the loss of contour details. In order to fully harness the potential of the vision transformer, we integrate both CA and SA to ensure semantic completeness and enhance attention towards contour details, thereby achieving more comprehensive and fine-grained segmentation results."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.886,
+                0.469,
+                0.902
+            ],
+            "angle": 0,
+            "content": "The Selection of \\(F_{HR}\\) and \\(F_{LR}\\). In our base model,"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.501,
+                0.089,
+                0.892,
+                0.171
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>FHR</td><td>FLR</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>Fenc1/8</td><td>Fenc1/16</td><td>4.1357</td><td>1.4280</td><td>0.3760</td><td>0.1753</td></tr><tr><td>Fenc1/4</td><td>Fenc1/8</td><td>3.9970</td><td>1.3343</td><td>0.3328</td><td>0.1679</td></tr><tr><td>Fenc1/4</td><td>Fenc1/16</td><td>3.9730</td><td>1.2949</td><td>0.3215</td><td>0.1665</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.499,
+                0.181,
+                0.892,
+                0.224
+            ],
+            "angle": 0,
+            "content": "Table 6. Ablation Study on the Selection of \\(F_{HR}\\) and \\(F_{LR}\\) with the test set of VideoMatte240K-JPEG-SD (resolution at \\(224 \\times 224\\)). Bold indicates the best performance among these models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.26,
+                0.892,
+                0.457
+            ],
+            "angle": 0,
+            "content": "considering the large resolution of feature maps would bring large computation cost, we use the \\( F_{enc}^{\\frac{1}{8}} \\) as \\( F_{HR} \\) (high-resolution features) and use the \\( F_{enc}^{\\frac{1}{16}} \\) as \\( F_{LR} \\) (low-resolution features). Then the \\( F_{HR} \\) serves as the source of contour flow, providing high-frequency contour detail features. Correspondingly, the \\( F_{LR} \\) serves as the source of semantic flow, providing low-frequency semantic information. As shown in the Tab. 6, after using \\( F_{enc}^{\\frac{1}{4}} \\) with higher resolution as \\( F_{HR} \\), our model can capture more high-frequency contour detail features and its performance is better. And after using \\( F_{enc}^{\\frac{1}{16}} \\) with lower resolution as \\( F_{LR} \\), our model can use more concentrated low-frequency semantic information, which makes all indicators optimal."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.485,
+                0.619,
+                0.501
+            ],
+            "angle": 0,
+            "content": "5. Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.516,
+                0.892,
+                0.773
+            ],
+            "angle": 0,
+            "content": "In this paper, we propose a strategy that use of cross-attention module between different resolutions guides the EFormer to autonomously locate and capture high-frequency contour features. Without manual parameter tuning, the proposed method can adaptability capture high-frequency details neglected by previous transformer-based methods. Based on the cross-attention module, we further construct a semantic and contour detector (SCD), which can enhance the transformer's attention to both the high-frequency contour and low-frequency semantic features. Additionally, we build a contour-edge extraction branch (CEEB) and a semantic extraction branch (SEB) to extract finer high-frequency contour details and more comprehensive portrait's low-frequency semantic information. Extensive experiments show that our method outperforms previous portrait matting solutions and exhibits robust performance during testing and inference."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.801,
+                0.661,
+                0.818
+            ],
+            "angle": 0,
+            "content": "Acknowledgement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.83,
+                0.892,
+                0.888
+            ],
+            "angle": 0,
+            "content": "The work was jointly supported by the National Natural Science Foundations of China under grant No.62272364, the Teaching Reform Project of Shaanxi Higher Continuing Education under Grant No.21XJZ004."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "3887"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.08,
+                0.09,
+                0.174,
+                0.106
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.115,
+                0.47,
+                0.171
+            ],
+            "angle": 0,
+            "content": "[1] Yagiz Aksoy, Tunc Ozan Aydin, and Marc Pollefeys. Designing effective inter-pixel information flow for natural image matting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 29-37, 2017. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.173,
+                0.472,
+                0.214
+            ],
+            "angle": 0,
+            "content": "[2] Yagiz Aksoy, Tae-Hyun Oh, Sylvain Paris, Marc Pollefeys, and Wojciech Matusik. Semantic soft segmentation. ACM Transactions on Graphics (TOG), 37(4):1-13, 2018. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.216,
+                0.47,
+                0.272
+            ],
+            "angle": 0,
+            "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.274,
+                0.47,
+                0.327
+            ],
+            "angle": 0,
+            "content": "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.33,
+                0.47,
+                0.385
+            ],
+            "angle": 0,
+            "content": "[5] 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. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.388,
+                0.47,
+                0.457
+            ],
+            "angle": 0,
+            "content": "[6] 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. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.459,
+                0.47,
+                0.541
+            ],
+            "angle": 0,
+            "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.544,
+                0.47,
+                0.598
+            ],
+            "angle": 0,
+            "content": "[8] Dazhao Du, Bing Su, Yu Li, Zhongang Qi, Lingyu Si, and Ying Shan. Efficient u-transformer with boundary-aware loss for action segmentation. arXiv preprint arXiv:2205.13425, 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.601,
+                0.47,
+                0.67
+            ],
+            "angle": 0,
+            "content": "[9] Xiaoxue Feng, Xiaohui Liang, and Zili Zhang. A cluster sampling method for image matting via sparse coding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 204-219. Springer, 2016. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.672,
+                0.469,
+                0.714
+            ],
+            "angle": 0,
+            "content": "[10] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaou Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056. IEEE, 2011. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.716,
+                0.469,
+                0.77
+            ],
+            "angle": 0,
+            "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.772,
+                0.469,
+                0.814
+            ],
+            "angle": 0,
+            "content": "[12] 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. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.816,
+                0.469,
+                0.871
+            ],
+            "angle": 0,
+            "content": "[13] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.873,
+                0.469,
+                0.902
+            ],
+            "angle": 0,
+            "content": "[14] Jubin Johnson, Ehsan Shahrian Varnousfaderani, Hisham Cholakkal, and Deepu Rajan. Sparse coding for alpha mat"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.115,
+                0.472,
+                0.902
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.533,
+                0.093,
+                0.892,
+                0.12
+            ],
+            "angle": 0,
+            "content": "ting. IEEE Transactions on Image Processing, 25(7):3032-3043, 2016. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.122,
+                0.892,
+                0.177
+            ],
+            "angle": 0,
+            "content": "[15] Levent Karacan, Aykut Erdem, and Erkut Erdem. Image matting with kl-divergence based sparse sampling. In Proceedings of the IEEE international conference on computer vision, pages 424-432, 2015. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.179,
+                0.892,
+                0.247
+            ],
+            "angle": 0,
+            "content": "[16] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4412–4421, 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.249,
+                0.892,
+                0.318
+            ],
+            "angle": 0,
+            "content": "[17] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1140-1147, 2022. 2, 3, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.32,
+                0.892,
+                0.375
+            ],
+            "angle": 0,
+            "content": "[18] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.377,
+                0.892,
+                0.418
+            ],
+            "angle": 0,
+            "content": "[19] Anat Levin, Alex Rav-Acha, and Dani Lischinski. Spectral matting. IEEE transactions on pattern analysis and machine intelligence, 30(10):1699-1712, 2008. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.42,
+                0.892,
+                0.462
+            ],
+            "angle": 0,
+            "content": "[20] Yaoyi Li and Hongtao Lu. Natural image matting via guided contextual attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11450-11457, 2020. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.463,
+                0.892,
+                0.545
+            ],
+            "angle": 0,
+            "content": "[21] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 2, 3, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.547,
+                0.892,
+                0.616
+            ],
+            "angle": 0,
+            "content": "[22] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 2, 3, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.618,
+                0.892,
+                0.673
+            ],
+            "angle": 0,
+            "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.675,
+                0.892,
+                0.745
+            ],
+            "angle": 0,
+            "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.746,
+                0.892,
+                0.773
+            ],
+            "angle": 0,
+            "content": "[25] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.775,
+                0.892,
+                0.871
+            ],
+            "angle": 0,
+            "content": "[26] Olivier Petit, Nicolas Thome, Clement Rambour, Loic Themyr, Toby Collins, and Luc Soler. U-net transformer: Self and cross attention for medical image segmentation. In Machine Learning in Medical Imaging: 12th International Workshop, MLMI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, September 27, 2021, Proceedings 12, pages 267-276. Springer, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.873,
+                0.892,
+                0.902
+            ],
+            "angle": 0,
+            "content": "[27] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.504,
+                0.093,
+                0.892,
+                0.902
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "3888"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.11,
+                0.093,
+                0.468,
+                0.12
+            ],
+            "angle": 0,
+            "content": "the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021.3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.122,
+                0.469,
+                0.189
+            ],
+            "angle": 0,
+            "content": "[28] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.191,
+                0.469,
+                0.244
+            ],
+            "angle": 0,
+            "content": "[29] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xinchao Wang, and Shuicheng Yan. Inception transformer. Advances in Neural Information Processing Systems, 35:23495-23509, 2022. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.247,
+                0.469,
+                0.314
+            ],
+            "angle": 0,
+            "content": "[30] Yanan Sun, Guanzhi Wang, Qiao Gu, Chi-Keung Tang, and Yu-Wing Tai. Deep video matting via spatio-temporal alignment and aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6975–6984, 2021. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.317,
+                0.469,
+                0.385
+            ],
+            "angle": 0,
+            "content": "[31] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5463-5474, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.386,
+                0.469,
+                0.453
+            ],
+            "angle": 0,
+            "content": "[32] Jian Wang, Chenhui Gou, Qiman Wu, Haocheng Feng, Junyu Han, Errui Ding, and Jingdong Wang. Rrformer: Efficient design for real-time semantic segmentation with transformer. Advances in Neural Information Processing Systems, 35:7423-7436, 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.456,
+                0.469,
+                0.496
+            ],
+            "angle": 0,
+            "content": "[33] Yu Wang, Yi Niu, Peiyong Duan, Jianwei Lin, and Yuanjie Zheng. Deep propagation based image matting. In IJCAI, pages 999-1006, 2018. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.498,
+                0.469,
+                0.564
+            ],
+            "angle": 0,
+            "content": "[34] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.568,
+                0.469,
+                0.621
+            ],
+            "angle": 0,
+            "content": "[35] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 1, 2, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.624,
+                0.469,
+                0.678
+            ],
+            "angle": 0,
+            "content": "[36] Haichao Yu, Ning Xu, Zilong Huang, Yuqian Zhou, and Humphrey Shi. High-resolution deep image matting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3217-3224, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.679,
+                0.469,
+                0.747
+            ],
+            "angle": 0,
+            "content": "[37] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5128-5137, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.749,
+                0.469,
+                0.843
+            ],
+            "angle": 0,
+            "content": "[38] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.081,
+                0.846,
+                0.469,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[39] Bingke Zhu, Yingying Chen, Jinqiao Wang, Si Liu, Bo Zhang, and Ming Tang. Fast deep matting for portrait animation on mobile phone. In Proceedings of the 25th ACM international conference on Multimedia, pages 297-305, 2017. 3"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.093,
+                0.469,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.946,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "3889"
+        }
+    ]
+]

2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/652e59cf-51c6-401e-a44e-7a8d54a66599_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:470ef394bca49ec8b7d6948b4f4feedb5f7ea565943a75289eaddda24fbf4a91
+size 7257131

2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/full.md

@@ -0,0 +1,318 @@
+# EFormer: Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting
+
+Zitao Wang Qiguang Miao* Yue Xi* Peipei Zhao
+
+Xidian University, China
+
+{21181214174@stu.，qgmiao@，xiyue@，zhpp2023@}xidian.edu.cn
+
+# Abstract
+
+The portrait matting task aims to extract an alpha matte with complete semantics and finely detailed contours. In comparison to CNN-based approaches, transformers with self-attention module have a better capacity to capture long-range dependencies and low-frequency semantic information of a portrait. However, recent research shows that the self-attention mechanism struggles with modeling high-frequency contour information and capturing fine contour details, which can lead to bias while predicting the portrait's contours. To deal with this issue, we propose EFormer to enhance the model's attention towards both the low-frequency semantic and high-frequency contour features. For the high-frequency contours, our research demonstrates that cross-attention module between different resolutions can guide our model to allocate attention appropriately to these contour regions. Supported by this, we can successfully extract the high-frequency detail information around the portrait's contours, which were previously ignored by self-attention. Based on the cross-attention module, we further build a semantic and contour detector (SCD) to accurately capture both the low-frequency semantic and high-frequency contour features. And we design a contour-edge extraction branch and semantic extraction branch to extract refined high-frequency contour features and complete low-frequency semantic information, respectively. Finally, we fuse the two kinds of features and leverage the segmentation head to generate a predicted portrait matte. Experiments on VideoMatte240K (JPEG SD Format) and Adobe Image Matting (AIM) datasets demonstrate that EFormer outperforms previous portrait matte methods.
+
+# 1. Introduction
+
+Portrait matting aims to extract a precise alpha matte of persons from natural images. It has gathered great momentum in vision applications, including but not limited to photo
+
+![](images/c9937fb43e03a7633aee3ae5e531b87101ba94ad97cd2f40462a3b2430c6fa14.jpg)
+
+![](images/3b06508ac7a09bba239e870787572c7c30df658ab82ad01b9fb750bf1c4e71f8.jpg)  
+Figure 1. (I) Visualization with Grad-CAM of output from each layer within different transformer blocks. (a) The features captured from Self-attention and MLP layers in the original transformer block. (b) In comparison, the features captured from each layer in the proposed transformer block. (II) Comparisons of different transformer blocks. (c) The original transformer block in ViT. (d) The proposed transformer block for portrait matte.
+
+editing and background replacement.
+
+Existing methods[20, 30, 33, 35] attempt to estimate a portrait matte with Convolutional Neural Networks (CNNs) under the guidance of a pre-determined trimap, which is a three-class map that indicates the foreground, background, and unknown region. However, it is time-consuming and labor-intensive to obtain abundant trimaps by manually an-
+
+notating images. Some methods eliminate the trimap, such as background matting series[21, 28] uses a separate background image instead of the trimap as an auxiliary reference for model prediction. However, in practical cases, it is necessary to provide two aligned images for the model, which include one with only the background and the other one including people. An alternative solution[17, 22, 23] is to rely solely on RGB images as input for predicting portrait matte. Although the previous models have made strides in portrait matte, there is a constraint on the receptive field in CNN-based approaches.
+
+In contrast, Vision Transformer (ViT)[7] and Swin Transformer (Swin) [24] for semantic segmentation employ a self-attention mechanism, which conducts interactions among all pixels in an image to obtain a global receptive field. Previous research [25] shows that ViT and its variants can capture the low-frequency components in images effectively, such as global shapes and structures of a scene or object. However, they usually neglect high-frequency ones in images, such as edges and textures. They work similarly to low-pass filters.
+
+As shown in Fig. 1(a), we can observe the transformer block with self-attention clearly pays little attention to the portrait's high-frequency contour regions. We believe that this phenomenon is attributed to the excessive propagation of global information because global attention is performed in each transformer block during the calculation of multiple blocks. So, the global information is propagated throughout the model in this iterative calculation, ultimately leading it to only focus on low-frequency semantic features with stronger clustering in portraits. Over time, the model gradually ignores the discrete high-frequency detail features around the contours. Therefore, it is desired to capture low- and high-frequency components simultaneously to improve the segmentation performance.
+
+To achieve this goal, Inception Transformer (IFormer)[29] splits all channels into a convolutional path and self-attention path, respectively, to capture high and low frequencies. However, it is required for a channel ratio of each block at different levels in the IFormer. These channel ratios typically require manual adjustment, resulting in significant uncertainty when aiming for optimal model performance. To enhance the model's stability and adaptive capability, we propose a novel architecture EFormer without manually adjusted parameters. In EFormer, we use the cross-attention layer between different resolution features, as shown in Fig. 1(d), which is replaced by convolution to handle high-frequency detail features.
+
+Our research further demonstrates that the cross-attention module between different resolutions can formulate a more reasonable attention allocation mechanism, which can accurately capture contour features, as shown in Fig. 1(b). Particularly, EFormer with a cross-attention layer
+
+not only can not lose the semantic information within the portrait but also can successfully capture the high-frequency details around the portrait's contours, which were previously ignored by self-attention. It is worth noting that high-frequency components only around the boundary between foreground and background are critical to the improvement of segmentation performance. Therefore, we follow a coarse-to-fine manner. Initially, EFormer needs to locate all high-frequency regions in the image. Then, it filters and extracts the high-frequency regions around the foreground contours.
+
+Based on the cross-attention module, we design a semantic and contour detector (SCD) in EFormer, which cascades the cross-attention layer and self-attention layer to sequentially locate the contour and semantic features. Firstly, the model adjusts its attention to contour features with the guidance of a cross-attention module. Secondly, the self-attention module captures semantic information within a portrait's contours and reverses filtering and correcting the contour information output by the cross-attention layer. This allows the model to gradually capture and match the high-frequency contour information and the low-frequency semantic information. Then, we use multi-layer perceptron (MLP) to build the contour-edge extraction branch (CEEB) and the semantic extraction branch (SEB), respectively. With this support, we independently purify the contour flow and the semantic flow to obtain finer contour information and more comprehensive portrait semantic information, as shown in Fig. 1(b). Finally, we fuse the contour features and semantic features and send them to the segmentation head to estimate the portrait matte. We perform extensive experiments on VideoMatte240K (JPEG SD Format)[21], Adobe Image Matting (AIM)[35], and BG10K. Comparisons with other state-of-the-art models and ablation studies verify that our model performs better than previous works.
+
+To summarize, our contributions are as follows:
+
+- We propose EFormer: an approach that can enhance the transformer's attention towards both semantic and contour features of the foreground.  
+- We further build a semantic and contour detector (SCD) to accurately capture the semantic and contour features and design two separate extraction branches to purify the contour flow and the semantic flow.  
+- We leverage the cross-attention module between different resolution features in transformer block to autonomously capture and extract high-frequency detail features, which are sparsely distributed along the foreground contour.
+
+# 2. Related Works
+
+# 2.1. Image Matting
+
+Image matting is an essential task in the field of computer vision that aims to accurately estimate the foreground in an
+
+image. Mathematically speaking, an image  $I$  is a combination of an unknown foreground image  $F$  and a background image  $B$  with a probability coefficient alpha matte maps  $\alpha$ .
+
+$$
+I = \alpha F + (1 - \alpha) B \tag {1}
+$$
+
+Previous image matting solutions have predominantly focused on low-level features, such as color cues[9, 10, 14, 15] or propagation[1, 2, 19], to distinguish the transition areas between foregrounds and backgrounds. However, such traditional matting algorithms commonly struggle to perform properly in complex scenes.
+
+With the significant advancements making in deep learning, numerous methods based on convolutional neural networks (CNNs) are proposed, leading to notable successes. Some approaches[36, 37, 39] incorporate auxiliary trimap supervisions to enhance matting performance, while other methods[22, 23] leverage trimap-free solutions to estimate alpha matts from image feature maps using an end-to-end segmentation network. In the background matting series[21, 28], an auxiliary input of the background image is used to compute alpha matte. Furthermore, MODNet[17] proposes an end-to-end manner without any auxiliary input. Although prior works based on CNN kernels have improved the accuracy of image matting, the prediction of models is in the limited receptive field.
+
+# 2.2. Vision Transformer
+
+Different from CNNs, vision transformers with self-attention mechanisms can capture long-term dependencies. Therefore, in the realm of computer vision, vision transformers have gained significant attention in recent times. For pixel-level prediction tasks, such as segmentation, SETR[38], SegFormer[34] and DPT[27] apply transformer as an encoder to attain feature maps of images. Their performances demonstrate the transformer with self-attention is capable of building complete contextual information through global interactions. However, recent studies[25, 29] indicate that transformer "yet is incompetent in capturing high frequencies that predominantly convey local information".
+
+Compared to self-attention, cross-attention has more potential to induce customized features. Max-Deeplab[31], MaskFormer[5, 6] and SeMask[13] use query-based methods through cross-attention, which are inspired by DETR[3]. They view segmentation as a set prediction problem. U-Transformer[26] and EUT[8] modify the cross-attention in the original transformer to leverage the information from the encoder, allowing a fine spatial recovery in the decoder. The cross-resolution attention employed by RTFormer[32] enables the gathering of comprehensive contextual information for high-resolution features.
+
+In this work, we propose a strategy that using of crossattention between different resolutions guides the model to
+
+autonomously locate and capture high-frequency features, particularly those near the portrait's contours. Contrary to IFormer[29], this method eliminates the need for predefined parameters to enhance the model's adaptive capabilities. At the same time, this approach can significantly improve the model's performance in capturing contour details.
+
+# 2.3. Refinement for Segmentation
+
+In the field of portrait segmentation, refining the high-frequency details near the contours is crucial. Many prior studies on segmentation refinement depend on convolutional networks or MLPs that are designed specifically for this purpose.
+
+PointRend[18], based on the coarse predictions of Mask R-CNN[12], calculates confidence scores as the point-wise uncertainty measure. And it focuses on the points with the higher uncertainty according to set hyperparameters. Finally, PointRend uses a shared MLP to refine the labels of the selected points. Although PointRend has implemented segmentation refinement to some extent, the model still requires manual adjustment of certain hyperparameters to assist in filtering uncertain points. Transfiner[16] processes detected error-prone and incoherent regions with transformer, which include high-frequency regions strewn along object boundaries. However, during this procedure, excessive focus is placed on the local details and incoherent regions, without effectively associating the local features of the object with the overall semantic information.
+
+We believe that locating effective detail distribution areas requires information interaction between local details and global semantics. To achieve this, we design a semantic and contour detector that adopts a cascade structure of cross-attention and self-attention. This approach uses global semantic information garnered from self-attention to autonomously select out local details situated at the borders of the contours. On this base, we further utilize the MLP-based extraction branch to purify contour details, thereby significantly enhancing the effect of refinement.
+
+# 3. Model
+
+In this section, we introduce the process of establishing the complete model architecture of EFormer, which comprises the backbone for encoder, the transformer block for decoder, and the prediction stage, as shown in Fig. 2.
+
+# 3.1. Backbone for Encoder
+
+The CNN backbone takes RGB images  $I \in \mathbb{R}^{B \times 3 \times H \times W}$  as input, where  $B$  is the number of images, to obtain a pyramid of feature maps  $F = \left\{F_{enc\frac{1}{4}}, F_{enc\frac{1}{8}}, F_{enc\frac{1}{16}}\right\}$  with dimension  $F_{enc\frac{1}{i}} \in \mathbb{R}^{B \times C_i \times \frac{H}{i} \times \frac{W}{i}}$ . As we all know, in the pyramid of feature maps, shallow features have high resolution and rich details, while deep features have low reso
+
+![](images/19dc18b25a5ce7108cc5db14d5b7e0dec35cccd2f842cb026d2fcf2d6a35cae3.jpg)  
+Figure 2. The Architecture of the EFormer. It contains a backbone for encoder, a transformer block for decoder, and the prediction stage, where the transformer block includes a semantic and contour detector, a semantic extraction branch, and a contour-edge extraction branch.
+
+lation but more concentrated semantic information. Based on this, our research makes the cross-attention mechanism to filter the high-frequency contour features from high-resolution features, using low-frequency semantic information in low-resolution features. However, since the large resolution of feature maps would bring large computation cost, we use the  $F_{enc\frac{1}{8}}$  as  $F_{HR}$  (high-resolution features) and use the  $F_{enc\frac{1}{16}}$  as  $F_{LR}$  (low-resolution features) in our base model. The impact of other choices of  $F_{HR}$  and  $F_{LR}$  on model performance is demonstrated in ablation experiments Sec. 4.4. As shown in Fig. 2, the  $F_{HR}$  serves as the source of contour flow, providing high-frequency contour detail features for the following processes. Correspondingly, the  $F_{LR}$  serves as the source of semantic flow, providing low-frequency semantic information. Then, the  $F_{HR}$  and  $F_{LR}$  are projected to same dimension  $C$  and respectively flattened into high-resolution feature embedding as  $F_{HR}^{em}$  and low-resolution feature embedding as  $F_{LR}^{em}$ , where  $F_{HR}^{em} \in R^{N\times B\times C}$ ,  $F_{LR}^{em} \in R^{N\times B\times C}$ , and  $N = \frac{H}{8}\times \frac{W}{8}$ . Both of them are utilized as inputs for the transformer block.
+
+# 3.2. Transformer Block for Decoder
+
+We propose a novel transformer block in the model decoding stage to enhance the ability to capture and extract visual features. It not only does not reduce the focus on low-frequency semantic features but also significantly increases the attention to high-frequency contour features. Its detailed architecture is depicted in Fig. 3. The novel transformer block first uses a Semantic and Contour De
+
+tector (SCD) to locate and capture high-frequency contour features and low-frequency semantic features. To achieve this, we cascade a cross-attention layer (CA) and a self-attention layer (SA) in the SCD. Then, as shown in Fig. 3, a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB) are further built to handle high-frequency contour feature flow and low-frequency semantic feature flow respectively. These branches can separately purify more refined high-frequency contour features and extract more complete low-frequency semantic features. This ensures an outcome that is optimal for both types of features.
+
+Semantic and Contour Detector (SCD). There is a fact that low-frequency semantic features far outperform high-frequency contour detail features in terms of quantity and distribution density. So, in the global information propagation process, the low-frequency semantic information easily dominates the feature representations of the original transformer block with self-attention module.
+
+To prevent the loss of high-frequency contour detail information, we first use a cross-attention layer with different resolution features rather than a self-attention layer. The cross-attention layer can capture high-frequency detail features with high semantic correlation in high-resolution features, based on the low-frequency semantic information in low-resolution features. Technically, after getting  $F_{HR}^{em}$  and  $F_{LR}^{em}$ , we use Layer Norm (LN) processing them to maintain consistency. Due to the sensitivity of the attention mechanism towards positional information, we use absolute and
+
+![](images/8b731e62de48786a607962e324bfde9d2bf61d970716c65951847e6993d27c69.jpg)  
+Figure 3. Our transformer block. It includes a Semantic and Contour Detector (SCD), a Semantic Extraction Branch (SEB), and a Contour-Edge Extraction Branch (CEEB). The detailed composition of the module is shown in the figure.
+
+learnable position encoding (PE) in the SCD. The input  $K$ ,  $Q$ ,  $V$  of the cross-attention layer (CA) is calculated as
+
+$$
+K = L N \left(F _ {H R} ^ {e m}\right) + P E \tag {2}
+$$
+
+$$
+Q = L N \left(F _ {L R} ^ {e m}\right) + P E \tag {3}
+$$
+
+$$
+V = L N \left(F _ {L R} ^ {e m} + F _ {H R} ^ {e m}\right) \tag {4}
+$$
+
+By performing cross-attention calculations using features of different resolutions, it is possible to automatically locate and capture the discrete contours along the portrait and allocate more attention to corresponding positions. The contour-edge features can be derived from the output of the CA.
+
+$$
+F _ {\text {c o n t o u r - e d g e}} = C A (K, Q, V) \tag {5}
+$$
+
+Although CA pays extra attention to contour-edge features, it still captures the overall semantics. By the residual connection between CA and  $V$ , both the semantic and contour-edge features of the portrait are enhanced, especially the
+
+contour-edge features.
+
+$$
+F _ {\text {e n h a n c e}} = C A (K, Q, V) + V \tag {6}
+$$
+
+Since semantic information is still present, we use the self-attention layer (SA) after CA. SA can locate the semantic feature within the portrait and contours, thereby emphasizing the semantic attributes at their respective positions. After using LN to normalize  $F_{\text{enhance}}$ , the input  $K', Q', V'$  of the SA is calculated as
+
+$$
+K ^ {\prime} = Q ^ {\prime} = L N \left(F _ {\text {e n h a n c e}}\right) + P E \tag {7}
+$$
+
+$$
+V ^ {\prime} = L N \left(F _ {\text {e n h a n c e}}\right) \tag {8}
+$$
+
+$$
+F _ {\text {c o n t o u r - s e m a n t i c}} = S A \left(K ^ {\prime}, Q ^ {\prime}, V ^ {\prime}\right) + V ^ {\prime} \tag {9}
+$$
+
+Through SA filtering the semantic and contour features, the residual connection between  $V'$  and the output of SA, we can produce  $F_{contour - semantic}$  as the output of the semantic and contour detector.
+
+Feature Extraction Branches. After SCD, we use multi-layer perceptron (MLP) to build a Contour-Edge Extraction Branch (CEEB) and a Semantic Extraction Branch (SEB). To prevent the loss of high-frequency contour detail information, we aggregate  $F_{contour - semantic}$  getting from SCD and  $F_{HR}^{em}$  in the contour flow. After getting the merged feature normalized by LN, we employ CEEB to purify it, in order to further extract and refine high-frequency contour details individually. Similarly, to obtain more comprehensive and consistent portrait semantic features separately, we merge  $F_{contour - semantic}$  and  $F_{LR}^{em}$  in the semantic flow. Then, we normalize the aggregated feature by applying LN and use SEB to filter it.
+
+$$
+F _ {\text {c o n t o u r}} = M L P \left(L N \left(F _ {H R} ^ {\text {e m}} + F _ {\text {c o n t o u r - s e m a n t i c}}\right)\right) \tag {10}
+$$
+
+$$
+F _ {\text {s e m a n t i c}} = M L P \left(L N \left(F _ {L R} ^ {\text {e m}} + F _ {\text {c o n t o u r - s e m a n t i c}}\right)\right) \tag {11}
+$$
+
+# 3.3. Prediction Stage
+
+After getting key characteristics  $F_{semantic} \in R^{N \times B \times C}$  and  $F_{contour} \in R^{N \times B \times C}$  from transformer block, where  $N = \frac{H}{8} \times \frac{W}{8}$ , we transform them into  $F_{semantic}' \in R^{B \times C \times \frac{H}{8} \times \frac{W}{8}}$  and  $F_{contour}' \in R^{B \times C \times \frac{H}{8} \times \frac{W}{8}}$  by converting the vector into the matrix. To achieve improved outcomes in portrait segmentation, we propose fusing the two kinds of features and ultimately acquiring the predicted portrait matte from the segmentation head module.
+
+Fuse. We attain the semantic feature through the residual connection between  $F_{LR}^{\prime}$  and  $F_{semantic}^{\prime}$  of semantic flow. Similarly, we get the contour feature from contour flow and subsequently merge the  $F_{semantic}^{\prime}$  and  $F_{contour}^{\prime}$  to generate the fused features  $F_{contour - semantic}^{\prime}$ .
+
+$$
+F _ {\text {s e m a n t i c}} ^ {\prime} = \operatorname {C o n v} \left(F _ {\text {s e m a n t i c}} ^ {\prime} + F _ {L R} ^ {\prime}\right) \tag {12}
+$$
+
+$$
+F _ {\text {c o n t o u r}} ^ {\prime} = \operatorname {C o n v} \left(F _ {\text {c o n t o u r}} ^ {\prime} + F _ {H R} ^ {\prime}\right) \tag {13}
+$$
+
+$$
+F _ {\text {c o n t o u r - s e m a n t i c}} ^ {\prime} = \operatorname {F u s e} \left(F _ {\text {s e m a n t i c}} ^ {\prime} + F _ {\text {c o n t o u r}} ^ {\prime}\right) \tag {14}
+$$
+
+Segmentation Head. Finally, We gather  $F_{contour - semantic}^{\prime}$  and  $F_{enc\frac{1}{4}}$  with higher resolution to predict portrait matte in 1/4 scale of the input image and upscale it to the original scale through bi-linear interpolation as the output of our model.
+
+$$
+m a t t e = H e a d \left(F _ {c o n t o u r - s e m a n t i c} ^ {\prime} + F _ {e n c \frac {1}{4}}\right) \tag {15}
+$$
+
+# 4. Experiments
+
+# 4.1. Dataset and Evaluation
+
+Dataset. We utilize composite training data from two sources: the foreground image dataset VideoMatte240K (JPEG SD Format)[21] and Adobe Image Matting (AIM)[35]. For the background images, we select BG10K following the approach of BGMv2[21], which is a collection of photographs depicting various life scenes without any human portraits. We select foregrounds from VideoMatte240K-JPEG-SD and AIM, and backgrounds from BG10K to composite image datasets. We split VideoMatte240K-JPEG-SD into 234,982/3,007/2720 image sets with image resolutions of  $224 \times 224$  and  $512 \times 288$  for training, validating, and testing our model. Similarly, BG10K is split into 9000/1000 image sets, and AIM is split into 214/10 image sets for training and testing, where AIM only includes the portrait images with resolutions at  $512 \times 512$ . Finally, we use the aforementioned testing datasets to compare and evaluate the performance of EFormer against other models.
+
+Evaluation. We mainly consider portrait matting accuracy for evaluation. For portrait matting accuracy, we use Mean Absolute Difference (MAD), Mean Squared Error (MSE), Gradient (Grad), and Connectivity (Conn) as evaluation metrics. We also scale MAD, MSE, Grad, and Conn by  $10^{3}$ ,  $10^{3}$ ,  $10^{-3}$ , and  $10^{-3}$  respectively, for convenience of reference. For all these metrics, the lower number represents better performance.
+
+# 4.2. Implementation Details
+
+Training Settings. For training the network, we use a single RTX 3090 GPU with batch size at 24. The optimizer is AdamW and the initial learning rate is set to  $10^{-4}$ . All models in ablation studies are trained for 25 epochs, with the learning rate decaying by a factor of 0.8 every 5 epochs. Additionally, to augment the data, each input image is subjected to random horizontal flipping.
+
+Backbone. The backbone we use for the network is the pretrained ResNet50[11], since it is the most frequently used backbone in prior works. We use the implementation and weights from torchvision.
+
+<table><tr><td>Model</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>DeepLabv3[4]</td><td>14.4700</td><td>9.6700</td><td>8.5500</td><td>1.6900</td></tr><tr><td>MODnet[17]</td><td>10.3900</td><td>5.6500</td><td>2.0200</td><td>1.0400</td></tr><tr><td>BGMv2[21]</td><td>4.1858</td><td>1.7934</td><td>1.3011</td><td>0.5326</td></tr><tr><td>RVM[22]</td><td>5.9900</td><td>1.1700</td><td>1.1000</td><td>0.3400</td></tr><tr><td>EFormer</td><td>2.3097</td><td>0.6637</td><td>0.4580</td><td>0.2740</td></tr></table>
+
+Table 1. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at  $512 \times 288$ .  
+
+<table><tr><td>Model</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>DeepLabv3[4]</td><td>29.64</td><td>23.78</td><td>20.17</td><td>7.71</td></tr><tr><td>MODnet[17]</td><td>21.66</td><td>14.27</td><td>5.37</td><td>5.23</td></tr><tr><td>BGMv2[21]</td><td>44.61</td><td>39.08</td><td>5.54</td><td>11.60</td></tr><tr><td>RVM[22]</td><td>14.84</td><td>8.93</td><td>4.35</td><td>3.83</td></tr><tr><td>EFormer</td><td>7.47</td><td>2.13</td><td>2.83</td><td>1.90</td></tr></table>
+
+Table 2. Comparisons on the test set of AIM. Bold indicates the best performance among these models under the inputs with the same resolution at  $512 \times 512$ .
+
+Transformer Block for Decoder. The transformer block is based on the transformer decoder of DETR. We continuously use four transformer blocks in the decoder of the network. We set the channel  $C = 256$  and the number of attention heads  $M = 8$  for each multi-head attention module.
+
+Prediction Stage. After obtaining the semantic and contour feature maps of the portrait from the stacked transformer blocks, we fuse both of them. Then, the alpha mattes predicted by the segmentation head are upsampled to the original size of the input image with bi-linear interpolation.
+
+Loss function. To compute losses during the training stage, we use a Binary CrossEntropy Loss function.
+
+$$
+L _ {b c e} = - (g \times \log (m) + (1 - g) \times \log (1 - m)) \tag {16}
+$$
+
+Where,  $m$  is the prediction of the network,  $g$  is the ground truth.
+
+# 4.3. Comparisons to State-of-the-art Methods
+
+To demonstrate the progressiveness of EFormer, we compare it against the latest state-of-the-art trimap-free portrait matting solutions. These are including DeepLabV3[4] with ResNet101[11] backbone, BGMv2[21] with ResNet50[11] backbone, MODNet[17], and RVM[22]. Our assessment is based on the testing datasets of VideoMatte240K-JPEG-SD and AIM, respectively. As the training files for MODNet are not available, we use its official weights. Furthermore, since BGMv2, DeepLabV3, and RVM have already been trained on all datasets, our method can be compared with them in a fair manner.
+
+![](images/a44cfc197d899a897cdae63917124a1992d807f4b3ea9c1a1cb1af45bf9a7ef0.jpg)  
+Figure 4. Visualization of portrait matte predictions from MODnet, BGMv2, and EFormer under challenging image from the test set of VideoMatte240K (JPEG SD Format). Our model shows a better ability to distinguish ambiguous foreground contours, as indicated by the red box. Please zoom in to view the detailed information.
+
+<table><tr><td>Model</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>DeepLabv3[4]</td><td>18.1586</td><td>14.7623</td><td>6.0015</td><td>2.8824</td></tr><tr><td>MODnet[17]</td><td>13.1400</td><td>8.9656</td><td>2.6028</td><td>2.1543</td></tr><tr><td>BGMv2[21]</td><td>7.0471</td><td>2.3543</td><td>1.5621</td><td>1.0651</td></tr><tr><td>RVM[22]</td><td>6.1357</td><td>2.2647</td><td>0.6523</td><td>0.2866</td></tr><tr><td>EFormer</td><td>4.1357</td><td>1.4280</td><td>0.3760</td><td>0.1753</td></tr></table>
+
+As shown in the Tab. 1, Tab. 2 and Tab. 3, our model exhibits superior performance in portrait matte tasks. It has a stronger capability to distinguish the contour of a portrait. And it can finely segment edge details (e.g. fingers), as exemplified in Fig. 4.
+
+# 4.4. Ablation Studies
+
+Training Epochs. In the ablation study with an image resolution of  $224 \times 224$ , we conduct 25 training epochs. Every 5 epochs of training, we evaluate the training effect of the model. At the same time, the learning rate is reduced by 0.8 times. The experimental results indicate that our model achieves optimal performance in the 20 epoch, as shown in the Tab. 4.
+
+Table 3. Comparisons on the test set of VideoMatte240K-JPEG-SD. Bold indicates the best performance among these models under the inputs with the same resolution at  $224 \times 224$ .  
+
+<table><tr><td></td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>5-epoch</td><td>4.3953</td><td>1.4824</td><td>0.4089</td><td>0.1865</td></tr><tr><td>10-epoch</td><td>4.3108</td><td>1.4766</td><td>0.3908</td><td>0.1834</td></tr><tr><td>15-epoch</td><td>4.2867</td><td>1.5238</td><td>0.4087</td><td>0.1833</td></tr><tr><td>20-epoch</td><td>4.1357</td><td>1.4280</td><td>0.3760</td><td>0.1753</td></tr><tr><td>25-epoch</td><td>4.1811</td><td>1.4797</td><td>0.3858</td><td>0.1779</td></tr></table>
+
+Table 4. Ablation Study on the Training Epochs with the test set of VideoMatte240K-JPEG-SD (resolution at  $224 \times 224$ ). **Bold** indicates the best performance among these models. We use the result of 20-epoch training as a strong baseline model.
+
+The Functional Role of Each Module. To demonstrate the effectiveness of the model in capturing and extracting contour details and portrait semantics, we analyze the outputs of CA, SA, CEEB, and SEB, and utilize Class Activation Map (CAM) to visualize the attention allocation of each layer's features, as shown in Fig. 5.
+
+From the first line of Fig. 5, it can be evident that after the cross-attention layer, the model clearly strengthens its attention to the contours of the portrait. Following the contour prior, the self-attention layer captures the portrait's semantics thoroughly. CEEB and SEB models provide sharper contour and more comprehensive semantic information.
+
+Comparing the second and first lines of Fig. 5, it is ob
+
+![](images/f4a86092d25db2d4a4c9dd6b16240410a2a59c5e7287d4d91926926071d90d7f.jpg)  
+Figure 5. We use Grad-CAM to visualize the attention distribution of various features in CA, SA, CEEB, and SEB. The model in the first row contains four complete modules. For comparison, the SA layer and CA layer were removed from the models in the second and third rows, respectively.
+
+<table><tr><td>CA</td><td>SA</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>✓</td><td></td><td>4.2762</td><td>1.5344</td><td>0.4092</td><td>0.1823</td></tr><tr><td></td><td>✓</td><td>4.1710</td><td>1.4354</td><td>0.3773</td><td>0.1769</td></tr><tr><td>✓</td><td>✓</td><td>4.1357</td><td>1.4280</td><td>0.3760</td><td>0.1753</td></tr></table>
+
+served that removing the self-attention layer considerably weakens the semantic information extracted by SEB. In the comparison between the third and first lines of Fig. 5, it can be indicated that removing the cross-attention layer significantly weakens the contour information extracted by CEEB.
+
+After qualitative analysis, the prediction results of all the above models are evaluated. It is concluded that the model with CA, SA, CEEB, and SEB is both effective and rational, as demonstrated in the Tab. 5.
+
+As is well known, when it comes to extracting features for portrait segmentation tasks, semantic information plays a significant role, in contrast, contour details account for a relatively small proportion. Owing to this, as shown in the Tab. 5 and Fig. 5, the model without SA leads to a significant increase in all metrics due to the lack of amount of semantic information, while the model without CA experiences a slight increase in all indicators due to the loss of contour details. In order to fully harness the potential of the vision transformer, we integrate both CA and SA to ensure semantic completeness and enhance attention towards contour details, thereby achieving more comprehensive and fine-grained segmentation results.
+
+The Selection of  $F_{HR}$  and  $F_{LR}$ . In our base model,
+
+Table 5. Ablation Study on the Functional Role of CA and SA with the test set of VideoMatte240K-JPEG-SD (resolution at  $224 \times 224$ ). **Bold** indicates the best performance among these models.  
+
+<table><tr><td>FHR</td><td>FLR</td><td>MAD↓</td><td>MSE↓</td><td>Grad↓</td><td>Conn↓</td></tr><tr><td>Fenc1/8</td><td>Fenc1/16</td><td>4.1357</td><td>1.4280</td><td>0.3760</td><td>0.1753</td></tr><tr><td>Fenc1/4</td><td>Fenc1/8</td><td>3.9970</td><td>1.3343</td><td>0.3328</td><td>0.1679</td></tr><tr><td>Fenc1/4</td><td>Fenc1/16</td><td>3.9730</td><td>1.2949</td><td>0.3215</td><td>0.1665</td></tr></table>
+
+Table 6. Ablation Study on the Selection of  $F_{HR}$  and  $F_{LR}$  with the test set of VideoMatte240K-JPEG-SD (resolution at  $224 \times 224$ ). Bold indicates the best performance among these models.
+
+considering the large resolution of feature maps would bring large computation cost, we use the  $F_{enc}^{\frac{1}{8}}$  as  $F_{HR}$  (high-resolution features) and use the  $F_{enc}^{\frac{1}{16}}$  as  $F_{LR}$  (low-resolution features). Then the  $F_{HR}$  serves as the source of contour flow, providing high-frequency contour detail features. Correspondingly, the  $F_{LR}$  serves as the source of semantic flow, providing low-frequency semantic information. As shown in the Tab. 6, after using  $F_{enc}^{\frac{1}{4}}$  with higher resolution as  $F_{HR}$ , our model can capture more high-frequency contour detail features and its performance is better. And after using  $F_{enc}^{\frac{1}{16}}$  with lower resolution as  $F_{LR}$ , our model can use more concentrated low-frequency semantic information, which makes all indicators optimal.
+
+# 5. Conclusion
+
+In this paper, we propose a strategy that use of cross-attention module between different resolutions guides the EFormer to autonomously locate and capture high-frequency contour features. Without manual parameter tuning, the proposed method can adaptability capture high-frequency details neglected by previous transformer-based methods. Based on the cross-attention module, we further construct a semantic and contour detector (SCD), which can enhance the transformer's attention to both the high-frequency contour and low-frequency semantic features. Additionally, we build a contour-edge extraction branch (CEEB) and a semantic extraction branch (SEB) to extract finer high-frequency contour details and more comprehensive portrait's low-frequency semantic information. Extensive experiments show that our method outperforms previous portrait matting solutions and exhibits robust performance during testing and inference.
+
+# Acknowledgement
+
+The work was jointly supported by the National Natural Science Foundations of China under grant No.62272364, the Teaching Reform Project of Shaanxi Higher Continuing Education under Grant No.21XJZ004.
+
+# References
+
+[1] Yagiz Aksoy, Tunc Ozan Aydin, and Marc Pollefeys. Designing effective inter-pixel information flow for natural image matting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 29-37, 2017. 3  
+[2] Yagiz Aksoy, Tae-Hyun Oh, Sylvain Paris, Marc Pollefeys, and Wojciech Matusik. Semantic soft segmentation. ACM Transactions on Graphics (TOG), 37(4):1-13, 2018. 3  
+[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3  
+[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6, 7  
+[5] 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. 3  
+[6] 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. 3  
+[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2  
+[8] Dazhao Du, Bing Su, Yu Li, Zhongang Qi, Lingyu Si, and Ying Shan. Efficient u-transformer with boundary-aware loss for action segmentation. arXiv preprint arXiv:2205.13425, 2022. 3  
+[9] Xiaoxue Feng, Xiaohui Liang, and Zili Zhang. A cluster sampling method for image matting via sparse coding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 204-219. Springer, 2016. 3  
+[10] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaou Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056. IEEE, 2011. 3  
+[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6  
+[12] 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. 3  
+[13] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. 3  
+[14] Jubin Johnson, Ehsan Shahrian Varnousfaderani, Hisham Cholakkal, and Deepu Rajan. Sparse coding for alpha mat
+
+ting. IEEE Transactions on Image Processing, 25(7):3032-3043, 2016. 3  
+[15] Levent Karacan, Aykut Erdem, and Erkut Erdem. Image matting with kl-divergence based sparse sampling. In Proceedings of the IEEE international conference on computer vision, pages 424-432, 2015. 3  
+[16] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4412–4421, 2022. 3  
+[17] Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1140-1147, 2022. 2, 3, 6, 7  
+[18] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3  
+[19] Anat Levin, Alex Rav-Acha, and Dani Lischinski. Spectral matting. IEEE transactions on pattern analysis and machine intelligence, 30(10):1699-1712, 2008. 3  
+[20] Yaoyi Li and Hongtao Lu. Natural image matting via guided contextual attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11450-11457, 2020. 1  
+[21] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 2, 3, 6, 7  
+[22] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 2, 3, 6, 7  
+[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2, 3  
+[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2  
+[25] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 2, 3  
+[26] Olivier Petit, Nicolas Thome, Clement Rambour, Loic Themyr, Toby Collins, and Luc Soler. U-net transformer: Self and cross attention for medical image segmentation. In Machine Learning in Medical Imaging: 12th International Workshop, MLMI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, September 27, 2021, Proceedings 12, pages 267-276. Springer, 2021. 3  
+[27] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of
+
+the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021.3  
+[28] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 2, 3  
+[29] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xinchao Wang, and Shuicheng Yan. Inception transformer. Advances in Neural Information Processing Systems, 35:23495-23509, 2022. 2, 3  
+[30] Yanan Sun, Guanzhi Wang, Qiao Gu, Chi-Keung Tang, and Yu-Wing Tai. Deep video matting via spatio-temporal alignment and aggregation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6975–6984, 2021. 1  
+[31] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5463-5474, 2021. 3  
+[32] Jian Wang, Chenhui Gou, Qiman Wu, Haocheng Feng, Junyu Han, Errui Ding, and Jingdong Wang. Rrformer: Efficient design for real-time semantic segmentation with transformer. Advances in Neural Information Processing Systems, 35:7423-7436, 2022. 3  
+[33] Yu Wang, Yi Niu, Peiyong Duan, Jianwei Lin, and Yuanjie Zheng. Deep propagation based image matting. In IJCAI, pages 999-1006, 2018. 1  
+[34] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 3  
+[35] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 1, 2, 6  
+[36] Haichao Yu, Ning Xu, Zilong Huang, Yuqian Zhou, and Humphrey Shi. High-resolution deep image matting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3217-3224, 2021. 3  
+[37] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5128-5137, 2021. 3  
+[38] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 3  
+[39] Bingke Zhu, Yingying Chen, Jinqiao Wang, Si Liu, Bo Zhang, and Ming Tang. Fast deep matting for portrait animation on mobile phone. In Proceedings of the 25th ACM international conference on Multimedia, pages 297-305, 2017. 3

2024/EFormer_ Enhanced Transformer towards Semantic-Contour Features of Foreground for Portraits Matting/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c099c386318137ef5f3e4a8c16184ed3601bd5298d3b1897237a6d3c51933f41
+size 540061

2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/174c2199-8aa8-4e45-8d2e-205eb2e2e099_content_list.json

@@ -0,0 +1,1703 @@
+[
+    {
+        "type": "text",
+        "text": "EGTR: Extracting Graph from Transformer for Scene Graph Generation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            130,
+            854,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jinbae Im<sup>1</sup> JeongYeon Nam<sup>1</sup> Nokyung Park<sup>1,2,3*</sup> Hyungmin Lee<sup>2</sup> Seunghyun Park<sup>1</sup>  \n<sup>1</sup>NAVER Cloud AI <sup>2</sup>NAVER <sup>3</sup>Korea University  \n{jinbae.im, jy.nam, nokyung.park99, hyungmin.lee, seung.park}@navercorp.com",
+        "bbox": [
+            122,
+            179,
+            843,
+            239
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            268,
+            313,
+            284
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Scene Graph Generation (SGG) is a challenging task of detecting objects and predicting relationships between objects. After DETR was developed, one-stage SGG models based on a one-stage object detector have been actively studied. However, complex modeling is used to predict the relationship between objects, and the inherent relationship between object queries learned in the multi-head self-attention of the object detector has been neglected. We propose a lightweight one-stage SGG model that extracts the relation graph from the various relationships learned in the multi-head self-attention layers of the DETR decoder. By fully utilizing the self-attention by-products, the relation graph can be extracted effectively with a shallow relation extraction head. Considering the dependency of the relation extraction task on the object detection task, we propose a novel relation smoothing technique that adjusts the relation label adaptively according to the quality of the detected objects. By the relation smoothing, the model is trained according to the continuous curriculum that focuses on object detection task at the beginning of training and performs multi-task learning as the object detection performance gradually improves. Furthermore, we propose a connectivity prediction task that predicts whether a relation exists between object pairs as an auxiliary task of the relation extraction. We demonstrate the effectiveness and efficiency of our method for the Visual Genome and Open Image V6 datasets. Our code is publicly available at https://github.com/naver-ai/egtr.",
+        "bbox": [
+            75,
+            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": "Scene Graph Generation (SGG) [8] aims to generate a scene graph that represents objects as nodes and relationships between objects as edges from an image as shown in Fig. 1b. SGG is a challenging task as it is required to go beyond simply detecting objects but predicting the relationships between them based on a comprehensive understanding of the",
+        "bbox": [
+            75,
+            782,
+            468,
+            875
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/da086494a3493dff7e071faf8fea070928900aae618d983c549510da6ad11f8c.jpg",
+        "image_caption": [
+            "(a) Visual Genome example"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            271,
+            746,
+            395
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/9fcc74a969bc741c22cf5e31e09c3ed6a0b106a111765290803f5a92ea5dbd50.jpg",
+        "image_caption": [
+            "(b) Scene Graph"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            764,
+            268,
+            883,
+            323
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/d3d669486085177c87b65b065112e8bb0b5abe8f60a9d5e1549828477fc1cbc7.jpg",
+        "image_caption": [
+            "(c) Attention Graph",
+            "Figure 1. Motivation. SGG task aims to predict scene graph (Fig. 1b) with objects as nodes and relations as edges. We draw a plausible attention graph (Fig. 1c) simply by connecting objects with high attention weights to edges from the self-attention layers of the pre-trained DETR. It shows the potential for the self-attention from the object detector to contain rich information that aids in predicting the relations of the scene graph."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            767,
+            343,
+            883,
+            397
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "scene. Since a scene graph provides structural information of the image, it can be used for various vision tasks that require a higher level of understanding and reasoning about images such as image captioning [5, 10, 41], image retrieval [8, 28], and visual question answering [14, 43].",
+        "bbox": [
+            496,
+            535,
+            892,
+            611
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Most previous studies [16, 19, 33, 39, 42] took two-stage SGG approaches that detect objects first and then predict their relations. However, these approaches incurred high computational costs and risks of error propagation from the object detection. To address the drawbacks of the two-stage approaches, one-stage SGG models that perform object detection and relation prediction at once [3, 9, 17, 20, 26, 29, 35] have been studied recently, leveraging one-stage object detectors such as DETR [1]. Since edges in a scene graph can be represented as subject-predicate-object triplets, many studies embraced triplet-based approaches as shown in Figs. 2a and 2b. However, object-triplet detection models [3, 17] required sophisticated triplet detectors to obtain necessary information for triplet queries from the object detector. Moreover, triplet detection models [9, 35] lacked the ability to detect objects without relation such as \"tree\" in Fig. 1b, by focusing solely on the triplet detection without an object detector. Considering that objects without relations account for more than",
+        "bbox": [
+            496,
+            613,
+            893,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Most work was done during the internship at Naver Cloud AI.",
+        "bbox": [
+            94,
+            886,
+            442,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "24229",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/eeef0756f110b56051eef1ee222a4df8ecbb04eaa0125bb5bc46ca221ac47ed5.jpg",
+        "image_caption": [
+            "(a) Object-Triplet Detection Models"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            107,
+            89,
+            349,
+            205
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/17fbf1da620e9dc2c721e805545b7c7f43c92cb1db5cdc6e1494d2fbdc0930f5.jpg",
+        "image_caption": [
+            "(b) Triplet Detection Models",
+            "Figure 2. Comparison with existing one-stage SGG models. (a) Object-Triplet Detection Models introduce additional triplet queries and a triplet detector to the object detector. The triplet detector requires additional modules to incorporate information from the object detector into the triplet queries. (b) Triplet Detection Models focus on detecting triplets directly without an object detector. Objects without relations may not be detected. (c) Relation Extraction Models extract relations from the object detector without a separate triplet detector. In particular, ours extracts relations more effectively by utilizing by-products from the self-attention of the object detector."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            385,
+            90,
+            571,
+            205
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/1200dc7f83785412b59a2e3649fd9623129020a4a233efc8ba94d41c3b085407.jpg",
+        "image_caption": [
+            "(c) Relation Extraction Models"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            616,
+            90,
+            859,
+            205
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "$42\\%$  in Visual Genome [12] data, their priority lies in detecting a sub-graph rather than the complete scene graph.",
+        "bbox": [
+            75,
+            309,
+            467,
+            339
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address the shortcomings of the existing one-stage SGG models, we focus on the relationships between objects inherent in the object detector. As shown in Fig. 1a, objects are related to each other. For instance, when a horse appears in a scene, a person is likely to appear in the scene, and clothing such as hats, jackets, and pants often depicts the person's current situation. From this intuition, there has been a long belief [2, 7, 22] that modeling relationship or context between objects would be beneficial to object detection task. Accordingly, recent one-stage object detectors [1, 32, 46] have incorporated self-attention layers [36] to implicitly model the relationships among the object queries. We hypothesize that self-attention between object queries learned in a one-stage object detector might contain valuable information for predicting triplet outputs. In our preliminary investigation, we are able to extract a plausible attention graph by simply connecting two object queries with high attention weights from the pre-trained DETR, as shown in Fig. 1c. It shows the potential that the attention weights between object queries can be interpreted as relations between them.",
+        "bbox": [
+            73,
+            340,
+            468,
+            656
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "From the findings, we propose a lightweight one-stage scene graph generator EGTR, which stands for Extracting Graph from TRansformer. We design the model to comprehensively leverage the by-products of the object detector, eliminating the need for a separate triplet detector, as depicted in Fig. 2c. From the multi-head self-attention layers of the object detector, we regard an attention query and key, where their relations are learned in the attention weights, as the subject entity and object entity, respectively. Subsequently, we leverage a shallow classifier to predict relations between them. Due to the abundant information about the relationships among the objects present in the by-products derived from all layers of self-attention, we can effectively extract the scene graph.",
+        "bbox": [
+            75,
+            657,
+            468,
+            869
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Since the relation extraction task is dependent on the object detection task, we speculate that performing relation ex",
+        "bbox": [
+            75,
+            869,
+            468,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "traction without sufficiently learned representations of the object queries might be harmful. Therefore, we devise a novel adaptive smoothing technique that smoothes the value of the ground truth relation label based on the object detection performance. With the adaptive smoothing, the model is trained with a continuous curriculum that initially focuses on object detection and gradually performs multi-task learning. Furthermore, we propose a connectivity prediction task as an auxiliary task for relation extraction, aiming to predict the existence of any relationship between a subject entity and an object entity. This auxiliary task facilitates the acquisition of representations for the relation extraction.",
+        "bbox": [
+            496,
+            309,
+            890,
+            489
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To verify the effectiveness of the proposed method, we conduct experiments on two representative SGG datasets: Visual Genome [12] and Open Images V6 [13]. By actively utilizing the by-products of the object detector, EGTR shows the best object detection performance and comparable triplet detection performance, with the fewest parameters and the fastest inference speed.",
+        "bbox": [
+            496,
+            492,
+            890,
+            597
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our main contributions can be summarized as follows:",
+        "bbox": [
+            517,
+            599,
+            880,
+            612
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose EGTR that generates scene graphs efficiently and effectively by utilizing the multi-head self-attention by-products derived from the object detector.",
+            "- We present adaptive smoothing, enabling effective multitask learning for both object detection and relation extraction. In addition, the proposed connectivity prediction offers clues to the relation extraction.",
+            "- Our comprehensive experiments show the superiority of the proposed model framework and the effectiveness of the devised training techniques."
+        ],
+        "bbox": [
+            500,
+            614,
+            893,
+            766
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            782,
+            640,
+            799
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "SGG models can be categorized into two groups: two-stage models and one-stage models. For two-stage models [4, 11, 15, 16, 19, 23-25, 30, 33, 37, 39, 42, 45], separate object detection model and relation prediction model are trained sequentially. They usually detect  $N$  objects from the off-the-shelf object detector such as Faster R-CNN [27],",
+        "bbox": [
+            496,
+            810,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "24230",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "and then all possible combinations of the detected objects are fed into the relation prediction model to predict the relations between each object pair. Although they showed high relation extraction performance, they have the inherent limitation that the object detector that helps with relation extraction is trained separately, resulting in a significant increase in model complexity.",
+        "bbox": [
+            75,
+            90,
+            470,
+            196
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As for the one-stage models [3, 9, 17, 20, 26, 29, 35], the object detection and relation prediction are trained in an end-to-end manner. Early studies [20, 26] proposed fully convolutional SGG models and took pixel-based approaches. After DETR [1] brought a huge success as a Transformer [36]-based one-stage object detector, many one-stage SGG studies are based on one-stage object detectors [1, 32, 46]. They efficiently modeled SGG by introducing object queries or triplet queries. We categorize them into three distinct groups: (a) object-triplet detection models, (b) triplet prediction models, and (c) relation extraction models, as shown in Fig. 2.",
+        "bbox": [
+            75,
+            200,
+            470,
+            383
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Object-Triplet Detection Models. Object-triplet detection models are characterized by introducing additional triplet queries and building a triplet predictor on top of the object detector as shown in Fig. 2a. RelTR [3] introduced paired subject queries and object queries, and SGTR [17] introduced compositional queries decoupled into subjects, objects, and predicates. As the introduced queries are initialized without prior cues from the object detector, the triplet predictor requires modules to incorporate information from the outputs of the object detector and modules for triplet queries to exchange information with each other. It led to the triplet predictor having an intricate structure. In contrast, we refrain from introducing additional queries and regard the attention queries and attention keys, whose relationships are learned in the object detector, as subject and object queries, respectively.",
+        "bbox": [
+            75,
+            392,
+            468,
+            633
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Triplet Detection Models. Triplet detection models directly detect triplets using triplet queries without an object detector, as shown in Fig. 2b. Iterative SGG [9] introduced subject, object, and predicate queries and modeled the conditional dependencies among them with separate subject, object, and predicate multi-layer Transformer decoders. Inspired by Sparse R-CNN [32], Structured Sparse R-CNN (SSR-CNN) [35] designed triplet queries consisting of subject box, object box, subject, object, and predicate queries. Since it was difficult to train the model only with sparse triplet annotations, SSR-CNN introduced somewhat intricate training details: it detected object pairs with a Siamese Sparse RCNN model and auxiliary queries and performed additional triplet matching using the detected object pairs as a pseudo label. Although they show excellent triplet detection performance, the model architecture became even more intricate to detect subjects and objects separately since",
+        "bbox": [
+            75,
+            643,
+            470,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "an explicit object detector is not used. Last but not least, the triplet prediction models focus on detecting a sub-graph composed only of objects with relations, overlooking objects in an image that lack explicit relations.",
+        "bbox": [
+            496,
+            90,
+            890,
+            151
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Relation Extraction Models. Relation extraction models extract a scene graph using a lightweight relation predictor without separate triplet queries or triplet detector. Relationformer [29] added a special \"[rln]\" token to capture global information in conjunction with object queries. They concatenated the final hidden representations of the object query pairs and the relation token, followed by a shallow fully connected network for relation prediction. In this work, in conjunction with the final hidden representations, we use the inherent relationship information among the object queries learned in the multi-head self-attention layer of the object detector. Furthermore, we propose training techniques to boost multi-task learning, leading to significantly improved performance with architectural simplicity.",
+        "bbox": [
+            496,
+            156,
+            890,
+            369
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            500,
+            382,
+            591,
+            398
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            500,
+            407,
+            640,
+            422
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we first introduce the formulation of the SGG task and then provide a brief review of the one-stage object detector that serves as the basis for our study.",
+        "bbox": [
+            496,
+            431,
+            890,
+            476
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1.1 Formulation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            494,
+            642,
+            510
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Scene graph generation is a task of generating a scene graph  $\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}$  from an image, where  $\\mathcal{V}$  denotes the node set consisting of objects and  $\\mathcal{E}$  denotes the edge set that represents the relation between objects. Each object entity  $v_{i}\\in \\mathcal{V}$  has an object category label  $v_{i}^{c}$  from a set of object categories  $\\mathcal{C}_v$  and box coordinates  $v_{i}^{b}$ . Each relation  $e_j\\in \\mathcal{E}$  represents the  $j$ -th triplet  $(s_j,p_j,o_j)$ , where subject  $s_j$  and object  $o_j$  indicate related object entities and predicate  $p_j$  has a relation category label  $p_j^c$  from a set of predicate categories  $\\mathcal{C}_p$ . Generating  $\\mathcal{V}$  and  $\\mathcal{E}$  correspond to object detection and relation extraction, respectively.",
+        "bbox": [
+            496,
+            518,
+            890,
+            686
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1.2 One-stage Object Detector",
+        "text_level": 1,
+        "bbox": [
+            500,
+            705,
+            740,
+            720
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our model framework is deeply influenced by DETR [1], a one-stage object detector. In DETR, representations for input images are learned through CNN and Transformer [36] encoder. Transformer decoder enhances object queries by utilizing self- and cross-attention mechanisms. The final object detection heads predict object category labels and box coordinates from the contextualized object queries.",
+        "bbox": [
+            496,
+            728,
+            890,
+            835
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Backbone. CNN backbone generates  $C$ -dimensional feature map  $F \\in \\mathbb{R}^{C \\times H_F \\times W_F}$  from an input image  $x_{\\mathrm{img}} \\in \\mathbb{R}^{3 \\times H_0 \\times W_0}$ . For ResNet-50 backbone,  $H_F$  and  $W_F$  are set to  $H_0 / 32$  and  $W_0 / 32$ , respectively.",
+        "bbox": [
+            496,
+            839,
+            890,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "24231",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/8ecf05244754a50b2be1aece4d420418de000403424f1f62df4891ed67db2042.jpg",
+        "image_caption": [
+            "Figure 3. The overall architecture of EGTR. We present a novel lightweight relation extractor, EGTR, which fully utilizes the self-attention of the DETR decoder. We extract query and key representations from each self-attention layer and concatenate them pairwise to represent relations between them. Additionally, we leverage the last hidden representation in the same manner. To effectively aggregate information, we apply a gated sum and then predict relation with a shallow relation head."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            88,
+            704,
+            280
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/051d3620529018711d7ac2cc82099c6a645af5e74a4b87f1aedd5638cdb66c02.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            709,
+            93,
+            885,
+            183
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/cbe093ab64a6516b8c1b4273701a8dbd74801876a7784e78ebbac8263356f574.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            714,
+            188,
+            885,
+            279
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Transformer Encoder. Transformer encoder enhances image representations based on multi-head self-attention. After applying a  $1 \\times 1$  convolution to reduce dimension from  $C$  to  $d_{\\mathrm{model}}$ , the feature map is flattened to get an input sequence with length  $H_{F}W_{F}$  for the Transformer encoder. Additional positional encodings are used to reflect the spatial information of the feature maps.",
+        "bbox": [
+            75,
+            367,
+            470,
+            474
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Transformer Decoder. Transformer decoder takes  $N$  object queries as input and returns their representations. Each object query detects an individual object, and  $N$  is usually set large enough to cover all objects in the image. Through alternating self-attention and cross-attention layers, object queries learn features of object candidates in the input image. Note that causal attention mask is not applied to the self-attention layer; therefore, fully connected  $N \\times N$  attention weights among the  $N$  object queries are learned from the multi-head self-attention as follows:",
+        "bbox": [
+            75,
+            479,
+            468,
+            630
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nA _ {h} ^ {l} = f \\left(Q _ {h} ^ {l}, K _ {h} ^ {l}\\right) = \\operatorname {s o f t m a x} \\left(Q _ {h} ^ {l} K _ {h} ^ {T} / \\sqrt {d _ {\\text {h e a d}}}\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            102,
+            641,
+            468,
+            664
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $A_h^l \\in \\mathbb{R}^{N \\times N}$  denotes the attention weights from the  $h$ -th head in the  $l$ -th layer.  $Q_h^l \\in \\mathbb{R}^{N \\times d_{\\mathrm{head}}}$  and  $K_h^l \\in \\mathbb{R}^{N \\times d_{\\mathrm{head}}}$  are attention queries and keys, respectively.",
+        "bbox": [
+            76,
+            674,
+            470,
+            722
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Object Detection Heads. Object detection heads detect  $N$  object candidates  $\\{\\hat{v}_i\\}_{i = 1}^N$  from the last layer representations of the object queries  $Z^{L}$ . For each object query, a linear layer is used to predict the object category label  $\\hat{v}_i^c\\in \\mathbb{R}^{N\\times |\\mathcal{C}_v|}$ , and three-layer perceptron with ReLU activation is used to obtain the box coordinates  $\\hat{v}_i^b\\in \\mathbb{R}^{N\\times 4}$ .",
+        "bbox": [
+            76,
+            727,
+            468,
+            820
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Object Detection Loss. To match  $N$  detected object candidates  $\\{\\hat{v}_i\\}_{i=1}^N$  and  $M$  ground truth objects  $\\{v_i\\}_{i=1}^M$ , Carion et al. [1] pad the ground truth objects with  $\\phi$  (no object) and find the best permutation of predicted objects that minimizes total bipartite matching costs. From the permuted",
+        "bbox": [
+            75,
+            824,
+            470,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "predictions  $\\{\\hat{v}_i^{\\prime}\\}_{i = 1}^{N}$  , the loss is calculated as follows:",
+        "bbox": [
+            498,
+            366,
+            852,
+            383
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {o d}} = \\Sigma_ {i = 1} ^ {N} \\left[ \\lambda_ {c} \\mathcal {L} _ {c} \\left(\\hat {v} _ {i} ^ {\\prime c}, v _ {i} ^ {c}\\right) + \\mathbb {1} _ {v _ {i} ^ {c} \\neq \\phi} \\left(\\lambda_ {b} \\mathcal {L} _ {b} \\left(\\hat {v} _ {i} ^ {\\prime b}, v _ {i} ^ {b}\\right)\\right) \\right], \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            511,
+            393,
+            890,
+            414
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{L}_c$  and  $\\mathcal{L}_b$  denote loss function for category labels and box coordinates, respectively.",
+        "bbox": [
+            498,
+            425,
+            890,
+            455
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. EGTR",
+        "text_level": 1,
+        "bbox": [
+            500,
+            467,
+            588,
+            482
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We propose a novel lightweight relation extractor, EGTR, which exploits the self-attention of DETR decoder, as depicted in Fig. 3. Since the self-attention weights in Eq. (1) contain  $N \\times N$  bidirectional relationships among the  $N$  object queries, our relation extractor aims to extract the predicate information from the self-attention weights in the entire  $L$  layers, by considering the attention queries and keys as subjects and objects, respectively.",
+        "bbox": [
+            496,
+            491,
+            890,
+            612
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In order to preserve rich information learned in the self-attention layer, we devise another relation function  $f$  between the attention queries and keys instead of the dot product attention in Eq. (1). As concatenation preserves representations completely, we concatenate the representations on all  $N \\times N$  pairs of attention queries and keys as shown in Fig. 3 (a) to get the relation representations of the  $l$ -th layer  $R_{a}^{l} \\in \\mathbb{R}^{N \\times N \\times 2d_{\\mathrm{model}}}$ . Before pairwise concatenation, to help queries and keys play the role of subjects and objects, a linear projection is added as follows:",
+        "bbox": [
+            496,
+            613,
+            892,
+            763
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR _ {a} ^ {l} = \\left[ Q ^ {l} W _ {S} ^ {l}; K ^ {l} W _ {O} ^ {l} \\right], \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            616,
+            776,
+            890,
+            795
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $Q^{l}\\in \\mathbb{R}^{N\\times d_{\\mathrm{model}}}$  and  $K^l\\in \\mathbb{R}^{N\\times d_{\\mathrm{model}}}$  refer to attention queries and keys of the  $l$  -th layer, and  $[\\cdot ;\\cdot ]$  denotes the pairwise concatenation.  $W_{S}^{l}$  and  $W_{O}^{l}$  are linear weights of shape  $\\mathbb{R}^{d_{\\mathrm{model}}\\times d_{\\mathrm{model}}}$ .",
+        "bbox": [
+            496,
+            806,
+            890,
+            868
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We also utilize the last layer representations of the object queries  $Z^{L} \\in \\mathbb{R}^{N \\times d_{\\mathrm{model}}}$ , which are used for the object",
+        "bbox": [
+            496,
+            869,
+            890,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "24232",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "detection in the same manner:",
+        "bbox": [
+            76,
+            90,
+            279,
+            104
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR _ {z} = \\left[ Z ^ {L} W _ {S}; Z ^ {L} W _ {O} \\right], \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            191,
+            116,
+            468,
+            133
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $W_{S}, W_{O}$  are linear weights of shape  $\\mathbb{R}^{d_{\\mathrm{model}} \\times d_{\\mathrm{model}}}$ . To effectively use the various information learned in all layers, we introduce a gating mechanism as follows:",
+        "bbox": [
+            76,
+            146,
+            468,
+            191
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ng _ {a} ^ {l} = \\sigma \\left(R _ {a} ^ {l} W _ {G}\\right), g _ {z} = \\sigma \\left(R _ {z} W _ {G}\\right), \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            156,
+            200,
+            468,
+            220
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $g_{a}^{l}$  and  $g_{z} \\in \\mathbb{R}^{N \\times N \\times 1}$  represent the gate values obtained through a single linear layer  $W_{G} \\in \\mathbb{R}^{2d_{\\mathrm{model}} \\times 1}$  for  $R_{a}^{l}$  and  $R_{z}$ , respectively. Finally, we extract the relation graph from the gated summation of the relation representations across all layers as follows:",
+        "bbox": [
+            76,
+            229,
+            470,
+            308
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {G} = \\sigma \\left(\\mathrm {M L P} _ {\\mathrm {r e l}} \\left(\\Sigma_ {l = 1} ^ {L} \\left(g _ {a} ^ {l} * R _ {a} ^ {l}\\right) + g _ {z} * R _ {z}\\right)\\right), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            125,
+            332,
+            468,
+            352
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\hat{G} \\in \\mathbb{R}^{N \\times N \\times |\\mathcal{C}_p|}$  denotes predicted relation graph and  $\\mathrm{MLP}_{\\mathrm{rel}}$  is a three-layer perceptron with ReLU activation. Note that we use sigmoid function  $\\sigma$  so that multiple relationships can exist between objects.",
+        "bbox": [
+            76,
+            363,
+            468,
+            425
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3. Learning and Inference",
+        "text_level": 1,
+        "bbox": [
+            76,
+            434,
+            294,
+            450
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To train EGTR, we perform multi-task learning. In addition to object detection and relation extraction, we devise connectivity prediction, an auxiliary task for relation extraction. The overall loss for the framework is as follows:",
+        "bbox": [
+            76,
+            457,
+            468,
+            516
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {o d}} + \\lambda_ {\\mathrm {r e l}} \\mathcal {L} _ {\\mathrm {r e l}} + \\lambda_ {\\mathrm {c o n}} \\mathcal {L} _ {\\mathrm {c o n}}, \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            168,
+            530,
+            468,
+            546
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{L}_{\\mathrm{od}}$  is object detection loss in Eq. (2).  $\\mathcal{L}_{\\mathrm{rel}}$  and  $\\mathcal{L}_{\\mathrm{con}}$  are loss functions for the relation extraction and connectivity prediction, respectively. Since explicit object detection loss is used, EGTR has the capability to detect all nodes in the scene graph. The details of the loss for each task are provided below.",
+        "bbox": [
+            76,
+            556,
+            468,
+            647
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3.1 Relation Extraction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            665,
+            269,
+            679
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use binary cross-entropy loss for the relation extraction. To match predicted graph  $\\hat{G} \\in \\mathbb{R}^{N \\times N \\times |\\mathcal{C}_p|}$  and ground truth triplet set  $\\mathcal{E}$ , we first encode  $\\mathcal{E}$  as a one-hot ground truth graph  $G \\in \\mathbb{R}^{N \\times N \\times |\\mathcal{C}_p|}$  by padding the regions that do not correspond to relations between the ground truth objects as zero. Then, we permute indices of the predicted graph using the permutation found in the object detection. From the permuted graph  $\\hat{G}'$ , the loss for the relation extraction is calculated as  $\\mathcal{L}_{\\mathrm{rel}} = \\mathcal{L}_r(\\hat{G}', G)$ .",
+        "bbox": [
+            76,
+            689,
+            468,
+            824
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "However, since the size of the graph is proportional to the square of  $N$ , the sparsity of the ground truth graph  $G$  is too severe. For instance, the density is only  $10^{-14}$  when  $N$  is set to 200 for the Visual Genome [12] validation dataset. Therefore, we divide  $G$  into three regions: (1) GT region,",
+        "bbox": [
+            76,
+            825,
+            468,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/8b4b9a348b44f666252103d9ddd29526093a28ee47d8308828587aeab359d27e.jpg",
+        "image_caption": [],
+        "image_footnote": [
+            "girl - on - horse",
+            "jirl - riding"
+        ],
+        "bbox": [
+            519,
+            94,
+            676,
+            220
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/150501b135af7d479b690a4d0f111b17fd90658d9f4ed8d284146a7bbe9e14b7.jpg",
+        "image_caption": [
+            "(subject - predicate - object)",
+            "non-matching region",
+            "Figure 4. Example of graph  $G$  region. The target graph has a shape of  $N \\times N \\times |\\mathcal{V}_c|$ . The graph is severely sparse since the number of object queries  $N$  is set large enough to cover objects in the image. We split  $G$  into GT, negative, and non-matching regions."
+        ],
+        "image_footnote": [
+            "girl - wearing - pant"
+        ],
+        "bbox": [
+            676,
+            92,
+            877,
+            257
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "(2) negative region, and (3) non-matching region, as shown in Fig. 4. The GT region indicates the ground truth triplets where the value of  $G$  is 1. The negative region consists of the triplets in which both subjects and objects are composed of the ground truth objects, but no relation exists between them. The non-matching region represents the zero-padded region. It is paired with a region consisting of the object candidates that do not match the ground truth objects and match  $\\phi$  in Eq. (2). For each region, we apply different techniques to effectively train the relation extraction with object detection as subsequently described.",
+        "bbox": [
+            498,
+            364,
+            892,
+            532
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Adaptive Smoothing. We propose a novel adaptive smoothing for the GT region. For  $G_{ijk}$  belonging to the GT region, the model is trained to predict the  $k$ -th predicate category between subject entity  $v_i$  and object entity  $v_j$ . However, since  $\\hat{v}_i'$  and  $\\hat{v}_j'$ , which match  $v_i$  and  $v_j$  respectively, do not have enough representations about the corresponding ground truth object at the beginning of the training, it may not be appropriate to predict the probability of the predicate as 1. Moreover, even when the object detection performance is reasonably assured, the detection performance may still vary for individual object candidates. Therefore, we reflect the detection performance of each object candidate on the relation label via adaptive smoothing.",
+        "bbox": [
+            498,
+            537,
+            890,
+            733
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We first measure the uncertainty of each object candidate with the corresponding bipartite matching cost. For object candidate  $\\hat{v}_i'$ , we define the uncertainty as follows:",
+        "bbox": [
+            498,
+            733,
+            890,
+            777
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nu _ {i} = \\sigma \\left(\\operatorname {c o s t} _ {i} - \\operatorname {c o s t} _ {\\min } + \\sigma^ {- 1} (\\alpha)\\right), \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            785,
+            890,
+            803
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathrm{cost}_i$  denotes the matching cost and  $\\mathrm{cost}_{\\mathrm{min}}$  indicates the matching cost when  $\\hat{v}_i^\\prime$  perfectly matches  $v_{i}$  .  $\\alpha$  is a nonnegative hyperparameter representing the minimum uncertainty. We set the value of  $G_{ijk}$  to  $(1 - u_{i})(1 - u_{j})$  taking the uncertainty into account. By using the relation label adjusted by the uncertainty, the multi-task learning of object",
+        "bbox": [
+            498,
+            810,
+            892,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "24233",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "detection and relation extraction is dynamically regulated according to the quality of the detected objects.",
+        "bbox": [
+            76,
+            90,
+            468,
+            121
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Negative and Non-matching Sampling. Rather than employing all negatives, we sample them from the negative region. Inspired by the hard negative mining introduced by Liu et al. [21], we sort all negatives based on the predicted relation score  $\\hat{G}_{ijk}^{\\prime}$  and choose the top  $k_{\\mathrm{neg}} \\times |\\mathcal{E}|$  most challenging negatives. Similarly, we extract  $k_{\\mathrm{non}} \\times |\\mathcal{E}|$  hard samples from the non-matching region. As the non-matching region typically encompasses a substantial part of the graph  $G$ , this approach notably diminishes sparsity.",
+        "bbox": [
+            76,
+            126,
+            467,
+            262
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.3.2 Connectivity Prediction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            281,
+            294,
+            295
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Influenced by Graph-RCNN [40] that predicted a relatedness to prune object pairs, we propose a connectivity prediction that predicts whether at least one edge between two object nodes exists for the relation extraction. We get a connectivity graph  $\\hat{E} \\in \\mathbb{R}^{N \\times N \\times 1}$  in a similar way to get the relation graph in Eq. (6) with the same relation source representations. Instead of  $\\mathrm{MLP}_{\\mathrm{rel}}$  for multi-label prediction, we use another  $\\mathrm{MLP}_{\\mathrm{con}}$  for binary classification. We calculate the binary cross entropy loss from the permutated connectivity graph as follows:  $\\mathcal{L}_{\\mathrm{con}} = \\mathcal{L}_r(\\hat{E}', E)$ .",
+        "bbox": [
+            75,
+            305,
+            467,
+            455
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.3.3 Inference",
+        "text_level": 1,
+        "bbox": [
+            76,
+            474,
+            196,
+            488
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For model inference, we get triplet scores by multiplying predicate score  $\\hat{G}_{ijk}$  by the corresponding class scores of  $\\hat{v}_i^c$  and  $\\hat{v}_j^c$ . Note that we set  $\\hat{G}_{iik}$  to 0 to prevent self-connections in which the subject and object are the same entity. In addition, we enhance the triplet scores by utilizing the connectivity score  $\\hat{E}_{ijk}$ . By multiplying the connectivity score, we can effectively filter out the triplets that are not likely to have a relation between the subject and object.",
+        "bbox": [
+            76,
+            500,
+            467,
+            621
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            636,
+            207,
+            652
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Datasets and Evaluation Settings",
+        "text_level": 1,
+        "bbox": [
+            76,
+            660,
+            364,
+            676
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We conduct experiments on two SGG datasets. We describe the datasets and evaluation settings for each dataset. Detailed settings for each dataset are presented in the supplementary materials.",
+        "bbox": [
+            76,
+            684,
+            467,
+            744
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Visual Genome. Visual Genome [12] is the most representative SGG dataset, consisting of 57K training images, 5K validation images, and 26K test images. We follow the popular Visual Genome split, retaining the most frequent 150 object categories and 50 relation categories. We adopt Scene Graph Detection (SGDet) evaluation settings and report Recall@k (R@k) which is class agnostic and mean Recall@k (mR@k) that aggregates the recalls for each predicate category. Following Motifs [42], these metrics are measures with graph constraint, which means each",
+        "bbox": [
+            75,
+            750,
+            467,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "object pair can have a single predicate category. Since they are only related to triplet detection of the scene graph, we report AP50 that evaluates the detection performance of all objects appearing in the scene. Furthermore, we report the number of model parameters and Frames Per Second (FPS) to measure efficiency.",
+        "bbox": [
+            496,
+            90,
+            890,
+            181
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Open Image V6. Open Image V6 [13] is also widely used dataset, comprising of 126K images for training, 2K for validation, and 5K for test set. It includes 601 object categories and 30 relation categories. For evaluation, we adopt both recall and weighted mean AP (wmAP) following standard settings. For recall evaluation, micro-R@50 is adopted. The WMAP is evaluated with two settings:  $\\mathrm{wmAP}_{\\mathrm{rel}}$  for predicting boxes of subject entity and object entity separately and  $\\mathrm{wmAP}_{\\mathrm{phr}}$  for predicting a union box of them. The final score is calculated by  $0.2 \\times \\text{micro-R} @50 + 0.4 \\times \\mathrm{wmAP}_{\\mathrm{rel}} + 0.4 \\times \\mathrm{wmAP}_{\\mathrm{phr}}$ .",
+        "bbox": [
+            496,
+            186,
+            890,
+            354
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            500,
+            366,
+            715,
+            381
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We employ Deformable DETR [46] that improves the convergence speed of the DETR with ResNet-50 [6] as a backbone. It is worth noting that our approach can be extended to any object detector that incorporates self-attention mechanisms between object features, including models like DETR [1], Sparse R-CNN [32], and others. We follow the configurations of the original Deformable DETR, except that we use only 200 object queries. To speed up convergence, we first train the object detector with the target dataset and subsequently train the SGG task using the pretrained object detector, similar to prior work [17]. To calculate the overall loss in Equation 7,  $\\lambda_{\\mathrm{rel}}$  is set to 15.  $\\lambda_{\\mathrm{con}}$  is set to 30 and 90 for the Visual Genome and Open Image V6, respectively. We set  $\\alpha$  as  $10^{-14}$  for adaptive smoothing and both  $k_{\\mathrm{neg}}$  and  $k_{\\mathrm{non}}$  are set to 80.",
+        "bbox": [
+            496,
+            388,
+            890,
+            617
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Results",
+        "text_level": 1,
+        "bbox": [
+            500,
+            628,
+            591,
+            642
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We present quantitative results and perform a comparative analysis of our proposed framework with representative SGG models. Additional experimental results are included in the supplementary materials.",
+        "bbox": [
+            496,
+            652,
+            890,
+            713
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Visual Genome. Visual Genome results are shown in Tab. 1. Our proposed method demonstrates competitive performance with the current one-stage SGG models that have 1.5 to 6.5 times larger parameters with the fastest inference speed. In particular, our method achieves the highest object detection performance and shows comparable performance in the triplet detection compared to SSR-CNN [35], the state-of-the-art method. The results demonstrate that our method can generate scene graphs in an efficient and effective way by exploiting the by-products of the object detector. By applying logit adjustment [35], a technique for predicting tail predicate classes well, our",
+        "bbox": [
+            496,
+            719,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "24234",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/a16fe17cf96c9225171bc6281cd360eb2c9e469b302579803d440aef71ae7091.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Model</td><td># params (M)</td><td>FPS</td><td>AP50</td><td>R@20</td><td>R@50</td><td>R@100</td><td>mR@20</td><td>mR@50</td><td>mR@100</td></tr><tr><td rowspan=\"8\">two-stage</td><td>IMP (EBM) [31, 39]</td><td>322.2</td><td>2.0</td><td>28.1</td><td>18.1</td><td>25.9</td><td>31.2</td><td>2.8</td><td>4.2</td><td>5.4</td></tr><tr><td>VTransE [44]</td><td>312.3</td><td>3.5</td><td>-</td><td>24.5</td><td>31.3</td><td>35.5</td><td>5.1</td><td>6.8</td><td>8.0</td></tr><tr><td>Motifs [42]</td><td>369.9</td><td>1.9</td><td>28.1</td><td>25.1</td><td>32.1</td><td>36.9</td><td>4.1</td><td>5.5</td><td>6.8</td></tr><tr><td>VCTree [33]</td><td>361.5</td><td>0.8</td><td>28.1</td><td>24.8</td><td>31.8</td><td>36.1</td><td>4.9</td><td>6.6</td><td>7.7</td></tr><tr><td>VCTree (TDE) [33, 34]</td><td>361.3</td><td>0.8</td><td>28.1</td><td>14.0</td><td>19.4</td><td>23.2</td><td>6.9</td><td>9.3</td><td>11.1</td></tr><tr><td>VCTree (EBM) [31, 33]</td><td>372.5</td><td>-</td><td>28.1</td><td>24.2</td><td>31.4</td><td>35.9</td><td>5.7</td><td>7.7</td><td>9.1</td></tr><tr><td>GPS-Net [19]</td><td>-</td><td>-</td><td>-</td><td>-</td><td>31.1</td><td>35.9</td><td>-</td><td>6.7</td><td>8.6</td></tr><tr><td>BGNN [16]</td><td>341.9</td><td>1.7</td><td>29.0</td><td>23.3</td><td>31.0</td><td>35.8</td><td>7.5</td><td>10.7</td><td>12.6</td></tr><tr><td rowspan=\"11\">one-stage</td><td>FCSGG [20]</td><td>87.1</td><td>6.0</td><td>28.5</td><td>16.1</td><td>21.3</td><td>25.1</td><td>2.7</td><td>3.6</td><td>4.2</td></tr><tr><td>RelTR [7]</td><td>63.7</td><td>13.4</td><td>26.4</td><td>21.2</td><td>27.5</td><td>-</td><td>6.8</td><td>10.8</td><td>-</td></tr><tr><td>SGTR [17]</td><td>117.1</td><td>6.2</td><td>25.4</td><td>-</td><td>24.6</td><td>28.4</td><td>-</td><td>12.0</td><td>15.2</td></tr><tr><td>Relationformer [29]</td><td>92.9</td><td>8.5</td><td>26.3</td><td>22.2</td><td>28.4</td><td>31.3</td><td>4.6</td><td>9.3</td><td>10.7</td></tr><tr><td>Iterative SGG [9]</td><td>93.5</td><td>6.0</td><td>27.7†</td><td>-</td><td>29.7</td><td>32.1</td><td>-</td><td>8.0</td><td>8.8</td></tr><tr><td>SSR-CNN [35]</td><td>274.3</td><td>4.0</td><td>23.8†</td><td>25.8</td><td>32.7</td><td>36.9</td><td>6.1</td><td>8.4</td><td>10.0</td></tr><tr><td>SSR-CNN [35] LA,τ=0.3</td><td>274.3</td><td>4.0</td><td>23.8†</td><td>18.4</td><td>23.3</td><td>26.5</td><td>13.5</td><td>17.9</td><td>21.4</td></tr><tr><td>EGTR (Ours)</td><td>42.5</td><td>14.7</td><td>30.8</td><td>23.5</td><td>30.2</td><td>34.3</td><td>5.5</td><td>7.9</td><td>10.1</td></tr><tr><td>EGTR (Ours) LA,τ=0.7</td><td>42.5</td><td>14.7</td><td>30.8</td><td>15.7</td><td>18.7</td><td>20.5</td><td>12.1</td><td>17.8</td><td>21.7</td></tr><tr><td>EGTR (Ours) LA,τ=0.5</td><td>42.5</td><td>14.7</td><td>30.8</td><td>19.7</td><td>24.2</td><td>26.7</td><td>11.0</td><td>17.1</td><td>21.4</td></tr><tr><td>EGTR (Ours) LA,τ=0.3</td><td>42.5</td><td>14.7</td><td>30.8</td><td>22.4</td><td>28.2</td><td>31.7</td><td>8.8</td><td>14.0</td><td>18.3</td></tr></table>",
+        "bbox": [
+            76,
+            88,
+            893,
+            356
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/7064226fa7dbe87e3156582cfc815f2a9dc9c4a9cd4fd7effeb5b3b0549607d9.jpg",
+        "table_caption": [
+            "Table 1. Graph-Constraint results on the test set of Visual Genome dataset. We report the results of representative two-stage SGG models based on the Faster R-CNN [27] object detector with a ResNeXt-101-FPN [18, 38] backbone and concurrent one-stage SGG models. Among the one-stage models, the best is highlighted in bold, and the second-best is indicated with underlining. LA denotes logit adjustment proposed in SSR-CNN [35].  $\\dagger$  represents that we concatenate the predicted subjects and objects and then apply non-maximum-suppression with a threshold of 0.3. Italic indicates the evaluation of metrics using publicly available model checkpoints. FPS is measured with a single V100 for images resized to a minimum of 600 for the shortest side and a maximum of 1000 for the longest side."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>score</td><td>micro-R@50</td><td>wmAPrel</td><td>wmAPphr</td></tr><tr><td>Motifs [42]</td><td>38.9</td><td>71.6</td><td>29.9</td><td>31.6</td></tr><tr><td>VCTree [33]</td><td>40.2</td><td>74.1</td><td>34.2</td><td>33.1</td></tr><tr><td>GPS-Net [19]</td><td>41.7</td><td>74.8</td><td>32.9</td><td>34.0</td></tr><tr><td>BGNN [16]</td><td>42.1</td><td>75.0</td><td>33.5</td><td>34.2</td></tr><tr><td>RelTR [7]</td><td>43.0</td><td>71.7</td><td>34.2</td><td>37.5</td></tr><tr><td>SGTR [17]</td><td>42.3</td><td>59.9</td><td>37.0</td><td>38.7</td></tr><tr><td>SSR-CNN [35]</td><td>49.4</td><td>76.7</td><td>41.5</td><td>43.6</td></tr><tr><td>EGTR (Ours)</td><td>48.6</td><td>75.0</td><td>42.0</td><td>41.9</td></tr></table>",
+        "bbox": [
+            76,
+            457,
+            468,
+            595
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Results on test set of Open Image V6. The score is a weighted sum of micro-R@50,  $\\mathrm{wmAP}_{\\mathrm{rel}}$ , and  $\\mathrm{wmAP}_{\\mathrm{phr}}$ .",
+        "bbox": [
+            76,
+            602,
+            468,
+            632
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "method performs favorably in the trade-off between  $\\mathrm{R}@\\mathbf{k}$  and  $\\mathrm{mR}@\\mathbf{k}$ , showing superiority over existing state-of-the-art models. Note that for triplet detection models [9, 35] without an explicit object detector, AP50 is measured by applying non-maximum-suppression (NMS) to the union of predicted subjects and objects set. AP50 performance is discussed further in the section 4.5.",
+        "bbox": [
+            75,
+            643,
+            468,
+            750
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Open Image V6. As depicted in Tab. 2, the experiments conducted on the Open Image V6 dataset also demonstrate competitive performance, underscoring the effectiveness and robustness of our method across different datasets.",
+        "bbox": [
+            75,
+            753,
+            468,
+            816
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4. Ablation Studies",
+        "text_level": 1,
+        "bbox": [
+            76,
+            825,
+            241,
+            842
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We analyze the effects of our model components. All of the ablation studies are done with the Visual Genome dataset.",
+        "bbox": [
+            76,
+            849,
+            468,
+            878
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Relation Sources. In Tab. 3, we investigate the source of",
+        "bbox": [
+            76,
+            885,
+            468,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "relations used for the relation extractor. To verify the benefits of using attention queries and keys, we change  $Q^l$  and  $K^l$  in Eq. (3) to  $Z^l$ , the hidden states of each layer. Remarkably, when simply using the hidden states of all layers, the triplet detection performance decreases. Surprisingly, the performance is lower than that of using only  $R_z$  from the last layer. On the other hand, when only attention by-products are used as a relation source, the performance is similar to that of using only hidden states of the last layer, the most contextualized representations for detecting objects. Incorporating both attention by-products and final hidden states yields improved performance by amalgamating varied information. The results underscore that the attention by-products of the object detectors contain rich information for relation extraction.",
+        "bbox": [
+            496,
+            460,
+            890,
+            686
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Training Techniques. In Tab. 4, we ablate proposed techniques. Without any techniques, our model framework shows low performance. When each technique is applied, it substantially improves both R@50 and mR@50, and the best performance is achieved when all techniques are combined. The results demonstrate that our proposed techniques help to train the model framework effectively. Additional experimental results are provided in the supplementary materials.",
+        "bbox": [
+            496,
+            695,
+            890,
+            832
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Sampling Methodology. In Tab. 5, we investigate the impact of hard sampling methods for the negative and non-matching regions. Hard negatives and hard non-matchings show a trade-off that improves the performance of R@50",
+        "bbox": [
+            496,
+            839,
+            893,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "24235",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/130b3fcdca977aff0793d48b4bd73ce8c49708db37c5250079679f4f3392cdfc.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Rl^l a source</td><td>Rl^l a</td><td>Rz</td><td>R@50</td><td>mR@50</td></tr><tr><td>Ql &amp; Kl</td><td>✓</td><td>✓</td><td>30.2</td><td>7.9</td></tr><tr><td>Zl</td><td>✓</td><td>✓</td><td>29.6</td><td>7.4</td></tr><tr><td>-</td><td></td><td>✓</td><td>29.9</td><td>7.6</td></tr><tr><td>Ql &amp; Kl</td><td>✓</td><td></td><td>29.8</td><td>7.7</td></tr></table>",
+        "bbox": [
+            78,
+            88,
+            470,
+            167
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/9da1d5b0b3751bc9139a7ea1b3f68db2ac24a2b648ee5e40c17b42a22563ab5c.jpg",
+        "table_caption": [
+            "Table 3. Ablation study on relation source."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>adaptive smoothing</td><td>\\( \\mathcal{L}_{\\text{con}} \\)</td><td>sampling</td><td>R@50</td><td>mR@50</td></tr><tr><td></td><td></td><td></td><td>26.6</td><td>5.3</td></tr><tr><td>✓</td><td></td><td></td><td>28.3</td><td>6.5</td></tr><tr><td></td><td>✓</td><td></td><td>29.6</td><td>7.0</td></tr><tr><td></td><td></td><td>✓</td><td>28.9</td><td>7.1</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>30.2</td><td>7.9</td></tr></table>",
+        "bbox": [
+            78,
+            203,
+            468,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/ca8a4a2118b2bf6c879fa9b0eeed7043ab378d3428d9275d525d53208d7986e0.jpg",
+        "table_caption": [
+            "Table 4. Ablation study on proposed techniques."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>hard negative</td><td>hard non-matching</td><td>R@50</td><td>mR@50</td></tr><tr><td>✓</td><td>✓</td><td>30.2</td><td>7.9</td></tr><tr><td>✓</td><td></td><td>30.0</td><td>7.1</td></tr><tr><td></td><td>✓</td><td>29.6</td><td>7.7</td></tr></table>",
+        "bbox": [
+            78,
+            328,
+            468,
+            393
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "and mR@50, respectively, and additional performance improvements are achieved for both measures when they are used together. For hard negatives, we speculate that R@50 performance is improved by selecting negatives for the tail predicate class, which is likely to exist in reality but is not in the annotation. For hard non-matchings, mR@50 performance seems to have improved by choosing object candidates that are akin to the ground truth objects but unmatched with them. This encourages the model to predict diverse predicate classes by preventing duplicate object pairs, which are likely to predict head predicate classes, from being predicted.",
+        "bbox": [
+            75,
+            428,
+            468,
+            609
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5. Discussion",
+        "text_level": 1,
+        "bbox": [
+            76,
+            619,
+            194,
+            633
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Regarding the highest AP50 score, we posit that it is because our proposed method focuses on not only objects with relations but also objects without any relations. To validate this, we calculate AP50 for two subsets of the ground-truth objects: those with relations and those without relations. We compare our method with two triplet detection-based models. Tab. 6 illustrates that our method achieves significantly high  $\\mathrm{AP50}_{\\mathrm{no - rel}}$ . Considering that triplet detection performance is related to  $\\mathrm{AP50}_{\\mathrm{rel}}$ , while triplet detection-based models only focus on objects with relations to detect triplets well, our method is capable of extracting the complete scene graph, including objects without relations.",
+        "bbox": [
+            75,
+            643,
+            468,
+            824
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "For further analysis, we visualize bounding boxes for subjects and objects from the top 20 predictions in Fig. 5. For SSR-CNN [35], which directly predicts triplets, we observe a significant number of overlapping bounding boxes. On the other hand, our proposed model leverages predic",
+        "bbox": [
+            75,
+            824,
+            468,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/e58fc5b81d9148d9476052a5decff82b2473d1214d9d6f90d01e0685faefade3.jpg",
+        "table_caption": [
+            "Table 5. Ablation study on sampling methods."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>AP50</td><td>\\( AP50_{\\text{rel}} \\)</td><td>\\( AP50_{\\text{no-rel}} \\)</td></tr><tr><td>Iterative SGG [9]†</td><td>27.7</td><td>24.3</td><td>7.8</td></tr><tr><td>SSR-CNN [35]†</td><td>23.8</td><td>20.2</td><td>7.4</td></tr><tr><td>EGTR (Ours)</td><td>30.8</td><td>24.3</td><td>10.7</td></tr></table>",
+        "bbox": [
+            500,
+            88,
+            890,
+            152
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6. AP50 for two subsets of objects.  $\\mathrm{AP50}_{\\mathrm{rel}}$  and  $\\mathrm{AP50}_{\\mathrm{no - rel}}$  denote scores evaluated only on objects having at least one edge and no edge, respectively. Note that each measure operates on its unique scale since the ground-truth objects used for each measure differ.  $\\dagger$  indicates that additional NMS is applied to the union of subjects and objects set.",
+        "bbox": [
+            498,
+            160,
+            893,
+            244
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/24bd9ff20bfc213d9d6ec0c5211af0c429e35efa9df27fb0358e28fccd39745f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            257,
+            694,
+            358
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/18ae33dd39cb2090a7c41cafda42b6943398cbcb2bce6166315195b836fe3f0f.jpg",
+        "image_caption": [
+            "(a) SSR-CNN [35]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            358,
+            694,
+            455
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/46306f91c2dbb914d9f6359cae5848fe4882b7587c39fc49c4b75c459bcd9dfb.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            694,
+            258,
+            885,
+            358
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/bcd320527c7035d3e2f40585be25959ccfdf58231515632ecd35636a8f686e10.jpg",
+        "image_caption": [
+            "(b) EGTR (Ours)",
+            "Figure 5. Comparison of detected subjects and objects. We select the top 20 predictions of SSR-CNN and EGTR and visualize bounding boxes of detected subjects and objects."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            694,
+            358,
+            885,
+            455
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "tions from an explicit object detector, reducing redundant predictions. Furthermore, our method detects various subjects and objects that appear in the scene rather than focusing on objects that are more likely to have relations.",
+        "bbox": [
+            498,
+            532,
+            890,
+            592
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            606,
+            617,
+            621
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this study, we propose the lightweight one-stage scene graph generator EGTR. We significantly reduce the model complexity by harnessing the relationships among the object queries learned from the self-attention layers of the object decoder. Furthermore, we devise a novel adaptive smoothing technique that helps multi-task learning of object detection and relation extraction by adjusting the relation label according to the object detection performance. As an auxiliary task of relation extraction, connectivity prediction contributes to the effective learning of EGTR. We conduct extensive experiments and ablation studies, and the results demonstrate that EGTR achieves the highest object detection performance and competitive triplet detection capabilities with the fastest inference speed.",
+        "bbox": [
+            496,
+            631,
+            890,
+            843
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements. We thank Taeho Kil, Geewook Kim, and the anonymous reviewers for their insightful comments and suggestions.",
+        "bbox": [
+            498,
+            847,
+            890,
+            893
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "24236",
+        "bbox": [
+            478,
+            945,
+            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] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 2, 3, 4, 6",
+            "[2] Zhe Chen, Shaoli Huang, and Dacheng Tao. Context refinement for object detection. In ECCV, pages 71-86, 2018. 2",
+            "[3] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE TPAMI, 2023. 1, 3",
+            "[4] Naina Dhingra, Florian Ritter, and Andreas Kunz. Bgt-net: Bidirectional gru transformer network for scene graph generation. In CVPR, pages 2150-2159, 2021. 2",
+            "[5] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In ICMLC, pages 225-229, 2018. 1",
+            "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6",
+            "[7] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In CVPR, pages 3588-3597, 2018. 2, 7",
+            "[8] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, pages 3668-3678, 2015. 1",
+            "[9] Siddhesh Khandelwal and Leonid Sigal. Iterative scene graph generation. In NeurIPS, 2022. 1, 3, 7, 8",
+            "[10] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In CVPR, pages 6271-6280, 2019. 1",
+            "[11] Rajat Koner, Suprosanna Shit, and Volker Tresp. Relation transformer network. ECCV, pages 422-439, 2022. 2",
+            "[12] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 2, 5, 6",
+            "[13] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 2, 6",
+            "[14] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In ICCV, pages 10313-10322, 2019. 1",
+            "[15] Lin Li, Guikun Chen, Jun Xiao, Yi Yang, Chunping Wang, and Long Chen. Compositional feature augmentation for unbiased scene graph generation. In ICCV, pages 21685-21695, 2023. 2",
+            "[16] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In CVPR, pages 11109-11119, 2021. 1, 2, 7"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[17] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In CVPR, pages 19486-19496, 2022. 1, 3, 6, 7",
+            "[18] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 7",
+            "[19] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In CVPR, pages 3746-3753, 2020. 1, 2, 7",
+            "[20] Hengyue Liu, Ning Yan, Masood Mortazavi, and Bir Bhanu. Fully convolutional scene graph generation. In CVPR, pages 11546-11556, 2021. 1, 3, 7",
+            "[21] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 6",
+            "[22] Yong Liu, Ruiping Wang, Shiguang Shan, and Xilin Chen. Structure inference net: Object detection using scene-level context and instance-level relationships. In CVPR, pages 6985-6994, 2018. 2",
+            "[23] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In ECCV, pages 852–869, 2016. 2",
+            "[24] Yichao Lu, Himanshu Rai, Jason Chang, Boris Knyazev, Guangwei Yu, Shashank Shekhar, Graham W Taylor, and Maksims Volkovs. Context-aware scene graph generation with seq2seq transformers. In ICCV, pages 15931-15941, 2021.",
+            "[25] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In ICCV, pages 13296-13307, 2023. 2",
+            "[26] Alejandro Newell and Jia Deng. Pixels to graphs by associative embedding. NeurIPS, 30, 2017. 1, 3",
+            "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 2, 7",
+            "[28] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80. Association for Computational Linguistics, 2015. 1",
+            "[29] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In ECCV, pages 422-439, 2022. 1, 3, 7",
+            "[30] Gopika Sudhakaran, Devendra Singh Dhami, Kristian Kersting, and Stefan Roth. Vision relation transformer for unbiased scene graph generation. In ICCV, pages 21882-21893, 2023. 2",
+            "[31] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In CVPR, pages 13936-13945, 2021. 7"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "24237",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[32] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, and Ping Luo. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 14454-14463, 2021. 2, 3, 6",
+            "[33] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In CVPR, pages 6619-6628, 2019. 1, 2, 7",
+            "[34] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In CVPR, pages 3716-3725, 2020. 7",
+            "[35] Yao Teng and Limin Wang. Structured sparse r-cnn for direct scene graph generation. In CVPR, pages 19437-19446, 2022. 1, 3, 6, 7, 8",
+            "[36] 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. 2, 3",
+            "[37] Sanghyun Woo, Dahun Kim, Donghyeon Cho, and In So Kweon. Linknet: Relational embedding for scene graph. NeurIPS, 31, 2018. 2",
+            "[38] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 1492-1500, 2017. 7",
+            "[39] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In CVPR, pages 5410–5419, 2017. 1, 2, 7",
+            "[40] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In ECCV, pages 670-685, 2018. 6",
+            "[41] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In CVPR, pages 10685–10694, 2019. 1",
+            "[42] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In CVPR, pages 5831-5840, 2018. 1, 2, 6, 7",
+            "[43] Cheng Zhang, Wei-Lun Chao, and Dong Xuan. An empirical study on leveraging scene graphs for visual question answering. In BMVC, 2019. 1",
+            "[44] Hanwang Zhang, Zawlin Kyaw, Shih-Fu Chang, and Tat-Seng Chua. Visual translation embedding network for visual relation detection. In CVPR, page 3107-3115, 2017. 7",
+            "[45] Chaofan Zheng, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, Jingkuan Song, and Lianli Gao. Learning to generate scene graph from head to tail. In ICME, pages 1-6. IEEE, 2022. 2",
+            "[46] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 3, 6"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            787
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "24238",
+        "bbox": [
+            478,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/174c2199-8aa8-4e45-8d2e-205eb2e2e099_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48415d99182105a59d0b4f6135ceef5c08353ce230019ec47ed4a7853f3f29d7
+size 3379274

2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/full.md

@@ -0,0 +1,340 @@
+# EGTR: Extracting Graph from Transformer for Scene Graph Generation
+
+Jinbae Im<sup>1</sup> JeongYeon Nam<sup>1</sup> Nokyung Park<sup>1,2,3*</sup> Hyungmin Lee<sup>2</sup> Seunghyun Park<sup>1</sup>  
+<sup>1</sup>NAVER Cloud AI <sup>2</sup>NAVER <sup>3</sup>Korea University  
+{jinbae.im, jy.nam, nokyung.park99, hyungmin.lee, seung.park}@navercorp.com
+
+# Abstract
+
+Scene Graph Generation (SGG) is a challenging task of detecting objects and predicting relationships between objects. After DETR was developed, one-stage SGG models based on a one-stage object detector have been actively studied. However, complex modeling is used to predict the relationship between objects, and the inherent relationship between object queries learned in the multi-head self-attention of the object detector has been neglected. We propose a lightweight one-stage SGG model that extracts the relation graph from the various relationships learned in the multi-head self-attention layers of the DETR decoder. By fully utilizing the self-attention by-products, the relation graph can be extracted effectively with a shallow relation extraction head. Considering the dependency of the relation extraction task on the object detection task, we propose a novel relation smoothing technique that adjusts the relation label adaptively according to the quality of the detected objects. By the relation smoothing, the model is trained according to the continuous curriculum that focuses on object detection task at the beginning of training and performs multi-task learning as the object detection performance gradually improves. Furthermore, we propose a connectivity prediction task that predicts whether a relation exists between object pairs as an auxiliary task of the relation extraction. We demonstrate the effectiveness and efficiency of our method for the Visual Genome and Open Image V6 datasets. Our code is publicly available at https://github.com/naver-ai/egtr.
+
+# 1. Introduction
+
+Scene Graph Generation (SGG) [8] aims to generate a scene graph that represents objects as nodes and relationships between objects as edges from an image as shown in Fig. 1b. SGG is a challenging task as it is required to go beyond simply detecting objects but predicting the relationships between them based on a comprehensive understanding of the
+
+![](images/da086494a3493dff7e071faf8fea070928900aae618d983c549510da6ad11f8c.jpg)  
+(a) Visual Genome example
+
+![](images/9fcc74a969bc741c22cf5e31e09c3ed6a0b106a111765290803f5a92ea5dbd50.jpg)  
+(b) Scene Graph
+
+![](images/d3d669486085177c87b65b065112e8bb0b5abe8f60a9d5e1549828477fc1cbc7.jpg)  
+(c) Attention Graph  
+Figure 1. Motivation. SGG task aims to predict scene graph (Fig. 1b) with objects as nodes and relations as edges. We draw a plausible attention graph (Fig. 1c) simply by connecting objects with high attention weights to edges from the self-attention layers of the pre-trained DETR. It shows the potential for the self-attention from the object detector to contain rich information that aids in predicting the relations of the scene graph.
+
+scene. Since a scene graph provides structural information of the image, it can be used for various vision tasks that require a higher level of understanding and reasoning about images such as image captioning [5, 10, 41], image retrieval [8, 28], and visual question answering [14, 43].
+
+Most previous studies [16, 19, 33, 39, 42] took two-stage SGG approaches that detect objects first and then predict their relations. However, these approaches incurred high computational costs and risks of error propagation from the object detection. To address the drawbacks of the two-stage approaches, one-stage SGG models that perform object detection and relation prediction at once [3, 9, 17, 20, 26, 29, 35] have been studied recently, leveraging one-stage object detectors such as DETR [1]. Since edges in a scene graph can be represented as subject-predicate-object triplets, many studies embraced triplet-based approaches as shown in Figs. 2a and 2b. However, object-triplet detection models [3, 17] required sophisticated triplet detectors to obtain necessary information for triplet queries from the object detector. Moreover, triplet detection models [9, 35] lacked the ability to detect objects without relation such as "tree" in Fig. 1b, by focusing solely on the triplet detection without an object detector. Considering that objects without relations account for more than
+
+![](images/eeef0756f110b56051eef1ee222a4df8ecbb04eaa0125bb5bc46ca221ac47ed5.jpg)  
+(a) Object-Triplet Detection Models
+
+![](images/17fbf1da620e9dc2c721e805545b7c7f43c92cb1db5cdc6e1494d2fbdc0930f5.jpg)  
+(b) Triplet Detection Models  
+Figure 2. Comparison with existing one-stage SGG models. (a) Object-Triplet Detection Models introduce additional triplet queries and a triplet detector to the object detector. The triplet detector requires additional modules to incorporate information from the object detector into the triplet queries. (b) Triplet Detection Models focus on detecting triplets directly without an object detector. Objects without relations may not be detected. (c) Relation Extraction Models extract relations from the object detector without a separate triplet detector. In particular, ours extracts relations more effectively by utilizing by-products from the self-attention of the object detector.
+
+![](images/1200dc7f83785412b59a2e3649fd9623129020a4a233efc8ba94d41c3b085407.jpg)  
+(c) Relation Extraction Models
+
+$42\%$  in Visual Genome [12] data, their priority lies in detecting a sub-graph rather than the complete scene graph.
+
+To address the shortcomings of the existing one-stage SGG models, we focus on the relationships between objects inherent in the object detector. As shown in Fig. 1a, objects are related to each other. For instance, when a horse appears in a scene, a person is likely to appear in the scene, and clothing such as hats, jackets, and pants often depicts the person's current situation. From this intuition, there has been a long belief [2, 7, 22] that modeling relationship or context between objects would be beneficial to object detection task. Accordingly, recent one-stage object detectors [1, 32, 46] have incorporated self-attention layers [36] to implicitly model the relationships among the object queries. We hypothesize that self-attention between object queries learned in a one-stage object detector might contain valuable information for predicting triplet outputs. In our preliminary investigation, we are able to extract a plausible attention graph by simply connecting two object queries with high attention weights from the pre-trained DETR, as shown in Fig. 1c. It shows the potential that the attention weights between object queries can be interpreted as relations between them.
+
+From the findings, we propose a lightweight one-stage scene graph generator EGTR, which stands for Extracting Graph from TRansformer. We design the model to comprehensively leverage the by-products of the object detector, eliminating the need for a separate triplet detector, as depicted in Fig. 2c. From the multi-head self-attention layers of the object detector, we regard an attention query and key, where their relations are learned in the attention weights, as the subject entity and object entity, respectively. Subsequently, we leverage a shallow classifier to predict relations between them. Due to the abundant information about the relationships among the objects present in the by-products derived from all layers of self-attention, we can effectively extract the scene graph.
+
+Since the relation extraction task is dependent on the object detection task, we speculate that performing relation ex
+
+traction without sufficiently learned representations of the object queries might be harmful. Therefore, we devise a novel adaptive smoothing technique that smoothes the value of the ground truth relation label based on the object detection performance. With the adaptive smoothing, the model is trained with a continuous curriculum that initially focuses on object detection and gradually performs multi-task learning. Furthermore, we propose a connectivity prediction task as an auxiliary task for relation extraction, aiming to predict the existence of any relationship between a subject entity and an object entity. This auxiliary task facilitates the acquisition of representations for the relation extraction.
+
+To verify the effectiveness of the proposed method, we conduct experiments on two representative SGG datasets: Visual Genome [12] and Open Images V6 [13]. By actively utilizing the by-products of the object detector, EGTR shows the best object detection performance and comparable triplet detection performance, with the fewest parameters and the fastest inference speed.
+
+Our main contributions can be summarized as follows:
+
+- We propose EGTR that generates scene graphs efficiently and effectively by utilizing the multi-head self-attention by-products derived from the object detector.  
+- We present adaptive smoothing, enabling effective multitask learning for both object detection and relation extraction. In addition, the proposed connectivity prediction offers clues to the relation extraction.  
+- Our comprehensive experiments show the superiority of the proposed model framework and the effectiveness of the devised training techniques.
+
+# 2. Related Work
+
+SGG models can be categorized into two groups: two-stage models and one-stage models. For two-stage models [4, 11, 15, 16, 19, 23-25, 30, 33, 37, 39, 42, 45], separate object detection model and relation prediction model are trained sequentially. They usually detect  $N$  objects from the off-the-shelf object detector such as Faster R-CNN [27],
+
+and then all possible combinations of the detected objects are fed into the relation prediction model to predict the relations between each object pair. Although they showed high relation extraction performance, they have the inherent limitation that the object detector that helps with relation extraction is trained separately, resulting in a significant increase in model complexity.
+
+As for the one-stage models [3, 9, 17, 20, 26, 29, 35], the object detection and relation prediction are trained in an end-to-end manner. Early studies [20, 26] proposed fully convolutional SGG models and took pixel-based approaches. After DETR [1] brought a huge success as a Transformer [36]-based one-stage object detector, many one-stage SGG studies are based on one-stage object detectors [1, 32, 46]. They efficiently modeled SGG by introducing object queries or triplet queries. We categorize them into three distinct groups: (a) object-triplet detection models, (b) triplet prediction models, and (c) relation extraction models, as shown in Fig. 2.
+
+Object-Triplet Detection Models. Object-triplet detection models are characterized by introducing additional triplet queries and building a triplet predictor on top of the object detector as shown in Fig. 2a. RelTR [3] introduced paired subject queries and object queries, and SGTR [17] introduced compositional queries decoupled into subjects, objects, and predicates. As the introduced queries are initialized without prior cues from the object detector, the triplet predictor requires modules to incorporate information from the outputs of the object detector and modules for triplet queries to exchange information with each other. It led to the triplet predictor having an intricate structure. In contrast, we refrain from introducing additional queries and regard the attention queries and attention keys, whose relationships are learned in the object detector, as subject and object queries, respectively.
+
+Triplet Detection Models. Triplet detection models directly detect triplets using triplet queries without an object detector, as shown in Fig. 2b. Iterative SGG [9] introduced subject, object, and predicate queries and modeled the conditional dependencies among them with separate subject, object, and predicate multi-layer Transformer decoders. Inspired by Sparse R-CNN [32], Structured Sparse R-CNN (SSR-CNN) [35] designed triplet queries consisting of subject box, object box, subject, object, and predicate queries. Since it was difficult to train the model only with sparse triplet annotations, SSR-CNN introduced somewhat intricate training details: it detected object pairs with a Siamese Sparse RCNN model and auxiliary queries and performed additional triplet matching using the detected object pairs as a pseudo label. Although they show excellent triplet detection performance, the model architecture became even more intricate to detect subjects and objects separately since
+
+an explicit object detector is not used. Last but not least, the triplet prediction models focus on detecting a sub-graph composed only of objects with relations, overlooking objects in an image that lack explicit relations.
+
+Relation Extraction Models. Relation extraction models extract a scene graph using a lightweight relation predictor without separate triplet queries or triplet detector. Relationformer [29] added a special "[rln]" token to capture global information in conjunction with object queries. They concatenated the final hidden representations of the object query pairs and the relation token, followed by a shallow fully connected network for relation prediction. In this work, in conjunction with the final hidden representations, we use the inherent relationship information among the object queries learned in the multi-head self-attention layer of the object detector. Furthermore, we propose training techniques to boost multi-task learning, leading to significantly improved performance with architectural simplicity.
+
+# 3. Method
+
+# 3.1. Preliminaries
+
+In this section, we first introduce the formulation of the SGG task and then provide a brief review of the one-stage object detector that serves as the basis for our study.
+
+# 3.1.1 Formulation
+
+Scene graph generation is a task of generating a scene graph  $\mathcal{G} = \{\mathcal{V},\mathcal{E}\}$  from an image, where  $\mathcal{V}$  denotes the node set consisting of objects and  $\mathcal{E}$  denotes the edge set that represents the relation between objects. Each object entity  $v_{i}\in \mathcal{V}$  has an object category label  $v_{i}^{c}$  from a set of object categories  $\mathcal{C}_v$  and box coordinates  $v_{i}^{b}$ . Each relation  $e_j\in \mathcal{E}$  represents the  $j$ -th triplet  $(s_j,p_j,o_j)$ , where subject  $s_j$  and object  $o_j$  indicate related object entities and predicate  $p_j$  has a relation category label  $p_j^c$  from a set of predicate categories  $\mathcal{C}_p$ . Generating  $\mathcal{V}$  and  $\mathcal{E}$  correspond to object detection and relation extraction, respectively.
+
+# 3.1.2 One-stage Object Detector
+
+Our model framework is deeply influenced by DETR [1], a one-stage object detector. In DETR, representations for input images are learned through CNN and Transformer [36] encoder. Transformer decoder enhances object queries by utilizing self- and cross-attention mechanisms. The final object detection heads predict object category labels and box coordinates from the contextualized object queries.
+
+Backbone. CNN backbone generates  $C$ -dimensional feature map  $F \in \mathbb{R}^{C \times H_F \times W_F}$  from an input image  $x_{\mathrm{img}} \in \mathbb{R}^{3 \times H_0 \times W_0}$ . For ResNet-50 backbone,  $H_F$  and  $W_F$  are set to  $H_0 / 32$  and  $W_0 / 32$ , respectively.
+
+![](images/8ecf05244754a50b2be1aece4d420418de000403424f1f62df4891ed67db2042.jpg)  
+Figure 3. The overall architecture of EGTR. We present a novel lightweight relation extractor, EGTR, which fully utilizes the self-attention of the DETR decoder. We extract query and key representations from each self-attention layer and concatenate them pairwise to represent relations between them. Additionally, we leverage the last hidden representation in the same manner. To effectively aggregate information, we apply a gated sum and then predict relation with a shallow relation head.
+
+![](images/051d3620529018711d7ac2cc82099c6a645af5e74a4b87f1aedd5638cdb66c02.jpg)
+
+![](images/cbe093ab64a6516b8c1b4273701a8dbd74801876a7784e78ebbac8263356f574.jpg)
+
+Transformer Encoder. Transformer encoder enhances image representations based on multi-head self-attention. After applying a  $1 \times 1$  convolution to reduce dimension from  $C$  to  $d_{\mathrm{model}}$ , the feature map is flattened to get an input sequence with length  $H_{F}W_{F}$  for the Transformer encoder. Additional positional encodings are used to reflect the spatial information of the feature maps.
+
+Transformer Decoder. Transformer decoder takes  $N$  object queries as input and returns their representations. Each object query detects an individual object, and  $N$  is usually set large enough to cover all objects in the image. Through alternating self-attention and cross-attention layers, object queries learn features of object candidates in the input image. Note that causal attention mask is not applied to the self-attention layer; therefore, fully connected  $N \times N$  attention weights among the  $N$  object queries are learned from the multi-head self-attention as follows:
+
+$$
+A _ {h} ^ {l} = f \left(Q _ {h} ^ {l}, K _ {h} ^ {l}\right) = \operatorname {s o f t m a x} \left(Q _ {h} ^ {l} K _ {h} ^ {T} / \sqrt {d _ {\text {h e a d}}}\right), \tag {1}
+$$
+
+where  $A_h^l \in \mathbb{R}^{N \times N}$  denotes the attention weights from the  $h$ -th head in the  $l$ -th layer.  $Q_h^l \in \mathbb{R}^{N \times d_{\mathrm{head}}}$  and  $K_h^l \in \mathbb{R}^{N \times d_{\mathrm{head}}}$  are attention queries and keys, respectively.
+
+Object Detection Heads. Object detection heads detect  $N$  object candidates  $\{\hat{v}_i\}_{i = 1}^N$  from the last layer representations of the object queries  $Z^{L}$ . For each object query, a linear layer is used to predict the object category label  $\hat{v}_i^c\in \mathbb{R}^{N\times |\mathcal{C}_v|}$ , and three-layer perceptron with ReLU activation is used to obtain the box coordinates  $\hat{v}_i^b\in \mathbb{R}^{N\times 4}$ .
+
+Object Detection Loss. To match  $N$  detected object candidates  $\{\hat{v}_i\}_{i=1}^N$  and  $M$  ground truth objects  $\{v_i\}_{i=1}^M$ , Carion et al. [1] pad the ground truth objects with  $\phi$  (no object) and find the best permutation of predicted objects that minimizes total bipartite matching costs. From the permuted
+
+predictions  $\{\hat{v}_i^{\prime}\}_{i = 1}^{N}$  , the loss is calculated as follows:
+
+$$
+\mathcal {L} _ {\mathrm {o d}} = \Sigma_ {i = 1} ^ {N} \left[ \lambda_ {c} \mathcal {L} _ {c} \left(\hat {v} _ {i} ^ {\prime c}, v _ {i} ^ {c}\right) + \mathbb {1} _ {v _ {i} ^ {c} \neq \phi} \left(\lambda_ {b} \mathcal {L} _ {b} \left(\hat {v} _ {i} ^ {\prime b}, v _ {i} ^ {b}\right)\right) \right], \tag {2}
+$$
+
+where  $\mathcal{L}_c$  and  $\mathcal{L}_b$  denote loss function for category labels and box coordinates, respectively.
+
+# 3.2. EGTR
+
+We propose a novel lightweight relation extractor, EGTR, which exploits the self-attention of DETR decoder, as depicted in Fig. 3. Since the self-attention weights in Eq. (1) contain  $N \times N$  bidirectional relationships among the  $N$  object queries, our relation extractor aims to extract the predicate information from the self-attention weights in the entire  $L$  layers, by considering the attention queries and keys as subjects and objects, respectively.
+
+In order to preserve rich information learned in the self-attention layer, we devise another relation function  $f$  between the attention queries and keys instead of the dot product attention in Eq. (1). As concatenation preserves representations completely, we concatenate the representations on all  $N \times N$  pairs of attention queries and keys as shown in Fig. 3 (a) to get the relation representations of the  $l$ -th layer  $R_{a}^{l} \in \mathbb{R}^{N \times N \times 2d_{\mathrm{model}}}$ . Before pairwise concatenation, to help queries and keys play the role of subjects and objects, a linear projection is added as follows:
+
+$$
+R _ {a} ^ {l} = \left[ Q ^ {l} W _ {S} ^ {l}; K ^ {l} W _ {O} ^ {l} \right], \tag {3}
+$$
+
+where  $Q^{l}\in \mathbb{R}^{N\times d_{\mathrm{model}}}$  and  $K^l\in \mathbb{R}^{N\times d_{\mathrm{model}}}$  refer to attention queries and keys of the  $l$  -th layer, and  $[\cdot ;\cdot ]$  denotes the pairwise concatenation.  $W_{S}^{l}$  and  $W_{O}^{l}$  are linear weights of shape  $\mathbb{R}^{d_{\mathrm{model}}\times d_{\mathrm{model}}}$ .
+
+We also utilize the last layer representations of the object queries  $Z^{L} \in \mathbb{R}^{N \times d_{\mathrm{model}}}$ , which are used for the object
+
+detection in the same manner:
+
+$$
+R _ {z} = \left[ Z ^ {L} W _ {S}; Z ^ {L} W _ {O} \right], \tag {4}
+$$
+
+where  $W_{S}, W_{O}$  are linear weights of shape  $\mathbb{R}^{d_{\mathrm{model}} \times d_{\mathrm{model}}}$ . To effectively use the various information learned in all layers, we introduce a gating mechanism as follows:
+
+$$
+g _ {a} ^ {l} = \sigma \left(R _ {a} ^ {l} W _ {G}\right), g _ {z} = \sigma \left(R _ {z} W _ {G}\right), \tag {5}
+$$
+
+where  $g_{a}^{l}$  and  $g_{z} \in \mathbb{R}^{N \times N \times 1}$  represent the gate values obtained through a single linear layer  $W_{G} \in \mathbb{R}^{2d_{\mathrm{model}} \times 1}$  for  $R_{a}^{l}$  and  $R_{z}$ , respectively. Finally, we extract the relation graph from the gated summation of the relation representations across all layers as follows:
+
+$$
+\hat {G} = \sigma \left(\mathrm {M L P} _ {\mathrm {r e l}} \left(\Sigma_ {l = 1} ^ {L} \left(g _ {a} ^ {l} * R _ {a} ^ {l}\right) + g _ {z} * R _ {z}\right)\right), \tag {6}
+$$
+
+where  $\hat{G} \in \mathbb{R}^{N \times N \times |\mathcal{C}_p|}$  denotes predicted relation graph and  $\mathrm{MLP}_{\mathrm{rel}}$  is a three-layer perceptron with ReLU activation. Note that we use sigmoid function  $\sigma$  so that multiple relationships can exist between objects.
+
+# 3.3. Learning and Inference
+
+To train EGTR, we perform multi-task learning. In addition to object detection and relation extraction, we devise connectivity prediction, an auxiliary task for relation extraction. The overall loss for the framework is as follows:
+
+$$
+\mathcal {L} = \mathcal {L} _ {\mathrm {o d}} + \lambda_ {\mathrm {r e l}} \mathcal {L} _ {\mathrm {r e l}} + \lambda_ {\mathrm {c o n}} \mathcal {L} _ {\mathrm {c o n}}, \tag {7}
+$$
+
+where  $\mathcal{L}_{\mathrm{od}}$  is object detection loss in Eq. (2).  $\mathcal{L}_{\mathrm{rel}}$  and  $\mathcal{L}_{\mathrm{con}}$  are loss functions for the relation extraction and connectivity prediction, respectively. Since explicit object detection loss is used, EGTR has the capability to detect all nodes in the scene graph. The details of the loss for each task are provided below.
+
+# 3.3.1 Relation Extraction
+
+We use binary cross-entropy loss for the relation extraction. To match predicted graph  $\hat{G} \in \mathbb{R}^{N \times N \times |\mathcal{C}_p|}$  and ground truth triplet set  $\mathcal{E}$ , we first encode  $\mathcal{E}$  as a one-hot ground truth graph  $G \in \mathbb{R}^{N \times N \times |\mathcal{C}_p|}$  by padding the regions that do not correspond to relations between the ground truth objects as zero. Then, we permute indices of the predicted graph using the permutation found in the object detection. From the permuted graph  $\hat{G}'$ , the loss for the relation extraction is calculated as  $\mathcal{L}_{\mathrm{rel}} = \mathcal{L}_r(\hat{G}', G)$ .
+
+However, since the size of the graph is proportional to the square of  $N$ , the sparsity of the ground truth graph  $G$  is too severe. For instance, the density is only  $10^{-14}$  when  $N$  is set to 200 for the Visual Genome [12] validation dataset. Therefore, we divide  $G$  into three regions: (1) GT region,
+
+![](images/8b4b9a348b44f666252103d9ddd29526093a28ee47d8308828587aeab359d27e.jpg)  
+girl - on - horse  
+jirl - riding
+
+(subject - predicate - object)  
+non-matching region  
+Figure 4. Example of graph  $G$  region. The target graph has a shape of  $N \times N \times |\mathcal{V}_c|$ . The graph is severely sparse since the number of object queries  $N$  is set large enough to cover objects in the image. We split  $G$  into GT, negative, and non-matching regions.  
+![](images/150501b135af7d479b690a4d0f111b17fd90658d9f4ed8d284146a7bbe9e14b7.jpg)  
+girl - wearing - pant
+
+(2) negative region, and (3) non-matching region, as shown in Fig. 4. The GT region indicates the ground truth triplets where the value of  $G$  is 1. The negative region consists of the triplets in which both subjects and objects are composed of the ground truth objects, but no relation exists between them. The non-matching region represents the zero-padded region. It is paired with a region consisting of the object candidates that do not match the ground truth objects and match  $\phi$  in Eq. (2). For each region, we apply different techniques to effectively train the relation extraction with object detection as subsequently described.
+
+Adaptive Smoothing. We propose a novel adaptive smoothing for the GT region. For  $G_{ijk}$  belonging to the GT region, the model is trained to predict the  $k$ -th predicate category between subject entity  $v_i$  and object entity  $v_j$ . However, since  $\hat{v}_i'$  and  $\hat{v}_j'$ , which match  $v_i$  and  $v_j$  respectively, do not have enough representations about the corresponding ground truth object at the beginning of the training, it may not be appropriate to predict the probability of the predicate as 1. Moreover, even when the object detection performance is reasonably assured, the detection performance may still vary for individual object candidates. Therefore, we reflect the detection performance of each object candidate on the relation label via adaptive smoothing.
+
+We first measure the uncertainty of each object candidate with the corresponding bipartite matching cost. For object candidate  $\hat{v}_i'$ , we define the uncertainty as follows:
+
+$$
+u _ {i} = \sigma \left(\operatorname {c o s t} _ {i} - \operatorname {c o s t} _ {\min } + \sigma^ {- 1} (\alpha)\right), \tag {8}
+$$
+
+where  $\mathrm{cost}_i$  denotes the matching cost and  $\mathrm{cost}_{\mathrm{min}}$  indicates the matching cost when  $\hat{v}_i^\prime$  perfectly matches  $v_{i}$  .  $\alpha$  is a nonnegative hyperparameter representing the minimum uncertainty. We set the value of  $G_{ijk}$  to  $(1 - u_{i})(1 - u_{j})$  taking the uncertainty into account. By using the relation label adjusted by the uncertainty, the multi-task learning of object
+
+detection and relation extraction is dynamically regulated according to the quality of the detected objects.
+
+Negative and Non-matching Sampling. Rather than employing all negatives, we sample them from the negative region. Inspired by the hard negative mining introduced by Liu et al. [21], we sort all negatives based on the predicted relation score  $\hat{G}_{ijk}^{\prime}$  and choose the top  $k_{\mathrm{neg}} \times |\mathcal{E}|$  most challenging negatives. Similarly, we extract  $k_{\mathrm{non}} \times |\mathcal{E}|$  hard samples from the non-matching region. As the non-matching region typically encompasses a substantial part of the graph  $G$ , this approach notably diminishes sparsity.
+
+# 3.3.2 Connectivity Prediction
+
+Influenced by Graph-RCNN [40] that predicted a relatedness to prune object pairs, we propose a connectivity prediction that predicts whether at least one edge between two object nodes exists for the relation extraction. We get a connectivity graph  $\hat{E} \in \mathbb{R}^{N \times N \times 1}$  in a similar way to get the relation graph in Eq. (6) with the same relation source representations. Instead of  $\mathrm{MLP}_{\mathrm{rel}}$  for multi-label prediction, we use another  $\mathrm{MLP}_{\mathrm{con}}$  for binary classification. We calculate the binary cross entropy loss from the permutated connectivity graph as follows:  $\mathcal{L}_{\mathrm{con}} = \mathcal{L}_r(\hat{E}', E)$ .
+
+# 3.3.3 Inference
+
+For model inference, we get triplet scores by multiplying predicate score  $\hat{G}_{ijk}$  by the corresponding class scores of  $\hat{v}_i^c$  and  $\hat{v}_j^c$ . Note that we set  $\hat{G}_{iik}$  to 0 to prevent self-connections in which the subject and object are the same entity. In addition, we enhance the triplet scores by utilizing the connectivity score  $\hat{E}_{ijk}$ . By multiplying the connectivity score, we can effectively filter out the triplets that are not likely to have a relation between the subject and object.
+
+# 4. Experiments
+
+# 4.1. Datasets and Evaluation Settings
+
+We conduct experiments on two SGG datasets. We describe the datasets and evaluation settings for each dataset. Detailed settings for each dataset are presented in the supplementary materials.
+
+Visual Genome. Visual Genome [12] is the most representative SGG dataset, consisting of 57K training images, 5K validation images, and 26K test images. We follow the popular Visual Genome split, retaining the most frequent 150 object categories and 50 relation categories. We adopt Scene Graph Detection (SGDet) evaluation settings and report Recall@k (R@k) which is class agnostic and mean Recall@k (mR@k) that aggregates the recalls for each predicate category. Following Motifs [42], these metrics are measures with graph constraint, which means each
+
+object pair can have a single predicate category. Since they are only related to triplet detection of the scene graph, we report AP50 that evaluates the detection performance of all objects appearing in the scene. Furthermore, we report the number of model parameters and Frames Per Second (FPS) to measure efficiency.
+
+Open Image V6. Open Image V6 [13] is also widely used dataset, comprising of 126K images for training, 2K for validation, and 5K for test set. It includes 601 object categories and 30 relation categories. For evaluation, we adopt both recall and weighted mean AP (wmAP) following standard settings. For recall evaluation, micro-R@50 is adopted. The WMAP is evaluated with two settings:  $\mathrm{wmAP}_{\mathrm{rel}}$  for predicting boxes of subject entity and object entity separately and  $\mathrm{wmAP}_{\mathrm{phr}}$  for predicting a union box of them. The final score is calculated by  $0.2 \times \text{micro-R} @50 + 0.4 \times \mathrm{wmAP}_{\mathrm{rel}} + 0.4 \times \mathrm{wmAP}_{\mathrm{phr}}$ .
+
+# 4.2. Implementation Details
+
+We employ Deformable DETR [46] that improves the convergence speed of the DETR with ResNet-50 [6] as a backbone. It is worth noting that our approach can be extended to any object detector that incorporates self-attention mechanisms between object features, including models like DETR [1], Sparse R-CNN [32], and others. We follow the configurations of the original Deformable DETR, except that we use only 200 object queries. To speed up convergence, we first train the object detector with the target dataset and subsequently train the SGG task using the pretrained object detector, similar to prior work [17]. To calculate the overall loss in Equation 7,  $\lambda_{\mathrm{rel}}$  is set to 15.  $\lambda_{\mathrm{con}}$  is set to 30 and 90 for the Visual Genome and Open Image V6, respectively. We set  $\alpha$  as  $10^{-14}$  for adaptive smoothing and both  $k_{\mathrm{neg}}$  and  $k_{\mathrm{non}}$  are set to 80.
+
+# 4.3. Results
+
+We present quantitative results and perform a comparative analysis of our proposed framework with representative SGG models. Additional experimental results are included in the supplementary materials.
+
+Visual Genome. Visual Genome results are shown in Tab. 1. Our proposed method demonstrates competitive performance with the current one-stage SGG models that have 1.5 to 6.5 times larger parameters with the fastest inference speed. In particular, our method achieves the highest object detection performance and shows comparable performance in the triplet detection compared to SSR-CNN [35], the state-of-the-art method. The results demonstrate that our method can generate scene graphs in an efficient and effective way by exploiting the by-products of the object detector. By applying logit adjustment [35], a technique for predicting tail predicate classes well, our
+
+<table><tr><td></td><td>Model</td><td># params (M)</td><td>FPS</td><td>AP50</td><td>R@20</td><td>R@50</td><td>R@100</td><td>mR@20</td><td>mR@50</td><td>mR@100</td></tr><tr><td rowspan="8">two-stage</td><td>IMP (EBM) [31, 39]</td><td>322.2</td><td>2.0</td><td>28.1</td><td>18.1</td><td>25.9</td><td>31.2</td><td>2.8</td><td>4.2</td><td>5.4</td></tr><tr><td>VTransE [44]</td><td>312.3</td><td>3.5</td><td>-</td><td>24.5</td><td>31.3</td><td>35.5</td><td>5.1</td><td>6.8</td><td>8.0</td></tr><tr><td>Motifs [42]</td><td>369.9</td><td>1.9</td><td>28.1</td><td>25.1</td><td>32.1</td><td>36.9</td><td>4.1</td><td>5.5</td><td>6.8</td></tr><tr><td>VCTree [33]</td><td>361.5</td><td>0.8</td><td>28.1</td><td>24.8</td><td>31.8</td><td>36.1</td><td>4.9</td><td>6.6</td><td>7.7</td></tr><tr><td>VCTree (TDE) [33, 34]</td><td>361.3</td><td>0.8</td><td>28.1</td><td>14.0</td><td>19.4</td><td>23.2</td><td>6.9</td><td>9.3</td><td>11.1</td></tr><tr><td>VCTree (EBM) [31, 33]</td><td>372.5</td><td>-</td><td>28.1</td><td>24.2</td><td>31.4</td><td>35.9</td><td>5.7</td><td>7.7</td><td>9.1</td></tr><tr><td>GPS-Net [19]</td><td>-</td><td>-</td><td>-</td><td>-</td><td>31.1</td><td>35.9</td><td>-</td><td>6.7</td><td>8.6</td></tr><tr><td>BGNN [16]</td><td>341.9</td><td>1.7</td><td>29.0</td><td>23.3</td><td>31.0</td><td>35.8</td><td>7.5</td><td>10.7</td><td>12.6</td></tr><tr><td rowspan="11">one-stage</td><td>FCSGG [20]</td><td>87.1</td><td>6.0</td><td>28.5</td><td>16.1</td><td>21.3</td><td>25.1</td><td>2.7</td><td>3.6</td><td>4.2</td></tr><tr><td>RelTR [7]</td><td>63.7</td><td>13.4</td><td>26.4</td><td>21.2</td><td>27.5</td><td>-</td><td>6.8</td><td>10.8</td><td>-</td></tr><tr><td>SGTR [17]</td><td>117.1</td><td>6.2</td><td>25.4</td><td>-</td><td>24.6</td><td>28.4</td><td>-</td><td>12.0</td><td>15.2</td></tr><tr><td>Relationformer [29]</td><td>92.9</td><td>8.5</td><td>26.3</td><td>22.2</td><td>28.4</td><td>31.3</td><td>4.6</td><td>9.3</td><td>10.7</td></tr><tr><td>Iterative SGG [9]</td><td>93.5</td><td>6.0</td><td>27.7†</td><td>-</td><td>29.7</td><td>32.1</td><td>-</td><td>8.0</td><td>8.8</td></tr><tr><td>SSR-CNN [35]</td><td>274.3</td><td>4.0</td><td>23.8†</td><td>25.8</td><td>32.7</td><td>36.9</td><td>6.1</td><td>8.4</td><td>10.0</td></tr><tr><td>SSR-CNN [35] LA,τ=0.3</td><td>274.3</td><td>4.0</td><td>23.8†</td><td>18.4</td><td>23.3</td><td>26.5</td><td>13.5</td><td>17.9</td><td>21.4</td></tr><tr><td>EGTR (Ours)</td><td>42.5</td><td>14.7</td><td>30.8</td><td>23.5</td><td>30.2</td><td>34.3</td><td>5.5</td><td>7.9</td><td>10.1</td></tr><tr><td>EGTR (Ours) LA,τ=0.7</td><td>42.5</td><td>14.7</td><td>30.8</td><td>15.7</td><td>18.7</td><td>20.5</td><td>12.1</td><td>17.8</td><td>21.7</td></tr><tr><td>EGTR (Ours) LA,τ=0.5</td><td>42.5</td><td>14.7</td><td>30.8</td><td>19.7</td><td>24.2</td><td>26.7</td><td>11.0</td><td>17.1</td><td>21.4</td></tr><tr><td>EGTR (Ours) LA,τ=0.3</td><td>42.5</td><td>14.7</td><td>30.8</td><td>22.4</td><td>28.2</td><td>31.7</td><td>8.8</td><td>14.0</td><td>18.3</td></tr></table>
+
+Table 1. Graph-Constraint results on the test set of Visual Genome dataset. We report the results of representative two-stage SGG models based on the Faster R-CNN [27] object detector with a ResNeXt-101-FPN [18, 38] backbone and concurrent one-stage SGG models. Among the one-stage models, the best is highlighted in bold, and the second-best is indicated with underlining. LA denotes logit adjustment proposed in SSR-CNN [35].  $\dagger$  represents that we concatenate the predicted subjects and objects and then apply non-maximum-suppression with a threshold of 0.3. Italic indicates the evaluation of metrics using publicly available model checkpoints. FPS is measured with a single V100 for images resized to a minimum of 600 for the shortest side and a maximum of 1000 for the longest side.  
+
+<table><tr><td>Model</td><td>score</td><td>micro-R@50</td><td>wmAPrel</td><td>wmAPphr</td></tr><tr><td>Motifs [42]</td><td>38.9</td><td>71.6</td><td>29.9</td><td>31.6</td></tr><tr><td>VCTree [33]</td><td>40.2</td><td>74.1</td><td>34.2</td><td>33.1</td></tr><tr><td>GPS-Net [19]</td><td>41.7</td><td>74.8</td><td>32.9</td><td>34.0</td></tr><tr><td>BGNN [16]</td><td>42.1</td><td>75.0</td><td>33.5</td><td>34.2</td></tr><tr><td>RelTR [7]</td><td>43.0</td><td>71.7</td><td>34.2</td><td>37.5</td></tr><tr><td>SGTR [17]</td><td>42.3</td><td>59.9</td><td>37.0</td><td>38.7</td></tr><tr><td>SSR-CNN [35]</td><td>49.4</td><td>76.7</td><td>41.5</td><td>43.6</td></tr><tr><td>EGTR (Ours)</td><td>48.6</td><td>75.0</td><td>42.0</td><td>41.9</td></tr></table>
+
+Table 2. Results on test set of Open Image V6. The score is a weighted sum of micro-R@50,  $\mathrm{wmAP}_{\mathrm{rel}}$ , and  $\mathrm{wmAP}_{\mathrm{phr}}$ .
+
+method performs favorably in the trade-off between  $\mathrm{R}@\mathbf{k}$  and  $\mathrm{mR}@\mathbf{k}$ , showing superiority over existing state-of-the-art models. Note that for triplet detection models [9, 35] without an explicit object detector, AP50 is measured by applying non-maximum-suppression (NMS) to the union of predicted subjects and objects set. AP50 performance is discussed further in the section 4.5.
+
+Open Image V6. As depicted in Tab. 2, the experiments conducted on the Open Image V6 dataset also demonstrate competitive performance, underscoring the effectiveness and robustness of our method across different datasets.
+
+# 4.4. Ablation Studies
+
+We analyze the effects of our model components. All of the ablation studies are done with the Visual Genome dataset.
+
+Relation Sources. In Tab. 3, we investigate the source of
+
+relations used for the relation extractor. To verify the benefits of using attention queries and keys, we change  $Q^l$  and  $K^l$  in Eq. (3) to  $Z^l$ , the hidden states of each layer. Remarkably, when simply using the hidden states of all layers, the triplet detection performance decreases. Surprisingly, the performance is lower than that of using only  $R_z$  from the last layer. On the other hand, when only attention by-products are used as a relation source, the performance is similar to that of using only hidden states of the last layer, the most contextualized representations for detecting objects. Incorporating both attention by-products and final hidden states yields improved performance by amalgamating varied information. The results underscore that the attention by-products of the object detectors contain rich information for relation extraction.
+
+Training Techniques. In Tab. 4, we ablate proposed techniques. Without any techniques, our model framework shows low performance. When each technique is applied, it substantially improves both R@50 and mR@50, and the best performance is achieved when all techniques are combined. The results demonstrate that our proposed techniques help to train the model framework effectively. Additional experimental results are provided in the supplementary materials.
+
+Sampling Methodology. In Tab. 5, we investigate the impact of hard sampling methods for the negative and non-matching regions. Hard negatives and hard non-matchings show a trade-off that improves the performance of R@50
+
+<table><tr><td>Rl^l a source</td><td>Rl^l a</td><td>Rz</td><td>R@50</td><td>mR@50</td></tr><tr><td>Ql &amp; Kl</td><td>✓</td><td>✓</td><td>30.2</td><td>7.9</td></tr><tr><td>Zl</td><td>✓</td><td>✓</td><td>29.6</td><td>7.4</td></tr><tr><td>-</td><td></td><td>✓</td><td>29.9</td><td>7.6</td></tr><tr><td>Ql &amp; Kl</td><td>✓</td><td></td><td>29.8</td><td>7.7</td></tr></table>
+
+Table 3. Ablation study on relation source.  
+
+<table><tr><td>adaptive smoothing</td><td>\( \mathcal{L}_{\text{con}} \)</td><td>sampling</td><td>R@50</td><td>mR@50</td></tr><tr><td></td><td></td><td></td><td>26.6</td><td>5.3</td></tr><tr><td>✓</td><td></td><td></td><td>28.3</td><td>6.5</td></tr><tr><td></td><td>✓</td><td></td><td>29.6</td><td>7.0</td></tr><tr><td></td><td></td><td>✓</td><td>28.9</td><td>7.1</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>30.2</td><td>7.9</td></tr></table>
+
+Table 4. Ablation study on proposed techniques.  
+
+<table><tr><td>hard negative</td><td>hard non-matching</td><td>R@50</td><td>mR@50</td></tr><tr><td>✓</td><td>✓</td><td>30.2</td><td>7.9</td></tr><tr><td>✓</td><td></td><td>30.0</td><td>7.1</td></tr><tr><td></td><td>✓</td><td>29.6</td><td>7.7</td></tr></table>
+
+and mR@50, respectively, and additional performance improvements are achieved for both measures when they are used together. For hard negatives, we speculate that R@50 performance is improved by selecting negatives for the tail predicate class, which is likely to exist in reality but is not in the annotation. For hard non-matchings, mR@50 performance seems to have improved by choosing object candidates that are akin to the ground truth objects but unmatched with them. This encourages the model to predict diverse predicate classes by preventing duplicate object pairs, which are likely to predict head predicate classes, from being predicted.
+
+# 4.5. Discussion
+
+Regarding the highest AP50 score, we posit that it is because our proposed method focuses on not only objects with relations but also objects without any relations. To validate this, we calculate AP50 for two subsets of the ground-truth objects: those with relations and those without relations. We compare our method with two triplet detection-based models. Tab. 6 illustrates that our method achieves significantly high  $\mathrm{AP50}_{\mathrm{no - rel}}$ . Considering that triplet detection performance is related to  $\mathrm{AP50}_{\mathrm{rel}}$ , while triplet detection-based models only focus on objects with relations to detect triplets well, our method is capable of extracting the complete scene graph, including objects without relations.
+
+For further analysis, we visualize bounding boxes for subjects and objects from the top 20 predictions in Fig. 5. For SSR-CNN [35], which directly predicts triplets, we observe a significant number of overlapping bounding boxes. On the other hand, our proposed model leverages predic
+
+Table 5. Ablation study on sampling methods.  
+
+<table><tr><td>Model</td><td>AP50</td><td>\( AP50_{\text{rel}} \)</td><td>\( AP50_{\text{no-rel}} \)</td></tr><tr><td>Iterative SGG [9]†</td><td>27.7</td><td>24.3</td><td>7.8</td></tr><tr><td>SSR-CNN [35]†</td><td>23.8</td><td>20.2</td><td>7.4</td></tr><tr><td>EGTR (Ours)</td><td>30.8</td><td>24.3</td><td>10.7</td></tr></table>
+
+Table 6. AP50 for two subsets of objects.  $\mathrm{AP50}_{\mathrm{rel}}$  and  $\mathrm{AP50}_{\mathrm{no - rel}}$  denote scores evaluated only on objects having at least one edge and no edge, respectively. Note that each measure operates on its unique scale since the ground-truth objects used for each measure differ.  $\dagger$  indicates that additional NMS is applied to the union of subjects and objects set.
+
+![](images/24bd9ff20bfc213d9d6ec0c5211af0c429e35efa9df27fb0358e28fccd39745f.jpg)
+
+![](images/18ae33dd39cb2090a7c41cafda42b6943398cbcb2bce6166315195b836fe3f0f.jpg)  
+(a) SSR-CNN [35]
+
+![](images/46306f91c2dbb914d9f6359cae5848fe4882b7587c39fc49c4b75c459bcd9dfb.jpg)
+
+![](images/bcd320527c7035d3e2f40585be25959ccfdf58231515632ecd35636a8f686e10.jpg)  
+(b) EGTR (Ours)  
+Figure 5. Comparison of detected subjects and objects. We select the top 20 predictions of SSR-CNN and EGTR and visualize bounding boxes of detected subjects and objects.
+
+tions from an explicit object detector, reducing redundant predictions. Furthermore, our method detects various subjects and objects that appear in the scene rather than focusing on objects that are more likely to have relations.
+
+# 5. Conclusion
+
+In this study, we propose the lightweight one-stage scene graph generator EGTR. We significantly reduce the model complexity by harnessing the relationships among the object queries learned from the self-attention layers of the object decoder. Furthermore, we devise a novel adaptive smoothing technique that helps multi-task learning of object detection and relation extraction by adjusting the relation label according to the object detection performance. As an auxiliary task of relation extraction, connectivity prediction contributes to the effective learning of EGTR. We conduct extensive experiments and ablation studies, and the results demonstrate that EGTR achieves the highest object detection performance and competitive triplet detection capabilities with the fastest inference speed.
+
+Acknowledgements. We thank Taeho Kil, Geewook Kim, and the anonymous reviewers for their insightful comments and suggestions.
+
+# References
+
+[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 2, 3, 4, 6  
+[2] Zhe Chen, Shaoli Huang, and Dacheng Tao. Context refinement for object detection. In ECCV, pages 71-86, 2018. 2  
+[3] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE TPAMI, 2023. 1, 3  
+[4] Naina Dhingra, Florian Ritter, and Andreas Kunz. Bgt-net: Bidirectional gru transformer network for scene graph generation. In CVPR, pages 2150-2159, 2021. 2  
+[5] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In ICMLC, pages 225-229, 2018. 1  
+[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6  
+[7] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In CVPR, pages 3588-3597, 2018. 2, 7  
+[8] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, pages 3668-3678, 2015. 1  
+[9] Siddhesh Khandelwal and Leonid Sigal. Iterative scene graph generation. In NeurIPS, 2022. 1, 3, 7, 8  
+[10] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In CVPR, pages 6271-6280, 2019. 1  
+[11] Rajat Koner, Suprosanna Shit, and Volker Tresp. Relation transformer network. ECCV, pages 422-439, 2022. 2  
+[12] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 2, 5, 6  
+[13] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 2, 6  
+[14] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In ICCV, pages 10313-10322, 2019. 1  
+[15] Lin Li, Guikun Chen, Jun Xiao, Yi Yang, Chunping Wang, and Long Chen. Compositional feature augmentation for unbiased scene graph generation. In ICCV, pages 21685-21695, 2023. 2  
+[16] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In CVPR, pages 11109-11119, 2021. 1, 2, 7
+
+[17] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In CVPR, pages 19486-19496, 2022. 1, 3, 6, 7  
+[18] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 7  
+[19] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In CVPR, pages 3746-3753, 2020. 1, 2, 7  
+[20] Hengyue Liu, Ning Yan, Masood Mortazavi, and Bir Bhanu. Fully convolutional scene graph generation. In CVPR, pages 11546-11556, 2021. 1, 3, 7  
+[21] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 6  
+[22] Yong Liu, Ruiping Wang, Shiguang Shan, and Xilin Chen. Structure inference net: Object detection using scene-level context and instance-level relationships. In CVPR, pages 6985-6994, 2018. 2  
+[23] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In ECCV, pages 852–869, 2016. 2  
+[24] Yichao Lu, Himanshu Rai, Jason Chang, Boris Knyazev, Guangwei Yu, Shashank Shekhar, Graham W Taylor, and Maksims Volkovs. Context-aware scene graph generation with seq2seq transformers. In ICCV, pages 15931-15941, 2021.  
+[25] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In ICCV, pages 13296-13307, 2023. 2  
+[26] Alejandro Newell and Jia Deng. Pixels to graphs by associative embedding. NeurIPS, 30, 2017. 1, 3  
+[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 2, 7  
+[28] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80. Association for Computational Linguistics, 2015. 1  
+[29] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In ECCV, pages 422-439, 2022. 1, 3, 7  
+[30] Gopika Sudhakaran, Devendra Singh Dhami, Kristian Kersting, and Stefan Roth. Vision relation transformer for unbiased scene graph generation. In ICCV, pages 21882-21893, 2023. 2  
+[31] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In CVPR, pages 13936-13945, 2021. 7
+
+[32] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, and Ping Luo. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 14454-14463, 2021. 2, 3, 6  
+[33] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In CVPR, pages 6619-6628, 2019. 1, 2, 7  
+[34] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In CVPR, pages 3716-3725, 2020. 7  
+[35] Yao Teng and Limin Wang. Structured sparse r-cnn for direct scene graph generation. In CVPR, pages 19437-19446, 2022. 1, 3, 6, 7, 8  
+[36] 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. 2, 3  
+[37] Sanghyun Woo, Dahun Kim, Donghyeon Cho, and In So Kweon. Linknet: Relational embedding for scene graph. NeurIPS, 31, 2018. 2  
+[38] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 1492-1500, 2017. 7  
+[39] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In CVPR, pages 5410–5419, 2017. 1, 2, 7  
+[40] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In ECCV, pages 670-685, 2018. 6  
+[41] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In CVPR, pages 10685–10694, 2019. 1  
+[42] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In CVPR, pages 5831-5840, 2018. 1, 2, 6, 7  
+[43] Cheng Zhang, Wei-Lun Chao, and Dong Xuan. An empirical study on leveraging scene graphs for visual question answering. In BMVC, 2019. 1  
+[44] Hanwang Zhang, Zawlin Kyaw, Shih-Fu Chang, and Tat-Seng Chua. Visual translation embedding network for visual relation detection. In CVPR, page 3107-3115, 2017. 7  
+[45] Chaofan Zheng, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, Jingkuan Song, and Lianli Gao. Learning to generate scene graph from head to tail. In ICME, pages 1-6. IEEE, 2022. 2  
+[46] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 3, 6

2024/EGTR_ Extracting Graph from Transformer for Scene Graph Generation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbb00bc197d6b92493feb5957d82b027f749dddc3f93b2bf8cd49b8f8b50cbcb
+size 510266

2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/f8a85da7-9221-4a37-baf0-e5a3b46705e3_content_list.json

@@ -0,0 +1,1976 @@
+[
+    {
+        "type": "text",
+        "text": "EMAGE: Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling",
+        "text_level": 1,
+        "bbox": [
+            130,
+            128,
+            841,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Haiyang Liu $^{1*}$  Zihao Zhu $^{2*}$  Giorgio Becherini $^{4}$  Yichen Peng $^{5}$  Mingyang Su $^{6}$  You Zhou $^{3}$  Xuefei Zhe $^{3}$  Naoya Iwamoto $^{3}$  Bo Zheng $^{3}$  Michael J. Black $^{3}$",
+        "bbox": [
+            143,
+            200,
+            821,
+            241
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ The University of Tokyo  $^{2}$ Keio University  $^{3}$ Independent Researcher  $^{4}$ Max Planck Institute for Intelligent Systems  $^{5}$ JAIST  $^{6}$ Tsinghua University",
+        "bbox": [
+            155,
+            255,
+            789,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/77aa6d75a5af7f91c15f89f767785f7436b00b41c1d6074cbaa3e215dafb7c19.jpg",
+        "image_caption": [
+            "Figure 1. EMAGE. We present a Masked Audio-Conditioned Gesture Modeling framework, along with a new holistic gesture dataset, BEAT2 (BEAT-SMPLX-FLAME), for jointly generating facial expressions, local body dynamics, hand movements and global translations, conditioned on audio and a partially or completely masked gestures. The gray denotes visible gestures, and blue represents our outputs."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            116,
+            327,
+            857,
+            505
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            565,
+            313,
+            583
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We propose EMAGE, a framework to generate full-body human gestures from audio and masked gestures, encompassing facial, local body, hands, and global movements. To achieve this, we first introduce BEAT2 (BEATSMLPX-FLAME), a new mesh-level holistic co-speech dataset. BEAT2 combines a MoShed SMPL-X body with FLAME head parameters and further refines the modeling of head, neck, and finger movements, offering a community-standardized, high-quality 3D motion captured dataset. EMAGE leverages masked body gesture priors during training to boost inference performance. It involves a Masked Audio Gesture Transformer, facilitating joint training on audio-to-gesture generation and masked gesture reconstruction to effectively encode audio and body gesture hints. Encoded body hints from masked gestures are then separately employed to generate facial and body movements. Moreover, EMAGE adaptively merges speech features from the audio's rhythm and content and utilizes four compositional VQ-VAEs to enhance the results' fidelity and diversity. Experiments demonstrate that EMAGE generates",
+        "bbox": [
+            75,
+            598,
+            473,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "holistic gestures with state-of-the-art performance and is flexible in accepting predefined spatial-temporal gesture inputs, generating complete, audio-synchronized results. Our code and dataset are available.<sup>1</sup>",
+        "bbox": [
+            496,
+            566,
+            893,
+            628
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            500,
+            657,
+            633,
+            672
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The path towards full-body co-speech gesture generation presents many remaining challenges. Despite the development of various baselines, e.g., audio conditioned models for face [15, 20, 61], body [13, 25, 67], and hand movements [33, 39, 41], along with a few attempts at merged models [28, 65], the limited availability of comprehensive data and models poses an ongoing obstacle. To make progress in this direction, we present a new framework that can incorporate partial spatial-temporal predefined gestures and autonomously complete the remaining frames in synchronization with the audio. This provides a new tool for creating realistic animations of digital humans",
+        "bbox": [
+            496,
+            681,
+            893,
+            866
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Equal Contribution.  \n1https://pantomatrix.github.io/EMAGE/",
+        "bbox": [
+            522,
+            875,
+            810,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "1144",
+        "bbox": [
+            482,
+            944,
+            516,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[57, 63, 64, 71].",
+            "To achieve this, we first require a comprehensive gesture dataset. While several datasets are available for audio-to-body [67], audio-to-body&face [28], body-to-hands [39, 41], and common actions [16, 46], integrating them poses challenges due to varying data formats and evaluation metrics across different sub-areas. This motivates us to establish a unified, community-standard benchmark for training and evaluating holistic gestures. We build upon the BEAT dataset [39], which provides the most comprehensive mocap co-speech gesture dataset with diverse modalities. Despite its size and richness, the data in BEAT is not presented in a standardized format. The dataset comprises iPhone ARKit blendshape weights and Vicon skeletons, posing challenges when using the data for training; e.g., the lack of a mesh representation prevents the use of a vertex loss [19, 52]. Additionally, conducting direct human perceptual evaluations on these diverse modalities is inherently complex. On the other hand, SMPL-X [50] and FLAME [35] are common mesh standards for the academic community for other human dynamics-related tasks, e.g., action [52] and talking head generation [14]. Driven by the motivation to facilitate the sharing of knowledge across different tasks, we present BEAT-SMPLX-FLAME (BEAT2), which consists of two main components: i) Refined SMPL-X body shape and pose parameters from MoSh++ [44, 46] with hard-coded physical priors. ii) Optimized high-quality FLAME head parameters. By integrating SMPL-X's body and FLAME's head, BEAT2 facilitates a comprehensive training and evaluation benchmark across multiple sub-areas of co-speech human animation generation.",
+            "With the holistic data from BEAT2, we aim to enhance the coherence of individual body parts while ensuring accurate cross-modal alignment between audio and motion; see Figure 1. This leads to the design of Expressive Masked Audio-conditioned GEsture modeling, denoted as EMAGE, a spatial and temporal transformer-based framework. EMAGE first aggregates spatial features from masked body joints. Then, it reconstructs the latent space of pretrained gestures with a switchable gesture temporal self-attention and audio-gesture cross-attention. The selection of different forward paths enables modeling the gesture-to-gesture and audio-to-gesture prior separately and effectively. Once the reconstructed latent features are obtained, EMAGE decodes local face and body gestures from four compositional pretrained Vector Quantized-Variational AutoEncoders (VQ-VAEs) and decodes global translations from a pretrained Global Motion Predictor. The compositional VQ-VAEs are key to preserving audio-related motions during the reconstruction. Through these designs, EMAGE achieves state-of-the-art performance in generating body and face gestures. It takes as input audio and partially initialized gestures, and recovers audio"
+        ],
+        "bbox": [
+            76,
+            90,
+            472,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "synchronized, coherent gestures. Additionally, we show how EMAGE can flexibly incorporate additional, nonholistic, datasets to improve results.",
+            "Overall, our contributions are as follows: (1) We release BEAT2, a representation-unified, mesh-level dataset that leverages MoShed SMPL-X body and FLAME head parameters. (2) We propose EMAGE, a simple yet effective holistic gesture generation framework that generates coherent gestures with partial gesture and audio priors. (3) EMAGE achieves state-of-the-art performance in generating both body and face gestures with only four frames of seed gestures. (4) EMAGE showcases the use of additional non-holistic gesture datasets for training, e.g., Trinity and AMASS, to further improve the fidelity and diversity of the results, effectively leveraging data from different datasets."
+        ],
+        "bbox": [
+            496,
+            90,
+            893,
+            318
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            330,
+            640,
+            347
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Co-Speech Animation Datasets are categorized into two types; see Table 1: pseudo-labeled (PGT) and motion-captured (mocap). For PGT, the Speech2Gesture Dataset [25] utilizes OpenPose [11] to extract 2D poses from News and Teaching videos. Subsequent works extend the dataset to 3D pose [28] and SMPL-X [65]. The TED Dataset [66], which extracts 2D pose from TED videos, has been extended to include 3D pose [67], fingers [40], and SMPL-X [45]. Similarly, 2D, 3D landmarks, and meshes are estimated from multi-view recorded videos [60] as PGT for talking-head generation. Although pseudo-labeled approaches could theoretically extract infinite data, their accuracy is limited. For instance, the average error (in mm) for the SOTA monocular 3D pose estimation algorithm [23] on the Human3.6M dataset [30] is 33.4, while for Vicon mocap it is 0.142 [47]. For mocap datasets, Trinity [3] features one male actor with 4h of data, and TalkingWithHands [32] collects conversation scenarios for two speakers. ZEEG [24] considers one speaker with 12 styles. For the face, 3D scanning is accurate but limited in quantity due to cost, e.g., less than 3h for BIWI [21], VOCA [14], and MeshTalk [56]. This leads to a balance between performance and quantity in the ARKit dataset [51]. The aforementioned datasets address the body and face separately. On the other hand, BEAT [39] includes both 3D pose body and ARKit [8] facial blendshape weights for the first time. However, BEAT lacks mesh data.",
+            "Co-Speech Animation Models are categorized based on their outputs; see Table 2. In addition to the baselines and datasets [14, 25, 32, 48, 56, 67], several frameworks for body gestures [2, 33, 53, 54] have improved the performance of the baselines. They are trained and evaluated by selecting specific upper body joints [4, 5, 38-40, 49, 70], or all body joints [62, 72]. Recent improvements in facial gestures drive the vertices with transformers and discrete face priors [19, 61]. However, these methods only address ei"
+        ],
+        "bbox": [
+            496,
+            357,
+            893,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "1145",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/66e6b8d9cd0bc32106e05b2751c2989dd9fc80f1283577f7cd48d24a19ffe185.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Trinity Mocap 2017 [22]</td><td>S2G S2G-2D 2019 [25]</td><td>Seq2Seq TED-2D 2019 [66]</td><td>TWH Mocap 2019 [32]</td><td>Trimodal TED-3D 2020 [67]</td><td>VOCA Scan 2020 [14]</td><td>MeshTalk Scan 2021 [56]</td><td>Hahibie et al. S2G-3D 2021 [28]</td><td>HA2G TED-3D+ 2022 [40]</td><td>BEAT Mocap 2022 [39]</td><td>ZEEG Mocap 2022 [24]</td><td>Yoon. et al. TED-SMPL 2023 [45]</td><td>Talkshow S2G-SMPL 2023 [65]</td><td>BEAT2 Ours 2024</td></tr><tr><td>Head</td><td>-</td><td>2D</td><td>-</td><td>-</td><td>-</td><td>3D Scan</td><td>3D Scan</td><td>3D</td><td>-</td><td>ARKit</td><td>-</td><td>-</td><td>PGT-Mesh</td><td>MC-Mesh</td></tr><tr><td>Upper Body</td><td>3D</td><td>2D</td><td>2D</td><td>3D</td><td>3D</td><td>-</td><td>-</td><td>3D</td><td>3D</td><td>3D</td><td>3D</td><td>PGT-Mesh</td><td>PGT-Mesh</td><td>MC-Mesh</td></tr><tr><td>Hands</td><td>3D</td><td>2D</td><td>-</td><td>3D</td><td>-</td><td>-</td><td>-</td><td>3D</td><td>3D</td><td>3D</td><td>3D</td><td>PGT-Mesh</td><td>PGT-Mesh</td><td>MC-Mesh</td></tr><tr><td>Lower Body</td><td>3D</td><td>-</td><td>-</td><td>3D</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>3D</td><td>3D</td><td>-</td><td>-</td><td>MC-Mesh</td></tr><tr><td>Global Motion</td><td>3D</td><td>-</td><td>-</td><td>3D</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>3D</td><td>3D</td><td>-</td><td>-</td><td>MC-Mesh</td></tr><tr><td>Duration (hours)</td><td>4</td><td>60</td><td>97</td><td>20</td><td>97</td><td>0.5</td><td>2</td><td>38</td><td>33</td><td>76</td><td>4</td><td>30</td><td>27</td><td>60</td></tr></table>",
+        "bbox": [
+            81,
+            87,
+            883,
+            170
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/e89e22ee13f61b8b34cf4be94f8611a9fa64537145333d53cd95adc63906479a.jpg",
+        "table_caption": [
+            "Table 1. Comparison of Co-Speech Gesture Datasets. We summarize gesture and face datasets in talking scenarios. 'PGT' and 'MC' denote Pseudo Ground Truth and Motion Capture, respectively. The BEAT2 (Ours) dataset is the largest dataset for motion-captured data, providing holistic, academic community standard mesh-level information."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td>S2G</td><td>TriModal</td><td>B2H</td><td>TWH</td><td>Hahibie et al.</td><td>DisCo</td><td>HA2G</td><td>Face Former 2022[19]</td><td>Rhythmic</td><td>BEAT CaMN 2022[39]</td><td>Diff Gesture 2023[72]</td><td>Code Talker 2023[61]</td><td>Talk Show 2023[65]</td><td>DiffStyle Gesture 2023[62]</td><td>Body Former 2023[49]</td><td>Lively Speaker 2023[70]</td><td>EMAEG (Ours) 2024</td></tr><tr><td>2019 [25]</td><td>2020 [67]</td><td>2020[48]</td><td>2021[32]</td><td>2021[28]</td><td>2022[38]</td><td>2022[40]</td><td>2022[5]</td><td>2022[5]</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>Masked</td></tr><tr><td>Input Gestures</td><td>-</td><td>-</td><td>Body</td><td>Body</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td></tr><tr><td>Head</td><td>-</td><td>-</td><td>-</td><td>M</td><td>-</td><td>-</td><td>S</td><td>-</td><td>-</td><td>-</td><td>S</td><td>S</td><td>-</td><td>-</td><td>-</td><td>C</td><td></td></tr><tr><td>Upper</td><td>S</td><td>S</td><td>-</td><td>M</td><td>S</td><td>C</td><td>-</td><td>S</td><td>C</td><td>S</td><td>-</td><td>C</td><td>S</td><td>S</td><td>S</td><td>C</td><td></td></tr><tr><td>Hands</td><td>-</td><td>-</td><td>S</td><td>S</td><td>M</td><td>S</td><td>C</td><td>-</td><td>-</td><td>C</td><td>S</td><td>-</td><td>C</td><td>S</td><td>-</td><td>S</td><td>C</td></tr><tr><td>Lower + Global</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>S</td><td>-</td><td>-</td><td>C</td><td></td></tr></table>",
+        "bbox": [
+            81,
+            228,
+            883,
+            296
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Comparison of Co-Speech Gesture Models. We compare with previous methods for generating face or body motion trained on co-speech datasets. The first row lists their inputs, and the subsequent rows list their outputs, respectively. Different decoder designs are denoted by the initials 'S' for Single, 'M' for Multiple, and 'C' for Cascaded. To the best of our knowledge, EMAGE (Ours) is the first to accept audio and partially or completely masked gestures, generating full-body audio-synchronized results.",
+        "bbox": [
+            75,
+            301,
+            893,
+            357
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ther the face or body. Applying these methods to the full body will yield sub-optimal results as audio is differently correlated with face and body dynamics. Most similar to our work, Habibie et al. [28] employ a single audio encoder and multiple decoders for generating face and body gestures. TalkSHOW [65] demonstrates the advantage of separating the audio encoders and decoding the body and hands using quantized codes in an auto-regressive manner. However, it lacks lower body and global motion, and there is no shared information for body and facial gesture generation. Moreover, it cannot accept partially masked body hints due to the design of a fully auto-regressive model.",
+        "bbox": [
+            75,
+            383,
+            470,
+            565
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Masked Representation Learning was first demonstrated to be effective in Natural Language Processing, with BERT-based models [17, 31, 42] boosting the performance of learned word embeddings in downstream tasks through a combination of masked language modeling and transformer architecture. Subsequently, Masked AutoEncoders [29] expanded masked image modeling to computer vision by removing and inpainting image patches. This concept of masked representation learning has since been employed in other modalities, e.g., video [12, 18, 43], audio [26, 36, 37], and point clouds [55, 68, 69]. Most related to our work, MotionBERT [73] proposes a spatial-temporal transformer to learn a robust motion representation for classification-based tasks by masking 2D pose. In contrast to their method, we target robust motion features for conditional motion generation tasks, requiring a balance in training between multiple modalities, e.g., audio and gesture.",
+        "bbox": [
+            75,
+            566,
+            470,
+            825
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. BEAT2",
+        "text_level": 1,
+        "bbox": [
+            76,
+            843,
+            163,
+            858
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This section introduces how we obtain unified mesh-level data, i.e., SMPL-X [50] and FLAME [35] parameters, from",
+        "bbox": [
+            75,
+            869,
+            470,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "the original BEAT dataset [39]. BEAT utilizes a Vicon motion capture system [47] and releases 78-joint skeleton-level Biovision Hierarchy (BVH) [10] motion files. Their facial capture system uses ARKit [8] with a depth camera on the iPhone 12 Pro, extracting  $\\mathbb{R}^{51}$  blendshape weights. These blendshapes are designed based on the Facial Action Coding System (FACS), which is widely adopted. However, both the motion and facial data lack mesh-level details (see Figure 2), e.g., shapes and vertices.",
+        "bbox": [
+            496,
+            383,
+            893,
+            521
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Body Parameters Initialization via MoSh++",
+        "text_level": 1,
+        "bbox": [
+            498,
+            542,
+            870,
+            559
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We initialize the SMPL-X body shape and pose parameters from BEAT mocap marker data using MoSh++ [46]. Given the captured markers positions  $\\mathbf{m} \\in \\mathbb{R}^{T \\times K \\times 3}$ , predefined markers position offsets  $\\mathbf{d} \\in \\mathbb{R}^{K \\times 3}$ , and user-defined vertices-to-markers function  $\\mathcal{H}$ , we aim to obtain body shape  $\\beta \\in \\mathbb{R}^{300}$ , poses  $\\theta \\in \\mathbb{R}^{T \\times 55 \\times 3}$ , and translation parameters  $\\gamma \\in \\mathbb{R}^{T \\times 3}$ . The optimization uses a differentiable surface vertex mapping function  $S(\\beta, \\theta, \\gamma)$  and vertex normal function  $\\mathcal{N}(\\beta, \\theta_0, \\gamma_0)$ . For each frame, the latent marker  $\\tilde{\\mathbf{m}} \\in \\mathbb{R}^{T \\times K \\times 3}$  is calculated as",
+        "bbox": [
+            496,
+            569,
+            890,
+            720
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\tilde {\\mathbf {m}} _ {i} \\equiv \\mathcal {S} _ {\\mathcal {H}} (\\beta , \\theta_ {i}, \\gamma_ {i}) + \\mathbf {d N} _ {\\mathcal {H}} (\\beta , \\theta_ {i}, \\gamma_ {i}). \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            565,
+            742,
+            890,
+            760
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We first select 12 frames to optimize and fix  $\\mathbf{d}$  and  $\\beta$ , then optimize  $\\theta_{i}$ ,  $\\gamma_{i}$  for  $i \\in (1:T)$  by minimizing the loss terms based on  $\\|\\tilde{\\mathbf{m}}_i - \\mathbf{m}_i\\|^2$ , including Data Term, Surface Distance Energy, Marker Initialization Regularization, Pose and Shape Priors, Velocity Constancy, and Soft-Tissue Term (see supplementary materials). The overall objective function is the weighted sum of these terms to balance accuracy and plausibility.",
+        "bbox": [
+            496,
+            779,
+            893,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "1146",
+        "bbox": [
+            482,
+            944,
+            516,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/1a29c1ea8463bf00c2c3f184a817a4f1c708628a8ef1d47d76d7f19f87bcfc5d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            90,
+            197,
+            220
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/0d3097e295ab5f39d373170046e3ede29bd44caca6a90f7a1e16e8cbbc74f8e2.jpg",
+        "image_caption": [
+            "BEAT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            133,
+            224,
+            197,
+            273
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/3cf8e0bff276205e84046de305c5ed3d778dd92008b38ecc417fbe3aa563ac59.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            202,
+            92,
+            297,
+            220
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/9c115889ca59e8c11153c9150f32df01371750dbb710af599fab43cfc7f8e73d.jpg",
+        "image_caption": [
+            "Retargeted"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            202,
+            224,
+            297,
+            273
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/0b2f09f2374956161351194f45e2e2532c9d05df8cf0f908d5ce265ba34ba94c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            308,
+            92,
+            398,
+            220
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/f8f4640aefd7e52843ec5ee7f210087f3efb995a8208587356d71d31ebcdc3c4.jpg",
+        "image_caption": [
+            "Moshed",
+            "Figure 2. Comparison of Data between BEAT2 and Others. BEAT-SMPLX-FLAME presents a new mesh-level, motion-captured, holistic co-speech gesture dataset with 60h of data. Left: We compare our refined SMPL-X body parameters (denoted as Refined MoSh) with the original BEAT skeleton [39], Retargeted from AutoRegPro, and initial results of  $\\mathrm{Mosh}++$  [50]. The refined results show correct neck flexion, appropriate head and neck shape ratios, and detailed finger representations. Right: Visualization of blendshape weights from the original BEAT dataset [39] with ARKit's template, Wrapped-based, and handcrafted optimization. Our final handcrafted FLAME blendshape-based optimization demonstrates both accurate lip movement details and natural mouth shapes."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            308,
+            224,
+            403,
+            273
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/3c1b9aa5d1182a6c0f9d831ead367f4923e1919823cf79b76dca4f59fb4552c0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            408,
+            92,
+            506,
+            220
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/998179254e8dbd0a0cd08de2c91b898bf38655be992f223fb73fe5128e9c24bd.jpg",
+        "image_caption": [
+            "Refined Moshed (Ours)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            408,
+            223,
+            506,
+            273
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/2e2280169812212b1d236e2a6a3e66b5ab26bd14574a2bb1d9abdb43d01b8b65.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            95,
+            612,
+            220
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/0823b4da28a4d4e58bba88dec90cc1bb5f4de875f981e5c1269d82ed5df385a6.jpg",
+        "image_caption": [
+            "BEAT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            224,
+            617,
+            273
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/0a5a6dee70f74b4a51ebca6acff78622320dfb2cf09a947e6ba026a25c523992.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            620,
+            98,
+            718,
+            220
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/4e2e0a6d3965907893b8e602a4bd4865e5ecea413958654dd3f083be29da6d78.jpg",
+        "image_caption": [
+            "Wrapped Optimum"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            620,
+            224,
+            722,
+            273
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/aa583dca0754a84b1df2c39c7b1147b4db1387c0d69b19fcce8f8d91e9e9466b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            725,
+            98,
+            826,
+            220
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/3da42aa9860e70ce3b694a6054783d9f404c5a344355f108f4502a33f7839eda.jpg",
+        "image_caption": [
+            "Handcraft Optimum (Ours)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            727,
+            224,
+            831,
+            273
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Body Parameters Refinement",
+        "text_level": 1,
+        "bbox": [
+            76,
+            393,
+            339,
+            409
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "MoSh++ produces unnatural head shapes due to the fact that head markers were worn on a helmet and it sometimes produces unnatural finger poses. Consequently, we refine body shape and pose parameters with three simple yet effective hard-coded physical rules. i) The neck and head length should approximate 1/7 of the total length of the body. ii) Fingers, except for the thumb, should not bend backward. iii) We employ the Kolmogorov-Smirnov test, which reveals the data is similar to a normal distribution. Subsequently, we apply a Gaussian truncation approach, where all data points falling outside the  $3\\sigma$  range are adjusted to align with the  $3\\sigma$  threshold and blended with the adjacent 10 frames. We compare our refined body parameters (denoted as MoSh Refined) with Original BEAT, Retargeted SMPL-X, and MoSh SMPL-X in Figure 2.",
+        "bbox": [
+            75,
+            422,
+            472,
+            650
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Blendshape Weights to FLAME Parameters",
+        "text_level": 1,
+        "bbox": [
+            76,
+            675,
+            450,
+            691
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Given the ARKit blendshape weights  $\\mathbf{b}_{\\mathrm{ARKit}} \\in \\mathbb{R}^{T \\times 51}$ , we aim to derive a transformation matrix  $\\mathbf{W} \\in \\mathbb{R}^{51 \\times 103}$  that maps these into FLAME parameters  $\\mathbf{b}_{\\mathrm{FLAME}} \\in \\mathbb{R}^{T \\times (100 + 3)}$ , where 100 represents the number of dimensions for expression parameters and 3 for jaw movement. Due to discrepancies between the iPhone's and FLAME's template mesh topology, simply optimizing FLAME parameters by minimizing the geometric error between the ARKit template vertices and the wrapped FLAME vertices does not yield satisfactory results. Instead, we release a set of handcrafted blendshape templates  $\\mathbf{v}_{\\mathrm{template}} \\in \\mathbb{R}^{52}$  on FLAME following ARKit's FACS configuration. This approach allows direct driving of the FLAME topology ver",
+        "bbox": [
+            75,
+            703,
+            472,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "tices  $\\mathbf{v}\\in \\mathbb{R}^{T\\times 5023\\times 3}$  from the given blendshape weights,",
+        "bbox": [
+            498,
+            392,
+            879,
+            410
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {v} = \\mathbf {v} _ {\\text {t e m p l a t e}} ^ {0} + \\sum_ {j = 1} ^ {5 1} \\mathbf {b} _ {\\text {A R K i t}, j} \\cdot \\mathbf {v} _ {\\text {t e m p l a t e}} ^ {j} \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            571,
+            422,
+            890,
+            465
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{b}_{\\mathrm{ARKit},j}$  is the weight of the  $j$ -th ARKit blend-shape and  $\\mathbf{v}_{\\mathrm{template}}^j$  is the vertex position of the FLAME model template corresponding to the  $j$ -th blendshape. The term  $\\mathbf{v}_{\\mathrm{template}}^0$  represents the original template vertex position of the FLAME model. We optimize  $\\mathbf{W}$  by minimizing  $\\| \\tilde{\\mathbf{v}}_j - \\mathbf{v}_j\\| _2$ , where  $\\tilde{\\mathbf{v}}$  is obtained from FLAME's LBS  $\\mathcal{V}(\\mathbf{b}_{\\mathrm{FLAME}})$ . The comparisons between ARKit data, wrap-based, and our approach are shown in Figure 2.",
+        "bbox": [
+            498,
+            477,
+            890,
+            601
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. IMAGE",
+        "text_level": 1,
+        "bbox": [
+            500,
+            617,
+            598,
+            632
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We introduce the details of EMAGE, Expressive Masked Audio-Conditioned GEsture Modeling (see Figure 3). Given gestures  $\\mathbf{g} \\in \\mathbb{R}^{T \\times (55 \\times 6 + 100 + 4 + 3)}$ , representing 55 joints in Rot6D,  $\\mathbb{R}^{100}$  FLAME parameters,  $\\mathbb{R}^4$  foot contact labels,  $\\mathbb{R}^3$  global translations, and speech audio  $\\mathbf{a} \\in \\mathbb{R}^{T \\times sk}$ , where  $sk = sr_{\\mathrm{audio}} / fps_{\\mathrm{gestures}}$ , EMAGE jointly optimizes masked gesture reconstruction and audio-conditioned gesture generation. This optimization enhances performance during inference and enables the use of partially masked gestures to complete holistic gestures. To achieve this, we first model the quantized latent space in Sec. 4.1, following [61, 65]. We then design a separated speech audio encoder using Content Rhythm Self-Attention (CRA) as detailed in Sec. 4.2. Subsequently, we learn the body hints from masked gestures via the Masked Audio Gesture Transformer, explained in Sec. 4.3. Finally, the decoding strategy for different body segments is discussed in Sec. 4.4.",
+        "bbox": [
+            496,
+            643,
+            890,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "1147",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/03d3d4b036b0ca4357b992552ab291dfb862bda77305f32207eae0b1502e2b37.jpg",
+        "image_caption": [
+            "Figure 3. EMAGE leverages two training paths: Masked Gesture Reconstruction (MG2G) and Audio-Conditioned Gesture Generation (A2G). The MG2G path focuses on encoding robust body hints through a spatial-temporal transformer gesture encoder and cross-attention gesture decoder. In contrast, the A2G path utilizes these body hints, and separated audio encoders to decode pretrained face and body latent features. A key component in this process is a switchable cross-attention layer, crucial for merging body hints and audio features. This fusion allows the features to be effectively disentangled and utilized for gesture decoding. Once the gesture latent features are reconstructed, EMAGE employs a pretrained VQ-Decoder to decode face and local body motions. Additionally, a pretrained Global Motion Predictor is used to estimate global body translations, further enhancing the model's capability to generate realistic and coherent gestures."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            94,
+            867,
+            272
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/fcfbe7a0829c0edac6c9bb4d65277fe8bae5b40a7eca41f562d382b9fffccf6f.jpg",
+        "image_caption": [
+            "Figure 4. Details of CRA and Pretrained VQ-VAEs. Left: Content Rhythm Attention fuses speech rhythm (onset and amplitude) with content (pretrained word embeddings from text scripts) adaptively. This results in a preference for either content or rhythm in specific frames, which encourages the generation of semantical-aware gestures. Right: We pretrain four compositional VQ-VAEs by reconstructing face, upper body, hands and lower body separately to disentangle audio-agnostic gestures explicitly."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            151,
+            404,
+            271,
+            518
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/20c78e60af2f8a1e4a54cfc9be2acb3cffddc5a758e0ff797f181e0b8b1facf5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            276,
+            402,
+            395,
+            518
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Compositional Discrete Face and Body Prior",
+        "text_level": 1,
+        "bbox": [
+            76,
+            672,
+            454,
+            690
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We model full-body gestures in the separated quantized latent space (Figure 4) for several reasons. Similar to [65], the body and hands improves the diversity of the results. However, we additionally separate the face and lower body due to their differing correlations with audio, i.e., using a single VQ-VAE to encode both the upper and lower body may lead the model to overlook gestures that occur less frequently - regardless of their relation to the audio. Specifically, a single model might focus only on recovering lower body motion, neglecting elbow movements in a speaker who is constantly walking during the conversation.",
+        "bbox": [
+            75,
+            700,
+            467,
+            866
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The separated quantized latent space  $Q = \\{\\mathbf{q}_{\\mathrm{f}}, \\mathbf{q}_{u}, \\mathbf{q}_{\\mathrm{h}}, \\mathbf{q}_{\\mathrm{l}}\\}$  for the Face  $\\mathbf{g}_{\\mathrm{f}} \\in \\mathbb{R}^{T \\times 106}$ , Upper",
+        "bbox": [
+            76,
+            869,
+            468,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "body  $\\mathbf{g}_{\\mathrm{u}} \\in \\mathbb{R}^{T \\times 78}$ , Hands  $\\mathbf{g}_{\\mathrm{h}} \\in \\mathbb{R}^{T \\times 180}$ , and Lower body  $\\mathbf{g}_{\\mathrm{l}} \\in \\mathbb{R}^{T \\times (54 + 4)}$  are from four VQ-VAEs. The VQ-VAE for each is optimized by jointly optimizing the following loss terms,",
+        "bbox": [
+            498,
+            402,
+            890,
+            462
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nq _ {i} = \\underset {q _ {i} \\in \\mathbf {q}} {\\arg \\min } \\| z _ {j} - q _ {i} \\| ^ {2} \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            612,
+            464,
+            890,
+            491
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V Q - V A E}} = \\mathcal {L} _ {r e c} (\\mathbf {g}, \\hat {\\mathbf {g}}) + \\mathcal {L} _ {\\mathrm {v e l}} \\left(\\mathbf {g} ^ {\\prime}, \\hat {\\mathbf {g}} ^ {\\prime}\\right) + \\mathcal {L} _ {\\mathrm {a c c}} \\left(\\mathbf {g} ^ {\\prime \\prime}, \\hat {\\mathbf {g}} ^ {\\prime \\prime}\\right) \\\\ + \\left\\| \\operatorname {s g} [ \\mathbf {z} ] - \\mathbf {q} \\right\\| ^ {2} + \\left\\| \\mathbf {z} - \\operatorname {s g} [ \\mathbf {q} ] \\right\\| ^ {2}, \\tag {4} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            512,
+            890,
+            551
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{z}$  is the encoded  $\\mathbf{g}$  with a temporal window size  $w = 1$ .  $\\mathcal{L}_{\\mathrm{rec}}$ ,  $\\mathcal{L}_{\\mathrm{vel}}$ , and  $\\mathcal{L}_{\\mathrm{acc}}$  are Geodesic [58] and L1 losses, sg is a stop gradient operation, we set the weight of the commitment loss [59] (the last term) to 1 in this paper.",
+        "bbox": [
+            498,
+            563,
+            890,
+            625
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Content and Rhythm Self-Attention",
+        "text_level": 1,
+        "bbox": [
+            500,
+            633,
+            810,
+            650
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Given the speech audio, s, inspired by [4], we employ onset o and amplitude a as explicit audio rhythm, alongside the pretrained embeddings [9] e from transcripts as content. Different from previous approaches, which typically add the rhythm and content features, we leverage self-attention to merge these features adaptively. This approach is driven by the observation that for specific frames, the gestures are more related to content (semantic-aware) or rhythm (beat-aware). The rhythm and content features are first encoded into time-aligned features  $\\mathbf{r}_{1:T}$  and  $\\mathbf{c}_{1:T}$ , using a Temporal Convolutional Network (TCN) and linear mapping, respectively. For each time step  $t\\in \\{1,\\dots,T\\}$ , we merge the rhythm and content features by:",
+        "bbox": [
+            496,
+            657,
+            890,
+            854
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathbf {f} _ {1: T} = \\alpha \\times \\mathbf {r} _ {1: T} + (1 - \\alpha) \\times \\mathbf {c} _ {1: T}, \\tag {5} \\\\ \\alpha = \\operatorname {S o f t m a x} (\\mathcal {A T} (\\mathbf {r} _ {1: T}, \\mathbf {c} _ {1: T})), \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            867,
+            890,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "1148",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/b5ffedfff937a78fa40fca7cb43e4752e73f12773787d02e95ac5a687967d09f.jpg",
+        "image_caption": [
+            "Figure 5. Comparison of Forward Path Designs. Straightforward fusion module (a) merges audio features without refined body features and recombines audio features based only on position embedding. The Self-Attention decoder module (b), adopted in previous MLM models [17, 31], is limited for tasks requiring auto-regressive inference. Our design (c) considers effective audio feature fusion and auto-regressive decoding."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            84,
+            90,
+            207,
+            234
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/1cacf29fb5e6624bf438a25165fccad902e7718042aa87ac7732c73e36efd625.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            212,
+            90,
+            334,
+            234
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/32cfc4eb4ef2edd24fa8beeb6c6b37217aecf0976a8d36ece3aaf17f2d0afc5b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            338,
+            90,
+            454,
+            236
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{A}\\mathcal{T}$  is a 2-layer MLP. We apply two separate CRA encoders for the face and body.",
+        "bbox": [
+            75,
+            361,
+            468,
+            391
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Masked Audio-Conditioned Gesture Modeling",
+        "text_level": 1,
+        "bbox": [
+            76,
+            398,
+            467,
+            416
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We propose a Masked Audio Gesture Transformer to leverage different training paths (see Fig. 5 for the motivation behind the architecture designs). Given temporal and spatially masked gestures  $\\overline{\\mathbf{g}}\\in \\mathbb{R}^{T\\times 337}$ , we first replace the masked tokens with learned masked embeddings  $e_{\\mathrm{mask}}\\in \\mathbb{R}^{256}$ , as the value zero still represents specific motion content, e.g., a T-pose. We linearly increase the ratio of masked joints and frames from 0 to  $95\\%$  according to the training epochs.",
+        "bbox": [
+            75,
+            422,
+            468,
+            542
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A spatial convolutional  $\\mathcal{SC}$  encoder initially summarizes the spatial information, and compress the spatial feature into  $\\mathbb{R}^{T\\times 256}$ . We then employ a temporal self-attention (without feed-forward)  $\\mathcal{TSA}$  to refine the summarized spatial features",
+        "bbox": [
+            75,
+            544,
+            468,
+            616
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {h} = \\mathcal {T S A} (\\mathcal {S C} (\\overline {{\\mathbf {g}}}) + \\mathbf {p} _ {t}), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            186,
+            618,
+            468,
+            633
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{h} \\in \\mathbb{R}^{T \\times 512}$  represents encoded body hints, and  $\\mathbf{p}_t$  is the sum of a learned speaker embedding and PPE [19]. Then, a straightforward temporal cross-attention Transformer decoder  $\\mathcal{TCA}\\mathcal{T}$  is adopted for the reconstructed latent  $\\hat{\\mathbf{q}}_{\\mathrm{mg2g}}$ ,",
+        "bbox": [
+            75,
+            638,
+            468,
+            715
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {q}} _ {\\mathrm {m g} 2 \\mathrm {g}} = \\mathcal {T C A T} (\\mathbf {h} + \\mathbf {p} _ {t}). \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            184,
+            715,
+            468,
+            731
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We minimize the L1 distance in the latent space,",
+        "bbox": [
+            76,
+            736,
+            397,
+            751
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {m g} 2 \\mathrm {g}} = \\| \\hat {\\mathbf {q}} _ {\\mathrm {m g} 2 \\mathrm {g}} - \\mathbf {q} \\|. \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            196,
+            760,
+            468,
+            777
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Masked gesture reconstruction encodes effective body hints, with the key being to leverage these body hints for gesture generation. We employ a selective fusion for audio and body hints via a temporal cross-attention  $\\mathcal{TCA}$ . Subsequently, we use the merged audio-gesture features for audio-conditioned gesture latent reconstruction,",
+        "bbox": [
+            75,
+            785,
+            468,
+            875
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {q}} _ {\\mathrm {a} 2 \\mathrm {g}} = \\mathcal {T C A T} \\left(\\mathcal {T C A} \\left(\\mathbf {h} + \\mathbf {p} _ {t}, \\mathbf {f} _ {\\text {b o d y}}\\right), \\overline {{\\mathbf {g}}} + \\mathbf {p} _ {t}\\right). \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            119,
+            885,
+            468,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/1bd2e05a0ddf11f82d1922cf9aa70f0429d1d52a384c056486df192d8633e7ff.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Top.-B</td><td>Top.-F</td><td>Body</td><td>Face</td></tr><tr><td>VOCA[14]</td><td>-</td><td>FLAME</td><td>-</td><td>38.3 ± 5.63</td></tr><tr><td>AMASS[46]</td><td>SMPL-X</td><td>FLAME</td><td>42.0 ± 3.60</td><td>-</td></tr><tr><td>TalkSHOW[65]</td><td>SMPL-X</td><td>SMPL-X</td><td>14.4 ± 2.19</td><td>26.1 ± 6.42</td></tr><tr><td>BEAT2 (Ours)</td><td>SMPL-X</td><td>FLAME</td><td>43.6 ± 3.38</td><td>35.7 ± 5.91</td></tr></table>",
+        "bbox": [
+            522,
+            88,
+            867,
+            143
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 3. User Preference Win Rate between Existing Datasets. 'Top' denotes the topology of the mesh. The results show our BEAT2 dataset outperforms the existing PGT dataset [65] in both body and face aspects. It also performs slightly better on body and lower on face when compared with previous mocap [46] and 3D scan-based datasets [14], respectively.",
+        "bbox": [
+            498,
+            150,
+            890,
+            233
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We optimize both the latent code class classification cross-entropy loss  $\\mathcal{L}_{\\mathrm{a2g - rec}}$  and the latent reconstruction loss  $\\mathcal{L}_{\\mathrm{a2g - c1s}}$  to encourage diverse results.",
+        "bbox": [
+            498,
+            251,
+            890,
+            297
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4. Face and Translation Decoding",
+        "text_level": 1,
+        "bbox": [
+            500,
+            305,
+            774,
+            321
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Considering that the face is weakly related to body motion, applying the same operation - recombining the audio features based on body hints - is not reasonable. Therefore, we directly concatenate the masked body hints with the audio features for the final decoding of facial latent,",
+        "bbox": [
+            498,
+            328,
+            890,
+            402
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {q}} _ {\\mathrm {f}} = \\mathcal {T} \\mathcal {C A T} \\left(\\mathbf {f} _ {\\text {f a c e}} \\oplus \\mathbf {h}, \\mathbf {p} _ {t}\\right). \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            599,
+            412,
+            890,
+            429
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Once we have obtained the local lower body motion  $\\tilde{\\mathbf{g}}_l\\in$ $\\mathbb{R}^{T\\times (54 + 4)}$  from the VQ-Decoder, we estimate the global translations  $\\tilde{\\mathbf{t}}\\in \\mathbb{R}^{T\\times 3}$  with a pretrained Global Motion Predictor  $\\tilde{\\mathbf{t}} = \\mathcal{G}(\\tilde{\\mathbf{g}}_l)$ , which significantly reduces foot sliding.",
+        "bbox": [
+            498,
+            438,
+            890,
+            500
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            511,
+            630,
+            527
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We separate the evaluation into two categories: dataset quality and model ability. After removing data with low finger quality from BEAT, BEAT2 is reduced to 60 hours. We further split it into BEAT2-Standard (27 hours) and BEAT2-Additional (33 hours), based on the type of speech and conversation sections [39]. While acted gestures in speech sections are more diverse and expressive, spontaneous gestures in conversational sections tend to be more natural yet less varied. We report the results on BEAT2-Standard Speaker2 with an  $85\\% / 7.5\\% / 7.5\\%$  train/val/test split.",
+        "bbox": [
+            496,
+            537,
+            890,
+            688
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Dataset Quality Evaluation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            696,
+            743,
+            712
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We compare our dataset with the state-of-the-art pseudo ground truth (PGT) dataset, TalkSHOW [65], for both face and body. Additionally, we compare it with the AMASS [46] dataset for the body and VOCA [14] for the face as references. Due to the varying lengths of sequences in each dataset, we sample 100 comparison pairs, each with an equal duration ranging from 2 to 4 seconds. In a perceptual study, each participant evaluates a random set of 40 pairs during a 10-minute session by selecting the sequence they consider to have the best captured quality. In total, 60 participants were invited. It is important to note that participants are instructed to compare only the upper body results,",
+        "bbox": [
+            496,
+            719,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "1149",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "... In my spare time ... ... if I feel ok ...",
+        "bbox": [
+            137,
+            112,
+            380,
+            126
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "... I will go for a walk or hike in nature. I try to organize something",
+        "bbox": [
+            413,
+            112,
+            830,
+            127
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2c125761dd9cae7c4576e912ae7f448a11376383f3058ef27dd260925960ebc0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            135,
+            130,
+            256,
+            199
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/cdb9bcf28bebd72500a014eecdaa75e44c0ee0d34d340bf78adbc004c8817654.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            277,
+            130,
+            395,
+            199
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/bae710f5cf209e06c839b96c23eb6e0f4e8f3ae177079db6f6be7e4871ae7a24.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            418,
+            130,
+            539,
+            199
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/c1c57afefb94e35ad11b6de08d396967a3fa5faca4bb4d28920f253113b08c23.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            560,
+            131,
+            679,
+            199
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/428f0f0053095cf331f75a2ee392f3922abd1e4ccefa68d335d26cba72ee5c79.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            707,
+            132,
+            825,
+            199
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/0e571774cfd0d6a43f3e034ffbc64ca28c24d7885bf47b0ac76ba27000ba1cb3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            130,
+            200,
+            271,
+            273
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/6ac6fc9288d18bedb3545a7e62f6007ef4f7c006b15af8b108bf094d1ad7880d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            284,
+            200,
+            395,
+            272
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/5f6862fca91f66b90bd24f08769f95c06dfa042121411312abe32d886c6e6098.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            418,
+            200,
+            539,
+            273
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/331dc7ed8f3b856ee1e11e203e33219dc639d6472ce45416330386342a33996a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            560,
+            200,
+            676,
+            273
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/fd3a413577dc8a0433bca6a0d076bf3a615e42cf9e70297b2b0c83b45ad34d34.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            202,
+            818,
+            273
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d5f6d5f60c872172245a54c48b4f67219ea87e02dbdecbd6b9a8a9f105928657.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            137,
+            276,
+            251,
+            345
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/50f47b1881a4f057c10224c1452d9152a3706b0b2443921ac8022fe89d6e58c5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            281,
+            275,
+            390,
+            345
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/783ca256b44282fda31bf3021b3fb18df2884846cbacbae94c1c54303861afca.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            416,
+            275,
+            535,
+            345
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7e6a2efeb4f1b1f039e4b4103e35d22570d17ac002aa7b76b5737e7ad19b9a58.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            549,
+            273,
+            676,
+            345
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/3b5bb70b6c5c3fbf791f74724eb0bfd3ce4c78419828d7f4465313421d034016.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            694,
+            273,
+            825,
+            347
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/6870b2a16ab908948f4c9a29bfe7b7a29afcad6abcba862bc36131580f62c209.jpg",
+        "image_caption": [
+            "Figure 6. Subjective Examples. Top to down: GroundTruth, Generated W/O Body Hints, Generated With Body Hints, Visible Body Hints, EMAGE generates diverse, semantic-aware and audio synchronized body gestures, e.g. raise the both hands for \"spare time\", relaxing gestures for \"hike in nature\". Besides, as shown in the third and bottom rows, EMAGE is flexible to accept non-audio-synchronized body hints on any frames and joints, to guide or customize the generated gestures explicitly, e.g., repeat a similar motion for raising hands, change the walking orientations, etc. Note that in this figure, The gray joints of generated results are not copies of visible hints."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            132,
+            348,
+            259,
+            415
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/6d0dd81e175b8e70e7d2cbaa3207c1c29eec10056fee93189c6f6b8833554f74.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            277,
+            348,
+            395,
+            415
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/fd6f3eee973bb1ecc58436590341bf1e98b4a7419d4e84fc2f7ed5c9472f580e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            416,
+            347,
+            535,
+            415
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/fe89c02a59451a9e1f4d73c03f7d1acf521b5838104d33ec8fb5dc3acbf675a8.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            550,
+            345,
+            663,
+            421
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/87fdaad99801ac186d6b6a5c6f959383750e13087ad63f5e0cb766c482d4920a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            692,
+            348,
+            831,
+            420
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/faacea25f48cb1d51502542f3db4f1848e16a4bbb87af4b0c969e82300ef07f1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>FGD ↓</td><td>BC ↑</td><td>Diversity ↑</td><td>MSE ↓</td><td>LVD ↓</td></tr><tr><td>FaceFormer[19]</td><td>-</td><td>-</td><td>-</td><td>7.787</td><td>7.593</td></tr><tr><td>CodeTalker[61]</td><td>-</td><td>-</td><td>-</td><td>8.026</td><td>7.766</td></tr><tr><td>S2G[25]</td><td>28.15</td><td>4.683</td><td>5.971</td><td>-</td><td>-</td></tr><tr><td>Trimodal[67]</td><td>12.41</td><td>5.933</td><td>7.724</td><td>-</td><td>-</td></tr><tr><td>HA2G[40]</td><td>12.32</td><td>6.779</td><td>8.626</td><td>-</td><td>-</td></tr><tr><td>DisCo[38]</td><td>9.417</td><td>6.439</td><td>9.912</td><td>-</td><td>-</td></tr><tr><td>CaMN[39]</td><td>6.644</td><td>6.769</td><td>10.86</td><td>-</td><td>-</td></tr><tr><td>DiffStyleGesture[62]</td><td>8.811</td><td>7.241</td><td>11.49</td><td>-</td><td>-</td></tr><tr><td>Habibie et al.[28]</td><td>9.040</td><td>7.716</td><td>8.213</td><td>8.614</td><td>8.043</td></tr><tr><td>TalkSHOW[65]</td><td>6.209</td><td>6.947</td><td>13.47</td><td>7.791</td><td>7.771</td></tr><tr><td>EMAGE (Ours)</td><td>5.512</td><td>7.724</td><td>13.06</td><td>7.680</td><td>7.556</td></tr></table>",
+        "bbox": [
+            81,
+            517,
+            465,
+            661
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "as TalkSHOW contains only the upper body. The results are shown in Table 3.",
+        "bbox": [
+            76,
+            753,
+            468,
+            784
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2. Model Ability Evaluation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            795,
+            307,
+            811
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We report results using BEATv1.3 and BEATv2. The latter has facial data refined by animators and hand data filtered by annotators.",
+        "bbox": [
+            76,
+            819,
+            468,
+            864
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Metrics. We adopt FGD[67] to evaluate the realism of the body gestures. Then, we measure Diversity[33] by calcu",
+        "bbox": [
+            76,
+            869,
+            468,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/897c87d22557e913147a4e036f9b9610a58050783ae0433b85d3687dcc096855.jpg",
+        "table_caption": [
+            "Table 4. Quantitative evaluation on BEATv2. We report FGD  $\\times 10^{-1}$ , BC  $\\times 10^{-1}$ , Diversity, MSE  $\\times 10^{-8}$ , and LVD  $\\times 10^{-5}$ . For body gestures, EMAGE significantly improves the FGD, indicating that the generated results are closer to the GT. This shows the effectiveness of body hints from masked gesture modeling."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Holistic</td><td>Body</td><td>Face</td></tr><tr><td>Habibie et al.[28]</td><td>12.4 ± 3.70</td><td>15.9 ± 6.49</td><td>10.8 ± 3.19</td></tr><tr><td>TalkSHOW[65]</td><td>34.9 ± 5.79</td><td>40.4 ± 8.22</td><td>33.2 ± 6.03</td></tr><tr><td>EMAGE (Ours)</td><td>52.7 ± 7.91</td><td>44.7 ± 8.68</td><td>56.0 ± 7.80</td></tr></table>",
+        "bbox": [
+            531,
+            518,
+            859,
+            566
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 5. User Preference Win Rate On Generated Results. The results indicate that our generated outcomes are perceived as more realistic and believable, with a  $14\\%$  and  $23\\%$  higher user preference for body and face gestures, respectively.",
+        "bbox": [
+            498,
+            571,
+            890,
+            628
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "lating the average L1 distance between multiple body gesture clips, and use BC [34] to assess speech-motion synchrony. For the face, we calculate the vertex MSE[61] to measure the positional distance and the vertex L1 difference LVD [65] between the GT and generated facial vertices.",
+        "bbox": [
+            498,
+            670,
+            890,
+            746
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Compared Methods. We first compare our method with representative state-of-the-art methods in body gesture generation [25, 38-40, 62, 67] and talking head generation [19, 61] by reproducing their methods for body and face, respectively. In addition, we reproduce two previous state-of-the-art holistic pipelines, Habibie et al. [28] and TalkSHOW [65], whose original implementations are limited to the upper body. We add a lower-body decoder to Habibie et al. and a lower-body VQ-VAE to TalkSHOW.",
+        "bbox": [
+            496,
+            772,
+            892,
+            907
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "1150",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/ec4acad0cc10bb45d1f30e78a3cac08c72cd08e88bef35a1db35c54e8a1d34f8.jpg",
+        "image_caption": [
+            "Figure 7. Comparison of Generated Facial Movements. Compared with previous state-of-the-art talking face generation methods FaceFormer [19] and CodeTalker [61] as well as holistic gesture generation methods Habibie et al. [28] and TalkSHOW [65]. Note that CodeTalker has higher MSE than IMAGE on BEATv2 (Table 4, lower is better) but is subjectively realistic. IMAGE gets good lip motions by leveraging both the face model and masked body hints."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            87,
+            436,
+            256
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2.1 Quantitative and Qualitative Analysis",
+        "text_level": 1,
+        "bbox": [
+            76,
+            383,
+            393,
+            398
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 4, with a four-frame seed pose, our method outperforms previous SOTA algorithms. For qualitative results, see Figures 6 and 7. Furthermore, we conduct a perceptual study. Maintaining the same setup of 60 participants, each participant evaluates 40 pairs of 10-second results do decide which is most believable; this gives the win rate shown in Table 5.",
+        "bbox": [
+            75,
+            401,
+            470,
+            506
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2.2 Ablation Analysis",
+        "text_level": 1,
+        "bbox": [
+            76,
+            527,
+            256,
+            542
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Performance of Baseline. As shown in Table 6, we start with a teacher-force transformer-based baseline. This baseline, inspired by FaceFormer [19], adopts a 1-layer cross-attention transformer decoder and replaces the audio features from Wav2Vec2 [6] with our custom TCN [7] and trainable word embeddings [9].",
+        "bbox": [
+            75,
+            551,
+            468,
+            643
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of Multiple VQ-VAEs. Simply applying one VQ-VAE [27, 59] for full-body movements, including the face, decreases performance in facial movements. This is because the VQ-VAE is trained to minimize the average loss for the full body, and some speakers' most frequent movements are not related to the audio. Implementing separated VQ-VAEs allows the model to better leverage the advantages of discrete priors.",
+        "bbox": [
+            75,
+            643,
+            470,
+            763
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of Content Rhythm Self-Attention. Adaptive fusion of rhythm and content features shows improvement in both FGD and Alignment. It selectively focuses the current motion more on rhythm or content features based on the training data's distribution. Moreover, we observed more semantic-aware results when applying CRA.",
+        "bbox": [
+            75,
+            763,
+            468,
+            854
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of Masked Body Gesture Hints. The improvement across all objective metrics demonstrates that our model effectively leverages spatial-temporal gesture priors, reducing",
+        "bbox": [
+            75,
+            854,
+            470,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/b5e3474bf149b65160747babbc2213f218c7780bfb8d61b4f14ff101b36b7dd3.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>FGD ↓</td><td>BC ↑</td><td>Diversity ↑</td><td>MSE ↓</td><td>LVD ↓</td></tr><tr><td>Ground Truth</td><td>0</td><td>6.896</td><td>13.074</td><td>0</td><td>0</td></tr><tr><td>Reconstruction</td><td>3.913</td><td>6.758</td><td>13.145</td><td>0.841</td><td>6.389</td></tr><tr><td>Baseline</td><td>13.080</td><td>6.941</td><td>8.3145</td><td>1.442</td><td>9.317</td></tr><tr><td>+ VQVAE</td><td>9.787</td><td>6.673</td><td>10.624</td><td>1.619</td><td>9.473</td></tr><tr><td>+ 4 VQVAE</td><td>7.397</td><td>6.698</td><td>12.544</td><td>1.243</td><td>8.938</td></tr><tr><td>+ CRA</td><td>6.833</td><td>6.783</td><td>12.676</td><td>1.186</td><td>8.715</td></tr><tr><td>+ Masked Hints</td><td>5.423</td><td>6.794</td><td>13.057</td><td>1.180</td><td>9.015</td></tr></table>",
+        "bbox": [
+            506,
+            87,
+            885,
+            193
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/3082b043011fb1ad1fe944210d5da70299672d88000c5ade9d11baf995a271b7.jpg",
+        "table_caption": [
+            "Table 6. Abliation Analysis on BEATv1.3"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>body</td><td>hands</td><td>face</td><td>FGD ↓</td><td>BC ↑</td><td>Diversity ↑</td></tr><tr><td>BESTX</td><td>✓</td><td>✓</td><td>✓</td><td>5.423</td><td>6.794</td><td>13.057</td></tr><tr><td>+ Trinity[22]</td><td>✓</td><td></td><td></td><td>5.319</td><td>6.843</td><td>13.346</td></tr><tr><td>+ AMASS[46]</td><td>✓</td><td>✓</td><td></td><td>5.174</td><td>6.769</td><td>14.318</td></tr></table>",
+        "bbox": [
+            521,
+            231,
+            870,
+            281
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 7. Training IMAGE on Multiple Datasets. IMAGE demonstrates flexibility by training on multiple datasets, even when only a subset of holistic gestures is available. This approach further improves the objective metrics on the BEATv1.3 test set.",
+        "bbox": [
+            498,
+            285,
+            890,
+            340
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the likelihood of incorrect gesture sampling. Furthermore, masked gesture modeling is key to enabling the network to accept predefined gestures in specific frames.",
+        "bbox": [
+            498,
+            362,
+            890,
+            409
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2.3 Ability of Multi-Dataset Training.",
+        "text_level": 1,
+        "bbox": [
+            500,
+            431,
+            789,
+            448
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We demonstrate that EMAGE can effectively combine multiple non-holistic datasets for training by jointly training with the Trinity [22] and AMASS [46] datasets, by using only the upper body and audio pairs from Trinity and only body and hands from AMASS. Separate VQ-VAEs are trained for BEAT2, Trinity, and AMASS, with separate MLP heads implemented for codebook classification. The results in Table 7 show that incorporating data improves performance on the BEATv1.3 test set.",
+        "bbox": [
+            496,
+            460,
+            890,
+            597
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            612,
+            617,
+            628
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this work, we present IMAGE, a framework designed to accept partial gestures as input for completing audio-synchronized holistic gestures. It demonstrates that leveraging masked gesture reconstruction can significantly enhance the performance of audio-conditioned gesture generation. Furthermore, the design of IMAGE enables the training on multiple datasets, further improving performance. Alongside IMAGE, we release BEAT2, the largest multi-modal gesture dataset consistent with SMPL-X and FLAME. We hope that BEAT2 will contribute to knowledge and model sharing across various subareas.",
+        "bbox": [
+            496,
+            641,
+            890,
+            806
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Disclosures: While You Zhou, Xuefei Zhe, Naoya Iwamoto, and Bo Zheng are employees of Huawei Tokyo Research Center, this work was done on their own time with the approval of their employer. MJB CoI: https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt",
+        "bbox": [
+            496,
+            809,
+            890,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "1151",
+        "bbox": [
+            483,
+            944,
+            513,
+            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] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2",
+            "[2] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In European Conference on Computer Vision, pages 248–265. Springer, 2020. 2",
+            "[3] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, pages 487-496. Wiley Online Library, 2020. 2",
+            "[4] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffu-clip: Gesture diffusion model with clip latents. ACM Trans. Graph. 2, 5",
+            "[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 2, 3",
+            "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 8",
+            "[7] Shaojie Bai, J Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arxiv 2018. arXiv preprint arXiv:1803.01271, 2, 1803. 8",
+            "[8] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 2, 3",
+            "[9] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146, 2017. 5, 8",
+            "[10] Biovision BVH. Biovision bvh. In https://research.cs.wisc.edu/graphics /Courses/ cs-838-1999/Jeff/BVH.html, 1999. 3",
+            "[11] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: realtime multi-person 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence, 43(1):172-186, 2019. 2",
+            "[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 3",
+            "[13] Kiran Chhatre, Radek Danček, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, and Timo Bolkart. Emotional speech-driven 3d body animation via disentangled latent diffusion. arXiv preprint arXiv:2312.04466, 2023. 1"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[14] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael Black. Capture, learning, and synthesis of 3D speaking styles. Computer Vision and Pattern Recognition (CVPR), pages 10101-10111, 2019. 2, 3, 6",
+            "[15] Radek Dančecek, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. ACM, 2023. 1",
+            "[16] Anna Deichler, Kiran Chhatre, Christopher Peters, and Jonas Beskow. Spatio-temporal priors in 3d human motion. 2021. 2",
+            "[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3, 6",
+            "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3",
+            "[19] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8",
+            "[20] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 1",
+            "[21] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia, 12(6):591-598, 2010. 2",
+            "[22] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 93-98, 2018. 3, 8",
+            "[23] Erik Gartner, Mykhaylo Andriluka, Erwin Coumans, and Cristian Sminchisescu. Differentiable dynamics for articulated 3d human motion reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13190-13200, 2022. 2",
+            "[24] Saeed Ghorbani, Ylva Ferstl, Daniel Holden, Nikolaus F. Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. Computer Graphics Forum, 42(1):206-216, 2023. 2, 3",
+            "[25] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 1, 2, 3, 7",
+            "[26] Yuan Gong, Yu-An Chung, and James Glass. Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778, 2021.3"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "1152",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[27] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 8, 4",
+            "[28] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. arXiv preprint arXiv:2102.06837, 2021. 1, 2, 3, 7, 8",
+            "[29] 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. 3",
+            "[30] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 2",
+            "[31] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 3, 6",
+            "[32] Gilwoo Lee, Zhiwei Deng, Shugao Ma, Takaaki Shiratori, Siddhartha S Srinivasa, and Yaser Sheikh. Talking with hands  $16.2\\mathrm{m}$ : A large-scale dataset of synchronized body-finger motion and audio for conversational motion analysis and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 763-772, 2019. 2, 3",
+            "[33] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 1, 2, 7, 3",
+            "[34] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 7, 2",
+            "[35] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2, 3",
+            "[36] Haiyang Liu and Cheng Zhang. Reinforcement learning based neural architecture search for audio tagging. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2020. 3",
+            "[37] Haiyang Liu and Jihan Zhang. Improving ultrasound tongue image reconstruction from lip images using self-supervised learning and attention mechanism. arXiv preprint arXiv:2106.11769, 2021. 3",
+            "[38] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM Interna-"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "tional Conference on Multimedia, pages 3764-3773, 2022. 2,3,7",
+            "[39] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. arXiv preprint arXiv:2203.05297, 2022. 1, 2, 3, 4, 6, 7",
+            "[40] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2, 3, 7",
+            "[41] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 1, 2",
+            "[42] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3",
+            "[43] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3",
+            "[44] Matthew M. Loper, Naureen Mahmood, and Michael J. Black. MoSh: Motion and shape capture from sparse markers. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 33(6):220:1-220:13, 2014. 2",
+            "[45] Shuhong Lu, Youngwoo Yoon, and Andrew Feng. Co-speech gesture synthesis using discrete gesture token learning. arXiv preprint arXiv:2303.12822, 2023. 2, 3",
+            "[46] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, 2019. 2, 3, 6, 8",
+            "[47] Tim Massey. Vicon study of dynamic object tracking accuracy. In https://www.vicon.com/resources/blog/vicon-study-of-dynamic-object-tracking-accuracy/, 2020. 2, 3",
+            "[48] Evonne Ng, Shiry Ginosar, Trevor Darrell, and Hanbyul Joo. Body2hands: Learning to infer 3d hands from conversational gesture body dynamics. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11865-11874, 2021. 2, 3",
+            "[49] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. Bodyformer: Semantics-guided 3d body gesture synthesis with transformer. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 3",
+            "[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face,"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "1153",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3, 4",
+            "[51] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2",
+            "[52] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2",
+            "[53] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. Emotiongesture: Audio-driven diverse emotional co-speech 3d gesture generation. arXiv preprint arXiv:2305.18891, 2023. 2",
+            "[54] Xingqun Qi, Jiahao Pan, Peng Li, Ruibin Yuan, Xiaowei Chi, Mengfei Li, Wenhan Luo, Wei Xue, Shanghang Zhang, Qifeng Liu, et al. Weakly-supervised emotion transition learning for diverse 3d co-speech gesture generation. arXiv preprint arXiv:2311.17532, 2023. 2",
+            "[55] Guocheng Qian, Xingdi Zhang, Abdullah Hamdi, and Bernard Ghanem. Pix4point: Image pretrained transformers for 3d point cloud understanding. 2022. 3",
+            "[56] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 2, 3",
+            "[57] Kaede Shiohara, Xingchao Yang, and Takafumi Taketomi. Blendface: Re-designing identity encoders for face-swapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7634-7644, 2023. 2",
+            "[58] Tommi Tykkälä, Cédric Audras, and Andrew I Comport. Direct iterative closest point for real-time visual odometry. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 2050-2056. IEEE, 2011. 5",
+            "[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5, 8",
+            "[60] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, 2020. 2",
+            "[61] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 1, 2, 3, 4, 7, 8",
+            "[62] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven co-speech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2, 3, 7"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[63] Xingchao Yang and Takafumi Taketomi. Bareskinnet: Demakeup and de-lighting via 3d face reconstruction. In Computer Graphics Forum, pages 623-634. Wiley Online Library, 2022. 2",
+            "[64] Xingchao Yang, Takafumi Taketomi, and Yoshihiro Kanamori. Makeup extraction of 3d representation via illumination-aware image decomposition. In Computer Graphics Forum, pages 293-307. Wiley Online Library, 2023. 2",
+            "[65] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8",
+            "[66] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 3",
+            "[67] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 1, 2, 3, 7",
+            "[68] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 3",
+            "[69] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 3",
+            "[70] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20807-20817, 2023. 2, 3",
+            "[71] Yang Zhou, Jimei Yang, Dingzeyu Li, Jun Saito, Deepali Aneja, and Evangelos Kalogerakis. Audio-driven neural gesture reenactment with video motion graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3418-3428, 2022. 2",
+            "[72] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3",
+            "[73] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            835
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "1154",
+        "bbox": [
+            483,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/f8a85da7-9221-4a37-baf0-e5a3b46705e3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63c7969f18cbc47e99af485c52313d8c199c32c0e5487a5241335591e7d46854
+size 4858502

2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/full.md

@@ -0,0 +1,394 @@
+# EMAGE: Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling
+
+Haiyang Liu $^{1*}$  Zihao Zhu $^{2*}$  Giorgio Becherini $^{4}$  Yichen Peng $^{5}$  Mingyang Su $^{6}$  You Zhou $^{3}$  Xuefei Zhe $^{3}$  Naoya Iwamoto $^{3}$  Bo Zheng $^{3}$  Michael J. Black $^{3}$
+
+$^{1}$ The University of Tokyo  $^{2}$ Keio University  $^{3}$ Independent Researcher  $^{4}$ Max Planck Institute for Intelligent Systems  $^{5}$ JAIST  $^{6}$ Tsinghua University
+
+![](images/77aa6d75a5af7f91c15f89f767785f7436b00b41c1d6074cbaa3e215dafb7c19.jpg)  
+Figure 1. EMAGE. We present a Masked Audio-Conditioned Gesture Modeling framework, along with a new holistic gesture dataset, BEAT2 (BEAT-SMPLX-FLAME), for jointly generating facial expressions, local body dynamics, hand movements and global translations, conditioned on audio and a partially or completely masked gestures. The gray denotes visible gestures, and blue represents our outputs.
+
+# Abstract
+
+We propose EMAGE, a framework to generate full-body human gestures from audio and masked gestures, encompassing facial, local body, hands, and global movements. To achieve this, we first introduce BEAT2 (BEATSMLPX-FLAME), a new mesh-level holistic co-speech dataset. BEAT2 combines a MoShed SMPL-X body with FLAME head parameters and further refines the modeling of head, neck, and finger movements, offering a community-standardized, high-quality 3D motion captured dataset. EMAGE leverages masked body gesture priors during training to boost inference performance. It involves a Masked Audio Gesture Transformer, facilitating joint training on audio-to-gesture generation and masked gesture reconstruction to effectively encode audio and body gesture hints. Encoded body hints from masked gestures are then separately employed to generate facial and body movements. Moreover, EMAGE adaptively merges speech features from the audio's rhythm and content and utilizes four compositional VQ-VAEs to enhance the results' fidelity and diversity. Experiments demonstrate that EMAGE generates
+
+holistic gestures with state-of-the-art performance and is flexible in accepting predefined spatial-temporal gesture inputs, generating complete, audio-synchronized results. Our code and dataset are available.<sup>1</sup>
+
+# 1. Introduction
+
+The path towards full-body co-speech gesture generation presents many remaining challenges. Despite the development of various baselines, e.g., audio conditioned models for face [15, 20, 61], body [13, 25, 67], and hand movements [33, 39, 41], along with a few attempts at merged models [28, 65], the limited availability of comprehensive data and models poses an ongoing obstacle. To make progress in this direction, we present a new framework that can incorporate partial spatial-temporal predefined gestures and autonomously complete the remaining frames in synchronization with the audio. This provides a new tool for creating realistic animations of digital humans
+
+[57, 63, 64, 71].  
+To achieve this, we first require a comprehensive gesture dataset. While several datasets are available for audio-to-body [67], audio-to-body&face [28], body-to-hands [39, 41], and common actions [16, 46], integrating them poses challenges due to varying data formats and evaluation metrics across different sub-areas. This motivates us to establish a unified, community-standard benchmark for training and evaluating holistic gestures. We build upon the BEAT dataset [39], which provides the most comprehensive mocap co-speech gesture dataset with diverse modalities. Despite its size and richness, the data in BEAT is not presented in a standardized format. The dataset comprises iPhone ARKit blendshape weights and Vicon skeletons, posing challenges when using the data for training; e.g., the lack of a mesh representation prevents the use of a vertex loss [19, 52]. Additionally, conducting direct human perceptual evaluations on these diverse modalities is inherently complex. On the other hand, SMPL-X [50] and FLAME [35] are common mesh standards for the academic community for other human dynamics-related tasks, e.g., action [52] and talking head generation [14]. Driven by the motivation to facilitate the sharing of knowledge across different tasks, we present BEAT-SMPLX-FLAME (BEAT2), which consists of two main components: i) Refined SMPL-X body shape and pose parameters from MoSh++ [44, 46] with hard-coded physical priors. ii) Optimized high-quality FLAME head parameters. By integrating SMPL-X's body and FLAME's head, BEAT2 facilitates a comprehensive training and evaluation benchmark across multiple sub-areas of co-speech human animation generation.  
+With the holistic data from BEAT2, we aim to enhance the coherence of individual body parts while ensuring accurate cross-modal alignment between audio and motion; see Figure 1. This leads to the design of Expressive Masked Audio-conditioned GEsture modeling, denoted as EMAGE, a spatial and temporal transformer-based framework. EMAGE first aggregates spatial features from masked body joints. Then, it reconstructs the latent space of pretrained gestures with a switchable gesture temporal self-attention and audio-gesture cross-attention. The selection of different forward paths enables modeling the gesture-to-gesture and audio-to-gesture prior separately and effectively. Once the reconstructed latent features are obtained, EMAGE decodes local face and body gestures from four compositional pretrained Vector Quantized-Variational AutoEncoders (VQ-VAEs) and decodes global translations from a pretrained Global Motion Predictor. The compositional VQ-VAEs are key to preserving audio-related motions during the reconstruction. Through these designs, EMAGE achieves state-of-the-art performance in generating body and face gestures. It takes as input audio and partially initialized gestures, and recovers audio
+
+synchronized, coherent gestures. Additionally, we show how EMAGE can flexibly incorporate additional, nonholistic, datasets to improve results.  
+Overall, our contributions are as follows: (1) We release BEAT2, a representation-unified, mesh-level dataset that leverages MoShed SMPL-X body and FLAME head parameters. (2) We propose EMAGE, a simple yet effective holistic gesture generation framework that generates coherent gestures with partial gesture and audio priors. (3) EMAGE achieves state-of-the-art performance in generating both body and face gestures with only four frames of seed gestures. (4) EMAGE showcases the use of additional non-holistic gesture datasets for training, e.g., Trinity and AMASS, to further improve the fidelity and diversity of the results, effectively leveraging data from different datasets.
+
+# 2. Related Work
+
+Co-Speech Animation Datasets are categorized into two types; see Table 1: pseudo-labeled (PGT) and motion-captured (mocap). For PGT, the Speech2Gesture Dataset [25] utilizes OpenPose [11] to extract 2D poses from News and Teaching videos. Subsequent works extend the dataset to 3D pose [28] and SMPL-X [65]. The TED Dataset [66], which extracts 2D pose from TED videos, has been extended to include 3D pose [67], fingers [40], and SMPL-X [45]. Similarly, 2D, 3D landmarks, and meshes are estimated from multi-view recorded videos [60] as PGT for talking-head generation. Although pseudo-labeled approaches could theoretically extract infinite data, their accuracy is limited. For instance, the average error (in mm) for the SOTA monocular 3D pose estimation algorithm [23] on the Human3.6M dataset [30] is 33.4, while for Vicon mocap it is 0.142 [47]. For mocap datasets, Trinity [3] features one male actor with 4h of data, and TalkingWithHands [32] collects conversation scenarios for two speakers. ZEEG [24] considers one speaker with 12 styles. For the face, 3D scanning is accurate but limited in quantity due to cost, e.g., less than 3h for BIWI [21], VOCA [14], and MeshTalk [56]. This leads to a balance between performance and quantity in the ARKit dataset [51]. The aforementioned datasets address the body and face separately. On the other hand, BEAT [39] includes both 3D pose body and ARKit [8] facial blendshape weights for the first time. However, BEAT lacks mesh data.  
+Co-Speech Animation Models are categorized based on their outputs; see Table 2. In addition to the baselines and datasets [14, 25, 32, 48, 56, 67], several frameworks for body gestures [2, 33, 53, 54] have improved the performance of the baselines. They are trained and evaluated by selecting specific upper body joints [4, 5, 38-40, 49, 70], or all body joints [62, 72]. Recent improvements in facial gestures drive the vertices with transformers and discrete face priors [19, 61]. However, these methods only address ei
+
+<table><tr><td></td><td>Trinity Mocap 2017 [22]</td><td>S2G S2G-2D 2019 [25]</td><td>Seq2Seq TED-2D 2019 [66]</td><td>TWH Mocap 2019 [32]</td><td>Trimodal TED-3D 2020 [67]</td><td>VOCA Scan 2020 [14]</td><td>MeshTalk Scan 2021 [56]</td><td>Hahibie et al. S2G-3D 2021 [28]</td><td>HA2G TED-3D+ 2022 [40]</td><td>BEAT Mocap 2022 [39]</td><td>ZEEG Mocap 2022 [24]</td><td>Yoon. et al. TED-SMPL 2023 [45]</td><td>Talkshow S2G-SMPL 2023 [65]</td><td>BEAT2 Ours 2024</td></tr><tr><td>Head</td><td>-</td><td>2D</td><td>-</td><td>-</td><td>-</td><td>3D Scan</td><td>3D Scan</td><td>3D</td><td>-</td><td>ARKit</td><td>-</td><td>-</td><td>PGT-Mesh</td><td>MC-Mesh</td></tr><tr><td>Upper Body</td><td>3D</td><td>2D</td><td>2D</td><td>3D</td><td>3D</td><td>-</td><td>-</td><td>3D</td><td>3D</td><td>3D</td><td>3D</td><td>PGT-Mesh</td><td>PGT-Mesh</td><td>MC-Mesh</td></tr><tr><td>Hands</td><td>3D</td><td>2D</td><td>-</td><td>3D</td><td>-</td><td>-</td><td>-</td><td>3D</td><td>3D</td><td>3D</td><td>3D</td><td>PGT-Mesh</td><td>PGT-Mesh</td><td>MC-Mesh</td></tr><tr><td>Lower Body</td><td>3D</td><td>-</td><td>-</td><td>3D</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>3D</td><td>3D</td><td>-</td><td>-</td><td>MC-Mesh</td></tr><tr><td>Global Motion</td><td>3D</td><td>-</td><td>-</td><td>3D</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>3D</td><td>3D</td><td>-</td><td>-</td><td>MC-Mesh</td></tr><tr><td>Duration (hours)</td><td>4</td><td>60</td><td>97</td><td>20</td><td>97</td><td>0.5</td><td>2</td><td>38</td><td>33</td><td>76</td><td>4</td><td>30</td><td>27</td><td>60</td></tr></table>
+
+Table 1. Comparison of Co-Speech Gesture Datasets. We summarize gesture and face datasets in talking scenarios. 'PGT' and 'MC' denote Pseudo Ground Truth and Motion Capture, respectively. The BEAT2 (Ours) dataset is the largest dataset for motion-captured data, providing holistic, academic community standard mesh-level information.  
+
+<table><tr><td rowspan="2"></td><td>S2G</td><td>TriModal</td><td>B2H</td><td>TWH</td><td>Hahibie et al.</td><td>DisCo</td><td>HA2G</td><td>Face Former 2022[19]</td><td>Rhythmic</td><td>BEAT CaMN 2022[39]</td><td>Diff Gesture 2023[72]</td><td>Code Talker 2023[61]</td><td>Talk Show 2023[65]</td><td>DiffStyle Gesture 2023[62]</td><td>Body Former 2023[49]</td><td>Lively Speaker 2023[70]</td><td>EMAEG (Ours) 2024</td></tr><tr><td>2019 [25]</td><td>2020 [67]</td><td>2020[48]</td><td>2021[32]</td><td>2021[28]</td><td>2022[38]</td><td>2022[40]</td><td>2022[5]</td><td>2022[5]</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>Masked</td></tr><tr><td>Input Gestures</td><td>-</td><td>-</td><td>Body</td><td>Body</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td></tr><tr><td>Head</td><td>-</td><td>-</td><td>-</td><td>M</td><td>-</td><td>-</td><td>S</td><td>-</td><td>-</td><td>-</td><td>S</td><td>S</td><td>-</td><td>-</td><td>-</td><td>C</td><td></td></tr><tr><td>Upper</td><td>S</td><td>S</td><td>-</td><td>M</td><td>S</td><td>C</td><td>-</td><td>S</td><td>C</td><td>S</td><td>-</td><td>C</td><td>S</td><td>S</td><td>S</td><td>C</td><td></td></tr><tr><td>Hands</td><td>-</td><td>-</td><td>S</td><td>S</td><td>M</td><td>S</td><td>C</td><td>-</td><td>-</td><td>C</td><td>S</td><td>-</td><td>C</td><td>S</td><td>-</td><td>S</td><td>C</td></tr><tr><td>Lower + Global</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>S</td><td>-</td><td>-</td><td>C</td><td></td></tr></table>
+
+Table 2. Comparison of Co-Speech Gesture Models. We compare with previous methods for generating face or body motion trained on co-speech datasets. The first row lists their inputs, and the subsequent rows list their outputs, respectively. Different decoder designs are denoted by the initials 'S' for Single, 'M' for Multiple, and 'C' for Cascaded. To the best of our knowledge, EMAGE (Ours) is the first to accept audio and partially or completely masked gestures, generating full-body audio-synchronized results.
+
+ther the face or body. Applying these methods to the full body will yield sub-optimal results as audio is differently correlated with face and body dynamics. Most similar to our work, Habibie et al. [28] employ a single audio encoder and multiple decoders for generating face and body gestures. TalkSHOW [65] demonstrates the advantage of separating the audio encoders and decoding the body and hands using quantized codes in an auto-regressive manner. However, it lacks lower body and global motion, and there is no shared information for body and facial gesture generation. Moreover, it cannot accept partially masked body hints due to the design of a fully auto-regressive model.
+
+Masked Representation Learning was first demonstrated to be effective in Natural Language Processing, with BERT-based models [17, 31, 42] boosting the performance of learned word embeddings in downstream tasks through a combination of masked language modeling and transformer architecture. Subsequently, Masked AutoEncoders [29] expanded masked image modeling to computer vision by removing and inpainting image patches. This concept of masked representation learning has since been employed in other modalities, e.g., video [12, 18, 43], audio [26, 36, 37], and point clouds [55, 68, 69]. Most related to our work, MotionBERT [73] proposes a spatial-temporal transformer to learn a robust motion representation for classification-based tasks by masking 2D pose. In contrast to their method, we target robust motion features for conditional motion generation tasks, requiring a balance in training between multiple modalities, e.g., audio and gesture.
+
+# 3. BEAT2
+
+This section introduces how we obtain unified mesh-level data, i.e., SMPL-X [50] and FLAME [35] parameters, from
+
+the original BEAT dataset [39]. BEAT utilizes a Vicon motion capture system [47] and releases 78-joint skeleton-level Biovision Hierarchy (BVH) [10] motion files. Their facial capture system uses ARKit [8] with a depth camera on the iPhone 12 Pro, extracting  $\mathbb{R}^{51}$  blendshape weights. These blendshapes are designed based on the Facial Action Coding System (FACS), which is widely adopted. However, both the motion and facial data lack mesh-level details (see Figure 2), e.g., shapes and vertices.
+
+# 3.1. Body Parameters Initialization via MoSh++
+
+We initialize the SMPL-X body shape and pose parameters from BEAT mocap marker data using MoSh++ [46]. Given the captured markers positions  $\mathbf{m} \in \mathbb{R}^{T \times K \times 3}$ , predefined markers position offsets  $\mathbf{d} \in \mathbb{R}^{K \times 3}$ , and user-defined vertices-to-markers function  $\mathcal{H}$ , we aim to obtain body shape  $\beta \in \mathbb{R}^{300}$ , poses  $\theta \in \mathbb{R}^{T \times 55 \times 3}$ , and translation parameters  $\gamma \in \mathbb{R}^{T \times 3}$ . The optimization uses a differentiable surface vertex mapping function  $S(\beta, \theta, \gamma)$  and vertex normal function  $\mathcal{N}(\beta, \theta_0, \gamma_0)$ . For each frame, the latent marker  $\tilde{\mathbf{m}} \in \mathbb{R}^{T \times K \times 3}$  is calculated as
+
+$$
+\tilde {\mathbf {m}} _ {i} \equiv \mathcal {S} _ {\mathcal {H}} (\beta , \theta_ {i}, \gamma_ {i}) + \mathbf {d N} _ {\mathcal {H}} (\beta , \theta_ {i}, \gamma_ {i}). \tag {1}
+$$
+
+We first select 12 frames to optimize and fix  $\mathbf{d}$  and  $\beta$ , then optimize  $\theta_{i}$ ,  $\gamma_{i}$  for  $i \in (1:T)$  by minimizing the loss terms based on  $\|\tilde{\mathbf{m}}_i - \mathbf{m}_i\|^2$ , including Data Term, Surface Distance Energy, Marker Initialization Regularization, Pose and Shape Priors, Velocity Constancy, and Soft-Tissue Term (see supplementary materials). The overall objective function is the weighted sum of these terms to balance accuracy and plausibility.
+
+![](images/1a29c1ea8463bf00c2c3f184a817a4f1c708628a8ef1d47d76d7f19f87bcfc5d.jpg)
+
+![](images/0d3097e295ab5f39d373170046e3ede29bd44caca6a90f7a1e16e8cbbc74f8e2.jpg)  
+BEAT
+
+![](images/3cf8e0bff276205e84046de305c5ed3d778dd92008b38ecc417fbe3aa563ac59.jpg)
+
+![](images/9c115889ca59e8c11153c9150f32df01371750dbb710af599fab43cfc7f8e73d.jpg)  
+Retargeted
+
+![](images/0b2f09f2374956161351194f45e2e2532c9d05df8cf0f908d5ce265ba34ba94c.jpg)
+
+![](images/f8f4640aefd7e52843ec5ee7f210087f3efb995a8208587356d71d31ebcdc3c4.jpg)  
+Moshed  
+Figure 2. Comparison of Data between BEAT2 and Others. BEAT-SMPLX-FLAME presents a new mesh-level, motion-captured, holistic co-speech gesture dataset with 60h of data. Left: We compare our refined SMPL-X body parameters (denoted as Refined MoSh) with the original BEAT skeleton [39], Retargeted from AutoRegPro, and initial results of  $\mathrm{Mosh}++$  [50]. The refined results show correct neck flexion, appropriate head and neck shape ratios, and detailed finger representations. Right: Visualization of blendshape weights from the original BEAT dataset [39] with ARKit's template, Wrapped-based, and handcrafted optimization. Our final handcrafted FLAME blendshape-based optimization demonstrates both accurate lip movement details and natural mouth shapes.
+
+![](images/3c1b9aa5d1182a6c0f9d831ead367f4923e1919823cf79b76dca4f59fb4552c0.jpg)
+
+![](images/998179254e8dbd0a0cd08de2c91b898bf38655be992f223fb73fe5128e9c24bd.jpg)  
+Refined Moshed (Ours)
+
+![](images/2e2280169812212b1d236e2a6a3e66b5ab26bd14574a2bb1d9abdb43d01b8b65.jpg)
+
+![](images/0823b4da28a4d4e58bba88dec90cc1bb5f4de875f981e5c1269d82ed5df385a6.jpg)  
+BEAT
+
+![](images/0a5a6dee70f74b4a51ebca6acff78622320dfb2cf09a947e6ba026a25c523992.jpg)
+
+![](images/4e2e0a6d3965907893b8e602a4bd4865e5ecea413958654dd3f083be29da6d78.jpg)  
+Wrapped Optimum
+
+![](images/aa583dca0754a84b1df2c39c7b1147b4db1387c0d69b19fcce8f8d91e9e9466b.jpg)
+
+![](images/3da42aa9860e70ce3b694a6054783d9f404c5a344355f108f4502a33f7839eda.jpg)  
+Handcraft Optimum (Ours)
+
+# 3.2. Body Parameters Refinement
+
+MoSh++ produces unnatural head shapes due to the fact that head markers were worn on a helmet and it sometimes produces unnatural finger poses. Consequently, we refine body shape and pose parameters with three simple yet effective hard-coded physical rules. i) The neck and head length should approximate 1/7 of the total length of the body. ii) Fingers, except for the thumb, should not bend backward. iii) We employ the Kolmogorov-Smirnov test, which reveals the data is similar to a normal distribution. Subsequently, we apply a Gaussian truncation approach, where all data points falling outside the  $3\sigma$  range are adjusted to align with the  $3\sigma$  threshold and blended with the adjacent 10 frames. We compare our refined body parameters (denoted as MoSh Refined) with Original BEAT, Retargeted SMPL-X, and MoSh SMPL-X in Figure 2.
+
+# 3.3. Blendshape Weights to FLAME Parameters
+
+Given the ARKit blendshape weights  $\mathbf{b}_{\mathrm{ARKit}} \in \mathbb{R}^{T \times 51}$ , we aim to derive a transformation matrix  $\mathbf{W} \in \mathbb{R}^{51 \times 103}$  that maps these into FLAME parameters  $\mathbf{b}_{\mathrm{FLAME}} \in \mathbb{R}^{T \times (100 + 3)}$ , where 100 represents the number of dimensions for expression parameters and 3 for jaw movement. Due to discrepancies between the iPhone's and FLAME's template mesh topology, simply optimizing FLAME parameters by minimizing the geometric error between the ARKit template vertices and the wrapped FLAME vertices does not yield satisfactory results. Instead, we release a set of handcrafted blendshape templates  $\mathbf{v}_{\mathrm{template}} \in \mathbb{R}^{52}$  on FLAME following ARKit's FACS configuration. This approach allows direct driving of the FLAME topology ver
+
+tices  $\mathbf{v}\in \mathbb{R}^{T\times 5023\times 3}$  from the given blendshape weights,
+
+$$
+\mathbf {v} = \mathbf {v} _ {\text {t e m p l a t e}} ^ {0} + \sum_ {j = 1} ^ {5 1} \mathbf {b} _ {\text {A R K i t}, j} \cdot \mathbf {v} _ {\text {t e m p l a t e}} ^ {j} \tag {2}
+$$
+
+where  $\mathbf{b}_{\mathrm{ARKit},j}$  is the weight of the  $j$ -th ARKit blend-shape and  $\mathbf{v}_{\mathrm{template}}^j$  is the vertex position of the FLAME model template corresponding to the  $j$ -th blendshape. The term  $\mathbf{v}_{\mathrm{template}}^0$  represents the original template vertex position of the FLAME model. We optimize  $\mathbf{W}$  by minimizing  $\| \tilde{\mathbf{v}}_j - \mathbf{v}_j\| _2$ , where  $\tilde{\mathbf{v}}$  is obtained from FLAME's LBS  $\mathcal{V}(\mathbf{b}_{\mathrm{FLAME}})$ . The comparisons between ARKit data, wrap-based, and our approach are shown in Figure 2.
+
+# 4. IMAGE
+
+We introduce the details of EMAGE, Expressive Masked Audio-Conditioned GEsture Modeling (see Figure 3). Given gestures  $\mathbf{g} \in \mathbb{R}^{T \times (55 \times 6 + 100 + 4 + 3)}$ , representing 55 joints in Rot6D,  $\mathbb{R}^{100}$  FLAME parameters,  $\mathbb{R}^4$  foot contact labels,  $\mathbb{R}^3$  global translations, and speech audio  $\mathbf{a} \in \mathbb{R}^{T \times sk}$ , where  $sk = sr_{\mathrm{audio}} / fps_{\mathrm{gestures}}$ , EMAGE jointly optimizes masked gesture reconstruction and audio-conditioned gesture generation. This optimization enhances performance during inference and enables the use of partially masked gestures to complete holistic gestures. To achieve this, we first model the quantized latent space in Sec. 4.1, following [61, 65]. We then design a separated speech audio encoder using Content Rhythm Self-Attention (CRA) as detailed in Sec. 4.2. Subsequently, we learn the body hints from masked gestures via the Masked Audio Gesture Transformer, explained in Sec. 4.3. Finally, the decoding strategy for different body segments is discussed in Sec. 4.4.
+
+![](images/03d3d4b036b0ca4357b992552ab291dfb862bda77305f32207eae0b1502e2b37.jpg)  
+Figure 3. EMAGE leverages two training paths: Masked Gesture Reconstruction (MG2G) and Audio-Conditioned Gesture Generation (A2G). The MG2G path focuses on encoding robust body hints through a spatial-temporal transformer gesture encoder and cross-attention gesture decoder. In contrast, the A2G path utilizes these body hints, and separated audio encoders to decode pretrained face and body latent features. A key component in this process is a switchable cross-attention layer, crucial for merging body hints and audio features. This fusion allows the features to be effectively disentangled and utilized for gesture decoding. Once the gesture latent features are reconstructed, EMAGE employs a pretrained VQ-Decoder to decode face and local body motions. Additionally, a pretrained Global Motion Predictor is used to estimate global body translations, further enhancing the model's capability to generate realistic and coherent gestures.
+
+![](images/fcfbe7a0829c0edac6c9bb4d65277fe8bae5b40a7eca41f562d382b9fffccf6f.jpg)  
+Figure 4. Details of CRA and Pretrained VQ-VAEs. Left: Content Rhythm Attention fuses speech rhythm (onset and amplitude) with content (pretrained word embeddings from text scripts) adaptively. This results in a preference for either content or rhythm in specific frames, which encourages the generation of semantical-aware gestures. Right: We pretrain four compositional VQ-VAEs by reconstructing face, upper body, hands and lower body separately to disentangle audio-agnostic gestures explicitly.
+
+![](images/20c78e60af2f8a1e4a54cfc9be2acb3cffddc5a758e0ff797f181e0b8b1facf5.jpg)
+
+# 4.1. Compositional Discrete Face and Body Prior
+
+We model full-body gestures in the separated quantized latent space (Figure 4) for several reasons. Similar to [65], the body and hands improves the diversity of the results. However, we additionally separate the face and lower body due to their differing correlations with audio, i.e., using a single VQ-VAE to encode both the upper and lower body may lead the model to overlook gestures that occur less frequently - regardless of their relation to the audio. Specifically, a single model might focus only on recovering lower body motion, neglecting elbow movements in a speaker who is constantly walking during the conversation.
+
+The separated quantized latent space  $Q = \{\mathbf{q}_{\mathrm{f}}, \mathbf{q}_{u}, \mathbf{q}_{\mathrm{h}}, \mathbf{q}_{\mathrm{l}}\}$  for the Face  $\mathbf{g}_{\mathrm{f}} \in \mathbb{R}^{T \times 106}$ , Upper
+
+body  $\mathbf{g}_{\mathrm{u}} \in \mathbb{R}^{T \times 78}$ , Hands  $\mathbf{g}_{\mathrm{h}} \in \mathbb{R}^{T \times 180}$ , and Lower body  $\mathbf{g}_{\mathrm{l}} \in \mathbb{R}^{T \times (54 + 4)}$  are from four VQ-VAEs. The VQ-VAE for each is optimized by jointly optimizing the following loss terms,
+
+$$
+q _ {i} = \underset {q _ {i} \in \mathbf {q}} {\arg \min } \| z _ {j} - q _ {i} \| ^ {2} \tag {3}
+$$
+
+$$
+\begin{array}{l} \mathcal {L} _ {\mathrm {V Q - V A E}} = \mathcal {L} _ {r e c} (\mathbf {g}, \hat {\mathbf {g}}) + \mathcal {L} _ {\mathrm {v e l}} \left(\mathbf {g} ^ {\prime}, \hat {\mathbf {g}} ^ {\prime}\right) + \mathcal {L} _ {\mathrm {a c c}} \left(\mathbf {g} ^ {\prime \prime}, \hat {\mathbf {g}} ^ {\prime \prime}\right) \\ + \left\| \operatorname {s g} [ \mathbf {z} ] - \mathbf {q} \right\| ^ {2} + \left\| \mathbf {z} - \operatorname {s g} [ \mathbf {q} ] \right\| ^ {2}, \tag {4} \\ \end{array}
+$$
+
+where  $\mathbf{z}$  is the encoded  $\mathbf{g}$  with a temporal window size  $w = 1$ .  $\mathcal{L}_{\mathrm{rec}}$ ,  $\mathcal{L}_{\mathrm{vel}}$ , and  $\mathcal{L}_{\mathrm{acc}}$  are Geodesic [58] and L1 losses, sg is a stop gradient operation, we set the weight of the commitment loss [59] (the last term) to 1 in this paper.
+
+# 4.2. Content and Rhythm Self-Attention
+
+Given the speech audio, s, inspired by [4], we employ onset o and amplitude a as explicit audio rhythm, alongside the pretrained embeddings [9] e from transcripts as content. Different from previous approaches, which typically add the rhythm and content features, we leverage self-attention to merge these features adaptively. This approach is driven by the observation that for specific frames, the gestures are more related to content (semantic-aware) or rhythm (beat-aware). The rhythm and content features are first encoded into time-aligned features  $\mathbf{r}_{1:T}$  and  $\mathbf{c}_{1:T}$ , using a Temporal Convolutional Network (TCN) and linear mapping, respectively. For each time step  $t\in \{1,\dots,T\}$ , we merge the rhythm and content features by:
+
+$$
+\begin{array}{l} \mathbf {f} _ {1: T} = \alpha \times \mathbf {r} _ {1: T} + (1 - \alpha) \times \mathbf {c} _ {1: T}, \tag {5} \\ \alpha = \operatorname {S o f t m a x} (\mathcal {A T} (\mathbf {r} _ {1: T}, \mathbf {c} _ {1: T})), \\ \end{array}
+$$
+
+![](images/b5ffedfff937a78fa40fca7cb43e4752e73f12773787d02e95ac5a687967d09f.jpg)  
+Figure 5. Comparison of Forward Path Designs. Straightforward fusion module (a) merges audio features without refined body features and recombines audio features based only on position embedding. The Self-Attention decoder module (b), adopted in previous MLM models [17, 31], is limited for tasks requiring auto-regressive inference. Our design (c) considers effective audio feature fusion and auto-regressive decoding.
+
+![](images/1cacf29fb5e6624bf438a25165fccad902e7718042aa87ac7732c73e36efd625.jpg)
+
+![](images/32cfc4eb4ef2edd24fa8beeb6c6b37217aecf0976a8d36ece3aaf17f2d0afc5b.jpg)
+
+where  $\mathcal{A}\mathcal{T}$  is a 2-layer MLP. We apply two separate CRA encoders for the face and body.
+
+# 4.3. Masked Audio-Conditioned Gesture Modeling
+
+We propose a Masked Audio Gesture Transformer to leverage different training paths (see Fig. 5 for the motivation behind the architecture designs). Given temporal and spatially masked gestures  $\overline{\mathbf{g}}\in \mathbb{R}^{T\times 337}$ , we first replace the masked tokens with learned masked embeddings  $e_{\mathrm{mask}}\in \mathbb{R}^{256}$ , as the value zero still represents specific motion content, e.g., a T-pose. We linearly increase the ratio of masked joints and frames from 0 to  $95\%$  according to the training epochs.
+
+A spatial convolutional  $\mathcal{SC}$  encoder initially summarizes the spatial information, and compress the spatial feature into  $\mathbb{R}^{T\times 256}$ . We then employ a temporal self-attention (without feed-forward)  $\mathcal{TSA}$  to refine the summarized spatial features
+
+$$
+\mathbf {h} = \mathcal {T S A} (\mathcal {S C} (\overline {{\mathbf {g}}}) + \mathbf {p} _ {t}), \tag {6}
+$$
+
+where  $\mathbf{h} \in \mathbb{R}^{T \times 512}$  represents encoded body hints, and  $\mathbf{p}_t$  is the sum of a learned speaker embedding and PPE [19]. Then, a straightforward temporal cross-attention Transformer decoder  $\mathcal{TCA}\mathcal{T}$  is adopted for the reconstructed latent  $\hat{\mathbf{q}}_{\mathrm{mg2g}}$ ,
+
+$$
+\hat {\mathbf {q}} _ {\mathrm {m g} 2 \mathrm {g}} = \mathcal {T C A T} (\mathbf {h} + \mathbf {p} _ {t}). \tag {7}
+$$
+
+We minimize the L1 distance in the latent space,
+
+$$
+\mathcal {L} _ {\mathrm {m g} 2 \mathrm {g}} = \| \hat {\mathbf {q}} _ {\mathrm {m g} 2 \mathrm {g}} - \mathbf {q} \|. \tag {8}
+$$
+
+Masked gesture reconstruction encodes effective body hints, with the key being to leverage these body hints for gesture generation. We employ a selective fusion for audio and body hints via a temporal cross-attention  $\mathcal{TCA}$ . Subsequently, we use the merged audio-gesture features for audio-conditioned gesture latent reconstruction,
+
+$$
+\hat {\mathbf {q}} _ {\mathrm {a} 2 \mathrm {g}} = \mathcal {T C A T} \left(\mathcal {T C A} \left(\mathbf {h} + \mathbf {p} _ {t}, \mathbf {f} _ {\text {b o d y}}\right), \overline {{\mathbf {g}}} + \mathbf {p} _ {t}\right). \tag {9}
+$$
+
+<table><tr><td></td><td>Top.-B</td><td>Top.-F</td><td>Body</td><td>Face</td></tr><tr><td>VOCA[14]</td><td>-</td><td>FLAME</td><td>-</td><td>38.3 ± 5.63</td></tr><tr><td>AMASS[46]</td><td>SMPL-X</td><td>FLAME</td><td>42.0 ± 3.60</td><td>-</td></tr><tr><td>TalkSHOW[65]</td><td>SMPL-X</td><td>SMPL-X</td><td>14.4 ± 2.19</td><td>26.1 ± 6.42</td></tr><tr><td>BEAT2 (Ours)</td><td>SMPL-X</td><td>FLAME</td><td>43.6 ± 3.38</td><td>35.7 ± 5.91</td></tr></table>
+
+Table 3. User Preference Win Rate between Existing Datasets. 'Top' denotes the topology of the mesh. The results show our BEAT2 dataset outperforms the existing PGT dataset [65] in both body and face aspects. It also performs slightly better on body and lower on face when compared with previous mocap [46] and 3D scan-based datasets [14], respectively.
+
+We optimize both the latent code class classification cross-entropy loss  $\mathcal{L}_{\mathrm{a2g - rec}}$  and the latent reconstruction loss  $\mathcal{L}_{\mathrm{a2g - c1s}}$  to encourage diverse results.
+
+# 4.4. Face and Translation Decoding
+
+Considering that the face is weakly related to body motion, applying the same operation - recombining the audio features based on body hints - is not reasonable. Therefore, we directly concatenate the masked body hints with the audio features for the final decoding of facial latent,
+
+$$
+\hat {\mathbf {q}} _ {\mathrm {f}} = \mathcal {T} \mathcal {C A T} \left(\mathbf {f} _ {\text {f a c e}} \oplus \mathbf {h}, \mathbf {p} _ {t}\right). \tag {10}
+$$
+
+Once we have obtained the local lower body motion  $\tilde{\mathbf{g}}_l\in$ $\mathbb{R}^{T\times (54 + 4)}$  from the VQ-Decoder, we estimate the global translations  $\tilde{\mathbf{t}}\in \mathbb{R}^{T\times 3}$  with a pretrained Global Motion Predictor  $\tilde{\mathbf{t}} = \mathcal{G}(\tilde{\mathbf{g}}_l)$ , which significantly reduces foot sliding.
+
+# 5. Experiments
+
+We separate the evaluation into two categories: dataset quality and model ability. After removing data with low finger quality from BEAT, BEAT2 is reduced to 60 hours. We further split it into BEAT2-Standard (27 hours) and BEAT2-Additional (33 hours), based on the type of speech and conversation sections [39]. While acted gestures in speech sections are more diverse and expressive, spontaneous gestures in conversational sections tend to be more natural yet less varied. We report the results on BEAT2-Standard Speaker2 with an  $85\% / 7.5\% / 7.5\%$  train/val/test split.
+
+# 5.1. Dataset Quality Evaluation
+
+We compare our dataset with the state-of-the-art pseudo ground truth (PGT) dataset, TalkSHOW [65], for both face and body. Additionally, we compare it with the AMASS [46] dataset for the body and VOCA [14] for the face as references. Due to the varying lengths of sequences in each dataset, we sample 100 comparison pairs, each with an equal duration ranging from 2 to 4 seconds. In a perceptual study, each participant evaluates a random set of 40 pairs during a 10-minute session by selecting the sequence they consider to have the best captured quality. In total, 60 participants were invited. It is important to note that participants are instructed to compare only the upper body results,
+
+... In my spare time ... ... if I feel ok ...
+
+... I will go for a walk or hike in nature. I try to organize something
+
+![](images/2c125761dd9cae7c4576e912ae7f448a11376383f3058ef27dd260925960ebc0.jpg)
+
+![](images/cdb9bcf28bebd72500a014eecdaa75e44c0ee0d34d340bf78adbc004c8817654.jpg)
+
+![](images/bae710f5cf209e06c839b96c23eb6e0f4e8f3ae177079db6f6be7e4871ae7a24.jpg)
+
+![](images/c1c57afefb94e35ad11b6de08d396967a3fa5faca4bb4d28920f253113b08c23.jpg)
+
+![](images/428f0f0053095cf331f75a2ee392f3922abd1e4ccefa68d335d26cba72ee5c79.jpg)
+
+![](images/0e571774cfd0d6a43f3e034ffbc64ca28c24d7885bf47b0ac76ba27000ba1cb3.jpg)
+
+![](images/6ac6fc9288d18bedb3545a7e62f6007ef4f7c006b15af8b108bf094d1ad7880d.jpg)
+
+![](images/5f6862fca91f66b90bd24f08769f95c06dfa042121411312abe32d886c6e6098.jpg)
+
+![](images/331dc7ed8f3b856ee1e11e203e33219dc639d6472ce45416330386342a33996a.jpg)
+
+![](images/fd3a413577dc8a0433bca6a0d076bf3a615e42cf9e70297b2b0c83b45ad34d34.jpg)
+
+![](images/d5f6d5f60c872172245a54c48b4f67219ea87e02dbdecbd6b9a8a9f105928657.jpg)
+
+![](images/50f47b1881a4f057c10224c1452d9152a3706b0b2443921ac8022fe89d6e58c5.jpg)
+
+![](images/783ca256b44282fda31bf3021b3fb18df2884846cbacbae94c1c54303861afca.jpg)
+
+![](images/7e6a2efeb4f1b1f039e4b4103e35d22570d17ac002aa7b76b5737e7ad19b9a58.jpg)
+
+![](images/3b5bb70b6c5c3fbf791f74724eb0bfd3ce4c78419828d7f4465313421d034016.jpg)
+
+![](images/6870b2a16ab908948f4c9a29bfe7b7a29afcad6abcba862bc36131580f62c209.jpg)  
+Figure 6. Subjective Examples. Top to down: GroundTruth, Generated W/O Body Hints, Generated With Body Hints, Visible Body Hints, EMAGE generates diverse, semantic-aware and audio synchronized body gestures, e.g. raise the both hands for "spare time", relaxing gestures for "hike in nature". Besides, as shown in the third and bottom rows, EMAGE is flexible to accept non-audio-synchronized body hints on any frames and joints, to guide or customize the generated gestures explicitly, e.g., repeat a similar motion for raising hands, change the walking orientations, etc. Note that in this figure, The gray joints of generated results are not copies of visible hints.
+
+![](images/6d0dd81e175b8e70e7d2cbaa3207c1c29eec10056fee93189c6f6b8833554f74.jpg)
+
+![](images/fd6f3eee973bb1ecc58436590341bf1e98b4a7419d4e84fc2f7ed5c9472f580e.jpg)
+
+![](images/fe89c02a59451a9e1f4d73c03f7d1acf521b5838104d33ec8fb5dc3acbf675a8.jpg)
+
+![](images/87fdaad99801ac186d6b6a5c6f959383750e13087ad63f5e0cb766c482d4920a.jpg)
+
+<table><tr><td></td><td>FGD ↓</td><td>BC ↑</td><td>Diversity ↑</td><td>MSE ↓</td><td>LVD ↓</td></tr><tr><td>FaceFormer[19]</td><td>-</td><td>-</td><td>-</td><td>7.787</td><td>7.593</td></tr><tr><td>CodeTalker[61]</td><td>-</td><td>-</td><td>-</td><td>8.026</td><td>7.766</td></tr><tr><td>S2G[25]</td><td>28.15</td><td>4.683</td><td>5.971</td><td>-</td><td>-</td></tr><tr><td>Trimodal[67]</td><td>12.41</td><td>5.933</td><td>7.724</td><td>-</td><td>-</td></tr><tr><td>HA2G[40]</td><td>12.32</td><td>6.779</td><td>8.626</td><td>-</td><td>-</td></tr><tr><td>DisCo[38]</td><td>9.417</td><td>6.439</td><td>9.912</td><td>-</td><td>-</td></tr><tr><td>CaMN[39]</td><td>6.644</td><td>6.769</td><td>10.86</td><td>-</td><td>-</td></tr><tr><td>DiffStyleGesture[62]</td><td>8.811</td><td>7.241</td><td>11.49</td><td>-</td><td>-</td></tr><tr><td>Habibie et al.[28]</td><td>9.040</td><td>7.716</td><td>8.213</td><td>8.614</td><td>8.043</td></tr><tr><td>TalkSHOW[65]</td><td>6.209</td><td>6.947</td><td>13.47</td><td>7.791</td><td>7.771</td></tr><tr><td>EMAGE (Ours)</td><td>5.512</td><td>7.724</td><td>13.06</td><td>7.680</td><td>7.556</td></tr></table>
+
+as TalkSHOW contains only the upper body. The results are shown in Table 3.
+
+# 5.2. Model Ability Evaluation
+
+We report results using BEATv1.3 and BEATv2. The latter has facial data refined by animators and hand data filtered by annotators.
+
+Metrics. We adopt FGD[67] to evaluate the realism of the body gestures. Then, we measure Diversity[33] by calcu
+
+Table 4. Quantitative evaluation on BEATv2. We report FGD  $\times 10^{-1}$ , BC  $\times 10^{-1}$ , Diversity, MSE  $\times 10^{-8}$ , and LVD  $\times 10^{-5}$ . For body gestures, EMAGE significantly improves the FGD, indicating that the generated results are closer to the GT. This shows the effectiveness of body hints from masked gesture modeling.  
+
+<table><tr><td></td><td>Holistic</td><td>Body</td><td>Face</td></tr><tr><td>Habibie et al.[28]</td><td>12.4 ± 3.70</td><td>15.9 ± 6.49</td><td>10.8 ± 3.19</td></tr><tr><td>TalkSHOW[65]</td><td>34.9 ± 5.79</td><td>40.4 ± 8.22</td><td>33.2 ± 6.03</td></tr><tr><td>EMAGE (Ours)</td><td>52.7 ± 7.91</td><td>44.7 ± 8.68</td><td>56.0 ± 7.80</td></tr></table>
+
+Table 5. User Preference Win Rate On Generated Results. The results indicate that our generated outcomes are perceived as more realistic and believable, with a  $14\%$  and  $23\%$  higher user preference for body and face gestures, respectively.
+
+lating the average L1 distance between multiple body gesture clips, and use BC [34] to assess speech-motion synchrony. For the face, we calculate the vertex MSE[61] to measure the positional distance and the vertex L1 difference LVD [65] between the GT and generated facial vertices.
+
+Compared Methods. We first compare our method with representative state-of-the-art methods in body gesture generation [25, 38-40, 62, 67] and talking head generation [19, 61] by reproducing their methods for body and face, respectively. In addition, we reproduce two previous state-of-the-art holistic pipelines, Habibie et al. [28] and TalkSHOW [65], whose original implementations are limited to the upper body. We add a lower-body decoder to Habibie et al. and a lower-body VQ-VAE to TalkSHOW.
+
+![](images/ec4acad0cc10bb45d1f30e78a3cac08c72cd08e88bef35a1db35c54e8a1d34f8.jpg)  
+Figure 7. Comparison of Generated Facial Movements. Compared with previous state-of-the-art talking face generation methods FaceFormer [19] and CodeTalker [61] as well as holistic gesture generation methods Habibie et al. [28] and TalkSHOW [65]. Note that CodeTalker has higher MSE than IMAGE on BEATv2 (Table 4, lower is better) but is subjectively realistic. IMAGE gets good lip motions by leveraging both the face model and masked body hints.
+
+# 5.2.1 Quantitative and Qualitative Analysis
+
+As shown in Table 4, with a four-frame seed pose, our method outperforms previous SOTA algorithms. For qualitative results, see Figures 6 and 7. Furthermore, we conduct a perceptual study. Maintaining the same setup of 60 participants, each participant evaluates 40 pairs of 10-second results do decide which is most believable; this gives the win rate shown in Table 5.
+
+# 5.2.2 Ablation Analysis
+
+Performance of Baseline. As shown in Table 6, we start with a teacher-force transformer-based baseline. This baseline, inspired by FaceFormer [19], adopts a 1-layer cross-attention transformer decoder and replaces the audio features from Wav2Vec2 [6] with our custom TCN [7] and trainable word embeddings [9].
+
+Effect of Multiple VQ-VAEs. Simply applying one VQ-VAE [27, 59] for full-body movements, including the face, decreases performance in facial movements. This is because the VQ-VAE is trained to minimize the average loss for the full body, and some speakers' most frequent movements are not related to the audio. Implementing separated VQ-VAEs allows the model to better leverage the advantages of discrete priors.
+
+Effect of Content Rhythm Self-Attention. Adaptive fusion of rhythm and content features shows improvement in both FGD and Alignment. It selectively focuses the current motion more on rhythm or content features based on the training data's distribution. Moreover, we observed more semantic-aware results when applying CRA.
+
+Effect of Masked Body Gesture Hints. The improvement across all objective metrics demonstrates that our model effectively leverages spatial-temporal gesture priors, reducing
+
+<table><tr><td></td><td>FGD ↓</td><td>BC ↑</td><td>Diversity ↑</td><td>MSE ↓</td><td>LVD ↓</td></tr><tr><td>Ground Truth</td><td>0</td><td>6.896</td><td>13.074</td><td>0</td><td>0</td></tr><tr><td>Reconstruction</td><td>3.913</td><td>6.758</td><td>13.145</td><td>0.841</td><td>6.389</td></tr><tr><td>Baseline</td><td>13.080</td><td>6.941</td><td>8.3145</td><td>1.442</td><td>9.317</td></tr><tr><td>+ VQVAE</td><td>9.787</td><td>6.673</td><td>10.624</td><td>1.619</td><td>9.473</td></tr><tr><td>+ 4 VQVAE</td><td>7.397</td><td>6.698</td><td>12.544</td><td>1.243</td><td>8.938</td></tr><tr><td>+ CRA</td><td>6.833</td><td>6.783</td><td>12.676</td><td>1.186</td><td>8.715</td></tr><tr><td>+ Masked Hints</td><td>5.423</td><td>6.794</td><td>13.057</td><td>1.180</td><td>9.015</td></tr></table>
+
+Table 6. Abliation Analysis on BEATv1.3  
+
+<table><tr><td></td><td>body</td><td>hands</td><td>face</td><td>FGD ↓</td><td>BC ↑</td><td>Diversity ↑</td></tr><tr><td>BESTX</td><td>✓</td><td>✓</td><td>✓</td><td>5.423</td><td>6.794</td><td>13.057</td></tr><tr><td>+ Trinity[22]</td><td>✓</td><td></td><td></td><td>5.319</td><td>6.843</td><td>13.346</td></tr><tr><td>+ AMASS[46]</td><td>✓</td><td>✓</td><td></td><td>5.174</td><td>6.769</td><td>14.318</td></tr></table>
+
+Table 7. Training IMAGE on Multiple Datasets. IMAGE demonstrates flexibility by training on multiple datasets, even when only a subset of holistic gestures is available. This approach further improves the objective metrics on the BEATv1.3 test set.
+
+the likelihood of incorrect gesture sampling. Furthermore, masked gesture modeling is key to enabling the network to accept predefined gestures in specific frames.
+
+# 5.2.3 Ability of Multi-Dataset Training.
+
+We demonstrate that EMAGE can effectively combine multiple non-holistic datasets for training by jointly training with the Trinity [22] and AMASS [46] datasets, by using only the upper body and audio pairs from Trinity and only body and hands from AMASS. Separate VQ-VAEs are trained for BEAT2, Trinity, and AMASS, with separate MLP heads implemented for codebook classification. The results in Table 7 show that incorporating data improves performance on the BEATv1.3 test set.
+
+# 6. Conclusion
+
+In this work, we present IMAGE, a framework designed to accept partial gestures as input for completing audio-synchronized holistic gestures. It demonstrates that leveraging masked gesture reconstruction can significantly enhance the performance of audio-conditioned gesture generation. Furthermore, the design of IMAGE enables the training on multiple datasets, further improving performance. Alongside IMAGE, we release BEAT2, the largest multi-modal gesture dataset consistent with SMPL-X and FLAME. We hope that BEAT2 will contribute to knowledge and model sharing across various subareas.
+
+Disclosures: While You Zhou, Xuefei Zhe, Naoya Iwamoto, and Bo Zheng are employees of Huawei Tokyo Research Center, this work was done on their own time with the approval of their employer. MJB CoI: https://files.is.tue.mpg.de/black/CoI_CVPR_2024.txt
+
+# References
+
+[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2  
+[2] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In European Conference on Computer Vision, pages 248–265. Springer, 2020. 2  
+[3] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, pages 487-496. Wiley Online Library, 2020. 2  
+[4] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffu-clip: Gesture diffusion model with clip latents. ACM Trans. Graph. 2, 5  
+[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 2, 3  
+[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 8  
+[7] Shaojie Bai, J Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arxiv 2018. arXiv preprint arXiv:1803.01271, 2, 1803. 8  
+[8] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK-scenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 2, 3  
+[9] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146, 2017. 5, 8  
+[10] Biovision BVH. Biovision bvh. In https://research.cs.wisc.edu/graphics /Courses/ cs-838-1999/Jeff/BVH.html, 1999. 3  
+[11] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: realtime multi-person 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence, 43(1):172-186, 2019. 2  
+[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 3  
+[13] Kiran Chhatre, Radek Danček, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, and Timo Bolkart. Emotional speech-driven 3d body animation via disentangled latent diffusion. arXiv preprint arXiv:2312.04466, 2023. 1
+
+[14] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael Black. Capture, learning, and synthesis of 3D speaking styles. Computer Vision and Pattern Recognition (CVPR), pages 10101-10111, 2019. 2, 3, 6  
+[15] Radek Dančecek, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. ACM, 2023. 1  
+[16] Anna Deichler, Kiran Chhatre, Christopher Peters, and Jonas Beskow. Spatio-temporal priors in 3d human motion. 2021. 2  
+[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3, 6  
+[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3  
+[19] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8  
+[20] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 1  
+[21] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia, 12(6):591-598, 2010. 2  
+[22] Ylva Ferstl and Rachel McDonnell. Investigating the use of recurrent motion modelling for speech gesture generation. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 93-98, 2018. 3, 8  
+[23] Erik Gartner, Mykhaylo Andriluka, Erwin Coumans, and Cristian Sminchisescu. Differentiable dynamics for articulated 3d human motion reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13190-13200, 2022. 2  
+[24] Saeed Ghorbani, Ylva Ferstl, Daniel Holden, Nikolaus F. Troje, and Marc-Andre Carbonneau. Zeroeggs: Zero-shot example-based gesture generation from speech. Computer Graphics Forum, 42(1):206-216, 2023. 2, 3  
+[25] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 1, 2, 3, 7  
+[26] Yuan Gong, Yu-An Chung, and James Glass. Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778, 2021.3
+
+[27] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 8, 4  
+[28] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. arXiv preprint arXiv:2102.06837, 2021. 1, 2, 3, 7, 8  
+[29] 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. 3  
+[30] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 2  
+[31] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 3, 6  
+[32] Gilwoo Lee, Zhiwei Deng, Shugao Ma, Takaaki Shiratori, Siddhartha S Srinivasa, and Yaser Sheikh. Talking with hands  $16.2\mathrm{m}$ : A large-scale dataset of synchronized body-finger motion and audio for conversational motion analysis and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 763-772, 2019. 2, 3  
+[33] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 1, 2, 7, 3  
+[34] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 7, 2  
+[35] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2, 3  
+[36] Haiyang Liu and Cheng Zhang. Reinforcement learning based neural architecture search for audio tagging. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2020. 3  
+[37] Haiyang Liu and Jihan Zhang. Improving ultrasound tongue image reconstruction from lip images using self-supervised learning and attention mechanism. arXiv preprint arXiv:2106.11769, 2021. 3  
+[38] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM Interna-
+
+tional Conference on Multimedia, pages 3764-3773, 2022. 2,3,7  
+[39] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. arXiv preprint arXiv:2203.05297, 2022. 1, 2, 3, 4, 6, 7  
+[40] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2, 3, 7  
+[41] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 1, 2  
+[42] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3  
+[43] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3  
+[44] Matthew M. Loper, Naureen Mahmood, and Michael J. Black. MoSh: Motion and shape capture from sparse markers. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 33(6):220:1-220:13, 2014. 2  
+[45] Shuhong Lu, Youngwoo Yoon, and Andrew Feng. Co-speech gesture synthesis using discrete gesture token learning. arXiv preprint arXiv:2303.12822, 2023. 2, 3  
+[46] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, 2019. 2, 3, 6, 8  
+[47] Tim Massey. Vicon study of dynamic object tracking accuracy. In https://www.vicon.com/resources/blog/vicon-study-of-dynamic-object-tracking-accuracy/, 2020. 2, 3  
+[48] Evonne Ng, Shiry Ginosar, Trevor Darrell, and Hanbyul Joo. Body2hands: Learning to infer 3d hands from conversational gesture body dynamics. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11865-11874, 2021. 2, 3  
+[49] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. Bodyformer: Semantics-guided 3d body gesture synthesis with transformer. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 3  
+[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face,
+
+and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3, 4  
+[51] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2  
+[52] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2  
+[53] Xingqun Qi, Chen Liu, Lincheng Li, Jie Hou, Haoran Xin, and Xin Yu. Emotiongesture: Audio-driven diverse emotional co-speech 3d gesture generation. arXiv preprint arXiv:2305.18891, 2023. 2  
+[54] Xingqun Qi, Jiahao Pan, Peng Li, Ruibin Yuan, Xiaowei Chi, Mengfei Li, Wenhan Luo, Wei Xue, Shanghang Zhang, Qifeng Liu, et al. Weakly-supervised emotion transition learning for diverse 3d co-speech gesture generation. arXiv preprint arXiv:2311.17532, 2023. 2  
+[55] Guocheng Qian, Xingdi Zhang, Abdullah Hamdi, and Bernard Ghanem. Pix4point: Image pretrained transformers for 3d point cloud understanding. 2022. 3  
+[56] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 2, 3  
+[57] Kaede Shiohara, Xingchao Yang, and Takafumi Taketomi. Blendface: Re-designing identity encoders for face-swapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7634-7644, 2023. 2  
+[58] Tommi Tykkälä, Cédric Audras, and Andrew I Comport. Direct iterative closest point for real-time visual odometry. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 2050-2056. IEEE, 2011. 5  
+[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5, 8  
+[60] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, 2020. 2  
+[61] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 1, 2, 3, 4, 7, 8  
+[62] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven co-speech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2, 3, 7
+
+[63] Xingchao Yang and Takafumi Taketomi. Bareskinnet: Demakeup and de-lighting via 3d face reconstruction. In Computer Graphics Forum, pages 623-634. Wiley Online Library, 2022. 2  
+[64] Xingchao Yang, Takafumi Taketomi, and Yoshihiro Kanamori. Makeup extraction of 3d representation via illumination-aware image decomposition. In Computer Graphics Forum, pages 293-307. Wiley Online Library, 2023. 2  
+[65] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8  
+[66] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 3  
+[67] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 1, 2, 3, 7  
+[68] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 3  
+[69] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 3  
+[70] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20807-20817, 2023. 2, 3  
+[71] Yang Zhou, Jimei Yang, Dingzeyu Li, Jun Saito, Deepali Aneja, and Evangelos Kalogerakis. Audio-driven neural gesture reenactment with video motion graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3418-3428, 2022. 2  
+[72] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3  
+[73] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3

2024/EMAGE_ Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e22bf9b2357d77a2b649b18fefa1f07a46110c1b49e9db0788c106ab73de39e
+size 588813

2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/32047b53-ccf5-41e9-bc53-4e712089f958_content_list.json

@@ -0,0 +1,1618 @@
+[
+    {
+        "type": "text",
+        "text": "EMCAD: Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation",
+        "text_level": 1,
+        "bbox": [
+            96,
+            128,
+            875,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Md Mostafijur Rahman, Mustafa Munir, and Radu Marculescu",
+        "bbox": [
+            236,
+            203,
+            732,
+            220
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The University of Texas at Austin",
+        "bbox": [
+            351,
+            222,
+            617,
+            238
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Austin, Texas, USA",
+        "bbox": [
+            405,
+            239,
+            565,
+            255
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "mostafijur.rahman, mmunir, radum@utexas.edu",
+        "bbox": [
+            289,
+            258,
+            673,
+            273
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            309,
+            313,
+            325
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "An efficient and effective decoding mechanism is crucial in medical image segmentation, especially in scenarios with limited computational resources. However, these decoding mechanisms usually come with high computational costs. To address this concern, we introduce EMCAD, a new efficient multi-scale convolutional attention decoder, designed to optimize both performance and computational efficiency. EMCAD leverages a unique multi-scale depth-wise convolution block, significantly enhancing feature maps through multi-scale convolutions. EMCAD also employs channel, spatial, and grouped (large-kernel) gated attention mechanisms, which are highly effective at capturing intricate spatial relationships while focusing on salient regions. By employing group and depth-wise convolution, EMCAD is very efficient and scales well (e.g., only 1.91M parameters and 0.381G FLOPs are needed when using a standard encoder). Our rigorous evaluations across 12 datasets that belong to six medical image segmentation tasks reveal that EMCAD achieves state-of-the-art (SOTA) performance with  $79.4\\%$  and  $80.3\\%$  reduction in #Params and #FLOPs, respectively. Moreover, EMCAD's adaptability to different encoders and versatility across segmentation tasks further establish EMCAD as a promising tool, advancing the field towards more efficient and accurate medical image analysis. Our implementation is available at https://github.com/SLDGroup/EMCAD.",
+        "bbox": [
+            76,
+            340,
+            472,
+            734
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            753,
+            209,
+            768
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In the realm of medical diagnostics and therapeutic strategies, automated segmentation of medical images is vital, as it classifies pixels to identify critical regions such as lesions, tumors, or entire organs. A variety of U-shaped convolutional neural network (CNN) architectures [20, 24, 37, 41, 44, 62], notably UNet [44], UNet++ [62], UNet3+ [24], and nnU-Net [19], have become standard techniques for this purpose, achieving high-quality, high-resolution segmen",
+        "bbox": [
+            75,
+            779,
+            470,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tation output. Attention mechanisms [12, 17, 20, 41, 57] have also been integrated into these models to enhance feature maps and improve pixel-level classification. Although attention-based models have shown improved performance, they still face significant challenges due to the computationally expensive convolutional blocks that are typically used in conjunction with attention mechanisms.",
+        "bbox": [
+            496,
+            310,
+            893,
+            415
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recently, vision transformers [18] have shown promise in medical image segmentation tasks [5, 8, 17, 42, 43, 52, 54, 61] by capturing long-range dependencies among pixels through Self-attention (SA) mechanisms. Hierarchical vision transformers like Swin [34], PVT [55, 56], MaxViT [49], MERIT [43], ConvFormer [33], and MetaFormer [59] have been introduced to further improve the performance in this field. While the SA excels at capturing global information, it is less adept at understanding the local spatial context [13, 28]. To address this limitation, some approaches have integrated local convolutional attention within the decoders to better grasp spatial details. Nevertheless, these methods can still be computationally demanding because they frequently employ costly convolutional blocks. This limits their applicability to real-world scenarios where computational resources are restricted.",
+        "bbox": [
+            496,
+            416,
+            893,
+            656
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address the aforementioned limitations, we introduce EMCAD, an efficient multi-scale convolutional attention decoding using a new multi-scale depth-wise convolution block. More precisely, EMCAD enhances the feature maps via efficient multi-scale convolutions, while incorporating complex spatial relationships and local attention through the use of channel, spatial, and grouped (large-kernel) gated attention mechanisms. Our contributions are as follows:",
+        "bbox": [
+            496,
+            657,
+            893,
+            777
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "- New Efficient Multi-scale Convolutional Decoder: We introduce an efficient multi-scale cascaded fullyconvolutional attention decoder (EMCAD) for 2D medical image segmentation; this takes the multi-stage features of vision encoders and progressively enhances the multi-scale and multi-resolution spatial representations. EMCAD has only 0.506M parameters and 0.11G FLOPs for a tiny encoder with #channels = [32, 64, 160, 256],",
+        "bbox": [
+            500,
+            779,
+            893,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "11769",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/3af94799489a0efcba7e12ed56d3a523e2bdb917ee5fa27705063e9a6672a011.jpg",
+        "image_caption": [
+            "Figure 1. Average DICE scores vs. #FLOPs for different methods over 10 binary medical image segmentation datasets. As shown, our approaches (PVT-EMCAD-B0 and PVT-EMCAD-B2) have the lowest #FLOPs, yet the highest DICE scores."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            109,
+            89,
+            441,
+            273
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "while it has 1.91M parameters and 0.381G FLOPs for a standard encoder with #channels = [64, 128, 320, 512].",
+        "bbox": [
+            89,
+            347,
+            468,
+            378
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Efficient Multi-scale Convolutional Attention Module: We introduce MSCAM, a new efficient multi-scale convolutional attention module that performs depth-wise convolutions at multiple scales; this refines the feature maps produced by vision encoders and enables capturing multi-scale salient features by suppressing irrelevant regions. The use of depth-wise convolutions makes MSCAM very efficient.",
+            "- Large-kernel Grouped Attention Gate: We introduce a new grouped attention gate to fuse refined features with the features from skip connections. By using larger kernel  $(3 \\times 3)$  group convolutions instead of point-wise convolutions in the design, we capture salient features in a larger local context with less computation.",
+            "- Improved Performance: We empirically show that EM-CAD can be used with any hierarchical vision encoder (e.g., PVTv2-B0, PVTv2-B2 [56]), while significantly improving the performance of 2D medical image segmentation. EMCAD produces better results than SOTA methods with a significantly lower computational cost (as shown in Figure 1) on 12 medical image segmentation benchmarks that belong to six different tasks."
+        ],
+        "bbox": [
+            76,
+            378,
+            468,
+            710
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The remaining of this paper is organized as follows: Section 2 summarizes related work. Section 3 describes the proposed method. Section 4 explains our experimental setup and results on 12 medical image segmentation benchmarks. Section 5 covers different ablation experiments. Lastly, Section 6 concludes the paper.",
+        "bbox": [
+            75,
+            710,
+            468,
+            803
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            819,
+            218,
+            834
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Vision encoders",
+        "text_level": 1,
+        "bbox": [
+            76,
+            845,
+            235,
+            859
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Convolutional Neural Networks (CNNs) [21-23, 32, 35, 45-48] have been foundational as encoders due to their pro",
+        "bbox": [
+            75,
+            869,
+            468,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "ficiency in handling spatial relationships in images. More precisely, AlexNet [32] and VGG [46] pave the way, leveraging deep layers of convolutions to extract features progressively. GoogleNet [47] introduces the inception module, allowing more efficient computation of representations across various scales. ResNet [21] introduces residual connections, enabling the training of networks with substantially more layers by addressing the vanishing gradients problem. MobileNets [22, 45] bring CNNs to mobile devices through lightweight, depth-wise separable convolutions. EfficientNet [48] introduces a scalable architectural design to CNNs with compound scaling. Although CNNs are pivotal for many vision applications, they generally lack the ability to capture long-range dependencies within images due to their inherent local receptive fields.",
+        "bbox": [
+            496,
+            90,
+            890,
+            316
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recently, Vision Transformers (ViTs), pioneered by Dosovitskiy et al. [18], enabled the learning of long-range relationships among pixels using Self-attention (SA). Since then, ViTs have been enhanced by integrating CNN features [49, 56], developing novel self-attention (SA) blocks [34, 49], and introducing new architectural designs [55, 58]. The Swin Transformer [34] incorporates a sliding window attention mechanism, while SegFormer [58] leverages MixFFN blocks for hierarchical structures. PVT [55] uses spatial reduction attention, refined in PVTv2 [56] with overlapping patch embedding and a linear complexity attention layer. MaxViT [49] introduces a multi-axis self-attention to form a hierarchical CNN-transformer encoder. Although ViTs address the CNNs limitation in capturing long-range pixel dependencies [21-23, 32, 35, 45-48], they face challenges in capturing the local spatial relationships among pixels. In this paper, we aim to overcome these limitations by introducing a new multi-scale cascaded attention decoder that refines feature maps and incorporates local attention using a multi-scale convolutional attention module.",
+        "bbox": [
+            496,
+            319,
+            892,
+            621
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Medical image segmentation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            633,
+            756,
+            648
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Medical image segmentation involves pixel-wise classification to identify various anatomical structures like lesions, tumors, or organs within different imaging modalities such as endoscopy, MRI, or CT scans [8]. U-shaped networks [7, 19, 24, 26, 37, 41, 44, 62] are particularly favored due to their simple but effective encoder-decoder design. The UNet [44] pioneered this approach with its use of skip connections to fuse features at different resolution stages. UNet++ [62] evolves this design by incorporating nested encoder-decoder pathways with dense skip connections. Expanding on these ideas, UNet  $3+$  [24] introduces comprehensive skip pathways that facilitate full-scale feature integration. Further advancement comes with DC-UNet [37], which integrates a multi-resolution convolution scheme and residual paths into its skip connections. The DeepLab series, including DeepLabv3 [10] and DeepLabv3+ [11], in",
+        "bbox": [
+            496,
+            659,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "11770",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "produce atrous convolutions and spatial pyramid pooling to handle multi-scale information. SegNet [2] uses pooling indices to upsample feature maps, preserving the boundary details. nnU-Net [19] automatically configures hyperparameters based on the specific dataset characteristics, using standard 2D and 3D UNets. Collectively, these U-shaped models have become a benchmark for success in the domain of medical image segmentation.",
+        "bbox": [
+            75,
+            90,
+            468,
+            210
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Recently, vision transformers have emerged as a formidable force in medical image segmentation, harnessing the ability to capture pixel relationships at global scales [5, 8, 17, 42, 43, 52, 58, 61]. TransUNet [8] presents a novel blend of CNNs for local feature extraction and transformers for global context, enhancing both local and global feature capture. Swin-Unet [5] extends this by incorporating Swin Transformer blocks [34] into a U-shaped model for both encoding and decoding processes. Building on these concepts, MERIT [43] introduces a multi-scale hierarchical transformer, which employs SA across different window sizes, thus enhancing the model capacity to capture multiscale features critical for medical image segmentation.",
+        "bbox": [
+            75,
+            212,
+            470,
+            409
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The integration of attention mechanisms has been investigated within CNNs [20, 41] and transformer-based systems [17] for enhancing medical image segmentation. PraNet [20] employs a reverse attention strategy for feature refinement. PolypPVT [17] leverages PVTv2 [56] as its backbone encoder and incorporates CBAM [57] within its decoding stages. The CASCADE [42] presents a novel cascaded decoder, combining channel [23] and spatial [9] attention to refine features at multiple stages, extracted from a transformer encoder, culminating in high-resolution segmentation outputs. While CASCADE achieves notable performance in segmenting medical images by integrating local and global insights from transformers, it is computationally inefficient due to the use of triple  $3 \\times 3$  convolution layers at each decoder stage. In addition to this, it uses single-scale convolutions during decoding. Our new proposal involves the adoption of multi-scale depth-wise convolutions to mitigate these constraints.",
+        "bbox": [
+            75,
+            410,
+            470,
+            680
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Methodology",
+        "text_level": 1,
+        "bbox": [
+            76,
+            690,
+            212,
+            708
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we first introduce our new EMCAD decoder and then explain two transformer-based architectures (i.e., PVT-EMCAD-B0 and PVT-EMCAD-B2) incorporating our proposed decoder.",
+        "bbox": [
+            75,
+            715,
+            468,
+            777
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Efficient multi-scale convolutional attention decoding (EMCAD)",
+        "text_level": 1,
+        "bbox": [
+            76,
+            785,
+            468,
+            816
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce our efficient multi-scale convolutional decoding (EMCAD) to process the multi-stage features extracted from pretrained hierarchical vision encoders for high-resolution semantic segmentation. As shown in Figure 2(b), EMCAD consists of efficient multi",
+        "bbox": [
+            75,
+            825,
+            468,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "scale convolutional attention modules (MSCAMs) to robustly enhance the feature maps, large-kernel grouped attention gates (LGAGs) to refine feature maps fusing with the skip connection via gated attention mechanism, efficient up-convolution blocks (EUCBs) for up-sampling followed by enhancement of feature maps, and segmentation heads (SHs) to produce the segmentation outputs.",
+        "bbox": [
+            496,
+            90,
+            890,
+            196
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "More specifically, we use four MSCAMs to refine pyramid features (i.e.,  $X1$ ,  $X2$ ,  $X3$ ,  $X4$  in Figure 2) extracted from the four stages of the encoder. After each MSCAM, we use an SH to produce a segmentation map of that stage. Subsequently, we upscale the refined feature maps using EUCBs and add them to the outputs from the corresponding LGAGs. Finally, we add four different segmentation maps to produce the final segmentation output. Different modules of our decoder are described next.",
+        "bbox": [
+            496,
+            199,
+            892,
+            335
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1.1 Large-kernel grouped attention gate (LGAG)",
+        "text_level": 1,
+        "bbox": [
+            498,
+            361,
+            867,
+            378
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We introduce a new large-kernel grouped attention gate (LGAG) to progressively combine feature maps with attention coefficients, which are learned by the network to allow higher activation of relevant features and suppression of irrelevant ones. This process employs a gating signal derived from higher-level features to control the flow of information across different stages of the network, thus enhancing its precision for medical image segmentation. Unlike Attention UNet [41] which uses  $1 \\times 1$  convolution to process gating signal  $g$  (features from skip connections) and input feature map  $x$  (upsampled features), in our  $q_{att}(.)$  function, we process  $g$  and  $x$  by applying separate  $3 \\times 3$  group convolutions  $GC_{g}(.)$  and  $GC_{x}(.)$ , respectively. These convolved features are then normalized using batch normalization  $(BN(.))$  [27] and merged through element-wise addition. The resultant feature map is activated through a ReLU  $(R(.))$  layer [39]. Afterward, we apply a  $1 \\times 1$  convolution  $(C(.))$  followed by  $BN(.)$  layer to get a single channel feature map. We then pass the resultant single-channel feature map through a Sigmoid  $(\\sigma(.))$  activation function to yield the attention coefficients. The output of this transformation is used to scale the input feature  $x$  through elementwise multiplication, producing the attention-gated feature  $LGAG(g, x)$ . The  $LGAG(.)$  (Figure 2(g)) can be formulated as in Equations 1 and 2:",
+        "bbox": [
+            496,
+            391,
+            892,
+            768
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nq _ {a t t} (g, x) = R \\left(B N \\left(G C _ {g} (g) + B N \\left(G C _ {x} (x)\\right)\\right)\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            782,
+            890,
+            799
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL G A G (g, x) = x * \\sigma \\left(B N \\left(C \\left(q _ {a t t} (g, x)\\right)\\right)\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            549,
+            832,
+            890,
+            848
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Due to using  $3 \\times 3$  kernel group convolutions in  $q_{att}(.)$ , our LGAG captures comparatively larger spatial contexts with less computational cost.",
+        "bbox": [
+            496,
+            859,
+            892,
+            905
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "11771",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/680916196f86f563262f187b1e9498563b3b45b7b797b45690ae9a0e76ac7c8c.jpg",
+        "image_caption": [
+            "Figure 2. Hierarchical encoder with newly proposed EMCAD decoder architecture. (a) CNN or transformer encoder with four hierarchical stages, (b) EMCAD decoder, (c) Efficient up-convolution block (EUCB), (d) Multi-scale convolutional attention module (MSCAM), (e) Multi-scale convolution block (MSCB), (f) Multi-scale (parallel) depth-wise convolution (MSDC), (g) Large-kernel grouped attention gate (LGAG), (h) Channel attention block (CAB), and (i) Spatial attention block (SAB). X1, X2, X3, and X4 are the features from the four stages of the hierarchical encoder. p1, p2, p3, and p4 are output segmentation maps from four stages of our decoder."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            87,
+            854,
+            354
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1.2 Multi-scale convolutional attention module (MSCAM)",
+        "text_level": 1,
+        "bbox": [
+            76,
+            436,
+            470,
+            465
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We introduce an efficient multi-scale convolutional attention module to refine the feature maps. MSCAM consists of a channel attention block  $(CAB(\\cdot))$  to put emphasis on pertinent channels, a spatial attention block [9]  $(SAB(\\cdot))$  to capture the local contextual information, and an efficient multi-scale convolution block  $(MSCB(\\cdot))$  to enhance the feature maps preserving contextual relationships. The  $MSCAM(\\cdot)$  (Figure 2(d)) is given in Equation 3:",
+        "bbox": [
+            75,
+            474,
+            468,
+            597
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nM S C A M (x) = M S C B (S A B (C A B (x))) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            124,
+            604,
+            468,
+            622
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $x$  is the input tensor. Due to using depth-wise convolution in multiple scales, our MSCAM is more effective with significantly lower computational cost than the convolutional attention module (CAM) proposed in [42].",
+        "bbox": [
+            75,
+            628,
+            468,
+            689
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Multi-scale Convolution Block (MSCB): We introduce an efficient multi-scale convolution block to enhance the features generated by our cascaded expanding path. In our MSCB, we follow the design of the inverted residual block (IRB) of MobileNetV2 [45]. However, unlike IRB, our MSCB performs depth-wise convolution at multiple scales and uses channel_shuffle [60] to shuffle channels across groups. More specifically, in our MSCB, we first expand the number of channels (i.e., expansion_factor = 2) using a point-wise  $(1 \\times 1)$  convolution layers  $PWC_{1}(\\cdot)$  followed by a batch normalization layer  $BN(\\cdot)$  and a ReLU6 [31] activation layer  $R6(\\cdot)$ . We then use a multi-scale depth-wise convolution  $MSDC(\\cdot)$  to capture both multi-scale and multi-resolution contexts. As depth-wise convolution overlooks",
+        "bbox": [
+            75,
+            689,
+            470,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "the relationships among channels, we use a channel_shuffle operation to incorporate relationships among channels. Afterward, we use another point-wise convolution  $PWC_{2}(.)$  followed by a  $BN(.)$  to transform back the original #channels, which also encodes dependency among channels. The  $MSCB(\\cdot)$  (Figure 2(e)) is formulated as in Equation 4:",
+        "bbox": [
+            498,
+            436,
+            890,
+            529
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nM S C B (x) = B N \\left(P W C _ {2} \\left(C S \\left(M S D C \\left(R 6 \\left(B N \\left(P W C _ {1} (x)\\right)\\right)\\right)\\right)\\right)\\right) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            511,
+            537,
+            890,
+            551
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where parallel  $MSDC(.)$  (Figure 2(f)) for different kernel sizes  $(KS)$  can be formulated using Equation 5:",
+        "bbox": [
+            498,
+            560,
+            890,
+            592
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nM S D C (x) = \\sum_ {k s \\in K S} D W C B _ {k s} (x) \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            571,
+            599,
+            890,
+            617
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $DWCB_{ks}(x) = R6(BN(DWC_{ks}(x)))$ . Here,  $DWC_{ks}()$  is a depth-wise convolution with the kernel size  $ks$ .  $BN()$  and  $R6()$  are batch normalization and ReLU6 activation, respectively. Additionally, our sequential  $MSDC()$  uses the recursively updated input  $x$ , where the input  $x$  is residually connected to the previous  $DWCB_{ks}()$  for better regularization as in Equation 6:",
+        "bbox": [
+            498,
+            625,
+            890,
+            731
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nx = x + D W C B _ {k s} (x) \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            614,
+            739,
+            890,
+            756
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Channel Attention Block (CAB): We use channel attention block to assign different levels of importance to each channel, thus emphasizing more relevant features while suppressing less useful ones. Basically, the CAB identifies which feature maps to focus on (and then refine them). Following [57], in CAB, we first apply the adaptive maximum pooling  $(P_{m}(\\cdot))$  and adaptive average pooling  $(P_{a}(\\cdot))$  to the spatial dimensions (i.e., height and width) to extract the most significant feature of the entire feature map per",
+        "bbox": [
+            496,
+            763,
+            890,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "11772",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "channel. Then, for each pooled feature map, we reduce the number of channels  $r = 1 / 16$  times separately using a point-wise convolution  $(C_1(\\cdot))$  followed by a ReLU activation  $(R)$ . Afterward, we recover the original channels using another point-wise convolution  $(C_2(\\cdot))$ . We then add both recovered feature maps and apply Sigmoid  $(\\sigma)$  activation to estimate attention weights. Finally, we incorporate these weights to input  $x$  using the Hadamard product  $(\\otimes)$ . The  $CAB(\\cdot)$  (Figure 2(h)) is defined using Equation 7:",
+        "bbox": [
+            75,
+            90,
+            472,
+            227
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nC A B (x) = \\sigma \\left(C _ {2} \\left(R \\left(C _ {1} \\left(P _ {m} (x)\\right)\\right)\\right) + C _ {2} \\left(R \\left(C _ {1} \\left(P _ {a} (x)\\right)\\right)\\right)\\right) * x \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            94,
+            237,
+            468,
+            252
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Spatial Attention Block (SAB): We use spatial attention to mimic the attentional processes of the human brain by focusing on specific parts of an input image. Basically, the SAB determines where to focus in a feature map; then it enhances those features. This process enhances the model's ability to recognize and respond to relevant spatial features, which is crucial for image segmentation where the context and location of objects significantly influence the output. In SAB, we first pool maximum  $(Ch_{max}(\\cdot))$  and average  $(Ch_{avg}(\\cdot))$  values along the channel dimension to pay attention to local features. Then, we use a large kernel (i.e.,  $7\\times 7$  as in [17]) convolution layer to enhance local contextual relationships among features. Afterward, we apply the Sigmoid activation  $(\\sigma)$  to calculate attention weights. Finally, we feed these weights to the input  $x$  (using Hadamard product  $(\\otimes)$  to attend information in a more targeted way. The  $SAB(.)$  (Figure 2(i)) is defined using Equation 8:",
+        "bbox": [
+            75,
+            260,
+            472,
+            496
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS A B (x) = \\sigma \\left(L K C \\left(\\left[ C h _ {m a x} (x), C h _ {a v g} (x) \\right]\\right)\\right) * x \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            114,
+            506,
+            468,
+            522
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.1.3 Efficient up-convolution block (EUCB)",
+        "text_level": 1,
+        "bbox": [
+            76,
+            534,
+            401,
+            550
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use an efficient up-convolution block to progressively upsample the feature maps of the current stage to match the dimension and resolution of the feature maps from the next skip connection. The EUCB first uses an UpSampling  $Up(\\cdot)$  with scale-factor 2 to upscale the feature maps. Then, it enhances the upscaled feature maps by applying a  $3 \\times 3$  depth-wise convolution  $DWC(\\cdot)$  followed by a  $BN(\\cdot)$  and a  $ReLU(\\cdot)$  activation. Finally, a  $1 \\times 1$  convolution  $C_{1 \\times 1}(\\cdot)$  is used to reduce the #channels to match with the next stage. The  $EUCB(\\cdot)$  (Figure 2(c)) is formulated as in Equation 9:",
+        "bbox": [
+            75,
+            559,
+            472,
+            710
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nE U C B (x) = C _ {1 \\times 1} \\left(R e L U \\left(B N \\left(D W C (U p (x))\\right)\\right)\\right) \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            89,
+            737,
+            468,
+            753
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Due to using depth-wise convolution instead of  $3 \\times 3$  convolution, our EUCB is very efficient.",
+        "bbox": [
+            76,
+            762,
+            468,
+            792
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.1.4 Segmentation head (SH)",
+        "text_level": 1,
+        "bbox": [
+            76,
+            800,
+            302,
+            816
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use segmentation heads to produce the segmentation outputs from the refined feature maps of four stages of the decoder. The SH layer applies a  $1 \\times 1$  convolution  $Conv_{1 \\times 1}(\\cdot)$  to the refined feature maps having  $ch_i$  channels ( $ch_i$  is the #channels in the feature map of stage  $i$ ) and",
+        "bbox": [
+            76,
+            825,
+            472,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "produces output with #channels equal to #classes in target dataset for multi-class but 1 channel for binary segmentation. The  $SH(\\cdot)$  is formulated as in Equation 10:",
+        "bbox": [
+            496,
+            90,
+            892,
+            137
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS H (x) = \\operatorname {C o n v} _ {1 \\times 1} (x) \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            614,
+            142,
+            892,
+            159
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2. Overall architecture",
+        "text_level": 1,
+        "bbox": [
+            498,
+            169,
+            692,
+            184
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To show the generalization, effectiveness, and ability to process multi-scale features for medical image segmentation, we integrate our EMCAD decoder alongside tiny (PVTv2-B0) and standard (PVTv2-B2) networks of PVTv2 [56]. However, our decoder is adaptable and seamlessly compatible with other hierarchical backbone networks.",
+        "bbox": [
+            496,
+            191,
+            890,
+            281
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "PVTv2 differs from conventional transformer patch embedding modules by applying convolutional operations for consistent spatial information capture. Using PVTv2-b0 (Tiny) and PVTv2-b2 (Standard) encoders [56], we develop the PVT-EMCAD-B0 and PVT-EMCAD-B2 architectures. To adopt PVTv2, we first extract the features (X1, X2, X3, and X4) from four layers and feed them (i.e., X4 in the upsample path and X3, X2, X1 in the skip connections) into our EMCAD decoder as shown in Figure 2(a-b). EMCAD then processes them and produces four segmentation maps that correspond to the four stages of the encoder network.",
+        "bbox": [
+            496,
+            282,
+            892,
+            449
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3. Multi-stage loss and outputs aggregation",
+        "text_level": 1,
+        "bbox": [
+            498,
+            453,
+            848,
+            469
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our EMCAD decoder's four segmentation heads produce four prediction maps  $p_1, p_2, p_3$ , and  $p_4$  across its stages.",
+        "bbox": [
+            496,
+            476,
+            890,
+            506
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Loss aggregation: We adopt a combinatorial approach to loss combination called MUTATION, inspired by the work of MERIT [43] for multi-class segmentation. This involves calculating the loss for all possible combinations of predictions derived from 4 heads, totaling  $2^{4} - 1 = 15$  unique predictions, and then summing these losses. We focus on minimizing this cumulative combinatorial loss during the training process. For binary segmentation, we optimize the additive loss like [42] with an additional term  $\\mathcal{L}_{p_1 + p_2 + p_3 + p_4}$  as in Equation 11:",
+        "bbox": [
+            496,
+            507,
+            892,
+            657
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\alpha \\mathcal {L} _ {p _ {1}} + \\beta \\mathcal {L} _ {p _ {2}} + \\gamma \\mathcal {L} _ {p _ {3}} + \\zeta \\mathcal {L} _ {p _ {4}} + \\delta \\mathcal {L} _ {p _ {1} + p _ {2} + p _ {3} + p _ {4}} \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            517,
+            665,
+            890,
+            680
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{L}_{p_1},\\mathcal{L}_{p_2},\\mathcal{L}_{p_3}$  , and  $\\mathcal{L}_{p_4}$  are the losses of each individual prediction maps.  $\\alpha = \\beta = \\gamma = \\zeta = \\delta = 1.0$  are the weights assigned to each loss.",
+        "bbox": [
+            496,
+            681,
+            890,
+            727
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Output segmentation maps aggregation: We consider the prediction map,  $p_4$ , from the last stage of our decoder as the final segmentation map. Then, we obtain the final segmentation output by employing a Sigmoid function for binary or a Softmax function for multi-class segmentation.",
+        "bbox": [
+            496,
+            728,
+            890,
+            803
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            498,
+            815,
+            633,
+            832
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present the details of our implementation followed by a comparative analysis of our PVT-EMCAD-B0 and PVT-EMCAD-B2 against SOTA methods. Datasets and evaluation metrics are in Supplementary Section 7.",
+        "bbox": [
+            496,
+            840,
+            890,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "11773",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/5013ef6128dea5916bb2dcbb8344229ac8db20e6b47ae03b79b7b457e9a3f673.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td rowspan=\"2\">#Params</td><td rowspan=\"2\">#FLOPs</td><td colspan=\"5\">Polyp</td><td colspan=\"2\">Skin Lesion</td><td colspan=\"2\">Cell</td><td rowspan=\"2\">BUSI</td><td rowspan=\"2\">Avg.</td></tr><tr><td>Clinic</td><td>Colon</td><td>ETIS</td><td>Kvasir</td><td>BKAI</td><td>ISIC17</td><td>ISIC18</td><td>DSB18</td><td>EM</td></tr><tr><td>UNet [44]</td><td>34.53M</td><td>65.53G</td><td>92.11</td><td>83.95</td><td>76.85</td><td>82.87</td><td>85.05</td><td>83.07</td><td>86.67</td><td>92.23</td><td>95.46</td><td>74.04</td><td>85.23</td></tr><tr><td>UNet++ [62]</td><td>9.16M</td><td>34.65G</td><td>92.17</td><td>87.88</td><td>77.40</td><td>83.36</td><td>84.07</td><td>82.98</td><td>87.46</td><td>91.97</td><td>95.48</td><td>74.76</td><td>85.75</td></tr><tr><td>AttnUNet [41]</td><td>34.88M</td><td>66.64G</td><td>92.20</td><td>86.46</td><td>76.84</td><td>83.49</td><td>84.07</td><td>83.66</td><td>87.05</td><td>92.22</td><td>95.55</td><td>74.48</td><td>85.60</td></tr><tr><td>DeepLabv3+ [10]</td><td>39.76M</td><td>14.92G</td><td>93.24</td><td>91.92</td><td>90.73</td><td>89.06</td><td>89.74</td><td>83.84</td><td>88.64</td><td>92.14</td><td>94.96</td><td>76.81</td><td>89.11</td></tr><tr><td>PraNet [20]</td><td>32.55M</td><td>6.93G</td><td>91.71</td><td>89.16</td><td>83.84</td><td>84.82</td><td>85.56</td><td>83.03</td><td>88.56</td><td>89.89</td><td>92.37</td><td>75.14</td><td>86.41</td></tr><tr><td>CaraNet [38]</td><td>46.64M</td><td>11.48G</td><td>94.08</td><td>91.19</td><td>90.25</td><td>89.74</td><td>89.71</td><td>85.02</td><td>90.18</td><td>89.15</td><td>92.78</td><td>77.34</td><td>88.94</td></tr><tr><td>UACANet-L [30]</td><td>69.16M</td><td>31.51G</td><td>94.16</td><td>91.02</td><td>89.77</td><td>90.17</td><td>90.35</td><td>83.72</td><td>89.76</td><td>88.86</td><td>89.28</td><td>76.96</td><td>88.41</td></tr><tr><td>SSFormer-L [54]</td><td>66.22M</td><td>17.28G</td><td>94.18</td><td>92.11</td><td>90.16</td><td>91.47</td><td>91.14</td><td>85.28</td><td>90.25</td><td>92.03</td><td>94.95</td><td>78.76</td><td>90.03</td></tr><tr><td>PolypPVT [17]</td><td>25.11M</td><td>5.30G</td><td>94.13</td><td>91.53</td><td>89.93</td><td>91.56</td><td>91.17</td><td>85.56</td><td>90.36</td><td>90.69</td><td>94.40</td><td>79.35</td><td>89.87</td></tr><tr><td>TransUNet [8]</td><td>105.32M</td><td>38.52G</td><td>93.90</td><td>91.63</td><td>87.79</td><td>91.08</td><td>89.17</td><td>85.00</td><td>89.16</td><td>92.04</td><td>95.27</td><td>78.30</td><td>89.33</td></tr><tr><td>SwinUNet [5]</td><td>27.17M</td><td>6.2G</td><td>92.42</td><td>89.27</td><td>85.10</td><td>89.59</td><td>87.61</td><td>83.97</td><td>89.26</td><td>91.03</td><td>94.47</td><td>77.38</td><td>88.01</td></tr><tr><td>TransFuse [61]</td><td>143.74M</td><td>82.71G</td><td>93.62</td><td>90.35</td><td>86.91</td><td>90.24</td><td>87.47</td><td>84.89</td><td>89.62</td><td>90.85</td><td>94.35</td><td>79.36</td><td>88.77</td></tr><tr><td>UNeXt [50]</td><td>1.47M</td><td>0.57G</td><td>90.20</td><td>83.84</td><td>74.03</td><td>77.88</td><td>77.93</td><td>82.74</td><td>87.78</td><td>86.01</td><td>93.81</td><td>74.71</td><td>82.89</td></tr><tr><td>PVT-CASCADE [42]</td><td>34.12M</td><td>7.62G</td><td>94.53</td><td>91.60</td><td>91.03</td><td>92.05</td><td>92.14</td><td>85.50</td><td>90.41</td><td>92.35</td><td>95.42</td><td>79.21</td><td>90.42</td></tr><tr><td>PVT-EMCAD-B0 (Ours)</td><td>3.92M</td><td>0.84G</td><td>94.60</td><td>91.71</td><td>91.65</td><td>91.95</td><td>91.30</td><td>85.67</td><td>90.70</td><td>92.46</td><td>95.35</td><td>79.80</td><td>90.52</td></tr><tr><td>PVT-EMCAD-B2 (Ours)</td><td>26.76M</td><td>5.6G</td><td>95.21</td><td>92.31</td><td>92.29</td><td>92.75</td><td>92.96</td><td>85.95</td><td>90.96</td><td>92.74</td><td>95.53</td><td>80.25</td><td>91.10</td></tr></table>",
+        "bbox": [
+            80,
+            89,
+            890,
+            324
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/fdfb3380337452f6615e0035d575276581729e149378c3ca5dcfb7cce8888e93.jpg",
+        "table_caption": [
+            "Table 1. Results of binary medical image segmentation (i.e., polyp, skin lesion, cell, and breast cancer). We reproduce the results of SOTA methods using their publicly available implementation with our train-val-test splits of 80:10:10. #FLOPs of all the methods are reported for  $256 \\times 256$  inputs, except Swin-UNet ( $224 \\times 224$ ). All results are averaged over five runs. Best results are shown in bold."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Architectures</td><td colspan=\"3\">Average</td><td rowspan=\"2\">Aorta</td><td rowspan=\"2\">GB</td><td rowspan=\"2\">KL</td><td rowspan=\"2\">KR</td><td rowspan=\"2\">Liver</td><td rowspan=\"2\">PC</td><td rowspan=\"2\">SP</td><td rowspan=\"2\">SM</td></tr><tr><td>DICE↑</td><td>HD95↓</td><td>mIoU↑</td></tr><tr><td>UNet [44]</td><td>70.11</td><td>44.69</td><td>59.39</td><td>84.00</td><td>56.70</td><td>72.41</td><td>62.64</td><td>86.98</td><td>48.73</td><td>81.48</td><td>67.96</td></tr><tr><td>AttnUNet [41]</td><td>71.70</td><td>34.47</td><td>61.38</td><td>82.61</td><td>61.94</td><td>76.07</td><td>70.42</td><td>87.54</td><td>46.70</td><td>80.67</td><td>67.66</td></tr><tr><td>R50+UNet [8]</td><td>74.68</td><td>36.87</td><td>-</td><td>84.18</td><td>62.84</td><td>79.19</td><td>71.29</td><td>93.35</td><td>48.23</td><td>84.41</td><td>73.92</td></tr><tr><td>R50+AttnUNet [8]</td><td>75.57</td><td>36.97</td><td>-</td><td>55.92</td><td>63.91</td><td>79.20</td><td>72.71</td><td>93.56</td><td>49.37</td><td>87.19</td><td>74.95</td></tr><tr><td>SSFormer [54]</td><td>78.01</td><td>25.72</td><td>67.23</td><td>82.78</td><td>63.74</td><td>80.72</td><td>78.11</td><td>93.53</td><td>61.53</td><td>87.07</td><td>76.61</td></tr><tr><td>PolypPVT [17]</td><td>78.08</td><td>25.61</td><td>67.43</td><td>82.34</td><td>66.14</td><td>81.21</td><td>73.78</td><td>94.37</td><td>59.34</td><td>88.05</td><td>79.4</td></tr><tr><td>TransUNet [8]</td><td>77.61</td><td>26.9</td><td>67.32</td><td>86.56</td><td>60.43</td><td>80.54</td><td>78.53</td><td>94.33</td><td>58.47</td><td>87.06</td><td>75.00</td></tr><tr><td>SwinUNet [5]</td><td>77.58</td><td>27.32</td><td>66.88</td><td>81.76</td><td>65.95</td><td>82.32</td><td>79.22</td><td>93.73</td><td>53.81</td><td>88.04</td><td>75.79</td></tr><tr><td>MT-UNet [53]</td><td>78.59</td><td>26.59</td><td>-</td><td>87.92</td><td>64.99</td><td>81.47</td><td>77.29</td><td>93.06</td><td>59.46</td><td>87.75</td><td>76.81</td></tr><tr><td>MISSformer [25]</td><td>81.96</td><td>18.20</td><td>-</td><td>86.99</td><td>68.65</td><td>85.21</td><td>82.00</td><td>94.41</td><td>65.67</td><td>91.92</td><td>80.81</td></tr><tr><td>PVT-CASCADE [42]</td><td>81.06</td><td>20.23</td><td>70.88</td><td>83.01</td><td>70.59</td><td>82.23</td><td>80.37</td><td>94.08</td><td>64.43</td><td>90.1</td><td>83.69</td></tr><tr><td>TransCASCADE [42]</td><td>82.68</td><td>17.34</td><td>73.48</td><td>86.63</td><td>68.48</td><td>87.66</td><td>84.56</td><td>94.43</td><td>65.33</td><td>90.79</td><td>83.52</td></tr><tr><td>PVT-EMCAD-B0 (Ours)</td><td>81.97</td><td>17.39</td><td>72.64</td><td>87.21</td><td>66.62</td><td>87.48</td><td>83.96</td><td>94.57</td><td>62.00</td><td>92.66</td><td>81.22</td></tr><tr><td>PVT-EMCAD-B2 (Ours)</td><td>83.63</td><td>15.68</td><td>74.65</td><td>88.14</td><td>68.87</td><td>88.08</td><td>84.10</td><td>95.26</td><td>68.51</td><td>92.17</td><td>83.92</td></tr></table>",
+        "bbox": [
+            93,
+            377,
+            877,
+            619
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Results of abdomen organ segmentation on Synapse Multi-organ dataset. DICE scores are reported for individual organs. Results of UNet, AttnUNet, PolypPVT, SSFormerPVT, TransUNet, and SwinUNet are taken from [42].  $\\uparrow (\\downarrow)$  denotes the higher (lower) the better.  $^{\\prime} - ^{\\prime}$  means missing data from the source. EMCAD results are averaged over five runs. Best results are shown in bold.",
+        "bbox": [
+            76,
+            623,
+            893,
+            666
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Implementation details",
+        "text_level": 1,
+        "bbox": [
+            76,
+            676,
+            290,
+            694
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We implement our network and conduct experiments using Pytorch 1.11.0 on a single NVIDIA RTX A6000 GPU with 48GB of memory. We utilize ImageNet [16] pretrained PVTv2-b0 and PVTv2-b2 [56] as encoders. In the MSDC of our decoder, we set the multi-scale kernels [1,3,5] through an ablation study. We use the parallel arrangement of depth-wise convolutions in all experiments. Our models are trained using the AdamW optimizer [36] with a learning rate and weight decay of  $1e - 4$ . We generally train for 200 epochs with a batch size of 16, except for Synapse multi-organ (300 epochs, batch size 6) and ACDC cardiac organ (400 epochs, batch size 12), saving the best model based on the DICE score. We resize images",
+        "bbox": [
+            75,
+            703,
+            470,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "to  $352 \\times 352$  and use a multi-scale  $\\{0.75, 1.0, 1.25\\}$  training strategy with a gradient clip limit of 0.5 for ClinicDB [3], Kvasir [29], ColonDB [51], ETIS [51], BKAI [40], ISIC17 [15], and ISIC18 [15], while we resize images to  $256 \\times 256$  for BUSI [1], EM [6], and DSB18 [4]. For Synapse and ACDC datasets, images are resized to  $224 \\times 224$ , with random rotation and flipping augmentations, optimizing a combined Cross-entropy (0.3) and DICE (0.7) loss. For binary segmentation, we utilize the combined weighted BinaryCrossEntropy (BCE) and weighted IoU loss function.",
+        "bbox": [
+            496,
+            676,
+            890,
+            830
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Results",
+        "text_level": 1,
+        "bbox": [
+            500,
+            844,
+            591,
+            858
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We compare our architectures (i.e., PVT-EMCAD-B0 and PVT-EMCAD-B2) with SOTA CNN and transformer-based",
+        "bbox": [
+            500,
+            869,
+            893,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "11774",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/68e95713cdd7aa1342c98b587e762e015c98e20ad9f521d47ceae1aa69432ee7.jpg",
+        "image_caption": [
+            "Figure 3. Average DICE scores vs. #Params for different methods over 10 binary medical image segmentation datasets. As shown, our proposed approaches (PVT-EMCAD-B0 and PVT-EMCAD-B2) have the fewest parameters, yet the highest DICE scores."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            109,
+            89,
+            439,
+            277
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/cc9b5fe31b2429853d4989b4ae359486428262b5afc60e7caacd71fa24d74474.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>Avg. DICE</td><td>RV</td><td>Myo</td><td>LV</td></tr><tr><td>R50+UNet [8]</td><td>87.55</td><td>87.10</td><td>80.63</td><td>94.92</td></tr><tr><td>R50+AttnUNet [8]</td><td>86.75</td><td>87.58</td><td>79.20</td><td>93.47</td></tr><tr><td>ViT+CUP [8]</td><td>81.45</td><td>81.46</td><td>70.71</td><td>92.18</td></tr><tr><td>R50+ViT+CUP [8]</td><td>87.57</td><td>86.07</td><td>81.88</td><td>94.75</td></tr><tr><td>TransUNet [8]</td><td>89.71</td><td>86.67</td><td>87.27</td><td>95.18</td></tr><tr><td>SwinUNet [5]</td><td>88.07</td><td>85.77</td><td>84.42</td><td>94.03</td></tr><tr><td>MT-UNet [53]</td><td>90.43</td><td>86.64</td><td>89.04</td><td>95.62</td></tr><tr><td>MISSformer [25]</td><td>90.86</td><td>89.55</td><td>88.04</td><td>94.99</td></tr><tr><td>PVT-CASCADE [42]</td><td>91.46</td><td>89.97</td><td>88.9</td><td>95.50</td></tr><tr><td>TransCASCade [42]</td><td>91.63</td><td>90.25</td><td>89.14</td><td>95.50</td></tr><tr><td>Cascaded MERIT [43]</td><td>91.85</td><td>90.23</td><td>89.53</td><td>95.80</td></tr><tr><td>PVT-EMCAD-B0 (Ours)</td><td>91.34±0.2</td><td>89.37</td><td>88.99</td><td>95.65</td></tr><tr><td>PVT-EMCAD-B2 (Ours)</td><td>92.12±0.2</td><td>90.65</td><td>89.68</td><td>96.02</td></tr></table>",
+        "bbox": [
+            78,
+            343,
+            472,
+            547
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "segmentation methods on 12 datasets that belong to six medical image segmentation tasks. Qualitative results are in the Supplementary Section 7.3.",
+        "bbox": [
+            75,
+            606,
+            468,
+            654
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2.1 Results of binary medical image segmentation",
+        "text_level": 1,
+        "bbox": [
+            75,
+            664,
+            450,
+            681
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Results for different methods on 10 binary medical image segmentation datasets are shown in Table 1 and Figure 1. Our PVT-EMCAD-B2 attains the highest average DICE score  $(91.10\\%)$  with only  $26.76\\mathrm{M}$  parameters and 5.6G FLOPs. The multi-scale depth-wise convolution in our EMCAD decoder, combined with the transformer encoder, contributes to these performance gains.",
+        "bbox": [
+            75,
+            688,
+            468,
+            794
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Polyp segmentation: Table 1 reveals that our PVT-EMCAD-B2 surpasses all SOTA methods in five polyp segmentation datasets. PVT-EMCAD-B2 achieves DICE score improvements of  $1.08\\%$ ,  $0.78\\%$ ,  $2.36\\%$ ,  $1.19\\%$ , and  $1.79\\%$  over PolypPVT in ClinicDB, ColonDB, ETIS, Kvasir, and BKAI-IGI, despite having slightly more parameters and FLOPs. The smallest model UNeXt, ex",
+        "bbox": [
+            75,
+            795,
+            470,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/2d653adeec59f708f8225a0f18e1acdfba7cd70d71d3dab6b3fe686aa3f65563.jpg",
+        "table_caption": [
+            "Table 3. Results of cardiac organ segmentation on ACDC dataset. DICE scores  $(\\%)$  are reported for individual organs. We get the results of SwinUNet from [42]. Best results are shown in bold."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"3\">Components</td><td colspan=\"2\">#FLOPs(G)</td><td rowspan=\"2\">#Params (M)</td><td rowspan=\"2\">Avg DICE</td></tr><tr><td>Cascaded</td><td>LGAG</td><td>MSCAM</td><td>224</td><td>256</td></tr><tr><td>No</td><td>No</td><td>No</td><td>0</td><td>0</td><td>0</td><td>80.10±0.2</td></tr><tr><td>Yes</td><td>No</td><td>No</td><td>0.100</td><td>0.131</td><td>0.224</td><td>81.08±0.2</td></tr><tr><td>Yes</td><td>Yes</td><td>No</td><td>0.108</td><td>0.141</td><td>0.235</td><td>81.92±0.2</td></tr><tr><td>Yes</td><td>No</td><td>Yes</td><td>0.373</td><td>0.487</td><td>1.898</td><td>82.86±0.3</td></tr><tr><td>Yes</td><td>Yes</td><td>Yes</td><td>0.381</td><td>0.498</td><td>1.91</td><td>83.63±0.3</td></tr></table>",
+        "bbox": [
+            501,
+            89,
+            893,
+            186
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Effect of different components of EMCAD with PVTv2-b2 encoder on Synapse multi-organ dataset. #FLOPs are reported for input resolution of  $224 \\times 224$  and  $256 \\times 256$ . All results are averaged over five runs. Best results are shown in bold.",
+        "bbox": [
+            496,
+            189,
+            893,
+            247
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "hibits the worst performance in all five polyp segmentation datasets. Our smaller model with only 3.92M parameters and 0.84G FLOPs also outperforms all the methods except PVT-CASCADE (in Kvasir and BKAI-IGH) and SSFormer-L (in ColonDB), which achieve the best performance among SOTA methods. In conclusion, our PVT-EMCAD-B2 achieves the new SOTA results in these five polyp segmentation datasets.",
+        "bbox": [
+            496,
+            255,
+            890,
+            376
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Skin lesion segmentation: Table 1 shows PVT-EMCAD-B2's strong performance on ISIC17 and ISIC18 skin lesion segmentation datasets, achieving DICE scores of  $85.95\\%$  and  $90.96\\%$ , surpassing DeepLabV3+ by  $2.11\\%$  and  $2.32\\%$ . It also beats the nearest method PVT-CASCADE by  $0.45\\%$  and  $0.55\\%$  in ISIC17 and ISIC18, respectively, though our decoder is significantly more efficient than CASCADE. Our PVT-EMCAD-B0 also shows huge potential in point care applications like skin lesion segmentation with only 3.92M parameters and 0.84G FLOPs.",
+        "bbox": [
+            496,
+            377,
+            892,
+            527
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Cell segmentation: To evaluate our method's effectiveness in biological imaging, we use DSB18 [4] for cell nuclei and EM [6] for cell structure segmentation. As Table 1 indicates, our PVT-EMCAD-B2 sets a SOTA benchmark in cell nuclei segmentation on DSB18, outperforming DeepLabv3+, TransFuse, and PVT-CASCADE. On the EM dataset, PVT-EMCAD-B2 secures the second-best DICE score (95.53%), offering significantly lower computational costs than the top-performing AttnUNet (95.55%).",
+        "bbox": [
+            496,
+            529,
+            893,
+            665
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Breast cancer segmentation: We conduct experiments on the BUSI dataset for breast cancer segmentation in ultrasound images. Our PVT-EMCAD-B2 achieves the SOTA DICE score  $(80.25\\%)$  on this dataset. Furthermore, our PVT-EMCAD-B0 outperforms the computationally similar method UNeXt by a notable margin of  $5.54\\%$ .",
+        "bbox": [
+            496,
+            665,
+            893,
+            756
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2.2 Results of abdomen organ segmentation",
+        "text_level": 1,
+        "bbox": [
+            498,
+            770,
+            831,
+            786
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2 shows that our PVT-EMCAD-B2 excels in abdomen organ segmentation on the Synapse multi-organ dataset, achieving the highest average DICE score of  $83.63\\%$  and surpassing all SOTA CNN- and transformer-based methods. It outperforms PVT-CASCADE by  $2.57\\%$  in DICE score and 4.55 in HD95 distance, indicating superior organ boundary location. Our EMCAD decoder boosts individ",
+        "bbox": [
+            496,
+            794,
+            893,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "11775",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/960dbdb51d3a1ddc506c2d15b7d43fb9faa404d4b0f8e5c2be86685d76f5ca3f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Conv. kernels</td><td>[1]</td><td>[3]</td><td>[5]</td><td>[1,3]</td><td>[3,3]</td><td>[1,3,5]</td><td>[3,3,3]</td><td>[3,5,7]</td><td>[1,3,5,7]</td><td>[1,3,5,7,9]</td></tr><tr><td>Synapse</td><td>82.43</td><td>82.79</td><td>82.74</td><td>82.98</td><td>82.81</td><td>83.63</td><td>82.92</td><td>83.11</td><td>83.57</td><td>83.34</td></tr><tr><td>ClinicDB</td><td>94.81</td><td>94.90</td><td>94.98</td><td>95.13</td><td>95.06</td><td>95.21</td><td>95.15</td><td>95.03</td><td>95.18</td><td>95.07</td></tr></table>",
+        "bbox": [
+            109,
+            88,
+            857,
+            147
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/795666650d4a54a2666e7382c68cfd032201b42d5155ab078918ba97932efa6c.jpg",
+        "table_caption": [
+            "Table 5. Effect of multi-scale kernels in the depth-wise convolution of MSDC on ClinicDB and Synapse multi-organ datasets. We use the PVTv2-b2 encoder for these experiments. All results are averaged over five runs. Best results are highlighted in bold."
+        ],
+        "table_footnote": [
+            "Table 6. Comparison with the baseline decoder on Synapse Multi-organ dataset. We only report the #FLOPs (with input resolution of  $224 \\times 224$ ) and the #parameters of the decoders. All the results are averaged over five runs. Best results are shown in bold."
+        ],
+        "table_body": "<table><tr><td>Encoders</td><td>Decoders</td><td>#FLOPs(G)</td><td>#Params(M)</td><td>DICE (%)</td></tr><tr><td>PVTv2-B0</td><td>CASCADE</td><td>0.439</td><td>2.32</td><td>80.54</td></tr><tr><td>PVTv2-B0</td><td>EMCAD (Ours)</td><td>0.110</td><td>0.507</td><td>81.97</td></tr><tr><td>PVTv2-B2</td><td>CASCADE</td><td>1.93</td><td>9.27</td><td>82.78</td></tr><tr><td>PVTv2-B2</td><td>EMCAD (Ours)</td><td>0.381</td><td>1.91</td><td>83.63</td></tr></table>",
+        "bbox": [
+            78,
+            189,
+            470,
+            267
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "ual organ segmentation, significantly outperforming SOTA methods on six of eight organs.",
+        "bbox": [
+            76,
+            330,
+            467,
+            361
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.2.3 Results of cardiac organ segmentation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            364,
+            395,
+            380
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3 shows the DICE scores of our PVT-EMCAD-B2 and PVT-EMCAD-B0 along with other SOTA methods, on the MRI images of the ACDC dataset for cardiac organ segmentation. Our PVT-EMCAD-B2 achieves the highest average DICE score of  $92.12\\%$ , thus improving about  $0.27\\%$  over Cascaded MERIT though our network has significantly lower computational cost. Besides, PVT-EMCAD-B2 has better DICE scores in all three organ segmentations.",
+        "bbox": [
+            75,
+            388,
+            467,
+            508
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Ablation Studies",
+        "text_level": 1,
+        "bbox": [
+            76,
+            518,
+            240,
+            534
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this section, we conduct ablation studies to explore different aspects of our architectures and the experimental framework. More ablations are in Supplementary Section 8.",
+        "bbox": [
+            75,
+            542,
+            467,
+            588
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.1. Effect of different components of EMCAD",
+        "text_level": 1,
+        "bbox": [
+            76,
+            592,
+            436,
+            608
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We conduct a set of experiments on the Synapse multi-organ dataset to understand the effect of different components of our EMCAD decoder. We start with only the encoder and add different modules such as Cascaded structure, LGAG, and MSCAM to understand their effect. Table 4 exhibits that the cascaded structure of the decoder helps to improve performance over the non-cascaded one. The incorporation of LGAG and MSCAM improves performance, however, MSCAM proves to be more effective. When both the LGAG and MSCAM modules are used together, it produces the best DICE score of  $83.63\\%$ . It is also evident that there is about  $3.53\\%$  improvement in the DICE score with an additional 0.381G FLOPs and 1.91M parameters.",
+        "bbox": [
+            75,
+            616,
+            467,
+            811
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2. Effect of multi-scale kernels in MSCAM",
+        "text_level": 1,
+        "bbox": [
+            76,
+            816,
+            419,
+            832
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We have conducted another set of experiments on Synapse multi-organ and ClinicDB datasets to understand the effect of different multi-scale kernels used for depth-wise convolutions in MSDC. Table 5 reports these results which show",
+        "bbox": [
+            75,
+            839,
+            467,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "that performance improves from  $1 \\times 1$  to  $3 \\times 3$  kernel. When  $1 \\times 1$  kernel is used together with  $3 \\times 3$  it improves more than when using them alone. However, when two  $3 \\times 3$  kernels are used together, performance drops. The incorporation of a  $5 \\times 5$  kernel with  $1 \\times 1$  and  $3 \\times 3$  kernels further improves the performance and it achieves the best results in both Synapse multi-organ and ClinicDB datasets. If we add additional larger kernels (e.g.,  $7 \\times 7$ ,  $9 \\times 9$ ), the performance of both datasets drops. Based on these empirical observations, we choose [1, 3, 5] kernels in all our experiments.",
+        "bbox": [
+            496,
+            191,
+            890,
+            343
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3. Comparison with the baseline decoder",
+        "text_level": 1,
+        "bbox": [
+            498,
+            352,
+            828,
+            368
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In Table 6, we report the experimental results with the computational complexity of our EMCAD decoder and a baseline decoder, namely CASCADE. From Table 6, we can see that our EMCAD decoder with PVTv2-b2 requires  $80.3\\%$  fewer FLOPs and  $79.4\\%$  fewer parameters to outperform (by  $0.85\\%$ ) the respective CASCADE decoder. Similarly, our EMCAD decoder with PVTv2-B0 achieves  $1.43\\%$  better DICE score than the CASCADE decoder with  $78.1\\%$  fewer parameters and  $74.9\\%$  fewer FLOPs.",
+        "bbox": [
+            496,
+            376,
+            890,
+            512
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusions",
+        "text_level": 1,
+        "bbox": [
+            500,
+            522,
+            625,
+            537
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we have presented EMCAD, a new and efficient multi-scale convolutional attention decoder designed for multi-stage feature aggregation and refinement in medical image segmentation. EMCAD employs a multi-scale depth-wise convolution block, which is key for capturing diverse scale information within feature maps, a critical factor for precision in medical image segmentation. This design choice, using depth-wise convolutions instead of standard  $3 \\times 3$  convolution blocks, makes EMCAD notably efficient.",
+        "bbox": [
+            496,
+            547,
+            890,
+            683
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our experiments reveal that EMCAD surpasses the recent CASCADE decoder in DICE scores with  $79.4\\%$  fewer parameters and  $80.3\\%$  less FLOPs. Our extensive experiments also confirm EMCAD's superior performance compared to SOTA methods across 12 public datasets covering six different 2D medical image segmentation tasks. EMCAD's compatibility with smaller encoders makes it an excellent fit for point-of-care applications while maintaining high performance. We anticipate that our EMCAD decoder will be a valuable asset in enhancing a variety of medical image segmentation and semantic segmentation tasks.",
+        "bbox": [
+            496,
+            685,
+            890,
+            849
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements: This work is supported in part by the NSF grant CNS 2007284, and in part by the iMAGiNE Consortium (https://imagine.utexas.edu/).",
+        "bbox": [
+            496,
+            851,
+            890,
+            893
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "11776",
+        "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] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6, 1",
+            "[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 39(12):2481-2495, 2017. 3",
+            "[3] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Comput. Med. Imaging Graph., 43:99-111, 2015. 6, 1",
+            "[4] Juan C Caicedo, Allen Goodman, Kyle W Karhohs, Beth A Cimini, Jeanelle Ackerman, Marzieh Haghighi, CherKeng Heng, Tim Becker, Minh Doan, Claire McQuin, et al. Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature methods, 16(12):1247-1253, 2019. 6, 7, 1",
+            "[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537, 2021. 1, 3, 6, 7",
+            "[6] Albert Cardona, Stephan Saalfeld, Stephan Preibisch, Benjamin Schmid, Anchi Cheng, Jim Pulokas, Pavel Tomancak, and Volker Hartenstein. An integrated micro-and macroarchitectural analysis of the drosophila brain by computer-assisted serial section electron microscopy. PLoS biology, 8(10):e1000502, 2010. 6, 7, 1",
+            "[7] Gongping Chen, Lei Li, Yu Dai, Jianxun Zhang, and Moi Hoon Yap. Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images. IEEE Trans. Med. Imaging, 2022. 2",
+            "[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 2, 3, 6, 7",
+            "[9] Long Chen, Hanwang Zhang, Jun Xiao, Liqiang Nie, Jian Shao, Wei Liu, and Tat-Seng Chua. Sca-cnn: Spatial and channel-wise attention in convolutional networks for image captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5659-5667, 2017. 3, 4",
+            "[10] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2017. 2, 6",
+            "[11] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Eur. Conf. Comput. Vis., pages 801-818, 2018. 2",
+            "[12] Shuhan Chen, Xiuli Tan, Ben Wang, and Xuelong Hu. Reverse attention for salient object detection. In Eur. Conf. Comput. Vis., pages 234-250, 2018. 1",
+            "[13] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional po"
+        ],
+        "bbox": [
+            78,
+            114,
+            468,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "sitional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 1",
+            "[14] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 1",
+            "[15] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In IEEE Int. Symp. Biomed. Imaging, pages 168-172. IEEE, 2018. 6, 1",
+            "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., pages 248-255. IEEE, 2009. 6",
+            "[17] Bo Dong, Wenhai Wang, Deng-Ping Fan, Jinpeng Li, Huazhu Fu, and Ling Shao. Polyp-pvt: Polyp segmentation with pyramid vision transformers. arXiv preprint arXiv:2108.06932, 2021. 1, 3, 5, 6",
+            "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2",
+            "[19] Isensee et al. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 1, 2, 3",
+            "[20] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 263-273. Springer, 2020. 1, 3, 6",
+            "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 2",
+            "[22] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2",
+            "[23] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7132-7141, 2018. 2, 3",
+            "[24] Huimin Huang, Lanfen Lin, Ruofeng Tong, Hongjie Hu, Qiaowei Zhang, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, and Jian Wu. Unet 3+: A full-scale connected unet for medical image segmentation. In ICASSP, pages 1055–1059. IEEE, 2020. 1, 2",
+            "[25] Xiaohong Huang, Zhifang Deng, Dandan Li, and Xueguang Yuan. Missformer: An effective medical image segmentation transformer. arXiv preprint arXiv:2109.07162, 2021. 6, 7"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "11777",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[26] Nabil Ibtehaz and Daisuke Kihara. Acc-unet: A completely convolutional unet model for the 2020s. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 692-702. Springer, 2023. 2",
+            "[27] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Int. Conf. Mach. Learn., pages 448-456. pmlr, 2015. 3",
+            "[28] Md Amirul Islam, Sen Jia, and Neil DB Bruce. How much position information do convolutional neural networks encode? arXiv preprint arXiv:2001.08248, 2020. 1",
+            "[29] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In Int. Conf. Multimedia Model., pages 451-462. Springer, 2020. 6, 1",
+            "[30] Taehun Kim, Hyemin Lee, and Daijin Kim. Uacanet: Uncertainty augmented context attention for polyp segmentation. In ACM Int. Conf. Multimedia, pages 2167-2175, 2021. 6",
+            "[31] Alex Krizhevsky and Geoff Hinton. Convolutional deep belief networks on CIFar-10. Unpublished manuscript, 40(7): 1-9, 2010. 4",
+            "[32] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Adv. Neural Inform. Process. Syst., 25, 2012. 2",
+            "[33] Xian Lin, Zengqiang Yan, Xianbo Deng, Chuansheng Zheng, and Li Yu. Convformer: Plug-and-play cnn-style transformers for improving medical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 642-651. Springer, 2023. 1",
+            "[34] 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 Int. Conf. Comput. Vis., pages 10012-10022, 2021. 1, 2, 3",
+            "[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11976-11986, 2022. 2",
+            "[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6",
+            "[37] Ange Lou, Shuyue Guan, and Murray Loew. Dc-unet: rethinking the u-net architecture with dual channel efficient cnn for medical image segmentation. In Med. Imaging 2021: Image Process., pages 758-768. SPIE, 2021. 1, 2",
+            "[38] Ange Lou, Shuyue Guan, Hanseok Ko, and Murray H Loew. Caranet: context axial reverse attention network for segmentation of small medical objects. In Med. Imaging 2022: Image Process., pages 81-92. SPIE, 2022. 6",
+            "[39] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Int. Conf. Mach. Learn., pages 807-814, 2010. 3",
+            "[40] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Adv. Vis. Comput. - Int. Symp., pages 15-28. Springer, 2021. 6, 1"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[41] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 1, 2, 3, 6",
+            "[42] Md Mostafijur Rahman and Radu Marculescu. Medical image segmentation via cascaded attention decoding. In IEEE/CVF Winter Conf. Appl. Comput. Vis., pages 6222-6231, 2023. 1, 3, 4, 5, 6, 7",
+            "[43] Md Mostafijur Rahman and Radu Marculescu. Multi-scale hierarchical vision transformer with cascaded attention decoding for medical image segmentation. In Med. Imaging Deep Learn., 2023. 1, 3, 5, 7",
+            "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 234-241. Springer, 2015. 1, 2, 6",
+            "[45] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4510-4520, 2018. 2, 4",
+            "[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2",
+            "[47] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1-9, 2015. 2",
+            "[48] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Int. Conf. Mach. Learn., pages 6105-6114. PMLR, 2019. 2",
+            "[49] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In Eur. Conf. Comput. Vis., pages 459-479. Springer, 2022. 1, 2",
+            "[50] Jeya Maria Jose Valanarasu and Vishal M Patel. Unext: Mlp-based rapid medical image segmentation network. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 23-33. Springer, 2022. 6, 1",
+            "[51] David Vázquez, Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Antonio M López, Adriana Romero, Michal Drozdzal, and Aaron Courville. A benchmark for endoluminal scene segmentation of colonoscopy images. J. Healthc. Eng., 2017, 2017. 6, 1",
+            "[52] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer. In AAAI, pages 2441-2449, 2022. 1, 3",
+            "[53] Hongyi Wang, Shiao Xie, Lanfen Lin, Yutaro Iwamoto, Xian-Hua Han, Yen-Wei Chen, and Ruofeng Tong. Mixed transformer u-net for medical image segmentation. In ICASSP, pages 2390–2394. IEEE, 2022. 6, 7",
+            "[54] Jinfeng Wang, Qiming Huang, Feilong Tang, Jia Meng, Jionglong Su, and Sifan Song. Stepwise feature fusion: Local guides global. arXiv preprint arXiv:2203.03635, 2022. 1, 6"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "11778",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[55] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Int. Conf. Comput. Vis., pages 568-578, 2021. 1, 2",
+            "[56] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Comput. Vis. Media, 8(3):415-424, 2022. 1, 2, 3, 5, 6",
+            "[57] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Eur. Conf. Comput. Vis., pages 3–19, 2018. 1, 3, 4",
+            "[58] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Adv. Neural Inform. Process. Syst., 34:12077-12090, 2021. 2, 3",
+            "[59] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10819-10829, 2022. 1",
+            "[60] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6848-6856, 2018. 4",
+            "[61] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 14-24. Springer, 2021. 1, 3, 6",
+            "[62] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learn. Med. Image Anal. Multimodal Learn. Clin. Decis. Support, pages 3-11. Springer, 2018. 1, 2, 6"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            571
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "11779",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/32047b53-ccf5-41e9-bc53-4e712089f958_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a18399b13fcea07f75053cd8135ed365e90e488b3c0e39d0c73e750565af542
+size 1077730

2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/full.md

@@ -0,0 +1,332 @@
+# EMCAD: Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation
+
+Md Mostafijur Rahman, Mustafa Munir, and Radu Marculescu
+
+The University of Texas at Austin
+
+Austin, Texas, USA
+
+mostafijur.rahman, mmunir, radum@utexas.edu
+
+# Abstract
+
+An efficient and effective decoding mechanism is crucial in medical image segmentation, especially in scenarios with limited computational resources. However, these decoding mechanisms usually come with high computational costs. To address this concern, we introduce EMCAD, a new efficient multi-scale convolutional attention decoder, designed to optimize both performance and computational efficiency. EMCAD leverages a unique multi-scale depth-wise convolution block, significantly enhancing feature maps through multi-scale convolutions. EMCAD also employs channel, spatial, and grouped (large-kernel) gated attention mechanisms, which are highly effective at capturing intricate spatial relationships while focusing on salient regions. By employing group and depth-wise convolution, EMCAD is very efficient and scales well (e.g., only 1.91M parameters and 0.381G FLOPs are needed when using a standard encoder). Our rigorous evaluations across 12 datasets that belong to six medical image segmentation tasks reveal that EMCAD achieves state-of-the-art (SOTA) performance with  $79.4\%$  and  $80.3\%$  reduction in #Params and #FLOPs, respectively. Moreover, EMCAD's adaptability to different encoders and versatility across segmentation tasks further establish EMCAD as a promising tool, advancing the field towards more efficient and accurate medical image analysis. Our implementation is available at https://github.com/SLDGroup/EMCAD.
+
+# 1. Introduction
+
+In the realm of medical diagnostics and therapeutic strategies, automated segmentation of medical images is vital, as it classifies pixels to identify critical regions such as lesions, tumors, or entire organs. A variety of U-shaped convolutional neural network (CNN) architectures [20, 24, 37, 41, 44, 62], notably UNet [44], UNet++ [62], UNet3+ [24], and nnU-Net [19], have become standard techniques for this purpose, achieving high-quality, high-resolution segmen
+
+tation output. Attention mechanisms [12, 17, 20, 41, 57] have also been integrated into these models to enhance feature maps and improve pixel-level classification. Although attention-based models have shown improved performance, they still face significant challenges due to the computationally expensive convolutional blocks that are typically used in conjunction with attention mechanisms.
+
+Recently, vision transformers [18] have shown promise in medical image segmentation tasks [5, 8, 17, 42, 43, 52, 54, 61] by capturing long-range dependencies among pixels through Self-attention (SA) mechanisms. Hierarchical vision transformers like Swin [34], PVT [55, 56], MaxViT [49], MERIT [43], ConvFormer [33], and MetaFormer [59] have been introduced to further improve the performance in this field. While the SA excels at capturing global information, it is less adept at understanding the local spatial context [13, 28]. To address this limitation, some approaches have integrated local convolutional attention within the decoders to better grasp spatial details. Nevertheless, these methods can still be computationally demanding because they frequently employ costly convolutional blocks. This limits their applicability to real-world scenarios where computational resources are restricted.
+
+To address the aforementioned limitations, we introduce EMCAD, an efficient multi-scale convolutional attention decoding using a new multi-scale depth-wise convolution block. More precisely, EMCAD enhances the feature maps via efficient multi-scale convolutions, while incorporating complex spatial relationships and local attention through the use of channel, spatial, and grouped (large-kernel) gated attention mechanisms. Our contributions are as follows:
+
+- New Efficient Multi-scale Convolutional Decoder: We introduce an efficient multi-scale cascaded fullyconvolutional attention decoder (EMCAD) for 2D medical image segmentation; this takes the multi-stage features of vision encoders and progressively enhances the multi-scale and multi-resolution spatial representations. EMCAD has only 0.506M parameters and 0.11G FLOPs for a tiny encoder with #channels = [32, 64, 160, 256],
+
+![](images/3af94799489a0efcba7e12ed56d3a523e2bdb917ee5fa27705063e9a6672a011.jpg)  
+Figure 1. Average DICE scores vs. #FLOPs for different methods over 10 binary medical image segmentation datasets. As shown, our approaches (PVT-EMCAD-B0 and PVT-EMCAD-B2) have the lowest #FLOPs, yet the highest DICE scores.
+
+while it has 1.91M parameters and 0.381G FLOPs for a standard encoder with #channels = [64, 128, 320, 512].
+
+- Efficient Multi-scale Convolutional Attention Module: We introduce MSCAM, a new efficient multi-scale convolutional attention module that performs depth-wise convolutions at multiple scales; this refines the feature maps produced by vision encoders and enables capturing multi-scale salient features by suppressing irrelevant regions. The use of depth-wise convolutions makes MSCAM very efficient.  
+- Large-kernel Grouped Attention Gate: We introduce a new grouped attention gate to fuse refined features with the features from skip connections. By using larger kernel  $(3 \times 3)$  group convolutions instead of point-wise convolutions in the design, we capture salient features in a larger local context with less computation.  
+- Improved Performance: We empirically show that EM-CAD can be used with any hierarchical vision encoder (e.g., PVTv2-B0, PVTv2-B2 [56]), while significantly improving the performance of 2D medical image segmentation. EMCAD produces better results than SOTA methods with a significantly lower computational cost (as shown in Figure 1) on 12 medical image segmentation benchmarks that belong to six different tasks.
+
+The remaining of this paper is organized as follows: Section 2 summarizes related work. Section 3 describes the proposed method. Section 4 explains our experimental setup and results on 12 medical image segmentation benchmarks. Section 5 covers different ablation experiments. Lastly, Section 6 concludes the paper.
+
+# 2. Related Work
+
+# 2.1. Vision encoders
+
+Convolutional Neural Networks (CNNs) [21-23, 32, 35, 45-48] have been foundational as encoders due to their pro
+
+ficiency in handling spatial relationships in images. More precisely, AlexNet [32] and VGG [46] pave the way, leveraging deep layers of convolutions to extract features progressively. GoogleNet [47] introduces the inception module, allowing more efficient computation of representations across various scales. ResNet [21] introduces residual connections, enabling the training of networks with substantially more layers by addressing the vanishing gradients problem. MobileNets [22, 45] bring CNNs to mobile devices through lightweight, depth-wise separable convolutions. EfficientNet [48] introduces a scalable architectural design to CNNs with compound scaling. Although CNNs are pivotal for many vision applications, they generally lack the ability to capture long-range dependencies within images due to their inherent local receptive fields.
+
+Recently, Vision Transformers (ViTs), pioneered by Dosovitskiy et al. [18], enabled the learning of long-range relationships among pixels using Self-attention (SA). Since then, ViTs have been enhanced by integrating CNN features [49, 56], developing novel self-attention (SA) blocks [34, 49], and introducing new architectural designs [55, 58]. The Swin Transformer [34] incorporates a sliding window attention mechanism, while SegFormer [58] leverages MixFFN blocks for hierarchical structures. PVT [55] uses spatial reduction attention, refined in PVTv2 [56] with overlapping patch embedding and a linear complexity attention layer. MaxViT [49] introduces a multi-axis self-attention to form a hierarchical CNN-transformer encoder. Although ViTs address the CNNs limitation in capturing long-range pixel dependencies [21-23, 32, 35, 45-48], they face challenges in capturing the local spatial relationships among pixels. In this paper, we aim to overcome these limitations by introducing a new multi-scale cascaded attention decoder that refines feature maps and incorporates local attention using a multi-scale convolutional attention module.
+
+# 2.2. Medical image segmentation
+
+Medical image segmentation involves pixel-wise classification to identify various anatomical structures like lesions, tumors, or organs within different imaging modalities such as endoscopy, MRI, or CT scans [8]. U-shaped networks [7, 19, 24, 26, 37, 41, 44, 62] are particularly favored due to their simple but effective encoder-decoder design. The UNet [44] pioneered this approach with its use of skip connections to fuse features at different resolution stages. UNet++ [62] evolves this design by incorporating nested encoder-decoder pathways with dense skip connections. Expanding on these ideas, UNet  $3+$  [24] introduces comprehensive skip pathways that facilitate full-scale feature integration. Further advancement comes with DC-UNet [37], which integrates a multi-resolution convolution scheme and residual paths into its skip connections. The DeepLab series, including DeepLabv3 [10] and DeepLabv3+ [11], in
+
+produce atrous convolutions and spatial pyramid pooling to handle multi-scale information. SegNet [2] uses pooling indices to upsample feature maps, preserving the boundary details. nnU-Net [19] automatically configures hyperparameters based on the specific dataset characteristics, using standard 2D and 3D UNets. Collectively, these U-shaped models have become a benchmark for success in the domain of medical image segmentation.
+
+Recently, vision transformers have emerged as a formidable force in medical image segmentation, harnessing the ability to capture pixel relationships at global scales [5, 8, 17, 42, 43, 52, 58, 61]. TransUNet [8] presents a novel blend of CNNs for local feature extraction and transformers for global context, enhancing both local and global feature capture. Swin-Unet [5] extends this by incorporating Swin Transformer blocks [34] into a U-shaped model for both encoding and decoding processes. Building on these concepts, MERIT [43] introduces a multi-scale hierarchical transformer, which employs SA across different window sizes, thus enhancing the model capacity to capture multiscale features critical for medical image segmentation.
+
+The integration of attention mechanisms has been investigated within CNNs [20, 41] and transformer-based systems [17] for enhancing medical image segmentation. PraNet [20] employs a reverse attention strategy for feature refinement. PolypPVT [17] leverages PVTv2 [56] as its backbone encoder and incorporates CBAM [57] within its decoding stages. The CASCADE [42] presents a novel cascaded decoder, combining channel [23] and spatial [9] attention to refine features at multiple stages, extracted from a transformer encoder, culminating in high-resolution segmentation outputs. While CASCADE achieves notable performance in segmenting medical images by integrating local and global insights from transformers, it is computationally inefficient due to the use of triple  $3 \times 3$  convolution layers at each decoder stage. In addition to this, it uses single-scale convolutions during decoding. Our new proposal involves the adoption of multi-scale depth-wise convolutions to mitigate these constraints.
+
+# 3. Methodology
+
+In this section, we first introduce our new EMCAD decoder and then explain two transformer-based architectures (i.e., PVT-EMCAD-B0 and PVT-EMCAD-B2) incorporating our proposed decoder.
+
+# 3.1. Efficient multi-scale convolutional attention decoding (EMCAD)
+
+In this section, we introduce our efficient multi-scale convolutional decoding (EMCAD) to process the multi-stage features extracted from pretrained hierarchical vision encoders for high-resolution semantic segmentation. As shown in Figure 2(b), EMCAD consists of efficient multi
+
+scale convolutional attention modules (MSCAMs) to robustly enhance the feature maps, large-kernel grouped attention gates (LGAGs) to refine feature maps fusing with the skip connection via gated attention mechanism, efficient up-convolution blocks (EUCBs) for up-sampling followed by enhancement of feature maps, and segmentation heads (SHs) to produce the segmentation outputs.
+
+More specifically, we use four MSCAMs to refine pyramid features (i.e.,  $X1$ ,  $X2$ ,  $X3$ ,  $X4$  in Figure 2) extracted from the four stages of the encoder. After each MSCAM, we use an SH to produce a segmentation map of that stage. Subsequently, we upscale the refined feature maps using EUCBs and add them to the outputs from the corresponding LGAGs. Finally, we add four different segmentation maps to produce the final segmentation output. Different modules of our decoder are described next.
+
+# 3.1.1 Large-kernel grouped attention gate (LGAG)
+
+We introduce a new large-kernel grouped attention gate (LGAG) to progressively combine feature maps with attention coefficients, which are learned by the network to allow higher activation of relevant features and suppression of irrelevant ones. This process employs a gating signal derived from higher-level features to control the flow of information across different stages of the network, thus enhancing its precision for medical image segmentation. Unlike Attention UNet [41] which uses  $1 \times 1$  convolution to process gating signal  $g$  (features from skip connections) and input feature map  $x$  (upsampled features), in our  $q_{att}(.)$  function, we process  $g$  and  $x$  by applying separate  $3 \times 3$  group convolutions  $GC_{g}(.)$  and  $GC_{x}(.)$ , respectively. These convolved features are then normalized using batch normalization  $(BN(.))$  [27] and merged through element-wise addition. The resultant feature map is activated through a ReLU  $(R(.))$  layer [39]. Afterward, we apply a  $1 \times 1$  convolution  $(C(.))$  followed by  $BN(.)$  layer to get a single channel feature map. We then pass the resultant single-channel feature map through a Sigmoid  $(\sigma(.))$  activation function to yield the attention coefficients. The output of this transformation is used to scale the input feature  $x$  through elementwise multiplication, producing the attention-gated feature  $LGAG(g, x)$ . The  $LGAG(.)$  (Figure 2(g)) can be formulated as in Equations 1 and 2:
+
+$$
+q _ {a t t} (g, x) = R \left(B N \left(G C _ {g} (g) + B N \left(G C _ {x} (x)\right)\right)\right) \tag {1}
+$$
+
+$$
+L G A G (g, x) = x * \sigma \left(B N \left(C \left(q _ {a t t} (g, x)\right)\right)\right) \tag {2}
+$$
+
+Due to using  $3 \times 3$  kernel group convolutions in  $q_{att}(.)$ , our LGAG captures comparatively larger spatial contexts with less computational cost.
+
+![](images/680916196f86f563262f187b1e9498563b3b45b7b797b45690ae9a0e76ac7c8c.jpg)  
+Figure 2. Hierarchical encoder with newly proposed EMCAD decoder architecture. (a) CNN or transformer encoder with four hierarchical stages, (b) EMCAD decoder, (c) Efficient up-convolution block (EUCB), (d) Multi-scale convolutional attention module (MSCAM), (e) Multi-scale convolution block (MSCB), (f) Multi-scale (parallel) depth-wise convolution (MSDC), (g) Large-kernel grouped attention gate (LGAG), (h) Channel attention block (CAB), and (i) Spatial attention block (SAB). X1, X2, X3, and X4 are the features from the four stages of the hierarchical encoder. p1, p2, p3, and p4 are output segmentation maps from four stages of our decoder.
+
+# 3.1.2 Multi-scale convolutional attention module (MSCAM)
+
+We introduce an efficient multi-scale convolutional attention module to refine the feature maps. MSCAM consists of a channel attention block  $(CAB(\cdot))$  to put emphasis on pertinent channels, a spatial attention block [9]  $(SAB(\cdot))$  to capture the local contextual information, and an efficient multi-scale convolution block  $(MSCB(\cdot))$  to enhance the feature maps preserving contextual relationships. The  $MSCAM(\cdot)$  (Figure 2(d)) is given in Equation 3:
+
+$$
+M S C A M (x) = M S C B (S A B (C A B (x))) \tag {3}
+$$
+
+where  $x$  is the input tensor. Due to using depth-wise convolution in multiple scales, our MSCAM is more effective with significantly lower computational cost than the convolutional attention module (CAM) proposed in [42].
+
+Multi-scale Convolution Block (MSCB): We introduce an efficient multi-scale convolution block to enhance the features generated by our cascaded expanding path. In our MSCB, we follow the design of the inverted residual block (IRB) of MobileNetV2 [45]. However, unlike IRB, our MSCB performs depth-wise convolution at multiple scales and uses channel_shuffle [60] to shuffle channels across groups. More specifically, in our MSCB, we first expand the number of channels (i.e., expansion_factor = 2) using a point-wise  $(1 \times 1)$  convolution layers  $PWC_{1}(\cdot)$  followed by a batch normalization layer  $BN(\cdot)$  and a ReLU6 [31] activation layer  $R6(\cdot)$ . We then use a multi-scale depth-wise convolution  $MSDC(\cdot)$  to capture both multi-scale and multi-resolution contexts. As depth-wise convolution overlooks
+
+the relationships among channels, we use a channel_shuffle operation to incorporate relationships among channels. Afterward, we use another point-wise convolution  $PWC_{2}(.)$  followed by a  $BN(.)$  to transform back the original #channels, which also encodes dependency among channels. The  $MSCB(\cdot)$  (Figure 2(e)) is formulated as in Equation 4:
+
+$$
+M S C B (x) = B N \left(P W C _ {2} \left(C S \left(M S D C \left(R 6 \left(B N \left(P W C _ {1} (x)\right)\right)\right)\right)\right)\right) \tag {4}
+$$
+
+where parallel  $MSDC(.)$  (Figure 2(f)) for different kernel sizes  $(KS)$  can be formulated using Equation 5:
+
+$$
+M S D C (x) = \sum_ {k s \in K S} D W C B _ {k s} (x) \tag {5}
+$$
+
+where  $DWCB_{ks}(x) = R6(BN(DWC_{ks}(x)))$ . Here,  $DWC_{ks}()$  is a depth-wise convolution with the kernel size  $ks$ .  $BN()$  and  $R6()$  are batch normalization and ReLU6 activation, respectively. Additionally, our sequential  $MSDC()$  uses the recursively updated input  $x$ , where the input  $x$  is residually connected to the previous  $DWCB_{ks}()$  for better regularization as in Equation 6:
+
+$$
+x = x + D W C B _ {k s} (x) \tag {6}
+$$
+
+Channel Attention Block (CAB): We use channel attention block to assign different levels of importance to each channel, thus emphasizing more relevant features while suppressing less useful ones. Basically, the CAB identifies which feature maps to focus on (and then refine them). Following [57], in CAB, we first apply the adaptive maximum pooling  $(P_{m}(\cdot))$  and adaptive average pooling  $(P_{a}(\cdot))$  to the spatial dimensions (i.e., height and width) to extract the most significant feature of the entire feature map per
+
+channel. Then, for each pooled feature map, we reduce the number of channels  $r = 1 / 16$  times separately using a point-wise convolution  $(C_1(\cdot))$  followed by a ReLU activation  $(R)$ . Afterward, we recover the original channels using another point-wise convolution  $(C_2(\cdot))$ . We then add both recovered feature maps and apply Sigmoid  $(\sigma)$  activation to estimate attention weights. Finally, we incorporate these weights to input  $x$  using the Hadamard product  $(\otimes)$ . The  $CAB(\cdot)$  (Figure 2(h)) is defined using Equation 7:
+
+$$
+C A B (x) = \sigma \left(C _ {2} \left(R \left(C _ {1} \left(P _ {m} (x)\right)\right)\right) + C _ {2} \left(R \left(C _ {1} \left(P _ {a} (x)\right)\right)\right)\right) * x \tag {7}
+$$
+
+Spatial Attention Block (SAB): We use spatial attention to mimic the attentional processes of the human brain by focusing on specific parts of an input image. Basically, the SAB determines where to focus in a feature map; then it enhances those features. This process enhances the model's ability to recognize and respond to relevant spatial features, which is crucial for image segmentation where the context and location of objects significantly influence the output. In SAB, we first pool maximum  $(Ch_{max}(\cdot))$  and average  $(Ch_{avg}(\cdot))$  values along the channel dimension to pay attention to local features. Then, we use a large kernel (i.e.,  $7\times 7$  as in [17]) convolution layer to enhance local contextual relationships among features. Afterward, we apply the Sigmoid activation  $(\sigma)$  to calculate attention weights. Finally, we feed these weights to the input  $x$  (using Hadamard product  $(\otimes)$  to attend information in a more targeted way. The  $SAB(.)$  (Figure 2(i)) is defined using Equation 8:
+
+$$
+S A B (x) = \sigma \left(L K C \left(\left[ C h _ {m a x} (x), C h _ {a v g} (x) \right]\right)\right) * x \tag {8}
+$$
+
+# 3.1.3 Efficient up-convolution block (EUCB)
+
+We use an efficient up-convolution block to progressively upsample the feature maps of the current stage to match the dimension and resolution of the feature maps from the next skip connection. The EUCB first uses an UpSampling  $Up(\cdot)$  with scale-factor 2 to upscale the feature maps. Then, it enhances the upscaled feature maps by applying a  $3 \times 3$  depth-wise convolution  $DWC(\cdot)$  followed by a  $BN(\cdot)$  and a  $ReLU(\cdot)$  activation. Finally, a  $1 \times 1$  convolution  $C_{1 \times 1}(\cdot)$  is used to reduce the #channels to match with the next stage. The  $EUCB(\cdot)$  (Figure 2(c)) is formulated as in Equation 9:
+
+$$
+E U C B (x) = C _ {1 \times 1} \left(R e L U \left(B N \left(D W C (U p (x))\right)\right)\right) \tag {9}
+$$
+
+Due to using depth-wise convolution instead of  $3 \times 3$  convolution, our EUCB is very efficient.
+
+# 3.1.4 Segmentation head (SH)
+
+We use segmentation heads to produce the segmentation outputs from the refined feature maps of four stages of the decoder. The SH layer applies a  $1 \times 1$  convolution  $Conv_{1 \times 1}(\cdot)$  to the refined feature maps having  $ch_i$  channels ( $ch_i$  is the #channels in the feature map of stage  $i$ ) and
+
+produces output with #channels equal to #classes in target dataset for multi-class but 1 channel for binary segmentation. The  $SH(\cdot)$  is formulated as in Equation 10:
+
+$$
+S H (x) = \operatorname {C o n v} _ {1 \times 1} (x) \tag {10}
+$$
+
+# 3.2. Overall architecture
+
+To show the generalization, effectiveness, and ability to process multi-scale features for medical image segmentation, we integrate our EMCAD decoder alongside tiny (PVTv2-B0) and standard (PVTv2-B2) networks of PVTv2 [56]. However, our decoder is adaptable and seamlessly compatible with other hierarchical backbone networks.
+
+PVTv2 differs from conventional transformer patch embedding modules by applying convolutional operations for consistent spatial information capture. Using PVTv2-b0 (Tiny) and PVTv2-b2 (Standard) encoders [56], we develop the PVT-EMCAD-B0 and PVT-EMCAD-B2 architectures. To adopt PVTv2, we first extract the features (X1, X2, X3, and X4) from four layers and feed them (i.e., X4 in the upsample path and X3, X2, X1 in the skip connections) into our EMCAD decoder as shown in Figure 2(a-b). EMCAD then processes them and produces four segmentation maps that correspond to the four stages of the encoder network.
+
+# 3.3. Multi-stage loss and outputs aggregation
+
+Our EMCAD decoder's four segmentation heads produce four prediction maps  $p_1, p_2, p_3$ , and  $p_4$  across its stages.
+
+Loss aggregation: We adopt a combinatorial approach to loss combination called MUTATION, inspired by the work of MERIT [43] for multi-class segmentation. This involves calculating the loss for all possible combinations of predictions derived from 4 heads, totaling  $2^{4} - 1 = 15$  unique predictions, and then summing these losses. We focus on minimizing this cumulative combinatorial loss during the training process. For binary segmentation, we optimize the additive loss like [42] with an additional term  $\mathcal{L}_{p_1 + p_2 + p_3 + p_4}$  as in Equation 11:
+
+$$
+\mathcal {L} _ {\text {t o t a l}} = \alpha \mathcal {L} _ {p _ {1}} + \beta \mathcal {L} _ {p _ {2}} + \gamma \mathcal {L} _ {p _ {3}} + \zeta \mathcal {L} _ {p _ {4}} + \delta \mathcal {L} _ {p _ {1} + p _ {2} + p _ {3} + p _ {4}} \tag {11}
+$$
+
+where  $\mathcal{L}_{p_1},\mathcal{L}_{p_2},\mathcal{L}_{p_3}$  , and  $\mathcal{L}_{p_4}$  are the losses of each individual prediction maps.  $\alpha = \beta = \gamma = \zeta = \delta = 1.0$  are the weights assigned to each loss.
+
+Output segmentation maps aggregation: We consider the prediction map,  $p_4$ , from the last stage of our decoder as the final segmentation map. Then, we obtain the final segmentation output by employing a Sigmoid function for binary or a Softmax function for multi-class segmentation.
+
+# 4. Experiments
+
+In this section, we present the details of our implementation followed by a comparative analysis of our PVT-EMCAD-B0 and PVT-EMCAD-B2 against SOTA methods. Datasets and evaluation metrics are in Supplementary Section 7.
+
+<table><tr><td rowspan="2">Methods</td><td rowspan="2">#Params</td><td rowspan="2">#FLOPs</td><td colspan="5">Polyp</td><td colspan="2">Skin Lesion</td><td colspan="2">Cell</td><td rowspan="2">BUSI</td><td rowspan="2">Avg.</td></tr><tr><td>Clinic</td><td>Colon</td><td>ETIS</td><td>Kvasir</td><td>BKAI</td><td>ISIC17</td><td>ISIC18</td><td>DSB18</td><td>EM</td></tr><tr><td>UNet [44]</td><td>34.53M</td><td>65.53G</td><td>92.11</td><td>83.95</td><td>76.85</td><td>82.87</td><td>85.05</td><td>83.07</td><td>86.67</td><td>92.23</td><td>95.46</td><td>74.04</td><td>85.23</td></tr><tr><td>UNet++ [62]</td><td>9.16M</td><td>34.65G</td><td>92.17</td><td>87.88</td><td>77.40</td><td>83.36</td><td>84.07</td><td>82.98</td><td>87.46</td><td>91.97</td><td>95.48</td><td>74.76</td><td>85.75</td></tr><tr><td>AttnUNet [41]</td><td>34.88M</td><td>66.64G</td><td>92.20</td><td>86.46</td><td>76.84</td><td>83.49</td><td>84.07</td><td>83.66</td><td>87.05</td><td>92.22</td><td>95.55</td><td>74.48</td><td>85.60</td></tr><tr><td>DeepLabv3+ [10]</td><td>39.76M</td><td>14.92G</td><td>93.24</td><td>91.92</td><td>90.73</td><td>89.06</td><td>89.74</td><td>83.84</td><td>88.64</td><td>92.14</td><td>94.96</td><td>76.81</td><td>89.11</td></tr><tr><td>PraNet [20]</td><td>32.55M</td><td>6.93G</td><td>91.71</td><td>89.16</td><td>83.84</td><td>84.82</td><td>85.56</td><td>83.03</td><td>88.56</td><td>89.89</td><td>92.37</td><td>75.14</td><td>86.41</td></tr><tr><td>CaraNet [38]</td><td>46.64M</td><td>11.48G</td><td>94.08</td><td>91.19</td><td>90.25</td><td>89.74</td><td>89.71</td><td>85.02</td><td>90.18</td><td>89.15</td><td>92.78</td><td>77.34</td><td>88.94</td></tr><tr><td>UACANet-L [30]</td><td>69.16M</td><td>31.51G</td><td>94.16</td><td>91.02</td><td>89.77</td><td>90.17</td><td>90.35</td><td>83.72</td><td>89.76</td><td>88.86</td><td>89.28</td><td>76.96</td><td>88.41</td></tr><tr><td>SSFormer-L [54]</td><td>66.22M</td><td>17.28G</td><td>94.18</td><td>92.11</td><td>90.16</td><td>91.47</td><td>91.14</td><td>85.28</td><td>90.25</td><td>92.03</td><td>94.95</td><td>78.76</td><td>90.03</td></tr><tr><td>PolypPVT [17]</td><td>25.11M</td><td>5.30G</td><td>94.13</td><td>91.53</td><td>89.93</td><td>91.56</td><td>91.17</td><td>85.56</td><td>90.36</td><td>90.69</td><td>94.40</td><td>79.35</td><td>89.87</td></tr><tr><td>TransUNet [8]</td><td>105.32M</td><td>38.52G</td><td>93.90</td><td>91.63</td><td>87.79</td><td>91.08</td><td>89.17</td><td>85.00</td><td>89.16</td><td>92.04</td><td>95.27</td><td>78.30</td><td>89.33</td></tr><tr><td>SwinUNet [5]</td><td>27.17M</td><td>6.2G</td><td>92.42</td><td>89.27</td><td>85.10</td><td>89.59</td><td>87.61</td><td>83.97</td><td>89.26</td><td>91.03</td><td>94.47</td><td>77.38</td><td>88.01</td></tr><tr><td>TransFuse [61]</td><td>143.74M</td><td>82.71G</td><td>93.62</td><td>90.35</td><td>86.91</td><td>90.24</td><td>87.47</td><td>84.89</td><td>89.62</td><td>90.85</td><td>94.35</td><td>79.36</td><td>88.77</td></tr><tr><td>UNeXt [50]</td><td>1.47M</td><td>0.57G</td><td>90.20</td><td>83.84</td><td>74.03</td><td>77.88</td><td>77.93</td><td>82.74</td><td>87.78</td><td>86.01</td><td>93.81</td><td>74.71</td><td>82.89</td></tr><tr><td>PVT-CASCADE [42]</td><td>34.12M</td><td>7.62G</td><td>94.53</td><td>91.60</td><td>91.03</td><td>92.05</td><td>92.14</td><td>85.50</td><td>90.41</td><td>92.35</td><td>95.42</td><td>79.21</td><td>90.42</td></tr><tr><td>PVT-EMCAD-B0 (Ours)</td><td>3.92M</td><td>0.84G</td><td>94.60</td><td>91.71</td><td>91.65</td><td>91.95</td><td>91.30</td><td>85.67</td><td>90.70</td><td>92.46</td><td>95.35</td><td>79.80</td><td>90.52</td></tr><tr><td>PVT-EMCAD-B2 (Ours)</td><td>26.76M</td><td>5.6G</td><td>95.21</td><td>92.31</td><td>92.29</td><td>92.75</td><td>92.96</td><td>85.95</td><td>90.96</td><td>92.74</td><td>95.53</td><td>80.25</td><td>91.10</td></tr></table>
+
+Table 1. Results of binary medical image segmentation (i.e., polyp, skin lesion, cell, and breast cancer). We reproduce the results of SOTA methods using their publicly available implementation with our train-val-test splits of 80:10:10. #FLOPs of all the methods are reported for  $256 \times 256$  inputs, except Swin-UNet ( $224 \times 224$ ). All results are averaged over five runs. Best results are shown in bold.  
+
+<table><tr><td rowspan="2">Architectures</td><td colspan="3">Average</td><td rowspan="2">Aorta</td><td rowspan="2">GB</td><td rowspan="2">KL</td><td rowspan="2">KR</td><td rowspan="2">Liver</td><td rowspan="2">PC</td><td rowspan="2">SP</td><td rowspan="2">SM</td></tr><tr><td>DICE↑</td><td>HD95↓</td><td>mIoU↑</td></tr><tr><td>UNet [44]</td><td>70.11</td><td>44.69</td><td>59.39</td><td>84.00</td><td>56.70</td><td>72.41</td><td>62.64</td><td>86.98</td><td>48.73</td><td>81.48</td><td>67.96</td></tr><tr><td>AttnUNet [41]</td><td>71.70</td><td>34.47</td><td>61.38</td><td>82.61</td><td>61.94</td><td>76.07</td><td>70.42</td><td>87.54</td><td>46.70</td><td>80.67</td><td>67.66</td></tr><tr><td>R50+UNet [8]</td><td>74.68</td><td>36.87</td><td>-</td><td>84.18</td><td>62.84</td><td>79.19</td><td>71.29</td><td>93.35</td><td>48.23</td><td>84.41</td><td>73.92</td></tr><tr><td>R50+AttnUNet [8]</td><td>75.57</td><td>36.97</td><td>-</td><td>55.92</td><td>63.91</td><td>79.20</td><td>72.71</td><td>93.56</td><td>49.37</td><td>87.19</td><td>74.95</td></tr><tr><td>SSFormer [54]</td><td>78.01</td><td>25.72</td><td>67.23</td><td>82.78</td><td>63.74</td><td>80.72</td><td>78.11</td><td>93.53</td><td>61.53</td><td>87.07</td><td>76.61</td></tr><tr><td>PolypPVT [17]</td><td>78.08</td><td>25.61</td><td>67.43</td><td>82.34</td><td>66.14</td><td>81.21</td><td>73.78</td><td>94.37</td><td>59.34</td><td>88.05</td><td>79.4</td></tr><tr><td>TransUNet [8]</td><td>77.61</td><td>26.9</td><td>67.32</td><td>86.56</td><td>60.43</td><td>80.54</td><td>78.53</td><td>94.33</td><td>58.47</td><td>87.06</td><td>75.00</td></tr><tr><td>SwinUNet [5]</td><td>77.58</td><td>27.32</td><td>66.88</td><td>81.76</td><td>65.95</td><td>82.32</td><td>79.22</td><td>93.73</td><td>53.81</td><td>88.04</td><td>75.79</td></tr><tr><td>MT-UNet [53]</td><td>78.59</td><td>26.59</td><td>-</td><td>87.92</td><td>64.99</td><td>81.47</td><td>77.29</td><td>93.06</td><td>59.46</td><td>87.75</td><td>76.81</td></tr><tr><td>MISSformer [25]</td><td>81.96</td><td>18.20</td><td>-</td><td>86.99</td><td>68.65</td><td>85.21</td><td>82.00</td><td>94.41</td><td>65.67</td><td>91.92</td><td>80.81</td></tr><tr><td>PVT-CASCADE [42]</td><td>81.06</td><td>20.23</td><td>70.88</td><td>83.01</td><td>70.59</td><td>82.23</td><td>80.37</td><td>94.08</td><td>64.43</td><td>90.1</td><td>83.69</td></tr><tr><td>TransCASCADE [42]</td><td>82.68</td><td>17.34</td><td>73.48</td><td>86.63</td><td>68.48</td><td>87.66</td><td>84.56</td><td>94.43</td><td>65.33</td><td>90.79</td><td>83.52</td></tr><tr><td>PVT-EMCAD-B0 (Ours)</td><td>81.97</td><td>17.39</td><td>72.64</td><td>87.21</td><td>66.62</td><td>87.48</td><td>83.96</td><td>94.57</td><td>62.00</td><td>92.66</td><td>81.22</td></tr><tr><td>PVT-EMCAD-B2 (Ours)</td><td>83.63</td><td>15.68</td><td>74.65</td><td>88.14</td><td>68.87</td><td>88.08</td><td>84.10</td><td>95.26</td><td>68.51</td><td>92.17</td><td>83.92</td></tr></table>
+
+Table 2. Results of abdomen organ segmentation on Synapse Multi-organ dataset. DICE scores are reported for individual organs. Results of UNet, AttnUNet, PolypPVT, SSFormerPVT, TransUNet, and SwinUNet are taken from [42].  $\uparrow (\downarrow)$  denotes the higher (lower) the better.  $^{\prime} - ^{\prime}$  means missing data from the source. EMCAD results are averaged over five runs. Best results are shown in bold.
+
+# 4.1. Implementation details
+
+We implement our network and conduct experiments using Pytorch 1.11.0 on a single NVIDIA RTX A6000 GPU with 48GB of memory. We utilize ImageNet [16] pretrained PVTv2-b0 and PVTv2-b2 [56] as encoders. In the MSDC of our decoder, we set the multi-scale kernels [1,3,5] through an ablation study. We use the parallel arrangement of depth-wise convolutions in all experiments. Our models are trained using the AdamW optimizer [36] with a learning rate and weight decay of  $1e - 4$ . We generally train for 200 epochs with a batch size of 16, except for Synapse multi-organ (300 epochs, batch size 6) and ACDC cardiac organ (400 epochs, batch size 12), saving the best model based on the DICE score. We resize images
+
+to  $352 \times 352$  and use a multi-scale  $\{0.75, 1.0, 1.25\}$  training strategy with a gradient clip limit of 0.5 for ClinicDB [3], Kvasir [29], ColonDB [51], ETIS [51], BKAI [40], ISIC17 [15], and ISIC18 [15], while we resize images to  $256 \times 256$  for BUSI [1], EM [6], and DSB18 [4]. For Synapse and ACDC datasets, images are resized to  $224 \times 224$ , with random rotation and flipping augmentations, optimizing a combined Cross-entropy (0.3) and DICE (0.7) loss. For binary segmentation, we utilize the combined weighted BinaryCrossEntropy (BCE) and weighted IoU loss function.
+
+# 4.2. Results
+
+We compare our architectures (i.e., PVT-EMCAD-B0 and PVT-EMCAD-B2) with SOTA CNN and transformer-based
+
+![](images/68e95713cdd7aa1342c98b587e762e015c98e20ad9f521d47ceae1aa69432ee7.jpg)  
+Figure 3. Average DICE scores vs. #Params for different methods over 10 binary medical image segmentation datasets. As shown, our proposed approaches (PVT-EMCAD-B0 and PVT-EMCAD-B2) have the fewest parameters, yet the highest DICE scores.
+
+<table><tr><td>Methods</td><td>Avg. DICE</td><td>RV</td><td>Myo</td><td>LV</td></tr><tr><td>R50+UNet [8]</td><td>87.55</td><td>87.10</td><td>80.63</td><td>94.92</td></tr><tr><td>R50+AttnUNet [8]</td><td>86.75</td><td>87.58</td><td>79.20</td><td>93.47</td></tr><tr><td>ViT+CUP [8]</td><td>81.45</td><td>81.46</td><td>70.71</td><td>92.18</td></tr><tr><td>R50+ViT+CUP [8]</td><td>87.57</td><td>86.07</td><td>81.88</td><td>94.75</td></tr><tr><td>TransUNet [8]</td><td>89.71</td><td>86.67</td><td>87.27</td><td>95.18</td></tr><tr><td>SwinUNet [5]</td><td>88.07</td><td>85.77</td><td>84.42</td><td>94.03</td></tr><tr><td>MT-UNet [53]</td><td>90.43</td><td>86.64</td><td>89.04</td><td>95.62</td></tr><tr><td>MISSformer [25]</td><td>90.86</td><td>89.55</td><td>88.04</td><td>94.99</td></tr><tr><td>PVT-CASCADE [42]</td><td>91.46</td><td>89.97</td><td>88.9</td><td>95.50</td></tr><tr><td>TransCASCade [42]</td><td>91.63</td><td>90.25</td><td>89.14</td><td>95.50</td></tr><tr><td>Cascaded MERIT [43]</td><td>91.85</td><td>90.23</td><td>89.53</td><td>95.80</td></tr><tr><td>PVT-EMCAD-B0 (Ours)</td><td>91.34±0.2</td><td>89.37</td><td>88.99</td><td>95.65</td></tr><tr><td>PVT-EMCAD-B2 (Ours)</td><td>92.12±0.2</td><td>90.65</td><td>89.68</td><td>96.02</td></tr></table>
+
+segmentation methods on 12 datasets that belong to six medical image segmentation tasks. Qualitative results are in the Supplementary Section 7.3.
+
+# 4.2.1 Results of binary medical image segmentation
+
+Results for different methods on 10 binary medical image segmentation datasets are shown in Table 1 and Figure 1. Our PVT-EMCAD-B2 attains the highest average DICE score  $(91.10\%)$  with only  $26.76\mathrm{M}$  parameters and 5.6G FLOPs. The multi-scale depth-wise convolution in our EMCAD decoder, combined with the transformer encoder, contributes to these performance gains.
+
+Polyp segmentation: Table 1 reveals that our PVT-EMCAD-B2 surpasses all SOTA methods in five polyp segmentation datasets. PVT-EMCAD-B2 achieves DICE score improvements of  $1.08\%$ ,  $0.78\%$ ,  $2.36\%$ ,  $1.19\%$ , and  $1.79\%$  over PolypPVT in ClinicDB, ColonDB, ETIS, Kvasir, and BKAI-IGI, despite having slightly more parameters and FLOPs. The smallest model UNeXt, ex
+
+Table 3. Results of cardiac organ segmentation on ACDC dataset. DICE scores  $(\%)$  are reported for individual organs. We get the results of SwinUNet from [42]. Best results are shown in bold.  
+
+<table><tr><td colspan="3">Components</td><td colspan="2">#FLOPs(G)</td><td rowspan="2">#Params (M)</td><td rowspan="2">Avg DICE</td></tr><tr><td>Cascaded</td><td>LGAG</td><td>MSCAM</td><td>224</td><td>256</td></tr><tr><td>No</td><td>No</td><td>No</td><td>0</td><td>0</td><td>0</td><td>80.10±0.2</td></tr><tr><td>Yes</td><td>No</td><td>No</td><td>0.100</td><td>0.131</td><td>0.224</td><td>81.08±0.2</td></tr><tr><td>Yes</td><td>Yes</td><td>No</td><td>0.108</td><td>0.141</td><td>0.235</td><td>81.92±0.2</td></tr><tr><td>Yes</td><td>No</td><td>Yes</td><td>0.373</td><td>0.487</td><td>1.898</td><td>82.86±0.3</td></tr><tr><td>Yes</td><td>Yes</td><td>Yes</td><td>0.381</td><td>0.498</td><td>1.91</td><td>83.63±0.3</td></tr></table>
+
+Table 4. Effect of different components of EMCAD with PVTv2-b2 encoder on Synapse multi-organ dataset. #FLOPs are reported for input resolution of  $224 \times 224$  and  $256 \times 256$ . All results are averaged over five runs. Best results are shown in bold.
+
+hibits the worst performance in all five polyp segmentation datasets. Our smaller model with only 3.92M parameters and 0.84G FLOPs also outperforms all the methods except PVT-CASCADE (in Kvasir and BKAI-IGH) and SSFormer-L (in ColonDB), which achieve the best performance among SOTA methods. In conclusion, our PVT-EMCAD-B2 achieves the new SOTA results in these five polyp segmentation datasets.
+
+Skin lesion segmentation: Table 1 shows PVT-EMCAD-B2's strong performance on ISIC17 and ISIC18 skin lesion segmentation datasets, achieving DICE scores of  $85.95\%$  and  $90.96\%$ , surpassing DeepLabV3+ by  $2.11\%$  and  $2.32\%$ . It also beats the nearest method PVT-CASCADE by  $0.45\%$  and  $0.55\%$  in ISIC17 and ISIC18, respectively, though our decoder is significantly more efficient than CASCADE. Our PVT-EMCAD-B0 also shows huge potential in point care applications like skin lesion segmentation with only 3.92M parameters and 0.84G FLOPs.
+
+Cell segmentation: To evaluate our method's effectiveness in biological imaging, we use DSB18 [4] for cell nuclei and EM [6] for cell structure segmentation. As Table 1 indicates, our PVT-EMCAD-B2 sets a SOTA benchmark in cell nuclei segmentation on DSB18, outperforming DeepLabv3+, TransFuse, and PVT-CASCADE. On the EM dataset, PVT-EMCAD-B2 secures the second-best DICE score (95.53%), offering significantly lower computational costs than the top-performing AttnUNet (95.55%).
+
+Breast cancer segmentation: We conduct experiments on the BUSI dataset for breast cancer segmentation in ultrasound images. Our PVT-EMCAD-B2 achieves the SOTA DICE score  $(80.25\%)$  on this dataset. Furthermore, our PVT-EMCAD-B0 outperforms the computationally similar method UNeXt by a notable margin of  $5.54\%$ .
+
+# 4.2.2 Results of abdomen organ segmentation
+
+Table 2 shows that our PVT-EMCAD-B2 excels in abdomen organ segmentation on the Synapse multi-organ dataset, achieving the highest average DICE score of  $83.63\%$  and surpassing all SOTA CNN- and transformer-based methods. It outperforms PVT-CASCADE by  $2.57\%$  in DICE score and 4.55 in HD95 distance, indicating superior organ boundary location. Our EMCAD decoder boosts individ
+
+<table><tr><td>Conv. kernels</td><td>[1]</td><td>[3]</td><td>[5]</td><td>[1,3]</td><td>[3,3]</td><td>[1,3,5]</td><td>[3,3,3]</td><td>[3,5,7]</td><td>[1,3,5,7]</td><td>[1,3,5,7,9]</td></tr><tr><td>Synapse</td><td>82.43</td><td>82.79</td><td>82.74</td><td>82.98</td><td>82.81</td><td>83.63</td><td>82.92</td><td>83.11</td><td>83.57</td><td>83.34</td></tr><tr><td>ClinicDB</td><td>94.81</td><td>94.90</td><td>94.98</td><td>95.13</td><td>95.06</td><td>95.21</td><td>95.15</td><td>95.03</td><td>95.18</td><td>95.07</td></tr></table>
+
+Table 5. Effect of multi-scale kernels in the depth-wise convolution of MSDC on ClinicDB and Synapse multi-organ datasets. We use the PVTv2-b2 encoder for these experiments. All results are averaged over five runs. Best results are highlighted in bold.  
+
+<table><tr><td>Encoders</td><td>Decoders</td><td>#FLOPs(G)</td><td>#Params(M)</td><td>DICE (%)</td></tr><tr><td>PVTv2-B0</td><td>CASCADE</td><td>0.439</td><td>2.32</td><td>80.54</td></tr><tr><td>PVTv2-B0</td><td>EMCAD (Ours)</td><td>0.110</td><td>0.507</td><td>81.97</td></tr><tr><td>PVTv2-B2</td><td>CASCADE</td><td>1.93</td><td>9.27</td><td>82.78</td></tr><tr><td>PVTv2-B2</td><td>EMCAD (Ours)</td><td>0.381</td><td>1.91</td><td>83.63</td></tr></table>
+
+Table 6. Comparison with the baseline decoder on Synapse Multi-organ dataset. We only report the #FLOPs (with input resolution of  $224 \times 224$ ) and the #parameters of the decoders. All the results are averaged over five runs. Best results are shown in bold.
+
+ual organ segmentation, significantly outperforming SOTA methods on six of eight organs.
+
+# 4.2.3 Results of cardiac organ segmentation
+
+Table 3 shows the DICE scores of our PVT-EMCAD-B2 and PVT-EMCAD-B0 along with other SOTA methods, on the MRI images of the ACDC dataset for cardiac organ segmentation. Our PVT-EMCAD-B2 achieves the highest average DICE score of  $92.12\%$ , thus improving about  $0.27\%$  over Cascaded MERIT though our network has significantly lower computational cost. Besides, PVT-EMCAD-B2 has better DICE scores in all three organ segmentations.
+
+# 5. Ablation Studies
+
+In this section, we conduct ablation studies to explore different aspects of our architectures and the experimental framework. More ablations are in Supplementary Section 8.
+
+# 5.1. Effect of different components of EMCAD
+
+We conduct a set of experiments on the Synapse multi-organ dataset to understand the effect of different components of our EMCAD decoder. We start with only the encoder and add different modules such as Cascaded structure, LGAG, and MSCAM to understand their effect. Table 4 exhibits that the cascaded structure of the decoder helps to improve performance over the non-cascaded one. The incorporation of LGAG and MSCAM improves performance, however, MSCAM proves to be more effective. When both the LGAG and MSCAM modules are used together, it produces the best DICE score of  $83.63\%$ . It is also evident that there is about  $3.53\%$  improvement in the DICE score with an additional 0.381G FLOPs and 1.91M parameters.
+
+# 5.2. Effect of multi-scale kernels in MSCAM
+
+We have conducted another set of experiments on Synapse multi-organ and ClinicDB datasets to understand the effect of different multi-scale kernels used for depth-wise convolutions in MSDC. Table 5 reports these results which show
+
+that performance improves from  $1 \times 1$  to  $3 \times 3$  kernel. When  $1 \times 1$  kernel is used together with  $3 \times 3$  it improves more than when using them alone. However, when two  $3 \times 3$  kernels are used together, performance drops. The incorporation of a  $5 \times 5$  kernel with  $1 \times 1$  and  $3 \times 3$  kernels further improves the performance and it achieves the best results in both Synapse multi-organ and ClinicDB datasets. If we add additional larger kernels (e.g.,  $7 \times 7$ ,  $9 \times 9$ ), the performance of both datasets drops. Based on these empirical observations, we choose [1, 3, 5] kernels in all our experiments.
+
+# 5.3. Comparison with the baseline decoder
+
+In Table 6, we report the experimental results with the computational complexity of our EMCAD decoder and a baseline decoder, namely CASCADE. From Table 6, we can see that our EMCAD decoder with PVTv2-b2 requires  $80.3\%$  fewer FLOPs and  $79.4\%$  fewer parameters to outperform (by  $0.85\%$ ) the respective CASCADE decoder. Similarly, our EMCAD decoder with PVTv2-B0 achieves  $1.43\%$  better DICE score than the CASCADE decoder with  $78.1\%$  fewer parameters and  $74.9\%$  fewer FLOPs.
+
+# 6. Conclusions
+
+In this paper, we have presented EMCAD, a new and efficient multi-scale convolutional attention decoder designed for multi-stage feature aggregation and refinement in medical image segmentation. EMCAD employs a multi-scale depth-wise convolution block, which is key for capturing diverse scale information within feature maps, a critical factor for precision in medical image segmentation. This design choice, using depth-wise convolutions instead of standard  $3 \times 3$  convolution blocks, makes EMCAD notably efficient.
+
+Our experiments reveal that EMCAD surpasses the recent CASCADE decoder in DICE scores with  $79.4\%$  fewer parameters and  $80.3\%$  less FLOPs. Our extensive experiments also confirm EMCAD's superior performance compared to SOTA methods across 12 public datasets covering six different 2D medical image segmentation tasks. EMCAD's compatibility with smaller encoders makes it an excellent fit for point-of-care applications while maintaining high performance. We anticipate that our EMCAD decoder will be a valuable asset in enhancing a variety of medical image segmentation and semantic segmentation tasks.
+
+Acknowledgements: This work is supported in part by the NSF grant CNS 2007284, and in part by the iMAGiNE Consortium (https://imagine.utexas.edu/).
+
+# References
+
+[1] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6, 1  
+[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 39(12):2481-2495, 2017. 3  
+[3] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Comput. Med. Imaging Graph., 43:99-111, 2015. 6, 1  
+[4] Juan C Caicedo, Allen Goodman, Kyle W Karhohs, Beth A Cimini, Jeanelle Ackerman, Marzieh Haghighi, CherKeng Heng, Tim Becker, Minh Doan, Claire McQuin, et al. Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature methods, 16(12):1247-1253, 2019. 6, 7, 1  
+[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537, 2021. 1, 3, 6, 7  
+[6] Albert Cardona, Stephan Saalfeld, Stephan Preibisch, Benjamin Schmid, Anchi Cheng, Jim Pulokas, Pavel Tomancak, and Volker Hartenstein. An integrated micro-and macroarchitectural analysis of the drosophila brain by computer-assisted serial section electron microscopy. PLoS biology, 8(10):e1000502, 2010. 6, 7, 1  
+[7] Gongping Chen, Lei Li, Yu Dai, Jianxun Zhang, and Moi Hoon Yap. Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images. IEEE Trans. Med. Imaging, 2022. 2  
+[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 2, 3, 6, 7  
+[9] Long Chen, Hanwang Zhang, Jun Xiao, Liqiang Nie, Jian Shao, Wei Liu, and Tat-Seng Chua. Sca-cnn: Spatial and channel-wise attention in convolutional networks for image captioning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5659-5667, 2017. 3, 4  
+[10] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2017. 2, 6  
+[11] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Eur. Conf. Comput. Vis., pages 801-818, 2018. 2  
+[12] Shuhan Chen, Xiuli Tan, Ben Wang, and Xuelong Hu. Reverse attention for salient object detection. In Eur. Conf. Comput. Vis., pages 234-250, 2018. 1  
+[13] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional po
+
+sitional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 1  
+[14] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 1  
+[15] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In IEEE Int. Symp. Biomed. Imaging, pages 168-172. IEEE, 2018. 6, 1  
+[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., pages 248-255. IEEE, 2009. 6  
+[17] Bo Dong, Wenhai Wang, Deng-Ping Fan, Jinpeng Li, Huazhu Fu, and Ling Shao. Polyp-pvt: Polyp segmentation with pyramid vision transformers. arXiv preprint arXiv:2108.06932, 2021. 1, 3, 5, 6  
+[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2  
+[19] Isensee et al. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 1, 2, 3  
+[20] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 263-273. Springer, 2020. 1, 3, 6  
+[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 2  
+[22] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2  
+[23] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7132-7141, 2018. 2, 3  
+[24] Huimin Huang, Lanfen Lin, Ruofeng Tong, Hongjie Hu, Qiaowei Zhang, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, and Jian Wu. Unet 3+: A full-scale connected unet for medical image segmentation. In ICASSP, pages 1055–1059. IEEE, 2020. 1, 2  
+[25] Xiaohong Huang, Zhifang Deng, Dandan Li, and Xueguang Yuan. Missformer: An effective medical image segmentation transformer. arXiv preprint arXiv:2109.07162, 2021. 6, 7
+
+[26] Nabil Ibtehaz and Daisuke Kihara. Acc-unet: A completely convolutional unet model for the 2020s. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 692-702. Springer, 2023. 2  
+[27] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Int. Conf. Mach. Learn., pages 448-456. pmlr, 2015. 3  
+[28] Md Amirul Islam, Sen Jia, and Neil DB Bruce. How much position information do convolutional neural networks encode? arXiv preprint arXiv:2001.08248, 2020. 1  
+[29] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In Int. Conf. Multimedia Model., pages 451-462. Springer, 2020. 6, 1  
+[30] Taehun Kim, Hyemin Lee, and Daijin Kim. Uacanet: Uncertainty augmented context attention for polyp segmentation. In ACM Int. Conf. Multimedia, pages 2167-2175, 2021. 6  
+[31] Alex Krizhevsky and Geoff Hinton. Convolutional deep belief networks on CIFar-10. Unpublished manuscript, 40(7): 1-9, 2010. 4  
+[32] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Adv. Neural Inform. Process. Syst., 25, 2012. 2  
+[33] Xian Lin, Zengqiang Yan, Xianbo Deng, Chuansheng Zheng, and Li Yu. Convformer: Plug-and-play cnn-style transformers for improving medical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 642-651. Springer, 2023. 1  
+[34] 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 Int. Conf. Comput. Vis., pages 10012-10022, 2021. 1, 2, 3  
+[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11976-11986, 2022. 2  
+[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6  
+[37] Ange Lou, Shuyue Guan, and Murray Loew. Dc-unet: rethinking the u-net architecture with dual channel efficient cnn for medical image segmentation. In Med. Imaging 2021: Image Process., pages 758-768. SPIE, 2021. 1, 2  
+[38] Ange Lou, Shuyue Guan, Hanseok Ko, and Murray H Loew. Caranet: context axial reverse attention network for segmentation of small medical objects. In Med. Imaging 2022: Image Process., pages 81-92. SPIE, 2022. 6  
+[39] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Int. Conf. Mach. Learn., pages 807-814, 2010. 3  
+[40] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Adv. Vis. Comput. - Int. Symp., pages 15-28. Springer, 2021. 6, 1
+
+[41] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 1, 2, 3, 6  
+[42] Md Mostafijur Rahman and Radu Marculescu. Medical image segmentation via cascaded attention decoding. In IEEE/CVF Winter Conf. Appl. Comput. Vis., pages 6222-6231, 2023. 1, 3, 4, 5, 6, 7  
+[43] Md Mostafijur Rahman and Radu Marculescu. Multi-scale hierarchical vision transformer with cascaded attention decoding for medical image segmentation. In Med. Imaging Deep Learn., 2023. 1, 3, 5, 7  
+[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 234-241. Springer, 2015. 1, 2, 6  
+[45] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4510-4520, 2018. 2, 4  
+[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2  
+[47] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1-9, 2015. 2  
+[48] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Int. Conf. Mach. Learn., pages 6105-6114. PMLR, 2019. 2  
+[49] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In Eur. Conf. Comput. Vis., pages 459-479. Springer, 2022. 1, 2  
+[50] Jeya Maria Jose Valanarasu and Vishal M Patel. Unext: Mlp-based rapid medical image segmentation network. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 23-33. Springer, 2022. 6, 1  
+[51] David Vázquez, Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Antonio M López, Adriana Romero, Michal Drozdzal, and Aaron Courville. A benchmark for endoluminal scene segmentation of colonoscopy images. J. Healthc. Eng., 2017, 2017. 6, 1  
+[52] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer. In AAAI, pages 2441-2449, 2022. 1, 3  
+[53] Hongyi Wang, Shiao Xie, Lanfen Lin, Yutaro Iwamoto, Xian-Hua Han, Yen-Wei Chen, and Ruofeng Tong. Mixed transformer u-net for medical image segmentation. In ICASSP, pages 2390–2394. IEEE, 2022. 6, 7  
+[54] Jinfeng Wang, Qiming Huang, Feilong Tang, Jia Meng, Jionglong Su, and Sifan Song. Stepwise feature fusion: Local guides global. arXiv preprint arXiv:2203.03635, 2022. 1, 6
+
+[55] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Int. Conf. Comput. Vis., pages 568-578, 2021. 1, 2  
+[56] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Comput. Vis. Media, 8(3):415-424, 2022. 1, 2, 3, 5, 6  
+[57] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Eur. Conf. Comput. Vis., pages 3–19, 2018. 1, 3, 4  
+[58] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Adv. Neural Inform. Process. Syst., 34:12077-12090, 2021. 2, 3  
+[59] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10819-10829, 2022. 1  
+[60] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6848-6856, 2018. 4  
+[61] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In Int. Conf. Med. Image Comput. Comput. Assist. Interv., pages 14-24. Springer, 2021. 1, 3, 6  
+[62] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learn. Med. Image Anal. Multimodal Learn. Clin. Decis. Support, pages 3-11. Springer, 2018. 1, 2, 6

2024/EMCAD_ Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d41e832c1e30177b3d0d37c9c5b80d8bd020394b77ae8b7b06630db65a850b4
+size 666005

2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/ab68d1cd-73a3-4a4c-9896-716aeaaff436_content_list.json

@@ -0,0 +1,2133 @@
+[
+    {
+        "type": "text",
+        "text": "EMOPortraits: Emotion-enhanced Multimodal One-shot Head Avatars",
+        "text_level": 1,
+        "bbox": [
+            125,
+            130,
+            844,
+            150
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Nikita Drobyshev",
+        "bbox": [
+            86,
+            180,
+            230,
+            199
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Antoni Bigata Casademunt  \nStavros Petridis",
+        "bbox": [
+            271,
+            181,
+            486,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Konstantinos Vougioukas  \nMaja Pantic",
+        "bbox": [
+            519,
+            181,
+            728,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zoe Landgraf",
+        "bbox": [
+            769,
+            181,
+            880,
+            199
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{n.drobyshev, a/bigata-casademunt22, k.vougioukas, zoe.landgraf15, stavros.petridis04, m.pantic}@imperial.ac.uk Imperial College London",
+        "bbox": [
+            220,
+            219,
+            751,
+            270
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/5571c03b59a22946cb276d1deded7196a8c084b040541ce529bbee5da0acb26d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            304,
+            271,
+            416
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/6f78ae2acda6416384be5581a2bc2ae87413212cd8bfa3feb9df5869a0f8e39f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            269,
+            305,
+            415,
+            416
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/867efe1bddaf1887f13140417c41e46dc4c66e055a39bb4c7e189e5a87cba5b5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            415,
+            305,
+            560,
+            416
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/c12038d2254398ecb4c7ba204508b41706cc611de0c85c955e62dc16a81d8eb4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            560,
+            305,
+            704,
+            416
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/272336b8074a1488e724b3c0fbb531059f7b99cb096e5d5bc5ee91f1570dcb0f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            704,
+            305,
+            844,
+            416
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/e1eb172f4eb38c6ece771d39f7a6007b8dcd48c8729563a1872fa82a97e69526.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            420,
+            269,
+            527
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/d5150c201a58cf952e956203832059c28533b90018e0a66348c1f3f500048536.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            269,
+            420,
+            413,
+            527
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/316eae8164a7f14a42016f25d69f3acd05748b74e70f6c167f52db9458ece090.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            413,
+            420,
+            560,
+            527
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/479f825123fe7a77d8b249bb5eb77206881d58fc1981346f540388e496db4655.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            560,
+            420,
+            699,
+            527
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/8da0d6fcd24f450ef1006655f2860be5194a1760923b15e36c23cfbcd55b3cd1.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            699,
+            420,
+            844,
+            527
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/1d3be78635c31aa04786c174e323af741e93526bbfde5ad9472c0a7afda2dfaf.jpg",
+        "image_caption": [
+            "Source"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            530,
+            269,
+            642
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/a2ab0f9b76c3a8b179b604978468a109567c35f2493df4b12487308d887b0ea9.jpg",
+        "image_caption": [
+            "Driver"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            269,
+            530,
+            413,
+            642
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/7af6d1668fe3d655129ce10d232c0993eb2693facbf0ba6bfeccb40e09db4e40.jpg",
+        "image_caption": [
+            "Output"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            413,
+            530,
+            560,
+            642
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/37fa260d97e9e5d2b12fdb7dff66ab9aa5c86ba957ebcccf9309e28ed674827c.jpg",
+        "image_caption": [
+            "Driver"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            560,
+            530,
+            699,
+            642
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/6f169bc4bce76438b2a929bd4d15373bc33cedd96bc3fcaca9948d0a78a551f7.jpg",
+        "image_caption": [
+            "Output",
+            "Figure 1. Selected animation results for image-driven mode of our method."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            699,
+            530,
+            844,
+            642
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            691,
+            313,
+            707
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Head avatars animated by visual signals have gained popularity, particularly in cross-driving synthesis where the driver differs from the animated character, a challenging but highly practical approach. The recently presented MegaPortraits model has demonstrated state-of-the-art results in this domain. We conduct a deep examination and evaluation of this model, with a particular focus on its latent space for facial expression descriptors, and uncover several limitations with its ability to express intense face motions. To address these limitations, we propose substantial changes in both training pipeline and model architecture, to introduce our EMOPortraits model, where we:",
+        "bbox": [
+            75,
+            719,
+            472,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Enhance the model's capability to faithfully support intense, asymmetric face expressions, setting a new state-of-the-art result in the emotion transfer task, surpassing previous methods in both metrics and quality.",
+        "bbox": [
+            496,
+            712,
+            893,
+            772
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Incorporate speech-driven mode to our model, achieving top-tier performance in audio-driven facial animation, making it possible to drive source identity through diverse modalities, including visual signal, audio, or a blend of both.",
+        "bbox": [
+            496,
+            776,
+            893,
+            837
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, we propose a novel multi-view video dataset featuring a wide range of intense and asymmetric facial expressions, filling the gap with absence of such data in existing datasets.",
+        "bbox": [
+            496,
+            840,
+            893,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            807,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "8498",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            78,
+            90,
+            207,
+            104
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "1.1. Representation of facial expression",
+        "text_level": 1,
+        "bbox": [
+            78,
+            114,
+            379,
+            131
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Advancements in neural head technologies now enable the creation of realistic avatars from a few or even one image, with the latter being crucial when only one source image is available. Cross-driving synthesis, where avatars are animated with different identities is a key technique for virtual reality, filmmaking, photo animation, etc. However, accurate transfer of facial expressions, especially those that are intense and uneven, remains a substantial challenge in avatar animation, particularly for cross-driving synthesis. While previous research focused mainly on preserving identity and transferring moderate facial motions, our work also seeks to accurately drive high-intensity and asymmetric expressions.",
+        "bbox": [
+            76,
+            138,
+            468,
+            318
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our research is built on and extends the MegaPortraits [9] model, notable for state-of-the-art results in cross-driving synthesis. An additional advantage using this method is that, unlike many avatar systems that depend on limited [3, 25] predefined motion descriptors, MegaPortraits, learns expression representations from scratch, allowing for greater adaptability to a wider range of expressions. We investigate the latent expression spaces and training methods of MegaPortraits to enhance its ability to depict a broad spectrum of facial expressions. Our comprehensive analysis reveals that, while the original model shows limited effectiveness in intense motion representing, it has significant potential for improvement through targeted architectural modifications, adjustments in the training approach, and the integration of our novel dataset.",
+        "bbox": [
+            76,
+            320,
+            470,
+            545
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "1.2. Integrating speech driving",
+        "text_level": 1,
+        "bbox": [
+            76,
+            556,
+            315,
+            571
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We integrate speech in our model to be used either as a complementary or an alternative driver, a crucial aspect for applications like virtual assistance or mixed reality when a primary visual signal is absent.",
+        "bbox": [
+            76,
+            580,
+            468,
+            638
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "By improving disentanglement of facial expression latent space, we remove disturbing information from face motion descriptors and emphasize components that solely control lip movements. Based on this, we formulated a novel loss that helps to achieve desirable results. Furthermore, our method can generate plausible head rotations and blinks, which appear natural and enhance its applicability across various tasks. Thus, our final method can drive source identity through image, audio, or a mix of them.",
+        "bbox": [
+            76,
+            641,
+            468,
+            776
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "1.3. FEED: Facial Extreme Emotions Dataset",
+        "text_level": 1,
+        "bbox": [
+            76,
+            786,
+            426,
+            801
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Public data scarcity is a common obstacle in deep learning, the topic of human facial expressions and movements is not an exception. In particular, we note a lack of high quality video data capturing a wide range of facial expressions, with existing datasets not going beyond basic facial actions shown at Fig. 2. To address this, we collect a new dataset that",
+        "bbox": [
+            76,
+            810,
+            468,
+            898
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "includes basic expressions and also captures complex movements like blinks, winks, and head and tongue movements, along with varied extreme expressions which are difficult or impossible to categorize through basic facial actions. Given this variety in expressions in our dataset, we believe that it will be a valuable resource for research in human emotions and facial reconstruction fields. In summary, our main contributions are as follows:",
+        "bbox": [
+            496,
+            90,
+            893,
+            210
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We introduce our new EMOPortraits model for one-shot head avatars synthesis, that is capable of transferring intense facial expression, showing state-of-the-art results in cross-driving synthesis. We achieve this through careful latent facial expression space development as well as novel losses and a minimal amount of domain-specific data ( $\\sim$  0.1% of the train set).",
+            "- We integrate a speech-driving mode in our model, that demonstrates cutting-edge results in speech-driven animations. It functions effectively alongside visual signals or independently, also generating realistic head rotations and eye blinks.",
+            "- We present a unique multi-view dataset that spans a broad spectrum of extreme facial expressions, filling the gap of absence of such data in existing datasets."
+        ],
+        "bbox": [
+            500,
+            223,
+            890,
+            449
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            476,
+            640,
+            491
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Visual-driven head avatars",
+        "text_level": 1,
+        "bbox": [
+            500,
+            506,
+            741,
+            520
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In recent years, the domain of neural avatars has branched into two prominent sub-fields: person-specific and person-agnostic avatars. While person-specific methodologies [10, 11, 30, 38, 39] excel in delivering stunning realism and motion fidelity for a particular individual, they encounter challenges representing an arbitrary person. Moreover, these approach demands multiple frames of training data, distinct training for each avatar, and can struggle to replicate motions not encountered during training.",
+        "bbox": [
+            496,
+            532,
+            893,
+            669
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Person-agnostic methods, on the other hand, don't need train or fine-tuning for each new person and present an alternative approach to talking-head synthesis. Earlier works in this domain generated avatars in a few-shot technique [3, 33], while subsequent studies introduced one-shot capabilities [2, 8, 9, 22, 29, 31, 32, 34]. Many of these works employ predefined motion representations, such as 3DMM's blendshapes [8, 15, 31, 32]. In contrast, some learned latent emotion representation from scratch [3, 9, 22, 34, 35]. This latter approach holds potential for better motion representation as in such settings, expression descriptors become entirely trainable and free from inheriting the limitations typical for blendshapes as was shown in [3, 25]. In our work, we adopt this strategy, remarkably improving upon the methodology used in MegaPortraits[9].",
+        "bbox": [
+            496,
+            674,
+            893,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "8499",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/66a3d422459b2df9f0a60e5e7a3cd577b47d5f2fcd46c4308a5bbed92a3374f7.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">BFEs</td><td colspan=\"6\">FFED dataset internal structure</td></tr><tr><td>Head rotations</td><td>Winks</td><td>Eyes move</td><td>Asym. em.</td><td>Tongue em.</td><td>Extreme em.</td></tr><tr><td>Mode</td><td>Mild/Strong</td><td>Axes/random</td><td>Mild/Strong</td><td>Random</td><td>Strong</td><td>Strong</td><td>Extreme</td></tr><tr><td># of part.</td><td>21/23</td><td>21</td><td>23</td><td>21</td><td>23</td><td>23</td><td>23</td></tr><tr><td>Avg. length</td><td>1:38/1.52</td><td>0:58</td><td>0:49</td><td>0:33</td><td>1:41</td><td>1:53</td><td>3:02</td></tr></table>",
+        "bbox": [
+            89,
+            88,
+            457,
+            127
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/578c2462fd48de5b379bc63eeefef14e809e14ca69ec9aef8856235a65654fea.jpg",
+        "table_caption": [
+            "Table 1. The tasks we asked our participants to perform for our FEED dataset. Here BFEs stands for basic facial expressions, shown in Fig. 2"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"7\">Datasets comparison</td></tr><tr><td>#Actors</td><td>#Views</td><td>Basic exp.</td><td>Strong exp.</td><td>Extrme &amp; Assym. exp.</td><td>Tongue exp.</td><td>Resolution</td></tr><tr><td>SAVESS</td><td>4</td><td>1</td><td>✓</td><td>✘</td><td>✘</td><td>✘</td><td>1280 x 1024</td></tr><tr><td>RAVGESS</td><td>24</td><td>1</td><td>✓</td><td>✘</td><td>✘</td><td>✘</td><td>1920 x 1080</td></tr><tr><td>CREMA-D</td><td>91</td><td>1</td><td>✓</td><td>✘</td><td>✘</td><td>✘</td><td>1280 x 720</td></tr><tr><td>MMI</td><td>25</td><td>1</td><td>✓</td><td>✓</td><td>✘</td><td>✘</td><td>1920 x 1080</td></tr><tr><td>MEAD</td><td>48</td><td>7</td><td>✓</td><td>✓</td><td>✘</td><td>✘</td><td>1920 x 1080</td></tr><tr><td>Ours (FEED)</td><td>23</td><td>3</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>3840 x 2160</td></tr></table>",
+        "bbox": [
+            89,
+            195,
+            457,
+            257
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Comparisons with modern, high-quality audio-visual datasets created in controlled settings.",
+        "bbox": [
+            76,
+            267,
+            467,
+            295
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2. Face expression datasets",
+        "text_level": 1,
+        "bbox": [
+            76,
+            324,
+            302,
+            339
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Early image datasets[1, 19] predominantly offered annotated face expression data corresponded to up to 8 basic emotions. However, these datasets are not well-suited for training head avatars, which necessitate video or multi-view data.",
+        "bbox": [
+            76,
+            348,
+            468,
+            407
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The SAVEE [14] dataset captures facial expressions in speech, but involves just 4 actors. The MMI [20] dataset, more expansive in actor participation, offers spontaneous moderate-intensity expression but is restricted to single-view sequences. RAVDESS [18] distinguishes itself by capturing two motion intensities, but its limited recordings challenge broad applicability. CREMA-D [4] encompasses three expression intensities, and though MEAD [28] provides detailed multi-view data across three intensities, it's limited to the standard eight expression groups. Our novel FEED dataset aims to address the scientific community's demand for high-quality multi-view facial expression videos outside the standard categories: Joy, Fear, Sadness, Disgust, Anger Contempt and Surprise.",
+        "bbox": [
+            75,
+            409,
+            470,
+            621
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3. Speech-driven head avatars",
+        "text_level": 1,
+        "bbox": [
+            76,
+            633,
+            326,
+            648
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "While numerous studies introduce speech-driven avatars [12, 21, 24, 31, 36, 39, 40], few excel in producing high-quality talking heads with authentic rotations, blinks, and the capability to use both visual and audio inputs. For instance, Wav2Lip [21] aims at re-dubbing videos with accurate lip motions, but often falls short in realism with using single images. Diffused heads [24] struggles to generate long sequences and doesn't provide access to pose control. MakeItTalk [40] animates facial landmarks in a speaker-specific manner, yet struggles with head pose controls due to its non-reliance on 3D. PC-AVS [39] offers a solution, but demands a driving video for pose modulation. Newly developed models such as SadTalker [36] and StyleHEAT [31] have shown promising outcomes in generating high-quality talking heads. We conducted a quantitative evaluation of our audio-driven mode compared to the aforementioned models.",
+        "bbox": [
+            75,
+            659,
+            472,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/e582360968f3b94ebaa9e6716b25e04a6f5eb78834c851fa58aea8caced7cf0b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            529,
+            87,
+            867,
+            193
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/33d7247f6f7c8a4a0d29ea41ed87dcebf4c56350ecb13484916dc2b21c5c7926.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            527,
+            196,
+            867,
+            292
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/a0ffc5d39905b35e0710031436f2b4a76dde5773b3d2205c14a51ddcbc729cf5.jpg",
+        "image_caption": [
+            "Figure 2. Illustration of the problem in publicly available face expression data and selected examples from our FEED dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            294,
+            867,
+            369
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. FEED dataset",
+        "text_level": 1,
+        "bbox": [
+            500,
+            441,
+            640,
+            455
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As previously mentioned, the current publicly available datasets with human face expression videos fall short in capturing a broad range of facial manipulations (see illustration at Fig. 2). In our view, this gap could stem from the complexity of designing clear tasks for participants, coupled with the reality that not everyone is ready for performing extreme expressions on camera. To bridge this gap, our pioneering FEED dataset is designed to meet the scientific community's need for high-quality, multi-view emotion videos that extend beyond standard emotions. It contains various expressions, including strong asymmetric ones, tongue and cheeks movements, winks, head rotations, eye movements and more nuanced gestures. ;",
+        "bbox": [
+            496,
+            470,
+            893,
+            667
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our dataset consists of 520 multi-view videos of 23 subjects, captured with 3 cameras. As extreme face motions are complex and their perception can be heavily influenced by subtle differences, we use a high resolution of 4k for all video, capturing the whole face up to the level of individual hair strands and wrinkles, as shown in Fig. 2. For more examples, please refer to the supplementary materials. Our participants were asked to perform 7 tasks (see Tab. 1) to cover as many facial and head movements as possible.",
+        "bbox": [
+            496,
+            670,
+            893,
+            806
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In our comparison with other expression datasets detailed in Tab. 2, the MEAD dataset [28] emerges as the closest to ours regarding viewpoints and maximal expression intensity. However, despite MEAD's larger number of actors and viewpoints, it offers significantly less variety of facial expressions and have lower resolution than our dataset.",
+        "bbox": [
+            496,
+            810,
+            893,
+            898
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "8500",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4. Expression enhancement",
+        "text_level": 1,
+        "bbox": [
+            76,
+            89,
+            308,
+            107
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We found that the MegaPortraits [9] model fails to transfer intense expressions correctly, as shown in Fig. 9. Using our FEED dataset of strong and uneven facial expression to fine-tune a pre-trained MegaPortraits model didn't enhance final results as shown at Tab. 4. Training from scratch on FEED leads to fast overfitting due to the small number of identities presented in the dataset compared to VoxCeleb2 [7] used in [9] (23 vs  $\\sim 5000$ ). Our method effectively injects this small dataset into training, yielding desirable results. We list our key findings here, for all details, see supplementary materials. Our model's scheme, displayed in Fig. 4, shares similarities with [9], leading us to occasionally refer to certain elements of our scheme in the context of [9]'s pipeline.",
+        "bbox": [
+            75,
+            114,
+            473,
+            311
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1. Latent expression space",
+        "text_level": 1,
+        "bbox": [
+            76,
+            321,
+            295,
+            338
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We begin by exploring MegaPortraits' latent expression descriptors' space  $(\\mathbf{z}_{s / d}$  in Fig. 4), which is crucial for expression transfer, through PCA analysis. Inspired by [16], which demonstrated that the area under the cumulative explained variance curve of singular values (denoted here as  $\\mathbf{AUC}_{\\mathbf{z}}$ ) can serve as an effective metric for dimensionality collapse in some latent spaces, we applied this measure to our study.",
+        "bbox": [
+            75,
+            344,
+            470,
+            450
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "This metric allows us to forecast model performance: greater collapse suggests reduced entropy in the latent space, which correlates with lower representational quality. High quality of the expression space, produced by  $\\mathbf{E}_{\\text {motion }}$  (as shown in Fig. 4), is crucial to capture and differentiate between the nuances of strong facial expressions. Our ablation study confirms (see Tab. 4) that there is a notable correlation between how broad and isotropic model's latent space is (expressed by  $\\mathbf{AUC}_{\\mathbf{z}}$ ) and its final performance. Visual comparison is available in supplementary materials. As shown in Fig. 3, our model's latent space outperforms MegaPortraits in expression representation ability. This is supported by both visual (Fig. 9) and quantitative analysis (Tab. 3).",
+        "bbox": [
+            75,
+            450,
+            473,
+            648
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/43e0be9fd62862a686fc0a3749085dde5d7228aefbfb924c5e39ad2a1a627ee5.jpg",
+        "image_caption": [
+            "Figure 3. Comparison of latent spaces. Left plot shows that our model's latent space is wider and exhibits more even variance distribution. Also, as shown on the right plot, a greater number of principal components are involved in capturing variance across various thresholds. This implies a more robust representational capacity of expression space compared to [9]. The VoxCeleb2 test set was used for both plots."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            660,
+            450,
+            789
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/cfaa5cbc7efaf0aaec7497921709bfe2ffc0e67e9a55e614e647ce8545e2ac06.jpg",
+        "image_caption": [
+            "Figure 4. Method Overview. We use  $\\mathbf{E}_{\\mathrm{app}}$  to extract volume features  $\\mathbf{V}_s$  and a global descriptor  $\\mathbf{e}_s$  from the source image. Then  $\\mathbf{E}_{\\mathrm{motion}}$  or  $\\mathbf{E}_{\\mathrm{audio}}$  generates motion representations from source and driver, including head rotations  $\\mathbf{R}_{s / d}$ , translations  $\\mathbf{t}_{s / d}$ , and expression descriptors  $\\mathbf{z}_{s / d}$ . Using them, we predict warpings  $\\mathbf{w}_{s\\rightarrow}$  and  $\\mathbf{w}_{\\rightarrow d}$ . First warping and  $\\mathbf{G}_{3D}$  transform  $\\mathbf{V}_s$  into a canonical volume  $\\mathbf{V}_s^C$  by removing the source motions. Second warping and  $\\mathbf{G}_{2D}$  imposes the driver's motions and renders the final image."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            539,
+            87,
+            849,
+            200
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Both metrics shown in Fig. 3 were vital in our model's development since they highly correlate with the final model's ability to transfer intense expressions. This was particularly useful for finding adjustments described in Sec. 4.2 and Sec. 4.3, as early visual results during training are not very telling, but changes in the latent space are noticeable after a few epochs. We also found that in MegaPortraits, just a few principal components significantly affect variance, with only 18 components making up  $99\\%$  (Fig. 3), suggesting their  $512\\mathbf{z}$ 's dimension is oversized. For our model, we found a 128-dimensional  $\\mathbf{z}$  is optimal, offering solid performance and reducing overfitting in imbalanced data, as our ablation study reveals. We calculate the Explained Variance  $(\\mathrm{EV}_i)$  for the  $i^{th}$  component and  $(\\mathbf{AUC}_{\\mathbf{z}})$  using equations shown in Eq. (1), which involves standardizing the vector set  $\\mathcal{Z}$  for PCA and sorting the eigenvalues  $\\lambda$  in descending order.",
+        "bbox": [
+            496,
+            335,
+            893,
+            579
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {Z} _ {\\mathrm {s t d}} = \\frac {\\mathcal {Z} - \\bar {\\mathcal {Z}}}{\\sigma (\\bar {\\mathcal {Z}})}, \\quad \\boldsymbol {\\lambda} = \\operatorname {s o r t e d} _ {\\mathrm {d e s c}} \\left(\\operatorname {e i g} \\left(\\frac {1}{n - 1} \\mathbf {Z} _ {\\mathrm {s t d}} ^ {\\top} \\mathbf {Z} _ {\\mathrm {s t d}}\\right)\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            506,
+            593,
+            893,
+            633
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathrm {E V} _ {i} = \\frac {\\lambda_ {i}}{\\sum_ {j = 1} ^ {p} \\lambda_ {j}}, \\quad \\mathbf {A U C _ {z}} = \\frac {1}{d} \\frac {\\sum_ {i = 1} ^ {d} \\sum_ {k = 1} ^ {i} \\lambda_ {k}}{\\sum_ {j = 1} ^ {d} \\lambda_ {j}},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            506,
+            627,
+            820,
+            662
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2. Canonical volume",
+        "text_level": 1,
+        "bbox": [
+            500,
+            666,
+            673,
+            679
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Despite being designed to exclude face expression details, we found that canonical volume  $(\\mathbf{V}^C$  in Fig. 4) in MegaPortraits actually retains significant expression information from the source image, contributing to poor translation of intense expressions, as shown in our experiment detailed in Fig. 5. In this experiment, using portraits with varied expressions, we assessed the expression translation accuracy and how expression intensity affects the  $\\mathbf{V}^C$ s. Our findings suggest that the canonical volumes in MegaPortraits are not truly neutral. As confirmed by our ablation study (Tab. 4), creating an expression-free  $\\mathbf{V}^C$  is essential for precise translation of intense motions. We believe that a more neutral canonical volume improves tractability and effectiveness in expression translation tasks.",
+        "bbox": [
+            496,
+            688,
+            893,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "8501",
+        "bbox": [
+            480,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/65fd3cb501e87e562244924ab7734643dc949462bd6d5e71948526318373e41c.jpg",
+        "image_caption": [
+            "Figure 5. Canonical volumes  $(\\mathbf{V}^C)$  in MegaPortraits [9] are not expression-neutral. To show it we use three portraits: A (regular expression), B (intense expression), D (regular expression, new identity). Case 1 visualizes poor results in transferring B's intense expression to A using a self-driving mode, contrasting with the effective reconstruction of B when used as both driver and source in case 3. This discrepancy, indicative of expression leakage into the canonical volume, is further quantified by a  $43\\%$  relative difference in  $\\mathbf{V}^C$ s between B and C, contrary to expectations of their similarity for same identity. In cross-driving generation case 2, using B as the driver and D as the source yielded poor results. However, in case 4, after using additive operations on source canonical volumes, the output expression is much closer to driver than in case 2. This manipulation again confirms the significant retention of expression information in canonical volumes."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            89,
+            88,
+            465,
+            210
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To overcome this issue, we propose to match canonical volumes from the different images of the same person  $(\\mathbf{V}_{s^n}^C,\\mathbf{V}_{d^n}^C)$  during training as following:",
+        "bbox": [
+            75,
+            448,
+            470,
+            494
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C V}} ^ {\\mathrm {n}} = \\mathcal {L} _ {\\mathrm {M A E}} \\left(\\mathbf {V} _ {s ^ {n}} ^ {C}, \\mathbf {V} _ {d ^ {n}} ^ {C}\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            186,
+            505,
+            468,
+            526
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This loss ensures  $\\mathbf{V}^C$  remains stable and expression-independent, crucial for translating intense expressions, as shown in our ablation study (Tab. 4).",
+        "bbox": [
+            75,
+            529,
+            472,
+            575
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3. Source-driver mismatch loss",
+        "text_level": 1,
+        "bbox": [
+            76,
+            582,
+            331,
+            597
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In addition to maintaining expression-free  $\\mathbf{V}^{C}$ , it's vital to remove all identity information from the expression vector  $\\mathbf{z}$ . While contrastive losses address this in [9], our experiments indicate they are insufficient to prevent overfitting in our imbalanced data scenario, where emotionally intense images represent only  $0.1\\%$  of our training set but are sampled  $25\\%$  of the time. We introduce a novel self-supervised loss which mitigates identity information in latent expression vectors:",
+        "bbox": [
+            75,
+            604,
+            470,
+            727
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {s d m} (\\mathbf {z} _ {s}, \\mathbf {z} _ {d}) = w * \\max (0, \\cos (\\mathbf {z} _ {s}, \\mathbf {z} _ {d}) - m a r g i n), \\\\ w; \\text {m a r g i n} = \\left\\{ \\begin{array}{l l} 1; 0. 5 & \\text {i f} \\mathbf {z} _ {s}, \\mathbf {z} _ {d} \\text {a r e f r o m V C 2} \\\\ 1 0; 0. 2 5 & \\text {i f} \\mathbf {z} _ {s}, \\mathbf {z} _ {d} \\text {a r e f r o m F E E D} \\end{array} \\right. \\tag {3} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            84,
+            744,
+            468,
+            805
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Here,  $\\mathbf{z}_s, \\mathbf{z}_d$  are emotion vectors from source and driver images,  $w, \\text{margin}$  adjust the loss's intensity and strictness respectively. We increase  $w$  for FEED due to its higher sampling rate and decrease margin for more assured expression variance in  $s, d$ . Refer to Fig. 6 for a visual summary of all self-supervised losses enhancing our latent emotion space.",
+        "bbox": [
+            75,
+            810,
+            472,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/4168394fbcbb61f875f732ea64d8e9e00b593fc9b8b7198c39c30b7289b588f7.jpg",
+        "image_caption": [
+            "Figure 6. The visualisation of our self-supervised losses.  $\\mathcal{L}_{\\mathrm{cos}}$  is a contrastive loss, similar to one used in [9] but uses different pairs as described in Sec. 5.1. Our novel  $\\mathcal{L}_{sdm}$  is designed to prevent overfitting in an imbalanced dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            540,
+            87,
+            854,
+            335
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5. Incorporating speech-driving mode",
+        "text_level": 1,
+        "bbox": [
+            500,
+            426,
+            820,
+            445
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1. Latent space disentanglement",
+        "text_level": 1,
+        "bbox": [
+            500,
+            452,
+            764,
+            468
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "During our latent space analysis described in Sec. 4, we found that combining expression from one driver with the head pose from another resulted in poor performance for MegaPortraits model, as illustrated in Fig. 7 (bottom row). Indeed, the base model lacks a mechanism to intentionally prevent head pose leakage. Disentangled expression latent space, besides expanding the model's use cases, plays the crucial role for the speech-driving mode. When the latent vector  $\\mathbf{z}$  is entangled with head pose data, predicting it from speech becomes challenging because speech lacks head rotation information, unlike images. Moreover, we would like to have a full control over head rotations during speech mode. We were able to make our latent space disentangled by changing the way we sample images for  $\\mathcal{L}_{\\mathrm{cos}}$  Eq. (5).",
+        "bbox": [
+            496,
+            476,
+            893,
+            688
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The MegaPortraits model uses  $\\mathcal{L}_{\\mathrm{cos}}$  - a modified Large Margin Cosine Loss [26] to mitigate the transfer of appearance features to expression descriptor (latent vector  $\\mathbf{z}$  on Fig. 4), which is crucial for cross-driving mode. Without this, differences in appearance between the source and driver, like hairstyle or skin tone, leak from driver to output image.",
+        "bbox": [
+            496,
+            688,
+            893,
+            779
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For computing this loss, the authors utilize a supplementary source-driving pair  $(\\mathbf{x}_{s^{*}}$  and  $\\mathbf{x}_{d^{*}})$  from a different video with another identity, ensuring a distinct appearance from the current  $\\mathbf{x}_s,\\mathbf{x}_d$  pair. The base model is then employed to produce cross-reenacted image  $(\\hat{\\mathbf{x}}_{s^{*}\\rightarrow d} = \\mathbf{G}_{\\mathrm{base}}(\\mathbf{x}_{s^{*}},\\mathbf{x}_d))$  Concurrently, they determine a separate motion descriptor,  $\\mathbf{z}_{d^{*}} = \\mathbf{E}_{\\mathrm{motion}}(\\mathbf{x}_{d^{*}})$ . Descriptors  $\\mathbf{z}_{* \\rightarrow d}$  from the respective forward passes are also used for positive and negative pairs.",
+        "bbox": [
+            496,
+            780,
+            895,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "8502",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/2fc771a3720ccaa6e2cde67511df1976c6cbeed5e2acdb994a7fe463b911909b.jpg",
+        "image_caption": [
+            "Figure 7. The illustration of disentanglement problem. Images in the right row indicate the expression descriptor  $\\mathbf{z}$  entangled with head pose for MegaPortraits, whereas we avoid it using changes proposed in Sec. 5.1."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            88,
+            467,
+            205
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Our innovation here is a new sampling strategy: for each  $\\mathbf{x}_s$ ,  $\\mathbf{x}_d$  pair, we sample one more random additional pair  $\\mathbf{x}_{s^m}$  and  $\\mathbf{x}_{d^m}$  apart from  $\\mathbf{x}_{s^*}$  and  $\\mathbf{x}_{d^*}$ . We then apply the model to produce the following cross-reenacted moved image using source identity from  $\\mathbf{x}_{s^*}$ , desired head pose from  $\\mathbf{x}_{s^m}$  and emotions from  $\\mathbf{x}_d$ , represented as:  $\\hat{\\mathbf{x}}_{s^{*m}\\rightarrow d} = \\mathbf{G}_{\\mathrm{base}}(\\mathbf{x}_{s^{*m}},\\mathbf{x}_d)$ . This creates a positive pair with the same desired emotion but varying head poses.",
+        "bbox": [
+            75,
+            297,
+            468,
+            417
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Motion descriptors are organized into positive pairs  $\\mathcal{P}$  for alignment, and negative pairs  $\\mathcal{N}$  for non-alignment:  $\\mathcal{P} = \\left\\{(\\mathbf{z}_{s\\rightarrow d},\\mathbf{z}_d),(\\mathbf{z}_{s^*\\rightarrow d},\\mathbf{z}_d),(\\mathbf{z}_{s^{*m}\\rightarrow d},\\mathbf{z}_d)\\right\\}$ , and  $\\mathcal{N} = \\left\\{(\\mathbf{z}_{s\\rightarrow d},\\mathbf{z}_{d^*}),(\\mathbf{z}_{s^*\\rightarrow d},\\mathbf{z}_{d^*}),(\\mathbf{z}_{s^{*m}\\rightarrow d},\\mathbf{z}_{d^*})\\right\\}$  (see Fig. 6). These pairs are used to calculate the following cosine distance:",
+        "bbox": [
+            76,
+            417,
+            470,
+            507
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nd \\left(\\mathbf {z} _ {i}, \\mathbf {z} _ {j}\\right) = s \\cdot \\left(\\left\\langle \\mathbf {z} _ {i}, \\mathbf {z} _ {j} \\right\\rangle - m\\right), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            169,
+            508,
+            468,
+            526
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where both  $s$  and  $m$  are hyperparameters. This distance is then used to calculate a large margin cosine loss (Cos-Face) [27]:",
+        "bbox": [
+            75,
+            530,
+            468,
+            575
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\cos} = - \\sum_ {(\\mathbf {z} _ {k}, \\mathbf {z} _ {l}) \\in \\mathcal {P}} \\log \\frac {\\exp \\left\\{d \\left(\\mathbf {z} _ {k} , \\mathbf {z} _ {l}\\right) \\right\\}}{\\exp \\left\\{d \\left(\\mathbf {z} _ {k} , \\mathbf {z} _ {l}\\right) \\right\\} + \\sum_ {(\\mathbf {z} _ {i} , \\mathbf {z} _ {j}) \\in \\mathcal {N}} \\exp \\left\\{d \\left(\\mathbf {z} _ {i} , \\mathbf {z} _ {j}\\right) \\right\\}} \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            86,
+            595,
+            468,
+            637
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "This loss prevents the head pose from leaking into the embeddings as shown in Fig. 7 (upper row)",
+        "bbox": [
+            75,
+            643,
+            468,
+            672
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "By doing this, we achieved the capability to interpret  $\\mathbf{z}_i$  as a latent vector encapsulating the emotional content from an image  $\\mathbf{x}_i$ . This advancement opened the door for an intriguing possibility during inference: the prediction of this vector could potentially be derived from an audio signal instead of a facial image using the motion encoder  $\\mathbf{E}_{\\mathrm{motion}}$ , effectively transforming the model into a speech-driven system. The critical challenge here is aligning the audio input with the pretrained latent space. We addressed this by introducing an additional audio encoder,  $\\mathbf{E}_{\\mathrm{aud}}$ , and employing  $\\mathbf{E}_{\\mathrm{motion}}$  as a teacher model for this new encoder. A crucial factor in achieving remarkable accuracy in lip synchronization was our ability to isolate specific components within the expression latent space, that are solely responsible for mouth movements. We describe it in the next subsection Sec. 5.2.",
+        "bbox": [
+            75,
+            674,
+            470,
+            898
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/426d2459ddac66dc468bd55fd6056ae9f8154ca624c50c1e768277a3b7502cea.jpg",
+        "image_caption": [
+            "Figure 8. In the upper row, the first principal component from a random subset of expression vectors  $\\mathcal{Z}$  affects not just the mouth but also blinks and gaze, making it difficult to interpret and isolate mouth movement components, needed for training  $\\mathbf{E}_{\\mathrm{aud}}$ . However, as shown in the bottom row, using  $\\mathcal{Z}_m$  from Sec. 5.2, the first principal component,  $\\mathbf{v}_1$ , focused on mouth movements, is effectively isolated,  $\\mathbf{z}_0$  - latent vector, corresponded to the central image."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            498,
+            88,
+            890,
+            227
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2. Mouth movements",
+        "text_level": 1,
+        "bbox": [
+            500,
+            356,
+            679,
+            371
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To train the audio encoder  $\\mathbf{E}_{\\mathrm{aud}}$  for predicting  $\\mathbf{z}$  from speech, we chose to use the motion encoder  $\\mathbf{E}_{\\mathrm{motion}}$  as a teacher model. However, direct matching vectors  $\\mathbf{z}_i^{\\mathrm{aud}}$  from  $\\mathbf{E}_{\\mathrm{aud}}$  and pseudo-ground truth  $\\mathbf{z}_i$  from  $\\mathbf{E}_{\\mathrm{motion}}$  gives poor results (see Tab. 6). We believe that the reason behind this is that speech-derived  $\\mathbf{z}_i^{\\mathrm{aud}}$ , while able to capture lip movements, struggles with other facial motions, especially in the upper face. Thus, we focused on isolating mouth movement components in the expression latent space using PCA analysis.",
+        "bbox": [
+            496,
+            378,
+            893,
+            516
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Employing PCA on a broad array of expression vectors  $\\mathcal{Z}$ , generated from images with varied facial emotions, yields principal components with their explained variances  $(\\mathrm{EV}_i)$ . Altering these components, particularly those with high  $\\mathrm{EV}_i$ , significantly changes facial expressions in an image. However, they typically affect a combination of facial features—Fig. 8 (upper row), not isolated areas like the mouth or eyes.",
+        "bbox": [
+            496,
+            516,
+            893,
+            621
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To focus on mouth movements, we created a unique set of  $\\mathcal{Z}_m$  from just one video of a person performing mouth manipulations and manually edited it to minimize upper face movements. This involved using a still upper face from the first frame in all subsequent frames, ensuring that the principal components mainly represented mouth movements. Applying PCA of  $\\mathcal{Z}_m$  reveals that most principal components are responsible for solely mouth movements. For illustration of the first component, see Fig. 8 (bottom row). Based on distilled components, we introduce mouth PCA mouth loss:",
+        "bbox": [
+            496,
+            622,
+            893,
+            772
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P C A}} \\left(v _ {i}, v _ {j}, n\\right) = \\frac {1}{n} \\sum_ {k = 1} ^ {n} \\left| \\mathrm {P C} _ {\\mathbf {v} _ {i}} (k) - \\mathrm {P C} _ {\\mathbf {v} _ {j}} (k) \\right| \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            539,
+            779,
+            890,
+            819
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Where  $n$  is the number of principal components to be considered,  $\\mathrm{PC}_{\\mathbf{z}_i}(k)$  and  $\\mathrm{PC}_{\\mathbf{z}_j}(k)$  are the  $k$ -th principal components of vectors  $\\mathbf{z}_i$  and  $\\mathbf{z}_j$  respectively. Implementing this loss has demonstrably enhanced the quality and accuracy of our results, as our ablation study confirms (see Tab. 6).",
+        "bbox": [
+            496,
+            824,
+            893,
+            898
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "8503",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            90,
+            209,
+            107
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this section, we describe experiments we set for comparison for both our image-driven and speech-driven mode. Please see implementation details such as all used losses, component's sizes and architecture details, training procedure, data preprocess in supplementary materials.",
+        "bbox": [
+            75,
+            114,
+            470,
+            191
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.1. Image-driven comparison",
+        "text_level": 1,
+        "bbox": [
+            76,
+            196,
+            310,
+            214
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Methods. We evaluated our image-driven model alongside various one-shot models. For this comparison, we included models like NOFA [32] and StyleHEAT [31], which use 3DMM expression blendshapes without disentanglement training, and models with trainable latent face motion representations like FOMM [22], UVA [23], MetaPortrait [35] and MegaPortraits [9]. It's noteworthy that NOFA and UVA require per-source optimization, unlike the others, including our model.",
+        "bbox": [
+            75,
+            220,
+            468,
+            354
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Evaluation metrics and data. For our evaluation, we use 100 random, non-child, single-person images from the FFHQ dataset as sources. From the MEAD dataset, we chose 100 images covering all high-intensity emotions except neutral, and another 100 from the FEED dataset's extreme emotion task, featuring 5 identities not seen during training of our model. Each source image was paired with 20 driving images from these sets, totaling 2000 pairs.",
+        "bbox": [
+            75,
+            356,
+            468,
+            477
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "For assessing reenactment quality, we employ various metrics including the Frechet Inception Distance (FID) [13] to measure the distributional discrepancy between synthetic and real images. Cosine similarity (CSIM) from a face recognition network [5] quantifies the identity preservation in generated images. Additionally, we conducted user studies to determine preferences regarding motion (UMTN) and appearance (UAPP) preservation, presenting participants with triplets—either a driving or source image and two method outputs—and asking them to choose the one with better motion or appearance preservation. Our user study includes 34 participants and around 300 unique questions. Both metrics numerically reflect preference percentages. As our visual results Fig. 9, and quantitative metrics Tab. 3 show, our method demonstrates exceptional ability in translating strong and extreme expressions, notably surpassing other image-driven avatar techniques in user preference for facial expression translation (UMTN) and FID scores.",
+        "bbox": [
+            75,
+            478,
+            470,
+            750
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.2. Speech-driven comparison",
+        "text_level": 1,
+        "bbox": [
+            76,
+            756,
+            316,
+            772
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Methods. To assess our model's speech-driven capabilities, we compared it with StyleHEAT [31], which also operates using both image and audio inputs, and audio-exclusive models including SadTalker [36], MakeItTalk [40], and PC-AVS [39].",
+        "bbox": [
+            75,
+            779,
+            468,
+            854
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Evaluation metrics and data. We employ the evaluation procedure described in [12]. We measure video realism using Frechet Inception Distance (FID) [13]. We also extract",
+        "bbox": [
+            76,
+            854,
+            470,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/f723b0a62eb5b4733b04b9bdc559c91c2d518e61d1a5e9187c6db9f9ecf35458.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 3. Our method notably outperforms others in the FID score and strongly leads in the user preference metric for face expression translation (UMTN). It excels in reliably translating strong and extreme expressions from the driving image, distinguishing it as the most capable among the compared methods. Additionally, our approach maintains the source image's shape and appearance (measured by CSIM and UAPP metrics) on par with MegaPortraits [9] when using strong and regular expressions from the MEAD dataset as the driving signal. However, we observe a minor deviation in identity preservation metrics for extreme emotions (using the FEED dataset). This difference may arise because our method attempts to transfer asymmetric facial expressions, while MegaPortraits might modify the source image less, or not at all, with a challenging driver (as depicted in Fig. 5, case 1), thereby achieving better identity preservation. UVA was excluded from the MEAD-driven user study due to the authors were able to provide only part of the test data. Also, MetaPortrait was excluded from both user studies because of late data provision by authors, offering only the basic model results, without their super-resolution module outputs."
+        ],
+        "table_body": "<table><tr><td>Driver:</td><td colspan=\"4\">MEAD</td><td colspan=\"4\">FEED</td></tr><tr><td>Method</td><td>FID↓</td><td>CSIM↑</td><td>UMTN↑</td><td>UAPP↑</td><td>FID↓</td><td>CSIM↑</td><td>UMTN↑</td><td>UAPP↑</td></tr><tr><td>MetaPortrait [35]</td><td>77.5</td><td>0.64</td><td>*</td><td>*</td><td>78.8</td><td>0.66</td><td>*</td><td>*</td></tr><tr><td>NOFA [32]</td><td>69.2</td><td>0.65</td><td>*</td><td>*</td><td>69.8</td><td>0.66</td><td>3.8</td><td>8.8</td></tr><tr><td>FOMM [22]</td><td>84.4</td><td>0.56</td><td>2.4</td><td>0.3</td><td>85.9</td><td>0.55</td><td>1.8</td><td>0.5</td></tr><tr><td>UVA [17]</td><td>78.7</td><td>0.68</td><td>0.9</td><td>6.4</td><td>79.5</td><td>0.68</td><td>0.6</td><td>9.4</td></tr><tr><td>StyleHEAT [31]</td><td>71.9</td><td>0.63</td><td>3.7</td><td>3.3</td><td>72.4</td><td>0.61</td><td>2.6</td><td>3.8</td></tr><tr><td>MegaPortraits [9]</td><td>61.1</td><td>0.73</td><td>22.7</td><td>46.6</td><td>62.8</td><td>0.73</td><td>16.7</td><td>42.8</td></tr><tr><td>Our (EMOPortraits)</td><td>59.6</td><td>0.74</td><td>70.3</td><td>43.4</td><td>60.2</td><td>0.70</td><td>74.6</td><td>34.7</td></tr></table>",
+        "bbox": [
+            504,
+            89,
+            890,
+            176
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/873156c9e6afd47be52e4d5a99fb0aa75c8f06eb0509809b0b49c314487dafaa.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 4. Our ablation study examines key aspects of the image-driven mode. After fine-tuning MegaPortraits on the FEED dataset, we observed a decline in identity preservation due to overfitting on FEED identities, without any improvement in emotion translation. The lack of  $\\mathcal{L}_{CV}^{n}$  primarily impacts the ability to translate extreme expression, as indicated by UMTN. Additionally, not using  $\\mathcal{L}_{sdm}$  or setting  $dim(\\mathbf{z}) = 512$  results in identity leakage, particularly evident in the latter scenario. For this ablation, we use the part of our driven by MEAD dataset and described in Sec. 6.1"
+        ],
+        "table_body": "<table><tr><td>Method</td><td>FID↓</td><td>CSIM↑</td><td>UMTN↓</td><td>UAPP↓</td><td>AUCz↓</td></tr><tr><td>MegaPortraits fine-tuned</td><td>62.3</td><td>0.63</td><td>8.5</td><td>12.4</td><td>0.89</td></tr><tr><td>Ours w/o LcnCV</td><td>61.4</td><td>0.67</td><td>9.2</td><td>17.9</td><td>0.88</td></tr><tr><td>Ours w/o Lsdm</td><td>61.8</td><td>0.59</td><td>11.3</td><td>6.7</td><td>0.85</td></tr><tr><td>Ours dim(z) = 512</td><td>63.4</td><td>0.44</td><td>25.6</td><td>1.1</td><td>0.77</td></tr><tr><td>Ours</td><td>59.6</td><td>0.74</td><td>45.4</td><td>61.9</td><td>0.80</td></tr></table>",
+        "bbox": [
+            509,
+            465,
+            883,
+            534
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "facial landmarks from outputs of all methods, frontalized, and normalized to position the mouth edges at  $(-1,0)$  and  $(1,0)$ . We compute mean absolute error (MAE) for both predicted mouth landmark positions  $(M_P)$  and velocities  $(M_V)$ , along with errors in facial geometry  $(F_P)$  and velocity  $(F_V)$ . For our evaluation, we use 100 randomly selected video sequences from the HDTF dataset [37], each up to 30 seconds. As the results show in Tab. 5, our speech-driven model performs on par with leading methods, excelling in realism and facial dynamics. SyncNet [6] is used to assess lip sync quality, providing offset (temporal misalignment between the audio and video) and confidence (certainty of audio-visual alignment) scores.",
+        "bbox": [
+            496,
+            703,
+            893,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "8504",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/78926baf7339f356df241bf8a88b63a9b83f4b6c3649de1e1ee938d86b29bfe5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            112,
+            87,
+            194,
+            151
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/249eb01f4131fdddf201108d1aabb80575d66581bff4a918301a2418cf1ced01.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            194,
+            88,
+            276,
+            151
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e6833fe36b387c24e85728fd42f64f6c1ebddbb0976d1cc0ff663463ff80b708.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            277,
+            88,
+            359,
+            151
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/babf75fa3d53bebbcf1c449f3409b684032610a97ed8f57c97e1cac8db6ce0b7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            362,
+            88,
+            441,
+            151
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/69919e8e1c6efec27db2fa17703209ebafd9fb8767da1ce3158618272fad3ff7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            442,
+            88,
+            522,
+            151
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/4e727456ce94a10c5cb0cf85dd00e0f157e05f4b06e02049f397c50584441bdd.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            88,
+            606,
+            151
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/40604b2aa996fc6d2acd23a2db2b3411973eea072538acf29f635c8bee37d1b6.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            609,
+            88,
+            687,
+            151
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/414e09e5425d5e35225c67a5ef070930a8035b17cd7948c4f042cab6758ae06a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            691,
+            88,
+            771,
+            151
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/fed734f68f0698bb790e0edb0feec83072cb67a77fe1500b92938e93eb4d9317.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            88,
+            854,
+            151
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/3404ad765f09bb27d600e94b7dfd8336954bec7b7f517c9d336de8e7692620ea.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            112,
+            157,
+            192,
+            220
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5d483df94f00a8f21a401b78a622031777a572b55a0ad53f3c54126f06199994.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            194,
+            157,
+            276,
+            220
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/2f5865bfe29388b182d03ed4d7a9d27dd59feab9fd7f2c0fb1094a9e3561bc51.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            277,
+            157,
+            359,
+            220
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/514cacf7cd21b690fc389cf3ac74b73b37cf9698264c21cad61806b3e19a0ea3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            361,
+            157,
+            439,
+            220
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6d511aa5d5a396393d77990f87760f332acc794d63f47c209639a816d5ea6435.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            442,
+            157,
+            522,
+            220
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/70d3e059bb62413c66fd955421186f77044e156d0e45ee6c2b8d47b983c07b88.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            157,
+            606,
+            220
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/18926e0a02b7ae432452adbefe702e8868b7a75fe3646dec0b6b5d9e908bf0d5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            609,
+            157,
+            687,
+            220
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b9020c9c3663e6a403c4af94a5a9fa47af6149c2161bc18ebf0b4b4174ee17e0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            689,
+            157,
+            771,
+            220
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/18cada0a23f99cc471fbaa33df513dc3bdc55efddbbf3f720784a087bb0d7786.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            157,
+            854,
+            220
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/abdd874750024f19fafebca20a06b1f70aa928706d99df413ba354979d75ba5c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            112,
+            229,
+            192,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/8c349ae9a21dac090ccda0bff743883591aeeca9e5450b427eac42b730eaa9f6.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            194,
+            229,
+            276,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/3c10c4d85aa3cca3c2b78f3e2c34e7a1a4bfbaa64af74b33787d9f6fe596d827.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            277,
+            229,
+            359,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/17e23469abdc5c7ab7a6a5550d1010eb601a5984b5074e13f879ce44ad308ecd.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            361,
+            229,
+            439,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6a6831ddbb77f1d43713ce4c0cf619c551296f51f84ded434ab7b7241a9f5c62.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            442,
+            229,
+            522,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/fd0efa99b095cb5c8d77d101660f6c7bb18b5ee8d5d2d7cb16d3cd628893f63d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            229,
+            606,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/73780beddb0fb11c3d6d1430fac4b29508bdfa7ada34735177e0e1aa0d8d7c4c.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            609,
+            229,
+            687,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/be464290184a7120b5a3faa800bfa8fea8bed08c5546356fdabcb1efb9302981.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            689,
+            229,
+            771,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/455b641de64cefbaf69d4b86b1a3d711d7191a180fdb1111b6aad142ca9b34d4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            229,
+            854,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/c20f63e03b90141305ab3f748fe6e81e7db5f7761224f2e498b225b5bd0fbdb8.jpg",
+        "image_caption": [
+            "Source"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            112,
+            301,
+            192,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/914f03cb4787c3219044fec13bdedab69d72544c0aaf586e94c11589a47cc118.jpg",
+        "image_caption": [
+            "Driver"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            194,
+            301,
+            276,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/2a4a11262eaefe3893566bd9f65ab0cf40f2dd90dbc5eb22abf893187532cf4e.jpg",
+        "image_caption": [
+            "FOMM [22]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            277,
+            301,
+            357,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/167c2eae9f425be2922aad926fe782a8fae893bef8708cefbea97fb2b42f3c59.jpg",
+        "image_caption": [
+            "UVA [17]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            359,
+            301,
+            439,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/8a024063cda079847e308c3b84df1cd3d92d6782b38b156413a97c18dec990a1.jpg",
+        "image_caption": [
+            "StyleHEAT [31]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            442,
+            301,
+            522,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ca68c0f5dd22b90c2d4067ccb3bef63902c3770d8b9a7a684fa690295510ba96.jpg",
+        "image_caption": [
+            "MetaPortrait [35]",
+            "Figure 9. A qualitative comparison of head avatar systems in cross-reenactment scenario."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            301,
+            606,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/2f6050cbd5ae51b790b8714d4b1c4c863640bf93ba51abf7f465583f4c63ce65.jpg",
+        "image_caption": [
+            "NOFA [32]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            609,
+            301,
+            687,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a60b6865628d9d29031984af0b311f49aa111e0d10c11263007409dd4919132c.jpg",
+        "image_caption": [
+            "MegaPortraits [9]"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            691,
+            301,
+            771,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5cd2ec3d01fca131da0568ab9ce41e09a269a4f1cdc110e312097c112488ba8f.jpg",
+        "image_caption": [
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            772,
+            301,
+            854,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/234c49cf3f3f200a5accfefb0aaf8eed6b3d111522a888153bde86d510c0a6e8.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>FID↓</td><td>\\( F_P \\downarrow \\)</td><td>\\( F_V \\downarrow \\)</td><td>\\( M_P \\downarrow \\)</td><td>\\( M_V \\downarrow \\)</td><td>Sync. offset ↓</td><td>Sync. conf. ↑</td></tr><tr><td>StyleHEAT [31]</td><td>62.4</td><td>1.34</td><td>0.41</td><td>2.98</td><td>1.67</td><td>0.41</td><td>7.22</td></tr><tr><td>PC-AVS [39]</td><td>71.5</td><td>2.78</td><td>0.51</td><td>2.56</td><td>1.04</td><td>3.79</td><td>8.42</td></tr><tr><td>SadTalker [36]</td><td>55.6</td><td>0.91</td><td>0.35</td><td>2.12</td><td>1.17</td><td>0.27</td><td>7.58</td></tr><tr><td>MakeIRTalk [40]</td><td>63.3</td><td>2.15</td><td>0.38</td><td>3.11</td><td>1.59</td><td>2.98</td><td>5.17</td></tr><tr><td>Ours (EMOPortraits)</td><td>28.5</td><td>0.82</td><td>0.33</td><td>2.53</td><td>1.38</td><td>0.07</td><td>5.77</td></tr></table>",
+        "bbox": [
+            81,
+            431,
+            467,
+            489
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/a33ea7fa3dd7e2f88d704e8bd1b0a93c9948c355710e8868de8a970281300f33.jpg",
+        "table_caption": [
+            "Table 5. Our model matches other leading speech-driven models, exceeding in realism (FID), facial geometry and velocity, but is slightly less accurate in mouth landmark positions and velocities. We also score best in audio-video sync offset, but lower in confidence. We believe that while SyncNet metrics are still helpful, they may not fully represent true synchronization quality, as generated videos can sometimes score higher than ground truth without necessarily being better synchronized. In our experimentation, ground truth videos scored an offset of 1.13 (which is worse than our method, SadTalker, and StyleHEAT) and a confidence score of 8.35 (less than PC-AVS). For qualitative analysis, please refer to our supplementary materials."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>FID↓</td><td>MP↓</td><td>MV↓</td><td>FP↓</td><td>FV↓</td></tr><tr><td>Ours w/o disentanglement</td><td>44.7</td><td>0.98</td><td>0.45</td><td>5.78</td><td>2.55</td></tr><tr><td>Ours w/o LPCA</td><td>29.3</td><td>0.84</td><td>0.35</td><td>3.13</td><td>1.88</td></tr><tr><td>Ours with MAE(\\( \\mathbf{z}_i, \\hat{\\mathbf{z}}_i^{\\text{speech}} \\))</td><td>28.2</td><td>0.84</td><td>0.34</td><td>2.61</td><td>1.45</td></tr><tr><td>Ours</td><td>28.5</td><td>0.82</td><td>0.33</td><td>2.53</td><td>1.38</td></tr></table>",
+        "bbox": [
+            81,
+            705,
+            467,
+            773
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6. Ablation study of our speech-driven mode, where in the first line, we show that it notably underperforms without head pose - face expression disentanglement of the latent space (see Sec. 5.1). Removing  $\\mathcal{L}_{\\mathrm{PCA}}$  leads to a significant reduction in generating realistic mouth movements. Replacing  $\\mathcal{L}_{\\mathrm{PCA}}$  with  $\\mathrm{MAE}(\\mathbf{z}_i,\\hat{\\mathbf{z}}_i^{\\mathrm{speech}})$  causes a noticeable, decline in mouth movement generation. This suggests that identifying mouth-related PCA components was beneficial in enhancing the outcome.",
+        "bbox": [
+            75,
+            785,
+            470,
+            896
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.3. Ablation study",
+        "text_level": 1,
+        "bbox": [
+            500,
+            426,
+            650,
+            441
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We conducted an extensive ablation study to evaluate the contributions of individual components within our method. For our main model, we present the evaluation of the importance of the proposed source-driver unmatched loss, canonical volume loss, and change of  $dim(\\mathbf{z})$  from 512 in original model to 128 Tab. 4. For our speech-driven mode, we show how vital the disentanglement of the latent space for speech-drive ability, how  $\\mathcal{L}_{\\mathrm{PCA}}$  is better than naive matching of PCA components from random  $\\mathcal{Z}$  and how it helps match speech and latent expression space Tab. 6. For more ablation analysis, plots and visual comparison, please refer to our supplementary.",
+        "bbox": [
+            496,
+            449,
+            895,
+            632
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            647,
+            619,
+            662
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we introduce EmoPortraits, a novel method for creating neural avatars with superior performance in image-driven, cross-identity emotion translation. Our speech-driven mode makes it possible to drive the facial animation through multiple conditions (video, audio, head motion). We collected FEED dataset which, we believe, will be a valuable asset for researchers in diverse human-centered studies.",
+        "bbox": [
+            496,
+            672,
+            893,
+            777
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "However, our method has some limitations. It doesn't generate the avatar's body or shoulders, limiting some use cases. We currently integrate our output with a source image body. Additionally, the model sometimes struggles with accurate expression translation and underperforms with extensive head rotation. These challenges are crucial for future enhancements and remain central to our ongoing research efforts.",
+        "bbox": [
+            496,
+            779,
+            895,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8505",
+        "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] Emad Barsoum, Cha Zhang, Cristian Canton Ferrer, and Zhengyou Zhang. Training deep networks for facial expression recognition with crowd-sourced label distribution. In Proceedings of the 18th ACM international conference on multimodal interaction, pages 279-283, 2016. 3",
+            "[2] Stella Bounareli, Christos Tzelepis, Vasileios Argyriou, Ioannis Patras, and Georgios Tzimiropoulos. Hyperreenact: One-shot reenactment via jointly learning to refine and retarget faces, 2023. 2",
+            "[3] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020. 2",
+            "[4] Houwei Cao, David G Cooper, Michael K Keutmann, Ruben C Gur, Ani Nenkova, and Ragini Verma. Crema-d: Crowd-sourced emotional multimodal actors dataset. IEEE transactions on affective computing, 5(4):377-390, 2014. 3",
+            "[5] Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age, 2018. 7",
+            "[6] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7",
+            "[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 4",
+            "[8] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing, 2021. 2",
+            "[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars, 2023. 2, 4, 5, 7, 8",
+            "[10] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction, 2020. 2",
+            "[11] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos, 2022. 2",
+            "[12] Siddharth Gururani, Arun Mallya, Ting-Chun Wang, Rafael Valle, and Ming-Yu Liu. Space: Speech-driven portrait animation with controllable expression, 2022. 3, 7",
+            "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, 2017. 7",
+            "[14] Philip Jackson and SJUoSG Haq. Surrey audio-visual expressed emotion (savee) database. University of Surrey: Guildford, UK, 2014. 3",
+            "[15] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars, 2022. 2"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[16] Alexander C. Li, Alexei A. Efros, and Deepak Pathak. Understanding collapse in non-contrastive siamese representation learning, 2022. 4",
+            "[17] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar, 2023. 7, 8",
+            "[18] Steven R Livingstone and Frank A Russo. The ryerson audiovisual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5):e0196391, 2018. 3",
+            "[19] Patrick Lucey, Jeffrey F Cohn, Takeo Kanade, Jason Saragih, Zara Ambadar, and Iain Matthews. The extended cohn-kanade dataset (ck+): A complete dataset for action unit and emotion-specified expression. In 2010 iiee computer society conference on computer vision and pattern recognition-workshops, pages 94-101. IEEE, 2010. 3",
+            "[20] Maja Pantic, Michel Valstar, Ron Rademaker, and Ludo Maat. Web-based database for facial expression analysis. In 2005 IEEE international conference on multimedia and Expo, pages 5-pp. IEEE, 2005. 3",
+            "[21] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 3",
+            "[22] Aliaksandr Siarohin, Stéphane Lathuilière, S. Tulyakov, Elisa Ricci, and N. Sebe. First order motion model for image animation. ArXiv, abs/2003.00196, 2019. 2, 7, 8",
+            "[23] Aliaksandr Siarohin, Willi Menapace, Ivan Skorokhodov, Kyle Olszewski, Hsin-Ying Lee, Jian Ren, Mengelei Chai, and Sergey Tulyakov. Unsupervised volumetric animation. arXiv preprint arXiv:2301.11326, 2023. 7",
+            "[24] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation, 2023. 3",
+            "[25] Mani Kumar Tellamekala, Ömer Ümer, Björn W. Schuller, Elisabeth Andre, Timo Giesbrecht, and Michel Valstar. Are 3d face shapes expressive enough for recognising continuous emotions and action unit intensities?, 2023. 2",
+            "[26] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition, 2018. 5",
+            "[27] H. Wang, Yitong Wang, Zheng Zhou, Xing Ji, Zhifeng Li, Dihong Gong, Jin Zhou, and Wenyu Liu. Cosface: Large margin cosine loss for deep face recognition. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5265-5274, 2018. 6",
+            "[28] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 3",
+            "[29] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "8506",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[30] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. AvatarMAV: Fast 3d head avatar reconstruction using motion-aware neural voxels. In *Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Proceedings*. ACM, 2023. 2",
+            "[31] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan, 2022. 2, 3, 7, 8",
+            "[32] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction, 2023. 2, 7, 8",
+            "[33] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models, 2019. 2",
+            "[34] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars, 2020. 2",
+            "[35] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation, 2023. 2, 7, 8",
+            "[36] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation, 2023. 3, 7, 8",
+            "[37] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7",
+            "[38] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos, 2022. 2",
+            "[39] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation, 2021. 2, 3, 7, 8",
+            "[40] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. MakeltTalk. ACM Transactions on Graphics, 39(6):1-15, 2020. 3, 7, 8"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            699
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "8507",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/ab68d1cd-73a3-4a4c-9896-716aeaaff436_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57c622f292fb82ba78bc221eb442f36191520e00ccc87bbdbba9030f7bced5a3
+size 2935821

2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/full.md

@@ -0,0 +1,399 @@
+# EMOPortraits: Emotion-enhanced Multimodal One-shot Head Avatars
+
+Nikita Drobyshev
+
+Antoni Bigata Casademunt  
+Stavros Petridis
+
+Konstantinos Vougioukas  
+Maja Pantic
+
+Zoe Landgraf
+
+{n.drobyshev, a/bigata-casademunt22, k.vougioukas, zoe.landgraf15, stavros.petridis04, m.pantic}@imperial.ac.uk Imperial College London
+
+![](images/5571c03b59a22946cb276d1deded7196a8c084b040541ce529bbee5da0acb26d.jpg)
+
+![](images/6f78ae2acda6416384be5581a2bc2ae87413212cd8bfa3feb9df5869a0f8e39f.jpg)
+
+![](images/867efe1bddaf1887f13140417c41e46dc4c66e055a39bb4c7e189e5a87cba5b5.jpg)
+
+![](images/c12038d2254398ecb4c7ba204508b41706cc611de0c85c955e62dc16a81d8eb4.jpg)
+
+![](images/272336b8074a1488e724b3c0fbb531059f7b99cb096e5d5bc5ee91f1570dcb0f.jpg)
+
+![](images/e1eb172f4eb38c6ece771d39f7a6007b8dcd48c8729563a1872fa82a97e69526.jpg)
+
+![](images/d5150c201a58cf952e956203832059c28533b90018e0a66348c1f3f500048536.jpg)
+
+![](images/316eae8164a7f14a42016f25d69f3acd05748b74e70f6c167f52db9458ece090.jpg)
+
+![](images/479f825123fe7a77d8b249bb5eb77206881d58fc1981346f540388e496db4655.jpg)
+
+![](images/8da0d6fcd24f450ef1006655f2860be5194a1760923b15e36c23cfbcd55b3cd1.jpg)
+
+![](images/1d3be78635c31aa04786c174e323af741e93526bbfde5ad9472c0a7afda2dfaf.jpg)  
+Source
+
+![](images/a2ab0f9b76c3a8b179b604978468a109567c35f2493df4b12487308d887b0ea9.jpg)  
+Driver
+
+![](images/7af6d1668fe3d655129ce10d232c0993eb2693facbf0ba6bfeccb40e09db4e40.jpg)  
+Output
+
+![](images/37fa260d97e9e5d2b12fdb7dff66ab9aa5c86ba957ebcccf9309e28ed674827c.jpg)  
+Driver
+
+![](images/6f169bc4bce76438b2a929bd4d15373bc33cedd96bc3fcaca9948d0a78a551f7.jpg)  
+Output  
+Figure 1. Selected animation results for image-driven mode of our method.
+
+# Abstract
+
+Head avatars animated by visual signals have gained popularity, particularly in cross-driving synthesis where the driver differs from the animated character, a challenging but highly practical approach. The recently presented MegaPortraits model has demonstrated state-of-the-art results in this domain. We conduct a deep examination and evaluation of this model, with a particular focus on its latent space for facial expression descriptors, and uncover several limitations with its ability to express intense face motions. To address these limitations, we propose substantial changes in both training pipeline and model architecture, to introduce our EMOPortraits model, where we:
+
+Enhance the model's capability to faithfully support intense, asymmetric face expressions, setting a new state-of-the-art result in the emotion transfer task, surpassing previous methods in both metrics and quality.
+
+Incorporate speech-driven mode to our model, achieving top-tier performance in audio-driven facial animation, making it possible to drive source identity through diverse modalities, including visual signal, audio, or a blend of both.
+
+Furthermore, we propose a novel multi-view video dataset featuring a wide range of intense and asymmetric facial expressions, filling the gap with absence of such data in existing datasets.
+
+# 1. Introduction
+
+# 1.1. Representation of facial expression
+
+Advancements in neural head technologies now enable the creation of realistic avatars from a few or even one image, with the latter being crucial when only one source image is available. Cross-driving synthesis, where avatars are animated with different identities is a key technique for virtual reality, filmmaking, photo animation, etc. However, accurate transfer of facial expressions, especially those that are intense and uneven, remains a substantial challenge in avatar animation, particularly for cross-driving synthesis. While previous research focused mainly on preserving identity and transferring moderate facial motions, our work also seeks to accurately drive high-intensity and asymmetric expressions.
+
+Our research is built on and extends the MegaPortraits [9] model, notable for state-of-the-art results in cross-driving synthesis. An additional advantage using this method is that, unlike many avatar systems that depend on limited [3, 25] predefined motion descriptors, MegaPortraits, learns expression representations from scratch, allowing for greater adaptability to a wider range of expressions. We investigate the latent expression spaces and training methods of MegaPortraits to enhance its ability to depict a broad spectrum of facial expressions. Our comprehensive analysis reveals that, while the original model shows limited effectiveness in intense motion representing, it has significant potential for improvement through targeted architectural modifications, adjustments in the training approach, and the integration of our novel dataset.
+
+# 1.2. Integrating speech driving
+
+We integrate speech in our model to be used either as a complementary or an alternative driver, a crucial aspect for applications like virtual assistance or mixed reality when a primary visual signal is absent.
+
+By improving disentanglement of facial expression latent space, we remove disturbing information from face motion descriptors and emphasize components that solely control lip movements. Based on this, we formulated a novel loss that helps to achieve desirable results. Furthermore, our method can generate plausible head rotations and blinks, which appear natural and enhance its applicability across various tasks. Thus, our final method can drive source identity through image, audio, or a mix of them.
+
+# 1.3. FEED: Facial Extreme Emotions Dataset
+
+Public data scarcity is a common obstacle in deep learning, the topic of human facial expressions and movements is not an exception. In particular, we note a lack of high quality video data capturing a wide range of facial expressions, with existing datasets not going beyond basic facial actions shown at Fig. 2. To address this, we collect a new dataset that
+
+includes basic expressions and also captures complex movements like blinks, winks, and head and tongue movements, along with varied extreme expressions which are difficult or impossible to categorize through basic facial actions. Given this variety in expressions in our dataset, we believe that it will be a valuable resource for research in human emotions and facial reconstruction fields. In summary, our main contributions are as follows:
+
+- We introduce our new EMOPortraits model for one-shot head avatars synthesis, that is capable of transferring intense facial expression, showing state-of-the-art results in cross-driving synthesis. We achieve this through careful latent facial expression space development as well as novel losses and a minimal amount of domain-specific data ( $\sim$  0.1% of the train set).  
+- We integrate a speech-driving mode in our model, that demonstrates cutting-edge results in speech-driven animations. It functions effectively alongside visual signals or independently, also generating realistic head rotations and eye blinks.  
+- We present a unique multi-view dataset that spans a broad spectrum of extreme facial expressions, filling the gap of absence of such data in existing datasets.
+
+# 2. Related Work
+
+# 2.1. Visual-driven head avatars
+
+In recent years, the domain of neural avatars has branched into two prominent sub-fields: person-specific and person-agnostic avatars. While person-specific methodologies [10, 11, 30, 38, 39] excel in delivering stunning realism and motion fidelity for a particular individual, they encounter challenges representing an arbitrary person. Moreover, these approach demands multiple frames of training data, distinct training for each avatar, and can struggle to replicate motions not encountered during training.
+
+Person-agnostic methods, on the other hand, don't need train or fine-tuning for each new person and present an alternative approach to talking-head synthesis. Earlier works in this domain generated avatars in a few-shot technique [3, 33], while subsequent studies introduced one-shot capabilities [2, 8, 9, 22, 29, 31, 32, 34]. Many of these works employ predefined motion representations, such as 3DMM's blendshapes [8, 15, 31, 32]. In contrast, some learned latent emotion representation from scratch [3, 9, 22, 34, 35]. This latter approach holds potential for better motion representation as in such settings, expression descriptors become entirely trainable and free from inheriting the limitations typical for blendshapes as was shown in [3, 25]. In our work, we adopt this strategy, remarkably improving upon the methodology used in MegaPortraits[9].
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">BFEs</td><td colspan="6">FFED dataset internal structure</td></tr><tr><td>Head rotations</td><td>Winks</td><td>Eyes move</td><td>Asym. em.</td><td>Tongue em.</td><td>Extreme em.</td></tr><tr><td>Mode</td><td>Mild/Strong</td><td>Axes/random</td><td>Mild/Strong</td><td>Random</td><td>Strong</td><td>Strong</td><td>Extreme</td></tr><tr><td># of part.</td><td>21/23</td><td>21</td><td>23</td><td>21</td><td>23</td><td>23</td><td>23</td></tr><tr><td>Avg. length</td><td>1:38/1.52</td><td>0:58</td><td>0:49</td><td>0:33</td><td>1:41</td><td>1:53</td><td>3:02</td></tr></table>
+
+Table 1. The tasks we asked our participants to perform for our FEED dataset. Here BFEs stands for basic facial expressions, shown in Fig. 2  
+
+<table><tr><td rowspan="2">Dataset</td><td colspan="7">Datasets comparison</td></tr><tr><td>#Actors</td><td>#Views</td><td>Basic exp.</td><td>Strong exp.</td><td>Extrme &amp; Assym. exp.</td><td>Tongue exp.</td><td>Resolution</td></tr><tr><td>SAVESS</td><td>4</td><td>1</td><td>✓</td><td>✘</td><td>✘</td><td>✘</td><td>1280 x 1024</td></tr><tr><td>RAVGESS</td><td>24</td><td>1</td><td>✓</td><td>✘</td><td>✘</td><td>✘</td><td>1920 x 1080</td></tr><tr><td>CREMA-D</td><td>91</td><td>1</td><td>✓</td><td>✘</td><td>✘</td><td>✘</td><td>1280 x 720</td></tr><tr><td>MMI</td><td>25</td><td>1</td><td>✓</td><td>✓</td><td>✘</td><td>✘</td><td>1920 x 1080</td></tr><tr><td>MEAD</td><td>48</td><td>7</td><td>✓</td><td>✓</td><td>✘</td><td>✘</td><td>1920 x 1080</td></tr><tr><td>Ours (FEED)</td><td>23</td><td>3</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>3840 x 2160</td></tr></table>
+
+Table 2. Comparisons with modern, high-quality audio-visual datasets created in controlled settings.
+
+# 2.2. Face expression datasets
+
+Early image datasets[1, 19] predominantly offered annotated face expression data corresponded to up to 8 basic emotions. However, these datasets are not well-suited for training head avatars, which necessitate video or multi-view data.
+
+The SAVEE [14] dataset captures facial expressions in speech, but involves just 4 actors. The MMI [20] dataset, more expansive in actor participation, offers spontaneous moderate-intensity expression but is restricted to single-view sequences. RAVDESS [18] distinguishes itself by capturing two motion intensities, but its limited recordings challenge broad applicability. CREMA-D [4] encompasses three expression intensities, and though MEAD [28] provides detailed multi-view data across three intensities, it's limited to the standard eight expression groups. Our novel FEED dataset aims to address the scientific community's demand for high-quality multi-view facial expression videos outside the standard categories: Joy, Fear, Sadness, Disgust, Anger Contempt and Surprise.
+
+# 2.3. Speech-driven head avatars
+
+While numerous studies introduce speech-driven avatars [12, 21, 24, 31, 36, 39, 40], few excel in producing high-quality talking heads with authentic rotations, blinks, and the capability to use both visual and audio inputs. For instance, Wav2Lip [21] aims at re-dubbing videos with accurate lip motions, but often falls short in realism with using single images. Diffused heads [24] struggles to generate long sequences and doesn't provide access to pose control. MakeItTalk [40] animates facial landmarks in a speaker-specific manner, yet struggles with head pose controls due to its non-reliance on 3D. PC-AVS [39] offers a solution, but demands a driving video for pose modulation. Newly developed models such as SadTalker [36] and StyleHEAT [31] have shown promising outcomes in generating high-quality talking heads. We conducted a quantitative evaluation of our audio-driven mode compared to the aforementioned models.
+
+![](images/e582360968f3b94ebaa9e6716b25e04a6f5eb78834c851fa58aea8caced7cf0b.jpg)
+
+![](images/33d7247f6f7c8a4a0d29ea41ed87dcebf4c56350ecb13484916dc2b21c5c7926.jpg)
+
+![](images/a0ffc5d39905b35e0710031436f2b4a76dde5773b3d2205c14a51ddcbc729cf5.jpg)  
+Figure 2. Illustration of the problem in publicly available face expression data and selected examples from our FEED dataset.
+
+# 3. FEED dataset
+
+As previously mentioned, the current publicly available datasets with human face expression videos fall short in capturing a broad range of facial manipulations (see illustration at Fig. 2). In our view, this gap could stem from the complexity of designing clear tasks for participants, coupled with the reality that not everyone is ready for performing extreme expressions on camera. To bridge this gap, our pioneering FEED dataset is designed to meet the scientific community's need for high-quality, multi-view emotion videos that extend beyond standard emotions. It contains various expressions, including strong asymmetric ones, tongue and cheeks movements, winks, head rotations, eye movements and more nuanced gestures. ;
+
+Our dataset consists of 520 multi-view videos of 23 subjects, captured with 3 cameras. As extreme face motions are complex and their perception can be heavily influenced by subtle differences, we use a high resolution of 4k for all video, capturing the whole face up to the level of individual hair strands and wrinkles, as shown in Fig. 2. For more examples, please refer to the supplementary materials. Our participants were asked to perform 7 tasks (see Tab. 1) to cover as many facial and head movements as possible.
+
+In our comparison with other expression datasets detailed in Tab. 2, the MEAD dataset [28] emerges as the closest to ours regarding viewpoints and maximal expression intensity. However, despite MEAD's larger number of actors and viewpoints, it offers significantly less variety of facial expressions and have lower resolution than our dataset.
+
+# 4. Expression enhancement
+
+We found that the MegaPortraits [9] model fails to transfer intense expressions correctly, as shown in Fig. 9. Using our FEED dataset of strong and uneven facial expression to fine-tune a pre-trained MegaPortraits model didn't enhance final results as shown at Tab. 4. Training from scratch on FEED leads to fast overfitting due to the small number of identities presented in the dataset compared to VoxCeleb2 [7] used in [9] (23 vs  $\sim 5000$ ). Our method effectively injects this small dataset into training, yielding desirable results. We list our key findings here, for all details, see supplementary materials. Our model's scheme, displayed in Fig. 4, shares similarities with [9], leading us to occasionally refer to certain elements of our scheme in the context of [9]'s pipeline.
+
+# 4.1. Latent expression space
+
+We begin by exploring MegaPortraits' latent expression descriptors' space  $(\mathbf{z}_{s / d}$  in Fig. 4), which is crucial for expression transfer, through PCA analysis. Inspired by [16], which demonstrated that the area under the cumulative explained variance curve of singular values (denoted here as  $\mathbf{AUC}_{\mathbf{z}}$ ) can serve as an effective metric for dimensionality collapse in some latent spaces, we applied this measure to our study.
+
+This metric allows us to forecast model performance: greater collapse suggests reduced entropy in the latent space, which correlates with lower representational quality. High quality of the expression space, produced by  $\mathbf{E}_{\text {motion }}$  (as shown in Fig. 4), is crucial to capture and differentiate between the nuances of strong facial expressions. Our ablation study confirms (see Tab. 4) that there is a notable correlation between how broad and isotropic model's latent space is (expressed by  $\mathbf{AUC}_{\mathbf{z}}$ ) and its final performance. Visual comparison is available in supplementary materials. As shown in Fig. 3, our model's latent space outperforms MegaPortraits in expression representation ability. This is supported by both visual (Fig. 9) and quantitative analysis (Tab. 3).
+
+![](images/43e0be9fd62862a686fc0a3749085dde5d7228aefbfb924c5e39ad2a1a627ee5.jpg)  
+Figure 3. Comparison of latent spaces. Left plot shows that our model's latent space is wider and exhibits more even variance distribution. Also, as shown on the right plot, a greater number of principal components are involved in capturing variance across various thresholds. This implies a more robust representational capacity of expression space compared to [9]. The VoxCeleb2 test set was used for both plots.
+
+![](images/cfaa5cbc7efaf0aaec7497921709bfe2ffc0e67e9a55e614e647ce8545e2ac06.jpg)  
+Figure 4. Method Overview. We use  $\mathbf{E}_{\mathrm{app}}$  to extract volume features  $\mathbf{V}_s$  and a global descriptor  $\mathbf{e}_s$  from the source image. Then  $\mathbf{E}_{\mathrm{motion}}$  or  $\mathbf{E}_{\mathrm{audio}}$  generates motion representations from source and driver, including head rotations  $\mathbf{R}_{s / d}$ , translations  $\mathbf{t}_{s / d}$ , and expression descriptors  $\mathbf{z}_{s / d}$ . Using them, we predict warpings  $\mathbf{w}_{s\rightarrow}$  and  $\mathbf{w}_{\rightarrow d}$ . First warping and  $\mathbf{G}_{3D}$  transform  $\mathbf{V}_s$  into a canonical volume  $\mathbf{V}_s^C$  by removing the source motions. Second warping and  $\mathbf{G}_{2D}$  imposes the driver's motions and renders the final image.
+
+Both metrics shown in Fig. 3 were vital in our model's development since they highly correlate with the final model's ability to transfer intense expressions. This was particularly useful for finding adjustments described in Sec. 4.2 and Sec. 4.3, as early visual results during training are not very telling, but changes in the latent space are noticeable after a few epochs. We also found that in MegaPortraits, just a few principal components significantly affect variance, with only 18 components making up  $99\%$  (Fig. 3), suggesting their  $512\mathbf{z}$ 's dimension is oversized. For our model, we found a 128-dimensional  $\mathbf{z}$  is optimal, offering solid performance and reducing overfitting in imbalanced data, as our ablation study reveals. We calculate the Explained Variance  $(\mathrm{EV}_i)$  for the  $i^{th}$  component and  $(\mathbf{AUC}_{\mathbf{z}})$  using equations shown in Eq. (1), which involves standardizing the vector set  $\mathcal{Z}$  for PCA and sorting the eigenvalues  $\lambda$  in descending order.
+
+$$
+\mathbf {Z} _ {\mathrm {s t d}} = \frac {\mathcal {Z} - \bar {\mathcal {Z}}}{\sigma (\bar {\mathcal {Z}})}, \quad \boldsymbol {\lambda} = \operatorname {s o r t e d} _ {\mathrm {d e s c}} \left(\operatorname {e i g} \left(\frac {1}{n - 1} \mathbf {Z} _ {\mathrm {s t d}} ^ {\top} \mathbf {Z} _ {\mathrm {s t d}}\right)\right), \tag {1}
+$$
+
+$$
+\mathrm {E V} _ {i} = \frac {\lambda_ {i}}{\sum_ {j = 1} ^ {p} \lambda_ {j}}, \quad \mathbf {A U C _ {z}} = \frac {1}{d} \frac {\sum_ {i = 1} ^ {d} \sum_ {k = 1} ^ {i} \lambda_ {k}}{\sum_ {j = 1} ^ {d} \lambda_ {j}},
+$$
+
+# 4.2. Canonical volume
+
+Despite being designed to exclude face expression details, we found that canonical volume  $(\mathbf{V}^C$  in Fig. 4) in MegaPortraits actually retains significant expression information from the source image, contributing to poor translation of intense expressions, as shown in our experiment detailed in Fig. 5. In this experiment, using portraits with varied expressions, we assessed the expression translation accuracy and how expression intensity affects the  $\mathbf{V}^C$ s. Our findings suggest that the canonical volumes in MegaPortraits are not truly neutral. As confirmed by our ablation study (Tab. 4), creating an expression-free  $\mathbf{V}^C$  is essential for precise translation of intense motions. We believe that a more neutral canonical volume improves tractability and effectiveness in expression translation tasks.
+
+![](images/65fd3cb501e87e562244924ab7734643dc949462bd6d5e71948526318373e41c.jpg)  
+Figure 5. Canonical volumes  $(\mathbf{V}^C)$  in MegaPortraits [9] are not expression-neutral. To show it we use three portraits: A (regular expression), B (intense expression), D (regular expression, new identity). Case 1 visualizes poor results in transferring B's intense expression to A using a self-driving mode, contrasting with the effective reconstruction of B when used as both driver and source in case 3. This discrepancy, indicative of expression leakage into the canonical volume, is further quantified by a  $43\%$  relative difference in  $\mathbf{V}^C$ s between B and C, contrary to expectations of their similarity for same identity. In cross-driving generation case 2, using B as the driver and D as the source yielded poor results. However, in case 4, after using additive operations on source canonical volumes, the output expression is much closer to driver than in case 2. This manipulation again confirms the significant retention of expression information in canonical volumes.
+
+To overcome this issue, we propose to match canonical volumes from the different images of the same person  $(\mathbf{V}_{s^n}^C,\mathbf{V}_{d^n}^C)$  during training as following:
+
+$$
+\mathcal {L} _ {\mathrm {C V}} ^ {\mathrm {n}} = \mathcal {L} _ {\mathrm {M A E}} \left(\mathbf {V} _ {s ^ {n}} ^ {C}, \mathbf {V} _ {d ^ {n}} ^ {C}\right), \tag {2}
+$$
+
+This loss ensures  $\mathbf{V}^C$  remains stable and expression-independent, crucial for translating intense expressions, as shown in our ablation study (Tab. 4).
+
+# 4.3. Source-driver mismatch loss
+
+In addition to maintaining expression-free  $\mathbf{V}^{C}$ , it's vital to remove all identity information from the expression vector  $\mathbf{z}$ . While contrastive losses address this in [9], our experiments indicate they are insufficient to prevent overfitting in our imbalanced data scenario, where emotionally intense images represent only  $0.1\%$  of our training set but are sampled  $25\%$  of the time. We introduce a novel self-supervised loss which mitigates identity information in latent expression vectors:
+
+$$
+\begin{array}{l} \mathcal {L} _ {s d m} (\mathbf {z} _ {s}, \mathbf {z} _ {d}) = w * \max (0, \cos (\mathbf {z} _ {s}, \mathbf {z} _ {d}) - m a r g i n), \\ w; \text {m a r g i n} = \left\{ \begin{array}{l l} 1; 0. 5 & \text {i f} \mathbf {z} _ {s}, \mathbf {z} _ {d} \text {a r e f r o m V C 2} \\ 1 0; 0. 2 5 & \text {i f} \mathbf {z} _ {s}, \mathbf {z} _ {d} \text {a r e f r o m F E E D} \end{array} \right. \tag {3} \\ \end{array}
+$$
+
+Here,  $\mathbf{z}_s, \mathbf{z}_d$  are emotion vectors from source and driver images,  $w, \text{margin}$  adjust the loss's intensity and strictness respectively. We increase  $w$  for FEED due to its higher sampling rate and decrease margin for more assured expression variance in  $s, d$ . Refer to Fig. 6 for a visual summary of all self-supervised losses enhancing our latent emotion space.
+
+![](images/4168394fbcbb61f875f732ea64d8e9e00b593fc9b8b7198c39c30b7289b588f7.jpg)  
+Figure 6. The visualisation of our self-supervised losses.  $\mathcal{L}_{\mathrm{cos}}$  is a contrastive loss, similar to one used in [9] but uses different pairs as described in Sec. 5.1. Our novel  $\mathcal{L}_{sdm}$  is designed to prevent overfitting in an imbalanced dataset.
+
+# 5. Incorporating speech-driving mode
+
+# 5.1. Latent space disentanglement
+
+During our latent space analysis described in Sec. 4, we found that combining expression from one driver with the head pose from another resulted in poor performance for MegaPortraits model, as illustrated in Fig. 7 (bottom row). Indeed, the base model lacks a mechanism to intentionally prevent head pose leakage. Disentangled expression latent space, besides expanding the model's use cases, plays the crucial role for the speech-driving mode. When the latent vector  $\mathbf{z}$  is entangled with head pose data, predicting it from speech becomes challenging because speech lacks head rotation information, unlike images. Moreover, we would like to have a full control over head rotations during speech mode. We were able to make our latent space disentangled by changing the way we sample images for  $\mathcal{L}_{\mathrm{cos}}$  Eq. (5).
+
+The MegaPortraits model uses  $\mathcal{L}_{\mathrm{cos}}$  - a modified Large Margin Cosine Loss [26] to mitigate the transfer of appearance features to expression descriptor (latent vector  $\mathbf{z}$  on Fig. 4), which is crucial for cross-driving mode. Without this, differences in appearance between the source and driver, like hairstyle or skin tone, leak from driver to output image.
+
+For computing this loss, the authors utilize a supplementary source-driving pair  $(\mathbf{x}_{s^{*}}$  and  $\mathbf{x}_{d^{*}})$  from a different video with another identity, ensuring a distinct appearance from the current  $\mathbf{x}_s,\mathbf{x}_d$  pair. The base model is then employed to produce cross-reenacted image  $(\hat{\mathbf{x}}_{s^{*}\rightarrow d} = \mathbf{G}_{\mathrm{base}}(\mathbf{x}_{s^{*}},\mathbf{x}_d))$  Concurrently, they determine a separate motion descriptor,  $\mathbf{z}_{d^{*}} = \mathbf{E}_{\mathrm{motion}}(\mathbf{x}_{d^{*}})$ . Descriptors  $\mathbf{z}_{* \rightarrow d}$  from the respective forward passes are also used for positive and negative pairs.
+
+![](images/2fc771a3720ccaa6e2cde67511df1976c6cbeed5e2acdb994a7fe463b911909b.jpg)  
+Figure 7. The illustration of disentanglement problem. Images in the right row indicate the expression descriptor  $\mathbf{z}$  entangled with head pose for MegaPortraits, whereas we avoid it using changes proposed in Sec. 5.1.
+
+Our innovation here is a new sampling strategy: for each  $\mathbf{x}_s$ ,  $\mathbf{x}_d$  pair, we sample one more random additional pair  $\mathbf{x}_{s^m}$  and  $\mathbf{x}_{d^m}$  apart from  $\mathbf{x}_{s^*}$  and  $\mathbf{x}_{d^*}$ . We then apply the model to produce the following cross-reenacted moved image using source identity from  $\mathbf{x}_{s^*}$ , desired head pose from  $\mathbf{x}_{s^m}$  and emotions from  $\mathbf{x}_d$ , represented as:  $\hat{\mathbf{x}}_{s^{*m}\rightarrow d} = \mathbf{G}_{\mathrm{base}}(\mathbf{x}_{s^{*m}},\mathbf{x}_d)$ . This creates a positive pair with the same desired emotion but varying head poses.
+
+Motion descriptors are organized into positive pairs  $\mathcal{P}$  for alignment, and negative pairs  $\mathcal{N}$  for non-alignment:  $\mathcal{P} = \left\{(\mathbf{z}_{s\rightarrow d},\mathbf{z}_d),(\mathbf{z}_{s^*\rightarrow d},\mathbf{z}_d),(\mathbf{z}_{s^{*m}\rightarrow d},\mathbf{z}_d)\right\}$ , and  $\mathcal{N} = \left\{(\mathbf{z}_{s\rightarrow d},\mathbf{z}_{d^*}),(\mathbf{z}_{s^*\rightarrow d},\mathbf{z}_{d^*}),(\mathbf{z}_{s^{*m}\rightarrow d},\mathbf{z}_{d^*})\right\}$  (see Fig. 6). These pairs are used to calculate the following cosine distance:
+
+$$
+d \left(\mathbf {z} _ {i}, \mathbf {z} _ {j}\right) = s \cdot \left(\left\langle \mathbf {z} _ {i}, \mathbf {z} _ {j} \right\rangle - m\right), \tag {4}
+$$
+
+where both  $s$  and  $m$  are hyperparameters. This distance is then used to calculate a large margin cosine loss (Cos-Face) [27]:
+
+$$
+\mathcal {L} _ {\cos} = - \sum_ {(\mathbf {z} _ {k}, \mathbf {z} _ {l}) \in \mathcal {P}} \log \frac {\exp \left\{d \left(\mathbf {z} _ {k} , \mathbf {z} _ {l}\right) \right\}}{\exp \left\{d \left(\mathbf {z} _ {k} , \mathbf {z} _ {l}\right) \right\} + \sum_ {(\mathbf {z} _ {i} , \mathbf {z} _ {j}) \in \mathcal {N}} \exp \left\{d \left(\mathbf {z} _ {i} , \mathbf {z} _ {j}\right) \right\}} \tag {5}
+$$
+
+This loss prevents the head pose from leaking into the embeddings as shown in Fig. 7 (upper row)
+
+By doing this, we achieved the capability to interpret  $\mathbf{z}_i$  as a latent vector encapsulating the emotional content from an image  $\mathbf{x}_i$ . This advancement opened the door for an intriguing possibility during inference: the prediction of this vector could potentially be derived from an audio signal instead of a facial image using the motion encoder  $\mathbf{E}_{\mathrm{motion}}$ , effectively transforming the model into a speech-driven system. The critical challenge here is aligning the audio input with the pretrained latent space. We addressed this by introducing an additional audio encoder,  $\mathbf{E}_{\mathrm{aud}}$ , and employing  $\mathbf{E}_{\mathrm{motion}}$  as a teacher model for this new encoder. A crucial factor in achieving remarkable accuracy in lip synchronization was our ability to isolate specific components within the expression latent space, that are solely responsible for mouth movements. We describe it in the next subsection Sec. 5.2.
+
+![](images/426d2459ddac66dc468bd55fd6056ae9f8154ca624c50c1e768277a3b7502cea.jpg)  
+Figure 8. In the upper row, the first principal component from a random subset of expression vectors  $\mathcal{Z}$  affects not just the mouth but also blinks and gaze, making it difficult to interpret and isolate mouth movement components, needed for training  $\mathbf{E}_{\mathrm{aud}}$ . However, as shown in the bottom row, using  $\mathcal{Z}_m$  from Sec. 5.2, the first principal component,  $\mathbf{v}_1$ , focused on mouth movements, is effectively isolated,  $\mathbf{z}_0$  - latent vector, corresponded to the central image.
+
+# 5.2. Mouth movements
+
+To train the audio encoder  $\mathbf{E}_{\mathrm{aud}}$  for predicting  $\mathbf{z}$  from speech, we chose to use the motion encoder  $\mathbf{E}_{\mathrm{motion}}$  as a teacher model. However, direct matching vectors  $\mathbf{z}_i^{\mathrm{aud}}$  from  $\mathbf{E}_{\mathrm{aud}}$  and pseudo-ground truth  $\mathbf{z}_i$  from  $\mathbf{E}_{\mathrm{motion}}$  gives poor results (see Tab. 6). We believe that the reason behind this is that speech-derived  $\mathbf{z}_i^{\mathrm{aud}}$ , while able to capture lip movements, struggles with other facial motions, especially in the upper face. Thus, we focused on isolating mouth movement components in the expression latent space using PCA analysis.
+
+Employing PCA on a broad array of expression vectors  $\mathcal{Z}$ , generated from images with varied facial emotions, yields principal components with their explained variances  $(\mathrm{EV}_i)$ . Altering these components, particularly those with high  $\mathrm{EV}_i$ , significantly changes facial expressions in an image. However, they typically affect a combination of facial features—Fig. 8 (upper row), not isolated areas like the mouth or eyes.
+
+To focus on mouth movements, we created a unique set of  $\mathcal{Z}_m$  from just one video of a person performing mouth manipulations and manually edited it to minimize upper face movements. This involved using a still upper face from the first frame in all subsequent frames, ensuring that the principal components mainly represented mouth movements. Applying PCA of  $\mathcal{Z}_m$  reveals that most principal components are responsible for solely mouth movements. For illustration of the first component, see Fig. 8 (bottom row). Based on distilled components, we introduce mouth PCA mouth loss:
+
+$$
+\mathcal {L} _ {\mathrm {P C A}} \left(v _ {i}, v _ {j}, n\right) = \frac {1}{n} \sum_ {k = 1} ^ {n} \left| \mathrm {P C} _ {\mathbf {v} _ {i}} (k) - \mathrm {P C} _ {\mathbf {v} _ {j}} (k) \right| \tag {6}
+$$
+
+Where  $n$  is the number of principal components to be considered,  $\mathrm{PC}_{\mathbf{z}_i}(k)$  and  $\mathrm{PC}_{\mathbf{z}_j}(k)$  are the  $k$ -th principal components of vectors  $\mathbf{z}_i$  and  $\mathbf{z}_j$  respectively. Implementing this loss has demonstrably enhanced the quality and accuracy of our results, as our ablation study confirms (see Tab. 6).
+
+# 6. Experiments
+
+In this section, we describe experiments we set for comparison for both our image-driven and speech-driven mode. Please see implementation details such as all used losses, component's sizes and architecture details, training procedure, data preprocess in supplementary materials.
+
+# 6.1. Image-driven comparison
+
+Methods. We evaluated our image-driven model alongside various one-shot models. For this comparison, we included models like NOFA [32] and StyleHEAT [31], which use 3DMM expression blendshapes without disentanglement training, and models with trainable latent face motion representations like FOMM [22], UVA [23], MetaPortrait [35] and MegaPortraits [9]. It's noteworthy that NOFA and UVA require per-source optimization, unlike the others, including our model.
+
+Evaluation metrics and data. For our evaluation, we use 100 random, non-child, single-person images from the FFHQ dataset as sources. From the MEAD dataset, we chose 100 images covering all high-intensity emotions except neutral, and another 100 from the FEED dataset's extreme emotion task, featuring 5 identities not seen during training of our model. Each source image was paired with 20 driving images from these sets, totaling 2000 pairs.
+
+For assessing reenactment quality, we employ various metrics including the Frechet Inception Distance (FID) [13] to measure the distributional discrepancy between synthetic and real images. Cosine similarity (CSIM) from a face recognition network [5] quantifies the identity preservation in generated images. Additionally, we conducted user studies to determine preferences regarding motion (UMTN) and appearance (UAPP) preservation, presenting participants with triplets—either a driving or source image and two method outputs—and asking them to choose the one with better motion or appearance preservation. Our user study includes 34 participants and around 300 unique questions. Both metrics numerically reflect preference percentages. As our visual results Fig. 9, and quantitative metrics Tab. 3 show, our method demonstrates exceptional ability in translating strong and extreme expressions, notably surpassing other image-driven avatar techniques in user preference for facial expression translation (UMTN) and FID scores.
+
+# 6.2. Speech-driven comparison
+
+Methods. To assess our model's speech-driven capabilities, we compared it with StyleHEAT [31], which also operates using both image and audio inputs, and audio-exclusive models including SadTalker [36], MakeItTalk [40], and PC-AVS [39].
+
+Evaluation metrics and data. We employ the evaluation procedure described in [12]. We measure video realism using Frechet Inception Distance (FID) [13]. We also extract
+
+<table><tr><td>Driver:</td><td colspan="4">MEAD</td><td colspan="4">FEED</td></tr><tr><td>Method</td><td>FID↓</td><td>CSIM↑</td><td>UMTN↑</td><td>UAPP↑</td><td>FID↓</td><td>CSIM↑</td><td>UMTN↑</td><td>UAPP↑</td></tr><tr><td>MetaPortrait [35]</td><td>77.5</td><td>0.64</td><td>*</td><td>*</td><td>78.8</td><td>0.66</td><td>*</td><td>*</td></tr><tr><td>NOFA [32]</td><td>69.2</td><td>0.65</td><td>*</td><td>*</td><td>69.8</td><td>0.66</td><td>3.8</td><td>8.8</td></tr><tr><td>FOMM [22]</td><td>84.4</td><td>0.56</td><td>2.4</td><td>0.3</td><td>85.9</td><td>0.55</td><td>1.8</td><td>0.5</td></tr><tr><td>UVA [17]</td><td>78.7</td><td>0.68</td><td>0.9</td><td>6.4</td><td>79.5</td><td>0.68</td><td>0.6</td><td>9.4</td></tr><tr><td>StyleHEAT [31]</td><td>71.9</td><td>0.63</td><td>3.7</td><td>3.3</td><td>72.4</td><td>0.61</td><td>2.6</td><td>3.8</td></tr><tr><td>MegaPortraits [9]</td><td>61.1</td><td>0.73</td><td>22.7</td><td>46.6</td><td>62.8</td><td>0.73</td><td>16.7</td><td>42.8</td></tr><tr><td>Our (EMOPortraits)</td><td>59.6</td><td>0.74</td><td>70.3</td><td>43.4</td><td>60.2</td><td>0.70</td><td>74.6</td><td>34.7</td></tr></table>
+
+Table 3. Our method notably outperforms others in the FID score and strongly leads in the user preference metric for face expression translation (UMTN). It excels in reliably translating strong and extreme expressions from the driving image, distinguishing it as the most capable among the compared methods. Additionally, our approach maintains the source image's shape and appearance (measured by CSIM and UAPP metrics) on par with MegaPortraits [9] when using strong and regular expressions from the MEAD dataset as the driving signal. However, we observe a minor deviation in identity preservation metrics for extreme emotions (using the FEED dataset). This difference may arise because our method attempts to transfer asymmetric facial expressions, while MegaPortraits might modify the source image less, or not at all, with a challenging driver (as depicted in Fig. 5, case 1), thereby achieving better identity preservation. UVA was excluded from the MEAD-driven user study due to the authors were able to provide only part of the test data. Also, MetaPortrait was excluded from both user studies because of late data provision by authors, offering only the basic model results, without their super-resolution module outputs.
+
+<table><tr><td>Method</td><td>FID↓</td><td>CSIM↑</td><td>UMTN↓</td><td>UAPP↓</td><td>AUCz↓</td></tr><tr><td>MegaPortraits fine-tuned</td><td>62.3</td><td>0.63</td><td>8.5</td><td>12.4</td><td>0.89</td></tr><tr><td>Ours w/o LcnCV</td><td>61.4</td><td>0.67</td><td>9.2</td><td>17.9</td><td>0.88</td></tr><tr><td>Ours w/o Lsdm</td><td>61.8</td><td>0.59</td><td>11.3</td><td>6.7</td><td>0.85</td></tr><tr><td>Ours dim(z) = 512</td><td>63.4</td><td>0.44</td><td>25.6</td><td>1.1</td><td>0.77</td></tr><tr><td>Ours</td><td>59.6</td><td>0.74</td><td>45.4</td><td>61.9</td><td>0.80</td></tr></table>
+
+Table 4. Our ablation study examines key aspects of the image-driven mode. After fine-tuning MegaPortraits on the FEED dataset, we observed a decline in identity preservation due to overfitting on FEED identities, without any improvement in emotion translation. The lack of  $\mathcal{L}_{CV}^{n}$  primarily impacts the ability to translate extreme expression, as indicated by UMTN. Additionally, not using  $\mathcal{L}_{sdm}$  or setting  $dim(\mathbf{z}) = 512$  results in identity leakage, particularly evident in the latter scenario. For this ablation, we use the part of our driven by MEAD dataset and described in Sec. 6.1
+
+facial landmarks from outputs of all methods, frontalized, and normalized to position the mouth edges at  $(-1,0)$  and  $(1,0)$ . We compute mean absolute error (MAE) for both predicted mouth landmark positions  $(M_P)$  and velocities  $(M_V)$ , along with errors in facial geometry  $(F_P)$  and velocity  $(F_V)$ . For our evaluation, we use 100 randomly selected video sequences from the HDTF dataset [37], each up to 30 seconds. As the results show in Tab. 5, our speech-driven model performs on par with leading methods, excelling in realism and facial dynamics. SyncNet [6] is used to assess lip sync quality, providing offset (temporal misalignment between the audio and video) and confidence (certainty of audio-visual alignment) scores.
+
+![](images/78926baf7339f356df241bf8a88b63a9b83f4b6c3649de1e1ee938d86b29bfe5.jpg)
+
+![](images/249eb01f4131fdddf201108d1aabb80575d66581bff4a918301a2418cf1ced01.jpg)
+
+![](images/e6833fe36b387c24e85728fd42f64f6c1ebddbb0976d1cc0ff663463ff80b708.jpg)
+
+![](images/babf75fa3d53bebbcf1c449f3409b684032610a97ed8f57c97e1cac8db6ce0b7.jpg)
+
+![](images/69919e8e1c6efec27db2fa17703209ebafd9fb8767da1ce3158618272fad3ff7.jpg)
+
+![](images/4e727456ce94a10c5cb0cf85dd00e0f157e05f4b06e02049f397c50584441bdd.jpg)
+
+![](images/40604b2aa996fc6d2acd23a2db2b3411973eea072538acf29f635c8bee37d1b6.jpg)
+
+![](images/414e09e5425d5e35225c67a5ef070930a8035b17cd7948c4f042cab6758ae06a.jpg)
+
+![](images/fed734f68f0698bb790e0edb0feec83072cb67a77fe1500b92938e93eb4d9317.jpg)
+
+![](images/3404ad765f09bb27d600e94b7dfd8336954bec7b7f517c9d336de8e7692620ea.jpg)
+
+![](images/5d483df94f00a8f21a401b78a622031777a572b55a0ad53f3c54126f06199994.jpg)
+
+![](images/2f5865bfe29388b182d03ed4d7a9d27dd59feab9fd7f2c0fb1094a9e3561bc51.jpg)
+
+![](images/514cacf7cd21b690fc389cf3ac74b73b37cf9698264c21cad61806b3e19a0ea3.jpg)
+
+![](images/6d511aa5d5a396393d77990f87760f332acc794d63f47c209639a816d5ea6435.jpg)
+
+![](images/70d3e059bb62413c66fd955421186f77044e156d0e45ee6c2b8d47b983c07b88.jpg)
+
+![](images/18926e0a02b7ae432452adbefe702e8868b7a75fe3646dec0b6b5d9e908bf0d5.jpg)
+
+![](images/b9020c9c3663e6a403c4af94a5a9fa47af6149c2161bc18ebf0b4b4174ee17e0.jpg)
+
+![](images/18cada0a23f99cc471fbaa33df513dc3bdc55efddbbf3f720784a087bb0d7786.jpg)
+
+![](images/abdd874750024f19fafebca20a06b1f70aa928706d99df413ba354979d75ba5c.jpg)
+
+![](images/8c349ae9a21dac090ccda0bff743883591aeeca9e5450b427eac42b730eaa9f6.jpg)
+
+![](images/3c10c4d85aa3cca3c2b78f3e2c34e7a1a4bfbaa64af74b33787d9f6fe596d827.jpg)
+
+![](images/17e23469abdc5c7ab7a6a5550d1010eb601a5984b5074e13f879ce44ad308ecd.jpg)
+
+![](images/6a6831ddbb77f1d43713ce4c0cf619c551296f51f84ded434ab7b7241a9f5c62.jpg)
+
+![](images/fd0efa99b095cb5c8d77d101660f6c7bb18b5ee8d5d2d7cb16d3cd628893f63d.jpg)
+
+![](images/73780beddb0fb11c3d6d1430fac4b29508bdfa7ada34735177e0e1aa0d8d7c4c.jpg)
+
+![](images/be464290184a7120b5a3faa800bfa8fea8bed08c5546356fdabcb1efb9302981.jpg)
+
+![](images/455b641de64cefbaf69d4b86b1a3d711d7191a180fdb1111b6aad142ca9b34d4.jpg)
+
+![](images/c20f63e03b90141305ab3f748fe6e81e7db5f7761224f2e498b225b5bd0fbdb8.jpg)  
+Source
+
+![](images/914f03cb4787c3219044fec13bdedab69d72544c0aaf586e94c11589a47cc118.jpg)  
+Driver
+
+![](images/2a4a11262eaefe3893566bd9f65ab0cf40f2dd90dbc5eb22abf893187532cf4e.jpg)  
+FOMM [22]
+
+![](images/167c2eae9f425be2922aad926fe782a8fae893bef8708cefbea97fb2b42f3c59.jpg)  
+UVA [17]
+
+![](images/8a024063cda079847e308c3b84df1cd3d92d6782b38b156413a97c18dec990a1.jpg)  
+StyleHEAT [31]
+
+![](images/ca68c0f5dd22b90c2d4067ccb3bef63902c3770d8b9a7a684fa690295510ba96.jpg)  
+MetaPortrait [35]  
+Figure 9. A qualitative comparison of head avatar systems in cross-reenactment scenario.
+
+![](images/2f6050cbd5ae51b790b8714d4b1c4c863640bf93ba51abf7f465583f4c63ce65.jpg)  
+NOFA [32]
+
+![](images/a60b6865628d9d29031984af0b311f49aa111e0d10c11263007409dd4919132c.jpg)  
+MegaPortraits [9]
+
+![](images/5cd2ec3d01fca131da0568ab9ce41e09a269a4f1cdc110e312097c112488ba8f.jpg)  
+Ours
+
+<table><tr><td>Method</td><td>FID↓</td><td>\( F_P \downarrow \)</td><td>\( F_V \downarrow \)</td><td>\( M_P \downarrow \)</td><td>\( M_V \downarrow \)</td><td>Sync. offset ↓</td><td>Sync. conf. ↑</td></tr><tr><td>StyleHEAT [31]</td><td>62.4</td><td>1.34</td><td>0.41</td><td>2.98</td><td>1.67</td><td>0.41</td><td>7.22</td></tr><tr><td>PC-AVS [39]</td><td>71.5</td><td>2.78</td><td>0.51</td><td>2.56</td><td>1.04</td><td>3.79</td><td>8.42</td></tr><tr><td>SadTalker [36]</td><td>55.6</td><td>0.91</td><td>0.35</td><td>2.12</td><td>1.17</td><td>0.27</td><td>7.58</td></tr><tr><td>MakeIRTalk [40]</td><td>63.3</td><td>2.15</td><td>0.38</td><td>3.11</td><td>1.59</td><td>2.98</td><td>5.17</td></tr><tr><td>Ours (EMOPortraits)</td><td>28.5</td><td>0.82</td><td>0.33</td><td>2.53</td><td>1.38</td><td>0.07</td><td>5.77</td></tr></table>
+
+Table 5. Our model matches other leading speech-driven models, exceeding in realism (FID), facial geometry and velocity, but is slightly less accurate in mouth landmark positions and velocities. We also score best in audio-video sync offset, but lower in confidence. We believe that while SyncNet metrics are still helpful, they may not fully represent true synchronization quality, as generated videos can sometimes score higher than ground truth without necessarily being better synchronized. In our experimentation, ground truth videos scored an offset of 1.13 (which is worse than our method, SadTalker, and StyleHEAT) and a confidence score of 8.35 (less than PC-AVS). For qualitative analysis, please refer to our supplementary materials.  
+
+<table><tr><td>Method</td><td>FID↓</td><td>MP↓</td><td>MV↓</td><td>FP↓</td><td>FV↓</td></tr><tr><td>Ours w/o disentanglement</td><td>44.7</td><td>0.98</td><td>0.45</td><td>5.78</td><td>2.55</td></tr><tr><td>Ours w/o LPCA</td><td>29.3</td><td>0.84</td><td>0.35</td><td>3.13</td><td>1.88</td></tr><tr><td>Ours with MAE(\( \mathbf{z}_i, \hat{\mathbf{z}}_i^{\text{speech}} \))</td><td>28.2</td><td>0.84</td><td>0.34</td><td>2.61</td><td>1.45</td></tr><tr><td>Ours</td><td>28.5</td><td>0.82</td><td>0.33</td><td>2.53</td><td>1.38</td></tr></table>
+
+Table 6. Ablation study of our speech-driven mode, where in the first line, we show that it notably underperforms without head pose - face expression disentanglement of the latent space (see Sec. 5.1). Removing  $\mathcal{L}_{\mathrm{PCA}}$  leads to a significant reduction in generating realistic mouth movements. Replacing  $\mathcal{L}_{\mathrm{PCA}}$  with  $\mathrm{MAE}(\mathbf{z}_i,\hat{\mathbf{z}}_i^{\mathrm{speech}})$  causes a noticeable, decline in mouth movement generation. This suggests that identifying mouth-related PCA components was beneficial in enhancing the outcome.
+
+# 6.3. Ablation study
+
+We conducted an extensive ablation study to evaluate the contributions of individual components within our method. For our main model, we present the evaluation of the importance of the proposed source-driver unmatched loss, canonical volume loss, and change of  $dim(\mathbf{z})$  from 512 in original model to 128 Tab. 4. For our speech-driven mode, we show how vital the disentanglement of the latent space for speech-drive ability, how  $\mathcal{L}_{\mathrm{PCA}}$  is better than naive matching of PCA components from random  $\mathcal{Z}$  and how it helps match speech and latent expression space Tab. 6. For more ablation analysis, plots and visual comparison, please refer to our supplementary.
+
+# 7. Conclusion
+
+In this paper, we introduce EmoPortraits, a novel method for creating neural avatars with superior performance in image-driven, cross-identity emotion translation. Our speech-driven mode makes it possible to drive the facial animation through multiple conditions (video, audio, head motion). We collected FEED dataset which, we believe, will be a valuable asset for researchers in diverse human-centered studies.
+
+However, our method has some limitations. It doesn't generate the avatar's body or shoulders, limiting some use cases. We currently integrate our output with a source image body. Additionally, the model sometimes struggles with accurate expression translation and underperforms with extensive head rotation. These challenges are crucial for future enhancements and remain central to our ongoing research efforts.
+
+# References
+
+[1] Emad Barsoum, Cha Zhang, Cristian Canton Ferrer, and Zhengyou Zhang. Training deep networks for facial expression recognition with crowd-sourced label distribution. In Proceedings of the 18th ACM international conference on multimodal interaction, pages 279-283, 2016. 3  
+[2] Stella Bounareli, Christos Tzelepis, Vasileios Argyriou, Ioannis Patras, and Georgios Tzimiropoulos. Hyperreenact: One-shot reenactment via jointly learning to refine and retarget faces, 2023. 2  
+[3] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020. 2  
+[4] Houwei Cao, David G Cooper, Michael K Keutmann, Ruben C Gur, Ani Nenkova, and Ragini Verma. Crema-d: Crowd-sourced emotional multimodal actors dataset. IEEE transactions on affective computing, 5(4):377-390, 2014. 3  
+[5] Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age, 2018. 7  
+[6] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7  
+[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 4  
+[8] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing, 2021. 2  
+[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars, 2023. 2, 4, 5, 7, 8  
+[10] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction, 2020. 2  
+[11] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos, 2022. 2  
+[12] Siddharth Gururani, Arun Mallya, Ting-Chun Wang, Rafael Valle, and Ming-Yu Liu. Space: Speech-driven portrait animation with controllable expression, 2022. 3, 7  
+[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, 2017. 7  
+[14] Philip Jackson and SJUoSG Haq. Surrey audio-visual expressed emotion (savee) database. University of Surrey: Guildford, UK, 2014. 3  
+[15] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars, 2022. 2
+
+[16] Alexander C. Li, Alexei A. Efros, and Deepak Pathak. Understanding collapse in non-contrastive siamese representation learning, 2022. 4  
+[17] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar, 2023. 7, 8  
+[18] Steven R Livingstone and Frank A Russo. The ryerson audiovisual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5):e0196391, 2018. 3  
+[19] Patrick Lucey, Jeffrey F Cohn, Takeo Kanade, Jason Saragih, Zara Ambadar, and Iain Matthews. The extended cohn-kanade dataset (ck+): A complete dataset for action unit and emotion-specified expression. In 2010 iiee computer society conference on computer vision and pattern recognition-workshops, pages 94-101. IEEE, 2010. 3  
+[20] Maja Pantic, Michel Valstar, Ron Rademaker, and Ludo Maat. Web-based database for facial expression analysis. In 2005 IEEE international conference on multimedia and Expo, pages 5-pp. IEEE, 2005. 3  
+[21] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 3  
+[22] Aliaksandr Siarohin, Stéphane Lathuilière, S. Tulyakov, Elisa Ricci, and N. Sebe. First order motion model for image animation. ArXiv, abs/2003.00196, 2019. 2, 7, 8  
+[23] Aliaksandr Siarohin, Willi Menapace, Ivan Skorokhodov, Kyle Olszewski, Hsin-Ying Lee, Jian Ren, Mengelei Chai, and Sergey Tulyakov. Unsupervised volumetric animation. arXiv preprint arXiv:2301.11326, 2023. 7  
+[24] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation, 2023. 3  
+[25] Mani Kumar Tellamekala, Ömer Ümer, Björn W. Schuller, Elisabeth Andre, Timo Giesbrecht, and Michel Valstar. Are 3d face shapes expressive enough for recognising continuous emotions and action unit intensities?, 2023. 2  
+[26] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition, 2018. 5  
+[27] H. Wang, Yitong Wang, Zheng Zhou, Xing Ji, Zhifeng Li, Dihong Gong, Jin Zhou, and Wenyu Liu. Cosface: Large margin cosine loss for deep face recognition. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5265-5274, 2018. 6  
+[28] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700–717. Springer, 2020. 3  
+[29] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2
+
+[30] Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. AvatarMAV: Fast 3d head avatar reconstruction using motion-aware neural voxels. In *Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Proceedings*. ACM, 2023. 2  
+[31] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan, 2022. 2, 3, 7, 8  
+[32] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction, 2023. 2, 7, 8  
+[33] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models, 2019. 2  
+[34] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars, 2020. 2  
+[35] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation, 2023. 2, 7, 8  
+[36] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation, 2023. 3, 7, 8  
+[37] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7  
+[38] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I m avatar: Implicit morphable head avatars from videos, 2022. 2  
+[39] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation, 2021. 2, 3, 7, 8  
+[40] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. MakeltTalk. ACM Transactions on Graphics, 39(6):1-15, 2020. 3, 7, 8

2024/EMOPortraits_ Emotion-enhanced Multimodal One-shot Head Avatars/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c295bbdd840ed661f28e73a7687010051a1f3e519cdd975a1b149536af22ed86
+size 672662

2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/197e9226-8724-4a9f-ac9a-1556459fa452_content_list.json

@@ -0,0 +1,1450 @@
+[
+    {
+        "type": "text",
+        "text": "ERMVP: Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments",
+        "text_level": 1,
+        "bbox": [
+            100,
+            161,
+            868,
+            203
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jingyu Zhang $^{1}$  Kun Yang $^{1}$  Yilei Wang $^{1}$  Hanqi Wang $^{1}$  Peng Sun $^{2,*}$  Liang Song $^{1,*}$ $^{1}$ Academy for Engineering and Technology, Fudan University  $^{2}$ Duke Kunshan University  \n{jingyuzhang22, yileiwang23}@m.fudan.edu.cn, songl@m.fudan.edu.cn",
+        "bbox": [
+            131,
+            226,
+            836,
+            277
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            308,
+            312,
+            324
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Collaborative perception enhances perception performance by enabling autonomous vehicles to exchange complementary information. Despite its potential to revolutionize the mobile industry, challenges in various environments, such as communication bandwidth limitations, localization errors and information aggregation inefficiencies, hinder its implementation in practical applications. In this work, we propose ERMVP, a communication-Efficient and collaboration-Robust Multi-Vehicle Perception method in challenging environments. Specifically, ERMVP has three distinct strengths: i) It utilizes the hierarchical feature sampling strategy to abstract a representative set of feature vectors, using less communication overhead for efficient communication; ii) It employs the sparse consensus features to execute precise spatial location calibrations, effectively mitigating the implications of vehicle localization errors; iii) A pioneering feature fusion and interaction paradigm is introduced to integrate holistic spatial semantics among different vehicles and data sources. To thoroughly validate our method, we conduct extensive experiments on real-world and simulated datasets. The results demonstrate that the proposed ERMVP is significantly superior to the state-of-the-art collaborative perception methods.",
+        "bbox": [
+            75,
+            338,
+            472,
+            657
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            77,
+            684,
+            208,
+            698
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Autonomous vehicles are widely recognized as a valuable means to enhance road safety and traffic efficiency. Equipped with lidar, cameras, and other sensors, these vehicles are capable of accurately sensing their surroundings to ensure safe and reliable operation. However, the single-vehicle perception system has inevitable drawbacks [26, 39], such as a limited sensor field of view that can be easily obstructed and the challenge of detecting distant objects due to sparse and low-resolution data. Recently, the",
+        "bbox": [
+            75,
+            707,
+            468,
+            834
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "advances in vehicle-to-vehicle (V2V) communication technologies [12, 17, 36] and deep learning [4-6, 10, 16, 24, 45] have spurred innovation and progress in the collaborative perception technology. This technology allows connected autonomous vehicles (CAVs) to share sensory data, leading to more comprehensive environmental perception.",
+        "bbox": [
+            499,
+            309,
+            890,
+            394
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Although collaborative perception technology shows great potential in transforming the mobility industry, its practical application faces several challenges, including communication bandwidth limitations [15, 35], localization errors [11, 25] and information aggregation inefficiencies [38, 47]. In practical situations, wireless communication resource and reliability constraints severely hamper the efficacy of delay-sensitive collaborative perception. While recent works [15, 42] have achieved a balance between perception performance and communication bandwidth through well-designed mechanisms, these methods have their limitations because they primarily considering information compression over spatial redundancy. This narrow focus exacerbates performance degradation at high compression ratios.",
+        "bbox": [
+            497,
+            394,
+            892,
+            600
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Further, complex dynamic environments lead to localization errors, which result in inaccurate relative transform estimates and spatial feature misalignment. This relative pose noise produces misleading features that adversely affect the effectiveness of collaborative perception. Existing methods [25, 34] attempt to optimize the overall pose through intensive computation, but the high latency makes them unsuitable for real-time dynamic perception. Meanwhile, collaborative methods [3, 38, 41, 42, 46] only focus on aggregated information, but overlook the inherent perceptual strengths of ego vehicle. This paradigm is vulnerable to the perturbations introduced by collaborative noise, including asynchronous motion blur and inaccurate projections. Such drawback becomes a bottleneck for achieving optimal perceptual performance. In contrast, ego-centric features may contain locally accurate spatial location information that is not affected by collaborative noise. Therefore, a priority for establishing a pragmatic collaborative perception system is to effectively overcome the above challenges.",
+        "bbox": [
+            497,
+            603,
+            894,
+            866
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            127,
+            49,
+            200,
+            85
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            252,
+            48,
+            791,
+            87
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding authors. Our code is available at https://github.com/Terry9a/ERMVP.",
+        "bbox": [
+            75,
+            842,
+            468,
+            866
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "12575",
+        "bbox": [
+            465,
+            887,
+            505,
+            897
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/17285042ad72afee503d390d9bf80d325dc5985f785ca8f0682e6d8e69e83e0c.jpg",
+        "image_caption": [
+            "Figure 1. The overall architecture of the proposed framework. The framework consists of six phases: metadata sharing and feature extraction, filter and merge feature sampling, feature spatial calibration, attention-based feature fusion, accuracy enhanced feature interaction and detection decoders. The details of each individual component are illustrated in Section 3."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            77,
+            122,
+            890,
+            371
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Based on these observations, we propose ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. From Figure 1, ERMVP uses four proposed core components to tackle existing challenges jointly. Specifically, (i) we first design an advanced filter and merge feature sampling strategy to tackle the limitations of wireless communication resources. This strategy considers both inter-class and intra-class redundancy relationships to abstract a refined set of feature vectors from the redundant features, using less communication overhead for efficient communication. (ii) Second, we introduce a plug-and-play feature spatial calibration module to mitigate the implications of vehicle localization errors. This module ingeniously utilizes consensus sparse foreground features to align the relative pose relationships between ego-vehicle and collaborators without any precise pose supervision. (iii) Furthermore, we present a pioneering feature fusion and interaction paradigm to integrate holistic spatial semantics. This paradigm comprises two key components: The first is an attention-based feature fusion module that alternates between local and global attention to fuse heterogeneous information from different vehicles. The second is an accuracy enhanced feature interaction strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by fused features. Through these tailored components, ERMVP represents a significant advancement towards pragmatic collaborative perception. To validate the effectiveness of the ER-MVP, we conduct extensive experiments on two collaborative 3D object detection datasets including V2V4Real [44]",
+        "bbox": [
+            75,
+            435,
+            472,
+            865
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "and OPV2V [43]. Comprehensive experimental results demonstrate that our method outperforms previous state-of-the-art methods under the bandwidth-limited noisy setting. The main contributions can be summarized as follows:",
+        "bbox": [
+            497,
+            435,
+            890,
+            491
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method, which addresses the communication bandwidth limitations, localization errors and information aggregation efficiency challenges.",
+            "- We develop a filter and merge feature sampling strategy to enhance communication efficiency, a feature spatial calibration module for accurate spatial feature alignment, and two information aggregation components to optimize the fusion process.",
+            "- We conduct extensive experiments on both real-world and simulated datasets. The results show the superiority of our method and the necessity of the proposed components."
+        ],
+        "bbox": [
+            499,
+            491,
+            890,
+            682
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            499,
+            697,
+            640,
+            711
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Multi-Agent Communication",
+        "text_level": 1,
+        "bbox": [
+            499,
+            720,
+            757,
+            733
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Communication has played a pivotal role in the development of robust multi-agent systems. Early multi-agent communication [20, 29, 32] relied on predefined protocols and heuristics to regulate communication between agents. However, these fixed methods are unsuitable in complex and dynamic settings. Advances in deep learning have inspired the development of advanced strategies for information exchange and collaboration among agents. For example, MAGIC [27] used a graph attention encoder to solve the",
+        "bbox": [
+            497,
+            741,
+            890,
+            866
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "12576",
+        "bbox": [
+            465,
+            887,
+            505,
+            897
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "problem of when and to whom to send messages. TMC [49] utilized temporal smoothing to ensure efficient communication. EC-MARL [2] tackled high-dimensional continuous control and partially observable states by introducing a rapid communication protocol to resolve task dilemmas. In comparison, our research focuses on LiDAR-based collaborative 3D object detection tasks in complex driving scenarios. We propose a hierarchical feature sampling strategy to achieve efficient communication across agents.",
+        "bbox": [
+            75,
+            125,
+            472,
+            251
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2. Collaborative Perception",
+        "text_level": 1,
+        "bbox": [
+            77,
+            261,
+            305,
+            275
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Factors including limited fields of view of sensors and physical obstructions in the environment can adversely impact the perception capabilities of individual agents. To address these challenges, collaborative perception within multi-agent systems has become a key technology. Utilizing newly available datasets [22, 43, 44], several novel collaborative perception methods have been proposed. For example, V2VNet [38] incorporated a graph neural network to fuse data from different agents. V2X-ViT [42] introduced a transformer architecture that combined information from vehicles and infrastructures. Where2comm [15] used feature spatial heterogeneity to reduce bandwidth utilization by transmitting sparse feature maps. CoBEVT [41] proposed the first multi-camera-based collaborative perception framework and designed the fused axis attention module to enable multi-view interactions. In this work, we propose an innovative feature calibration and interaction method that ensures robust vehicle collaboration in high-noise dynamic environments.",
+        "bbox": [
+            75,
+            282,
+            472,
+            545
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            77,
+            560,
+            169,
+            574
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This section introduces our efficient and robust collaborative perception method. Figure 1 shows the general procedure of our method, which is divided into six phases. We will detail each phase in the following introduction.",
+        "bbox": [
+            75,
+            584,
+            468,
+            641
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Metadata Sharing and Feature Extraction",
+        "text_level": 1,
+        "bbox": [
+            75,
+            650,
+            436,
+            666
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In multi-vehicle collaboration scenarios, one of the CAVs, known as the ego vehicle, constructs a communication graph. In this graph, the ego vehicle serves as the requester, while other vehicles within its communication range act as supporters. The ego vehicle broadcasts its metadata information, including position, heading, speed, and more. Supporters, upon receiving this metadata, project their local point cloud observations into the ego vehicle's coordinate system. Each vehicle then encodes these transformed point clouds into bird's eye view (BEV) features, yielding a visual representation. Given the  $i$ -th vehicle local observations  $\\mathbf{X}_i$ , the extracted features are represented as  $\\mathbf{F}_i = \\Phi_{\\mathrm{enc}}(\\mathbf{X}_i) \\in \\mathbb{R}^{H \\times W \\times C}$ , where  $\\Phi_{\\mathrm{enc}}(\\cdot)$  denotes the PointPillar [19] encoder shared by all vehicles and  $H$ ,  $W$ ,",
+        "bbox": [
+            75,
+            672,
+            472,
+            866
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "and  $C$  stand for the height, width, and channel of the feature map, respectively. Then, these extracted features are fed into the filter and merge feature sampling module.",
+        "bbox": [
+            497,
+            125,
+            890,
+            168
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Filter and Merge Feature Sampling",
+        "text_level": 1,
+        "bbox": [
+            497,
+            174,
+            808,
+            190
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Previous works have utilized well-designed mechanisms such as information entropy communication selection [37] and spatial heterogeneity map [15, 35] to reduce the required transmission bandwidth. However, these methods primarily focus on the inter-class redundancy between foreground and background features, ignoring the intra-class redundancy among features, which results in sub-optimal compression. To address this gap, we introduce an advanced Filtering and Merging Feature Sampling strategy (FMS). This strategy considers both inter-class and intra-class redundancy relationships, efficiently extracting a concise and distinctive set of feature vectors from the original feature maps, thus reducing communication overheads more effectively. FMS is composed of two core components as follows.",
+        "bbox": [
+            497,
+            195,
+            890,
+            402
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Filter Sampler. In object detection, foreground areas containing objects are more significant than the background areas. Therefore, we implement the idea of reducing spatial redundancy into a feature filter sampler module, aiming to preserve perceptually important yet sparse sets of feature vectors. Since explicitly learning a binary sampler is infeasible, we develop a confidence filter strategy. Initially, a detection confidence map is generated for the feature map. It reflects the perceptual importance of different spatial areas, with higher levels indicating potential object areas and lower levels typically denoting redundant background areas. For the feature map  $\\pmb{F}_i$  of the  $i$ -th vehicle, its confidence map  $C_i$  is defined as:",
+        "bbox": [
+            497,
+            403,
+            892,
+            583
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\boldsymbol {C} _ {i} = \\Phi_ {\\text {c o n - g e n}} \\left(\\boldsymbol {F} _ {i}\\right) \\in [ 0, 1 ] ^ {H \\times W}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            581,
+            591,
+            890,
+            609
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\Phi_{\\mathrm{con\\_gen}(\\cdot)}$  represents the confidence generation network with detection decoder structure. Then the confidence map is thresholded, followed by non-maximum suppression, resulting in a binary mask  $B$ . Utilizing this binary mask, we proceed to preserve the sparse foreground features  $\\widetilde{\\pmb{F}}_i = B\\odot \\pmb{F}_i$ . In order to cope with the dynamically changing environmental conditions and fully ensure the robustness of the system, the filtering threshold can be varied with the sensor data and the network state. We set  $\\alpha$  as the filter rate and the number of remaining feature vectors is  $L = \\alpha \\times HW$ .",
+        "bbox": [
+            497,
+            617,
+            890,
+            767
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Merge Sampler. Upon extracting a detailed foreground feature vector set with the filter sampler, we employ the merge sampler for additional optimization, refining similar or repetitive foreground feature vectors through weighted merging. The process is divided into three stages: information-driven feature grouping, attention-inspired feature merging, and index-based feature reconstruction.",
+        "bbox": [
+            497,
+            769,
+            890,
+            866
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "12577",
+        "bbox": [
+            465,
+            887,
+            505,
+            897
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "(a) Information-Driven Feature Grouping. Initially, a variant of the nearest neighbor clustering algorithm is applied to group the foreground feature vector set. Given a set of feature vector  $\\widetilde{\\boldsymbol{F}}_i = [x_1,x_2,\\dots ,x_L]^\\top$  and cluster center  $X_{c}$ , we compute the indicator  $\\delta_{i}$  for each feature vector.  $\\delta_{i}$  is calculated as the minimum feature distance minus the average pixel distance to any other cluster center vector, expressed as:",
+        "bbox": [
+            75,
+            125,
+            472,
+            238
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\delta_ {i} = \\min  _ {j: x _ {j} \\in X _ {c}} \\left(\\| x _ {i} - x _ {j} \\| _ {2} ^ {2} - \\gamma \\| p (x _ {i}), p (x _ {j}) \\| _ {2} ^ {2}\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            99,
+            248,
+            468,
+            273
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\delta_{i}$  denotes to which cluster the feature vector  $x_{i}$  should belong.  $p(\\cdot)$  means getting the position of the vector and  $\\gamma$  is a hyperparameter. Subsequently, we can divide all the feature vectors in  $\\widetilde{\\boldsymbol{F}}_i$  into  $K$  clusters represented by  $G = \\{G_{1},G_{2},\\dots ,G_{K}\\}$ . The total number of clusters is determined by dynamic clustering ratio  $\\beta$ , and is calculated as  $K = \\beta \\times L$ .",
+        "bbox": [
+            75,
+            283,
+            468,
+            380
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "(b) Attention-Inspired Feature Merging. A straightforward strategy for merging feature vectors is to average of each feature vector within a cluster. However, this scheme can be severely affected by outlier feature vectors. Drawing inspiration from attention mechanisms, we utilize the confidence score as a guide to quantify the significance of each feature. Thus, the merged feature vector  $\\tilde{x}_i$  for the  $i$ th cluster  $G_i$  is computed as:",
+        "bbox": [
+            75,
+            381,
+            468,
+            491
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\widetilde {x} _ {i} = \\frac {\\sum_ {j \\in G _ {i}} c _ {j} x _ {j}}{\\sum_ {j \\in G _ {i}} c _ {j}}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            205,
+            501,
+            468,
+            536
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $c_{j}$  and  $x_{j}$  represent the confidence score and the original feature vector, respectively. In the end, we obtain a final set of feature vectors  $\\mathbf{Z}_i = \\{\\widetilde{x}_1,\\widetilde{x}_2,\\dots ,\\widetilde{x}_K\\}$  that is necessary for transmission.",
+        "bbox": [
+            75,
+            545,
+            468,
+            600
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "(c) Index-Based Feature Reconstruction. During the feature grouping and merging process, each feature vector is allocated to a cluster, and every cluster is represented by a merged vector. We maintain a record of the index correspondence between original and merged feature vectors. Utilizing this index record, the ego-vehicle ensures that the merged feature vectors are mapped to their corresponding positions, leading to the reconstruction of feature maps.",
+        "bbox": [
+            75,
+            600,
+            468,
+            712
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Feature Spatial Calibration",
+        "text_level": 1,
+        "bbox": [
+            77,
+            720,
+            324,
+            735
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Localization errors [30, 31] may lead to the misalignment of feature maps among vehicles. Such misalignment causes the ego-vehicle to misinterpret the object's location, resulting in sub-optimal perception, as depicted in the upper right corner of Figure 2. To address this challenge, we introduce the Feature Spatial Calibration module (FSC) to facilitate precise feature alignment, as shown in Figure 2. The core idea of FSC is that overlapping viewpoints allow multiple vehicles to detect partially the same objects. The process",
+        "bbox": [
+            75,
+            741,
+            468,
+            866
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/7d58f4ddb6a0733b531a8909f7caf793b1c101be3a55015892f12e2ce2992a07.jpg",
+        "image_caption": [
+            "Figure 2. Illustration of the proposed FSC. By employing FSC, misaligned matching regions are corrected at the feature level, leading to more accurate predictions."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            502,
+            125,
+            890,
+            252
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "involves three phases: consensus matching, geometric verification, and error modulation.",
+        "bbox": [
+            497,
+            312,
+            890,
+            339
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Consensus Matching. After reconstructing the received features, the ego vehicle generates sparse feature maps. Within these maps, areas rich in information signify potential target regions, which correspond to the proposed matching areas. We denote the proposed matching regions of the ego vehicle as  $P$  and those identified by the collaborative vehicle under noisy pose conditions as  $Q$ . Utilizing  $P$  and  $Q$ , a weighted bipartite graph is constructed, wherein the weight of each edge is determined by the distance between the nodes, encapsulated in a cost matrix. The matching process is then converted into a linear assignment task, with the objective of identifying a matching result with the lowest cumulative edge weight. This procedure yields the matching pairs  $M$ .",
+        "bbox": [
+            497,
+            340,
+            890,
+            533
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Geometric Verification. Invalid matches may occur due to objects located in exclusive zones and detection noise. To tackle this issue, we utilize RANSAC to filter and sift a consistent set of matches aligned with expected geometric transformations. Initially, a random matching subset  $M_{s} \\subseteq M$  is selected, and then the transformation matrix  $\\Gamma_{s}$  is computed using singular value decomposition. When  $\\Gamma_{s}$  is applied to all pairs within  $M_{s}$ , and if the posttransformation distance between these pairs remains below the threshold  $\\eta$ , the set is considered correctly aligned. The threshold  $\\eta$  reflects the allowable localization error within the original collaborative framework. This process is iteratively conducted to identify the optimal transformation matrix that correlates with the maximum number of correct matches. Ultimately, a optimal refined transformation matrix  $\\Gamma_{r}$  is obtained and applied in subsequent spatial calibration operations, yielding the aligned feature  $\\widetilde{Z}_{j} = \\Gamma_{r}Z_{j}$ .",
+        "bbox": [
+            497,
+            534,
+            890,
+            769
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Error Modulation. To enhance the adaptability of the calibration method in various environments, we incorporate an error modulation strategy. This strategy aims to achieve a balance between localization errors and the estimation errors that emerge from the calibration process. It measures the overlap ratios between ego and collaborative features in both their adjusted and original states. Subsequently, fea",
+        "bbox": [
+            497,
+            770,
+            890,
+            866
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "12578",
+        "bbox": [
+            465,
+            887,
+            504,
+            897
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/d42d10cfc3c0721cb0a1f26165bbe6d235184a7bc9b1bd0e8df2217e6ae379c9.jpg",
+        "image_caption": [
+            "Figure 3. A example of local and global Attention in AFF. This shows how AFF processes both 3D local windows and the global query token for position-level and context-aware aggregation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            82,
+            122,
+            463,
+            252
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "tures demonstrating the most optimal overlap are selected as the final input.",
+        "bbox": [
+            75,
+            318,
+            468,
+            347
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4. Attention-based Feature Fusion",
+        "text_level": 1,
+        "bbox": [
+            75,
+            353,
+            356,
+            368
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In multi-vehicle collaboration scenarios, vehicles are able to capture heterogeneous information from different spatial regions. In order to efficiently fuse perceptual features from multiple vehicles, we propose an Attention-based Feature Fusion method (AFF). As shown in Figure 3, AFF utilizes alternating local and global attention to achieve position-level accurate matching in occlusion-variable traffic scenes and to capture the global semantics attention of road topology and traffic states. Specifically, we stack all the vehicles' features to  $S \\in \\mathbb{R}^{N \\times H \\times W \\times C}$ , where  $N$  is the number of vehicles. The features are divided into 3D non-overlapping windows, each of size  $N \\times P \\times P$ . The shape of the partition tensor is  $\\left(\\frac{H}{P} \\times \\frac{W}{P}, N \\times P^2, C\\right)$ . We perform 3D local self attention on the local window and implement region interaction within the window. In contrast, global attention goes beyond the limitations of the local view. It uses the extracted global query token and is shared between all windows to interact with the local key and value representations, thus helping to capture long-range dependencies of the features. Inspired by the efficiency of [13, 33], the global query token  $g$  is generated as follows:",
+        "bbox": [
+            75,
+            375,
+            470,
+            667
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {g} = \\mathcal {P} \\left(\\mathcal {C} _ {1 \\times 1} \\left(\\mathcal {S E} (\\mathcal {G} (\\mathcal {D} (\\mathbf {x}))) + \\mathbf {x}\\right)\\right), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            149,
+            675,
+            468,
+            692
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{D}$ ,  $\\mathcal{G}$ ,  $\\mathcal{SE}$ ,  $\\mathcal{C}_{1\\times 1}$  and  $\\mathcal{P}$  denote the depth-wise separable convolution, GELU activation function, Squeeze-and-Excitation operation,  $1\\times 1$  convolution and max pooling, respectively. These designs enable the global query token to provide advantages such as inductive bias and modeling of inter-channel dependencies. We then construct our proposed AFF module by combining this local and global attention with typical designs of Transformers, including LayerNorm [1], MLPs [9], and skip-connections [14]. This module allows the system to analyze spatial correlations from a local view while also capturing global feature responses, ensuring efficient and precise perception in dy",
+        "bbox": [
+            75,
+            700,
+            468,
+            866
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/1f09b47ad9f51cbbee866c451bf65bbf30029ce7fdb5145e4bb85834b8ec315b.jpg",
+        "image_caption": [
+            "Figure 4. The architecture of the proposed AFI component."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            124,
+            884,
+            254
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "namic, complex, and occlusion-variable traffic scenarios. Finally, we obtain the fused features  $\\pmb{H}_i$ .",
+        "bbox": [
+            497,
+            306,
+            889,
+            333
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.5. Accuracy Enhanced Feature Interaction",
+        "text_level": 1,
+        "bbox": [
+            499,
+            350,
+            843,
+            365
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Previous works [15, 21, 38, 42, 43] have demonstrated that fused features can provide richer semantic information, thereby enhancing perceptual performance. However, they may be affected by collaborative noise, such as asynchronous motion blur and inaccurate projections, which can compromise accurate position information and become a bottleneck to the optimal realization of perceptual performance. Ego-centric features may contain locally critical spatial location information without being affected by collaborative noise. To this end, we propose a novel Accuracy Enhanced Feature Interaction (AEI) strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by collaborative fused features. Firstly, we design a context cross attention module that is tailored for feature interaction, as shown in Figure 4. It considers the features from the ego-vehicle as queries, and obtains the context of the query location from the fused features as the key and value to achieve cross-attention. The computed output of the spatial location  $(i,j)$  is as follows:",
+        "bbox": [
+            497,
+            372,
+            890,
+            650
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ny _ {i j} = \\sum_ {a, b \\in \\mathcal {N} _ {k} (i, j)} \\operatorname {s o f t m a x} \\left(q _ {i j} ^ {\\top} k _ {a b}\\right) v _ {a b}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            563,
+            665,
+            890,
+            698
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{N}_k(i,j)$  is context position of  $(i,j)$  and  $q_{ij}, k_{ab}, v_{ab}$  represent the ego query, context key and value, respectively. As the relevant features of the same object are often located in similar locations across different feature maps, we prioritize the local context details of the query location. Additionally, considering that only a few positions within the sparse feature map hold significant information, focusing on these context-rich areas is computationally efficient. Following this, we employ the split attention [48] to adaptively fuse information from multiple branches, generating the enhanced output feature  $\\widetilde{H}_i$ .",
+        "bbox": [
+            497,
+            712,
+            892,
+            866
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "12579",
+        "bbox": [
+            465,
+            887,
+            505,
+            897
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/1de46cb1d7557a613601596d0c8251c9c9eee305c37b4dae042ecda1c4e0817a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td colspan=\"3\">V2V4Real</td><td colspan=\"3\">OPV2V</td></tr><tr><td>Noise Level σt/σr (m/°)</td><td>0.0/0.0</td><td>0.4/0.4</td><td>0.8/0.8</td><td>0.0/0.0</td><td>0.4/0.4</td><td>0.8/0.8</td></tr><tr><td>No Collaboration</td><td>41.64/23.43</td><td>41.64/23.43</td><td>41.64/23.43</td><td>73.25/58.22</td><td>73.25/58.22</td><td>73.25/58.22</td></tr><tr><td>Late Fusion</td><td>56.80/27.62</td><td>40.81/13.31</td><td>26.65/10.66</td><td>86.46/79.21</td><td>76.56/40.01</td><td>45.57/21.00</td></tr><tr><td>F-Cooper [3]</td><td>61.12/32.36</td><td>51.71/24.48</td><td>41.90/19.91</td><td>87.39/79.39</td><td>79.33/40.12</td><td>52.91/16.90</td></tr><tr><td>AttFuse [7]</td><td>64.41/34.32</td><td>58.42/30.02</td><td>49.37/25.44</td><td>90.50/81.80</td><td>83.54/52.43</td><td>68.37/40.96</td></tr><tr><td>Where2comm [15]</td><td>63.69/34.79</td><td>58.48/28.09</td><td>49.77/24.25</td><td>90.46/84.22</td><td>82.17/56.47</td><td>73.95/45.86</td></tr><tr><td>DiscoNet [21]</td><td>64.19/35.25</td><td>57.71/28.41</td><td>49.19/24.94</td><td>89.58/81.45</td><td>83.25/53.22</td><td>61.37/30.53</td></tr><tr><td>V2VNet [38]</td><td>65.70/35.38</td><td>62.11/30.63</td><td>54.73/25.57</td><td>91.35/82.43</td><td>85.43/54.16</td><td>70.43/31.28</td></tr><tr><td>CoAlign [25]</td><td>64.06/36.60</td><td>59.58/31.52</td><td>53.02/26.60</td><td>90.93/84.28</td><td>84.48/57.45</td><td>73.95/48.86</td></tr><tr><td>V2X-ViT [42]</td><td>66.42/37.34</td><td>63.33/32.38</td><td>57.15/28.22</td><td>91.74/83.31</td><td>82.91/54.73</td><td>61.70/26.07</td></tr><tr><td>CoBEVT [41]</td><td>66.01/37.36</td><td>58.63/29.09</td><td>49.22/23.66</td><td>91.71/85.98</td><td>85.49/60.64</td><td>64.63/29.59</td></tr><tr><td>ERMVP (Ours)</td><td>67.66/42.97</td><td>65.01/40.45</td><td>60.88/35.01</td><td>92.18/85.59</td><td>85.67/66.32</td><td>77.13/59.15</td></tr></table>",
+        "bbox": [
+            78,
+            122,
+            892,
+            309
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Overall performance on V2V4Real and OPV2V datasets with pose noises. The results are reported in AP@0.5/0.7.",
+        "bbox": [
+            117,
+            319,
+            848,
+            332
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.6. Decoder and Loss",
+        "text_level": 1,
+        "bbox": [
+            77,
+            339,
+            250,
+            353
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Based on the final fused feature  $\\widetilde{H}_i$ , we use the detection decoder to generate the final prediction output  $Y_{i} = \\Phi_{\\mathrm{dec}}(\\widetilde{H}_{i})$ . Each position of  $Y_{i}$  represents a rotated box with classes (c, x, y, h, w,  $\\cos \\alpha$ ,  $\\sin \\alpha$ ), corresponding to class confidence, position, size, and angle. These objects are the final output of the proposed collaborative perception system. Following existing work [19], we adopt the smooth  $L_{1}$  loss for regression and focal loss [23] for classification.",
+        "bbox": [
+            75,
+            362,
+            468,
+            476
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            77,
+            496,
+            208,
+            511
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Datasets and Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            77,
+            521,
+            363,
+            535
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Datasets. We validate the proposed ERMVP in the task of LiDAR-based 3D object detection on two benchmark datasets including V2V4real [44] and OPV2V [43]. V2V4real [44] is the first large-scale real-world dataset on V2V perception, collected by two vehicles equipped with multi-modal sensors driving together in a variety of scenarios. It covers a driving area of 410 kilometers, including 20K frames of point clouds, with training/validation/test sets containing 14210/2000/3986 frames, respectively. OPV2V [43] is a large-scale vehicle-to-vehicle collaborative perception dataset, simulated by OpenCDA [40] and Carla [8]. It contains 73 different scenarios, in which different numbers (2 to 7) of collaborative vehicles appear together, each equipped with a lidar sensor and 4 cameras. The dataset contains a total of 11,464 frames of point clouds and RGB images. The training/validation/test sets contain 6374, 1980, and 2170 frames, respectively.",
+        "bbox": [
+            75,
+            545,
+            468,
+            793
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics. We adopt the Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 and 0.7 to evaluate the detection performance. The calculation format of communication volume in [15] is used to count the message size by byte in the log scale with base 2.",
+        "bbox": [
+            75,
+            796,
+            468,
+            866
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            499,
+            339,
+            717,
+            353
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We implement the proposed ERMVP and comparison model on the Pytorch toolbox [28], and train them on a NVIDIA GeForceRTX 3090 GPU using the Adam optimizer [18]. The initial learning rate is  $2\\mathrm{e} - 3$  and it decays every 15 epochs with a factor of 0.1. All models are trained with 60 epochs and the batch size is set to 2. Early stopping is used to find the best epoch. We also add normal point cloud data augmentation for all experiments, including scaling, rotation, and flipping. All detection models are based on PointPillar [19] backbone to extract 2D features from point clouds and a  $0.4\\mathrm{m}$  width/length is used for each voxel. AFF component has 3 encoded layers and a window size of 8 for both local and global attention. The balance hyperparameter  $\\gamma$  is 0.1 and the error tolerance threshold  $\\eta$  is 0.25. Each vehicle has a communication range of 70 m based on [42], while vehicles outside of this broadcasting radius will be ignored. To simulate the localization and heading errors, we add Gaussian noise with a standard deviation of  $\\sigma_t$  for localization errors and  $\\sigma_r$  for heading errors.",
+        "bbox": [
+            497,
+            362,
+            890,
+            625
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3.Quantitative Evaluation",
+        "text_level": 1,
+        "bbox": [
+            499,
+            636,
+            721,
+            650
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Comparison of Detection Performance. Table 1 shows the performance comparison results of 3D detection on two datasets. We use the No Collaboration method as a baseline, which only uses ego vehicle's point clouds data without collaboration. Late Fusion allows vehicles to exchange detected outputs and utilizes non-maximum suppression to produce the final result. For intermediate fusion strategies, we evaluate the existing state-of-the-art (SOTA) methods: AttFuse [43], F-Cooper [3], V2VNet [38], DiscoNet [21], V2X-ViT [42], CoAlign (only fusion) [25], CoBEVT [41] and Where2comm [15]. The proposed ER-MVP achieves average improvements of  $12.48\\% / 18.47\\%$  and  $6.18\\% / 10.02\\%$  on two datasets in AP@0.5/0.7 compared with No Collaboration and Late Fusion methods, demonstrating the superiority of our collaborative method.",
+        "bbox": [
+            497,
+            657,
+            890,
+            866
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "12580",
+        "bbox": [
+            465,
+            887,
+            505,
+            897
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/a05276fe44b9498839d47f184cc6c82d59fe0ebf90f874dbf5bd3a548c86c415.jpg",
+        "image_caption": [
+            "(a) V2V4Real"
+        ],
+        "image_footnote": [
+            "Late Fusion FCooper V2VNet V2X-VIT CoBEVT -Where2comm -ERMVP"
+        ],
+        "bbox": [
+            80,
+            122,
+            282,
+            251
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/0d727aed98d8c717ad512713bc262dc41c0e12fc4ed41f3bfc943faddb1c8b02.jpg",
+        "image_caption": [
+            "Figure 5. Collaborative perception performance comparison on the V2V4Real and OPV2V with varying communication volume."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            282,
+            122,
+            482,
+            250
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f88bea390d557ccfa8148f6e1357029b30be1ff3a78f8f3b8d9d1900be3dafa5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            482,
+            122,
+            682,
+            250
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/517d86dc9d3835c14d405265d54458bd70a95d2611f426bab979bfa02ba7bde2.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            682,
+            122,
+            880,
+            250
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Meanwhile, it is superior to the SOTA collaborative perception method in both real-world and simulated scenarios: improves the SOTA collaborative performance by  $1.24\\%$  and  $5.61\\%$  on V2V4Real datasets in AP@0.5/0.7, and by  $0.44\\%$  on OPV2V datasets in AP@0.5. These results fully demonstrate the superiority of the ERMVP collaboration paradigm.",
+        "bbox": [
+            75,
+            321,
+            468,
+            419
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/aaa3e111652a9c67b6d7667c976a0c1cf556ed7598929d5b2d4b2facb1d4c5f2.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td colspan=\"2\">V2V4Real</td><td colspan=\"2\">OPV2V</td></tr><tr><td>Noise Level σt(m)</td><td>0.5</td><td>1.0</td><td>0.5</td><td>1.0</td></tr><tr><td>Late Fusion</td><td>11.02</td><td>10.13</td><td>53.88</td><td>38.80</td></tr><tr><td>F-Cooper [3]</td><td>24.60</td><td>18.41</td><td>32.24</td><td>12.71</td></tr><tr><td>AttFuse [43]</td><td>28.10</td><td>24.36</td><td>42.95</td><td>35.53</td></tr><tr><td>DiscoNet [21]</td><td>28.36</td><td>24.05</td><td>45.36</td><td>24.96</td></tr><tr><td>V2VNet [38]</td><td>29.57</td><td>23.67</td><td>47.48</td><td>25.82</td></tr><tr><td>Where2comm [15]</td><td>28.02</td><td>23.53</td><td>46.59</td><td>45.59</td></tr><tr><td>CoAlign [21]</td><td>29.84</td><td>25.82</td><td>49.09</td><td>46.66</td></tr><tr><td>V2X-ViT [42]</td><td>30.99</td><td>23.67</td><td>44.83</td><td>21.77</td></tr><tr><td>CoBEVT [41]</td><td>24.67</td><td>22.01</td><td>41.91</td><td>26.05</td></tr><tr><td>ERMVP (Ours)</td><td>39.26</td><td>33.94</td><td>77.25</td><td>76.09</td></tr></table>",
+        "bbox": [
+            78,
+            429,
+            455,
+            600
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Robustness to Localization and Heading Errors. To evaluate the sensitivity of existing methods to localization and heading errors, we use the same noise setting as [42] and conduct extensive experiments on two datasets. As shown in Tables 1 and 2, under ideal settings, Late Fusion and some advanced intermediate fusion methods can detect object vehicles with high accuracy, but their detection accuracy rapidly decreases as the standard deviation of errors increases. For example, when the localization and heading errors of OPV2V dataset are over  $0.4\\mathrm{m}$  and  $0.4^{\\circ}$ , the Disconet [21] and F-Cooper [3] methods fail and perform even worse than No Collaboration. Our method exceeds previous SOTA models at all noise levels and consistently outperforms the No Collaboration baseline, clearly demonstrating the robustness of ERMVP to pose errors. It is noteworthy that for localization errors, ERMVP can maintain consistent",
+        "bbox": [
+            75,
+            644,
+            470,
+            865
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "accuracy levels in high noise environments, showing superior robustness. When the localization error is  $1.0\\mathrm{m}$ , it outperforms the second-best method by  $29.43\\%$  and  $8.08\\%$  in AP@0.7 on OPV2V and V2V4Real datasets. The reasonable explanations are: (i) feature spatial calibration aligns mismatched collaborative features; (ii) accuracy enhanced feature interaction reduces performance degradation caused by collaborative noise.",
+        "bbox": [
+            497,
+            321,
+            890,
+            431
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Comparison of Communication Volume. To achieve pragmatic collaborative perception, it is crucial to evaluate the perceptual performance under different communication volumes. Figure 5 shows the results of the performance comparison under different bandwidth consumption conditions. Note that we do not consider any additional model/data/feature compression for a fair comparison. The results indicate that ERMVP achieves excellent perceptual performance and bandwidth consumption trade-off under all communication bandwidth conditions, consistently outperforming Where2comm [15]. At the same time, it achieves the same detection performance as the previous model [3, 38, 41, 42] with less communication volume on both real-world and simulated datasets.",
+        "bbox": [
+            497,
+            432,
+            892,
+            625
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4. Ablation Studies",
+        "text_level": 1,
+        "bbox": [
+            499,
+            631,
+            663,
+            644
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/69bd4eb33a169fcd8834f564b380acc1a803aa83bd6f57a98cc16dd8763cdf52.jpg",
+        "table_caption": [
+            "Table 2. Detection AP@0.7 on V2V4Real and OPV2V datasets with localization error."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>AFF</td><td>FSC</td><td>AEI</td><td>V2V4Real</td><td>OPV2V</td></tr><tr><td></td><td></td><td></td><td>51.71/24.48</td><td>79.33/40.12</td></tr><tr><td>✓</td><td></td><td></td><td>62.24/37.70</td><td>84.30/61.43</td></tr><tr><td>✓</td><td>✓</td><td></td><td>64.36/39.59</td><td>84.98/62.67</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>65.01/40.45</td><td>85.67/66.32</td></tr></table>",
+        "bbox": [
+            500,
+            653,
+            885,
+            726
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Ablation study results of the proposed core components on the both datasets with noise level of 0.4/0.4. AFF: Attention-based Feature Fusion; FSC: Feature Spatial Calibration; AEI: Accuracy Enhanced Feature Interaction.",
+        "bbox": [
+            497,
+            736,
+            890,
+            788
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Effect of Core Components. Table 3 details the contribution of each core components in our ERMVP framework. The base model is the naive position-wise maximum fusion method. We then assess the impact of each component by sequentially introducing i) attention-based feature",
+        "bbox": [
+            497,
+            796,
+            890,
+            866
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "12581",
+        "bbox": [
+            465,
+            887,
+            504,
+            897
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/280d2b5c3bb6552ed83f6681ca5e9ab65f17fefd5ae8bbd0b8c2945e3095852f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            87,
+            124,
+            284,
+            210
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e770195a486e1bf921acc083d6a82eecd9c532e4557f8586c9120e99f4b1c3dd.jpg",
+        "image_caption": [
+            "(a) V2VNet"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            87,
+            212,
+            282,
+            296
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/efd2f53dc9a564fb9dde0a5412ab504de41b34a537d009f58505a6de2f4afd40.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            285,
+            124,
+            482,
+            210
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/492842c4ba1259fbc06256306e088f9a970ff6743f2cbd8d52a3c6f44fd3ba2c.jpg",
+        "image_caption": [
+            "(b) CoBEVT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            285,
+            212,
+            482,
+            295
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/aefb2e149c7c44ff1a1486876c71877e2004bf57de183a250fdfb7e67f496884.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            484,
+            124,
+            680,
+            210
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/f4481015af92b97fb51c9f6721b939b1b571c6f3d179fb796d81053fc607f9bc.jpg",
+        "image_caption": [
+            "(c) V2X-ViT",
+            "Figure 6. Qualitative comparison results in real-world scenarios from the V2V4Real dataset with noise level of 0.4/0.4. Green and red boxes denote the ground truths and detection results, respectively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            485,
+            212,
+            680,
+            295
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/bba47526b06f4a957dc1c460977993b17132c871bd63b9b4bf3549c7e3f00d7f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            682,
+            124,
+            882,
+            210
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/cdb2cc3bba2d1bd3841d0f2755d0b7379f1b9c64db023463ded86abc01c94515.jpg",
+        "image_caption": [
+            "(d)ERMVP"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            684,
+            212,
+            880,
+            295
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "fusion (AFF), ii) feature spatial calibration (FSC), iii) accuracy enhanced feature interaction (AEF). The consistent rise in detection results over both datasets demonstrates the effectiveness of each introduced component. Notably, integrating all three components boosts detection performance by  $13.3\\%$  and  $15.97\\%$  on the V2V4Real dataset for AP@0.5 and AP@0.7, respectively.",
+        "bbox": [
+            75,
+            355,
+            468,
+            454
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b92763145d0f8cd9c12d9dd5457a2982c98ae21f38860f7e31289167e94214e1.jpg",
+        "image_caption": [
+            "Figure 7. Hyperparameter analysis of  $\\alpha$  and  $\\beta$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            89,
+            463,
+            267,
+            572
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/d71922f3b4b2034784cf84180fa4aae8983c8e9ae7ee30960da71fb8f0ab96ef.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            280,
+            463,
+            458,
+            572
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Hyperparameters  $\\alpha$  and  $\\beta$ . We explore the impact of the hyperparameters  $\\alpha$  and  $\\beta$  (as described in Section 3.2), which jointly determine the communication volume. Specifically, the feature compression ratio is represented as  $\\alpha \\times \\beta$ . The communication volume, in turn, is expressed as the product of the original feature size and the compression rate. By modulating the values of  $\\alpha$  and  $\\beta$  within the interval of  $10\\%$  to  $80\\%$ , we chart the resultant object detection performance (measured by AP@0.7) in Figure 7. As expected, a drop in either  $\\alpha$  or  $\\beta$  will impair ERMVP's efficacy. Yet, it's worth noting that ERMVP's performance generally exhibits robustness to these hyperparameters. The performance decline remains fairly minimal even under significant compression, such as a factor of 100. The results demonstrate the effectiveness of our proposed filter and merge feature sampling module. Therefore, ERMVP establishes an appropriate and effective communication paradigm among vehicles, ensuring an optimal tradeoff between perception and communication.",
+        "bbox": [
+            75,
+            603,
+            472,
+            866
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5. Qualitative Evaluation",
+        "text_level": 1,
+        "bbox": [
+            499,
+            355,
+            709,
+            369
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 6 displays the detection results of V2VNet [38], CoBEVT [41], V2X-ViT [42], and ERMVP in two scenarios. Clearly, ERMVP offers more precise and comprehensive detection than the prior SOTA methods. To begin with, ERMVP produces a greater number of predicted bounding boxes that are well aligned with the ground truths, while other methods show significant discrepancies. This demonstrates ERMVP's robustness, particularly in high-noise environments. Moreover, ERMVP detects more dynamic objects (more ground truth bounding boxes find matches). This suggests that ERMVP can effectively combine the inputs from nearby vehicles, leading to a thorough scene representation. Overall, these qualitative assessments confirm ERMVP's strengths in delivering accurate and comprehensive perception, especially in challenging conditions.",
+        "bbox": [
+            497,
+            376,
+            890,
+            583
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            499,
+            591,
+            616,
+            605
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. We introduce a filter and merge feature sampling strategy to efficient communication, a feature spatial calibration module to align features in fine-grid and two spatial information aggregation components to solve the bottleneck of the fusion method. Extensive experiments prove the superiority of ERMVP and the effectiveness of our components. The feasibility of our approach for other modalities will be explored in future work.",
+        "bbox": [
+            497,
+            614,
+            890,
+            752
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            499,
+            760,
+            658,
+            776
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This work is supported by the Shanghai Key Research Laboratory of NSAI, the Innovation Platform for Academicians of Hainan Province, Haikou, 570228, the Specific Research Fund of the Innovation Platform for Academicians of Hainan Province under Grant YSPTZX202314, and NFSC under grant 62250410368.",
+        "bbox": [
+            497,
+            783,
+            890,
+            866
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "12582",
+        "bbox": [
+            465,
+            887,
+            504,
+            897
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            124,
+            173,
+            138
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5",
+            "[2] Marwa Chafii, Salmane Naoumi, Reda Alami, Ebtesam Al-mazrouei, Mehdi Bennis, and Merouane Debbah. Emergent communication in multi-agent reinforcement learning for future wireless networks. arXiv preprint arXiv:2309.06021, 2023. 3",
+            "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 6, 7",
+            "[4] Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: deformable patch attack. In European conference on computer vision, pages 529-548. Springer, 2022. 1",
+            "[5] Zhaoyu Chen, Bo Li, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Query-efficient decision-based black-box patch attack. IEEE Transactions on Information Forensics and Security, 2023.",
+            "[6] Zhaoyu Chen, Bo Li, Shuang Wu, Kaixun Jiang, Shouhong Ding, and Wenqiang Zhang. Content-based unrestricted adversarial attack. Advances in Neural Information Processing Systems, 36, 2024. 1",
+            "[7] Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional feature fusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3560-3569, 2021. 6",
+            "[8] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6",
+            "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5",
+            "[10] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016. 1",
+            "[11] Jiaming Gu, Jingyu Zhang, Muyang Zhang, Weiliang Meng, Shibiao Xu, Jiguang Zhang, and Xiaopeng Zhang. Feaco: Reaching robust feature-level consensus in noisy pose conditions. In Proceedings of the 31st ACM International Conference on Multimedia, page 3628-3636, New York, NY, USA, 2023. Association for Computing Machinery. 1",
+            "[12] Monowar Hasan, Sibin Mohan, Takayuki Shimizu, and Hongsheng Lu. Securing vehicle-to-everything (v2x) communication platforms. IEEE Transactions on Intelligent Vehicles, 5(4):693-713, 2020. 1",
+            "[13] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In International Conference on Machine Learning, pages 12633-12646. PMLR, 2023. 5"
+        ],
+        "bbox": [
+            78,
+            147,
+            468,
+            864
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[14] 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. 5",
+            "[15] Yue Hu, Shaoheng Fang and Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 5, 6, 7",
+            "[16] Kaixun Jiang, Zhaoyu Chen, Hao Huang, Jiafeng Wang, Dingkang Yang, Bo Li, Yan Wang, and Wenqiang Zhang. Efficient decision-based black-box patch attacks on video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4379-4389, 2023. 1",
+            "[17] Hamza Khan, Sumudu Samarakoon, and Mehdi Bennis. Enhancing video streaming in vehicular networks via resource slicing. IEEE Transactions on Vehicular Technology, 69(4): 3513-3522, 2020. 1",
+            "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6",
+            "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 6",
+            "[20] Yiming Li, Bir Bhanu, and Wei Lin. Auction protocol for camera active control. In 2010 IEEE International Conference on Image Processing, pages 4325-4328. IEEE, 2010. 2",
+            "[21] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 5, 6, 7",
+            "[22] Yiming Li, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: A virtual collaborative perception dataset for autonomous driving. arXiv preprint arXiv:2202.08449, 2022. 3",
+            "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 6",
+            "[24] Yang Liu, Jing Liu, Kun Yang, Bobo Ju, Siao Liu, Yuzheng Wang, Dingkang Yang, Peng Sun, and Liang Song. Ampnet: Appearance-motion prototype network assisted automatic video anomaly detection system. IEEE Transactions on Industrial Informatics, 20(2):2843-2855, 2024. 1",
+            "[25] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 1, 6",
+            "[26] Zonglin Meng, Xin Xia, Runsheng Xu, Wei Liu, and Jiaqi Ma. Hydro-3d: Hybrid object detection and tracking for cooperative perception using 3d lidar. IEEE Transactions on Intelligent Vehicles, 2023. 1"
+        ],
+        "bbox": [
+            502,
+            125,
+            890,
+            865
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "12583",
+        "bbox": [
+            465,
+            887,
+            504,
+            897
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[27] Yaru Niu, Rohan R Paleja, and Matthew C Gombolay. Multiagent graph-attention communication and teaming. In AA-MAS, pages 964–973, 2021. 2",
+            "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6",
+            "[29] Faisal Qureshi and Demetri Terzopoulos. Smart camera networks in virtual reality. Proceedings of the IEEE, 96(10): 1640-1656, 2008. 2",
+            "[30] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3365-3374, 2021. 4",
+            "[31] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. An efficient and robust object-level cooperative perception framework for connected and automated driving. arXiv preprint arXiv:2210.06289, 2022. 4",
+            "[32] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth international conference on machine learning, pages 330-337, 1993. 2",
+            "[33] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International conference on machine learning, pages 10096-10106. PMLR, 2021. 5",
+            "[34] Nicholas Vadivelu, Mengye Ren, James Tu, Jingkang Wang, and Raquel Urtasun. Learning to communicate and correct pose errors. In Conference on Robot Learning, pages 1195-1210. PMLR, 2021. 1",
+            "[35] Binglu Wang, Lei Zhang, Zhaozhong Wang, Yongqiang Zhao, and Tianfei Zhou. Core: Cooperative reconstruction for multi-agent perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8710-8720, 2023. 1, 3",
+            "[36] Jian Wang, Yameng Shao, Yuming Ge, and Rundong Yu. A survey of vehicle to everything (v2x) testing. Sensors, 19(2): 334, 2019. 1",
+            "[37] Tianhang Wang, Guang Chen, Kai Chen, Zhengfa Liu, Bo Zhang, Alois Knoll, and Changjun Jiang. Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework. arXiv preprint arXiv:2303.12400, 2023. 3",
+            "[38] Tsun-Hsuan Wang, Sivabanan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 1, 3, 5, 6, 7, 8",
+            "[39] Qianyi Wu, Kaisiyuan Wang, Kejie Li, Jianmin Zheng, and Jianfei Cai. Objectsdf++: Improved object-compositional neural implicit surfaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21764-21774, 2023. 1"
+        ],
+        "bbox": [
+            78,
+            125,
+            468,
+            865
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[40] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6",
+            "[41] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers, 2022. 1, 3, 6, 7, 8",
+            "[42] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European Conference on Computer Vision, page 107-124. Springer, 2022. 1, 3, 5, 6, 7, 8",
+            "[43] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3, 5, 6, 7",
+            "[44] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 3, 6",
+            "[45] Kun Yang, Peng Sun, Jieyu Lin, Azzedine Boukerche, and Liang Song. A novel distributed task scheduling framework for supporting vehicular edge intelligence. In 2022 IEEE 42nd International Conference on Distributed Computing Systems (ICDCS), pages 972-982. IEEE, 2022. 1",
+            "[46] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception, 2023. 1",
+            "[47] Kun Yang, Dingkang Yang, Jingyu Zhang, Hanqi Wang, Peng Sun, and Liang Song. What2comm: Towards communication-efficient collaborative perception via feature decoupling. In Proceedings of the 31st ACM International Conference on Multimedia, page 7686-7695, New York, NY, USA, 2023. Association for Computing Machinery. 1",
+            "[48] Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R. Manmatha, Mu Li, and Alexander Smola. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2736-2746, 2022. 5",
+            "[49] Sai Qian Zhang, Qi Zhang, and Jieyu Lin. Succinct and robust multi-agent communication with temporal message control. Advances in Neural Information Processing Systems, 33:17271-17282, 2020. 3"
+        ],
+        "bbox": [
+            502,
+            125,
+            890,
+            781
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "12584",
+        "bbox": [
+            465,
+            887,
+            504,
+            897
+        ],
+        "page_idx": 9
+    }
+]

2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/197e9226-8724-4a9f-ac9a-1556459fa452_model.json

@@ -0,0 +1,2035 @@
+[
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.129,
+                0.051,
+                0.201,
+                0.086
+            ],
+            "angle": 0,
+            "content": "CVF"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.253,
+                0.049,
+                0.793,
+                0.088
+            ],
+            "angle": 0,
+            "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.102,
+                0.162,
+                0.87,
+                0.204
+            ],
+            "angle": 0,
+            "content": "ERMVP: Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.132,
+                0.228,
+                0.838,
+                0.278
+            ],
+            "angle": 0,
+            "content": "Jingyu Zhang\\(^{1}\\) Kun Yang\\(^{1}\\) Yilei Wang\\(^{1}\\) Hanqi Wang\\(^{1}\\) Peng Sun\\(^{2,*}\\) Liang Song\\(^{1,*}\\)  \n\\(^{1}\\)Academy for Engineering and Technology, Fudan University \\(^{2}\\)Duke Kunshan University  \n{jingyuzhang22, yileiwang23}@m.fudan.edu.cn, songl@m.fudan.edu.cn"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.235,
+                0.309,
+                0.314,
+                0.325
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.339,
+                0.473,
+                0.658
+            ],
+            "angle": 0,
+            "content": "Collaborative perception enhances perception performance by enabling autonomous vehicles to exchange complementary information. Despite its potential to revolutionize the mobile industry, challenges in various environments, such as communication bandwidth limitations, localization errors and information aggregation inefficiencies, hinder its implementation in practical applications. In this work, we propose ERMVP, a communication-Efficient and collaboration-Robust Multi-Vehicle Perception method in challenging environments. Specifically, ERMVP has three distinct strengths: i) It utilizes the hierarchical feature sampling strategy to abstract a representative set of feature vectors, using less communication overhead for efficient communication; ii) It employs the sparse consensus features to execute precise spatial location calibrations, effectively mitigating the implications of vehicle localization errors; iii) A pioneering feature fusion and interaction paradigm is introduced to integrate holistic spatial semantics among different vehicles and data sources. To thoroughly validate our method, we conduct extensive experiments on real-world and simulated datasets. The results demonstrate that the proposed ERMVP is significantly superior to the state-of-the-art collaborative perception methods."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.685,
+                0.21,
+                0.699
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.709,
+                0.47,
+                0.835
+            ],
+            "angle": 0,
+            "content": "Autonomous vehicles are widely recognized as a valuable means to enhance road safety and traffic efficiency. Equipped with lidar, cameras, and other sensors, these vehicles are capable of accurately sensing their surroundings to ensure safe and reliable operation. However, the single-vehicle perception system has inevitable drawbacks [26, 39], such as a limited sensor field of view that can be easily obstructed and the challenge of detecting distant objects due to sparse and low-resolution data. Recently, the"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.31,
+                0.892,
+                0.395
+            ],
+            "angle": 0,
+            "content": "advances in vehicle-to-vehicle (V2V) communication technologies [12, 17, 36] and deep learning [4-6, 10, 16, 24, 45] have spurred innovation and progress in the collaborative perception technology. This technology allows connected autonomous vehicles (CAVs) to share sensory data, leading to more comprehensive environmental perception."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.395,
+                0.893,
+                0.602
+            ],
+            "angle": 0,
+            "content": "Although collaborative perception technology shows great potential in transforming the mobility industry, its practical application faces several challenges, including communication bandwidth limitations [15, 35], localization errors [11, 25] and information aggregation inefficiencies [38, 47]. In practical situations, wireless communication resource and reliability constraints severely hamper the efficacy of delay-sensitive collaborative perception. While recent works [15, 42] have achieved a balance between perception performance and communication bandwidth through well-designed mechanisms, these methods have their limitations because they primarily considering information compression over spatial redundancy. This narrow focus exacerbates performance degradation at high compression ratios."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.604,
+                0.895,
+                0.868
+            ],
+            "angle": 0,
+            "content": "Further, complex dynamic environments lead to localization errors, which result in inaccurate relative transform estimates and spatial feature misalignment. This relative pose noise produces misleading features that adversely affect the effectiveness of collaborative perception. Existing methods [25, 34] attempt to optimize the overall pose through intensive computation, but the high latency makes them unsuitable for real-time dynamic perception. Meanwhile, collaborative methods [3, 38, 41, 42, 46] only focus on aggregated information, but overlook the inherent perceptual strengths of ego vehicle. This paradigm is vulnerable to the perturbations introduced by collaborative noise, including asynchronous motion blur and inaccurate projections. Such drawback becomes a bottleneck for achieving optimal perceptual performance. In contrast, ego-centric features may contain locally accurate spatial location information that is not affected by collaborative noise. Therefore, a priority for establishing a pragmatic collaborative perception system is to effectively overcome the above challenges."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.076,
+                0.843,
+                0.47,
+                0.867
+            ],
+            "angle": 0,
+            "content": "*Corresponding authors. Our code is available at https://github.com/Terry9a/ERMVP."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.466,
+                0.888,
+                0.506,
+                0.899
+            ],
+            "angle": 0,
+            "content": "12575"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.078,
+                0.123,
+                0.891,
+                0.372
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.382,
+                0.893,
+                0.421
+            ],
+            "angle": 0,
+            "content": "Figure 1. The overall architecture of the proposed framework. The framework consists of six phases: metadata sharing and feature extraction, filter and merge feature sampling, feature spatial calibration, attention-based feature fusion, accuracy enhanced feature interaction and detection decoders. The details of each individual component are illustrated in Section 3."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.437,
+                0.473,
+                0.866
+            ],
+            "angle": 0,
+            "content": "Based on these observations, we propose ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. From Figure 1, ERMVP uses four proposed core components to tackle existing challenges jointly. Specifically, (i) we first design an advanced filter and merge feature sampling strategy to tackle the limitations of wireless communication resources. This strategy considers both inter-class and intra-class redundancy relationships to abstract a refined set of feature vectors from the redundant features, using less communication overhead for efficient communication. (ii) Second, we introduce a plug-and-play feature spatial calibration module to mitigate the implications of vehicle localization errors. This module ingeniously utilizes consensus sparse foreground features to align the relative pose relationships between ego-vehicle and collaborators without any precise pose supervision. (iii) Furthermore, we present a pioneering feature fusion and interaction paradigm to integrate holistic spatial semantics. This paradigm comprises two key components: The first is an attention-based feature fusion module that alternates between local and global attention to fuse heterogeneous information from different vehicles. The second is an accuracy enhanced feature interaction strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by fused features. Through these tailored components, ERMVP represents a significant advancement towards pragmatic collaborative perception. To validate the effectiveness of the ER-MVP, we conduct extensive experiments on two collaborative 3D object detection datasets including V2V4Real [44]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.436,
+                0.892,
+                0.492
+            ],
+            "angle": 0,
+            "content": "and OPV2V [43]. Comprehensive experimental results demonstrate that our method outperforms previous state-of-the-art methods under the bandwidth-limited noisy setting. The main contributions can be summarized as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.492,
+                0.891,
+                0.56
+            ],
+            "angle": 0,
+            "content": "- We present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method, which addresses the communication bandwidth limitations, localization errors and information aggregation efficiency challenges."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.561,
+                0.892,
+                0.629
+            ],
+            "angle": 0,
+            "content": "- We develop a filter and merge feature sampling strategy to enhance communication efficiency, a feature spatial calibration module for accurate spatial feature alignment, and two information aggregation components to optimize the fusion process."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.63,
+                0.892,
+                0.684
+            ],
+            "angle": 0,
+            "content": "- We conduct extensive experiments on both real-world and simulated datasets. The results show the superiority of our method and the necessity of the proposed components."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.5,
+                0.492,
+                0.892,
+                0.684
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.698,
+                0.641,
+                0.712
+            ],
+            "angle": 0,
+            "content": "2. Related Work"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.721,
+                0.759,
+                0.735
+            ],
+            "angle": 0,
+            "content": "2.1. Multi-Agent Communication"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.742,
+                0.892,
+                0.867
+            ],
+            "angle": 0,
+            "content": "Communication has played a pivotal role in the development of robust multi-agent systems. Early multi-agent communication [20, 29, 32] relied on predefined protocols and heuristics to regulate communication between agents. However, these fixed methods are unsuitable in complex and dynamic settings. Advances in deep learning have inspired the development of advanced strategies for information exchange and collaboration among agents. For example, MAGIC [27] used a graph attention encoder to solve the"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.466,
+                0.888,
+                0.506,
+                0.899
+            ],
+            "angle": 0,
+            "content": "12576"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.126,
+                0.473,
+                0.252
+            ],
+            "angle": 0,
+            "content": "problem of when and to whom to send messages. TMC [49] utilized temporal smoothing to ensure efficient communication. EC-MARL [2] tackled high-dimensional continuous control and partially observable states by introducing a rapid communication protocol to resolve task dilemmas. In comparison, our research focuses on LiDAR-based collaborative 3D object detection tasks in complex driving scenarios. We propose a hierarchical feature sampling strategy to achieve efficient communication across agents."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.262,
+                0.307,
+                0.276
+            ],
+            "angle": 0,
+            "content": "2.2. Collaborative Perception"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.283,
+                0.473,
+                0.546
+            ],
+            "angle": 0,
+            "content": "Factors including limited fields of view of sensors and physical obstructions in the environment can adversely impact the perception capabilities of individual agents. To address these challenges, collaborative perception within multi-agent systems has become a key technology. Utilizing newly available datasets [22, 43, 44], several novel collaborative perception methods have been proposed. For example, V2VNet [38] incorporated a graph neural network to fuse data from different agents. V2X-ViT [42] introduced a transformer architecture that combined information from vehicles and infrastructures. Where2comm [15] used feature spatial heterogeneity to reduce bandwidth utilization by transmitting sparse feature maps. CoBEVT [41] proposed the first multi-camera-based collaborative perception framework and designed the fused axis attention module to enable multi-view interactions. In this work, we propose an innovative feature calibration and interaction method that ensures robust vehicle collaboration in high-noise dynamic environments."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.561,
+                0.17,
+                0.575
+            ],
+            "angle": 0,
+            "content": "3. Method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.585,
+                0.47,
+                0.642
+            ],
+            "angle": 0,
+            "content": "This section introduces our efficient and robust collaborative perception method. Figure 1 shows the general procedure of our method, which is divided into six phases. We will detail each phase in the following introduction."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.651,
+                0.437,
+                0.667
+            ],
+            "angle": 0,
+            "content": "3.1. Metadata Sharing and Feature Extraction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.673,
+                0.473,
+                0.868
+            ],
+            "angle": 0,
+            "content": "In multi-vehicle collaboration scenarios, one of the CAVs, known as the ego vehicle, constructs a communication graph. In this graph, the ego vehicle serves as the requester, while other vehicles within its communication range act as supporters. The ego vehicle broadcasts its metadata information, including position, heading, speed, and more. Supporters, upon receiving this metadata, project their local point cloud observations into the ego vehicle's coordinate system. Each vehicle then encodes these transformed point clouds into bird's eye view (BEV) features, yielding a visual representation. Given the \\(i\\)-th vehicle local observations \\(\\mathbf{X}_i\\), the extracted features are represented as \\(\\mathbf{F}_i = \\Phi_{\\mathrm{enc}}(\\mathbf{X}_i) \\in \\mathbb{R}^{H \\times W \\times C}\\), where \\(\\Phi_{\\mathrm{enc}}(\\cdot)\\) denotes the PointPillar [19] encoder shared by all vehicles and \\(H\\), \\(W\\),"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.126,
+                0.892,
+                0.169
+            ],
+            "angle": 0,
+            "content": "and \\( C \\) stand for the height, width, and channel of the feature map, respectively. Then, these extracted features are fed into the filter and merge feature sampling module."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.175,
+                0.809,
+                0.191
+            ],
+            "angle": 0,
+            "content": "3.2. Filter and Merge Feature Sampling"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.197,
+                0.892,
+                0.403
+            ],
+            "angle": 0,
+            "content": "Previous works have utilized well-designed mechanisms such as information entropy communication selection [37] and spatial heterogeneity map [15, 35] to reduce the required transmission bandwidth. However, these methods primarily focus on the inter-class redundancy between foreground and background features, ignoring the intra-class redundancy among features, which results in sub-optimal compression. To address this gap, we introduce an advanced Filtering and Merging Feature Sampling strategy (FMS). This strategy considers both inter-class and intra-class redundancy relationships, efficiently extracting a concise and distinctive set of feature vectors from the original feature maps, thus reducing communication overheads more effectively. FMS is composed of two core components as follows."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.404,
+                0.893,
+                0.584
+            ],
+            "angle": 0,
+            "content": "Filter Sampler. In object detection, foreground areas containing objects are more significant than the background areas. Therefore, we implement the idea of reducing spatial redundancy into a feature filter sampler module, aiming to preserve perceptually important yet sparse sets of feature vectors. Since explicitly learning a binary sampler is infeasible, we develop a confidence filter strategy. Initially, a detection confidence map is generated for the feature map. It reflects the perceptual importance of different spatial areas, with higher levels indicating potential object areas and lower levels typically denoting redundant background areas. For the feature map \\( \\pmb{F}_i \\) of the \\( i \\)-th vehicle, its confidence map \\( C_i \\) is defined as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.583,
+                0.592,
+                0.891,
+                0.61
+            ],
+            "angle": 0,
+            "content": "\\[\n\\boldsymbol {C} _ {i} = \\Phi_ {\\text {c o n - g e n}} \\left(\\boldsymbol {F} _ {i}\\right) \\in [ 0, 1 ] ^ {H \\times W}, \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.618,
+                0.892,
+                0.768
+            ],
+            "angle": 0,
+            "content": "where \\(\\Phi_{\\mathrm{con\\_gen}(\\cdot)}\\) represents the confidence generation network with detection decoder structure. Then the confidence map is thresholded, followed by non-maximum suppression, resulting in a binary mask \\(B\\). Utilizing this binary mask, we proceed to preserve the sparse foreground features \\(\\widetilde{\\pmb{F}}_i = B\\odot \\pmb{F}_i\\). In order to cope with the dynamically changing environmental conditions and fully ensure the robustness of the system, the filtering threshold can be varied with the sensor data and the network state. We set \\(\\alpha\\) as the filter rate and the number of remaining feature vectors is \\(L = \\alpha \\times HW\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.77,
+                0.892,
+                0.868
+            ],
+            "angle": 0,
+            "content": "Merge Sampler. Upon extracting a detailed foreground feature vector set with the filter sampler, we employ the merge sampler for additional optimization, refining similar or repetitive foreground feature vectors through weighted merging. The process is divided into three stages: information-driven feature grouping, attention-inspired feature merging, and index-based feature reconstruction."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.466,
+                0.888,
+                0.506,
+                0.899
+            ],
+            "angle": 0,
+            "content": "12577"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.126,
+                0.473,
+                0.239
+            ],
+            "angle": 0,
+            "content": "(a) Information-Driven Feature Grouping. Initially, a variant of the nearest neighbor clustering algorithm is applied to group the foreground feature vector set. Given a set of feature vector \\(\\widetilde{\\boldsymbol{F}}_i = [x_1,x_2,\\dots ,x_L]^\\top\\) and cluster center \\(X_{c}\\), we compute the indicator \\(\\delta_{i}\\) for each feature vector. \\(\\delta_{i}\\) is calculated as the minimum feature distance minus the average pixel distance to any other cluster center vector, expressed as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.1,
+                0.249,
+                0.47,
+                0.274
+            ],
+            "angle": 0,
+            "content": "\\[\n\\delta_ {i} = \\min  _ {j: x _ {j} \\in X _ {c}} \\left(\\| x _ {i} - x _ {j} \\| _ {2} ^ {2} - \\gamma \\| p (x _ {i}), p (x _ {j}) \\| _ {2} ^ {2}\\right), \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.285,
+                0.469,
+                0.381
+            ],
+            "angle": 0,
+            "content": "where \\(\\delta_{i}\\) denotes to which cluster the feature vector \\(x_{i}\\) should belong. \\(p(\\cdot)\\) means getting the position of the vector and \\(\\gamma\\) is a hyperparameter. Subsequently, we can divide all the feature vectors in \\(\\widetilde{\\boldsymbol{F}}_i\\) into \\(K\\) clusters represented by \\(G = \\{G_{1},G_{2},\\dots ,G_{K}\\}\\). The total number of clusters is determined by dynamic clustering ratio \\(\\beta\\), and is calculated as \\(K = \\beta \\times L\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.382,
+                0.47,
+                0.492
+            ],
+            "angle": 0,
+            "content": "(b) Attention-Inspired Feature Merging. A straightforward strategy for merging feature vectors is to average of each feature vector within a cluster. However, this scheme can be severely affected by outlier feature vectors. Drawing inspiration from attention mechanisms, we utilize the confidence score as a guide to quantify the significance of each feature. Thus, the merged feature vector \\(\\tilde{x}_i\\) for the \\(i\\)th cluster \\(G_i\\) is computed as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.206,
+                0.502,
+                0.47,
+                0.537
+            ],
+            "angle": 0,
+            "content": "\\[\n\\widetilde {x} _ {i} = \\frac {\\sum_ {j \\in G _ {i}} c _ {j} x _ {j}}{\\sum_ {j \\in G _ {i}} c _ {j}}, \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.546,
+                0.469,
+                0.601
+            ],
+            "angle": 0,
+            "content": "where \\(c_{j}\\) and \\(x_{j}\\) represent the confidence score and the original feature vector, respectively. In the end, we obtain a final set of feature vectors \\(\\mathbf{Z}_i = \\{\\widetilde{x}_1,\\widetilde{x}_2,\\dots ,\\widetilde{x}_K\\}\\) that is necessary for transmission."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.602,
+                0.469,
+                0.713
+            ],
+            "angle": 0,
+            "content": "(c) Index-Based Feature Reconstruction. During the feature grouping and merging process, each feature vector is allocated to a cluster, and every cluster is represented by a merged vector. We maintain a record of the index correspondence between original and merged feature vectors. Utilizing this index record, the ego-vehicle ensures that the merged feature vectors are mapped to their corresponding positions, leading to the reconstruction of feature maps."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.721,
+                0.325,
+                0.736
+            ],
+            "angle": 0,
+            "content": "3.3. Feature Spatial Calibration"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.742,
+                0.469,
+                0.867
+            ],
+            "angle": 0,
+            "content": "Localization errors [30, 31] may lead to the misalignment of feature maps among vehicles. Such misalignment causes the ego-vehicle to misinterpret the object's location, resulting in sub-optimal perception, as depicted in the upper right corner of Figure 2. To address this challenge, we introduce the Feature Spatial Calibration module (FSC) to facilitate precise feature alignment, as shown in Figure 2. The core idea of FSC is that overlapping viewpoints allow multiple vehicles to detect partially the same objects. The process"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.503,
+                0.126,
+                0.891,
+                0.254
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.499,
+                0.265,
+                0.892,
+                0.304
+            ],
+            "angle": 0,
+            "content": "Figure 2. Illustration of the proposed FSC. By employing FSC, misaligned matching regions are corrected at the feature level, leading to more accurate predictions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.313,
+                0.891,
+                0.34
+            ],
+            "angle": 0,
+            "content": "involves three phases: consensus matching, geometric verification, and error modulation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.341,
+                0.892,
+                0.534
+            ],
+            "angle": 0,
+            "content": "Consensus Matching. After reconstructing the received features, the ego vehicle generates sparse feature maps. Within these maps, areas rich in information signify potential target regions, which correspond to the proposed matching areas. We denote the proposed matching regions of the ego vehicle as \\( P \\) and those identified by the collaborative vehicle under noisy pose conditions as \\( Q \\). Utilizing \\( P \\) and \\( Q \\), a weighted bipartite graph is constructed, wherein the weight of each edge is determined by the distance between the nodes, encapsulated in a cost matrix. The matching process is then converted into a linear assignment task, with the objective of identifying a matching result with the lowest cumulative edge weight. This procedure yields the matching pairs \\( M \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.535,
+                0.892,
+                0.77
+            ],
+            "angle": 0,
+            "content": "Geometric Verification. Invalid matches may occur due to objects located in exclusive zones and detection noise. To tackle this issue, we utilize RANSAC to filter and sift a consistent set of matches aligned with expected geometric transformations. Initially, a random matching subset \\(M_{s} \\subseteq M\\) is selected, and then the transformation matrix \\(\\Gamma_{s}\\) is computed using singular value decomposition. When \\(\\Gamma_{s}\\) is applied to all pairs within \\(M_{s}\\), and if the posttransformation distance between these pairs remains below the threshold \\(\\eta\\), the set is considered correctly aligned. The threshold \\(\\eta\\) reflects the allowable localization error within the original collaborative framework. This process is iteratively conducted to identify the optimal transformation matrix that correlates with the maximum number of correct matches. Ultimately, a optimal refined transformation matrix \\(\\Gamma_{r}\\) is obtained and applied in subsequent spatial calibration operations, yielding the aligned feature \\(\\widetilde{Z}_{j} = \\Gamma_{r}Z_{j}\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.771,
+                0.892,
+                0.867
+            ],
+            "angle": 0,
+            "content": "Error Modulation. To enhance the adaptability of the calibration method in various environments, we incorporate an error modulation strategy. This strategy aims to achieve a balance between localization errors and the estimation errors that emerge from the calibration process. It measures the overlap ratios between ego and collaborative features in both their adjusted and original states. Subsequently, fea"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.466,
+                0.888,
+                0.505,
+                0.899
+            ],
+            "angle": 0,
+            "content": "12578"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.084,
+                0.123,
+                0.465,
+                0.254
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.263,
+                0.47,
+                0.303
+            ],
+            "angle": 0,
+            "content": "Figure 3. A example of local and global Attention in AFF. This shows how AFF processes both 3D local windows and the global query token for position-level and context-aware aggregation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.319,
+                0.47,
+                0.348
+            ],
+            "angle": 0,
+            "content": "tures demonstrating the most optimal overlap are selected as the final input."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.355,
+                0.357,
+                0.369
+            ],
+            "angle": 0,
+            "content": "3.4. Attention-based Feature Fusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.376,
+                0.471,
+                0.668
+            ],
+            "angle": 0,
+            "content": "In multi-vehicle collaboration scenarios, vehicles are able to capture heterogeneous information from different spatial regions. In order to efficiently fuse perceptual features from multiple vehicles, we propose an Attention-based Feature Fusion method (AFF). As shown in Figure 3, AFF utilizes alternating local and global attention to achieve position-level accurate matching in occlusion-variable traffic scenes and to capture the global semantics attention of road topology and traffic states. Specifically, we stack all the vehicles' features to \\( S \\in \\mathbb{R}^{N \\times H \\times W \\times C} \\), where \\( N \\) is the number of vehicles. The features are divided into 3D non-overlapping windows, each of size \\( N \\times P \\times P \\). The shape of the partition tensor is \\( \\left(\\frac{H}{P} \\times \\frac{W}{P}, N \\times P^2, C\\right) \\). We perform 3D local self attention on the local window and implement region interaction within the window. In contrast, global attention goes beyond the limitations of the local view. It uses the extracted global query token and is shared between all windows to interact with the local key and value representations, thus helping to capture long-range dependencies of the features. Inspired by the efficiency of [13, 33], the global query token \\( g \\) is generated as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.151,
+                0.676,
+                0.47,
+                0.693
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbf {g} = \\mathcal {P} \\left(\\mathcal {C} _ {1 \\times 1} \\left(\\mathcal {S E} (\\mathcal {G} (\\mathcal {D} (\\mathbf {x}))) + \\mathbf {x}\\right)\\right), \\tag {4}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.701,
+                0.47,
+                0.868
+            ],
+            "angle": 0,
+            "content": "where \\(\\mathcal{D}\\), \\(\\mathcal{G}\\), \\(\\mathcal{SE}\\), \\(\\mathcal{C}_{1\\times 1}\\) and \\(\\mathcal{P}\\) denote the depth-wise separable convolution, GELU activation function, Squeeze-and-Excitation operation, \\(1\\times 1\\) convolution and max pooling, respectively. These designs enable the global query token to provide advantages such as inductive bias and modeling of inter-channel dependencies. We then construct our proposed AFF module by combining this local and global attention with typical designs of Transformers, including LayerNorm [1], MLPs [9], and skip-connections [14]. This module allows the system to analyze spatial correlations from a local view while also capturing global feature responses, ensuring efficient and precise perception in dy"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.51,
+                0.125,
+                0.885,
+                0.255
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.518,
+                0.265,
+                0.871,
+                0.278
+            ],
+            "angle": 0,
+            "content": "Figure 4. The architecture of the proposed AFI component."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.307,
+                0.89,
+                0.334
+            ],
+            "angle": 0,
+            "content": "namic, complex, and occlusion-variable traffic scenarios. Finally, we obtain the fused features \\( \\pmb{H}_i \\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.351,
+                0.844,
+                0.366
+            ],
+            "angle": 0,
+            "content": "3.5. Accuracy Enhanced Feature Interaction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.374,
+                0.892,
+                0.652
+            ],
+            "angle": 0,
+            "content": "Previous works [15, 21, 38, 42, 43] have demonstrated that fused features can provide richer semantic information, thereby enhancing perceptual performance. However, they may be affected by collaborative noise, such as asynchronous motion blur and inaccurate projections, which can compromise accurate position information and become a bottleneck to the optimal realization of perceptual performance. Ego-centric features may contain locally critical spatial location information without being affected by collaborative noise. To this end, we propose a novel Accuracy Enhanced Feature Interaction (AEI) strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by collaborative fused features. Firstly, we design a context cross attention module that is tailored for feature interaction, as shown in Figure 4. It considers the features from the ego-vehicle as queries, and obtains the context of the query location from the fused features as the key and value to achieve cross-attention. The computed output of the spatial location \\((i,j)\\) is as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.564,
+                0.666,
+                0.892,
+                0.699
+            ],
+            "angle": 0,
+            "content": "\\[\ny _ {i j} = \\sum_ {a, b \\in \\mathcal {N} _ {k} (i, j)} \\operatorname {s o f t m a x} \\left(q _ {i j} ^ {\\top} k _ {a b}\\right) v _ {a b}, \\tag {5}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.713,
+                0.893,
+                0.868
+            ],
+            "angle": 0,
+            "content": "where \\(\\mathcal{N}_k(i,j)\\) is context position of \\((i,j)\\) and \\(q_{ij}, k_{ab}, v_{ab}\\) represent the ego query, context key and value, respectively. As the relevant features of the same object are often located in similar locations across different feature maps, we prioritize the local context details of the query location. Additionally, considering that only a few positions within the sparse feature map hold significant information, focusing on these context-rich areas is computationally efficient. Following this, we employ the split attention [48] to adaptively fuse information from multiple branches, generating the enhanced output feature \\(\\widetilde{H}_i\\)."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.466,
+                0.888,
+                0.506,
+                0.899
+            ],
+            "angle": 0,
+            "content": "12579"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.08,
+                0.123,
+                0.894,
+                0.31
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Dataset</td><td colspan=\"3\">V2V4Real</td><td colspan=\"3\">OPV2V</td></tr><tr><td>Noise Level σt/σr (m/°)</td><td>0.0/0.0</td><td>0.4/0.4</td><td>0.8/0.8</td><td>0.0/0.0</td><td>0.4/0.4</td><td>0.8/0.8</td></tr><tr><td>No Collaboration</td><td>41.64/23.43</td><td>41.64/23.43</td><td>41.64/23.43</td><td>73.25/58.22</td><td>73.25/58.22</td><td>73.25/58.22</td></tr><tr><td>Late Fusion</td><td>56.80/27.62</td><td>40.81/13.31</td><td>26.65/10.66</td><td>86.46/79.21</td><td>76.56/40.01</td><td>45.57/21.00</td></tr><tr><td>F-Cooper [3]</td><td>61.12/32.36</td><td>51.71/24.48</td><td>41.90/19.91</td><td>87.39/79.39</td><td>79.33/40.12</td><td>52.91/16.90</td></tr><tr><td>AttFuse [7]</td><td>64.41/34.32</td><td>58.42/30.02</td><td>49.37/25.44</td><td>90.50/81.80</td><td>83.54/52.43</td><td>68.37/40.96</td></tr><tr><td>Where2comm [15]</td><td>63.69/34.79</td><td>58.48/28.09</td><td>49.77/24.25</td><td>90.46/84.22</td><td>82.17/56.47</td><td>73.95/45.86</td></tr><tr><td>DiscoNet [21]</td><td>64.19/35.25</td><td>57.71/28.41</td><td>49.19/24.94</td><td>89.58/81.45</td><td>83.25/53.22</td><td>61.37/30.53</td></tr><tr><td>V2VNet [38]</td><td>65.70/35.38</td><td>62.11/30.63</td><td>54.73/25.57</td><td>91.35/82.43</td><td>85.43/54.16</td><td>70.43/31.28</td></tr><tr><td>CoAlign [25]</td><td>64.06/36.60</td><td>59.58/31.52</td><td>53.02/26.60</td><td>90.93/84.28</td><td>84.48/57.45</td><td>73.95/48.86</td></tr><tr><td>V2X-ViT [42]</td><td>66.42/37.34</td><td>63.33/32.38</td><td>57.15/28.22</td><td>91.74/83.31</td><td>82.91/54.73</td><td>61.70/26.07</td></tr><tr><td>CoBEVT [41]</td><td>66.01/37.36</td><td>58.63/29.09</td><td>49.22/23.66</td><td>91.71/85.98</td><td>85.49/60.64</td><td>64.63/29.59</td></tr><tr><td>ERMVP (Ours)</td><td>67.66/42.97</td><td>65.01/40.45</td><td>60.88/35.01</td><td>92.18/85.59</td><td>85.67/66.32</td><td>77.13/59.15</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.118,
+                0.32,
+                0.85,
+                0.333
+            ],
+            "angle": 0,
+            "content": "Table 1. Overall performance on V2V4Real and OPV2V datasets with pose noises. The results are reported in AP@0.5/0.7."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.34,
+                0.251,
+                0.354
+            ],
+            "angle": 0,
+            "content": "3.6. Decoder and Loss"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.363,
+                0.47,
+                0.477
+            ],
+            "angle": 0,
+            "content": "Based on the final fused feature \\(\\widetilde{H}_i\\), we use the detection decoder to generate the final prediction output \\(Y_{i} = \\Phi_{\\mathrm{dec}}(\\widetilde{H}_{i})\\). Each position of \\(Y_{i}\\) represents a rotated box with classes (c, x, y, h, w, \\(\\cos \\alpha\\), \\(\\sin \\alpha\\)), corresponding to class confidence, position, size, and angle. These objects are the final output of the proposed collaborative perception system. Following existing work [19], we adopt the smooth \\(L_{1}\\) loss for regression and focal loss [23] for classification."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.497,
+                0.21,
+                0.512
+            ],
+            "angle": 0,
+            "content": "4. Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.522,
+                0.364,
+                0.536
+            ],
+            "angle": 0,
+            "content": "4.1. Datasets and Evaluation Metrics"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.546,
+                0.47,
+                0.794
+            ],
+            "angle": 0,
+            "content": "Datasets. We validate the proposed ERMVP in the task of LiDAR-based 3D object detection on two benchmark datasets including V2V4real [44] and OPV2V [43]. V2V4real [44] is the first large-scale real-world dataset on V2V perception, collected by two vehicles equipped with multi-modal sensors driving together in a variety of scenarios. It covers a driving area of 410 kilometers, including 20K frames of point clouds, with training/validation/test sets containing 14210/2000/3986 frames, respectively. OPV2V [43] is a large-scale vehicle-to-vehicle collaborative perception dataset, simulated by OpenCDA [40] and Carla [8]. It contains 73 different scenarios, in which different numbers (2 to 7) of collaborative vehicles appear together, each equipped with a lidar sensor and 4 cameras. The dataset contains a total of 11,464 frames of point clouds and RGB images. The training/validation/test sets contain 6374, 1980, and 2170 frames, respectively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.797,
+                0.469,
+                0.867
+            ],
+            "angle": 0,
+            "content": "Evaluation Metrics. We adopt the Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 and 0.7 to evaluate the detection performance. The calculation format of communication volume in [15] is used to count the message size by byte in the log scale with base 2."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.34,
+                0.719,
+                0.355
+            ],
+            "angle": 0,
+            "content": "4.2. Implementation Details"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.363,
+                0.892,
+                0.626
+            ],
+            "angle": 0,
+            "content": "We implement the proposed ERMVP and comparison model on the Pytorch toolbox [28], and train them on a NVIDIA GeForceRTX 3090 GPU using the Adam optimizer [18]. The initial learning rate is \\(2\\mathrm{e} - 3\\) and it decays every 15 epochs with a factor of 0.1. All models are trained with 60 epochs and the batch size is set to 2. Early stopping is used to find the best epoch. We also add normal point cloud data augmentation for all experiments, including scaling, rotation, and flipping. All detection models are based on PointPillar [19] backbone to extract 2D features from point clouds and a \\(0.4\\mathrm{m}\\) width/length is used for each voxel. AFF component has 3 encoded layers and a window size of 8 for both local and global attention. The balance hyperparameter \\(\\gamma\\) is 0.1 and the error tolerance threshold \\(\\eta\\) is 0.25. Each vehicle has a communication range of 70 m based on [42], while vehicles outside of this broadcasting radius will be ignored. To simulate the localization and heading errors, we add Gaussian noise with a standard deviation of \\(\\sigma_t\\) for localization errors and \\(\\sigma_r\\) for heading errors."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.637,
+                0.722,
+                0.652
+            ],
+            "angle": 0,
+            "content": "4.3.Quantitative Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.659,
+                0.892,
+                0.867
+            ],
+            "angle": 0,
+            "content": "Comparison of Detection Performance. Table 1 shows the performance comparison results of 3D detection on two datasets. We use the No Collaboration method as a baseline, which only uses ego vehicle's point clouds data without collaboration. Late Fusion allows vehicles to exchange detected outputs and utilizes non-maximum suppression to produce the final result. For intermediate fusion strategies, we evaluate the existing state-of-the-art (SOTA) methods: AttFuse [43], F-Cooper [3], V2VNet [38], DiscoNet [21], V2X-ViT [42], CoAlign (only fusion) [25], CoBEVT [41] and Where2comm [15]. The proposed ER-MVP achieves average improvements of \\(12.48\\% / 18.47\\%\\) and \\(6.18\\% / 10.02\\%\\) on two datasets in AP@0.5/0.7 compared with No Collaboration and Late Fusion methods, demonstrating the superiority of our collaborative method."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.466,
+                0.888,
+                0.506,
+                0.899
+            ],
+            "angle": 0,
+            "content": "12580"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.082,
+                0.123,
+                0.283,
+                0.252
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.284,
+                0.123,
+                0.484,
+                0.251
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.484,
+                0.123,
+                0.683,
+                0.251
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.683,
+                0.123,
+                0.882,
+                0.251
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.259,
+                0.256,
+                0.341,
+                0.267
+            ],
+            "angle": 0,
+            "content": "(a) V2V4Real"
+        },
+        {
+            "type": "image_footnote",
+            "bbox": [
+                0.137,
+                0.273,
+                0.855,
+                0.291
+            ],
+            "angle": 0,
+            "content": "Late Fusion FCooper V2VNet V2X-VIT CoBEVT -Where2comm -ERMVP"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.102,
+                0.298,
+                0.866,
+                0.312
+            ],
+            "angle": 0,
+            "content": "Figure 5. Collaborative perception performance comparison on the V2V4Real and OPV2V with varying communication volume."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.322,
+                0.47,
+                0.42
+            ],
+            "angle": 0,
+            "content": "Meanwhile, it is superior to the SOTA collaborative perception method in both real-world and simulated scenarios: improves the SOTA collaborative performance by \\(1.24\\%\\) and \\(5.61\\%\\) on V2V4Real datasets in AP@0.5/0.7, and by \\(0.44\\%\\) on OPV2V datasets in AP@0.5. These results fully demonstrate the superiority of the ERMVP collaboration paradigm."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.08,
+                0.43,
+                0.456,
+                0.602
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Dataset</td><td colspan=\"2\">V2V4Real</td><td colspan=\"2\">OPV2V</td></tr><tr><td>Noise Level σt(m)</td><td>0.5</td><td>1.0</td><td>0.5</td><td>1.0</td></tr><tr><td>Late Fusion</td><td>11.02</td><td>10.13</td><td>53.88</td><td>38.80</td></tr><tr><td>F-Cooper [3]</td><td>24.60</td><td>18.41</td><td>32.24</td><td>12.71</td></tr><tr><td>AttFuse [43]</td><td>28.10</td><td>24.36</td><td>42.95</td><td>35.53</td></tr><tr><td>DiscoNet [21]</td><td>28.36</td><td>24.05</td><td>45.36</td><td>24.96</td></tr><tr><td>V2VNet [38]</td><td>29.57</td><td>23.67</td><td>47.48</td><td>25.82</td></tr><tr><td>Where2comm [15]</td><td>28.02</td><td>23.53</td><td>46.59</td><td>45.59</td></tr><tr><td>CoAlign [21]</td><td>29.84</td><td>25.82</td><td>49.09</td><td>46.66</td></tr><tr><td>V2X-ViT [42]</td><td>30.99</td><td>23.67</td><td>44.83</td><td>21.77</td></tr><tr><td>CoBEVT [41]</td><td>24.67</td><td>22.01</td><td>41.91</td><td>26.05</td></tr><tr><td>ERMVP (Ours)</td><td>39.26</td><td>33.94</td><td>77.25</td><td>76.09</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.076,
+                0.611,
+                0.47,
+                0.637
+            ],
+            "angle": 0,
+            "content": "Table 2. Detection AP@0.7 on V2V4Real and OPV2V datasets with localization error."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.646,
+                0.471,
+                0.866
+            ],
+            "angle": 0,
+            "content": "Robustness to Localization and Heading Errors. To evaluate the sensitivity of existing methods to localization and heading errors, we use the same noise setting as [42] and conduct extensive experiments on two datasets. As shown in Tables 1 and 2, under ideal settings, Late Fusion and some advanced intermediate fusion methods can detect object vehicles with high accuracy, but their detection accuracy rapidly decreases as the standard deviation of errors increases. For example, when the localization and heading errors of OPV2V dataset are over \\(0.4\\mathrm{m}\\) and \\(0.4^{\\circ}\\), the Disconet [21] and F-Cooper [3] methods fail and perform even worse than No Collaboration. Our method exceeds previous SOTA models at all noise levels and consistently outperforms the No Collaboration baseline, clearly demonstrating the robustness of ERMVP to pose errors. It is noteworthy that for localization errors, ERMVP can maintain consistent"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.322,
+                0.892,
+                0.432
+            ],
+            "angle": 0,
+            "content": "accuracy levels in high noise environments, showing superior robustness. When the localization error is \\(1.0\\mathrm{m}\\), it outperforms the second-best method by \\(29.43\\%\\) and \\(8.08\\%\\) in AP@0.7 on OPV2V and V2V4Real datasets. The reasonable explanations are: (i) feature spatial calibration aligns mismatched collaborative features; (ii) accuracy enhanced feature interaction reduces performance degradation caused by collaborative noise."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.433,
+                0.893,
+                0.626
+            ],
+            "angle": 0,
+            "content": "Comparison of Communication Volume. To achieve pragmatic collaborative perception, it is crucial to evaluate the perceptual performance under different communication volumes. Figure 5 shows the results of the performance comparison under different bandwidth consumption conditions. Note that we do not consider any additional model/data/feature compression for a fair comparison. The results indicate that ERMVP achieves excellent perceptual performance and bandwidth consumption trade-off under all communication bandwidth conditions, consistently outperforming Where2comm [15]. At the same time, it achieves the same detection performance as the previous model [3, 38, 41, 42] with less communication volume on both real-world and simulated datasets."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.633,
+                0.665,
+                0.646
+            ],
+            "angle": 0,
+            "content": "4.4. Ablation Studies"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.501,
+                0.654,
+                0.887,
+                0.728
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>AFF</td><td>FSC</td><td>AEI</td><td>V2V4Real</td><td>OPV2V</td></tr><tr><td></td><td></td><td></td><td>51.71/24.48</td><td>79.33/40.12</td></tr><tr><td>✓</td><td></td><td></td><td>62.24/37.70</td><td>84.30/61.43</td></tr><tr><td>✓</td><td>✓</td><td></td><td>64.36/39.59</td><td>84.98/62.67</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>65.01/40.45</td><td>85.67/66.32</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.498,
+                0.737,
+                0.892,
+                0.789
+            ],
+            "angle": 0,
+            "content": "Table 3. Ablation study results of the proposed core components on the both datasets with noise level of 0.4/0.4. AFF: Attention-based Feature Fusion; FSC: Feature Spatial Calibration; AEI: Accuracy Enhanced Feature Interaction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.797,
+                0.892,
+                0.868
+            ],
+            "angle": 0,
+            "content": "Effect of Core Components. Table 3 details the contribution of each core components in our ERMVP framework. The base model is the naive position-wise maximum fusion method. We then assess the impact of each component by sequentially introducing i) attention-based feature"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.466,
+                0.888,
+                0.505,
+                0.899
+            ],
+            "angle": 0,
+            "content": "12581"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.088,
+                0.125,
+                0.285,
+                0.211
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.088,
+                0.213,
+                0.284,
+                0.297
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.14,
+                0.301,
+                0.218,
+                0.312
+            ],
+            "angle": 0,
+            "content": "(a) V2VNet"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.287,
+                0.125,
+                0.483,
+                0.211
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.287,
+                0.213,
+                0.484,
+                0.296
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.34,
+                0.301,
+                0.424,
+                0.312
+            ],
+            "angle": 0,
+            "content": "(b) CoBEVT"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.485,
+                0.125,
+                0.682,
+                0.211
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.486,
+                0.213,
+                0.682,
+                0.296
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.54,
+                0.301,
+                0.621,
+                0.312
+            ],
+            "angle": 0,
+            "content": "(c) V2X-ViT"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.684,
+                0.125,
+                0.883,
+                0.211
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.685,
+                0.213,
+                0.882,
+                0.296
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.738,
+                0.301,
+                0.818,
+                0.312
+            ],
+            "angle": 0,
+            "content": "(d)ERMVP"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.322,
+                0.893,
+                0.349
+            ],
+            "angle": 0,
+            "content": "Figure 6. Qualitative comparison results in real-world scenarios from the V2V4Real dataset with noise level of 0.4/0.4. Green and red boxes denote the ground truths and detection results, respectively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.356,
+                0.47,
+                0.455
+            ],
+            "angle": 0,
+            "content": "fusion (AFF), ii) feature spatial calibration (FSC), iii) accuracy enhanced feature interaction (AEF). The consistent rise in detection results over both datasets demonstrates the effectiveness of each introduced component. Notably, integrating all three components boosts detection performance by \\(13.3\\%\\) and \\(15.97\\%\\) on the V2V4Real dataset for AP@0.5 and AP@0.7, respectively."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.09,
+                0.464,
+                0.268,
+                0.573
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.281,
+                0.464,
+                0.46,
+                0.573
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.135,
+                0.582,
+                0.411,
+                0.595
+            ],
+            "angle": 0,
+            "content": "Figure 7. Hyperparameter analysis of \\(\\alpha\\) and \\(\\beta\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.604,
+                0.473,
+                0.868
+            ],
+            "angle": 0,
+            "content": "Hyperparameters \\(\\alpha\\) and \\(\\beta\\). We explore the impact of the hyperparameters \\(\\alpha\\) and \\(\\beta\\) (as described in Section 3.2), which jointly determine the communication volume. Specifically, the feature compression ratio is represented as \\(\\alpha \\times \\beta\\). The communication volume, in turn, is expressed as the product of the original feature size and the compression rate. By modulating the values of \\(\\alpha\\) and \\(\\beta\\) within the interval of \\(10\\%\\) to \\(80\\%\\), we chart the resultant object detection performance (measured by AP@0.7) in Figure 7. As expected, a drop in either \\(\\alpha\\) or \\(\\beta\\) will impair ERMVP's efficacy. Yet, it's worth noting that ERMVP's performance generally exhibits robustness to these hyperparameters. The performance decline remains fairly minimal even under significant compression, such as a factor of 100. The results demonstrate the effectiveness of our proposed filter and merge feature sampling module. Therefore, ERMVP establishes an appropriate and effective communication paradigm among vehicles, ensuring an optimal tradeoff between perception and communication."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.356,
+                0.71,
+                0.37
+            ],
+            "angle": 0,
+            "content": "4.5. Qualitative Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.377,
+                0.892,
+                0.584
+            ],
+            "angle": 0,
+            "content": "Figure 6 displays the detection results of V2VNet [38], CoBEVT [41], V2X-ViT [42], and ERMVP in two scenarios. Clearly, ERMVP offers more precise and comprehensive detection than the prior SOTA methods. To begin with, ERMVP produces a greater number of predicted bounding boxes that are well aligned with the ground truths, while other methods show significant discrepancies. This demonstrates ERMVP's robustness, particularly in high-noise environments. Moreover, ERMVP detects more dynamic objects (more ground truth bounding boxes find matches). This suggests that ERMVP can effectively combine the inputs from nearby vehicles, leading to a thorough scene representation. Overall, these qualitative assessments confirm ERMVP's strengths in delivering accurate and comprehensive perception, especially in challenging conditions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.592,
+                0.618,
+                0.606
+            ],
+            "angle": 0,
+            "content": "5. Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.615,
+                0.892,
+                0.754
+            ],
+            "angle": 0,
+            "content": "In this paper, we present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. We introduce a filter and merge feature sampling strategy to efficient communication, a feature spatial calibration module to align features in fine-grid and two spatial information aggregation components to solve the bottleneck of the fusion method. Extensive experiments prove the superiority of ERMVP and the effectiveness of our components. The feasibility of our approach for other modalities will be explored in future work."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.761,
+                0.659,
+                0.777
+            ],
+            "angle": 0,
+            "content": "Acknowledgments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.784,
+                0.892,
+                0.867
+            ],
+            "angle": 0,
+            "content": "This work is supported by the Shanghai Key Research Laboratory of NSAI, the Innovation Platform for Academicians of Hainan Province, Haikou, 570228, the Specific Research Fund of the Innovation Platform for Academicians of Hainan Province under Grant YSPTZX202314, and NFSC under grant 62250410368."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.466,
+                0.888,
+                0.505,
+                0.898
+            ],
+            "angle": 0,
+            "content": "12582"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.08,
+                0.125,
+                0.174,
+                0.139
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.148,
+                0.469,
+                0.185
+            ],
+            "angle": 0,
+            "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.188,
+                0.469,
+                0.25
+            ],
+            "angle": 0,
+            "content": "[2] Marwa Chafii, Salmane Naoumi, Reda Alami, Ebtesam Al-mazrouei, Mehdi Bennis, and Merouane Debbah. Emergent communication in multi-agent reinforcement learning for future wireless networks. arXiv preprint arXiv:2309.06021, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.253,
+                0.469,
+                0.316
+            ],
+            "angle": 0,
+            "content": "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.318,
+                0.469,
+                0.369
+            ],
+            "angle": 0,
+            "content": "[4] Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: deformable patch attack. In European conference on computer vision, pages 529-548. Springer, 2022. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.371,
+                0.469,
+                0.421
+            ],
+            "angle": 0,
+            "content": "[5] Zhaoyu Chen, Bo Li, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Query-efficient decision-based black-box patch attack. IEEE Transactions on Information Forensics and Security, 2023."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.423,
+                0.469,
+                0.473
+            ],
+            "angle": 0,
+            "content": "[6] Zhaoyu Chen, Bo Li, Shuang Wu, Kaixun Jiang, Shouhong Ding, and Wenqiang Zhang. Content-based unrestricted adversarial attack. Advances in Neural Information Processing Systems, 36, 2024. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.476,
+                0.47,
+                0.526
+            ],
+            "angle": 0,
+            "content": "[7] Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional feature fusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3560-3569, 2021. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.528,
+                0.469,
+                0.578
+            ],
+            "angle": 0,
+            "content": "[8] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.581,
+                0.469,
+                0.655
+            ],
+            "angle": 0,
+            "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.659,
+                0.469,
+                0.684
+            ],
+            "angle": 0,
+            "content": "[10] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.686,
+                0.469,
+                0.761
+            ],
+            "angle": 0,
+            "content": "[11] Jiaming Gu, Jingyu Zhang, Muyang Zhang, Weiliang Meng, Shibiao Xu, Jiguang Zhang, and Xiaopeng Zhang. Feaco: Reaching robust feature-level consensus in noisy pose conditions. In Proceedings of the 31st ACM International Conference on Multimedia, page 3628-3636, New York, NY, USA, 2023. Association for Computing Machinery. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.764,
+                0.469,
+                0.813
+            ],
+            "angle": 0,
+            "content": "[12] Monowar Hasan, Sibin Mohan, Takayuki Shimizu, and Hongsheng Lu. Securing vehicle-to-everything (v2x) communication platforms. IEEE Transactions on Intelligent Vehicles, 5(4):693-713, 2020. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.816,
+                0.469,
+                0.865
+            ],
+            "angle": 0,
+            "content": "[13] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In International Conference on Machine Learning, pages 12633-12646. PMLR, 2023. 5"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.148,
+                0.47,
+                0.865
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.127,
+                0.892,
+                0.178
+            ],
+            "angle": 0,
+            "content": "[14] 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. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.18,
+                0.892,
+                0.242
+            ],
+            "angle": 0,
+            "content": "[15] Yue Hu, Shaoheng Fang and Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 5, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.245,
+                0.892,
+                0.318
+            ],
+            "angle": 0,
+            "content": "[16] Kaixun Jiang, Zhaoyu Chen, Hao Huang, Jiafeng Wang, Dingkang Yang, Bo Li, Yan Wang, and Wenqiang Zhang. Efficient decision-based black-box patch attacks on video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4379-4389, 2023. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.322,
+                0.892,
+                0.371
+            ],
+            "angle": 0,
+            "content": "[17] Hamza Khan, Sumudu Samarakoon, and Mehdi Bennis. Enhancing video streaming in vehicular networks via resource slicing. IEEE Transactions on Vehicular Technology, 69(4): 3513-3522, 2020. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.374,
+                0.892,
+                0.412
+            ],
+            "angle": 0,
+            "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.414,
+                0.892,
+                0.476
+            ],
+            "angle": 0,
+            "content": "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.478,
+                0.892,
+                0.527
+            ],
+            "angle": 0,
+            "content": "[20] Yiming Li, Bir Bhanu, and Wei Lin. Auction protocol for camera active control. In 2010 IEEE International Conference on Image Processing, pages 4325-4328. IEEE, 2010. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.531,
+                0.892,
+                0.58
+            ],
+            "angle": 0,
+            "content": "[21] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 5, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.582,
+                0.892,
+                0.631
+            ],
+            "angle": 0,
+            "content": "[22] Yiming Li, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: A virtual collaborative perception dataset for autonomous driving. arXiv preprint arXiv:2202.08449, 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.634,
+                0.892,
+                0.684
+            ],
+            "angle": 0,
+            "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.686,
+                0.892,
+                0.749
+            ],
+            "angle": 0,
+            "content": "[24] Yang Liu, Jing Liu, Kun Yang, Bobo Ju, Siao Liu, Yuzheng Wang, Dingkang Yang, Peng Sun, and Liang Song. Ampnet: Appearance-motion prototype network assisted automatic video anomaly detection system. IEEE Transactions on Industrial Informatics, 20(2):2843-2855, 2024. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.751,
+                0.892,
+                0.813
+            ],
+            "angle": 0,
+            "content": "[25] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 1, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.816,
+                0.892,
+                0.866
+            ],
+            "angle": 0,
+            "content": "[26] Zonglin Meng, Xin Xia, Runsheng Xu, Wei Liu, and Jiaqi Ma. Hydro-3d: Hybrid object detection and tracking for cooperative perception using 3d lidar. IEEE Transactions on Intelligent Vehicles, 2023. 1"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.503,
+                0.127,
+                0.892,
+                0.866
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.467,
+                0.888,
+                0.505,
+                0.899
+            ],
+            "angle": 0,
+            "content": "12583"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.127,
+                0.47,
+                0.165
+            ],
+            "angle": 0,
+            "content": "[27] Yaru Niu, Rohan R Paleja, and Matthew C Gombolay. Multiagent graph-attention communication and teaming. In AA-MAS, pages 964–973, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.168,
+                0.47,
+                0.243
+            ],
+            "angle": 0,
+            "content": "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.246,
+                0.469,
+                0.284
+            ],
+            "angle": 0,
+            "content": "[29] Faisal Qureshi and Demetri Terzopoulos. Smart camera networks in virtual reality. Proceedings of the IEEE, 96(10): 1640-1656, 2008. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.287,
+                0.469,
+                0.363
+            ],
+            "angle": 0,
+            "content": "[30] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3365-3374, 2021. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.365,
+                0.469,
+                0.415
+            ],
+            "angle": 0,
+            "content": "[31] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. An efficient and robust object-level cooperative perception framework for connected and automated driving. arXiv preprint arXiv:2210.06289, 2022. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.419,
+                0.469,
+                0.468
+            ],
+            "angle": 0,
+            "content": "[32] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth international conference on machine learning, pages 330-337, 1993. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.471,
+                0.469,
+                0.51
+            ],
+            "angle": 0,
+            "content": "[33] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International conference on machine learning, pages 10096-10106. PMLR, 2021. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.512,
+                0.469,
+                0.562
+            ],
+            "angle": 0,
+            "content": "[34] Nicholas Vadivelu, Mengye Ren, James Tu, Jingkang Wang, and Raquel Urtasun. Learning to communicate and correct pose errors. In Conference on Robot Learning, pages 1195-1210. PMLR, 2021. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.565,
+                0.469,
+                0.628
+            ],
+            "angle": 0,
+            "content": "[35] Binglu Wang, Lei Zhang, Zhaozhong Wang, Yongqiang Zhao, and Tianfei Zhou. Core: Cooperative reconstruction for multi-agent perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8710-8720, 2023. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.631,
+                0.469,
+                0.668
+            ],
+            "angle": 0,
+            "content": "[36] Jian Wang, Yameng Shao, Yuming Ge, and Rundong Yu. A survey of vehicle to everything (v2x) testing. Sensors, 19(2): 334, 2019. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.671,
+                0.469,
+                0.734
+            ],
+            "angle": 0,
+            "content": "[37] Tianhang Wang, Guang Chen, Kai Chen, Zhengfa Liu, Bo Zhang, Alois Knoll, and Changjun Jiang. Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework. arXiv preprint arXiv:2303.12400, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.737,
+                0.469,
+                0.801
+            ],
+            "angle": 0,
+            "content": "[38] Tsun-Hsuan Wang, Sivabanan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 1, 3, 5, 6, 7, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.804,
+                0.469,
+                0.866
+            ],
+            "angle": 0,
+            "content": "[39] Qianyi Wu, Kaisiyuan Wang, Kejie Li, Jianmin Zheng, and Jianfei Cai. Objectsdf++: Improved object-compositional neural implicit surfaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21764-21774, 2023. 1"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.127,
+                0.47,
+                0.866
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.127,
+                0.892,
+                0.19
+            ],
+            "angle": 0,
+            "content": "[40] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.192,
+                0.892,
+                0.242
+            ],
+            "angle": 0,
+            "content": "[41] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers, 2022. 1, 3, 6, 7, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.244,
+                0.892,
+                0.306
+            ],
+            "angle": 0,
+            "content": "[42] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European Conference on Computer Vision, page 107-124. Springer, 2022. 1, 3, 5, 6, 7, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.308,
+                0.892,
+                0.381
+            ],
+            "angle": 0,
+            "content": "[43] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3, 5, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.385,
+                0.892,
+                0.461
+            ],
+            "angle": 0,
+            "content": "[44] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 3, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.462,
+                0.892,
+                0.526
+            ],
+            "angle": 0,
+            "content": "[45] Kun Yang, Peng Sun, Jieyu Lin, Azzedine Boukerche, and Liang Song. A novel distributed task scheduling framework for supporting vehicular edge intelligence. In 2022 IEEE 42nd International Conference on Distributed Computing Systems (ICDCS), pages 972-982. IEEE, 2022. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.528,
+                0.892,
+                0.577
+            ],
+            "angle": 0,
+            "content": "[46] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception, 2023. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.579,
+                0.892,
+                0.654
+            ],
+            "angle": 0,
+            "content": "[47] Kun Yang, Dingkang Yang, Jingyu Zhang, Hanqi Wang, Peng Sun, and Liang Song. What2comm: Towards communication-efficient collaborative perception via feature decoupling. In Proceedings of the 31st ACM International Conference on Multimedia, page 7686-7695, New York, NY, USA, 2023. Association for Computing Machinery. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.655,
+                0.892,
+                0.731
+            ],
+            "angle": 0,
+            "content": "[48] Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R. Manmatha, Mu Li, and Alexander Smola. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2736-2746, 2022. 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.733,
+                0.892,
+                0.782
+            ],
+            "angle": 0,
+            "content": "[49] Sai Qian Zhang, Qi Zhang, and Jieyu Lin. Succinct and robust multi-agent communication with temporal message control. Advances in Neural Information Processing Systems, 33:17271-17282, 2020. 3"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.503,
+                0.127,
+                0.892,
+                0.782
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.467,
+                0.888,
+                0.505,
+                0.899
+            ],
+            "angle": 0,
+            "content": "12584"
+        }
+    ]
+]

2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/197e9226-8724-4a9f-ac9a-1556459fa452_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ecb2465d7f16469e301765389a7c4da1400282ac38680a92e63c4c85621f3e62
+size 1332501

2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/full.md

@@ -0,0 +1,280 @@
+# ERMVP: Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments
+
+Jingyu Zhang $^{1}$  Kun Yang $^{1}$  Yilei Wang $^{1}$  Hanqi Wang $^{1}$  Peng Sun $^{2,*}$  Liang Song $^{1,*}$ $^{1}$ Academy for Engineering and Technology, Fudan University  $^{2}$ Duke Kunshan University  
+{jingyuzhang22, yileiwang23}@m.fudan.edu.cn, songl@m.fudan.edu.cn
+
+# Abstract
+
+Collaborative perception enhances perception performance by enabling autonomous vehicles to exchange complementary information. Despite its potential to revolutionize the mobile industry, challenges in various environments, such as communication bandwidth limitations, localization errors and information aggregation inefficiencies, hinder its implementation in practical applications. In this work, we propose ERMVP, a communication-Efficient and collaboration-Robust Multi-Vehicle Perception method in challenging environments. Specifically, ERMVP has three distinct strengths: i) It utilizes the hierarchical feature sampling strategy to abstract a representative set of feature vectors, using less communication overhead for efficient communication; ii) It employs the sparse consensus features to execute precise spatial location calibrations, effectively mitigating the implications of vehicle localization errors; iii) A pioneering feature fusion and interaction paradigm is introduced to integrate holistic spatial semantics among different vehicles and data sources. To thoroughly validate our method, we conduct extensive experiments on real-world and simulated datasets. The results demonstrate that the proposed ERMVP is significantly superior to the state-of-the-art collaborative perception methods.
+
+# 1. Introduction
+
+Autonomous vehicles are widely recognized as a valuable means to enhance road safety and traffic efficiency. Equipped with lidar, cameras, and other sensors, these vehicles are capable of accurately sensing their surroundings to ensure safe and reliable operation. However, the single-vehicle perception system has inevitable drawbacks [26, 39], such as a limited sensor field of view that can be easily obstructed and the challenge of detecting distant objects due to sparse and low-resolution data. Recently, the
+
+advances in vehicle-to-vehicle (V2V) communication technologies [12, 17, 36] and deep learning [4-6, 10, 16, 24, 45] have spurred innovation and progress in the collaborative perception technology. This technology allows connected autonomous vehicles (CAVs) to share sensory data, leading to more comprehensive environmental perception.
+
+Although collaborative perception technology shows great potential in transforming the mobility industry, its practical application faces several challenges, including communication bandwidth limitations [15, 35], localization errors [11, 25] and information aggregation inefficiencies [38, 47]. In practical situations, wireless communication resource and reliability constraints severely hamper the efficacy of delay-sensitive collaborative perception. While recent works [15, 42] have achieved a balance between perception performance and communication bandwidth through well-designed mechanisms, these methods have their limitations because they primarily considering information compression over spatial redundancy. This narrow focus exacerbates performance degradation at high compression ratios.
+
+Further, complex dynamic environments lead to localization errors, which result in inaccurate relative transform estimates and spatial feature misalignment. This relative pose noise produces misleading features that adversely affect the effectiveness of collaborative perception. Existing methods [25, 34] attempt to optimize the overall pose through intensive computation, but the high latency makes them unsuitable for real-time dynamic perception. Meanwhile, collaborative methods [3, 38, 41, 42, 46] only focus on aggregated information, but overlook the inherent perceptual strengths of ego vehicle. This paradigm is vulnerable to the perturbations introduced by collaborative noise, including asynchronous motion blur and inaccurate projections. Such drawback becomes a bottleneck for achieving optimal perceptual performance. In contrast, ego-centric features may contain locally accurate spatial location information that is not affected by collaborative noise. Therefore, a priority for establishing a pragmatic collaborative perception system is to effectively overcome the above challenges.
+
+![](images/17285042ad72afee503d390d9bf80d325dc5985f785ca8f0682e6d8e69e83e0c.jpg)  
+Figure 1. The overall architecture of the proposed framework. The framework consists of six phases: metadata sharing and feature extraction, filter and merge feature sampling, feature spatial calibration, attention-based feature fusion, accuracy enhanced feature interaction and detection decoders. The details of each individual component are illustrated in Section 3.
+
+Based on these observations, we propose ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. From Figure 1, ERMVP uses four proposed core components to tackle existing challenges jointly. Specifically, (i) we first design an advanced filter and merge feature sampling strategy to tackle the limitations of wireless communication resources. This strategy considers both inter-class and intra-class redundancy relationships to abstract a refined set of feature vectors from the redundant features, using less communication overhead for efficient communication. (ii) Second, we introduce a plug-and-play feature spatial calibration module to mitigate the implications of vehicle localization errors. This module ingeniously utilizes consensus sparse foreground features to align the relative pose relationships between ego-vehicle and collaborators without any precise pose supervision. (iii) Furthermore, we present a pioneering feature fusion and interaction paradigm to integrate holistic spatial semantics. This paradigm comprises two key components: The first is an attention-based feature fusion module that alternates between local and global attention to fuse heterogeneous information from different vehicles. The second is an accuracy enhanced feature interaction strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by fused features. Through these tailored components, ERMVP represents a significant advancement towards pragmatic collaborative perception. To validate the effectiveness of the ER-MVP, we conduct extensive experiments on two collaborative 3D object detection datasets including V2V4Real [44]
+
+and OPV2V [43]. Comprehensive experimental results demonstrate that our method outperforms previous state-of-the-art methods under the bandwidth-limited noisy setting. The main contributions can be summarized as follows:
+
+- We present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method, which addresses the communication bandwidth limitations, localization errors and information aggregation efficiency challenges.  
+- We develop a filter and merge feature sampling strategy to enhance communication efficiency, a feature spatial calibration module for accurate spatial feature alignment, and two information aggregation components to optimize the fusion process.  
+- We conduct extensive experiments on both real-world and simulated datasets. The results show the superiority of our method and the necessity of the proposed components.
+
+# 2. Related Work
+
+# 2.1. Multi-Agent Communication
+
+Communication has played a pivotal role in the development of robust multi-agent systems. Early multi-agent communication [20, 29, 32] relied on predefined protocols and heuristics to regulate communication between agents. However, these fixed methods are unsuitable in complex and dynamic settings. Advances in deep learning have inspired the development of advanced strategies for information exchange and collaboration among agents. For example, MAGIC [27] used a graph attention encoder to solve the
+
+problem of when and to whom to send messages. TMC [49] utilized temporal smoothing to ensure efficient communication. EC-MARL [2] tackled high-dimensional continuous control and partially observable states by introducing a rapid communication protocol to resolve task dilemmas. In comparison, our research focuses on LiDAR-based collaborative 3D object detection tasks in complex driving scenarios. We propose a hierarchical feature sampling strategy to achieve efficient communication across agents.
+
+# 2.2. Collaborative Perception
+
+Factors including limited fields of view of sensors and physical obstructions in the environment can adversely impact the perception capabilities of individual agents. To address these challenges, collaborative perception within multi-agent systems has become a key technology. Utilizing newly available datasets [22, 43, 44], several novel collaborative perception methods have been proposed. For example, V2VNet [38] incorporated a graph neural network to fuse data from different agents. V2X-ViT [42] introduced a transformer architecture that combined information from vehicles and infrastructures. Where2comm [15] used feature spatial heterogeneity to reduce bandwidth utilization by transmitting sparse feature maps. CoBEVT [41] proposed the first multi-camera-based collaborative perception framework and designed the fused axis attention module to enable multi-view interactions. In this work, we propose an innovative feature calibration and interaction method that ensures robust vehicle collaboration in high-noise dynamic environments.
+
+# 3. Method
+
+This section introduces our efficient and robust collaborative perception method. Figure 1 shows the general procedure of our method, which is divided into six phases. We will detail each phase in the following introduction.
+
+# 3.1. Metadata Sharing and Feature Extraction
+
+In multi-vehicle collaboration scenarios, one of the CAVs, known as the ego vehicle, constructs a communication graph. In this graph, the ego vehicle serves as the requester, while other vehicles within its communication range act as supporters. The ego vehicle broadcasts its metadata information, including position, heading, speed, and more. Supporters, upon receiving this metadata, project their local point cloud observations into the ego vehicle's coordinate system. Each vehicle then encodes these transformed point clouds into bird's eye view (BEV) features, yielding a visual representation. Given the  $i$ -th vehicle local observations  $\mathbf{X}_i$ , the extracted features are represented as  $\mathbf{F}_i = \Phi_{\mathrm{enc}}(\mathbf{X}_i) \in \mathbb{R}^{H \times W \times C}$ , where  $\Phi_{\mathrm{enc}}(\cdot)$  denotes the PointPillar [19] encoder shared by all vehicles and  $H$ ,  $W$ ,
+
+and  $C$  stand for the height, width, and channel of the feature map, respectively. Then, these extracted features are fed into the filter and merge feature sampling module.
+
+# 3.2. Filter and Merge Feature Sampling
+
+Previous works have utilized well-designed mechanisms such as information entropy communication selection [37] and spatial heterogeneity map [15, 35] to reduce the required transmission bandwidth. However, these methods primarily focus on the inter-class redundancy between foreground and background features, ignoring the intra-class redundancy among features, which results in sub-optimal compression. To address this gap, we introduce an advanced Filtering and Merging Feature Sampling strategy (FMS). This strategy considers both inter-class and intra-class redundancy relationships, efficiently extracting a concise and distinctive set of feature vectors from the original feature maps, thus reducing communication overheads more effectively. FMS is composed of two core components as follows.
+
+Filter Sampler. In object detection, foreground areas containing objects are more significant than the background areas. Therefore, we implement the idea of reducing spatial redundancy into a feature filter sampler module, aiming to preserve perceptually important yet sparse sets of feature vectors. Since explicitly learning a binary sampler is infeasible, we develop a confidence filter strategy. Initially, a detection confidence map is generated for the feature map. It reflects the perceptual importance of different spatial areas, with higher levels indicating potential object areas and lower levels typically denoting redundant background areas. For the feature map  $\pmb{F}_i$  of the  $i$ -th vehicle, its confidence map  $C_i$  is defined as:
+
+$$
+\boldsymbol {C} _ {i} = \Phi_ {\text {c o n - g e n}} \left(\boldsymbol {F} _ {i}\right) \in [ 0, 1 ] ^ {H \times W}, \tag {1}
+$$
+
+where  $\Phi_{\mathrm{con\_gen}(\cdot)}$  represents the confidence generation network with detection decoder structure. Then the confidence map is thresholded, followed by non-maximum suppression, resulting in a binary mask  $B$ . Utilizing this binary mask, we proceed to preserve the sparse foreground features  $\widetilde{\pmb{F}}_i = B\odot \pmb{F}_i$ . In order to cope with the dynamically changing environmental conditions and fully ensure the robustness of the system, the filtering threshold can be varied with the sensor data and the network state. We set  $\alpha$  as the filter rate and the number of remaining feature vectors is  $L = \alpha \times HW$ .
+
+Merge Sampler. Upon extracting a detailed foreground feature vector set with the filter sampler, we employ the merge sampler for additional optimization, refining similar or repetitive foreground feature vectors through weighted merging. The process is divided into three stages: information-driven feature grouping, attention-inspired feature merging, and index-based feature reconstruction.
+
+(a) Information-Driven Feature Grouping. Initially, a variant of the nearest neighbor clustering algorithm is applied to group the foreground feature vector set. Given a set of feature vector  $\widetilde{\boldsymbol{F}}_i = [x_1,x_2,\dots ,x_L]^\top$  and cluster center  $X_{c}$ , we compute the indicator  $\delta_{i}$  for each feature vector.  $\delta_{i}$  is calculated as the minimum feature distance minus the average pixel distance to any other cluster center vector, expressed as:
+
+$$
+\delta_ {i} = \min  _ {j: x _ {j} \in X _ {c}} \left(\| x _ {i} - x _ {j} \| _ {2} ^ {2} - \gamma \| p (x _ {i}), p (x _ {j}) \| _ {2} ^ {2}\right), \tag {2}
+$$
+
+where  $\delta_{i}$  denotes to which cluster the feature vector  $x_{i}$  should belong.  $p(\cdot)$  means getting the position of the vector and  $\gamma$  is a hyperparameter. Subsequently, we can divide all the feature vectors in  $\widetilde{\boldsymbol{F}}_i$  into  $K$  clusters represented by  $G = \{G_{1},G_{2},\dots ,G_{K}\}$ . The total number of clusters is determined by dynamic clustering ratio  $\beta$ , and is calculated as  $K = \beta \times L$ .
+
+(b) Attention-Inspired Feature Merging. A straightforward strategy for merging feature vectors is to average of each feature vector within a cluster. However, this scheme can be severely affected by outlier feature vectors. Drawing inspiration from attention mechanisms, we utilize the confidence score as a guide to quantify the significance of each feature. Thus, the merged feature vector  $\tilde{x}_i$  for the  $i$ th cluster  $G_i$  is computed as:
+
+$$
+\widetilde {x} _ {i} = \frac {\sum_ {j \in G _ {i}} c _ {j} x _ {j}}{\sum_ {j \in G _ {i}} c _ {j}}, \tag {3}
+$$
+
+where  $c_{j}$  and  $x_{j}$  represent the confidence score and the original feature vector, respectively. In the end, we obtain a final set of feature vectors  $\mathbf{Z}_i = \{\widetilde{x}_1,\widetilde{x}_2,\dots ,\widetilde{x}_K\}$  that is necessary for transmission.
+
+(c) Index-Based Feature Reconstruction. During the feature grouping and merging process, each feature vector is allocated to a cluster, and every cluster is represented by a merged vector. We maintain a record of the index correspondence between original and merged feature vectors. Utilizing this index record, the ego-vehicle ensures that the merged feature vectors are mapped to their corresponding positions, leading to the reconstruction of feature maps.
+
+# 3.3. Feature Spatial Calibration
+
+Localization errors [30, 31] may lead to the misalignment of feature maps among vehicles. Such misalignment causes the ego-vehicle to misinterpret the object's location, resulting in sub-optimal perception, as depicted in the upper right corner of Figure 2. To address this challenge, we introduce the Feature Spatial Calibration module (FSC) to facilitate precise feature alignment, as shown in Figure 2. The core idea of FSC is that overlapping viewpoints allow multiple vehicles to detect partially the same objects. The process
+
+![](images/7d58f4ddb6a0733b531a8909f7caf793b1c101be3a55015892f12e2ce2992a07.jpg)  
+Figure 2. Illustration of the proposed FSC. By employing FSC, misaligned matching regions are corrected at the feature level, leading to more accurate predictions.
+
+involves three phases: consensus matching, geometric verification, and error modulation.
+
+Consensus Matching. After reconstructing the received features, the ego vehicle generates sparse feature maps. Within these maps, areas rich in information signify potential target regions, which correspond to the proposed matching areas. We denote the proposed matching regions of the ego vehicle as  $P$  and those identified by the collaborative vehicle under noisy pose conditions as  $Q$ . Utilizing  $P$  and  $Q$ , a weighted bipartite graph is constructed, wherein the weight of each edge is determined by the distance between the nodes, encapsulated in a cost matrix. The matching process is then converted into a linear assignment task, with the objective of identifying a matching result with the lowest cumulative edge weight. This procedure yields the matching pairs  $M$ .
+
+Geometric Verification. Invalid matches may occur due to objects located in exclusive zones and detection noise. To tackle this issue, we utilize RANSAC to filter and sift a consistent set of matches aligned with expected geometric transformations. Initially, a random matching subset  $M_{s} \subseteq M$  is selected, and then the transformation matrix  $\Gamma_{s}$  is computed using singular value decomposition. When  $\Gamma_{s}$  is applied to all pairs within  $M_{s}$ , and if the posttransformation distance between these pairs remains below the threshold  $\eta$ , the set is considered correctly aligned. The threshold  $\eta$  reflects the allowable localization error within the original collaborative framework. This process is iteratively conducted to identify the optimal transformation matrix that correlates with the maximum number of correct matches. Ultimately, a optimal refined transformation matrix  $\Gamma_{r}$  is obtained and applied in subsequent spatial calibration operations, yielding the aligned feature  $\widetilde{Z}_{j} = \Gamma_{r}Z_{j}$ .
+
+Error Modulation. To enhance the adaptability of the calibration method in various environments, we incorporate an error modulation strategy. This strategy aims to achieve a balance between localization errors and the estimation errors that emerge from the calibration process. It measures the overlap ratios between ego and collaborative features in both their adjusted and original states. Subsequently, fea
+
+![](images/d42d10cfc3c0721cb0a1f26165bbe6d235184a7bc9b1bd0e8df2217e6ae379c9.jpg)  
+Figure 3. A example of local and global Attention in AFF. This shows how AFF processes both 3D local windows and the global query token for position-level and context-aware aggregation.
+
+tures demonstrating the most optimal overlap are selected as the final input.
+
+# 3.4. Attention-based Feature Fusion
+
+In multi-vehicle collaboration scenarios, vehicles are able to capture heterogeneous information from different spatial regions. In order to efficiently fuse perceptual features from multiple vehicles, we propose an Attention-based Feature Fusion method (AFF). As shown in Figure 3, AFF utilizes alternating local and global attention to achieve position-level accurate matching in occlusion-variable traffic scenes and to capture the global semantics attention of road topology and traffic states. Specifically, we stack all the vehicles' features to  $S \in \mathbb{R}^{N \times H \times W \times C}$ , where  $N$  is the number of vehicles. The features are divided into 3D non-overlapping windows, each of size  $N \times P \times P$ . The shape of the partition tensor is  $\left(\frac{H}{P} \times \frac{W}{P}, N \times P^2, C\right)$ . We perform 3D local self attention on the local window and implement region interaction within the window. In contrast, global attention goes beyond the limitations of the local view. It uses the extracted global query token and is shared between all windows to interact with the local key and value representations, thus helping to capture long-range dependencies of the features. Inspired by the efficiency of [13, 33], the global query token  $g$  is generated as follows:
+
+$$
+\mathbf {g} = \mathcal {P} \left(\mathcal {C} _ {1 \times 1} \left(\mathcal {S E} (\mathcal {G} (\mathcal {D} (\mathbf {x}))) + \mathbf {x}\right)\right), \tag {4}
+$$
+
+where  $\mathcal{D}$ ,  $\mathcal{G}$ ,  $\mathcal{SE}$ ,  $\mathcal{C}_{1\times 1}$  and  $\mathcal{P}$  denote the depth-wise separable convolution, GELU activation function, Squeeze-and-Excitation operation,  $1\times 1$  convolution and max pooling, respectively. These designs enable the global query token to provide advantages such as inductive bias and modeling of inter-channel dependencies. We then construct our proposed AFF module by combining this local and global attention with typical designs of Transformers, including LayerNorm [1], MLPs [9], and skip-connections [14]. This module allows the system to analyze spatial correlations from a local view while also capturing global feature responses, ensuring efficient and precise perception in dy
+
+![](images/1f09b47ad9f51cbbee866c451bf65bbf30029ce7fdb5145e4bb85834b8ec315b.jpg)  
+Figure 4. The architecture of the proposed AFI component.
+
+namic, complex, and occlusion-variable traffic scenarios. Finally, we obtain the fused features  $\pmb{H}_i$ .
+
+# 3.5. Accuracy Enhanced Feature Interaction
+
+Previous works [15, 21, 38, 42, 43] have demonstrated that fused features can provide richer semantic information, thereby enhancing perceptual performance. However, they may be affected by collaborative noise, such as asynchronous motion blur and inaccurate projections, which can compromise accurate position information and become a bottleneck to the optimal realization of perceptual performance. Ego-centric features may contain locally critical spatial location information without being affected by collaborative noise. To this end, we propose a novel Accuracy Enhanced Feature Interaction (AEI) strategy that leverages the accurate positional information inherent in ego-centric features to enhance the rich semantic information provided by collaborative fused features. Firstly, we design a context cross attention module that is tailored for feature interaction, as shown in Figure 4. It considers the features from the ego-vehicle as queries, and obtains the context of the query location from the fused features as the key and value to achieve cross-attention. The computed output of the spatial location  $(i,j)$  is as follows:
+
+$$
+y _ {i j} = \sum_ {a, b \in \mathcal {N} _ {k} (i, j)} \operatorname {s o f t m a x} \left(q _ {i j} ^ {\top} k _ {a b}\right) v _ {a b}, \tag {5}
+$$
+
+where  $\mathcal{N}_k(i,j)$  is context position of  $(i,j)$  and  $q_{ij}, k_{ab}, v_{ab}$  represent the ego query, context key and value, respectively. As the relevant features of the same object are often located in similar locations across different feature maps, we prioritize the local context details of the query location. Additionally, considering that only a few positions within the sparse feature map hold significant information, focusing on these context-rich areas is computationally efficient. Following this, we employ the split attention [48] to adaptively fuse information from multiple branches, generating the enhanced output feature  $\widetilde{H}_i$ .
+
+<table><tr><td>Dataset</td><td colspan="3">V2V4Real</td><td colspan="3">OPV2V</td></tr><tr><td>Noise Level σt/σr (m/°)</td><td>0.0/0.0</td><td>0.4/0.4</td><td>0.8/0.8</td><td>0.0/0.0</td><td>0.4/0.4</td><td>0.8/0.8</td></tr><tr><td>No Collaboration</td><td>41.64/23.43</td><td>41.64/23.43</td><td>41.64/23.43</td><td>73.25/58.22</td><td>73.25/58.22</td><td>73.25/58.22</td></tr><tr><td>Late Fusion</td><td>56.80/27.62</td><td>40.81/13.31</td><td>26.65/10.66</td><td>86.46/79.21</td><td>76.56/40.01</td><td>45.57/21.00</td></tr><tr><td>F-Cooper [3]</td><td>61.12/32.36</td><td>51.71/24.48</td><td>41.90/19.91</td><td>87.39/79.39</td><td>79.33/40.12</td><td>52.91/16.90</td></tr><tr><td>AttFuse [7]</td><td>64.41/34.32</td><td>58.42/30.02</td><td>49.37/25.44</td><td>90.50/81.80</td><td>83.54/52.43</td><td>68.37/40.96</td></tr><tr><td>Where2comm [15]</td><td>63.69/34.79</td><td>58.48/28.09</td><td>49.77/24.25</td><td>90.46/84.22</td><td>82.17/56.47</td><td>73.95/45.86</td></tr><tr><td>DiscoNet [21]</td><td>64.19/35.25</td><td>57.71/28.41</td><td>49.19/24.94</td><td>89.58/81.45</td><td>83.25/53.22</td><td>61.37/30.53</td></tr><tr><td>V2VNet [38]</td><td>65.70/35.38</td><td>62.11/30.63</td><td>54.73/25.57</td><td>91.35/82.43</td><td>85.43/54.16</td><td>70.43/31.28</td></tr><tr><td>CoAlign [25]</td><td>64.06/36.60</td><td>59.58/31.52</td><td>53.02/26.60</td><td>90.93/84.28</td><td>84.48/57.45</td><td>73.95/48.86</td></tr><tr><td>V2X-ViT [42]</td><td>66.42/37.34</td><td>63.33/32.38</td><td>57.15/28.22</td><td>91.74/83.31</td><td>82.91/54.73</td><td>61.70/26.07</td></tr><tr><td>CoBEVT [41]</td><td>66.01/37.36</td><td>58.63/29.09</td><td>49.22/23.66</td><td>91.71/85.98</td><td>85.49/60.64</td><td>64.63/29.59</td></tr><tr><td>ERMVP (Ours)</td><td>67.66/42.97</td><td>65.01/40.45</td><td>60.88/35.01</td><td>92.18/85.59</td><td>85.67/66.32</td><td>77.13/59.15</td></tr></table>
+
+Table 1. Overall performance on V2V4Real and OPV2V datasets with pose noises. The results are reported in AP@0.5/0.7.
+
+# 3.6. Decoder and Loss
+
+Based on the final fused feature  $\widetilde{H}_i$ , we use the detection decoder to generate the final prediction output  $Y_{i} = \Phi_{\mathrm{dec}}(\widetilde{H}_{i})$ . Each position of  $Y_{i}$  represents a rotated box with classes (c, x, y, h, w,  $\cos \alpha$ ,  $\sin \alpha$ ), corresponding to class confidence, position, size, and angle. These objects are the final output of the proposed collaborative perception system. Following existing work [19], we adopt the smooth  $L_{1}$  loss for regression and focal loss [23] for classification.
+
+# 4. Experiments
+
+# 4.1. Datasets and Evaluation Metrics
+
+Datasets. We validate the proposed ERMVP in the task of LiDAR-based 3D object detection on two benchmark datasets including V2V4real [44] and OPV2V [43]. V2V4real [44] is the first large-scale real-world dataset on V2V perception, collected by two vehicles equipped with multi-modal sensors driving together in a variety of scenarios. It covers a driving area of 410 kilometers, including 20K frames of point clouds, with training/validation/test sets containing 14210/2000/3986 frames, respectively. OPV2V [43] is a large-scale vehicle-to-vehicle collaborative perception dataset, simulated by OpenCDA [40] and Carla [8]. It contains 73 different scenarios, in which different numbers (2 to 7) of collaborative vehicles appear together, each equipped with a lidar sensor and 4 cameras. The dataset contains a total of 11,464 frames of point clouds and RGB images. The training/validation/test sets contain 6374, 1980, and 2170 frames, respectively.
+
+Evaluation Metrics. We adopt the Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 and 0.7 to evaluate the detection performance. The calculation format of communication volume in [15] is used to count the message size by byte in the log scale with base 2.
+
+# 4.2. Implementation Details
+
+We implement the proposed ERMVP and comparison model on the Pytorch toolbox [28], and train them on a NVIDIA GeForceRTX 3090 GPU using the Adam optimizer [18]. The initial learning rate is  $2\mathrm{e} - 3$  and it decays every 15 epochs with a factor of 0.1. All models are trained with 60 epochs and the batch size is set to 2. Early stopping is used to find the best epoch. We also add normal point cloud data augmentation for all experiments, including scaling, rotation, and flipping. All detection models are based on PointPillar [19] backbone to extract 2D features from point clouds and a  $0.4\mathrm{m}$  width/length is used for each voxel. AFF component has 3 encoded layers and a window size of 8 for both local and global attention. The balance hyperparameter  $\gamma$  is 0.1 and the error tolerance threshold  $\eta$  is 0.25. Each vehicle has a communication range of 70 m based on [42], while vehicles outside of this broadcasting radius will be ignored. To simulate the localization and heading errors, we add Gaussian noise with a standard deviation of  $\sigma_t$  for localization errors and  $\sigma_r$  for heading errors.
+
+# 4.3.Quantitative Evaluation
+
+Comparison of Detection Performance. Table 1 shows the performance comparison results of 3D detection on two datasets. We use the No Collaboration method as a baseline, which only uses ego vehicle's point clouds data without collaboration. Late Fusion allows vehicles to exchange detected outputs and utilizes non-maximum suppression to produce the final result. For intermediate fusion strategies, we evaluate the existing state-of-the-art (SOTA) methods: AttFuse [43], F-Cooper [3], V2VNet [38], DiscoNet [21], V2X-ViT [42], CoAlign (only fusion) [25], CoBEVT [41] and Where2comm [15]. The proposed ER-MVP achieves average improvements of  $12.48\% / 18.47\%$  and  $6.18\% / 10.02\%$  on two datasets in AP@0.5/0.7 compared with No Collaboration and Late Fusion methods, demonstrating the superiority of our collaborative method.
+
+(a) V2V4Real  
+![](images/a05276fe44b9498839d47f184cc6c82d59fe0ebf90f874dbf5bd3a548c86c415.jpg)  
+Late Fusion FCooper V2VNet V2X-VIT CoBEVT -Where2comm -ERMVP
+
+![](images/0d727aed98d8c717ad512713bc262dc41c0e12fc4ed41f3bfc943faddb1c8b02.jpg)  
+Figure 5. Collaborative perception performance comparison on the V2V4Real and OPV2V with varying communication volume.
+
+![](images/f88bea390d557ccfa8148f6e1357029b30be1ff3a78f8f3b8d9d1900be3dafa5.jpg)
+
+![](images/517d86dc9d3835c14d405265d54458bd70a95d2611f426bab979bfa02ba7bde2.jpg)
+
+Meanwhile, it is superior to the SOTA collaborative perception method in both real-world and simulated scenarios: improves the SOTA collaborative performance by  $1.24\%$  and  $5.61\%$  on V2V4Real datasets in AP@0.5/0.7, and by  $0.44\%$  on OPV2V datasets in AP@0.5. These results fully demonstrate the superiority of the ERMVP collaboration paradigm.
+
+<table><tr><td>Dataset</td><td colspan="2">V2V4Real</td><td colspan="2">OPV2V</td></tr><tr><td>Noise Level σt(m)</td><td>0.5</td><td>1.0</td><td>0.5</td><td>1.0</td></tr><tr><td>Late Fusion</td><td>11.02</td><td>10.13</td><td>53.88</td><td>38.80</td></tr><tr><td>F-Cooper [3]</td><td>24.60</td><td>18.41</td><td>32.24</td><td>12.71</td></tr><tr><td>AttFuse [43]</td><td>28.10</td><td>24.36</td><td>42.95</td><td>35.53</td></tr><tr><td>DiscoNet [21]</td><td>28.36</td><td>24.05</td><td>45.36</td><td>24.96</td></tr><tr><td>V2VNet [38]</td><td>29.57</td><td>23.67</td><td>47.48</td><td>25.82</td></tr><tr><td>Where2comm [15]</td><td>28.02</td><td>23.53</td><td>46.59</td><td>45.59</td></tr><tr><td>CoAlign [21]</td><td>29.84</td><td>25.82</td><td>49.09</td><td>46.66</td></tr><tr><td>V2X-ViT [42]</td><td>30.99</td><td>23.67</td><td>44.83</td><td>21.77</td></tr><tr><td>CoBEVT [41]</td><td>24.67</td><td>22.01</td><td>41.91</td><td>26.05</td></tr><tr><td>ERMVP (Ours)</td><td>39.26</td><td>33.94</td><td>77.25</td><td>76.09</td></tr></table>
+
+Robustness to Localization and Heading Errors. To evaluate the sensitivity of existing methods to localization and heading errors, we use the same noise setting as [42] and conduct extensive experiments on two datasets. As shown in Tables 1 and 2, under ideal settings, Late Fusion and some advanced intermediate fusion methods can detect object vehicles with high accuracy, but their detection accuracy rapidly decreases as the standard deviation of errors increases. For example, when the localization and heading errors of OPV2V dataset are over  $0.4\mathrm{m}$  and  $0.4^{\circ}$ , the Disconet [21] and F-Cooper [3] methods fail and perform even worse than No Collaboration. Our method exceeds previous SOTA models at all noise levels and consistently outperforms the No Collaboration baseline, clearly demonstrating the robustness of ERMVP to pose errors. It is noteworthy that for localization errors, ERMVP can maintain consistent
+
+accuracy levels in high noise environments, showing superior robustness. When the localization error is  $1.0\mathrm{m}$ , it outperforms the second-best method by  $29.43\%$  and  $8.08\%$  in AP@0.7 on OPV2V and V2V4Real datasets. The reasonable explanations are: (i) feature spatial calibration aligns mismatched collaborative features; (ii) accuracy enhanced feature interaction reduces performance degradation caused by collaborative noise.
+
+Comparison of Communication Volume. To achieve pragmatic collaborative perception, it is crucial to evaluate the perceptual performance under different communication volumes. Figure 5 shows the results of the performance comparison under different bandwidth consumption conditions. Note that we do not consider any additional model/data/feature compression for a fair comparison. The results indicate that ERMVP achieves excellent perceptual performance and bandwidth consumption trade-off under all communication bandwidth conditions, consistently outperforming Where2comm [15]. At the same time, it achieves the same detection performance as the previous model [3, 38, 41, 42] with less communication volume on both real-world and simulated datasets.
+
+# 4.4. Ablation Studies
+
+Table 2. Detection AP@0.7 on V2V4Real and OPV2V datasets with localization error.  
+
+<table><tr><td>AFF</td><td>FSC</td><td>AEI</td><td>V2V4Real</td><td>OPV2V</td></tr><tr><td></td><td></td><td></td><td>51.71/24.48</td><td>79.33/40.12</td></tr><tr><td>✓</td><td></td><td></td><td>62.24/37.70</td><td>84.30/61.43</td></tr><tr><td>✓</td><td>✓</td><td></td><td>64.36/39.59</td><td>84.98/62.67</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>65.01/40.45</td><td>85.67/66.32</td></tr></table>
+
+Table 3. Ablation study results of the proposed core components on the both datasets with noise level of 0.4/0.4. AFF: Attention-based Feature Fusion; FSC: Feature Spatial Calibration; AEI: Accuracy Enhanced Feature Interaction.
+
+Effect of Core Components. Table 3 details the contribution of each core components in our ERMVP framework. The base model is the naive position-wise maximum fusion method. We then assess the impact of each component by sequentially introducing i) attention-based feature
+
+![](images/280d2b5c3bb6552ed83f6681ca5e9ab65f17fefd5ae8bbd0b8c2945e3095852f.jpg)
+
+![](images/e770195a486e1bf921acc083d6a82eecd9c532e4557f8586c9120e99f4b1c3dd.jpg)  
+(a) V2VNet
+
+![](images/efd2f53dc9a564fb9dde0a5412ab504de41b34a537d009f58505a6de2f4afd40.jpg)
+
+![](images/492842c4ba1259fbc06256306e088f9a970ff6743f2cbd8d52a3c6f44fd3ba2c.jpg)  
+(b) CoBEVT
+
+![](images/aefb2e149c7c44ff1a1486876c71877e2004bf57de183a250fdfb7e67f496884.jpg)
+
+![](images/f4481015af92b97fb51c9f6721b939b1b571c6f3d179fb796d81053fc607f9bc.jpg)  
+(c) V2X-ViT  
+Figure 6. Qualitative comparison results in real-world scenarios from the V2V4Real dataset with noise level of 0.4/0.4. Green and red boxes denote the ground truths and detection results, respectively.
+
+![](images/bba47526b06f4a957dc1c460977993b17132c871bd63b9b4bf3549c7e3f00d7f.jpg)
+
+![](images/cdb2cc3bba2d1bd3841d0f2755d0b7379f1b9c64db023463ded86abc01c94515.jpg)  
+(d)ERMVP
+
+fusion (AFF), ii) feature spatial calibration (FSC), iii) accuracy enhanced feature interaction (AEF). The consistent rise in detection results over both datasets demonstrates the effectiveness of each introduced component. Notably, integrating all three components boosts detection performance by  $13.3\%$  and  $15.97\%$  on the V2V4Real dataset for AP@0.5 and AP@0.7, respectively.
+
+![](images/b92763145d0f8cd9c12d9dd5457a2982c98ae21f38860f7e31289167e94214e1.jpg)  
+Figure 7. Hyperparameter analysis of  $\alpha$  and  $\beta$
+
+![](images/d71922f3b4b2034784cf84180fa4aae8983c8e9ae7ee30960da71fb8f0ab96ef.jpg)
+
+Hyperparameters  $\alpha$  and  $\beta$ . We explore the impact of the hyperparameters  $\alpha$  and  $\beta$  (as described in Section 3.2), which jointly determine the communication volume. Specifically, the feature compression ratio is represented as  $\alpha \times \beta$ . The communication volume, in turn, is expressed as the product of the original feature size and the compression rate. By modulating the values of  $\alpha$  and  $\beta$  within the interval of  $10\%$  to  $80\%$ , we chart the resultant object detection performance (measured by AP@0.7) in Figure 7. As expected, a drop in either  $\alpha$  or  $\beta$  will impair ERMVP's efficacy. Yet, it's worth noting that ERMVP's performance generally exhibits robustness to these hyperparameters. The performance decline remains fairly minimal even under significant compression, such as a factor of 100. The results demonstrate the effectiveness of our proposed filter and merge feature sampling module. Therefore, ERMVP establishes an appropriate and effective communication paradigm among vehicles, ensuring an optimal tradeoff between perception and communication.
+
+# 4.5. Qualitative Evaluation
+
+Figure 6 displays the detection results of V2VNet [38], CoBEVT [41], V2X-ViT [42], and ERMVP in two scenarios. Clearly, ERMVP offers more precise and comprehensive detection than the prior SOTA methods. To begin with, ERMVP produces a greater number of predicted bounding boxes that are well aligned with the ground truths, while other methods show significant discrepancies. This demonstrates ERMVP's robustness, particularly in high-noise environments. Moreover, ERMVP detects more dynamic objects (more ground truth bounding boxes find matches). This suggests that ERMVP can effectively combine the inputs from nearby vehicles, leading to a thorough scene representation. Overall, these qualitative assessments confirm ERMVP's strengths in delivering accurate and comprehensive perception, especially in challenging conditions.
+
+# 5. Conclusion
+
+In this paper, we present ERMVP, a communication-efficient and collaboration-robust multi-vehicle perception method in challenging environments. We introduce a filter and merge feature sampling strategy to efficient communication, a feature spatial calibration module to align features in fine-grid and two spatial information aggregation components to solve the bottleneck of the fusion method. Extensive experiments prove the superiority of ERMVP and the effectiveness of our components. The feasibility of our approach for other modalities will be explored in future work.
+
+# Acknowledgments
+
+This work is supported by the Shanghai Key Research Laboratory of NSAI, the Innovation Platform for Academicians of Hainan Province, Haikou, 570228, the Specific Research Fund of the Innovation Platform for Academicians of Hainan Province under Grant YSPTZX202314, and NFSC under grant 62250410368.
+
+# References
+
+[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5  
+[2] Marwa Chafii, Salmane Naoumi, Reda Alami, Ebtesam Al-mazrouei, Mehdi Bennis, and Merouane Debbah. Emergent communication in multi-agent reinforcement learning for future wireless networks. arXiv preprint arXiv:2309.06021, 2023. 3  
+[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 6, 7  
+[4] Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: deformable patch attack. In European conference on computer vision, pages 529-548. Springer, 2022. 1  
+[5] Zhaoyu Chen, Bo Li, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Query-efficient decision-based black-box patch attack. IEEE Transactions on Information Forensics and Security, 2023.  
+[6] Zhaoyu Chen, Bo Li, Shuang Wu, Kaixun Jiang, Shouhong Ding, and Wenqiang Zhang. Content-based unrestricted adversarial attack. Advances in Neural Information Processing Systems, 36, 2024. 1  
+[7] Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional feature fusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3560-3569, 2021. 6  
+[8] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6  
+[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5  
+[10] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016. 1  
+[11] Jiaming Gu, Jingyu Zhang, Muyang Zhang, Weiliang Meng, Shibiao Xu, Jiguang Zhang, and Xiaopeng Zhang. Feaco: Reaching robust feature-level consensus in noisy pose conditions. In Proceedings of the 31st ACM International Conference on Multimedia, page 3628-3636, New York, NY, USA, 2023. Association for Computing Machinery. 1  
+[12] Monowar Hasan, Sibin Mohan, Takayuki Shimizu, and Hongsheng Lu. Securing vehicle-to-everything (v2x) communication platforms. IEEE Transactions on Intelligent Vehicles, 5(4):693-713, 2020. 1  
+[13] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In International Conference on Machine Learning, pages 12633-12646. PMLR, 2023. 5
+
+[14] 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. 5  
+[15] Yue Hu, Shaoheng Fang and Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 5, 6, 7  
+[16] Kaixun Jiang, Zhaoyu Chen, Hao Huang, Jiafeng Wang, Dingkang Yang, Bo Li, Yan Wang, and Wenqiang Zhang. Efficient decision-based black-box patch attacks on video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4379-4389, 2023. 1  
+[17] Hamza Khan, Sumudu Samarakoon, and Mehdi Bennis. Enhancing video streaming in vehicular networks via resource slicing. IEEE Transactions on Vehicular Technology, 69(4): 3513-3522, 2020. 1  
+[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6  
+[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 6  
+[20] Yiming Li, Bir Bhanu, and Wei Lin. Auction protocol for camera active control. In 2010 IEEE International Conference on Image Processing, pages 4325-4328. IEEE, 2010. 2  
+[21] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 5, 6, 7  
+[22] Yiming Li, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: A virtual collaborative perception dataset for autonomous driving. arXiv preprint arXiv:2202.08449, 2022. 3  
+[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 6  
+[24] Yang Liu, Jing Liu, Kun Yang, Bobo Ju, Siao Liu, Yuzheng Wang, Dingkang Yang, Peng Sun, and Liang Song. Ampnet: Appearance-motion prototype network assisted automatic video anomaly detection system. IEEE Transactions on Industrial Informatics, 20(2):2843-2855, 2024. 1  
+[25] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 1, 6  
+[26] Zonglin Meng, Xin Xia, Runsheng Xu, Wei Liu, and Jiaqi Ma. Hydro-3d: Hybrid object detection and tracking for cooperative perception using 3d lidar. IEEE Transactions on Intelligent Vehicles, 2023. 1
+
+[27] Yaru Niu, Rohan R Paleja, and Matthew C Gombolay. Multiagent graph-attention communication and teaming. In AA-MAS, pages 964–973, 2021. 2  
+[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6  
+[29] Faisal Qureshi and Demetri Terzopoulos. Smart camera networks in virtual reality. Proceedings of the IEEE, 96(10): 1640-1656, 2008. 2  
+[30] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3365-3374, 2021. 4  
+[31] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. An efficient and robust object-level cooperative perception framework for connected and automated driving. arXiv preprint arXiv:2210.06289, 2022. 4  
+[32] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth international conference on machine learning, pages 330-337, 1993. 2  
+[33] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International conference on machine learning, pages 10096-10106. PMLR, 2021. 5  
+[34] Nicholas Vadivelu, Mengye Ren, James Tu, Jingkang Wang, and Raquel Urtasun. Learning to communicate and correct pose errors. In Conference on Robot Learning, pages 1195-1210. PMLR, 2021. 1  
+[35] Binglu Wang, Lei Zhang, Zhaozhong Wang, Yongqiang Zhao, and Tianfei Zhou. Core: Cooperative reconstruction for multi-agent perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8710-8720, 2023. 1, 3  
+[36] Jian Wang, Yameng Shao, Yuming Ge, and Rundong Yu. A survey of vehicle to everything (v2x) testing. Sensors, 19(2): 334, 2019. 1  
+[37] Tianhang Wang, Guang Chen, Kai Chen, Zhengfa Liu, Bo Zhang, Alois Knoll, and Changjun Jiang. Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework. arXiv preprint arXiv:2303.12400, 2023. 3  
+[38] Tsun-Hsuan Wang, Sivabanan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 1, 3, 5, 6, 7, 8  
+[39] Qianyi Wu, Kaisiyuan Wang, Kejie Li, Jianmin Zheng, and Jianfei Cai. Objectsdf++: Improved object-compositional neural implicit surfaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21764-21774, 2023. 1
+
+[40] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6  
+[41] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers, 2022. 1, 3, 6, 7, 8  
+[42] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European Conference on Computer Vision, page 107-124. Springer, 2022. 1, 3, 5, 6, 7, 8  
+[43] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3, 5, 6, 7  
+[44] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 3, 6  
+[45] Kun Yang, Peng Sun, Jieyu Lin, Azzedine Boukerche, and Liang Song. A novel distributed task scheduling framework for supporting vehicular edge intelligence. In 2022 IEEE 42nd International Conference on Distributed Computing Systems (ICDCS), pages 972-982. IEEE, 2022. 1  
+[46] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception, 2023. 1  
+[47] Kun Yang, Dingkang Yang, Jingyu Zhang, Hanqi Wang, Peng Sun, and Liang Song. What2comm: Towards communication-efficient collaborative perception via feature decoupling. In Proceedings of the 31st ACM International Conference on Multimedia, page 7686-7695, New York, NY, USA, 2023. Association for Computing Machinery. 1  
+[48] Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R. Manmatha, Mu Li, and Alexander Smola. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2736-2746, 2022. 5  
+[49] Sai Qian Zhang, Qi Zhang, and Jieyu Lin. Succinct and robust multi-agent communication with temporal message control. Advances in Neural Information Processing Systems, 33:17271-17282, 2020. 3

2024/ERMVP_ Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e939236d3cfabb74b721c68771c8ca4ad916b2610319aef2fd107ebc229a78fe
+size 524448

2024/ES3_ Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations/09b7366b-807f-4dfb-a9c8-45edf9472928_content_list.json

@@ -0,0 +1,1590 @@
+[
+    {
+        "type": "text",
+        "text": "$\\mathbf{ES}^{3}$ : Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations",
+        "text_level": 1,
+        "bbox": [
+            243,
+            128,
+            725,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yuanhang Zhang $^{1,2}$ , Shuang Yang $^{1,2}$ , Shiguang Shan $^{1,2}$ , Xilin Chen $^{1,2}$",
+        "bbox": [
+            209,
+            202,
+            759,
+            220
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>Key Laboratory of Intelligent Information Processing of Chinese Academy of Sciences (CAS), Institute of Computing Technology, CAS, Beijing 100190, China",
+        "bbox": [
+            107,
+            220,
+            861,
+            256
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ University of Chinese Academy of Sciences, Beijing 100049, China",
+        "bbox": [
+            212,
+            256,
+            756,
+            273
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "zhangyuanhang15@mails.ucas.ac.cn, {shuang.yang, sgshan, xlchen}@ict.ac.cn",
+        "bbox": [
+            183,
+            276,
+            781,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            325,
+            313,
+            343
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We propose a novel strategy,  $ES^3$ , for self-supervised learning of robust audio-visual speech representations from unlabeled talking face videos. While many recent approaches for this task primarily rely on guiding the learning process using the audio modality alone to capture information shared between audio and video, we reframe the problem as the acquisition of shared, unique (modality-specific) and synergistic speech information to address the inherent asymmetry between the modalities. Based on this formulation, we propose a novel \"evolving\" strategy that progressively builds joint audio-visual speech representations that are strong for both uni-modal (audio & visual) and bimodal (audio-visual) speech. First, we leverage the more easily learnable audio modality to initialize audio and visual representations by capturing audio-unique and shared speech information. Next, we incorporate video-unique speech information and bootstrap the audio-visual representations on top of the previously acquired shared knowledge. Finally, we maximize the total audio-visual speech information, including synergistic information to obtain robust and comprehensive representations. We implement  $ES^3$  as a simple Siamese framework and experiments on both English benchmarks and a newly contributed large-scale Mandarin dataset show its effectiveness. In particular, on LRS2-BBC, our smallest model is on par with SoTA models with only 1/2 parameters and 1/8 unlabeled data (223h).",
+        "bbox": [
+            76,
+            358,
+            472,
+            752
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            768,
+            209,
+            786
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The task of learning audio-visual speech representations has gained increasing importance due to its potential to support a wide range of relevant applications. Building on the success in other domains [9, 11, 12, 14, 22, 31], several effective self-supervised approaches have recently emerged for this task by pre-training models on paired or mixed audio, visual and text data with pretext tasks like speech unit",
+        "bbox": [
+            75,
+            794,
+            470,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "prediction [30, 59, 80]. However, learning representations in the audio-visual speech domain faces unique challenges brought by the inherent asymmetry between the two modalities involved. First, audio and video often convey distinct and complementary types of speech information. For example,  $\\mathsf{m} /$  and  $\\mathsf{n} /$  sound similar, but are visually different; on the contrary,  $\\mathsf{p} /$  and  $\\mathsf{b} /$  are visually similar but can be distinguished with audio [70]. Even more interesting is the presence of synergistic information – information that can further enhance speech perception [55] but only emerges when audio and video are considered together, and cannot be inferred when either modality is absent. Second, they manifest different amounts of noise and information. Specifically, in most existing audio-visual speech datasets, audio is characterized by high sampling rate and relatively clean speech samples, making audio representations readily learnable from these datasets. In contrast, the visual modality often presents several challenging factors, including missing (e.g. due to self-occlusion in profile faces or low image quality), redundant (e.g. due to appearance variations, makeup and beards etc.) and ambiguous (e.g. due to the lower sampling rate and the presence of homophones) visual information. Their presence complicates learning meaningful and robust representations solely from video. The difficulties encountered in the continuous efforts to improve the performance of visual speech recognition (VSR) [15, 24, 40, 41] models illustrate this persistent obstacle. Additionally, these differences between the two modalities lead to an imbalance between their learning dynamics, requiring additional manual settings [3, 39, 42, 61].",
+        "bbox": [
+            496,
+            327,
+            893,
+            782
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this work, we break down the problem of learning audio-visual speech representations into the learning of three different types of information: shared (across modalities), unique (modality-specific) and synergistic speech information. Based on our formulation, we propose a novel strategy called  $\\mathrm{ES^3}$  (short for Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations)",
+        "bbox": [
+            496,
+            794,
+            893,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "CVF",
+        "bbox": [
+            106,
+            2,
+            181,
+            42
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "header",
+        "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+        "bbox": [
+            236,
+            0,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "27069",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "to tackle the asymmetry between the two modalities for learning robust audio-visual speech representations. Specifically, our  $\\mathrm{ES}^3$  employs a three-stage progressive learning approach. Initially, it focuses on acquiring the easily learnable audio-unique and shared speech information. Then, it addresses the more challenging visual-unique speech information while incorporating it into the initial audio-visual joint representation. Lastly, it optimizes the final audio-visual representations by leveraging synergy, on top of the acquired shared and modality-unique information.",
+        "bbox": [
+            75,
+            90,
+            472,
+            242
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions can be summarized as follows:",
+        "bbox": [
+            96,
+            243,
+            421,
+            257
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose a new formulation for audio-visual speech representation learning by decomposing the task into the acquisition of shared (across modalities), unique (modality-specific) and synergistic speech information.",
+            "- We propose a novel strategy,  $\\mathrm{ES}^3$  to progressively learn the three types of information in a self-supervised manner with a simple Siamese framework. This evolving strategy significantly alleviates the previous burden of tuning balancing parameters.",
+            "- Extensive experiments on two widely used benchmarks show that  $\\mathrm{ES}^3$  rivals previous methods in auditory, visual and audio-visual speech recognition performance.",
+            "- We also contribute a new large-scale Mandarin audiovisual speech dataset named CAS-VSR-S101 to the community for evaluation of similar methods."
+        ],
+        "bbox": [
+            76,
+            258,
+            468,
+            484
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            496,
+            218,
+            511
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Self-supervised learning from multi-modal data. The arguably most successful paradigm for self-supervised multimodal learning in the past few years is contrastive learning, which exploits the common information encoded in different modalities [68], such as audio (A) [6, 7], image (I) [53, 66], video (V) [5, 8, 32, 67] and text (T) [4, 16, 43, 44, 49, 63]. These approaches focus on learning semantically aligned multi-modal representations that are effective for specific downstream tasks like cross-modal retrieval, but discard modality-unique information that is useful for other uni-modal tasks. To solve this problem, Gong et al. [25] combine contrastive learning with the masked prediction objective [28] to learn strong audio-visual representations suitable for both retrieval and joint classification. Liang et al. [35] propose a new algorithm, FACTORCL to learn self-supervised multi-modal representations that capture both shared and unique information and verify its effectiveness on multiple tasks. In this paper, we explicitly address capturing unique, shared and synergistic information for uni- and multi-modal downstream tasks.",
+        "bbox": [
+            75,
+            521,
+            472,
+            823
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Learning audio-visual speech representations. A main challenge of learning audio-visual speech representation is properly handling the aforementioned asymmetry between audio and video. An established approach is to leverage the easily learnable audio modality as a teacher to guide",
+        "bbox": [
+            75,
+            825,
+            472,
+            902
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "visual or audio-visual speech representation learning. Implementing additional measures to mitigate audio dominance and promote visual features during learning is vital for this paradigm. For instance, Shi et al. [59] and Zhang et al. [76] use different masking strengths for the two modalities, apply modality dropout, and enforce simple audio features by only using a weak linear projection layer. Others have explored freezing pre-trained audio representations [2, 38, 78]. However, such an audio-centric approach is sub-optimal and inefficient. First, it limits the learned representations to capturing only shared, i.e. modality-invariant information, or at most, shared and audio-unique information. Recently, a few attempts have been made to go beyond merely using audio as the anchor modality. Haliassos et al. [27] combine multiple audio-visual prediction tasks in place of purely audio-centric ones, and treats A-V asymmetry by having the visual stream predict only auditory but not visual targets. Lian et al. [34] perform a series of trials to determine an optimal schedule for audio and video input combination. Second, the process of exhaustively testing and identifying the optimal set of balancing parameters such as masking ratios, modality schedules and modality dropout rate is extremely time-consuming and computationally demanding. In contrast, our newly proposed evolving strategy identifies different types of task-relevant information and learns them progressively, while avoiding pitfalls and computational burdens (see supp).",
+        "bbox": [
+            496,
+            90,
+            893,
+            500
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3. Our Proposed ES<sup>3</sup>",
+        "text_level": 1,
+        "bbox": [
+            500,
+            503,
+            679,
+            522
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1. Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            500,
+            530,
+            640,
+            544
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The goal of  $\\mathrm{ES}^3$  is to learn audio-visual speech representations from unlabeled talking face videos. The inputs to our  $\\mathrm{ES}^3$  are paired audio-visual data  $X_{a,v} = \\{(X_{a,i},X_{v,i})\\mid i = 1,2,\\ldots ,N\\}$ . Each of the  $N$  samples consists of a sequence of audio frames  $X_{a,i} = \\operatorname {LogFBank}(\\mathbf{a}_i)\\in \\mathbb{R}^{F\\times T_{a,i}}$  and a temporally synchronized video sequence  $X_{v,i} = \\mathbf{v}_i\\in \\mathbb{R}^{C\\times H\\times W\\times T_{v,i}}$  showing the lips of a speaker, where  $F$  is the filter bank dimension,  $C$  is the number of video channels,  $T_{a,i}$  and  $T_{v,i}$  are the total number of audio and video frames in the  $i$ -th sample, and  $H$  and  $W$  are the width and height of each frame. Audio and video are sampled at 100 and 25 FPS, i.e. for a 1-second clip,  $T_{a} = 4T_{v} = 100$ . The goal is to learn a  $d$ -dimensional contextualized representation  $\\gamma \\in \\mathbb{R}^{d\\times T_v}$  capturing vital speech and context information within the two modality spaces.",
+        "bbox": [
+            496,
+            551,
+            893,
+            779
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Base architecture. The base architecture used in this paper is a relatively simple one and can be easily extended if needed. It includes two modality-specific encoders  $\\varphi_{a}$  and  $\\varphi_{v}$ , instantiated as a 2-layer CNN-based audio filter bank encoder and a lightweight VGG-(2+1)D video encoder [57, 58]. Here,  $\\varphi_{a}$  downsamples  $X_{a}$  by 4x to match the video frame rate;  $\\varphi_{v}$  has a temporal receptive field of 5 frames, or 200ms at 25 FPS, same as the most com",
+        "bbox": [
+            496,
+            780,
+            893,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "27070",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "monly adopted 3D-ResNet18 encoder [41, 59, 64], but has much fewer parameters. A modality-agnostic Transformer-based [71] modality fusion encoder  $\\theta$  equipped with rotary positional embeddings (RoPE) [65] follows the modality-specific encoders  $\\varphi_{a}$  and  $\\varphi_{v}$ . To allow  $\\theta$  to accept either audio, video, or audio-visual features, we adopt the simple sum operation to compute a fused audio-visual representation  $F_{\\mathrm{fused}}$  from modality features  $F_{m} = \\varphi_{m}(X_{m})$ , where  $m \\in \\{a, v\\}$ , to allow easy handling of missing modalities:",
+        "bbox": [
+            75,
+            90,
+            472,
+            227
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {\\text {f u s e d}} = \\left\\{ \\begin{array}{l l} F _ {a} + F _ {v} & \\text {i f i n p u t i s a u d i o - v i s u a l ,} \\\\ F _ {a} + \\mathbf {0} = F _ {a} & \\text {i f i n p u t i s a u d i o ,} \\\\ F _ {v} + \\mathbf {0} = F _ {v} & \\text {i f i n p u t i s v i d e o .} \\end{array} \\right. \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            94,
+            234,
+            468,
+            289
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The Transformer encoder  $\\theta$  then yields the final contextual features  $E = \\theta (F_{\\mathrm{fused}})$",
+        "bbox": [
+            75,
+            296,
+            468,
+            325
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Based on this base architecture, we build our Siamese based learning framework as shown in Fig. 1(a). A student model (left) is guided by a teacher model (right), which share the same base architecture. The student and teacher accept different versions of the same speech inputs (e.g. masked vs unmasked). The teacher model updates its parameters in an exponential moving average (EMA) manner, and the student is trained to regress the teacher outputs in general. In the subsequent sections, our objective is to optimize the student model with the base architecture to obtain the parameterized audio-visual speech representations  $\\gamma (\\mathbf{a},\\mathbf{v};\\varphi_{a},\\varphi_{v},\\theta) = E$",
+        "bbox": [
+            75,
+            325,
+            470,
+            508
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Problem Reformulation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            513,
+            295,
+            529
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the audio-visual speech representation learning task, audio and video inputs can be taken as two different views of the same underlying speech signal, carrying different types of information. The relationship between the speech information in the two modalities can be visualized as the bottom-left Venn diagram in Fig. 1(b). Owing to the asymmetry between the two modalities, there exists both modality-unique speech information, i.e. audio- and video-unique information  $U_{a}$  and  $U_{v}$ . There is also speech information  $R$  shared between the two modalities, since they share the same underlying speech production process. At the same time, information that only emerges when both audio and video modality are present is the synergistic information, which we denote by  $S$ . Our overall objective is to learn a representation that maximizes the total information  $R + U_{a} + U_{v} + S$  for bi-modal (audio-visual) speech representations, as well as  $R + U_{a}$  and  $R + U_{v}$  for uni-modal (audio & visual) speech representations.",
+        "bbox": [
+            75,
+            537,
+            468,
+            809
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "With this in mind, we reframe the task of learning audiovisual speech representations as a mutual information maximization task. Specifically, the two modalities, corresponding to source variables  $X_{a}$  and  $X_{v}$ , have mutual information with the target speech content variable  $Y$ , containing information useful for downstream tasks that we",
+        "bbox": [
+            75,
+            810,
+            470,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "wish to capture. Denote the joint mutual information between the sources and the target by  $I(X_{a},X_{v};Y)$ . The problem of learning robust audio-visual speech representations reduces to finding the set of parameters that maximizes  $I(\\gamma (X_a),\\gamma (X_v);Y) = I(E_{av};Y)^1$ . Similarly, the problem of learning robust audio and visual speech representations can be viewed as maximizing the mutual information  $I(\\gamma (X_a);Y) = I(E_a;Y)$  and  $I(\\gamma (X_v);Y) = I(E_v;Y)$ , respectively. Taking advantage of the Partial Information Decomposition (PID) framework [73], we can precisely decompose the multivariate mutual information  $I(E_{av};Y),I(E_a;Y)$  and  $I(E_{v};Y)$  into the previous types of information as:",
+        "bbox": [
+            496,
+            90,
+            893,
+            287
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR + U _ {a} = I \\left(E _ {a}; Y\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            640,
+            300,
+            890,
+            316
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR + U _ {v} = I \\left(E _ {v}; Y\\right), \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            642,
+            319,
+            890,
+            335
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nU _ {a} + S = I \\left(E _ {a}; Y \\mid X _ {v}\\right), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            642,
+            338,
+            890,
+            353
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nU _ {v} + S = I \\left(E _ {v}; Y \\mid X _ {a}\\right), \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            642,
+            357,
+            890,
+            372
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR - S = I \\left(E _ {a}; E _ {v}; Y\\right), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            650,
+            376,
+            890,
+            392
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR + U _ {a} + U _ {v} + S = I \\left(E _ {a v}; Y\\right), \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            571,
+            395,
+            890,
+            411
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $R, U_{a}, U_{v}$  and  $S$  are non-negative.",
+        "bbox": [
+            498,
+            422,
+            772,
+            436
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To optimize the above mutual information terms with a neural network parameterized by  $\\varphi_{a},\\varphi_{v}$  and  $\\theta$  we divide the optimization process into three progressive and nonconflicting stages. In stage 1, we learn audio-unique information  $U_{a}$  and shared information  $R$  from the informative audio modality to maximize  $I(E_{a};Y)$ . This process also helps to bootstrap visual representations by learning the shared information  $R$ . In stage 2, we turn to learning visual unique information  $U_{v}$  by maximizing  $I(E_{v};Y) = R + U_{v}$ , which is now easier with the visual information captured via  $R$  in stage 1. At the same time, we inject the learned visual information into the audio-visual joint representations to initialize  $I(E_{av};Y)$ . Finally, in stage 3 we optimize the joint audio-visual representation on top of the previously learned information by maximizing  $I(E_{av};Y)$ .",
+        "bbox": [
+            496,
+            438,
+            890,
+            667
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3. The Complete Learning Framework",
+        "text_level": 1,
+        "bbox": [
+            498,
+            675,
+            815,
+            691
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To learn network parameters that maximize the mutual information terms  $I(E;Y)$ , we leverage the popular self-supervised masked prediction task based on the simple Siamese framework to support the learning process via self-distillation. Inspired by the recent work of Baevski et al. [12], we create multiple masked copies for each training sample, encode the unmasked regions with the Transformer encoder, and train the student to recover contextualized targets of the masked regions with a lightweight 1D convolutional decoder, which are generated by an online EMA teacher. This task can be used to learn any mutual infor",
+        "bbox": [
+            496,
+            698,
+            890,
+            864
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>This equality holds since the architecture implemented in this work only models additive modality synergy stemming from sum fusion.",
+        "bbox": [
+            500,
+            875,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "27071",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/3e1e85f97e7651037acf05269d30db860a88065913e5f669cb25d5be1075bfef.jpg",
+        "image_caption": [
+            "Figure 1. Overview of the proposed method,  $\\mathbf{ES}^3$ . (a) We use a Siamese framework with an EMA teacher to implement an iteration-free training procedure. (b) Venn diagrams showing how the three types of information  $(R, U, \\text{and } S)$  are acquired in an evolving learning process. (c) Illustration of self-distillation with online clustering corresponding to  $\\mathcal{L}_{\\mathrm{distill}}(F_{\\mathrm{tea}}, F_{\\mathrm{stu}})$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            83,
+            92,
+            890,
+            385
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "mation  $I(E_{;}Y)$  when  $Y$  is the downstream speech understanding task. For example, training on auditory or visual speech acts as a proxy to maximize  $I(E_{a};Y)$  or  $I(E_{v};Y)$ . Self-distillation with online clustering [36]. We perform distillation from the right-hand side teacher to the left-hand side student in Fig. 1(a). Specifically, as illustrated in Fig. 1(c), for each layer  $k$  of the top  $K$  Transformer layers in  $\\theta$  with hidden dimension  $D$ , we learn a codebook  $\\mathbf{Q}^{k} = \\{\\mathbf{q}_{1}^{k},\\dots,\\mathbf{q}_{Q}^{k}\\}$  with  $Q$  codewords  $\\mathbf{q}_i^k\\in \\mathbb{R}^D$ , which is updated as follows: for each entry  $j$ , a set  $\\tilde{\\mathbf{Z}}_{\\mathrm{tea}}^k$  of the teacher output features  $\\tilde{\\mathbf{z}}_{\\mathrm{tea}}^k$  closest to the current representation of  $q$  in the codebook is found as",
+        "bbox": [
+            75,
+            438,
+            468,
+            622
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\tilde {\\mathbf {Z}} _ {\\mathrm {t e a}} ^ {k} = \\left\\{\\tilde {\\boldsymbol {z}} _ {\\mathrm {t e a}} ^ {k} \\Bigg | q = \\underset {j \\in \\{1, 2, \\dots , Q \\}} {\\arg \\min } | | \\tilde {\\boldsymbol {z}} _ {\\mathrm {t e a}} ^ {k} - \\mathbf {q} _ {j} ^ {k} | | _ {2} \\right\\}, \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            99,
+            632,
+            468,
+            667
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where the set indices  $q$  are used as discrete pseudo-labels to train the student model. Codewords are then updated in an EMA fashion as a weighted sum of the embeddings in  $\\tilde{\\mathbf{Z}}_{\\mathrm{tea}}^k$  with decay rate  $\\tau_{\\mathrm{code}}$ . The associated loss function is",
+        "bbox": [
+            76,
+            676,
+            468,
+            738
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t i l l}} \\left(F _ {\\text {t e a}}, F _ {\\text {s t u}}\\right) = \\sum_ {t \\in \\mathbb {M}} \\sum_ {k \\in (K - N, K ]} \\log p _ {\\psi_ {k}} \\left(q _ {t} \\mid \\boldsymbol {z} _ {t} ^ {K}\\right), \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            93,
+            747,
+            468,
+            792
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $F_{\\mathrm{tea}}$  and  $F_{\\mathrm{stu}}$  are teacher and student features,  $\\mathbb{M}$  is the set of masked time steps,  $\\psi_{k}$  is the softmax activations of linear projections for each target layer  $k$ ,  $q_{t}$  is the codeword index for time step  $t$  of the teacher  $\\tilde{\\theta}$ , and  $z_{t}^{K}$  is the output of the  $K$ -th and final layer of the student model  $\\theta$ . The teacher and the online clustering process do not require gradients. EMA teacher parameterization. To reduce GPU memory",
+        "bbox": [
+            75,
+            795,
+            468,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "usage, we apply EMA updates to the Transformer encoder  $\\theta$ , but not the earlier modality-specific encoders  $\\varphi_{a}$  and  $\\varphi_{v}$ , inspired by [50]. As shown in Fig. 1, the teacher model weights  $\\tilde{\\theta}$  are computed as an exponential moving average of the student model weights  $\\theta$ :",
+        "bbox": [
+            498,
+            439,
+            890,
+            513
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\tilde {\\theta} \\leftarrow \\eta \\tilde {\\theta} + (1 - \\eta) \\theta , \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            625,
+            523,
+            890,
+            542
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\eta = \\min \\{\\eta_{\\max}, (1 + \\text{step}) / (10 + \\text{step})\\}$  is a momentum parameter that slowly increases over time until a pre-determined value  $\\eta_{\\max}$  is reached.",
+        "bbox": [
+            498,
+            554,
+            890,
+            599
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4. The Learning Process",
+        "text_level": 1,
+        "bbox": [
+            500,
+            609,
+            702,
+            626
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We now elaborate on individual learning stages, based on the above Siamese framework and distillation loss.",
+        "bbox": [
+            498,
+            633,
+            890,
+            662
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Stage 1: Learning audio-unique information and shared information. In stage 1, we perform distillation from un-masked audio teacher model to the masked audio student model, yielding audio-unique information  $U_{a}$  and shared information  $R$  from audio. The distillation loss is:",
+        "bbox": [
+            498,
+            662,
+            890,
+            737
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a} \\rightarrow \\mathrm {a} ^ {\\prime}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {a}, E _ {a} ^ {\\prime}\\right). \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            606,
+            751,
+            890,
+            767
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Being readily learnable, audio has the potential to provide most of  $R$  for learning shared information. However, optimizing  $I(E_{a};Y)$  alone cannot help visual learning. Noting that the InfoNCE loss between two variables has been widely used to maximize a lower bound on the mutual information, we adopt the InfoNCE loss to learn shared information corresponding to  $I(E_{a};E_{v})$  to bootstrap visual learning with the shared information learned from audio.",
+        "bbox": [
+            496,
+            779,
+            890,
+            898
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "27072",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we apply it to a fine-grained frame-wise alignment task, where temporally aligned features from the same video are treated as positives, and unaligned ones negatives [19]. Since phonemes can appear repeatedly within the same video clip, leading to false negatives, we use a negative margin following Park et al. [47], which alleviates the effect of falsely matched pairs by enforcing a looser decision boundary. The objective function involved is:",
+        "bbox": [
+            75,
+            90,
+            472,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {N C E}} (H) = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\log \\frac {e ^ {(H _ {i , i} - m) / \\tau}}{e ^ {(H _ {i , i} - m) / \\tau} + \\sum_ {i \\neq j} e ^ {H _ {i , j} / \\tau}}, \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            86,
+            222,
+            468,
+            273
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $H$  is a matrix whose  $(i,j)$ -th element represents the cosine similarity between  $L^2$ -normalized audio and visual features at time step  $i$  and  $j$ , respectively,  $\\tau$  is a learnable temperature parameter, and  $m$  is a pre-determined margin. To promote the learning of shared features between the two modalities, we apply a symmetrized version of the above NCE loss as:",
+        "bbox": [
+            75,
+            273,
+            472,
+            378
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a} \\leftrightarrow \\mathrm {v}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {a} \\rightarrow \\mathrm {v}} + \\mathcal {L} _ {\\mathrm {v} \\rightarrow \\mathrm {a}}\\right) = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {N C E}} \\left(S _ {a v}\\right) + \\mathcal {L} _ {\\mathrm {N C E}} \\left(S _ {v a}\\right)\\right). \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            76,
+            386,
+            473,
+            426
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Here the similarity matrices  $S_{av} = \\langle F_a, F_v \\rangle / (||F_a|| \\cdot ||F_v||)$  and  $S_{va} = S_{av}^{\\top}$  are computed with features  $F_a$  and  $F_v$  from modality-specific encoders  $\\varphi_a$  and  $\\varphi_v$  instead of the contextual features from the Transformer encoder  $\\theta$  to avoid learning positional shortcuts. The total loss in stage 1 is",
+        "bbox": [
+            75,
+            426,
+            468,
+            503
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {s t a g e 1}} = \\mathcal {L} _ {\\mathrm {a} \\rightarrow \\mathrm {a} ^ {\\prime}} + \\mathcal {L} _ {\\mathrm {a} \\leftrightarrow \\mathrm {v}}. \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            181,
+            515,
+            468,
+            531
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Stage 2: Learning video-unique information. In stage 2, we proceed to learn video-unique information, which is now much easier than learning from scratch since it has been bootstrapped with shared knowledge from the audio in the previous stage. The audio encoder  $\\varphi_{a}$  does not receive any gradients in this stage and is essentially frozen, since it is expected to have already acquired the audio-specific knowledge well in the previous stage. We adopt a similar distillation loss as stage 1:",
+        "bbox": [
+            75,
+            537,
+            472,
+            675
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {v} \\rightarrow \\mathrm {v} ^ {\\prime}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {v}, E _ {v} ^ {\\prime}\\right). \\tag {15}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            183,
+            685,
+            468,
+            702
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "On top of the learned visual-unique information, we also perform distillation from unmasked video features to the joint features of unmasked audio-visual inputs in this stage to prepare for the learning of the final robust audio-visual representations in the next stage. For this purpose, the loss in this stage is set as:",
+        "bbox": [
+            75,
+            707,
+            468,
+            797
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a v} \\rightarrow \\mathrm {v}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {v}, E _ {a v}\\right), \\tag {16}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            179,
+            809,
+            468,
+            825
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {s t a g e 2}} = \\mathcal {L} _ {\\mathrm {v} \\rightarrow \\mathrm {v} ^ {\\prime}} + \\mathcal {L} _ {\\mathrm {a v} \\rightarrow \\mathrm {v}}. \\tag {17}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            179,
+            828,
+            468,
+            844
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To avoid repeated learning of only shared information between A and V again, the cross-modal loss  $\\mathcal{L}_{\\mathrm{av}\\to \\mathrm{v}}$  is asymmetric, strictly one-way from AV to V, and a learnable tem",
+        "bbox": [
+            75,
+            854,
+            468,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "perature  $\\tau_{v}$  Eq. (16) is introduced to the visual student outputs, i.e., we replace  $\\psi_{k}$  in Eq. (9) with softmax activations of temperature-scaled linear projections of  $E_{v}$ .",
+        "bbox": [
+            496,
+            90,
+            890,
+            136
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Stage 3: Learning synergistic information. Finally, we aim to learn synergistic audio-visual information which previous methods have not explicitly addressed. Our ultimate goal is to maximize total task-relevant multi-modal information  $I(E_{av};Y)$ , i.e. the sum of  $R$ ,  $U_{a}$ ,  $U_{v}$  and  $S$ , so that the representations are useful for both uni-modal and multi-modal tasks. Naively maximizing this objective without prior stages is problematic because the network tends to be biased towards learning audio-based representations (i.e. prioritizing  $R$  and  $U_{a}$ ) without considering the visual modality. By learning most of  $R$ ,  $U_{a}$  and  $U_{v}$  through the previous stages, our method is less susceptible to learning biased representations in the final stage. At this stage, we always mask identical positions in the audio and video inputs together so that the masked region never contains any unimodal information that can collapse or degrade training. To better recover the masked inputs, the model must cleverly leverage both modalities, which requires learning synergistic information that is not well-captured in prior stages. The loss function in stage 3 can be written as:",
+        "bbox": [
+            496,
+            136,
+            892,
+            438
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {s t a g e 3}} = \\mathcal {L} _ {\\mathrm {a v} \\rightarrow \\mathrm {a v} ^ {\\prime}} = \\mathcal {L} _ {\\text {d i s t i l l}} \\left(E _ {a v} ^ {\\prime}, E _ {a v}\\right). \\tag {18}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            557,
+            448,
+            890,
+            465
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "By progressively acquiring the three types of useful information, we build strong and robust audio-visual speech representations for both uni- and bi-modal speech, which we shall show in subsequent experiments.",
+        "bbox": [
+            496,
+            469,
+            890,
+            529
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.5. Fine-tuning Stage",
+        "text_level": 1,
+        "bbox": [
+            500,
+            537,
+            671,
+            554
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After pre-training, we add an extra linear layer after the fusion encoder  $\\theta$  to fine-tune the model end-to-end with the Lattice-Free Maximum Mutual Information (LF-MMI) criterion [26, 51], which is much more lightweight than seq2seq Encoder-Decoder models and allows highly efficient decoding. We adopt a \"compact\" [33] 1-state hidden Markov model (HMM) topology, using  $k2^2$  for differentiable Weighted Finite State Transducer (WFST)-based training. We use 624 bi-char units from text-based clustering [77] for English and tonal phone-level targets (initials and finals) for Chinese.",
+        "bbox": [
+            496,
+            561,
+            890,
+            727
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            739,
+            632,
+            757
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Datasets",
+        "text_level": 1,
+        "bbox": [
+            500,
+            763,
+            601,
+            779
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "LRS2-BBC [3]. LRS2-BBC is a 224-hour transcribed audio-visual English speech dataset collected from programs of various genres broadcast by the BBC. It is challenging due to the presence of extreme poses, as well as a large and diverse vocabulary. We use the official train / validation split (see supplementary materials) for evaluation.",
+        "bbox": [
+            496,
+            787,
+            890,
+            878
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "2https://github.com/k2-fsa/k2",
+        "bbox": [
+            516,
+            886,
+            712,
+            898
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "27073",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "LRS3-TED [1]. LRS3-TED is another audio-visual English speech dataset collected from TED and TEDx talks, with over 5.5k speakers and 439 hours of data. It covers diverse topics and is widely adopted for evaluating audio-visual speech representation learning methods for being the largest publicly available one of its kind. An official train / validation split is not given, so we use the widely adopted one prepared by Shi et al. [59], same as [27, 34, 80].",
+        "bbox": [
+            75,
+            90,
+            468,
+            210
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "CAS-VSR-S101. We also collect a new large-scale, in-the-wild Mandarin dataset with 101.1 hours of data. The videos are sourced from broadcast news and conversational programs in Chinese, covering a highly diverse set of topics, speakers and filming conditions (see supplementary).",
+        "bbox": [
+            75,
+            212,
+            468,
+            287
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            76,
+            297,
+            266,
+            314
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "More implementation details can be found in the supplementary materials.",
+        "bbox": [
+            75,
+            321,
+            467,
+            351
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "High- and low-resource setups. On LRS3-TED we follow the widely adopted setup of Shi et al. [59], which excludes the pre-train set during low-resource fine-tuning. To validate data scaling properties, previous works used VoxCeleb2-En [59], which is a 1326-hour subset of VoxCeleb2 [18] in English. However, not only are experiments of this scale computationally expensive, but more importantly, VoxCeleb2 is no longer available from its official source as of Feb. 2023. Therefore, we create a similar scaling protocol, using LRS3-TED (433h) as a pre-training source and LRS2-BBC (223h) for fine-tuning. For low-resource fine-tuning, we only use the 28-hour train subset of LRS2-BBC, while for high-resource fine-tuning, we include the pre-train subset, similar to [17].",
+        "bbox": [
+            75,
+            352,
+            467,
+            564
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation. Decoding is performed using beam search with a beam size of 50. The common evaluation metrics for ASR, VSR and AVSR are Word Error Rate (WER) and Character Error Rate (CER) for English and Chinese, respectively. As our bi-char LF-MMI models are trained on verbatim transcriptions for ease of lexicon construction, we report WERs both without and with using Whisper-style text normalization [54] (in parentheses) to compensate for innocuous errors related to text normalization.",
+        "bbox": [
+            75,
+            564,
+            467,
+            700
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Results on LRS2-BBC and LRS3-TED",
+        "text_level": 1,
+        "bbox": [
+            76,
+            710,
+            408,
+            724
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Low-resource setup. Low-resource results on LRS2-BBC and LRS3-TED can be found in Tabs. 1 and 2. Our smallest BASE* encoder is modality-agnostic, only half the size of the one in [27], which uses two separate encoders to perform AVSR. On LRS3-TED, although we employ the simple LF-MMI instead of cross-entropy (CE) based seq2seq criterion for fine-tuning, our BASE model rivals or achieves SoTA in all cases. When scaling to a LARGE model, we observe significant improvements in both ASR and AVSR. Meanwhile, VSR performance remains comparable to SoTA, with a moderate  $2\\%$  gap. This may sug",
+        "bbox": [
+            75,
+            734,
+            467,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "gest that with our current VGG-based visual encoder  $\\varphi_v$ , the benefits of scaling up the Transformer primarily manifest in ASR and AVSR.",
+        "bbox": [
+            496,
+            90,
+            890,
+            135
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "High-resource setup. High-resource results on LRS2-BBC and LRS3-TED can be found in Tabs. 1 and 3. On LRS2-BBC, with a 46M model and relying on its own 223h data for pre-training and fine-tuning, we achieve  $30.3\\% / 3.6\\% / 2.9\\%$  WER for VSR, ASR and AVSR. Notably, these results surpass a highly competitive supervised method [40] and other self-supervised methods that incorporate additional data. This indicates that our progressive learning approach can effectively exploit visual information. Furthermore, when we introduce additional 433h unlabeled data, we achieve even better performance:  $29.8\\% / 3.0\\% / 2.4\\%$  WER, which outperforms other methods that employ the same amount of data. When we switch to the LARGE model, performance rivals SoTA methods that employ 4x data (1759h) for pre-training. Similarly, we achieve or rival SoTA performance on LRS3-TED.",
+        "bbox": [
+            496,
+            137,
+            890,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4. Results on CAS-VSR-S101",
+        "text_level": 1,
+        "bbox": [
+            498,
+            390,
+            741,
+            404
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We also evaluate on our newly collected CAS-VSR-S101 dataset, which is smaller than both of the two previous datasets and in an even more ambiguous language (Chinese has more homophonous syllables and words than English). As shown in Tab. 4,  $\\mathrm{ES}^3$  still works out-of-the-box under this challenging setup, and yields noticeable performance gains over a strong fully supervised baseline [39]. Interestingly, our fully supervised baseline model beats [39] in ASR and AVSR but slightly lags behind in VSR, again hinting that our strategy can further benefit from more complex visual modeling architectures.",
+        "bbox": [
+            496,
+            412,
+            890,
+            580
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/d657ac7401c12c04bb6620d820470a30a41816cc760ec936a66a4bb09363df4e.jpg",
+        "table_caption": [
+            "Table 4. Results on CAS-VSR-S101. Here we use a BASE* model for the fully-supervised baseline and  $\\mathrm{ES}^3$ . The self-supervised models are pre-trained on training data and then finetuned for 50 epochs with the labels. The numbers before and after slashes correspond to CERs on the validation and test sets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>Criterion</td><td>Encoder Size</td><td>#Epochs</td><td>VSR</td><td>ASR</td><td>AVSR</td></tr><tr><td colspan=\"7\">Supervised</td></tr><tr><td>Baseline</td><td>LF-MMI</td><td>46M</td><td>100</td><td>60.0 / 57.4</td><td>15.4 / 12.6</td><td>14.6 / 12.1</td></tr><tr><td>E2E Conformer [39]</td><td>CTC+CE</td><td>79M</td><td>100</td><td>58.6 / 56.5</td><td>16.6 / 13.0</td><td>16.0 / 12.6</td></tr><tr><td colspan=\"7\">Self-supervised</td></tr><tr><td>ES3 Stage 1</td><td></td><td></td><td></td><td>61.2 / 59.0</td><td>14.7 / 11.9</td><td>-</td></tr><tr><td>ES3 Stage 2</td><td>LF-MMI</td><td>46M</td><td>50</td><td>59.1 / 56.7</td><td>-</td><td>13.9 / 11.4</td></tr><tr><td>ES3 Stage 3</td><td></td><td></td><td></td><td>58.1 / 55.6</td><td>14.4 / 11.6</td><td>13.4 / 11.0</td></tr></table>",
+        "bbox": [
+            503,
+            657,
+            890,
+            732
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.5. Analysis and Discussion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            739,
+            718,
+            756
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Representation quality probing. In Tab. 5, we demonstrate the effect of each stage on LRS2-BBC. ASR performance between stages 1 and 3 are close, indicating that stage 1 indeed acquires most of audio-unique information. Comparing the VSR results between stages 1 and 2 reveals a significant gap, suggesting that stage 2 also effectively learns visual-unique information, hence the minimal differences between the final VSR result and those of stage 2. Moreover, comparing the results of stage 1-2 and stage",
+        "bbox": [
+            496,
+            763,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "27074",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/927360e44f344c6270da4c7c7364195099ad0468b9eeae0f16e16f1fb37ff05a.jpg",
+        "table_caption": [
+            "Table 1. Results on LRS2-BBC. We pre-train a BASE* model with 223h unlabeled data from LRS2-BBC, as well as BASE and LARGE models with 433h unlabeled data (LRS3-TED) to demonstrate scaling properties. *: external language models."
+        ],
+        "table_footnote": [
+            "${}^{1}$  Not end-to-end; requires a GMM-HMM alignment stage.  ${}^{2}$  Uses an additional ASR model trained on LibriSpeech (960h).",
+            "3 Reproduced by Zhu et al. [80]. 4 Uses additional 3846h audio, 452h audio-text and 600M text data.",
+            "<sup>5</sup> Uses additional 60000h audio data and 1.28M unlabeled images."
+        ],
+        "table_body": "<table><tr><td>Methods</td><td>Unlabeled AV data</td><td>Labeled Data</td><td>Encoder Size</td><td>Criterion</td><td>VSR</td><td>ASR</td><td>AVSR</td></tr><tr><td colspan=\"8\">Supervised</td></tr><tr><td>Yu et al. [75]</td><td>-</td><td>1519h</td><td>-</td><td>LF-MMI1</td><td>48.9</td><td>6.7</td><td>5.9</td></tr><tr><td>Ma et al. [39]*</td><td>-</td><td>380h</td><td>79M</td><td>CTC+CE</td><td>37.9</td><td>3.9</td><td>3.7</td></tr><tr><td>Ma et al. [41]*</td><td>-</td><td>818h</td><td>186M</td><td>CTC+CE</td><td>27.9</td><td>2.6</td><td>-</td></tr><tr><td colspan=\"8\">Semi-supervised</td></tr><tr><td>Afouras et al. [2]*</td><td>777h</td><td>223h2</td><td>-</td><td>CTC</td><td>51.3</td><td>-</td><td>-</td></tr><tr><td>Ma et al. [41]*</td><td>2630h</td><td>818h</td><td>186M</td><td>CTC+CE</td><td>14.6</td><td>1.5</td><td>1.5</td></tr><tr><td colspan=\"8\">Self-supervised (BASE and BASE* models)</td></tr><tr><td>AV-HuBERT [59, 60]</td><td>1759h</td><td>223h</td><td>103M</td><td>CE</td><td>31.23</td><td>-</td><td>3.63</td></tr><tr><td>VATLM [80]</td><td>1759h4</td><td>223h</td><td>107M</td><td>CE</td><td>30.6</td><td>-</td><td>2.9</td></tr><tr><td>RAVeN [27]</td><td>433h</td><td>223h</td><td>97M</td><td>CTC+CE</td><td>32.1</td><td>3.9</td><td>-</td></tr><tr><td>Pan et al. [45]</td><td>-5</td><td>380h</td><td>399M</td><td>CTC+CE</td><td>43.2</td><td>2.7</td><td>2.6</td></tr><tr><td rowspan=\"4\">ES3(ours)</td><td>223h</td><td>28h</td><td>46M</td><td></td><td>40.2 (39.1)</td><td>6.0 (5.0)</td><td>5.7 (4.9)</td></tr><tr><td>433h</td><td>28h</td><td>102M</td><td rowspan=\"3\">LF-MMI</td><td>39.3 (38.2)</td><td>5.5 (4.8)</td><td>5.1 (4.1)</td></tr><tr><td>223h</td><td>223h</td><td>46M</td><td>31.4 (30.3)</td><td>4.3 (3.6)</td><td>3.8 (2.9)</td></tr><tr><td>433h</td><td>223h</td><td>102M</td><td>30.7 (29.8)</td><td>3.4 (3.0)</td><td>3.2 (2.4)</td></tr><tr><td colspan=\"8\">Self-supervised (LARGE models)</td></tr><tr><td rowspan=\"2\">ES3(ours)</td><td>433h</td><td>28h</td><td>317M</td><td rowspan=\"2\">LF-MMI</td><td>36.4 (35.4)</td><td>5.2 (4.4)</td><td>4.7 (4.0)</td></tr><tr><td>433h</td><td>223h</td><td>317M</td><td>26.7 (25.8)</td><td>3.1 (2.5)</td><td>3.1 (2.5)</td></tr><tr><td>AV-HuBERT [59, 60]</td><td>1759h</td><td>223h</td><td>325M</td><td>CE</td><td>25.53</td><td>-</td><td>2.53</td></tr><tr><td>VATLM [80]</td><td>1759h4</td><td>223h</td><td>332M</td><td>CE</td><td>24.3</td><td>-</td><td>2.3</td></tr><tr><td>RAVeN [27]</td><td>1759h</td><td>223h</td><td>671M</td><td>CTC+CE</td><td>23.2</td><td>2.5</td><td>-</td></tr></table>",
+        "bbox": [
+            161,
+            121,
+            810,
+            393
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/2dcdb6385a0d37b67d50823791dac44d7c521cec9ae185d63940dae40105ddae.jpg",
+        "table_caption": [
+            "Table 2. Low-resource results on LRS3-TED. We pre-train a BASE and LARGE model with 433h unlabeled data."
+        ],
+        "table_footnote": [
+            "$^{1}$  Reproduced by Lian et al. [34].  $^{2}$  Uses additional 3846h audio, 452h audio-text and 600M text data.",
+            "3 Zhang et al. [76] inject noise during pre-training and obtain better fine-tuning results even with its base model AV-HuBERT  $(47.1\\%)$ ."
+        ],
+        "table_body": "<table><tr><td>Methods</td><td>Unlabeled AV data</td><td>Labeled Data</td><td>Encoder Size</td><td>Criterion</td><td>VSR</td><td>ASR</td><td>AVSR</td></tr><tr><td colspan=\"8\">Self-supervised (BASE models)</td></tr><tr><td>AV-HuBERT [59, 60]</td><td>433h</td><td>30h</td><td>103M</td><td>CE</td><td>51.8</td><td>4.9</td><td>\\( {4.7}^{1} \\)</td></tr><tr><td>VATLM [80]</td><td>\\( {433}{\\mathrm{\\;h}}^{2} \\)</td><td>30h</td><td>107M</td><td>CE</td><td>48.0</td><td>-</td><td>3.6</td></tr><tr><td>RAVEn [27]</td><td>433h</td><td>30h</td><td>97M</td><td>CTC+CE</td><td>47.0</td><td>4.7</td><td>-</td></tr><tr><td>\\( {\\mathrm{{AV2vec}}}^{\\left\\lbrack {76}\\right\\rbrack  }{}^{3} \\)</td><td>433h</td><td>30h</td><td>103M</td><td>CE</td><td>45.0</td><td>-</td><td>5.8</td></tr><tr><td>AV-data2vec [34]</td><td>433h</td><td>30h</td><td>103M</td><td>CE</td><td>45.2</td><td>4.4</td><td>4.2</td></tr><tr><td>\\( {\\mathbf{{ES}}}^{3} \\) (ours)</td><td>433h</td><td>30h</td><td>102M</td><td>LF-MMI</td><td>45.5 (44.7)</td><td>3.9 (3.3)</td><td>3.6 (3.0)</td></tr><tr><td colspan=\"8\">Self-supervised (LARGE models)</td></tr><tr><td>AV-HuBERT</td><td>433h</td><td>30h</td><td>325M</td><td>CE</td><td>44.8</td><td>4.5</td><td>\\( {4.2}^{1} \\)</td></tr><tr><td>AV-data2vec</td><td>433h</td><td>30h</td><td>325M</td><td>CE</td><td>40.5</td><td>3.7</td><td>3.4</td></tr><tr><td>\\( {\\mathbf{{ES}}}^{3} \\) (ours)</td><td>433h</td><td>30h</td><td>317M</td><td>LF-MMI</td><td>43.5 (42.5)</td><td>3.8 (2.9)</td><td>2.9 (2.3)</td></tr></table>",
+        "bbox": [
+            161,
+            454,
+            810,
+            597
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3 demonstrates that the joint information learning in stage 3, which also incorporates synergistic information, benefits both uni-modal (ASR & VSR) and bi-modal (AVSR) tasks. Similar trends can also be observed on the more challenging Mandarin dataset CAS-VSR-S101 in Tab. 4, further supporting the motivation of our proposed strategy.",
+        "bbox": [
+            75,
+            627,
+            467,
+            719
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 5. Representation quality probing on LRS2-BBC. Results are reported on the validation set using the BASE* model under the low-resource setup for each pre-training stage.",
+        "bbox": [
+            75,
+            720,
+            467,
+            762
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/4b2ca0b3d50677d50b9757e16e0ca0bb8b8248f847b4005e729ff730ba474000.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Pre-training Stage</td><td>VSR</td><td>ASR</td><td>AVSR</td></tr><tr><td>Stage 1</td><td>58.3 (57.3)</td><td>11.5 (10.1)</td><td>-</td></tr><tr><td>Stage 2</td><td>48.7 (47.3)</td><td>-</td><td>12.9 (11.5)</td></tr><tr><td>Stage 3 (Full ES³)</td><td>48.1 (46.9)</td><td>11.4 (9.8)</td><td>10.1 (8.6)</td></tr></table>",
+        "bbox": [
+            109,
+            767,
+            439,
+            819
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Effect of automatically balancing shared and unique information learning. It is known that in a hardness-aware loss function, the temperature rules the strength of penalties on hard negative samples [72]. We track the automatically learned temperature parameter  $\\tau$  used for cross-modal",
+        "bbox": [
+            75,
+            825,
+            467,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "learning objectives  $\\mathcal{L}_{\\mathrm{a}\\leftrightarrow \\mathrm{v}}$  in Eq. (12) and  $\\mathcal{L}_{\\mathrm{av}\\rightarrow \\mathrm{v}}$  in Eq. (16) when pre-training on LRS2-BBC, as shown in Fig. 2. In stage 1,  $\\tau$  continues to fall until it reaches  $\\tau_{\\mathrm{min}}$ , while in stage 2 it briefly drops and then rises throughout training. The trend is in line with our intuition: in stage 1, audio information is easily learnable, so the model can fully exploit shared information by lowering  $\\tau$  without sacrificing audio-unique information learning; in stage 2, after learning some residual shared information, an increasing  $\\tau$  progressively relaxes  $\\mathcal{L}_{\\mathrm{av}\\rightarrow \\mathrm{v}}$  to yield more freedom to  $\\mathcal{L}_{\\mathrm{v}\\rightarrow \\mathrm{v}'}$ , i.e. video-unique information learning.",
+        "bbox": [
+            496,
+            627,
+            890,
+            794
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/dc65de78e349c393d51fed49ea503110b75c3317341ac1413d53fd9f30ec9c74.jpg",
+        "image_caption": [
+            "Figure 2. The temperature parameter paces learning automatically."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            560,
+            805,
+            826,
+            904
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "27075",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/7e5f9b935296553ba314447ad2237026f97cb39e4de41d77568a11274c9bd0ea.jpg",
+        "table_caption": [
+            "Table 3. High-resource results on LRS3-TED. We fine-tune on 433h labeled data of LRS3. †: uses test-time augmentation. *: uses external language models. ‡: noise injection during pre-training. We list more methods for comparison in the supplementary material."
+        ],
+        "table_footnote": [
+            "<sup>1</sup> Uses an additional ASR model trained on LibriSpeech (960h).",
+            "2 Reproduced by Lian et al. [34]."
+        ],
+        "table_body": "<table><tr><td>Methods</td><td>Year</td><td>Unlabeled AV data</td><td>Labeled Data</td><td>Backbone</td><td>Encoder Size</td><td>Criterion</td><td>VSR</td><td>ASR</td><td>AVSR</td></tr><tr><td colspan=\"10\">Supervised</td></tr><tr><td>Ma et al. [39]*</td><td>2021</td><td>-</td><td>595h</td><td>Conformer</td><td>79M</td><td>CTC+CE</td><td>43.3</td><td>2.3</td><td>2.3</td></tr><tr><td>Prajwal et al. [52]†*</td><td>2022</td><td>-</td><td>698h</td><td>Transformer</td><td>32M</td><td>CE</td><td>40.6</td><td>-</td><td>-</td></tr><tr><td>Ma et al. [40]*</td><td>2022</td><td>-</td><td>818h</td><td>Conformer</td><td>79M</td><td>CTC+CE</td><td>34.7</td><td>-</td><td>-</td></tr><tr><td colspan=\"10\">Semi-Supervised</td></tr><tr><td>Shillingford et al. [62]*</td><td>2019</td><td>-</td><td>3886h</td><td>RNN</td><td>-</td><td>CTC</td><td>55.1</td><td>-</td><td>-</td></tr><tr><td>Makino et al. [42]</td><td>2019</td><td>-</td><td>31kh</td><td>RNN</td><td>43M</td><td>Transducer</td><td>33.6</td><td>4.8</td><td>4.5</td></tr><tr><td>Afouras et al. [2]*</td><td>2020</td><td>344h</td><td>433h1</td><td>Jasper (CNN)</td><td>-</td><td>CTC</td><td>59.8</td><td>-</td><td>-</td></tr><tr><td>Serdyuk et al. [57]</td><td>2021</td><td>-</td><td>90kh</td><td>Transformer</td><td>-</td><td>Transducer</td><td>25.9</td><td>-</td><td>2.3</td></tr><tr><td>Ma et al. [40]*</td><td>2022</td><td>641h</td><td>818h</td><td>Conformer</td><td>79M</td><td>CTC+CE</td><td>31.5</td><td>-</td><td>-</td></tr><tr><td>Serdyuk et al. [58]</td><td>2022</td><td>-</td><td>90kh</td><td>Conformer</td><td>-</td><td>Transducer</td><td>17.0</td><td>1.6</td><td>1.6</td></tr><tr><td>Ma et al. [41]*</td><td>2023</td><td>2630h</td><td>818h</td><td>Conformer</td><td>186M</td><td>CTC+CE</td><td>19.1</td><td>1.0</td><td>0.9</td></tr><tr><td>Chang et al. [15]</td><td>2023</td><td>-</td><td>100kh</td><td>Conformer</td><td>98M</td><td>Transducer</td><td>12.8</td><td>-</td><td>0.9</td></tr><tr><td colspan=\"10\">Self-supervised (Base Models)</td></tr><tr><td>AV-HuBERT [59, 60]</td><td>2022</td><td>433h</td><td>433h</td><td>Transformer</td><td>103M</td><td>CE</td><td>44.0</td><td>3.0</td><td>2.82</td></tr><tr><td>AV2vec [76]‡</td><td>2023</td><td>433h</td><td>433h</td><td>Transformer</td><td>103M</td><td>CE</td><td>39.9</td><td>-</td><td>2.6</td></tr><tr><td>RAVEn [27]</td><td>2023</td><td>433h</td><td>433h</td><td>Transformer</td><td>97M</td><td>CTC+CE</td><td>39.1</td><td>2.2</td><td>-</td></tr><tr><td>ES3 (ours)</td><td>2023</td><td>433h</td><td>433h</td><td>Transformer</td><td>102M</td><td>LF-MMI</td><td>40.3 (39.2)</td><td>2.9 (2.4)</td><td>2.5 (2.0)</td></tr><tr><td>AV-data2vec [34]</td><td>2023</td><td>433h</td><td>433h</td><td>Transformer</td><td>103M</td><td>CE</td><td>39.0</td><td>2.0</td><td>1.8</td></tr><tr><td colspan=\"10\">Self-supervised (Large Models)</td></tr><tr><td>AV-HuBERT [59, 60]</td><td>2022</td><td>433h</td><td>433h</td><td>Transformer</td><td>325M</td><td>CE</td><td>41.6</td><td>2.7</td><td>2.52</td></tr><tr><td>AV-data2vec [34]</td><td>2023</td><td>433h</td><td>433h</td><td>Transformer</td><td>325M</td><td>CE</td><td>37.4</td><td>1.9</td><td>1.7</td></tr><tr><td>ES3 (ours)</td><td>2023</td><td>433h</td><td>433h</td><td>Transformer</td><td>317M</td><td>LF-MMI</td><td>37.1 (36.7)</td><td>2.8 (2.1)</td><td>2.1 (1.7)</td></tr></table>",
+        "bbox": [
+            140,
+            122,
+            831,
+            371
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Visualization of our learned psuedo-label sequences. We randomly select a sample from the LRS3-TED validation set and visualize the pseudo-label sequences, i.e. code indices  $q$  generated by the audio, visual and audio-visual EMA teacher  $\\tilde{\\theta}$  in each stage. Ground truth phones for each audio frame (10ms) are derived by forced alignment [79]. Fig. 3 shows that our learned codebook yields high-quality pseudo-labels that strongly correlate with phonemic units.",
+        "bbox": [
+            75,
+            388,
+            468,
+            508
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/5e00eb854a35c01a20449a847b6613fbe7a3476ae06be938215eb0055ab4b0f6.jpg",
+        "image_caption": [
+            "Figure 3. Pseudo-label sequences from each stage. Example of pseudo-labels generated from the last Transformer layer for the utterance \"so what does that mean\" by the audio, visual, and audio-visual teachers in each stage (BASE model). At the bottom we show the spectrogram for reference. Please note that our pseudolabels are derived at 40ms intervals, limited by video frame rate."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            515,
+            462,
+            633
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Training efficiency comparison with SoTA methods. Learning audio-visual speech representations has been extremely compute-intensive. Here we briefly compare  $\\mathrm{ES}^3$  with prominent SoTA methods. Pre-training a BASE/LARGE model on LRS3-TED using our  $\\mathrm{ES}^3$  for 100 epochs per stage only takes about 6/8.3 days in total with just 8 NVIDIA A40 GPUs. In contrast, AV-data2vec is trained for 1000k steps using 64 NVIDIA V100 GPUs for 4-5/6-7 days using BASE/LARGE with a batch size of 1280/2560s, or equivalently, 821/1642 epochs. They also experiment extensively with hyperparameters such as modality schedulers, which further increases the cost of",
+        "bbox": [
+            75,
+            719,
+            470,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "training. RAVEn performs an equivalent of 6.25/50 days of training on 8 A100s with BASE*/LARGE. It uses separate encoders for audio and video, so the resulting model cannot perform audio-visual fusion out-of-the-box. This also requires keeping track of four sets of parameters, which is very memory-intensive. By focusing on designing a clever learning strategy, we circumvent such manual tuning altogether and greatly reduce training costs. Finally, we use the simple LF-MMI criterion, and expect better performance from tuning beam size, including a larger external language model (see supplementary), or switching to the common heavy Attention-based Encoder-Decoder model.",
+        "bbox": [
+            496,
+            388,
+            893,
+            569
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusions",
+        "text_level": 1,
+        "bbox": [
+            500,
+            575,
+            627,
+            590
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We present a novel strategy,  $\\mathrm{ES}^3$ , for self-supervised learning of audio-visual speech representations. Unlike existing methods that primarily focus on learning information shared between audio and video, we take an \"evolving\" approach that also captures unique and synergistic information. Our experimental results demonstrate that  $\\mathrm{ES}^3$  indeed progressively learns unique, shared and synergistic information. Moreover, in terms of downstream performance, it rivals SoTA methods in terms of ASR, VSR and AVSR accuracy without burdensome manual balancing efforts.",
+        "bbox": [
+            496,
+            595,
+            890,
+            747
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments. This work is partially supported by the National Natural Science Foundation of China (No. 62276247, 62076250). We are grateful to the ICT computing platform, the Nanjing Institute of InforSuperBahn, and Capitalonline Data Service Co., Ltd for providing additional compute resources. We also thank Chenhao Wang, Mingmin Yang and Jingyun Xiao for helping to prepare CAS-VSR-S101, Bingquan Xia and Songtao Luo for help with Conformer and seq2seq baselines, Bingquan Xia and Feixiang Wang for help during the rebuttal phase, and Feixiang Wang and Zhaoxin Yuan for an early version of Fig. 1.",
+        "bbox": [
+            496,
+            747,
+            892,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "27076",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            119,
+            173,
+            136
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. LRS3-TED: a large-scale dataset for visual speech recognition. CoRR, abs/1809.00496, 2018. 6",
+            "[2] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. ASR is all you need: Cross-modal distillation for lip reading. In ICASSP, pages 2143-2147. IEEE, 2020. 2, 7, 8, 4, 5",
+            "[3] Triantafyllos Afouras, Joon Son Chung, Andrew W. Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE Trans. Pattern Anal. Mach. Intell., 44(12):8717-8727, 2022. 1, 5, 4",
+            "[4] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. In NeurlPS, 2020. 2",
+            "[5] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020. 2",
+            "[6] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In ICCV, pages 609-617. IEEE Computer Society, 2017. 2",
+            "[7] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In ECCV(1), pages 451-466. Springer, 2018. 2",
+            "[8] Yuki Markus Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. In NeurIPS, 2020. 2",
+            "[9] Alexei Baevski, Michael Auli, and Abdelrahman Mohamed. Effectiveness of self-supervised pre-training for speech recognition. CoRR, abs/1911.03912, 2019. 1",
+            "[10] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In NeurIPS, 2020. 3",
+            "[11] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. In ICML, pages 1298-1312. PMLR, 2022. 1",
+            "[12] Alexei Baevski, Arun Babu, Wei-Ning Hsu, and Michael Auli. Efficient self-supervised learning with contextualized target representations for vision, speech and language. In ICML, pages 1416-1429. PMLR, 2023. 1, 3",
+            "[13] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2D & 3D face alignment problem? (and a dataset of 230,000 3D facial landmarks). In ICCV, pages 1021-1030. IEEE Computer Society, 2017. 2",
+            "[14] 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 9630-9640. IEEE, 2021. 1",
+            "[15] Oscar Chang, Hank Liao, Dmitriy Serdyuk, Ankit Shahy, and Olivier Siohan. Conformer is all you need for visual"
+        ],
+        "bbox": [
+            78,
+            145,
+            470,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "speech recognition. In ICASSP, pages 10136-10140. IEEE, 2024. 1, 8, 5",
+            "[16] Brian Chen, Andrew Rouditchenko, Kevin Duarte, Hilde Kuehne, Samuel Thomas, Angie W. Boggust, Rameswar Panda, Brian Kingsbury, Rogério Feris, David Harwath, James R. Glass, Michael Picheny, and Shih-Fu Chang. Multimodal clustering networks for self-supervised learning from unlabeled videos. In ICCV, pages 7992-8001. IEEE, 2021. 2",
+            "[17] Xize Cheng, Tao Jin, Linjun Li, Wang Lin, Xinyu Duan, and Zhou Zhao. OpenSR: Open-modality speech recognition via maintaining multi-modality alignment. In ACL (1), pages 6592-6607. Association for Computational Linguistics, 2023. 6",
+            "[18] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. VoxCeleb2: Deep speaker recognition. In INTERSPEECH, pages 1086-1090. ISCA, 2018. 6",
+            "[19] Soo-Whan Chung, Joon Son Chung, and Hong-Goo Kang. Perfect match: Improved cross-modal embeddings for audiovisual synchronisation. In ICASSP, pages 3965–3969. IEEE, 2019. 5",
+            "[20] Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. CoRR, abs/2307.08691, 2023. 2",
+            "[21] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In NeurIPS, 2022. 2",
+            "[22] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT (1), pages 4171–4186. Association for Computational Linguistics, 2019. 1, 2",
+            "[23] Angela Fan, Edouard Grave, and Armand Joulin. Reducing Transformer depth on demand with structured dropout. In ICLR. OpenReview.net, 2020. 2",
+            "[24] Dalu Feng, Shuang Yang, and Shiguang Shan. An efficient software for building lip reading models without pains. In ICME Workshops, pages 1-2. IEEE, 2021. 1",
+            "[25] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In ICLR. OpenReview.net, 2023. 2",
+            "[26] Hossein Hadian, Hossein Sameti, Daniel Povey, and Sanjeev Khudanpur. End-to-end speech recognition using lattice-free MMI. In INTERSPEECH, pages 12-16. ISCA, 2018. 5",
+            "[27] Alexandros Haliassos, Pingchuan Ma, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Jointly learning visual and auditory speech representations from raw data. In ICLR. Open-Review.net, 2023. 2, 6, 7, 8, 1, 4, 5",
+            "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 15979-15988. IEEE, 2022. 2",
+            "[29] François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia A. Tomashenko, and Yannick Esteve. TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In SPECOM, pages 198-208. Springer, 2018. 2"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "27077",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[30] Wei-Ning Hsu and Bowen Shi. u-HuBERT: Unified mixed-modal speech pretraining and zero-shot transfer to unlabeled modality. In NeurIPS, 2022. 1",
+            "[31] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. HuBERT: Self-supervised speech representation learning by masked prediction of hidden units. IEEE ACM Trans. Audio Speech Lang. Process., 29:3451-3460, 2021. 1",
+            "[32] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurIPS, pages 7774-7785, 2018. 2",
+            "[33] Aleksandr Laptev, Somshubra Majumdar, and Boris Ginsburg. CTC variations through new WFST topologies. In INTERSPEECH, pages 1041-1045. ISCA, 2022. 5",
+            "[34] Jiachen Lian, Alexei Baevski, Wei-Ning Hsu, and Michael Auli. AV-data2vec: Self-supervised learning of audio-visual speech representations with contextualized target representations. In ASRU. IEEE, 2023. 2, 6, 7, 8, 1, 5",
+            "[35] Paul Pu Liang, Zihao Deng, Martin Q. Ma, James Y. Zou, Louis-Philippe Morency, and Ruslan Salakhutdinov. Factorized contrastive learning: Going beyond multi-view redundancy. In NeurIPS, 2023. 2",
+            "[36] Alexander H. Liu, Heng-Jui Chang, Michael Auli, Wei-Ning Hsu, and Jim Glass. DinoSR: Self-distillation and online clustering for self-supervised speech representation learning. In NeurIPS, 2023. 4, 2",
+            "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR (Poster)*. OpenReview.net, 2019. 2",
+            "[38] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Bjorn W. Schuller, and Maja Pantic. LiRA: Learning visual speech representations from audio through self-supervision. In *Interspeech*, pages 3011–3015. ISCA, 2021. 2",
+            "[39] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with Conformers. In ICASSP, pages 7613-7617. IEEE, 2021. 1, 6, 7, 8, 4, 5",
+            "[40] Pingchuan Ma, Stavros Petridis, and Maja Pantic. Visual speech recognition for multiple languages in the wild. Nat. Mac. Intell., 4(11):930-939, 2022. 1, 6, 8, 2, 4, 5",
+            "[41] Pingchuan Ma, Alexandros Haliassos, Adriana Fernandez-Lopez, Honglie Chen, Stavros Petridis, and Maja Pantic. Auto-AVSR: Audio-Visual speech recognition with automatic labels. In ICASSP, pages 1-5. IEEE, 2023. 1, 3, 7, 8, 4, 5",
+            "[42] Takaki Makino, Hank Liao, Yannis M. Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In ASRU, pages 905–912. IEEE, 2019. 1, 8, 5",
+            "[43] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640. IEEE, 2019. 2",
+            "[44] Arsha Nagrani, Chen Sun, David Ross, Rahul Sukthankar, Cordelia Schmid, and Andrew Zisserman. Speech2Action: Cross-modal supervision for action recognition. In CVPR, pages 10314-10323. Computer Vision Foundation / IEEE, 2020. 2"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[45] Xichen Pan, Peiyu Chen, Yichen Gong, Helong Zhou, Xinbing Wang, and Zhouhan Lin. Leveraging unimodal self-supervised learning for multimodal audio-visual speech recognition. In ACL (1), pages 4491–4503. Association for Computational Linguistics, 2022. 7, 4",
+            "[46] Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V. Le. SpecAugment: A simple data augmentation method for automatic speech recognition. In INTERSPEECH, pages 2613-2617. ISCA, 2019. 2",
+            "[47] Sooyoung Park, Arda Senocak, and Joon Son Chung. MarginNCE: Robust sound localization with a negative margin. In ICASSP, pages 1-5. IEEE, 2023. 5, 2",
+            "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Z. 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 NeurIPS, pages 8024-8035, 2019. 2",
+            "[49] Mandela Patrick, Yuki Markus Asano, Polina Kuznetsova, Ruth Fong, João F. Henriques, Geoffrey Zweig, and Andrea Vedaldi. On compositions of transformations in contrastive self-supervised learning. In ICCV, pages 9557-9567. IEEE, 2021. 2",
+            "[50] Trung Pham, Chaoning Zhang, Axi Niu, Kang Zhang, and Chang D. Yoo. On the pros and cons of momentum encoder in self-supervised visual representation learning. CoRR, abs/2208.05744, 2022. 4",
+            "[51] Daniel Povey, Vijayaditya Peddinti, Daniel Galvez, Pegah Ghahremani, Vimal Manohar, Xingyu Na, Yiming Wang, and Sanjeev Khudanpur. Purely sequence-trained neural networks for ASR based on lattice-free MMI. In INTER-SPEECH, pages 2751-2755. ISCA, 2016. 5",
+            "[52] K. R. Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In CVPR, pages 5152-5162. IEEE, 2022. 8, 4, 5",
+            "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2",
+            "[54] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICML, pages 28492-28518. PMLR, 2023. 6",
+            "[55] Jordi Robert-Ribes, Jean-Luc Schwartz, Tahar Lallouache, and Pierre Escudier. Complementarity and synergy in b-modal speech: Auditory, visual, and audio-visual identification of french oral vowels in noise. The Journal of the Acoustical Society of America, 103(6):3677-3689, 1998. 1",
+            "[56] Steffen Schneider, Alexei Baevski, Ronan Collobert, and Michael Auli. wav2vec: Unsupervised pre-training for speech recognition. In INTERSPEECH, pages 3465-3469. ISCA, 2019. 3"
+        ],
+        "bbox": [
+            501,
+            90,
+            892,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "27078",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[57] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Audiovisual speech recognition is worth  $32 \\times 32 \\times 8$  voxels. In ASRU, pages 796-802. IEEE, 2021. 2, 8, 5",
+            "[58] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video. In INTER-SPEECH, pages 2833–2837. ISCA, 2022. 2, 8, 5",
+            "[59] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In ICLR. OpenReview.net, 2022. 1, 2, 3, 6, 7, 8, 4, 5",
+            "[60] Bowen Shi, Wei-Ning Hsu, and Abdelrahman Mohamed. Robust self-supervised audio-visual speech recognition. In INTERSPEECH, pages 2118–2122. ISCA, 2022. 7, 8, 4, 5",
+            "[61] Bowen Shi, Abdelrahman Mohamed, and Wei-Ning Hsu. Learning lip-based audio-visual speaker embeddings with AV-HuBERT. In INTERSPEECH, pages 4785–4789. ISCA, 2022. 1",
+            "[62] Brendan Shillingford, Yannis M. Assael, Matthew W. Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, Marie Mulville, Misha Denil, Ben Coppin, Ben Laurie, Andrew W. Senior, and Nando de Freitas. Large-scale visual speech recognition. In INTERSPEECH, pages 4135-4139. ISCA, 2019. 8, 5",
+            "[63] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogério Feris, David Harwath, James R. Glass, and Hilde Kuehne. Everything at once - multi-modal fusion Transformer for video retrieval. In CVPR, pages 19988-19997. IEEE, 2022. 2",
+            "[64] Themos Stafylakis and Georgios Tzimiropoulos. Combining residual networks with LSTMs for lipreading. In INTER-SPEECH, pages 3652-3656. ISCA, 2017. 3",
+            "[65] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. CoRR, abs/2104.09864, 2021. 3",
+            "[66] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT: Pre-training of generic visual-linguistic representations. In ICLR. OpenReview.net, 2020. 2",
+            "[67] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. VideoBERT: A joint model for video and language representation learning. In ICCV, pages 7463-7472. IEEE, 2019. 2",
+            "[68] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In ECCV (11), pages 776-794. Springer, 2020. 2",
+            "[69] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image Transformers. In ICCV, pages 32-42. IEEE, 2021. 2",
+            "[70] Kristin J Van Engen, Avanti Dey, Mitchell S Sommers, and Jonathan E Peelle. Audiovisual speech perception: Moving beyond mcgurk. The Journal of the Acoustical Society of America, 152(6):3216-3225, 2022. 1",
+            "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. 3, 2"
+        ],
+        "bbox": [
+            78,
+            92,
+            467,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[72] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In CVPR, pages 2495-2504. Computer Vision Foundation / IEEE, 2021. 7",
+            "[73] Paul L. Williams and Randall D. Beer. Nonnegative decomposition of multivariate information. CoRR, abs/1004.2515, 2010. 3",
+            "[74] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In CVPR, pages 14421-14430. Computer Vision Foundation / IEEE, 2020. 5",
+            "[75] Jianwei Yu, Shi-Xiong Zhang, Jian Wu, Shahram Ghorbani, Bo Wu, Shiyin Kang, Shansong Liu, Xunying Liu, Helen Meng, and Dong Yu. Audio-visual recognition of overlapped speech for the LRS2 dataset. In ICASSP, pages 6984–6988. IEEE, 2020. 7, 4",
+            "[76] Jing-Xuan Zhang, Genshun Wan, Zhen-Hua Ling, Jia Pan, Jianqing Gao, and Cong Liu. Self-supervised audiovisual speech representations learning by multimodal self-distillation. In ICASSP, pages 1-5. IEEE, 2023. 2, 7, 8, 1, 5",
+            "[77] Xiaohui Zhang, Vimal Manohar, David Zhang, Frank Zhang, Yangyang Shi, Nayan Singhal, Julian Chan, Fuchun Peng, Yatharth Saraf, and Mike Seltzer. On lattice-free boosted MMI training of HMM and CTC-based full-context ASR models. In ASRU, pages 1026-1033. IEEE, 2021. 5",
+            "[78] Ya Zhao, Rui Xu, Xinchao Wang, Peng Hou, Haihong Tang, and Mingli Song. Hearing lips: Improving lip reading by distilling speech recognizers. In AAAI, pages 6917-6924. AAAI Press, 2020. 2",
+            "[79] Jian Zhu, Cong Zhang, and David Jurgens. Phone-to-audio alignment without text: A semi-supervised approach. In ICASSP, pages 8167-8171. IEEE, 2022. 8",
+            "[80] Qiushi Zhu, Long Zhou, Ziqiang Zhang, Shujie Liu, Binxing Jiao, Jie Zhang, Lirong Dai, Daxin Jiang, Jinyu Li, and Furu Wei. VATLM: Visual-Audio-Text pre-training with unified masked prediction for speech representation learning. IEEE Transactions on Multimedia, pages 1–11, 2023. 1, 6, 7, 4, 5"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            599
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "27079",
+        "bbox": [
+            478,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 10
+    }
+]