Add Batch f39567fb-a11c-4754-b3c0-ca410ec84d3d data

.gitattributes

@@ -4892,3 +4892,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2024/Structure-Aware[[:space:]]Sparse-View[[:space:]]X-ray[[:space:]]3D[[:space:]]Reconstruction/e0ae7df8-d3f1-46c2-a34a-52862215b941_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/Structure-Guided[[:space:]]Adversarial[[:space:]]Training[[:space:]]of[[:space:]]Diffusion[[:space:]]Models/b70a5045-dbc8-47eb-880e-4252d02e0e55_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/Structured[[:space:]]Gradient-based[[:space:]]Interpretations[[:space:]]via[[:space:]]Norm-Regularized[[:space:]]Adversarial[[:space:]]Training/25726c0b-b33a-4e50-8b53-f013deb7f334_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Structured[[:space:]]Model[[:space:]]Probing_[[:space:]]Empowering[[:space:]]Efficient[[:space:]]Transfer[[:space:]]Learning[[:space:]]by[[:space:]]Structured[[:space:]]Regularization/fcd62d92-907d-4f7f-bc80-b88db83abf7f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/StyLitGAN_[[:space:]]Image-Based[[:space:]]Relighting[[:space:]]via[[:space:]]Latent[[:space:]]Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Style[[:space:]]Aligned[[:space:]]Image[[:space:]]Generation[[:space:]]via[[:space:]]Shared[[:space:]]Attention/0cae3168-e63f-4f71-908a-5bdeae8d8200_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Style[[:space:]]Blind[[:space:]]Domain[[:space:]]Generalized[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]via[[:space:]]Covariance[[:space:]]Alignment[[:space:]]and[[:space:]]Semantic[[:space:]]Consistence[[:space:]]Contrastive[[:space:]]Learning/1e340b0f-2934-49e8-9933-238d4fb891ff_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Style[[:space:]]Injection[[:space:]]in[[:space:]]Diffusion_[[:space:]]A[[:space:]]Training-free[[:space:]]Approach[[:space:]]for[[:space:]]Adapting[[:space:]]Large-scale[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]Style[[:space:]]Transfer/43978830-73bd-4db5-8b8c-57888ce93c6e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/StyleCineGAN_[[:space:]]Landscape[[:space:]]Cinemagraph[[:space:]]Generation[[:space:]]using[[:space:]]a[[:space:]]Pre-trained[[:space:]]StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SuGaR_[[:space:]]Surface-Aligned[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Efficient[[:space:]]3D[[:space:]]Mesh[[:space:]]Reconstruction[[:space:]]and[[:space:]]High-Quality[[:space:]]Mesh[[:space:]]Rendering/8c37e125-7381-4a71-8f2b-015c3d8a066a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SubT-MRS[[:space:]]Dataset_[[:space:]]Pushing[[:space:]]SLAM[[:space:]]Towards[[:space:]]All-weather[[:space:]]Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Summarize[[:space:]]the[[:space:]]Past[[:space:]]to[[:space:]]Predict[[:space:]]the[[:space:]]Future_[[:space:]]Natural[[:space:]]Language[[:space:]]Descriptions[[:space:]]of[[:space:]]Context[[:space:]]Boost[[:space:]]Multimodal[[:space:]]Object[[:space:]]Interaction[[:space:]]Anticipation/5f051557-0067-4cd2-89f8-ff27dddafa6f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Super-Resolution[[:space:]]Reconstruction[[:space:]]from[[:space:]]Bayer-Pattern[[:space:]]Spike[[:space:]]Streams/ac654108-cae7-444a-a7e8-a0faf89a682d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SuperNormal_[[:space:]]Neural[[:space:]]Surface[[:space:]]Reconstruction[[:space:]]via[[:space:]]Multi-View[[:space:]]Normal[[:space:]]Integration/a4b33f00-d1f9-45e2-a62c-bb25956b58ec_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SuperPrimitive_[[:space:]]Scene[[:space:]]Reconstruction[[:space:]]at[[:space:]]a[[:space:]]Primitive[[:space:]]Level/556fb01c-1445-4c08-b116-0426ba52ea17_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SuperSVG_[[:space:]]Superpixel-based[[:space:]]Scalable[[:space:]]Vector[[:space:]]Graphics[[:space:]]Synthesis/3fa1e5c3-5890-4208-976c-19acdb43c796_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Supervised[[:space:]]Anomaly[[:space:]]Detection[[:space:]]for[[:space:]]Complex[[:space:]]Industrial[[:space:]]Images/ef358e69-3d45-4c6c-aa99-f5a5ceb11e08_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Suppress[[:space:]]and[[:space:]]Rebalance_[[:space:]]Towards[[:space:]]Generalized[[:space:]]Multi-Modal[[:space:]]Face[[:space:]]Anti-Spoofing/91e475e8-d317-465c-893c-024b420e20a0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SurMo_[[:space:]]Surface-based[[:space:]]4D[[:space:]]Motion[[:space:]]Modeling[[:space:]]for[[:space:]]Dynamic[[:space:]]Human[[:space:]]Rendering/6283bde0-3818-458b-8029-7eb693822789_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SurroundSDF_[[:space:]]Implicit[[:space:]]3D[[:space:]]Scene[[:space:]]Understanding[[:space:]]Based[[:space:]]on[[:space:]]Signed[[:space:]]Distance[[:space:]]Field/213de3e8-6a97-4d16-96b2-ac03f2f0441a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SwiftBrush_[[:space:]]One-Step[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Model[[:space:]]with[[:space:]]Variational[[:space:]]Score[[:space:]]Distillation/13853f95-872e-49e0-9add-f414db12f2de_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SwitchLight_[[:space:]]Co-design[[:space:]]of[[:space:]]Physics-driven[[:space:]]Architecture[[:space:]]and[[:space:]]Pre-training[[:space:]]Framework[[:space:]]for[[:space:]]Human[[:space:]]Portrait[[:space:]]Relighting/ce409cac-fd5a-4614-9e6d-87e790ca611e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Symphonize[[:space:]]3D[[:space:]]Semantic[[:space:]]Scene[[:space:]]Completion[[:space:]]with[[:space:]]Contextual[[:space:]]Instance[[:space:]]Queries/ba02978b-283d-434b-af7a-d4b604064e5e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SynFog_[[:space:]]A[[:space:]]Photo-realistic[[:space:]]Synthetic[[:space:]]Fog[[:space:]]Dataset[[:space:]]based[[:space:]]on[[:space:]]End-to-end[[:space:]]Imaging[[:space:]]Simulation[[:space:]]for[[:space:]]Advancing[[:space:]]Real-World[[:space:]]Defogging[[:space:]]in[[:space:]]Autonomous[[:space:]]Driving/30117727-713c-4f8f-a215-de682288f1dd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SynSP_[[:space:]]Synergy[[:space:]]of[[:space:]]Smoothness[[:space:]]and[[:space:]]Precision[[:space:]]in[[:space:]]Pose[[:space:]]Sequences[[:space:]]Refinement/e165eedf-6d86-4e7b-90c3-ccc6312e093e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SyncMask_[[:space:]]Synchronized[[:space:]]Attentional[[:space:]]Masking[[:space:]]for[[:space:]]Fashion-centric[[:space:]]Vision-Language[[:space:]]Pretraining/92bc41a7-a993-4d23-a197-14e6dd1e0509_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SyncTalk_[[:space:]]The[[:space:]]Devil[[:space:]]is[[:space:]]in[[:space:]]the[[:space:]]Synchronization[[:space:]]for[[:space:]]Talking[[:space:]]Head[[:space:]]Synthesis/89645a31-8829-4d8f-939e-0d17fcee95a5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Synergistic[[:space:]]Global-space[[:space:]]Camera[[:space:]]and[[:space:]]Human[[:space:]]Reconstruction[[:space:]]from[[:space:]]Videos/908ee619-1331-452b-9ccb-4e5b4d116ac0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Synthesize[[:space:]]Diagnose[[:space:]]and[[:space:]]Optimize_[[:space:]]Towards[[:space:]]Fine-Grained[[:space:]]Vision-Language[[:space:]]Understanding/eca2ceba-84d3-4751-b38a-1ba7e74c92ba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Synthesize[[:space:]]Step-by-Step_[[:space:]]Tools[[:space:]]Templates[[:space:]]and[[:space:]]LLMs[[:space:]]as[[:space:]]Data[[:space:]]Generators[[:space:]]for[[:space:]]Reasoning-Based[[:space:]]Chart[[:space:]]VQA/58e8331e-f572-44c9-b724-02ca5cd604a2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Systematic[[:space:]]Comparison[[:space:]]of[[:space:]]Semi-supervised[[:space:]]and[[:space:]]Self-supervised[[:space:]]Learning[[:space:]]for[[:space:]]Medical[[:space:]]Image[[:space:]]Classification/a33fa80f-bfa4-45b2-878d-abb01f3d724e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/T-VSL_[[:space:]]Text-Guided[[:space:]]Visual[[:space:]]Sound[[:space:]]Source[[:space:]]Localization[[:space:]]in[[:space:]]Mixtures/2ac9042c-6511-4205-8b71-c79e7f00a075_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/T4P_[[:space:]]Test-Time[[:space:]]Training[[:space:]]of[[:space:]]Trajectory[[:space:]]Prediction[[:space:]]via[[:space:]]Masked[[:space:]]Autoencoder[[:space:]]and[[:space:]]Actor-specific[[:space:]]Token[[:space:]]Memory/11bec0da-9290-4827-bb16-e062c2aadc7c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TACO_[[:space:]]Benchmarking[[:space:]]Generalizable[[:space:]]Bimanual[[:space:]]Tool-ACtion-Object[[:space:]]Understanding/b3378a53-d4f8-49d5-95a4-af3ae029dc3c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TAMM_[[:space:]]TriAdapter[[:space:]]Multi-Modal[[:space:]]Learning[[:space:]]for[[:space:]]3D[[:space:]]Shape[[:space:]]Understanding/f474f9b7-e45c-4a08-b76f-0b118ae61d7f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TASeg_[[:space:]]Temporal[[:space:]]Aggregation[[:space:]]Network[[:space:]]for[[:space:]]LiDAR[[:space:]]Semantic[[:space:]]Segmentation/745b72e4-2dd6-4811-ade0-f52413f14141_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TCP_Textual-based[[:space:]]Class-aware[[:space:]]Prompt[[:space:]]tuning[[:space:]]for[[:space:]]Visual-Language[[:space:]]Model/137896e1-f0a6-4b80-849c-9a14cbb3250c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TE-TAD_[[:space:]]Towards[[:space:]]Full[[:space:]]End-to-End[[:space:]]Temporal[[:space:]]Action[[:space:]]Detection[[:space:]]via[[:space:]]Time-Aligned[[:space:]]Coordinate[[:space:]]Expression/63fde73a-d557-47b6-aee9-074fa8fa7370_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TEA_[[:space:]]Test-time[[:space:]]Energy[[:space:]]Adaptation/c8c44200-6095-44b9-b188-9639c747d9ec_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TFMQ-DM_[[:space:]]Temporal[[:space:]]Feature[[:space:]]Maintenance[[:space:]]Quantization[[:space:]]for[[:space:]]Diffusion[[:space:]]Models/a91d56f7-042a-42e5-9c77-ba0dbaac2109_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/THRONE_[[:space:]]An[[:space:]]Object-based[[:space:]]Hallucination[[:space:]]Benchmark[[:space:]]for[[:space:]]the[[:space:]]Free-form[[:space:]]Generations[[:space:]]of[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models/da9cf0d5-8d8b-4df8-b396-af5ff1ab4d0d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TI2V-Zero_[[:space:]]Zero-Shot[[:space:]]Image[[:space:]]Conditioning[[:space:]]for[[:space:]]Text-to-Video[[:space:]]Diffusion[[:space:]]Models/17566a41-32ff-45d4-9eb1-0099d95d87b1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TIGER_[[:space:]]Time-Varying[[:space:]]Denoising[[:space:]]Model[[:space:]]for[[:space:]]3D[[:space:]]Point[[:space:]]Cloud[[:space:]]Generation[[:space:]]with[[:space:]]Diffusion[[:space:]]Process/8fe5d2c1-4f78-435e-8bde-f7a4afed8604_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TIM_[[:space:]]A[[:space:]]Time[[:space:]]Interval[[:space:]]Machine[[:space:]]for[[:space:]]Audio-Visual[[:space:]]Action[[:space:]]Recognition/03af8456-92c0-4a22-a6d9-94775506b39a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TRINS_[[:space:]]Towards[[:space:]]Multimodal[[:space:]]Language[[:space:]]Models[[:space:]]that[[:space:]]Can[[:space:]]Read/527c7573-3fb6-42ad-a14f-b7b957e97243_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TRIP_[[:space:]]Temporal[[:space:]]Residual[[:space:]]Learning[[:space:]]with[[:space:]]Image[[:space:]]Noise[[:space:]]Prior[[:space:]]for[[:space:]]Image-to-Video[[:space:]]Diffusion[[:space:]]Models/51572829-1348-4002-a40a-f8702f404317_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TTA-EVF_[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]for[[:space:]]Event-based[[:space:]]Video[[:space:]]Frame[[:space:]]Interpolation[[:space:]]via[[:space:]]Reliable[[:space:]]Pixel[[:space:]]and[[:space:]]Sample[[:space:]]Estimation/fa811509-bdf9-4c56-97db-551fcc2b0a3d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TULIP_[[:space:]]Multi-camera[[:space:]]3D[[:space:]]Precision[[:space:]]Assessment[[:space:]]of[[:space:]]Parkinson's[[:space:]]Disease/532b2907-27ca-48aa-bd4d-3e8291e221ab_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TULIP_[[:space:]]Transformer[[:space:]]for[[:space:]]Upsampling[[:space:]]of[[:space:]]LiDAR[[:space:]]Point[[:space:]]Clouds/52ac8b22-207c-41ff-adcb-c9f4c82bca08_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TUMTraf[[:space:]]V2X[[:space:]]Cooperative[[:space:]]Perception[[:space:]]Dataset/5f2a634b-45a2-4d92-81e8-cda538302a8f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Tackling[[:space:]]the[[:space:]]Singularities[[:space:]]at[[:space:]]the[[:space:]]Endpoints[[:space:]]of[[:space:]]Time[[:space:]]Intervals[[:space:]]in[[:space:]]Diffusion[[:space:]]Models/992f5837-40b9-47fe-b9bd-9d2e369d7fdb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Tactile-Augmented[[:space:]]Radiance[[:space:]]Fields/f9e9c348-b82c-415a-b07e-5b945a68c375_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Tailored[[:space:]]Visions_[[:space:]]Enhancing[[:space:]]Text-to-Image[[:space:]]Generation[[:space:]]with[[:space:]]Personalized[[:space:]]Prompt[[:space:]]Rewriting/7aad987a-39f2-4096-8aa1-a539bbaacb9c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Taming[[:space:]]Mode[[:space:]]Collapse[[:space:]]in[[:space:]]Score[[:space:]]Distillation[[:space:]]for[[:space:]]Text-to-3D[[:space:]]Generation/695a2b86-fc93-420e-a626-d23f402233d8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Taming[[:space:]]Self-Training[[:space:]]for[[:space:]]Open-Vocabulary[[:space:]]Object[[:space:]]Detection/08e6cd2d-6b02-4147-b946-b39f423452c4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Taming[[:space:]]Stable[[:space:]]Diffusion[[:space:]]for[[:space:]]Text[[:space:]]to[[:space:]]360[[:space:]]Panorama[[:space:]]Image[[:space:]]Generation/3259ff45-5fcb-4059-9736-190d9ce81cb7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Taming[[:space:]]the[[:space:]]Tail[[:space:]]in[[:space:]]Class-Conditional[[:space:]]GANs_[[:space:]]Knowledge[[:space:]]Sharing[[:space:]]via[[:space:]]Unconditional[[:space:]]Training[[:space:]]at[[:space:]]Lower[[:space:]]Resolutions/301b4dc8-6f1d-458c-a509-0fe38145fc73_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Targeted[[:space:]]Representation[[:space:]]Alignment[[:space:]]for[[:space:]]Open-World[[:space:]]Semi-Supervised[[:space:]]Learning/b33baf4c-8c46-4e13-a308-0177b1e2eee6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Task-Adaptive[[:space:]]Saliency[[:space:]]Guidance[[:space:]]for[[:space:]]Exemplar-free[[:space:]]Class[[:space:]]Incremental[[:space:]]Learning/ef2d4502-c3bc-43a6-a846-89e23e77cec3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Task-Aware[[:space:]]Encoder[[:space:]]Control[[:space:]]for[[:space:]]Deep[[:space:]]Video[[:space:]]Compression/b978be40-5174-4abc-9649-d7daa2d37c89_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Task-Conditioned[[:space:]]Adaptation[[:space:]]of[[:space:]]Visual[[:space:]]Features[[:space:]]in[[:space:]]Multi-Task[[:space:]]Policy[[:space:]]Learning/5ebc706d-8b34-4357-9c79-967e39bda2a5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Task-Customized[[:space:]]Mixture[[:space:]]of[[:space:]]Adapters[[:space:]]for[[:space:]]General[[:space:]]Image[[:space:]]Fusion/7c9be988-491f-4562-822c-11cfddbdaa7d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Task-Driven[[:space:]]Exploration_[[:space:]]Decoupling[[:space:]]and[[:space:]]Inter-Task[[:space:]]Feedback[[:space:]]for[[:space:]]Joint[[:space:]]Moment[[:space:]]Retrieval[[:space:]]and[[:space:]]Highlight[[:space:]]Detection/a6a833ff-f398-48fe-909b-70ff81bf1b94_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Task-Driven[[:space:]]Wavelets[[:space:]]using[[:space:]]Constrained[[:space:]]Empirical[[:space:]]Risk[[:space:]]Minimization/d6dc5c43-4c52-4779-8c6a-d513a47a71be_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Task-aligned[[:space:]]Part-aware[[:space:]]Panoptic[[:space:]]Segmentation[[:space:]]through[[:space:]]Joint[[:space:]]Object-Part[[:space:]]Representations/9dbd5f9f-08af-4d23-b3b6-162e326d318c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Task2Box_[[:space:]]Box[[:space:]]Embeddings[[:space:]]for[[:space:]]Modeling[[:space:]]Asymmetric[[:space:]]Task[[:space:]]Relationships/dd90eec4-a932-4045-a86b-cf06d264eadc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TeMO_[[:space:]]Towards[[:space:]]Text-Driven[[:space:]]3D[[:space:]]Stylization[[:space:]]for[[:space:]]Multi-Object[[:space:]]Meshes/d9ed6877-0911-45bb-b575-e2ba93e2db7d_origin.pdf filter=lfs diff=lfs merge=lfs -text

2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/fcd62d92-907d-4f7f-bc80-b88db83abf7f_content_list.json

@@ -0,0 +1,1671 @@
+[
+    {
+        "type": "text",
+        "text": "Structured Model Probing: Empowering Efficient Transfer Learning by Structured Regularization",
+        "text_level": 1,
+        "bbox": [
+            120,
+            106,
+            838,
+            151
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zhi-Fan Wu Chaojie Mao Xue Wang Jianwen Jiang Yiliang Lv Rong Jin Alibaba Group",
+        "bbox": [
+            140,
+            178,
+            815,
+            220
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            279,
+            313,
+            296
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite encouraging results from recent developments in transfer learning for adapting pre-trained model to downstream tasks, the performance of model probing is still lagging behind the state-of-the-art parameter efficient tuning methods. Our investigation reveals that existing model probing methods perform well for the easy case when the source domain (where models are pre-trained) and the adapted domain are similar, but fail for the difficult case when the two domains are significantly different. Simply incorporating features extracted from multiple layers and increasing complexity of the probing model can mitigate the gap in the difficult case, but degrades the performance in the easy case. To address this challenge, we propose structured model probing (SMP) that is able to deliver good performance for both cases through structured regularization. The regularization performs feature selection leveraging model structure as a prior, and controls the complexity of the probing model through the weights of selected structures. This enables us to construct a simple adaptation model, with a small number of selected features and a linear prediction model, for the easy case; and to automatically increase the complexity of adaptation model, with a large number of selected features and a non-linear model, for the difficult case. Our extensive empirical studies show that SMP significantly outperforms the state-of-the-art methods for parameter efficient tuning, and at the same time, still maintains the advantage of computational efficiency for probing-based methods.",
+        "bbox": [
+            76,
+            313,
+            473,
+            720
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            752,
+            209,
+            768
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The parameters of models have grown drastically with the rapid development of deep learning [5, 37], posing challenges for adapting pre-trained models to downstream tasks. The widely used transfer strategy, fully fine-tuning, becomes infeasible due to excessive computation and storage overhead. Recently, parameter efficient tuning [19, 23] has gained attention for reducing storage costs and improving tuning performance with few trainable parameters when lim",
+        "bbox": [
+            75,
+            779,
+            470,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/63f688131516d3ed72daf25e836154d3ffd2d13587e9edb41098c34830f2f3e3.jpg",
+        "image_caption": [
+            "Figure 1. Performance comparison on different dataset groups (left) and mean relative adaptation time (right) on VTAB-1k benchmark."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            282,
+            888,
+            426
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ited data are available for downstream tasks. Despite the progress, parameter efficient tuning methods still face significant computational overhead in training due to the need to perform forward and backward passes through the pretrained model at each training step.",
+        "bbox": [
+            496,
+            489,
+            893,
+            565
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Model probing, which leverages features extracted from the pre-trained model to perform adaptation, is promising due to its efficiency and simplicity: The training procedure is efficient since the pre-trained model is fixed and only performs one forward propagation to extract features, and probing model is decoupled from pre-trained model that makes it convenient to deploy small sized task-specific downstream models. However, it is a nuisance that probing methods are usually outperformed by state-of-the-art parameter efficient tuning methods. This is true even after several recent improvements made for the probing methods [16, 45]. In this paper, we aim to improve the performance of model probing significantly without sacrificing its simplicity and efficiency.",
+        "bbox": [
+            496,
+            566,
+            895,
+            763
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We first investigate when model probing methods perform worse than the tuning-based methods on widely used visual task adaptation benchmark VTAB-1k [44] as shown in Figure 1. Target domains are categorized as easy and difficult cases based on their similarity to the source domain where the pre-trained model is trained on (see Section 3.1 for details). We find simple linear probing yields comparable performance as the two tuning-based approaches (finetuning and VPT [23]) for easy cases, whereas a significant",
+        "bbox": [
+            496,
+            763,
+            895,
+            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,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "16849",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/8976b25f7d62a3b031921846e07f4362b5639759867d578f4903a604d17738bc.jpg",
+        "image_caption": [
+            "Figure 2. Adaptation pipelines of Model Tuning (left) and the proposed Structured Model Probing (right)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            76,
+            85,
+            401,
+            244
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/7160d55c197b3be1ab48dc688ed30909c72b2fb4a9022aedc0f7297b00ee4323.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            413,
+            85,
+            890,
+            244
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "performance gap is observed between linear probing and tuning-based methods for difficult cases. This suggests that model probing exhibits different behavior for various target domains, and performance gap arises for target domains that dissimilar to the source domain.",
+        "bbox": [
+            75,
+            286,
+            470,
+            361
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "This is not surprising that we expect a more complicated model than a linear probing to compensate the difference between the source domain and the target domain. In general, we can increase model complexity in two ways, either by introducing more features extracted from the pre-trained model, or by introducing a non-linear model for prediction. A main challenge is how to develop a unified framework of model probing that is able to handle both the easy and difficult adaptations effectively. This is because, according to our study in Section 3, for easy cases, a complicated probing model (with more features and non-linear probing) performs significantly worse than a simple linear probing.",
+        "bbox": [
+            75,
+            363,
+            468,
+            545
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To this end, we propose structured model probing, or SMP for short, that effectively address both easy and difficult cases of domain adaptation. The key to our unified framework is structured regularization. It performs feature selection leveraging the structure of the pre-trained model as a prior, and controls the complexity of the probing model through the weights assigned to selected features. Thus, for the easy case, this regularization constructs a simple linear model with a few selected features; and for the difficult case, it automatically introduces more features from the pre-trained model and more non-linearity in the prediction module based on the value of loss function, as shown in Figure 2. Our empirical studies show that the proposed structured model probing performs significantly better than state-of-the-art tuning-based approaches, while maintaining the advantage of computational efficiency.",
+        "bbox": [
+            75,
+            547,
+            470,
+            787
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our main contributions are as follows:",
+        "bbox": [
+            96,
+            789,
+            351,
+            803
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "1. We observe and demonstrate the conflicting behaviors of model probing: a complicated probing model with more features and non-linear transformation can improve the performance for difficult cases of domain adaptation, but it yields worse performance than linear probing for easy cases of domain adaptation.",
+        "bbox": [
+            78,
+            809,
+            470,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "2. We propose structured model probing (SMP) based on structured regularization. This regularization enables us to construct simple probing model for easy tasks, and automatically increase model complexity for difficult tasks, to simultaneously address both the easy and difficult cases of domain adaptation.",
+            "3. We conduct extensive empirical studies on various adaptation tasks. The proposed SMP method shows superior experimental results compared to state-of-the-art probing and tuning methods, and requires significantly lower training costs compared to tuning methods."
+        ],
+        "bbox": [
+            500,
+            286,
+            893,
+            453
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            470,
+            640,
+            486
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Transfer learning has been extensively studied in vision domain [34, 48]. It aims to improve the performance on target domain by leveraging the related information contained in source domain. Transfer learning based on pre-trained model [8, 20, 21] has received significant attention due to its simplicity and good empirical performance [23, 45].",
+        "bbox": [
+            496,
+            500,
+            893,
+            590
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Model tuning adapts the pre-trained model to new tasks via updating model parameters. Fully fine-tuning is a widely used strategies due to its efficacy and simplicity [9, 25, 44]. However, it suffers from huge storage and computation overhead due to the ever-expanding size of large-scale pre-trained model [37, 40]. Heuristic approaches have been proposed to improve fine-tuning performance [3, 6, 17, 18, 27]. Parameter efficient tuning [19] aims to reduce the storage requirement of transferred model by updating a few parameters attached to the pre-trained model [23, 47], like Prompt Tuning [29, 30] or Adapter [22]. While originally studied in natural language processing tasks due to the dominance of large-scale pre-trained language models [5, 14], there is growing interest in applying these methods to computer vision tasks [7, 23, 31, 41, 46, 47]. Though parameter efficient tuning reduces storage cost significantly, the training cost remains high due to the forward and backward propagation need to pass through the entire pre-trained model [11].",
+        "bbox": [
+            496,
+            594,
+            893,
+            867
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Model probing utilizes frozen extracted features for adapting the pre-trained model to new tasks, thus achieving effi",
+        "bbox": [
+            498,
+            869,
+            893,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "16850",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "cient adaptation. The most popular method, linear probing, suffers from performance degradation [25], which limits its application in transfer learning [20, 21]. Recent methods [16, 36, 45] have been proposed to improve probing performance by leveraging side network [38, 45] or incorporating intermediate representations [2, 16, 32], but there is still a significant performance gap compared to tuning methods. This motivates us to investigate the limitation of probing methods, and come up with a better solution.",
+        "bbox": [
+            75,
+            90,
+            472,
+            227
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Understanding the Limitation of Model Probing for Transfer Learning",
+        "text_level": 1,
+        "bbox": [
+            76,
+            242,
+            472,
+            277
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we analyze the performance of probing methods across various downstream domains to understand their limitations, motivated by the observation in Figure 1. We assess feature informativeness for different tasks and investigate the necessity of non-linear transformation.",
+        "bbox": [
+            75,
+            285,
+            470,
+            361
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Analysis Setup",
+        "text_level": 1,
+        "bbox": [
+            76,
+            371,
+            227,
+            387
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Pre-trained model and source data. The ViT-B/16 [15] is used as the backbone because of the extensive usage of transformer-based models. The model is pre-trained on ImageNet-21k [12] in a supervised learning manner.",
+        "bbox": [
+            75,
+            398,
+            468,
+            459
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Target data. We use the VTAB-1k benchmark, consisting of 19 datasets that cover natural images (natural), images captured by specialist equipment (specialized), and images generated from simulated environments (structured), as our target data for transfer learning. We follow the setting in [44] and use 1000 samples as the training set per task to evaluate adaptation with limited data as suggested in [44].",
+        "bbox": [
+            75,
+            462,
+            468,
+            568
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Domain similarity. Several previous works have proposed metrics to measure the similarity between source domain and target domain [1, 10, 13, 16, 33]. Here, we adopt the domain similarity measure defined in [16]:",
+        "bbox": [
+            75,
+            571,
+            468,
+            632
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nD o m a i n \\quad \\text {S i m i l a r i t y} = \\mathrm {A c c} _ {\\text {L i n e a r}} - \\mathrm {A c c} _ {\\text {S c r a t c h}}. \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            127,
+            645,
+            468,
+            660
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "A higher domain similarity suggests that the features extracted from a pre-trained model can be utilized for the target domain, leading to performance gains via a linear classifier compared to training from scratch. Conversely, a lower domain similarity indicates that training from scratch on the target domain may be more effective than using extracted features. Thus, the domain similarity is evaluated based on the benefit gained by a linear classifier compared to training from scratch. We also report the label-feature correlation [13] in Suppl. C.8, which exhibits a high Spearman rank correlation (0.854) with the domain similarity scores.",
+        "bbox": [
+            75,
+            672,
+            470,
+            839
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Domain similarity also reflects how difficult to adapt from source domain to target domain. As shown in Figure 4, tasks in natural and specialized groups have higher domain similarity to source domain where the pre-trained model is",
+        "bbox": [
+            75,
+            839,
+            470,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/e6ce50144cb8482b59edc659595afb130625e9a8932639916a34f16206715741.jpg",
+        "image_caption": [
+            "Figure 3. Performance of features extracted from different layers on Sun397 (easy) and Clevr-Dist (difficult)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            92,
+            696,
+            234
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/a2274185ec5d76b15ee835772fee332d9d4e0b99a6ae64cf014e11b7db3753ed.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            92,
+            885,
+            234
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "trained on, so we mark them as easy cases. And tasks in structured groups have lower domain similarity to source domain, thus they are labeled as difficult cases in Figure 1.",
+        "bbox": [
+            498,
+            297,
+            890,
+            342
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Informativeness of Extracted Features",
+        "text_level": 1,
+        "bbox": [
+            500,
+            353,
+            828,
+            368
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To quantify the informativeness of features extracted from different layers for downstream tasks, we use leave-one-out  $k$ -NN testing as the surrogate performance criterion:",
+        "bbox": [
+            498,
+            377,
+            890,
+            422
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\text {I n f o r m a t i v e n e s s} = \\frac {\\sum_ {i = 1} ^ {m} \\sum_ {x _ {j} \\in N (x _ {i})} \\mathbb {1} \\left\\{y _ {i} = y _ {j} \\right\\}}{k m}, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            517,
+            446,
+            890,
+            481
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $m$  is the number of samples,  $x$  is the extracted features of a sample,  $N(x_{i})$  represents  $k$  nearest neighbors of  $x_{i}$ . We simply set  $k = 1$ , which results in leave-one-out 1-NN testing. We choose  $k$ -NN, instead of linear classifier, for informativeness measurement because  $k$ -NN is a non-parametric classifier and thus avoids training additional parameters, making the measurement more robust.",
+        "bbox": [
+            498,
+            491,
+            893,
+            597
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In Figure 3, we plot the informativeness of features extracted from different layers on two tasks from different domains. For Sun397, an easy domain adaption task, we observe the best performance achieved by features extracted from deep layers. This explains the good performance of linear probing for easy domain adaption tasks, as observed in Figure 1. While for Clevr-Dist, a difficult task with target domain significantly different from the source domain, we observe that no layer has outstanding performance. The results echo the findings in [16].",
+        "bbox": [
+            496,
+            598,
+            893,
+            750
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, we conduct experiments on gradually incorporating features extracted from different layers of the pre-trained model, demonstrating the different behaviors on easy and difficult tasks, which can be found in Section 5.4. The results show for easy tasks, more features may lead to overfitting, but for difficult tasks, performance improves with more features. Therefore, when the target domain is very different from the source domain, we need to provide probing model with diverse features extracted from multiple layers of the pre-trained model.",
+        "bbox": [
+            496,
+            750,
+            893,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "16851",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/3939f6c2b6b291cfdf9500bca9d202299da8bc6a024f8e3cf50e6af13a5703c3.jpg",
+        "image_caption": [
+            "Figure 4. Performance of extracted features with linear/ non-linear transformation on VTAB-1k. Datasets are ordered from easy tasks (left) to difficult tasks (right) according to domain similarity."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            90,
+            885,
+            272
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. Effects of Non-linear Transformation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            330,
+            398,
+            345
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Since there is no significantly informative layer for difficult domain adaption tasks, a linear combination may not be enough for exploiting the information in extracted features. We further examine the efficacy of non-linear combination on extracted features, as neural networks achieve superior performance through non-linear feature transformation.",
+        "bbox": [
+            75,
+            353,
+            470,
+            444
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We employ an multiple-layer perceptron (MLP) for nonlinear transformation and compare its performance with linear transformation. Features extracted from different structures of the pre-trained model are concatenated together as input (See Section 4.1 for details).",
+        "bbox": [
+            75,
+            445,
+            470,
+            518
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In Figure 4, we clearly observe that adding non-linearity to feature transformation can significantly boost the performance for the difficult domain adaption tasks (i.e. the datasets on right end of Figure 4). This phenomenon indicates that simply performing linear transformation on extracted features can not exploit the full potential of representations extracted from the pre-trained model. On the other hand, we notice that the performance on easy tasks is not improved, and is even hurt with the introduction of non-linearity. We speculate that this is because the linear transformation can already make full use of extracted information, while non-linearity increases model capacity, which eventually leads to overfitting to the training data.",
+        "bbox": [
+            75,
+            520,
+            470,
+            717
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Discussion. Our investigation reveals that the optimal probing model differs depending on the domain of downstream task. Incorporating diverse features from multiple layers of the pre-trained model and performing non-linear transformation on them can significantly improve the performance on difficult cases. However, this yields worse results on easy tasks. On the contrary, simply performing linear combination on features from deep layers can achieve satisfactory performance for easy adaptation tasks, but not for difficult ones. The conflicting behaviors of model probing motivates us to develop a unified framework that can solve the adaptation problems for both easy and difficult adaptation tasks.",
+        "bbox": [
+            75,
+            719,
+            470,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. Structured Model Probing",
+        "text_level": 1,
+        "bbox": [
+            498,
+            329,
+            746,
+            347
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this section, we propose structured model probing, a unified probing method which can deal with both easy and difficult cases of adaptation tasks. We first present how to thoroughly extract features in a structured way, then introduce our probing model with structured regularization.",
+        "bbox": [
+            496,
+            357,
+            893,
+            434
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1. Structured Feature Extraction",
+        "text_level": 1,
+        "bbox": [
+            498,
+            449,
+            769,
+            465
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Incorporating diverse features extracted from the pre-trained model can improve the performance on difficult adaptation tasks for model probing, as shown in Section 3.2. To provide diverse features for the probing model, we extract features from all structures in a Vision Transformer (ViT). These structures include features before self-attention, features after self-attention, and features after Feed Forward Network (FFN), for each transformer layer. We also include tokenized image inputs and pre-logits. We use  $V = \\{v_{i}\\}_{i}^{|V|}$  represents all structures in ViT, and  $|V|$  is the total number of structures. For an input sample, we denote each set of tokens extracted from structure  $v_{i}$  as  $X_{i} \\in \\mathbb{R}^{t \\times c}$ , where  $t$  is the dimension of tokens and  $c$  is the dimension of channels.",
+        "bbox": [
+            496,
+            474,
+            893,
+            672
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To reduce feature redundancy while preserving diversity, we apply 1D average pooling along different dimensions on tokens  $X_{i}$ , resulting in two types of aggregation. To maintain channel information, we aggregate  $X_{i}$  along the token dimension, which produces  $x_{i}^{t}$ . And to preserve spatial information, we aggregate  $X_{i}$  along the channel dimension, resulting in  $x_{i}^{c}$ . Finally, we concatenate all structured features into a single vector,  $x = [x_{1}^{t}, x_{1}^{c}, \\dots, x_{|V|}^{t}, x_{|V|}^{c}]$ . We denote  $s_{i}$  as the index set of elements in  $x$ , and  $S = \\{s_{i} | i = 1\\dots |s|\\}$ .",
+        "bbox": [
+            496,
+            675,
+            895,
+            816
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2. Probing with Structured Regularization",
+        "text_level": 1,
+        "bbox": [
+            498,
+            829,
+            843,
+            845
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our structured regularization contains two components: structured sparsity regularizer and structured non-linearity regularizer. We introduce them as follows.",
+        "bbox": [
+            496,
+            854,
+            893,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "16852",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2.1 Structured Feature Selection",
+        "text_level": 1,
+        "bbox": [
+            76,
+            90,
+            331,
+            104
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As shown in Section 3.2, the informativeness of extracted features varies with tasks, and redundant features lead to performance degradation on some tasks. Thus, we propose to perform feature selection during the training stage of the probing model to prevent overfitting. A popular way is to use group lasso [16, 43] by constructing groups on feature dimension over the weight matrix of a linear classifier:",
+        "bbox": [
+            75,
+            114,
+            468,
+            220
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\min  _ {\\theta} L (\\theta) + \\lambda \\sum_ {i = 1} ^ {d} | | \\theta_ {i} | | _ {2}, \\text {w h e r e} | | \\theta_ {i} | | _ {2} = \\sqrt {\\sum_ {j = 1} ^ {n} \\theta_ {i j} ^ {2}}. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            89,
+            227,
+            468,
+            275
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use  $\\theta \\in \\mathbb{R}^{d\\times n}$  to represent weight matrix, where  $d$  is the feature dimension and  $n$  is the number of classes. However, due to the high dimensionality of features, the final achieved sparsity is hardly the optimal result. It is worth noting that the structure of pre-trained model provides a natural way to divide features as groups, and features extracted from the same structure have higher relevance to each other. Thus, we leverage the structure of the pre-trained model, to achieve the structured sparsity with a structured sparsity regularizer:",
+        "bbox": [
+            75,
+            301,
+            472,
+            439
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\min  _ {\\theta} L (\\theta) + \\lambda \\sum_ {s \\in S} \\frac {1}{| s |} \\| \\theta_ {s} \\| _ {2}, \\text {w h e r e} \\| \\theta_ {s} \\| _ {2} = \\sqrt {\\sum_ {i \\in s} \\sum_ {j = 1} ^ {n} \\theta_ {i j} ^ {2}}, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            78,
+            458,
+            473,
+            518
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "$s \\in S$  is the index set of a specific structure in feature as described in Section 4.1, and  $|s|$  represents its dimension. Compared with constructing groups by features, in our regularizer, we construct groups by leveraging the structure of the pre-trained model. This formalization can significantly reduce the combinations of selection and thus prevent overfitting. With such an advantage, our method achieves superior performance with a small number of samples under high-dimensional features.",
+        "bbox": [
+            75,
+            518,
+            470,
+            654
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2.2 Structured Non-linear Transformation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            672,
+            400,
+            686
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our discovery in Section 3.3 shows non-linear transformation can facilitate transfer performance for difficult tasks, while could reduce performance for easy adaptation tasks, mostly due to over-fitting. This means that we need a model that allows flexibly incorporating non-linearity to a linear model. We propose a structured non-linearity regularizer that nicely integrates feature selection (i.e. choosing the subset of informative features from the pre-trained model) with model selection (i.e. choosing either a linear model or nonlinear model for prediction) inspired by hierarchical sparse modeling [28, 42]. Specifically, we construct our model by the combination of a linear model and a neural network:",
+        "bbox": [
+            75,
+            696,
+            472,
+            876
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf (x) = \\theta^ {\\top} x + f _ {W} (x), \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            194,
+            883,
+            468,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\theta \\in \\mathbb{R}^{d\\times n}$  is a linear model, and  $f_{W}$  represents an MLP with model weights  $W$ . We use  $W^{(1)}$  to represent the input layer of the MLP, and the parameters of rest layers are denoted as  $W^{(2:)}$ .",
+        "bbox": [
+            496,
+            90,
+            890,
+            150
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We utilize the feature selection result  $||\\theta ||_2$ , the  $\\ell_2$  norm of weights assigned to the selected features, to regularize the complexity of the non-linear model:",
+        "bbox": [
+            496,
+            152,
+            890,
+            196
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- When a small number of features are good enough for prediction, we expect a small value for  $||\\theta||_2$ . By using  $||\\theta||_2$  to control the complexity of non-linear model, we expect the resulting model to be mostly linear.",
+            "- With increasing  $||\\theta||_2$ , more diverse structures from the pre-trained model are needed for prediction. By using  $||\\theta||_2$  to regularize the size of non-linear model, we expect the resulting model to be more non-linear."
+        ],
+        "bbox": [
+            500,
+            198,
+            890,
+            318
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "A way to achieve such control is to add a constraint to the objective function in Eq.(4) as follows:",
+        "bbox": [
+            500,
+            319,
+            890,
+            349
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\min  _ {\\theta , W} L (\\theta , W) + \\lambda \\sum_ {s \\in S} | | \\theta_ {s} | | _ {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            612,
+            361,
+            803,
+            392
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\text {s u b j e c t} \\left. | | W _ {s} ^ {(1)} | | _ {2} \\leq M _ {1} | | \\theta_ {s} | | _ {2}, s = 1 \\dots | S |, \\right. \\tag {6} \\\\ \\left| \\left| W ^ {(2:)} \\right| \\right| _ {2} \\leq M _ {2} \\left| \\left| \\theta \\right| \\right| _ {2}. \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            544,
+            392,
+            890,
+            436
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "There are two constraints in Eq.(6): the first constraint is used to maintain the input feature sparsity of the neural network, and the second term is used to control the non-linearity through the linear model. To achieve efficient implementation, we convert the constraint in Eq.(6) into a regularization term, which we call structured non-linearity regularizer:",
+        "bbox": [
+            496,
+            448,
+            893,
+            539
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\Omega (\\theta , W) = \\sum_ {s \\in S} \\max  \\left\\{\\left| \\left| W _ {s} ^ {(1)} \\right| \\right| _ {2} - M _ {1} \\left| \\left| \\theta_ {s} \\right| \\right| _ {2}, 0 \\right\\} \\tag {7} \\\\ + \\max  \\left\\{\\left| \\left| W ^ {(2:)} \\right| \\right| _ {2} - M _ {2} \\left| \\left| \\theta \\right| \\right| _ {2}, 0 \\right\\}. \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            526,
+            550,
+            890,
+            604
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This regularization term means that, when  $||W||_2 > M||\\theta||_2$ ,  $W$  receives penalty until it achieves similar complexity as  $M||\\theta||$ . Then our final objective function is:",
+        "bbox": [
+            500,
+            616,
+            890,
+            662
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\min  _ {\\theta , W} L (\\theta , W) + \\lambda_ {1} \\sum_ {s \\in S} | | \\theta_ {s} | | _ {2} + \\lambda_ {2} \\Omega (\\theta , W). \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            537,
+            675,
+            890,
+            705
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use  $\\lambda$  to control the penalty strength, and simply set  $\\lambda_{2} = 0.1$  works well in our experiments.",
+        "bbox": [
+            500,
+            719,
+            890,
+            750
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Loss function. Since our model  $f(x)$  consists of both linear and non-linear parts, we decouple these two components in loss function. This modification enables the structured sparsity regularizer to control the complexity of  $\\| \\theta \\|$  without the interference from the non-linear part:",
+        "bbox": [
+            496,
+            752,
+            890,
+            828
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL (\\theta , W) = \\operatorname {C E} \\left(\\theta^ {\\top} x, y\\right) + \\operatorname {C E} \\left(\\operatorname {s g} \\left(\\theta^ {\\top} x\\right) + f _ {W} (x), y\\right), \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            508,
+            839,
+            890,
+            859
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $y$  is the one-hot label vector of  $x$ , CE is cross-entropy loss and sg means stop-gradient.",
+        "bbox": [
+            496,
+            869,
+            890,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "16853",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/573bcd4ee05f2e73a3fd2ac396c5f700fde6e2bcc0248c0abd0cfb3fb1862da7.jpg",
+        "table_caption": [
+            "Table 1. Median test accuracy (%) over 3 seeds on the VTAB-1k benchmark using ViT-B/16 pretrained on ImageNet-21k. * indicates results are obtained from [23] and † indicates results are obtained from [16]."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"8\">Natural</td><td colspan=\"6\">Specialized</td><td colspan=\"7\">Structured</td><td></td><td></td></tr><tr><td>CTFAR-100</td><td>Caltech101</td><td>DTD</td><td>Flowers102</td><td>Pets</td><td>SVHN</td><td>Sun397</td><td>Mean (Natural)</td><td>Camelyon</td><td>EuroSAT</td><td>Resisc45</td><td>Retinopathy</td><td>Mean (Specialized)</td><td>Clevr-Count</td><td>Clevr-Dist</td><td>DMLab</td><td>KITTI-Dist</td><td>dSpr-Loc</td><td>dSpr-Ori</td><td>sNORB-Azim</td><td>sNORB-Elev</td><td>Mean (Structured)</td><td>Mean (All)</td></tr><tr><td colspan=\"22\">Tuning Methods</td><td></td><td></td></tr><tr><td>Scratch†</td><td>7.6</td><td>19.1</td><td>13.1</td><td>29.6</td><td>6.7</td><td>19.4</td><td>2.3</td><td>14.0</td><td>71.0</td><td>71.0</td><td>29.3</td><td>72.0</td><td>60.8</td><td>31.6</td><td>52.5</td><td>27.2</td><td>39.1</td><td>66.1</td><td>29.7</td><td>11.7</td><td>24.1</td><td>35.3</td><td>32.8</td></tr><tr><td>Fine-tuning*</td><td>68.9</td><td>87.7</td><td>64.3</td><td>97.2</td><td>86.9</td><td>87.4</td><td>38.8</td><td>75.9</td><td>79.7</td><td>95.7</td><td>84.2</td><td>73.9</td><td>83.7</td><td>56.3</td><td>58.6</td><td>41.7</td><td>65.5</td><td>57.5</td><td>46.7</td><td>25.7</td><td>29.1</td><td>47.6</td><td>65.6</td></tr><tr><td>Bias*</td><td>72.8</td><td>87.0</td><td>59.2</td><td>97.5</td><td>85.3</td><td>59.9</td><td>51.4</td><td>73.3</td><td>78.7</td><td>91.6</td><td>72.9</td><td>69.8</td><td>78.3</td><td>61.5</td><td>55.6</td><td>32.4</td><td>55.9</td><td>66.6</td><td>40.0</td><td>15.7</td><td>25.1</td><td>44.1</td><td>62.0</td></tr><tr><td>VPT*</td><td>78.8</td><td>90.8</td><td>65.8</td><td>98.0</td><td>88.3</td><td>78.1</td><td>49.6</td><td>78.5</td><td>81.8</td><td>96.1</td><td>83.4</td><td>68.4</td><td>82.4</td><td>68.5</td><td>60.0</td><td>46.5</td><td>72.8</td><td>73.6</td><td>47.9</td><td>32.9</td><td>37.8</td><td>55.0</td><td>69.4</td></tr><tr><td colspan=\"22\">Probing Methods</td><td></td><td></td></tr><tr><td>Linear</td><td>77.2</td><td>86.8</td><td>66.6</td><td>99.1</td><td>88.8</td><td>35.7</td><td>55.1</td><td>72.7</td><td>78.4</td><td>89.6</td><td>72.2</td><td>73.7</td><td>78.5</td><td>31.0</td><td>29.9</td><td>35.2</td><td>54.9</td><td>11.7</td><td>27.7</td><td>14.4</td><td>23.7</td><td>28.6</td><td>55.4</td></tr><tr><td>Side-Tuning*</td><td>60.7</td><td>60.8</td><td>53.6</td><td>95.5</td><td>66.7</td><td>34.9</td><td>35.3</td><td>58.2</td><td>58.5</td><td>87.7</td><td>65.2</td><td>61.0</td><td>68.1</td><td>27.6</td><td>22.6</td><td>31.3</td><td>51.7</td><td>8.2</td><td>14.4</td><td>9.8</td><td>21.8</td><td>23.4</td><td>45.6</td></tr><tr><td>Head2Toe</td><td>75.3</td><td>90.9</td><td>75.0</td><td>99.5</td><td>86.1</td><td>83.5</td><td>50.9</td><td>80.2</td><td>84.4</td><td>95.7</td><td>84.4</td><td>74.3</td><td>84.7</td><td>51.7</td><td>59.4</td><td>44.0</td><td>65.6</td><td>47.1</td><td>40.3</td><td>32.5</td><td>41.1</td><td>47.7</td><td>67.5</td></tr><tr><td>LinearStruc.</td><td>72.9</td><td>89.5</td><td>72.5</td><td>98.9</td><td>83.6</td><td>71.6</td><td>42.7</td><td>76.0</td><td>84.3</td><td>96.0</td><td>81.8</td><td>75.0</td><td>84.3</td><td>48.2</td><td>52.5</td><td>39.4</td><td>66.7</td><td>48.7</td><td>36.7</td><td>32.3</td><td>37.7</td><td>45.3</td><td>64.8</td></tr><tr><td>LinearStruc.w/FSR</td><td>78.2</td><td>90.1</td><td>73.4</td><td>99.3</td><td>88.1</td><td>71.7</td><td>45.4</td><td>78.0</td><td>83.9</td><td>96.1</td><td>81.9</td><td>74.1</td><td>84.0</td><td>48.5</td><td>52.8</td><td>39.6</td><td>64.4</td><td>51.4</td><td>38.6</td><td>32.9</td><td>37.0</td><td>45.7</td><td>65.7</td></tr><tr><td>LinearStruc.w/SSR</td><td>79.0</td><td>90.2</td><td>74.1</td><td>99.1</td><td>90.4</td><td>73.4</td><td>54.2</td><td>80.1</td><td>84.6</td><td>96.3</td><td>83.8</td><td>74.4</td><td>84.8</td><td>48.8</td><td>52.9</td><td>39.8</td><td>66.9</td><td>48.7</td><td>38.9</td><td>32.5</td><td>37.9</td><td>45.8</td><td>66.6</td></tr><tr><td>MLPStruc.</td><td>66.8</td><td>88.7</td><td>70.3</td><td>97.4</td><td>82.1</td><td>71.1</td><td>33.3</td><td>72.8</td><td>81.6</td><td>95.9</td><td>77.0</td><td>73.5</td><td>82.0</td><td>77.2</td><td>57.7</td><td>39.3</td><td>66.5</td><td>72.3</td><td>44.1</td><td>32.7</td><td>47.7</td><td>54.7</td><td>67.1</td></tr><tr><td>SMP w/o SNR</td><td>71.3</td><td>89.6</td><td>72.3</td><td>99.0</td><td>85.6</td><td>72.7</td><td>45.0</td><td>76.5</td><td>84.2</td><td>95.9</td><td>81.9</td><td>72.8</td><td>83.7</td><td>77.4</td><td>57.1</td><td>40.8</td><td>67.1</td><td>72.4</td><td>44.2</td><td>32.6</td><td>48.4</td><td>55.0</td><td>69.0</td></tr><tr><td>SMP</td><td>79.3</td><td>90.9</td><td>74.9</td><td>99.3</td><td>90.4</td><td>75.0</td><td>55.3</td><td>80.7</td><td>84.8</td><td>96.3</td><td>83.1</td><td>75.0</td><td>84.8</td><td>77.5</td><td>58.0</td><td>40.8</td><td>67.5</td><td>72.5</td><td>44.5</td><td>33.0</td><td>49.0</td><td>55.4</td><td>70.9</td></tr></table>",
+        "bbox": [
+            76,
+            116,
+            893,
+            431
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            441,
+            209,
+            458
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we evaluate the proposed structured model probing method on multiple datasets across various tasks, and provide detailed analysis of the proposed method.",
+        "bbox": [
+            76,
+            465,
+            470,
+            513
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation details. We evaluate SMP on the popular vision backbone, ViT-B/16 [15] pre-trained on ImageNet-21k [12], which is widely used across relative research works [16, 23]. We provide details of probing model and hyperparameter selection in the supplementary material.",
+        "bbox": [
+            76,
+            513,
+            470,
+            590
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Compared methods. We compare SMP with other commonly used and state-of-the-art tuning (Fine-tuning, Bias [6] and VPT [23]) and probing methods (Linear Probing, Side-Tuning [45] and Head2Toe [16]). All hyperparameters of baseline methods are carefully tuned, and further information is provided in supplementary material.",
+        "bbox": [
+            76,
+            593,
+            470,
+            685
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Experiments on VTAB-1k",
+        "text_level": 1,
+        "bbox": [
+            76,
+            693,
+            313,
+            709
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Experimental results on the VTAB-1k benchmark (Table 1) indicate that SMP outperforms other methods, achieving a mean performance of  $70.9\\%$ ,  $3.4\\%$  higher than Head2Toe and  $1.5\\%$  higher than VPT. SMP also significantly outperforms Linear Probing, Bias, and Side-Tuning. SMP achieves optimal results in all three dataset groups. Especially in Structured group, SMP attains a mean accuracy of  $55.4\\%$ , surpassing other probing methods by a significant margin. This demonstrates the incorporation of non-linear transformation is important for difficult adaptation tasks.",
+        "bbox": [
+            76,
+            715,
+            470,
+            867
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Ablation studies. We provide results of baseline methods that contribute to SMP, to validate the efficacy of structured",
+        "bbox": [
+            76,
+            869,
+            470,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "regularization and the contribution of each component.",
+        "bbox": [
+            500,
+            443,
+            861,
+            458
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "LinearStruc. represents extracting features by structured feature extraction proposed in Section 4.1, then training a linear classifier. Compare to Linear Probing, it gets  $9\\%$  average performance gain, indicating that solely extracting features from the last layer is inadequate for model probing.",
+        "bbox": [
+            496,
+            458,
+            893,
+            534
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We further compare feature sparsity regularization in Eq.(3) (represents by  $\\mathrm{Linear}_{\\mathrm{Struc.}} + \\mathrm{FSR}$ ) to the proposed structured sparsity regularizer (represents by  $\\mathrm{Linear}_{\\mathrm{Struc.}} + \\mathrm{SSR}$ ). Structured sparsity regularizer outperforms feature sparsity regularization, indicating that feature selection can mitigate the overfitting caused by high dimensions, and leveraging model structure as a prior can bring extra benefits.",
+        "bbox": [
+            496,
+            534,
+            893,
+            640
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We apply non-linear transformation on extracted features, represents by  $\\mathrm{MLP}_{\\text{Struc.}}$ . We find the performance on structured group gets a significant improvement over  $\\mathrm{Linear}_{\\text{Struc.}}$ , demonstrating the importance of non-linear transformation on difficult adaptation tasks. However, the performance on easy tasks, i.e., natural and specialized groups, is degraded, suggesting the risk of overfitting from non-linear transformation. We solve this problem by leveraging structured nonlinearity regularizer (SNR), which can automatically control the non-linearity through the norms of selected structures.",
+        "bbox": [
+            496,
+            640,
+            893,
+            791
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2. Experiments on Few-shot Learning",
+        "text_level": 1,
+        "bbox": [
+            500,
+            801,
+            807,
+            816
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we evaluate our method in few-shot learning scenario, where the number of training samples varies from 1 to 16 shots per class following existing studies. Five fine-grained recognition tasks are chosen, i.e., Food101 [4], CUB-200-2011 [39], Stanford Cars [26], Stanford Dogs [24]",
+        "bbox": [
+            496,
+            825,
+            893,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "16854",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/817f5dd48fe878aaf1b3177c862d35bfe2ed6a9e9e8e3980f397eaecfecb0609.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            80,
+            349,
+            242
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e2fe9524fb302c804246dd5f00e0babe0a03ccc14a1e9abb24d528eab516d278.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            351,
+            80,
+            617,
+            242
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/484d8aefe8aaa968b7c51fd933bef7233ff1cddb14dede596e27d1856b3004b2.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            619,
+            80,
+            883,
+            242
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/1dcfc2b477ee64adeaf90836b3949d8250e549672b924d614cbda5c630efdc49.jpg",
+        "image_caption": [
+            "Figure 5. Performance of few-shot learning on fine-grained visual recognition tasks. LinearStr. (F) indicates LinearStruc. w/ FSR."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            83,
+            244,
+            349,
+            406
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/47064e96137db55553faad763127478d043cbcae2ca099511d7e6dc3d2b89899.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            351,
+            244,
+            617,
+            406
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e7681287e02199c4d9a68a98c6b192a4e43bb4f3e93194a7da447464654dafba.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            619,
+            244,
+            883,
+            406
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "and Oxford Pets [35]. We compared SMP with two strong baselines, LinearStruc. and LinearStruc. w/ FSR, and two prominent methods, VPT (tuning) and Head2Toe (probing).",
+        "bbox": [
+            75,
+            450,
+            470,
+            496
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Experimental results presented in Figure 5 show that SMP consistently outperforms other methods in most cases, demonstrating its effectiveness in few-shot scenario, and its ability to improve performance with increasing training data. Notably, feature sparsity regularization performs worse than SMP, highlighting the importance of leveraging pre-trained model structure as prior knowledge in low-shot scenario.",
+        "bbox": [
+            75,
+            498,
+            468,
+            604
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3. Experiments on More Transfer Scenarios",
+        "text_level": 1,
+        "bbox": [
+            76,
+            619,
+            429,
+            636
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Larger downstream datasets. We provide average results on 9 full-size downstream datasets in Table 2. It can be observed even in the larger data regime, SMP still achieves competitive performance compared with tuning methods. Full results are presented in Suppl. B.1.",
+        "bbox": [
+            75,
+            648,
+            468,
+            724
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Different pre-trained models. We present VTAB-1k results on ViT-B/16 and ViT-L/14 pretrained by CLIP in Suppl. B.2. Despite the usage of stronger pre-trained models, SMP continuously outperforms baseline methods due to the incorporation of diverse features and non-linear transformation, as well as its flexible framework.",
+        "bbox": [
+            75,
+            729,
+            470,
+            819
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Different architectures. SMP is a versatile method that can be applied to convolutional neural networks as well. In Suppl. B.3, we present the VTAB-1k results obtained from ImageNet pre-trained ResNet-50 models, where SMP outperforms other methods and achieves superior performance.",
+        "bbox": [
+            75,
+            824,
+            472,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/b028cc30c23ebe2fa7447b605a5e3411ab10a9d73667bb5385e9281e3bc16f10.jpg",
+        "table_caption": [
+            "Table 2. Average accuracy (%) on 9 full-size downstream datasets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Fine-tuning</td><td>VPT</td><td>Linear</td><td>SMP</td></tr><tr><td>Average Accuracy</td><td>88.4</td><td>88.7</td><td>81.7</td><td>88.7</td></tr></table>",
+        "bbox": [
+            500,
+            467,
+            893,
+            510
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.4. Analysis and Discussion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            521,
+            718,
+            537
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this section, we conduct various experiments to analyze the components of the proposed SMP method. Full results of this section are presented in Suppl. C.",
+        "bbox": [
+            496,
+            546,
+            890,
+            590
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Visualization of group norms. We visualize the  $\\ell_2$  norm of each structure on different tasks in Figure 6, which helps us to understand the behavior of the proposed structured regularization. For Sun397, an easy adaptation task, we can observe the structures from deep layers have larger norms in Figure 6a. This validates our analysis in Section 3.2, that for easy adaptation tasks, features extracted from deep layers contain more information. For Clevr-Dist, a difficult adaptation task, the structures have similar norms overall, which means for difficult tasks, information contained in intermediate layers is also important, and diverse features can improve the performance, as we find in Section 3.2. This motivates the proposed structured non-linearity regularization, utilizing  $\\ell_2$  norm to control the complexity of non-linear model.",
+        "bbox": [
+            496,
+            595,
+            892,
+            805
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Retraining probing model by gradually incorporating structures. We retrain a probing model by gradually incorporating structures selected based on the magnitude of their  $\\ell_2$  norm, to validate their contribution to the prediction. Results are presented in Figure 7. For Sun397, an easy adaptation task, we can see in Figure 7a that there are redun",
+        "bbox": [
+            496,
+            810,
+            893,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "16855",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e99f4b82417c31d3c5087dcde2007334e8fda3dd38a873613a6132ae66d91ab8.jpg",
+        "image_caption": [
+            "(a) Sun397 (easy)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            84,
+            84,
+            477,
+            196
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/1fb3d64faad6bf3cad0400800a95c05e9b0fe2efb2efe6b342be0c3a1ad2b55b.jpg",
+        "image_caption": [
+            "(b) Clevr-Dist (difficult)",
+            "Figure 6.  $\\ell_2$  norm of structure groups on different tasks.",
+            "Figure 7. Retraining performance by gradually incorporating structures with larger  $\\ell_{2}$  norm obtained by SMP."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            84,
+            218,
+            475,
+            330
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "dant structures. And with a few selected structures, we can achieve comparable results compared with the original SMP, justifying the zero norm of some structured observed after performing structure regularization in Figure 6a. However, for Clevr-Dist, a difficult adaptation task, comparable performance is achieved until selecting 55 structures, suggesting intermediate structures also contribute to the final prediction. This demonstrates the necessity of proposed structured regularizer, which resolves the contradiction between easy and difficult domain adaptation tasks.",
+        "bbox": [
+            75,
+            391,
+            470,
+            542
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Impact of feature aggregation designs. The primary objective of performing token and channel aggregation is to minimize the redundancy in the extracted raw features, while maintaining feature diversity in the compact aggregated representation. To achieve this, two types of aggregation are designed to preserve channel and spatial information, respectively. We provide ablation studies in Suppl. C.5. Overall, channel information (aggregating along tokens) is essential for all datasets, while spatial information (aggregating along channels) is beneficial for tasks that are sensitive to spatial location, e.g., dSpr-Loc and sNORB-Azim.",
+        "bbox": [
+            75,
+            545,
+            470,
+            712
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Impact of different structures. To increase the feature diversity, we incorporate features extracted from various structures, coupled with a structured sparse regularizer to conduct model fitting and feature selection simultaneously. This flexible framework enables us to avoid the costly manual selection of candidate structures. We provide analysis of structures in Suppl. C.6. The results show that incorporating all candidate structured features yields best mean results, which demonstrates the efficacy of our method.",
+        "bbox": [
+            75,
+            715,
+            470,
+            851
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Cost of SMP. We compare the cost of SMP with baseline methods, including relative adaptation time, extra inference FLOPs, and extra parameters. The average results on VTAB-",
+        "bbox": [
+            75,
+            854,
+            470,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/c4addc513272d803bc245c903c01bcb20ac3f895ce7323006e5714743781c417.jpg",
+        "image_caption": [
+            "(a) Sun397 (easy)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            85,
+            694,
+            213
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/67152e2b9413ff984d8288c4e3f256e6f520b098e0b19564a1c7c9735b629693.jpg",
+        "image_caption": [
+            "(b) Clevr-Dist (difficult)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            700,
+            85,
+            887,
+            213
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/e757e527962b43def038b34ac3b8d4cde7b8ad9fd8cb10b9f250d99508f9085e.jpg",
+        "table_caption": [
+            "Table 3. Cost comparison (Average results on VTAB-1k). * indicates relative results compared with Fine-tuning."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Linear</td><td>FT</td><td>VPT</td><td>SMP</td></tr><tr><td>Adaptation time (rel.)*</td><td>0.05</td><td>1</td><td>0.97</td><td>0.15</td></tr><tr><td>Extra Inference FLOPs (G)</td><td>0</td><td>0</td><td>11.01</td><td>0.01</td></tr><tr><td>Extra parameters (M)</td><td>0.04</td><td>85.83</td><td>0.64</td><td>6.91</td></tr></table>",
+        "bbox": [
+            501,
+            316,
+            888,
+            393
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "1k are shown in Table 3. SMP only needs  $15\\%$  adaptation time on average compared with fine-tuning. On the contrary, the adaptation time of VPT is  $97\\%$  of fine-tuning. This results in a significantly lower adaptation cost of SMP compared to tuning methods. SMP is also very efficient during inference time, and only introduces 0.01G inference FLOPs compared with fine-tuning. It can be noticed that VPT introduces extra 11.01G inference FLOPs, which is far more than SMP. SMP requires 6.91M extra parameters on average, which is the  $8\\%$  of fine-tuning. Though the requirement of extra parameters is higher than VPT, we can leverage the retraining method proposed in Section 5.4, which can further reduce the extra parameters. Overall, SMP is a highly efficient method with low adaptation cost, fast inference, and fewer extra parameters, while maintaining superior performance compared to tuning-based methods.",
+        "bbox": [
+            496,
+            405,
+            893,
+            647
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            662,
+            617,
+            679
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we present structured model probing, an effective yet efficient probing method for transfer learning. Our investigation reveals that model probing behaves differently for easy and difficult adaptation tasks. To mitigate this conflicting behavior, we propose structured model probing, a method that is able to achieve good performance in both cases. With the proposed structured regularization, we can construct a simple adaptation model for easy adaptation cases, and increase the complexity of the adaptation model for difficult cases. Our method outperforms state-of-the-art tuning methods while maintaining computational efficiency. Our work highlights the potential of model probing, which motivates researchers to better exploit the use large-scale pre-trained model and apply model probing to diverse tasks.",
+        "bbox": [
+            496,
+            688,
+            893,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "16856",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            174,
+            106
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Alessandro Achille, Michael Lam, Rahul Tewari, Avinash Ravichandran, Subhransu Maji, Charless C. Fowlkes, Stefano Soatto, and Pietro Perona. Task2vec: Task embedding for meta-learning. In IEEE International Conference on Computer Vision (ICCV), pages 6429-6438, 2019. 3",
+            "[2] Alessandro Achille, Aditya Golatkar, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. LQF: linear quadratic fine-tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15729-15739, 2021. 3",
+            "[3] Armen Aghajanyan, Akshit Shrivastava, Anchit Gupta, Naman Goyal, Luke Zettlemoyer, and Sonal Gupta. Better finetuning by reducing representational collapse. In 9th International Conference on Learning Representations (ICLR), 2021. 2",
+            "[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446-461, 2014. 6",
+            "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020. 1, 2",
+            "[6] Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. Tinytl: Reduce memory, not parameters for efficient on-device learning. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020. 2, 6",
+            "[7] Shoufa Chen, Chongjian GE, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In Advances in Neural Information Processing Systems 35 (NeurIPS), 2022. 2",
+            "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 1597-1607, 2020. 2",
+            "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In IEEE International Conference on Computer Vision (ICCV), pages 9620-9629, 2021. 2",
+            "[10] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4109-4118, 2018. 3",
+            "[11] Mostafa Dehghani, Yi Tay, Anurag Arnab, Lucas Beyer, and Ashish Vaswani. The efficiency misnomer. In 10th International Conference on Learning Representations (ICLR), 2022. 2"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 3, 6",
+            "[13] Aditya Deshpande, Alessandro Achille, Avinash Ravichandran, Hao Li, Luca Zancato, Charless C. Fowlkes, Rahul Bhotika, Stefano Soatto, and Pietro Perona. A linearized framework and a new benchmark for model selection for fine-tuning. arXiv:2102.00084, 2021. 3",
+            "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 4171–4186, 2019. 2",
+            "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations (ICLR), 2021. 3, 6",
+            "[16] Utku Evci, Vincent Dumoulin, Hugo Larochelle, and Michael C. Mozer. Head2Toe: Utilizing intermediate representations for better transfer learning. In Proceedings of the 39th International Conference on Machine Learning (ICML), pages 6009-6033, 2022. 1, 3, 5, 6",
+            "[17] Weifeng Ge and Yizhou Yu. Borrowing treasures from the wealthy: Deep transfer learning through selective joint finetuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10-19, 2017. 2",
+            "[18] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogério Schmidt Feris. SpotTune: Transfer learning through adaptive fine-tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4805-4814, 2019. 2",
+            "[19] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In 10th International Conference on Learning Representations (ICLR), 2022. 1, 2",
+            "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2, 3",
+            "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2022. 2, 3",
+            "[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning (ICML), pages 2790-2799, 2019. 2",
+            "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser-Nam Lim."
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "16857",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6",
+            "[24] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition (CVPR Workshop), 2011. 6",
+            "[25] Simon Kornblith, Jonathon Shlens, and Quoc V. Le. Do better imagenet models transfer better? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2661-2671, 2019. 2, 3",
+            "[26] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshop (ICCV Workshop), pages 554–561, 2013. 6",
+            "[27] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In 10th International Conference on Learning Representations (ICLR), 2022. 2",
+            "[28] Ismael Lemhadri, Feng Ruan, Louis Abraham, and Robert Tibshirani. Lassonet: A neural network with feature sparsity. Journal of Machine Learning Research, 22:127:1-127:29, 2021. 5",
+            "[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3045-3059, 2021. 2",
+            "[30] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), pages 4582-4597, 2021. 2",
+            "[31] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5196-5205, 2022. 2",
+            "[32] Fangzhou Mu, Yingyu Liang, and Yin Li. Gradients as features for deep representation learning. In 8th International Conference on Learning Representations (ICLR), 2020. 3",
+            "[33] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. LEEP: A new measure to evaluate transferability of learned representations. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 7294-7305, 2020. 3",
+            "[34] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2010. 2",
+            "[35] Omkar M. Parkhi, Andreae Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3498-3505, 2012. 7",
+            "[36] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 2227-2237, 2018. 3"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[37] 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 Proceedings of the 38th International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1, 2",
+            "[38] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. LST: Ladder side-tuning for parameter and memory efficient transfer learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3",
+            "[39] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical report, California Institute of Technology, 2011. 6",
+            "[40] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In 10th International Conference on Learning Representations (ICLR), 2022. 2",
+            "[41] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer G. Dy, and Tomas Pfister. Learning to prompt for continual learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 2",
+            "[42] Xiaohan Yan and Jacob Bien. Hierarchical Sparse Modeling: A choice of two group lasso formulations. Statistical Science, 32(4):531 - 560. 5",
+            "[43] Ming Yuan and Yi Lin. Model selection and estimation in regression with grouped variables. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(1): 49-67, 2006. 5",
+            "[44] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv:1910.04867, 2019. 1, 2, 3",
+            "[45] Jeffrey O. Zhang, Alexander Sax, Amir Zamir, Leonidas J. Guibas, and Jitendra Malik. Side-tuning: A baseline for network adaptation via additive side networks. In European Conference on Computer Vision (ECCV), pages 698-714, 2020. 1, 2, 3, 6",
+            "[46] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16795-16804, 2022. 2",
+            "[47] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2",
+            "[48] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 2"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            867
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "16858",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/fcd62d92-907d-4f7f-bc80-b88db83abf7f_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1056be9bec23bb7ed3b5c99c890486b01ea67bbde403174f828d535d8654198
+size 351536

2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/full.md

@@ -0,0 +1,333 @@
+# Structured Model Probing: Empowering Efficient Transfer Learning by Structured Regularization
+
+Zhi-Fan Wu Chaojie Mao Xue Wang Jianwen Jiang Yiliang Lv Rong Jin Alibaba Group
+
+# Abstract
+
+Despite encouraging results from recent developments in transfer learning for adapting pre-trained model to downstream tasks, the performance of model probing is still lagging behind the state-of-the-art parameter efficient tuning methods. Our investigation reveals that existing model probing methods perform well for the easy case when the source domain (where models are pre-trained) and the adapted domain are similar, but fail for the difficult case when the two domains are significantly different. Simply incorporating features extracted from multiple layers and increasing complexity of the probing model can mitigate the gap in the difficult case, but degrades the performance in the easy case. To address this challenge, we propose structured model probing (SMP) that is able to deliver good performance for both cases through structured regularization. The regularization performs feature selection leveraging model structure as a prior, and controls the complexity of the probing model through the weights of selected structures. This enables us to construct a simple adaptation model, with a small number of selected features and a linear prediction model, for the easy case; and to automatically increase the complexity of adaptation model, with a large number of selected features and a non-linear model, for the difficult case. Our extensive empirical studies show that SMP significantly outperforms the state-of-the-art methods for parameter efficient tuning, and at the same time, still maintains the advantage of computational efficiency for probing-based methods.
+
+# 1. Introduction
+
+The parameters of models have grown drastically with the rapid development of deep learning [5, 37], posing challenges for adapting pre-trained models to downstream tasks. The widely used transfer strategy, fully fine-tuning, becomes infeasible due to excessive computation and storage overhead. Recently, parameter efficient tuning [19, 23] has gained attention for reducing storage costs and improving tuning performance with few trainable parameters when lim
+
+![](images/63f688131516d3ed72daf25e836154d3ffd2d13587e9edb41098c34830f2f3e3.jpg)  
+Figure 1. Performance comparison on different dataset groups (left) and mean relative adaptation time (right) on VTAB-1k benchmark.
+
+ited data are available for downstream tasks. Despite the progress, parameter efficient tuning methods still face significant computational overhead in training due to the need to perform forward and backward passes through the pretrained model at each training step.
+
+Model probing, which leverages features extracted from the pre-trained model to perform adaptation, is promising due to its efficiency and simplicity: The training procedure is efficient since the pre-trained model is fixed and only performs one forward propagation to extract features, and probing model is decoupled from pre-trained model that makes it convenient to deploy small sized task-specific downstream models. However, it is a nuisance that probing methods are usually outperformed by state-of-the-art parameter efficient tuning methods. This is true even after several recent improvements made for the probing methods [16, 45]. In this paper, we aim to improve the performance of model probing significantly without sacrificing its simplicity and efficiency.
+
+We first investigate when model probing methods perform worse than the tuning-based methods on widely used visual task adaptation benchmark VTAB-1k [44] as shown in Figure 1. Target domains are categorized as easy and difficult cases based on their similarity to the source domain where the pre-trained model is trained on (see Section 3.1 for details). We find simple linear probing yields comparable performance as the two tuning-based approaches (finetuning and VPT [23]) for easy cases, whereas a significant
+
+![](images/8976b25f7d62a3b031921846e07f4362b5639759867d578f4903a604d17738bc.jpg)  
+Figure 2. Adaptation pipelines of Model Tuning (left) and the proposed Structured Model Probing (right).
+
+![](images/7160d55c197b3be1ab48dc688ed30909c72b2fb4a9022aedc0f7297b00ee4323.jpg)
+
+performance gap is observed between linear probing and tuning-based methods for difficult cases. This suggests that model probing exhibits different behavior for various target domains, and performance gap arises for target domains that dissimilar to the source domain.
+
+This is not surprising that we expect a more complicated model than a linear probing to compensate the difference between the source domain and the target domain. In general, we can increase model complexity in two ways, either by introducing more features extracted from the pre-trained model, or by introducing a non-linear model for prediction. A main challenge is how to develop a unified framework of model probing that is able to handle both the easy and difficult adaptations effectively. This is because, according to our study in Section 3, for easy cases, a complicated probing model (with more features and non-linear probing) performs significantly worse than a simple linear probing.
+
+To this end, we propose structured model probing, or SMP for short, that effectively address both easy and difficult cases of domain adaptation. The key to our unified framework is structured regularization. It performs feature selection leveraging the structure of the pre-trained model as a prior, and controls the complexity of the probing model through the weights assigned to selected features. Thus, for the easy case, this regularization constructs a simple linear model with a few selected features; and for the difficult case, it automatically introduces more features from the pre-trained model and more non-linearity in the prediction module based on the value of loss function, as shown in Figure 2. Our empirical studies show that the proposed structured model probing performs significantly better than state-of-the-art tuning-based approaches, while maintaining the advantage of computational efficiency.
+
+Our main contributions are as follows:
+
+1. We observe and demonstrate the conflicting behaviors of model probing: a complicated probing model with more features and non-linear transformation can improve the performance for difficult cases of domain adaptation, but it yields worse performance than linear probing for easy cases of domain adaptation.
+
+2. We propose structured model probing (SMP) based on structured regularization. This regularization enables us to construct simple probing model for easy tasks, and automatically increase model complexity for difficult tasks, to simultaneously address both the easy and difficult cases of domain adaptation.  
+3. We conduct extensive empirical studies on various adaptation tasks. The proposed SMP method shows superior experimental results compared to state-of-the-art probing and tuning methods, and requires significantly lower training costs compared to tuning methods.
+
+# 2. Related Work
+
+Transfer learning has been extensively studied in vision domain [34, 48]. It aims to improve the performance on target domain by leveraging the related information contained in source domain. Transfer learning based on pre-trained model [8, 20, 21] has received significant attention due to its simplicity and good empirical performance [23, 45].
+
+Model tuning adapts the pre-trained model to new tasks via updating model parameters. Fully fine-tuning is a widely used strategies due to its efficacy and simplicity [9, 25, 44]. However, it suffers from huge storage and computation overhead due to the ever-expanding size of large-scale pre-trained model [37, 40]. Heuristic approaches have been proposed to improve fine-tuning performance [3, 6, 17, 18, 27]. Parameter efficient tuning [19] aims to reduce the storage requirement of transferred model by updating a few parameters attached to the pre-trained model [23, 47], like Prompt Tuning [29, 30] or Adapter [22]. While originally studied in natural language processing tasks due to the dominance of large-scale pre-trained language models [5, 14], there is growing interest in applying these methods to computer vision tasks [7, 23, 31, 41, 46, 47]. Though parameter efficient tuning reduces storage cost significantly, the training cost remains high due to the forward and backward propagation need to pass through the entire pre-trained model [11].
+
+Model probing utilizes frozen extracted features for adapting the pre-trained model to new tasks, thus achieving effi
+
+cient adaptation. The most popular method, linear probing, suffers from performance degradation [25], which limits its application in transfer learning [20, 21]. Recent methods [16, 36, 45] have been proposed to improve probing performance by leveraging side network [38, 45] or incorporating intermediate representations [2, 16, 32], but there is still a significant performance gap compared to tuning methods. This motivates us to investigate the limitation of probing methods, and come up with a better solution.
+
+# 3. Understanding the Limitation of Model Probing for Transfer Learning
+
+In this section, we analyze the performance of probing methods across various downstream domains to understand their limitations, motivated by the observation in Figure 1. We assess feature informativeness for different tasks and investigate the necessity of non-linear transformation.
+
+# 3.1. Analysis Setup
+
+Pre-trained model and source data. The ViT-B/16 [15] is used as the backbone because of the extensive usage of transformer-based models. The model is pre-trained on ImageNet-21k [12] in a supervised learning manner.
+
+Target data. We use the VTAB-1k benchmark, consisting of 19 datasets that cover natural images (natural), images captured by specialist equipment (specialized), and images generated from simulated environments (structured), as our target data for transfer learning. We follow the setting in [44] and use 1000 samples as the training set per task to evaluate adaptation with limited data as suggested in [44].
+
+Domain similarity. Several previous works have proposed metrics to measure the similarity between source domain and target domain [1, 10, 13, 16, 33]. Here, we adopt the domain similarity measure defined in [16]:
+
+$$
+D o m a i n \quad \text {S i m i l a r i t y} = \mathrm {A c c} _ {\text {L i n e a r}} - \mathrm {A c c} _ {\text {S c r a t c h}}. \tag {1}
+$$
+
+A higher domain similarity suggests that the features extracted from a pre-trained model can be utilized for the target domain, leading to performance gains via a linear classifier compared to training from scratch. Conversely, a lower domain similarity indicates that training from scratch on the target domain may be more effective than using extracted features. Thus, the domain similarity is evaluated based on the benefit gained by a linear classifier compared to training from scratch. We also report the label-feature correlation [13] in Suppl. C.8, which exhibits a high Spearman rank correlation (0.854) with the domain similarity scores.
+
+Domain similarity also reflects how difficult to adapt from source domain to target domain. As shown in Figure 4, tasks in natural and specialized groups have higher domain similarity to source domain where the pre-trained model is
+
+![](images/e6ce50144cb8482b59edc659595afb130625e9a8932639916a34f16206715741.jpg)  
+Figure 3. Performance of features extracted from different layers on Sun397 (easy) and Clevr-Dist (difficult).
+
+![](images/a2274185ec5d76b15ee835772fee332d9d4e0b99a6ae64cf014e11b7db3753ed.jpg)
+
+trained on, so we mark them as easy cases. And tasks in structured groups have lower domain similarity to source domain, thus they are labeled as difficult cases in Figure 1.
+
+# 3.2. Informativeness of Extracted Features
+
+To quantify the informativeness of features extracted from different layers for downstream tasks, we use leave-one-out  $k$ -NN testing as the surrogate performance criterion:
+
+$$
+\text {I n f o r m a t i v e n e s s} = \frac {\sum_ {i = 1} ^ {m} \sum_ {x _ {j} \in N (x _ {i})} \mathbb {1} \left\{y _ {i} = y _ {j} \right\}}{k m}, \tag {2}
+$$
+
+where  $m$  is the number of samples,  $x$  is the extracted features of a sample,  $N(x_{i})$  represents  $k$  nearest neighbors of  $x_{i}$ . We simply set  $k = 1$ , which results in leave-one-out 1-NN testing. We choose  $k$ -NN, instead of linear classifier, for informativeness measurement because  $k$ -NN is a non-parametric classifier and thus avoids training additional parameters, making the measurement more robust.
+
+In Figure 3, we plot the informativeness of features extracted from different layers on two tasks from different domains. For Sun397, an easy domain adaption task, we observe the best performance achieved by features extracted from deep layers. This explains the good performance of linear probing for easy domain adaption tasks, as observed in Figure 1. While for Clevr-Dist, a difficult task with target domain significantly different from the source domain, we observe that no layer has outstanding performance. The results echo the findings in [16].
+
+Furthermore, we conduct experiments on gradually incorporating features extracted from different layers of the pre-trained model, demonstrating the different behaviors on easy and difficult tasks, which can be found in Section 5.4. The results show for easy tasks, more features may lead to overfitting, but for difficult tasks, performance improves with more features. Therefore, when the target domain is very different from the source domain, we need to provide probing model with diverse features extracted from multiple layers of the pre-trained model.
+
+![](images/3939f6c2b6b291cfdf9500bca9d202299da8bc6a024f8e3cf50e6af13a5703c3.jpg)  
+Figure 4. Performance of extracted features with linear/ non-linear transformation on VTAB-1k. Datasets are ordered from easy tasks (left) to difficult tasks (right) according to domain similarity.
+
+# 3.3. Effects of Non-linear Transformation
+
+Since there is no significantly informative layer for difficult domain adaption tasks, a linear combination may not be enough for exploiting the information in extracted features. We further examine the efficacy of non-linear combination on extracted features, as neural networks achieve superior performance through non-linear feature transformation.
+
+We employ an multiple-layer perceptron (MLP) for nonlinear transformation and compare its performance with linear transformation. Features extracted from different structures of the pre-trained model are concatenated together as input (See Section 4.1 for details).
+
+In Figure 4, we clearly observe that adding non-linearity to feature transformation can significantly boost the performance for the difficult domain adaption tasks (i.e. the datasets on right end of Figure 4). This phenomenon indicates that simply performing linear transformation on extracted features can not exploit the full potential of representations extracted from the pre-trained model. On the other hand, we notice that the performance on easy tasks is not improved, and is even hurt with the introduction of non-linearity. We speculate that this is because the linear transformation can already make full use of extracted information, while non-linearity increases model capacity, which eventually leads to overfitting to the training data.
+
+Discussion. Our investigation reveals that the optimal probing model differs depending on the domain of downstream task. Incorporating diverse features from multiple layers of the pre-trained model and performing non-linear transformation on them can significantly improve the performance on difficult cases. However, this yields worse results on easy tasks. On the contrary, simply performing linear combination on features from deep layers can achieve satisfactory performance for easy adaptation tasks, but not for difficult ones. The conflicting behaviors of model probing motivates us to develop a unified framework that can solve the adaptation problems for both easy and difficult adaptation tasks.
+
+# 4. Structured Model Probing
+
+In this section, we propose structured model probing, a unified probing method which can deal with both easy and difficult cases of adaptation tasks. We first present how to thoroughly extract features in a structured way, then introduce our probing model with structured regularization.
+
+# 4.1. Structured Feature Extraction
+
+Incorporating diverse features extracted from the pre-trained model can improve the performance on difficult adaptation tasks for model probing, as shown in Section 3.2. To provide diverse features for the probing model, we extract features from all structures in a Vision Transformer (ViT). These structures include features before self-attention, features after self-attention, and features after Feed Forward Network (FFN), for each transformer layer. We also include tokenized image inputs and pre-logits. We use  $V = \{v_{i}\}_{i}^{|V|}$  represents all structures in ViT, and  $|V|$  is the total number of structures. For an input sample, we denote each set of tokens extracted from structure  $v_{i}$  as  $X_{i} \in \mathbb{R}^{t \times c}$ , where  $t$  is the dimension of tokens and  $c$  is the dimension of channels.
+
+To reduce feature redundancy while preserving diversity, we apply 1D average pooling along different dimensions on tokens  $X_{i}$ , resulting in two types of aggregation. To maintain channel information, we aggregate  $X_{i}$  along the token dimension, which produces  $x_{i}^{t}$ . And to preserve spatial information, we aggregate  $X_{i}$  along the channel dimension, resulting in  $x_{i}^{c}$ . Finally, we concatenate all structured features into a single vector,  $x = [x_{1}^{t}, x_{1}^{c}, \dots, x_{|V|}^{t}, x_{|V|}^{c}]$ . We denote  $s_{i}$  as the index set of elements in  $x$ , and  $S = \{s_{i} | i = 1\dots |s|\}$ .
+
+# 4.2. Probing with Structured Regularization
+
+Our structured regularization contains two components: structured sparsity regularizer and structured non-linearity regularizer. We introduce them as follows.
+
+# 4.2.1 Structured Feature Selection
+
+As shown in Section 3.2, the informativeness of extracted features varies with tasks, and redundant features lead to performance degradation on some tasks. Thus, we propose to perform feature selection during the training stage of the probing model to prevent overfitting. A popular way is to use group lasso [16, 43] by constructing groups on feature dimension over the weight matrix of a linear classifier:
+
+$$
+\min  _ {\theta} L (\theta) + \lambda \sum_ {i = 1} ^ {d} | | \theta_ {i} | | _ {2}, \text {w h e r e} | | \theta_ {i} | | _ {2} = \sqrt {\sum_ {j = 1} ^ {n} \theta_ {i j} ^ {2}}. \tag {3}
+$$
+
+We use  $\theta \in \mathbb{R}^{d\times n}$  to represent weight matrix, where  $d$  is the feature dimension and  $n$  is the number of classes. However, due to the high dimensionality of features, the final achieved sparsity is hardly the optimal result. It is worth noting that the structure of pre-trained model provides a natural way to divide features as groups, and features extracted from the same structure have higher relevance to each other. Thus, we leverage the structure of the pre-trained model, to achieve the structured sparsity with a structured sparsity regularizer:
+
+$$
+\min  _ {\theta} L (\theta) + \lambda \sum_ {s \in S} \frac {1}{| s |} \| \theta_ {s} \| _ {2}, \text {w h e r e} \| \theta_ {s} \| _ {2} = \sqrt {\sum_ {i \in s} \sum_ {j = 1} ^ {n} \theta_ {i j} ^ {2}}, \tag {4}
+$$
+
+$s \in S$  is the index set of a specific structure in feature as described in Section 4.1, and  $|s|$  represents its dimension. Compared with constructing groups by features, in our regularizer, we construct groups by leveraging the structure of the pre-trained model. This formalization can significantly reduce the combinations of selection and thus prevent overfitting. With such an advantage, our method achieves superior performance with a small number of samples under high-dimensional features.
+
+# 4.2.2 Structured Non-linear Transformation
+
+Our discovery in Section 3.3 shows non-linear transformation can facilitate transfer performance for difficult tasks, while could reduce performance for easy adaptation tasks, mostly due to over-fitting. This means that we need a model that allows flexibly incorporating non-linearity to a linear model. We propose a structured non-linearity regularizer that nicely integrates feature selection (i.e. choosing the subset of informative features from the pre-trained model) with model selection (i.e. choosing either a linear model or nonlinear model for prediction) inspired by hierarchical sparse modeling [28, 42]. Specifically, we construct our model by the combination of a linear model and a neural network:
+
+$$
+f (x) = \theta^ {\top} x + f _ {W} (x), \tag {5}
+$$
+
+where  $\theta \in \mathbb{R}^{d\times n}$  is a linear model, and  $f_{W}$  represents an MLP with model weights  $W$ . We use  $W^{(1)}$  to represent the input layer of the MLP, and the parameters of rest layers are denoted as  $W^{(2:)}$ .
+
+We utilize the feature selection result  $||\theta ||_2$ , the  $\ell_2$  norm of weights assigned to the selected features, to regularize the complexity of the non-linear model:
+
+- When a small number of features are good enough for prediction, we expect a small value for  $||\theta||_2$ . By using  $||\theta||_2$  to control the complexity of non-linear model, we expect the resulting model to be mostly linear.  
+- With increasing  $||\theta||_2$ , more diverse structures from the pre-trained model are needed for prediction. By using  $||\theta||_2$  to regularize the size of non-linear model, we expect the resulting model to be more non-linear.
+
+A way to achieve such control is to add a constraint to the objective function in Eq.(4) as follows:
+
+$$
+\min  _ {\theta , W} L (\theta , W) + \lambda \sum_ {s \in S} | | \theta_ {s} | | _ {2}
+$$
+
+$$
+\begin{array}{l} \text {s u b j e c t} \left. | | W _ {s} ^ {(1)} | | _ {2} \leq M _ {1} | | \theta_ {s} | | _ {2}, s = 1 \dots | S |, \right. \tag {6} \\ \left| \left| W ^ {(2:)} \right| \right| _ {2} \leq M _ {2} \left| \left| \theta \right| \right| _ {2}. \\ \end{array}
+$$
+
+There are two constraints in Eq.(6): the first constraint is used to maintain the input feature sparsity of the neural network, and the second term is used to control the non-linearity through the linear model. To achieve efficient implementation, we convert the constraint in Eq.(6) into a regularization term, which we call structured non-linearity regularizer:
+
+$$
+\begin{array}{l} \Omega (\theta , W) = \sum_ {s \in S} \max  \left\{\left| \left| W _ {s} ^ {(1)} \right| \right| _ {2} - M _ {1} \left| \left| \theta_ {s} \right| \right| _ {2}, 0 \right\} \tag {7} \\ + \max  \left\{\left| \left| W ^ {(2:)} \right| \right| _ {2} - M _ {2} \left| \left| \theta \right| \right| _ {2}, 0 \right\}. \\ \end{array}
+$$
+
+This regularization term means that, when  $||W||_2 > M||\theta||_2$ ,  $W$  receives penalty until it achieves similar complexity as  $M||\theta||$ . Then our final objective function is:
+
+$$
+\min  _ {\theta , W} L (\theta , W) + \lambda_ {1} \sum_ {s \in S} | | \theta_ {s} | | _ {2} + \lambda_ {2} \Omega (\theta , W). \tag {8}
+$$
+
+We use  $\lambda$  to control the penalty strength, and simply set  $\lambda_{2} = 0.1$  works well in our experiments.
+
+Loss function. Since our model  $f(x)$  consists of both linear and non-linear parts, we decouple these two components in loss function. This modification enables the structured sparsity regularizer to control the complexity of  $\| \theta \|$  without the interference from the non-linear part:
+
+$$
+L (\theta , W) = \operatorname {C E} \left(\theta^ {\top} x, y\right) + \operatorname {C E} \left(\operatorname {s g} \left(\theta^ {\top} x\right) + f _ {W} (x), y\right), \tag {9}
+$$
+
+where  $y$  is the one-hot label vector of  $x$ , CE is cross-entropy loss and sg means stop-gradient.
+
+Table 1. Median test accuracy (%) over 3 seeds on the VTAB-1k benchmark using ViT-B/16 pretrained on ImageNet-21k. * indicates results are obtained from [23] and † indicates results are obtained from [16].  
+
+<table><tr><td rowspan="2"></td><td colspan="8">Natural</td><td colspan="6">Specialized</td><td colspan="7">Structured</td><td></td><td></td></tr><tr><td>CTFAR-100</td><td>Caltech101</td><td>DTD</td><td>Flowers102</td><td>Pets</td><td>SVHN</td><td>Sun397</td><td>Mean (Natural)</td><td>Camelyon</td><td>EuroSAT</td><td>Resisc45</td><td>Retinopathy</td><td>Mean (Specialized)</td><td>Clevr-Count</td><td>Clevr-Dist</td><td>DMLab</td><td>KITTI-Dist</td><td>dSpr-Loc</td><td>dSpr-Ori</td><td>sNORB-Azim</td><td>sNORB-Elev</td><td>Mean (Structured)</td><td>Mean (All)</td></tr><tr><td colspan="22">Tuning Methods</td><td></td><td></td></tr><tr><td>Scratch†</td><td>7.6</td><td>19.1</td><td>13.1</td><td>29.6</td><td>6.7</td><td>19.4</td><td>2.3</td><td>14.0</td><td>71.0</td><td>71.0</td><td>29.3</td><td>72.0</td><td>60.8</td><td>31.6</td><td>52.5</td><td>27.2</td><td>39.1</td><td>66.1</td><td>29.7</td><td>11.7</td><td>24.1</td><td>35.3</td><td>32.8</td></tr><tr><td>Fine-tuning*</td><td>68.9</td><td>87.7</td><td>64.3</td><td>97.2</td><td>86.9</td><td>87.4</td><td>38.8</td><td>75.9</td><td>79.7</td><td>95.7</td><td>84.2</td><td>73.9</td><td>83.7</td><td>56.3</td><td>58.6</td><td>41.7</td><td>65.5</td><td>57.5</td><td>46.7</td><td>25.7</td><td>29.1</td><td>47.6</td><td>65.6</td></tr><tr><td>Bias*</td><td>72.8</td><td>87.0</td><td>59.2</td><td>97.5</td><td>85.3</td><td>59.9</td><td>51.4</td><td>73.3</td><td>78.7</td><td>91.6</td><td>72.9</td><td>69.8</td><td>78.3</td><td>61.5</td><td>55.6</td><td>32.4</td><td>55.9</td><td>66.6</td><td>40.0</td><td>15.7</td><td>25.1</td><td>44.1</td><td>62.0</td></tr><tr><td>VPT*</td><td>78.8</td><td>90.8</td><td>65.8</td><td>98.0</td><td>88.3</td><td>78.1</td><td>49.6</td><td>78.5</td><td>81.8</td><td>96.1</td><td>83.4</td><td>68.4</td><td>82.4</td><td>68.5</td><td>60.0</td><td>46.5</td><td>72.8</td><td>73.6</td><td>47.9</td><td>32.9</td><td>37.8</td><td>55.0</td><td>69.4</td></tr><tr><td colspan="22">Probing Methods</td><td></td><td></td></tr><tr><td>Linear</td><td>77.2</td><td>86.8</td><td>66.6</td><td>99.1</td><td>88.8</td><td>35.7</td><td>55.1</td><td>72.7</td><td>78.4</td><td>89.6</td><td>72.2</td><td>73.7</td><td>78.5</td><td>31.0</td><td>29.9</td><td>35.2</td><td>54.9</td><td>11.7</td><td>27.7</td><td>14.4</td><td>23.7</td><td>28.6</td><td>55.4</td></tr><tr><td>Side-Tuning*</td><td>60.7</td><td>60.8</td><td>53.6</td><td>95.5</td><td>66.7</td><td>34.9</td><td>35.3</td><td>58.2</td><td>58.5</td><td>87.7</td><td>65.2</td><td>61.0</td><td>68.1</td><td>27.6</td><td>22.6</td><td>31.3</td><td>51.7</td><td>8.2</td><td>14.4</td><td>9.8</td><td>21.8</td><td>23.4</td><td>45.6</td></tr><tr><td>Head2Toe</td><td>75.3</td><td>90.9</td><td>75.0</td><td>99.5</td><td>86.1</td><td>83.5</td><td>50.9</td><td>80.2</td><td>84.4</td><td>95.7</td><td>84.4</td><td>74.3</td><td>84.7</td><td>51.7</td><td>59.4</td><td>44.0</td><td>65.6</td><td>47.1</td><td>40.3</td><td>32.5</td><td>41.1</td><td>47.7</td><td>67.5</td></tr><tr><td>LinearStruc.</td><td>72.9</td><td>89.5</td><td>72.5</td><td>98.9</td><td>83.6</td><td>71.6</td><td>42.7</td><td>76.0</td><td>84.3</td><td>96.0</td><td>81.8</td><td>75.0</td><td>84.3</td><td>48.2</td><td>52.5</td><td>39.4</td><td>66.7</td><td>48.7</td><td>36.7</td><td>32.3</td><td>37.7</td><td>45.3</td><td>64.8</td></tr><tr><td>LinearStruc.w/FSR</td><td>78.2</td><td>90.1</td><td>73.4</td><td>99.3</td><td>88.1</td><td>71.7</td><td>45.4</td><td>78.0</td><td>83.9</td><td>96.1</td><td>81.9</td><td>74.1</td><td>84.0</td><td>48.5</td><td>52.8</td><td>39.6</td><td>64.4</td><td>51.4</td><td>38.6</td><td>32.9</td><td>37.0</td><td>45.7</td><td>65.7</td></tr><tr><td>LinearStruc.w/SSR</td><td>79.0</td><td>90.2</td><td>74.1</td><td>99.1</td><td>90.4</td><td>73.4</td><td>54.2</td><td>80.1</td><td>84.6</td><td>96.3</td><td>83.8</td><td>74.4</td><td>84.8</td><td>48.8</td><td>52.9</td><td>39.8</td><td>66.9</td><td>48.7</td><td>38.9</td><td>32.5</td><td>37.9</td><td>45.8</td><td>66.6</td></tr><tr><td>MLPStruc.</td><td>66.8</td><td>88.7</td><td>70.3</td><td>97.4</td><td>82.1</td><td>71.1</td><td>33.3</td><td>72.8</td><td>81.6</td><td>95.9</td><td>77.0</td><td>73.5</td><td>82.0</td><td>77.2</td><td>57.7</td><td>39.3</td><td>66.5</td><td>72.3</td><td>44.1</td><td>32.7</td><td>47.7</td><td>54.7</td><td>67.1</td></tr><tr><td>SMP w/o SNR</td><td>71.3</td><td>89.6</td><td>72.3</td><td>99.0</td><td>85.6</td><td>72.7</td><td>45.0</td><td>76.5</td><td>84.2</td><td>95.9</td><td>81.9</td><td>72.8</td><td>83.7</td><td>77.4</td><td>57.1</td><td>40.8</td><td>67.1</td><td>72.4</td><td>44.2</td><td>32.6</td><td>48.4</td><td>55.0</td><td>69.0</td></tr><tr><td>SMP</td><td>79.3</td><td>90.9</td><td>74.9</td><td>99.3</td><td>90.4</td><td>75.0</td><td>55.3</td><td>80.7</td><td>84.8</td><td>96.3</td><td>83.1</td><td>75.0</td><td>84.8</td><td>77.5</td><td>58.0</td><td>40.8</td><td>67.5</td><td>72.5</td><td>44.5</td><td>33.0</td><td>49.0</td><td>55.4</td><td>70.9</td></tr></table>
+
+# 5. Experiments
+
+In this section, we evaluate the proposed structured model probing method on multiple datasets across various tasks, and provide detailed analysis of the proposed method.
+
+Implementation details. We evaluate SMP on the popular vision backbone, ViT-B/16 [15] pre-trained on ImageNet-21k [12], which is widely used across relative research works [16, 23]. We provide details of probing model and hyperparameter selection in the supplementary material.
+
+Compared methods. We compare SMP with other commonly used and state-of-the-art tuning (Fine-tuning, Bias [6] and VPT [23]) and probing methods (Linear Probing, Side-Tuning [45] and Head2Toe [16]). All hyperparameters of baseline methods are carefully tuned, and further information is provided in supplementary material.
+
+# 5.1. Experiments on VTAB-1k
+
+Experimental results on the VTAB-1k benchmark (Table 1) indicate that SMP outperforms other methods, achieving a mean performance of  $70.9\%$ ,  $3.4\%$  higher than Head2Toe and  $1.5\%$  higher than VPT. SMP also significantly outperforms Linear Probing, Bias, and Side-Tuning. SMP achieves optimal results in all three dataset groups. Especially in Structured group, SMP attains a mean accuracy of  $55.4\%$ , surpassing other probing methods by a significant margin. This demonstrates the incorporation of non-linear transformation is important for difficult adaptation tasks.
+
+Ablation studies. We provide results of baseline methods that contribute to SMP, to validate the efficacy of structured
+
+regularization and the contribution of each component.
+
+LinearStruc. represents extracting features by structured feature extraction proposed in Section 4.1, then training a linear classifier. Compare to Linear Probing, it gets  $9\%$  average performance gain, indicating that solely extracting features from the last layer is inadequate for model probing.
+
+We further compare feature sparsity regularization in Eq.(3) (represents by  $\mathrm{Linear}_{\mathrm{Struc.}} + \mathrm{FSR}$ ) to the proposed structured sparsity regularizer (represents by  $\mathrm{Linear}_{\mathrm{Struc.}} + \mathrm{SSR}$ ). Structured sparsity regularizer outperforms feature sparsity regularization, indicating that feature selection can mitigate the overfitting caused by high dimensions, and leveraging model structure as a prior can bring extra benefits.
+
+We apply non-linear transformation on extracted features, represents by  $\mathrm{MLP}_{\text{Struc.}}$ . We find the performance on structured group gets a significant improvement over  $\mathrm{Linear}_{\text{Struc.}}$ , demonstrating the importance of non-linear transformation on difficult adaptation tasks. However, the performance on easy tasks, i.e., natural and specialized groups, is degraded, suggesting the risk of overfitting from non-linear transformation. We solve this problem by leveraging structured nonlinearity regularizer (SNR), which can automatically control the non-linearity through the norms of selected structures.
+
+# 5.2. Experiments on Few-shot Learning
+
+In this section, we evaluate our method in few-shot learning scenario, where the number of training samples varies from 1 to 16 shots per class following existing studies. Five fine-grained recognition tasks are chosen, i.e., Food101 [4], CUB-200-2011 [39], Stanford Cars [26], Stanford Dogs [24]
+
+![](images/817f5dd48fe878aaf1b3177c862d35bfe2ed6a9e9e8e3980f397eaecfecb0609.jpg)
+
+![](images/e2fe9524fb302c804246dd5f00e0babe0a03ccc14a1e9abb24d528eab516d278.jpg)
+
+![](images/484d8aefe8aaa968b7c51fd933bef7233ff1cddb14dede596e27d1856b3004b2.jpg)
+
+![](images/1dcfc2b477ee64adeaf90836b3949d8250e549672b924d614cbda5c630efdc49.jpg)  
+Figure 5. Performance of few-shot learning on fine-grained visual recognition tasks. LinearStr. (F) indicates LinearStruc. w/ FSR.
+
+![](images/47064e96137db55553faad763127478d043cbcae2ca099511d7e6dc3d2b89899.jpg)
+
+![](images/e7681287e02199c4d9a68a98c6b192a4e43bb4f3e93194a7da447464654dafba.jpg)
+
+and Oxford Pets [35]. We compared SMP with two strong baselines, LinearStruc. and LinearStruc. w/ FSR, and two prominent methods, VPT (tuning) and Head2Toe (probing).
+
+Experimental results presented in Figure 5 show that SMP consistently outperforms other methods in most cases, demonstrating its effectiveness in few-shot scenario, and its ability to improve performance with increasing training data. Notably, feature sparsity regularization performs worse than SMP, highlighting the importance of leveraging pre-trained model structure as prior knowledge in low-shot scenario.
+
+# 5.3. Experiments on More Transfer Scenarios
+
+Larger downstream datasets. We provide average results on 9 full-size downstream datasets in Table 2. It can be observed even in the larger data regime, SMP still achieves competitive performance compared with tuning methods. Full results are presented in Suppl. B.1.
+
+Different pre-trained models. We present VTAB-1k results on ViT-B/16 and ViT-L/14 pretrained by CLIP in Suppl. B.2. Despite the usage of stronger pre-trained models, SMP continuously outperforms baseline methods due to the incorporation of diverse features and non-linear transformation, as well as its flexible framework.
+
+Different architectures. SMP is a versatile method that can be applied to convolutional neural networks as well. In Suppl. B.3, we present the VTAB-1k results obtained from ImageNet pre-trained ResNet-50 models, where SMP outperforms other methods and achieves superior performance.
+
+Table 2. Average accuracy (%) on 9 full-size downstream datasets.  
+
+<table><tr><td></td><td>Fine-tuning</td><td>VPT</td><td>Linear</td><td>SMP</td></tr><tr><td>Average Accuracy</td><td>88.4</td><td>88.7</td><td>81.7</td><td>88.7</td></tr></table>
+
+# 5.4. Analysis and Discussion
+
+In this section, we conduct various experiments to analyze the components of the proposed SMP method. Full results of this section are presented in Suppl. C.
+
+Visualization of group norms. We visualize the  $\ell_2$  norm of each structure on different tasks in Figure 6, which helps us to understand the behavior of the proposed structured regularization. For Sun397, an easy adaptation task, we can observe the structures from deep layers have larger norms in Figure 6a. This validates our analysis in Section 3.2, that for easy adaptation tasks, features extracted from deep layers contain more information. For Clevr-Dist, a difficult adaptation task, the structures have similar norms overall, which means for difficult tasks, information contained in intermediate layers is also important, and diverse features can improve the performance, as we find in Section 3.2. This motivates the proposed structured non-linearity regularization, utilizing  $\ell_2$  norm to control the complexity of non-linear model.
+
+Retraining probing model by gradually incorporating structures. We retrain a probing model by gradually incorporating structures selected based on the magnitude of their  $\ell_2$  norm, to validate their contribution to the prediction. Results are presented in Figure 7. For Sun397, an easy adaptation task, we can see in Figure 7a that there are redun
+
+![](images/e99f4b82417c31d3c5087dcde2007334e8fda3dd38a873613a6132ae66d91ab8.jpg)  
+(a) Sun397 (easy)
+
+![](images/1fb3d64faad6bf3cad0400800a95c05e9b0fe2efb2efe6b342be0c3a1ad2b55b.jpg)  
+(b) Clevr-Dist (difficult)  
+Figure 6.  $\ell_2$  norm of structure groups on different tasks.  
+Figure 7. Retraining performance by gradually incorporating structures with larger  $\ell_{2}$  norm obtained by SMP.
+
+dant structures. And with a few selected structures, we can achieve comparable results compared with the original SMP, justifying the zero norm of some structured observed after performing structure regularization in Figure 6a. However, for Clevr-Dist, a difficult adaptation task, comparable performance is achieved until selecting 55 structures, suggesting intermediate structures also contribute to the final prediction. This demonstrates the necessity of proposed structured regularizer, which resolves the contradiction between easy and difficult domain adaptation tasks.
+
+Impact of feature aggregation designs. The primary objective of performing token and channel aggregation is to minimize the redundancy in the extracted raw features, while maintaining feature diversity in the compact aggregated representation. To achieve this, two types of aggregation are designed to preserve channel and spatial information, respectively. We provide ablation studies in Suppl. C.5. Overall, channel information (aggregating along tokens) is essential for all datasets, while spatial information (aggregating along channels) is beneficial for tasks that are sensitive to spatial location, e.g., dSpr-Loc and sNORB-Azim.
+
+Impact of different structures. To increase the feature diversity, we incorporate features extracted from various structures, coupled with a structured sparse regularizer to conduct model fitting and feature selection simultaneously. This flexible framework enables us to avoid the costly manual selection of candidate structures. We provide analysis of structures in Suppl. C.6. The results show that incorporating all candidate structured features yields best mean results, which demonstrates the efficacy of our method.
+
+Cost of SMP. We compare the cost of SMP with baseline methods, including relative adaptation time, extra inference FLOPs, and extra parameters. The average results on VTAB-
+
+![](images/c4addc513272d803bc245c903c01bcb20ac3f895ce7323006e5714743781c417.jpg)  
+(a) Sun397 (easy)
+
+![](images/67152e2b9413ff984d8288c4e3f256e6f520b098e0b19564a1c7c9735b629693.jpg)  
+(b) Clevr-Dist (difficult)
+
+Table 3. Cost comparison (Average results on VTAB-1k). * indicates relative results compared with Fine-tuning.  
+
+<table><tr><td></td><td>Linear</td><td>FT</td><td>VPT</td><td>SMP</td></tr><tr><td>Adaptation time (rel.)*</td><td>0.05</td><td>1</td><td>0.97</td><td>0.15</td></tr><tr><td>Extra Inference FLOPs (G)</td><td>0</td><td>0</td><td>11.01</td><td>0.01</td></tr><tr><td>Extra parameters (M)</td><td>0.04</td><td>85.83</td><td>0.64</td><td>6.91</td></tr></table>
+
+1k are shown in Table 3. SMP only needs  $15\%$  adaptation time on average compared with fine-tuning. On the contrary, the adaptation time of VPT is  $97\%$  of fine-tuning. This results in a significantly lower adaptation cost of SMP compared to tuning methods. SMP is also very efficient during inference time, and only introduces 0.01G inference FLOPs compared with fine-tuning. It can be noticed that VPT introduces extra 11.01G inference FLOPs, which is far more than SMP. SMP requires 6.91M extra parameters on average, which is the  $8\%$  of fine-tuning. Though the requirement of extra parameters is higher than VPT, we can leverage the retraining method proposed in Section 5.4, which can further reduce the extra parameters. Overall, SMP is a highly efficient method with low adaptation cost, fast inference, and fewer extra parameters, while maintaining superior performance compared to tuning-based methods.
+
+# 6. Conclusion
+
+In this paper, we present structured model probing, an effective yet efficient probing method for transfer learning. Our investigation reveals that model probing behaves differently for easy and difficult adaptation tasks. To mitigate this conflicting behavior, we propose structured model probing, a method that is able to achieve good performance in both cases. With the proposed structured regularization, we can construct a simple adaptation model for easy adaptation cases, and increase the complexity of the adaptation model for difficult cases. Our method outperforms state-of-the-art tuning methods while maintaining computational efficiency. Our work highlights the potential of model probing, which motivates researchers to better exploit the use large-scale pre-trained model and apply model probing to diverse tasks.
+
+# References
+
+[1] Alessandro Achille, Michael Lam, Rahul Tewari, Avinash Ravichandran, Subhransu Maji, Charless C. Fowlkes, Stefano Soatto, and Pietro Perona. Task2vec: Task embedding for meta-learning. In IEEE International Conference on Computer Vision (ICCV), pages 6429-6438, 2019. 3  
+[2] Alessandro Achille, Aditya Golatkar, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. LQF: linear quadratic fine-tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15729-15739, 2021. 3  
+[3] Armen Aghajanyan, Akshit Shrivastava, Anchit Gupta, Naman Goyal, Luke Zettlemoyer, and Sonal Gupta. Better finetuning by reducing representational collapse. In 9th International Conference on Learning Representations (ICLR), 2021. 2  
+[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446-461, 2014. 6  
+[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020. 1, 2  
+[6] Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. Tinytl: Reduce memory, not parameters for efficient on-device learning. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020. 2, 6  
+[7] Shoufa Chen, Chongjian GE, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In Advances in Neural Information Processing Systems 35 (NeurIPS), 2022. 2  
+[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 1597-1607, 2020. 2  
+[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In IEEE International Conference on Computer Vision (ICCV), pages 9620-9629, 2021. 2  
+[10] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4109-4118, 2018. 3  
+[11] Mostafa Dehghani, Yi Tay, Anurag Arnab, Lucas Beyer, and Ashish Vaswani. The efficiency misnomer. In 10th International Conference on Learning Representations (ICLR), 2022. 2
+
+[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 3, 6  
+[13] Aditya Deshpande, Alessandro Achille, Avinash Ravichandran, Hao Li, Luca Zancato, Charless C. Fowlkes, Rahul Bhotika, Stefano Soatto, and Pietro Perona. A linearized framework and a new benchmark for model selection for fine-tuning. arXiv:2102.00084, 2021. 3  
+[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 4171–4186, 2019. 2  
+[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations (ICLR), 2021. 3, 6  
+[16] Utku Evci, Vincent Dumoulin, Hugo Larochelle, and Michael C. Mozer. Head2Toe: Utilizing intermediate representations for better transfer learning. In Proceedings of the 39th International Conference on Machine Learning (ICML), pages 6009-6033, 2022. 1, 3, 5, 6  
+[17] Weifeng Ge and Yizhou Yu. Borrowing treasures from the wealthy: Deep transfer learning through selective joint finetuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10-19, 2017. 2  
+[18] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogério Schmidt Feris. SpotTune: Transfer learning through adaptive fine-tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4805-4814, 2019. 2  
+[19] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In 10th International Conference on Learning Representations (ICLR), 2022. 1, 2  
+[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2, 3  
+[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2022. 2, 3  
+[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning (ICML), pages 2790-2799, 2019. 2  
+[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser-Nam Lim.
+
+Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6  
+[24] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition (CVPR Workshop), 2011. 6  
+[25] Simon Kornblith, Jonathon Shlens, and Quoc V. Le. Do better imagenet models transfer better? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2661-2671, 2019. 2, 3  
+[26] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshop (ICCV Workshop), pages 554–561, 2013. 6  
+[27] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In 10th International Conference on Learning Representations (ICLR), 2022. 2  
+[28] Ismael Lemhadri, Feng Ruan, Louis Abraham, and Robert Tibshirani. Lassonet: A neural network with feature sparsity. Journal of Machine Learning Research, 22:127:1-127:29, 2021. 5  
+[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3045-3059, 2021. 2  
+[30] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), pages 4582-4597, 2021. 2  
+[31] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5196-5205, 2022. 2  
+[32] Fangzhou Mu, Yingyu Liang, and Yin Li. Gradients as features for deep representation learning. In 8th International Conference on Learning Representations (ICLR), 2020. 3  
+[33] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. LEEP: A new measure to evaluate transferability of learned representations. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 7294-7305, 2020. 3  
+[34] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2010. 2  
+[35] Omkar M. Parkhi, Andreae Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3498-3505, 2012. 7  
+[36] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 2227-2237, 2018. 3
+
+[37] 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 Proceedings of the 38th International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1, 2  
+[38] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. LST: Ladder side-tuning for parameter and memory efficient transfer learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3  
+[39] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical report, California Institute of Technology, 2011. 6  
+[40] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In 10th International Conference on Learning Representations (ICLR), 2022. 2  
+[41] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer G. Dy, and Tomas Pfister. Learning to prompt for continual learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 2  
+[42] Xiaohan Yan and Jacob Bien. Hierarchical Sparse Modeling: A choice of two group lasso formulations. Statistical Science, 32(4):531 - 560. 5  
+[43] Ming Yuan and Yi Lin. Model selection and estimation in regression with grouped variables. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(1): 49-67, 2006. 5  
+[44] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv:1910.04867, 2019. 1, 2, 3  
+[45] Jeffrey O. Zhang, Alexander Sax, Amir Zamir, Leonidas J. Guibas, and Jitendra Malik. Side-tuning: A baseline for network adaptation via additive side networks. In European Conference on Computer Vision (ECCV), pages 698-714, 2020. 1, 2, 3, 6  
+[46] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16795-16804, 2022. 2  
+[47] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2  
+[48] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 2

2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd932ef59971783b170d567a689bd7dbfd0fad6b7014abd5ca8fd871f2bcfd8f
+size 678822

2024/StyLitGAN_ Image-Based Relighting via Latent Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_content_list.json

@@ -0,0 +1,1348 @@
+[
+    {
+        "type": "text",
+        "text": "StyLitGAN: Image-based Relighting via Latent Control",
+        "text_level": 1,
+        "bbox": [
+            202,
+            130,
+            767,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Anand Bhattad James Soole D.A. Forsyth",
+        "bbox": [
+            274,
+            181,
+            694,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Illinois Urbana-Champaign",
+        "bbox": [
+            321,
+            199,
+            648,
+            217
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "https://anandbhattad.github.io/stylitgan/",
+        "bbox": [
+            241,
+            217,
+            725,
+            233
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Generated Image",
+        "bbox": [
+            133,
+            255,
+            227,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Relit-1  $(\\mathbf{w}^{+} + d_{1})$  Relit-2  $(\\mathbf{w}^{+} + d_{2})$  Relit-3  $(\\mathbf{w}^{+} + d_{3})$  Relit-4  $(\\mathbf{w}^{+} + d_{4})$  Relit-5  $(\\mathbf{w}^{+} + d_{5})$",
+        "bbox": [
+            245,
+            255,
+            841,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/ca50ceef1b4c4acb1b024de96b1a74ca80e39eb45597d829b2b3de7949c1e4ba.jpg",
+        "image_caption": [
+            "Figure 1. StyLitGAN identifies directional vectors  $(d_i)$  within StyleGAN's style space  $(\\mathcal{W}^{+})$  which, when added to the  $\\mathbf{w}^{+}$ style code, effectively modify the lighting of generated images while preserving their geometry and albedo. This process eliminates the need for per-image search or model fine-tuning. The first column displays images generated from StyleGAN2; subsequent columns illustrate the same scene, each relit using a specific direction. These relighting directions  $(d_i)$  are derived through a forward selection method, ensuring diversity and avoiding cherry-picking. The directional effects are consistent across different scenes: for instance,  $d_1$  activates an orange-tinged bedside lamp,  $d_2$  a less intense white-tinged lamp,  $d_3$  introduces strong directional light from the window, and so on, demonstrating diverse relighting capabilities of StyLitGAN."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            267,
+            846,
+            549
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            661,
+            313,
+            678
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We describe a novel method, StyLitGAN, for relighting and resurfacing images in the absence of labeled data. StyLitGAN generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects, without the need for paired or CGI data. StyLitGAN uses an intrinsic image method to decompose an image, followed by a search of the latent space of a pretrained StyleGAN to identify a set of directions. By prompting the model to fix one component (e.g., albedo) and vary another (e.g., shading), we generate relighted images by adding the identified directions to the latent style codes. Quantitative metrics of change in albedo and lighting diversity allow us to choose effective directions using a forward selection process. Qualitative evaluation confirms the effectiveness of our method.",
+        "bbox": [
+            73,
+            688,
+            473,
+            902
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            501,
+            662,
+            630,
+            679
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Scene appearance shifts dramatically with varying lighting conditions - a sunlit room takes on a different character as daylight fades, and interior spaces transform with the flick of a switch. Similarly, surface changes, like a wall's paint color, change not only the wall's appearance but also the overall image due to light reflection. Despite the impressive realism achieved by current generative models like StyleGAN [22-24], they fall short in dynamically controlling scene lighting, a key aspect of realistic image generation.",
+        "bbox": [
+            496,
+            688,
+            893,
+            824
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this work, we present StyLitGAN, a novel approach that extends the editing capabilities of StyleGAN [38, 45, 46, 53]. StyLitGAN uniquely manipulates style codes to selectively change lighting while preserving other image attributes like albedo and geometry. This selective editing",
+        "bbox": [
+            496,
+            824,
+            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,
+            44
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "4231",
+        "bbox": [
+            482,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "addresses a critical gap in current generative methods, which typically lack the precision to control individual scene components independently.",
+        "bbox": [
+            75,
+            90,
+            470,
+            136
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our method, StyLitGAN, first uses StyleGAN to produce a set of images and then decomposes these generated images into albedo, diffuse shading, and glossy effects using an off-the-shelf, self-supervised network [14]. We then search for style code edits by prompting StyleGAN to produce images that (a) are diverse, but (b) have the same albedo (and so geometry and material) as the original generated images. Our search selects the most effective relighting directions in a data-driven manner.",
+        "bbox": [
+            75,
+            136,
+            468,
+            271
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our approach generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects. Importantly, we observe that style code edits produce consistent effects across images. For instance, as seen in Fig. 1, adding the first and second directions tends to switch on bedside lamps (columns Relit-1 and Relit-2), while adding the fourth direction increases the light intensity from outside the window (column Relit-4). Since StyLitGAN can generate any image that a vanilla StyleGAN can, it also generate images that are out of distribution, one would expect FID scores to increase over StyleGAN; this happens.",
+        "bbox": [
+            75,
+            272,
+            470,
+            438
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The recent method GAN-control [39] controls lighting on face images using an attribute procedure. We show, in contrast to StyLitGAN, GAN-control fails on indoor scenes, likely because the attribute vocabulary is too easily subverted by the complex lighting effects in indoor scenes.",
+        "bbox": [
+            75,
+            438,
+            468,
+            513
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We demonstrate applications of StyLitGAN to standard vision problems. Using the Multi-Ilum dataset of [29], we show that predictions from a SOTA surface normal predictor [21] vary significantly when lighting is changed. Finetuning this normal predictor using StyLitGAN images suppresses this effect. The improvement is comparable with that obtained by finetuning with true multi-illuminant images (which are very difficult to obtain in quantity).",
+        "bbox": [
+            75,
+            513,
+            470,
+            635
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            648,
+            218,
+            664
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Image Manipulation: A significant literature deals with manipulating and editing images [3, 10, 11, 15, 17, 27, 33, 35, 44, 54]. Editing procedures for generative image models [16] are important, because they demand compact image representations with useful, disentangled interpretations. StyleGAN [22-24] is currently de facto state-of-the-art for editing generated images, likely because its mapping of initial noise vectors to style codes which control entire feature layers produces latent spaces that are heavily disentangled and so easy to manipulate. Recent editing methods include [8, 36, 38, 45, 46, 53], with a survey in [47]. The architecture can be adapted to incorporate spatial priors for authoring novel and edited images [13, 28, 43]. In contrast to this literature, we show how to fix one physically meaningful image factor while changing another. Doing so is difficult",
+        "bbox": [
+            75,
+            672,
+            470,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "because the latent spaces are not perfectly disentangled, and we must produce a diverse set of changes in only one factor. Relighting using StyleGAN: Relighting faces using StyleGAN can be achieved with Stylerig [43], but this method requires a 3D morphable face model. In contrast, StyLitGAN does not require a 3D model and can be extended to complex indoor scenes, which is not possible with Stylerig. Yang et al. [48] uses semantic label attributes to train a binary classifier to find latent space directions that represent indoor and natural lighting, but this method cannot produce diverse relighting effects. We also find Yang et al's relighting to change color or albedo. In contrast, StyLitGAN generates diverse realistic relighting effects without changing albedo and without requiring any labeled attributes.",
+        "bbox": [
+            496,
+            90,
+            893,
+            301
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "StyleFlow [1] and GAN-control [39] require a parametric model to express lighting, such as spherical harmonics. These methods are limited to relighting faces and do not result in realistic relighting of rooms. Our experiments using GAN-control for rooms result in large geometry and albedo change. In contrast, StyLitGAN can produce relighted images without changing geometry or albedo. We also note that rooms are more challenging to relight than faces due to significant long-scale inter-reflection effects, diverse shadow patterns, stylized luminaires, stylized surface albedos, and surface brightnesses that are not a function of surface normal alone. These factors make it difficult to apply GAN control directly to rooms. Also, none of these methods can resurface or recolor rooms, though StyLitGAN can also edit color or materials while preserving the scene's lighting.",
+        "bbox": [
+            496,
+            301,
+            893,
+            527
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Other Face Relighting methods use carefully collected supervisory data from light-stages or parametric spherical harmonics [30, 32, 37, 41, 52]. ShadeGAN [31], Rendering with Style [6], and Volux-GAN [42] use a volumetric rendering approach to learn the 3D structure of the face and the illumination encoding. Volux-GAN [42] also requires image decomposition from [32] that is trained using carefully curated light-stage data. In comparison, we neither require any explicit 3D modeling of the scene nor labeled and curated data for training the image decomposition model.",
+        "bbox": [
+            496,
+            529,
+            893,
+            680
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3. Approach",
+        "text_level": 1,
+        "bbox": [
+            500,
+            694,
+            607,
+            710
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We follow convention and manipulate StyleGAN [24] by adjusting the  $\\mathbf{w}^{+}$  latent variables. We do not modify StyleGAN weights, but instead, seek a set of lighting directions  $\\mathbf{d}_i$  (same shape as  $\\mathbf{w}^{+}$ ) which are independent of  $\\mathbf{w}^{+}$  and have desired effects on the generated image. We obtain these directions by constructing losses that capture the desired outcomes, and then search for directions that minimize these losses. We find all directions only once and use 2000 randomly generated images for this search. Once found, these lighting directions apply to all other generated images. Our search procedure only sees each image once.",
+        "bbox": [
+            496,
+            719,
+            893,
+            883
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Fig. 2 summarizes our procedure and we call our model",
+        "bbox": [
+            517,
+            885,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "4232",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/ef00eae8814e93b30e2a78df87c02ad12b5ad1ac95088f9a84c9e6b3b31bbaa6.jpg",
+        "image_caption": [
+            "Figure 2. How StyLitGAN works: We generate an image from random Gaussian noise using a pretrained StyleGAN. We also generate novel relighted versions (16 in our case) of the same image using randomly initialized latent directions (d) that are added to  $\\mathbf{w}^{+}$ latent style codes. We train a classifier (F) that takes in all the pairs of relighted and original images and predicts the relighting direction applied to them. We apply a distinction loss and jointly update the latent directions and the classifier. Next, we generate the decomposition of these images from a pretrained decomposition model (D). We then apply losses that force StyLitGAN to find latent directions such that the albedo does not change (consistency loss), but the image does (diversity loss)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            87,
+            844,
+            252
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "StyLitGAN. Our method consists of two stages. The first stage involves decomposing images using a pretrained model. The second stage jointly searches for directions and trains a classifier  $\\mathrm{F}$ . The classifier  $\\mathrm{F}$  predicts the latent direction applied to image pairs. It's a classification task with a fixed number of jointly learned latent directions. While classifier  $\\mathrm{F}$  doesn't directly know whether latent directions relate to lighting, our use of image decomposition losses ensures these directions are lighting-related. The consistency loss maintains albedo, and the diversity loss ensures diverse shading changes, both properties expected when changing lighting conditions. Thus, while jointly updating the classifier  $\\mathrm{F}$  and latent directions, these losses ensure the discovered directions pertain to relighting. We now elaborate on our search for directions and losses in detail.",
+        "bbox": [
+            75,
+            349,
+            472,
+            575
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Base StyleGAN Models: We use baseline pretrained models from [49] that use a dual-contrastive loss to train StyleGAN for bedrooms, faces, and churches. We also use baseline pretrained StyleGAN2 models from [13] for conference rooms, kitchens, dining rooms, and living rooms.",
+        "bbox": [
+            75,
+            579,
+            470,
+            655
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Decomposition: We decompose images into albedo, shading, and gloss maps (gloss only when available) as  $A \\times S + G$ , where  $A$  models albedo effects and  $S$  and  $G$  model shading and gloss effects respectively. We use the method of Forsyth and Rock [14] and its variant from Bhattad and Forsyth [3], which is easily adapted because it is self-supervised and uses only samples from statistical models derived from Land's Retinex theory [25]. By changing the statistical spatial model parameters, we can construct many decompositions using their approach. We evaluate many such decomposition models under several hyperparameter settings and create a large pool of relighting directions. We finalize our directions using a forward selection process that provides minimal albedo and geometry shift with a large relighting diversity (Section 4).",
+        "bbox": [
+            75,
+            656,
+            470,
+            868
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Relighting a scene should produce a new, realistic image where the shading has changed but the albedo has",
+        "bbox": [
+            76,
+            869,
+            470,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "not. Write  $I(\\mathbf{w}^{+})$  for the image produced by StyleGAN given style codes  $\\mathbf{w}^{+}$ , and  $A(I)$ ,  $S(I)$ , and  $G(I)$  for the albedo, shading, and gloss respectively recovered from image  $I$ . We search for multiple directions  $\\mathbf{d}_i$  such that: (a)  $A(I(\\mathbf{w}^{+} + \\mathbf{d}_i))$  is very close to  $A(I(\\mathbf{w}^{+}))$  – so the image is a relighted version of  $I(\\mathbf{w}^{+})$ , a property we call persistent consistency; (b) the images produced by the different directions are linearly independent – relighting diversity; (c) which direction was used can be determined from the image, so that different directions have visibly distinct effects – distinctive relighting; and (d) the new shading field (map) is not strongly correlated to the albedo – independent relighting. Not every shading field can be paired with a given albedo, otherwise there would be nothing to do. We assume that edited  $\\mathbf{w}^{+}$  will result in realistic images [7].",
+        "bbox": [
+            496,
+            349,
+            893,
+            577
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Recoloring: Alternatively, we may wish to edit scenes where the colors or materials of objects have changed, but the lighting hasn't. Because shading conveys a great deal of information about shape, we can find these edits using modified losses by seeking consistency in the shading field.",
+        "bbox": [
+            496,
+            577,
+            893,
+            652
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Persistent Consistency: The albedo decomposition of both the relighted scene:  $A_{R} = A(I(\\mathbf{w}^{+} + \\mathbf{d}_{i}))$  and the original:  $A_{O} = A(I(\\mathbf{w}^{+}))$  must be the same; where R refers to relighted images and O refers to StyleGAN generated images. We use a Huber loss and a perceptual feature loss [20, 51] from a VGG feature extractor (Φ) [40] at various feature layers (j) to preserve persistent effects (geometry, appearance, and texture) in the scene.",
+        "bbox": [
+            496,
+            657,
+            893,
+            777
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n s t}} \\left(A _ {O}, A _ {R}\\right) = \\left\\{ \\begin{array}{c} \\frac {1}{2} \\left[ A _ {O} - A _ {R} \\right] ^ {2} \\quad \\text {f o r} | A _ {O} - A _ {R} | \\leq \\delta , \\\\ \\delta \\left(| A _ {O} - A _ {R} | - \\delta / 2\\right) \\quad \\text {o t h e r w i s e .} \\end{array} \\right. (1) \\\\ \\mathcal {L} _ {p e r} \\left(A _ {O}, A _ {R}\\right) = \\left| \\left| \\Phi_ {j} \\left(A _ {O}\\right) - \\Phi_ {j} \\left(A _ {R}\\right) \\right| \\right| _ {2}. (2) \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            500,
+            787,
+            890,
+            849
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Relighting Diversity: We want the set of relighted images produced by the directions to be diverse on a long scale so that regions that were in shadow in one image might be",
+        "bbox": [
+            496,
+            854,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "4233",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "bright in another. For each  $S(\\mathbf{w}^{+} + \\mathbf{d}_{i})$ , we stack the two shading and gloss:  $S$  and  $G$ , and compute a smoothed and downsampled vector  $\\mathbf{t}_{i}$  from these maps. We then compute  $\\mathcal{L}_{div}(S,G)$  (diversity loss) which compels these  $\\mathbf{t}_{i}$  to be linearly independent and encourages diversity in relighting.",
+        "bbox": [
+            76,
+            90,
+            468,
+            167
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {d i v}} (S, G) = - \\log \\det  N \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            176,
+            468,
+            191
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $i^{th}$  &  $j^{th}$  component of  $\\mathbf{N}$  is  $t_i^\\top t_j$",
+        "bbox": [
+            96,
+            200,
+            370,
+            217
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Distinctive Relighting: A network might try to cheat by making minimal changes to the image. Directions  $\\mathbf{d}_i$  should have the property that  $\\mathbf{d}_i$  is easy to impute from  $I(\\mathbf{w}^{+} + \\mathbf{d}_{i})$ . Joint with the search for directions, we train a classifier to categorize the applied direction. This classifier accepts  $I(\\mathbf{w}^{+})$  and  $I(\\mathbf{w}^{+} + \\mathbf{d}_{i})$  and must predict  $i$ . Its cross-entropy supplies our loss:",
+        "bbox": [
+            76,
+            220,
+            468,
+            328
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\min  _ {l, F} \\mathcal {L} _ {\\text {d i s t}} (I (\\mathbf {w} ^ {+}), I (\\mathbf {w} ^ {+} + d _ {i})) \\\\ = - \\sum_ {i = 1} ^ {M} y _ {i} \\log F \\left(I \\left(\\mathbf {w} ^ {+}\\right), I \\left(\\mathbf {w} ^ {+} + d _ {i}\\right)\\right) \\tag {4} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            151,
+            337,
+            468,
+            398
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Saturation Penalty: Our diversity loss might cheat and obtain high diversity by generating blocks of over-saturated or under-saturated pixels. So, we apply a saturation penalty over several pixels within a certain threshold.",
+        "bbox": [
+            76,
+            404,
+            468,
+            465
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {s a t} = \\lambda_ {\\text {o v e r s a t}} \\left[ \\frac {1}{H \\times W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\max  (0, I _ {i, j} - s) ^ {2} \\right] \\tag {5} \\\\ + \\lambda_ {\\text {u n d e r s a t}} \\left[ \\frac {1}{H \\times W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\max  (0, s - I _ {i, j}) ^ {2} \\right] \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            88,
+            476,
+            468,
+            563
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda_{oversat}$  and  $\\lambda_{undersat}$  are the penalty weights for over-saturation and under-saturation respectively,  $H$  and  $W$  are the height and width of the images,  $I_{i,j}$  is the pixel intensity at pixel location  $(i,j)$ , and  $s$  is the saturation threshold (i.e., the maximum allowed pixel intensity). The penalty is computed as the mean squared difference between the pixel intensity and the saturation threshold.",
+        "bbox": [
+            75,
+            575,
+            468,
+            679
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Recoloring requires swapping albedo and shading components in all losses, except we do not use decorrelation loss while recoloring. Obtaining good results requires quite a careful choice of loss weights ( $\\lambda$  coefficients). We experiment with several  $\\lambda$  coefficients for both these edits (Section 4 and Supplementary).",
+        "bbox": [
+            75,
+            680,
+            470,
+            772
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. Model and Directions Selection",
+        "text_level": 1,
+        "bbox": [
+            76,
+            784,
+            362,
+            800
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We prompt StyleGAN to find style code directions that: (a) do not change the albedo, and (b) strongly change the image. We use many image decomposition models to obtain directions across multiple hyperparameter settings. We have no particular reason to believe that a single model will give only good directions or all good directions. We then find",
+        "bbox": [
+            75,
+            809,
+            470,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "a subset of admissible models. We must choose admissible models using a plot of albedo change versus diversity because there is no way to weigh these effects against one another. However, relatively few methods are admissible – see Figure 3. We then pool all directions from all of the admissible models and use forward selection to find a small set of polished directions in this pool.",
+        "bbox": [
+            496,
+            90,
+            893,
+            196
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Scoring Albedo Change: We use SuperPoints [19] to find 100 interest points in the original StyleGAN-generated image. Around each interest point, we form a  $8 \\times 8$  patch. We then compare these patches with patches in the same locations for multiple different relightings of that image. If the albedo in the image does not change, then each patch will have the same albedo but different lighting.",
+        "bbox": [
+            496,
+            204,
+            893,
+            309
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Given two color image patches  $\\mathbf{p}$  and  $\\mathbf{q}$ , viewed under different lights, we must measure the difference between their albedos  $d_{a}(\\mathbf{p},\\mathbf{q})$ . Write  $p_{ij}$  for the RGB vector at the  $i$ ,  $j$ 'th location ( $1 \\leq i \\leq M$ ,  $1 \\leq j \\leq N$ ) and write  $p_{ij,k}$  for the  $k$ 'th RGB component at that location. The intensity of the light may change without the albedo changing, so this problem is homogeneous (i.e. for  $\\lambda, \\mu > 0$ ,  $d_{a}(\\mathbf{p},\\mathbf{q}) = d_{a}(\\lambda \\mathbf{p},\\mu \\mathbf{q}))$ . Assume that the illumination intensity changes, but not color. The patches are small, so the illumination field on a patch can be modeled as a linear function, so there are albedos  $\\mathbf{a}$ ,  $\\mathbf{b}$  such that  $p_{ij} = (p_{x}i + p_{y}j + p_{c})a_{ij}$  and  $q_{ij} = (q_{x}i + q_{y}j + q_{c})b_{ij}$ . If the two patches have similar albedo, there will be  $p_{x}$  etc. such that  $p_{ij}' = (q_{x}i + q_{y}j + q_{c})p_{ij}$  is the same as  $q_{ij}' = (p_{x}i + p_{y}j + p_{c})q_{ij}$ . We measure the cosine distance",
+        "bbox": [
+            496,
+            311,
+            893,
+            539
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nd _ {a} (\\mathbf {p}, \\mathbf {q}) = 1 - \\max  _ {p _ {x}, \\dots q _ {c}} \\frac {\\sum_ {i j k} p _ {i j k} ^ {\\prime} q _ {i j k} ^ {\\prime}}{\\sqrt {\\sum_ {i j k} \\left(p _ {i j k} ^ {\\prime}\\right) ^ {2}} \\sqrt {\\sum_ {i j k} \\left(p _ {i j k} ^ {\\prime}\\right) ^ {2}}} \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            506,
+            551,
+            890,
+            597
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The relevant maximum can be calculated by analogy with canonical correlation analysis (Supplementary).",
+        "bbox": [
+            498,
+            608,
+            890,
+            640
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Scoring Lighting Diversity: Illumination cone theory [2] yields that any non-negative linear combination of  $k$  shadings is a physically plausible shading. To determine if an image is new, we relax the non-negativity constraint and so must ensure that it cannot be expressed as a linear combination of existing images. In turn, we seek a measure of the linear independence of a set of images. This measure should: be large when there is a strong linear dependency; and not grow too fast when the images are scaled. Write  $\\mathbf{x}_i$  for the  $i$ 'th image, and  $\\mathcal{X}$  for the matrix whose  $i$ ,  $j$ 'th component is  $\\mathbf{x}_i\\mathbf{x}_j$ . Then  $-\\log \\det \\mathcal{X}$  is very large when the  $\\mathbf{x}_i$  is close to linearly dependent, but does not scale too fast when the images are scaled.",
+        "bbox": [
+            496,
+            648,
+            890,
+            844
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Decomposition Models Investigated: We searched 25 instances in total obtained with different hyperparameter settings from three families of decomposition. The first family",
+        "bbox": [
+            498,
+            854,
+            893,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "4234",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/b5252d7045a48fc04cdc2dbabcad7e250e0ce1d2b8033904dcbac07430850133.jpg",
+        "image_caption": [
+            "Figure 3. Each model from StyLitGAN produces 16 directions. Models differ by choice of hyperparameters and intrinsic image decomposition. We evaluate models by albedo change and by diversity, averaged across a small fixed validation set of test scenes. As the figure shows, there is typically a payoff, but some models are not admissible. Figure 4 shows examples from some of the models considered here. We exclude inadmissible models, then pool all directions from all other models, and apply a forward selection procedure (section 4). This yields 16 strong relighting directions (the gold star). In our comparison with Yang et al. [48], we focus on the changes in albedo since their method identifies only a single relighting direction. This limits the scope for evaluating relighting diversity. Yang et al. [48] aggressively change scene albedo, while our StyLitGAN ensures only lighting changes. Additionally, we compare with GAN-control [39], which, while attempting relighting, often changes the scene layout, leading to large albedo change and increased diversity score due to layout variations."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            120,
+            90,
+            421,
+            321
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "is the SOTA unsupervised model of [14], which decomposes images into albedo and shading using example images drawn from statistical models. The second is a variant of that family that decomposes into albedo, shading, and gloss decomposition [3]. The third is an albedo, shading, and gloss decomposition that models fine edges in the albedo rather than the shading field. These models were chosen to represent a range of possible decompositions, but others could yield better results. The key point is that we can choose a model from a collection by a rational process.",
+        "bbox": [
+            75,
+            571,
+            470,
+            723
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Selecting Directions: Our approach for selecting directions involves creating a scatter plot of 25 instances with various hyperparameters and image decomposition models. We find 16 directions for each instance in our final experiments, and the search for 16 directions takes about 14 minutes on an A40 GPU. We experimented with different numbers of directions of order  $2^{n}$  for  $n = 2, 3, 4, 5, 6, 7$  and found that 16 directions  $(n = 4)$  strike a better balance between relighting diversity and albedo change. However, finding multiple directions is challenging because the search space is complex and high-dimensional, and we lack ground truth to supervise",
+        "bbox": [
+            75,
+            734,
+            472,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/7f5e3de54231b4937aaace146c50da9cea0c22c9600791c8d69cdae838a7e777.jpg",
+        "image_caption": [
+            "Figure 4. The bottom row shows scene relightings obtained using our final, forward selected, set of directions (star of Figure 3). For comparison, we also show scene relightings from different models obtained from StyLitGAN shown in Figure 3 (model numbers correspond to numbers on that figure). Note how most models are capable of producing some good directions, but not all directions from a given model may be good."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            88,
+            879,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "the search. Therefore, we apply a two-step process to find effective directions and filter out any bad directions.",
+        "bbox": [
+            496,
+            372,
+            890,
+            402
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We first identify and discard inadmissible models that are located behind the Pareto frontier. We then select the top 10 admissible models based on their average albedo change when applied over a large set of fixed validation images. Our goal is to select the best relighting directions from these admissible models. To achieve this, a forward selection process is employed, which involves selecting a subset of directions from the set of admissible directions.",
+        "bbox": [
+            496,
+            402,
+            892,
+            523
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Forward Selection Process: To select the best directions, we begin with all directions from the admissible models, resulting in 160 directions from 10 models. These directions are then filtered to remove \"bad\" directions that produce relighting similar to the original image or shading that does not vary across pixels, resulting in 108 directions.",
+        "bbox": [
+            496,
+            529,
+            893,
+            619
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Next, we use a greedy process to select the best 16 directions from the remaining 108 directions. We evaluate each direction one at a time and add it to the pool if it provides a large diversity score while incurring a small penalty for large albedo change. This process continues until the desired number of directions are selected. The forward selection process is fast and efficient, taking less than a minute.",
+        "bbox": [
+            496,
+            619,
+            893,
+            726
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The resulting scores from the forward select 16 directions are marked with a star in golden color in Fig. 3. The directions obtained with this process are significantly better than individual models alone. A qualitative ablation is in Fig. 4.",
+        "bbox": [
+            496,
+            726,
+            893,
+            786
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            800,
+            632,
+            816
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Qualitative Evaluation: StyLitGAN produces realistic images that are out of distribution but known to exist for straightforward physical reasons. Because they're out of distribution, current quantitative evaluation tools do not apply. We evaluate realism qualitatively. Further, there is no direct",
+        "bbox": [
+            496,
+            824,
+            893,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "4235",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/7166d8d77f767b023bc10ba80437020a9bbcd67714d16b616b5c6ea92a4fd88f.jpg",
+        "image_caption": [
+            "Generated Image",
+            "Relit-1  $(+d_{1})$",
+            "Relit-2  $(+d_{2})$",
+            "Relit-3  $(+d_{3})$",
+            "Relit-4  $(+d_{4})$",
+            "Relit-5  $(+d_{5})$",
+            "Relit-6  $(+d_{6})$",
+            "Relit-7  $(+d_{7})$",
+            "Figure 5. First column images generated by the original StyleGAN. Other columns show images obtained from  $\\mathbf{w}^{+} + \\mathbf{d}_{i}$ , our relighting directions added to the style codes  $(\\mathbf{w}^{+})$  of the images in the first column. These directions have been chosen to fix albedo, but change shading. Note: each row shows the same scenes but with different illumination. Lighting varies aggressively, and the individual latent direction  $\\mathbf{d}_{i}$  has persistent semantics - each column corresponds to a type of illumination. For example, the second  $(d_{1})$  and third  $(d_{2})$  column switches on bedside lamps. It is worth noting the presence of soft shadows, cast shadows, inter-reflections, and glossy changes across relights. Another important observation to note is that all relightings are with respect to world coordinates, not camera coordinates."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            99,
+            893,
+            489
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d1bc21dc64fd1c6702e3e66077b3aaa869db1ebc164fc3a4c97c19c1a94e7b06.jpg",
+        "image_caption": [
+            "Figure 6. Resurfacing Generated Images. Instead of relighting images, we can generate resurfaced images by swapping our consistency and diversity loss. We apply diversity loss to change the albedo and consistency loss to maintain the shading and global illumination. The first column shows images generated by the original StyleGAN, and the other columns show images obtained from  $\\mathbf{w}^{+} + \\mathbf{d}_{i}$  for our resurfacing or recoloring directions. Each column shows the same scene as in the first column, but with varying surface colors and materials, while the individual latent direction  $\\mathbf{d}_i$  retains its semantics."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            580,
+            870,
+            743
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "comparable method. However, we show relighting comparisons to a recent SOTA method that is physically motivated and trained with CGI data [26] by using a SOTA inversion method [5] (Fig. 14 in Supplementary). For relighting, our",
+        "bbox": [
+            75,
+            829,
+            472,
+            891
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "method should generate images that: are clearly relightings of a scene; fix geometry and albedo but visibly change shading; and display complicated illumination effects, including soft shadows, cast shadows and gloss. For resurfacing, our",
+        "bbox": [
+            498,
+            829,
+            893,
+            891
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "4236",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/2fd4b9a12e44e0c1910ee4e27fdfba0f92dd8940b5ac0a140ab0b22371286c64.jpg",
+        "image_caption": [
+            "Figure 7. Controllability: The first and last columns of the figure show relighted images generated using our Relit-1 and Relit-5 directions. The bottom section of the figure features a user-controllable slider that enables adjusting the weight of the relighting effects produced by these two directions. Moving the slider from left to right results in a seamless interpolation between the two lighting directions and provides precise control over the relighting of the generated images."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            99,
+            89,
+            870,
+            255
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/ea7eef6570ef1ec9bd1a3fc1856d73f98a7de1090b9b33827b3fb51fa5f55fae.jpg",
+        "image_caption": [
+            "Figure 8. Scaling Directions: The figure depicts the persistent and smooth effects of applying the direction at different scalar coefficients. We use our Relit-2 direction. A slider at the bottom allows the user to adjust the weight of the relighting direction, producing a seamless interpolation when increasing or decreasing the intensity of the chosen direction. The relighting effects range from a well-lit room with the bedside lamp off to weak external lighting with a bedside lamp on."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            319,
+            875,
+            486
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "method should generate images that: are clearly images of the original layout, but with different materials or colors or changes in furniture; and display illumination effects that are consistent with these color changes. As Figure 6 shows, our method meets these goals. Figure 7 and Figure 8 show interpolation sequences for a relighting between two directions and scaling only one direction. Note that the lighting changes smoothly, as one would expect. Figure 15 in Supplementary shows that our relighting and recoloring directions are largely disentangled. An ablation for individual losses is also given in Supplementary Figure 13.",
+        "bbox": [
+            75,
+            560,
+            472,
+            726
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Quantitative Evaluation: Figure 3 shows how we can evaluate albedo and lighting shifts. In Table 1 we show we can generate image datasets with increased FID [18, 34] (clean-FID) from the base comparison set. This is strong evidence our method can produce a set of images that is a strict superset of those that the vanilla StyleGAN can produce.",
+        "bbox": [
+            75,
+            727,
+            470,
+            816
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Generality: We have applied our method to StyleGAN trained on Conference Room, Kitchen, Living Room, Dining Room, Church, and Face datasets (results in Figure 10).",
+        "bbox": [
+            75,
+            816,
+            470,
+            863
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Comparison to GAN-control: GAN-control (GC) [39] represents lighting with a spherical harmonic predictor pre",
+        "bbox": [
+            75,
+            869,
+            472,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/53cbc0b29954b52114849d5ddae759d90d3ec9e264e16a1f6efdee545ce62eb0.jpg",
+        "table_caption": [
+            "Table 1. FID measures distribution shift and not realism. Our generated images are realistic and are out-of-distribution because of large illumination and color changes in the images. This results in large FID scores.  $KDL$  in the table is for kitchen, dining and living room which are jointly trained [13]."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Type</td><td>Bedroom</td><td>KDL</td><td>Conference</td><td>Church</td><td>Faces</td></tr><tr><td>StyleGAN (SG)</td><td>5.01</td><td>5.86</td><td>9.35</td><td>3.80</td><td>5.02</td></tr><tr><td>SG + Relighting (RL)</td><td>14.23</td><td>6.87</td><td>10.48</td><td>12.12</td><td>37.87</td></tr><tr><td>SG + Resurfacing (RS)</td><td>17.03</td><td>9.41</td><td>10.63</td><td>18.60</td><td>34.06</td></tr><tr><td>SG + RL + RS</td><td>21.39</td><td>11.68</td><td>12.71</td><td>21.08</td><td>37.40</td></tr></table>",
+        "bbox": [
+            501,
+            618,
+            888,
+            686
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "trained on a parameterized 3D face reconstruction model [9]. This model does not apply to indoor lighting (among other problems, it predicts all points with the same normal have the same shading). We trained a GC model on the LSUN Bedroom dataset with 2 subspaces  $z^k$  - illumination corresponding to spherical harmonic coefficients, and other to represent all other structural information. The model was trained for 800 epochs. GC produces images whose structure varies wildly with any lighting changes, resulting in large albedo changes (GC in Figures 3, 9). The difficulty appears to be that the attribute predictor is easily subverted; if the lighting representation cannot produce an image that is (say) dark on the left side, the albedo is adjusted instead.",
+        "bbox": [
+            496,
+            704,
+            893,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "4237",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/616e562904dbfc4791c0582ec40a42c8d2ed3c14537c46285ea6b0cee9bcadd9.jpg",
+        "image_caption": [
+            "Figure 9. GAN-control (GC) [39] cannot relight complex scenes like bedrooms; it completely changes the scene's layout."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            88,
+            450,
+            152
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/136a1a252d87a1153f2c7be876b8d88a7fcd64fcff843b5203840378e6914d27.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            194,
+            267,
+            330
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/4ee88bae8989c9c8a895b5d15a0e25021316baa5246f7afedb1db4702bcb771b.jpg",
+        "image_caption": [
+            "(c) Living Room Relighting"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            88,
+            333,
+            264,
+            483
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/41ebdd2d30f388edfd89eb5765d19b1696df76dcf09f9c25c465b7abb7791c2c.jpg",
+        "image_caption": [
+            "(e) Relighting outdoor Churches",
+            "Figure 10. StyLitGAN extends to finding relighting directions for StyleGANs trained on other datasets."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            497,
+            267,
+            633
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/345544a3e5e24f28b193a84fae75a982292eb0e373c4b94c3494e93f648e7955.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            281,
+            194,
+            460,
+            330
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/fdeb4c976808fe9c8c6d5bfed230d28b807fe438aab779b4a93459b9a4805f28.jpg",
+        "image_caption": [
+            "(b) Kitchen Relighting"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            281,
+            345,
+            459,
+            482
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/86a2438cb584c617b683c965ca94228ad9a05d7ece2cf54a21e1e3ca5fad0ccf.jpg",
+        "image_caption": [
+            "(d) Dining Room Relighting",
+            "(f) Portrait Relighting"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            281,
+            497,
+            460,
+            633
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Downstream Applications",
+        "text_level": 1,
+        "bbox": [
+            76,
+            691,
+            318,
+            709
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Lighting Variance in Surface Normal Prediction: The Multilum dataset [29] provides images of 1000 various indoor scenes, each under 25 lighting conditions, physically relit and captured. A SOTA normal predictor (Omnidata [12, 21]) applied to the test set produces surface normal predictions that vary significantly with changes of light in a fixed scene (Figure 11, purple bar). Finetuning Omnidata with the Multilum training dataset significantly reduces this variance (Figure 11, orange bar); but multiple lightings of a fixed scene are very hard to find. Finetuning with StyLitGAN relights produces comparable improvements; using seven distinct relights (Figure 11, pale orange bar) is slightly",
+        "bbox": [
+            75,
+            719,
+            472,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b18315a4c356fb926ef337ffe428642de1853a27d7aba075bb8b6ab8715c80e0.jpg",
+        "image_caption": [
+            "Figure 11. Normal variance to lighting reduces when fine-tuning on our relighting dataset. Purple boxplot shows normal variance under relighting for real test scenes from Multilum using the Omnidata normal predictor; orange shows the result of finetuning using Multilum training data; light orange and yellow show the result of finetuning using StyLitGAN images (7 and 25 per scene respectively). The measure is angular error in radians from the mean prediction of a scene for each relit image in the Multilum test set (30 scenes, 25 lightings each)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            550,
+            78,
+            836,
+            272
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/648cf2ffc644a4acbef8476ef546b65980865d6af2bf979b6e9b5ecb93fd8ac4.jpg",
+        "image_caption": [
+            "Figure 12. A surface normal predictor is finetuned to increase prediction consistency for relit images of the same scene. General performance shown on Taskonomy test images parallels that of the original model. Mean across buildings is given in the last column."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            542,
+            410,
+            846,
+            585
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "worse than using 25 distinct relights (Figure 11, yellow bar). The resulting improvement is not at the cost of base accuracy. Figure 12 compares the accuracy of various finetuned models on the Taskonomy test set [50] (recall this involves thousands of frames each in 10 blocks; we show results by block). Note that finetuned methods mostly show slight accuracy improvements over the base model, but losses in some blocks result in means that match.",
+        "bbox": [
+            496,
+            654,
+            893,
+            773
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In our follow-up work, we also show that StyleGAN \"knows\" intrinsic images and can be easily extracted [4].",
+        "bbox": [
+            496,
+            773,
+            893,
+            806
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgment",
+        "text_level": 1,
+        "bbox": [
+            500,
+            818,
+            651,
+            835
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We thank Aniruddha Kembhavi, Derek Hoiem, Min Jin Chong and Shenlong Wang. This material is based upon work supported by the National Science Foundation under Grant No. 2106825 and by a gift from Boeing Corporation.",
+        "bbox": [
+            496,
+            839,
+            893,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "4238",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            174,
+            106
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2",
+            "[2] Peter N Belhumeur and David J Kriegman. What is the set of images of an object under all possible illumination conditions? International Journal of Computer Vision, 1998. 4",
+            "[3] Anand Bhattachad and D.A. Forsyth. Cut-and-paste object insertion by enabling deep image prior for reshading. In 2022 International Conference on 3D Vision (3DV). IEEE, 2022. 2, 3, 5",
+            "[4] Anand Bhattachad, Daniel McKee, Derek Hoiem, and David Forsyth. Stylegan knows normal, depth, albedo, and more. Advances in Neural Information Processing Systems, 36, 2024. 8",
+            "[5] Anand Bhattachad, Viraj Shah, Derek Hoiem, and David A Forsyth. Make it so: Steering stylegan for any image inversion and editing. arXiv preprint arXiv:2304.14403, 2023. 6, 1",
+            "[6] Prashanth Chandran, Sebastian Winberg, Gaspard Zoss, Jérémy Riviere, Markus Gross, Paulo Gotardo, and Derek Bradley. Rendering with style: combining traditional and neural approaches for high-quality face rendering. ACM Transactions on Graphics (ToG), 40(6):1-14, 2021. 2",
+            "[7] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. arXiv preprint arXiv:2112.11641, 2021. 3",
+            "[8] Min Jin Chong, Hsin-Ying Lee, and David Forsyth. Stylegan of all trades: Image manipulation with only pretrained stylegan. arXiv preprint arXiv:2111.01619, 2021. 2",
+            "[9] 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.",
+            "[10] Aditya Deshpande, Jiajun Lu, Mao-Chuang Yeh, Min Jin Chong, and David Forsyth. Learning diverse image colorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6837-6845, 2017. 2",
+            "[11] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 341-346, 2001. 2",
+            "[12] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 8",
+            "[13] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A Efros. Blobgan: Spatially disentangled scene representations. arXiv preprint arXiv:2205.02837, 2022. 2, 3, 7",
+            "[14] D.A. Forsyth and Jason J Rock. Intrinsic image decomposition using paradigms. TPAMI, 2022 in press. 2, 3, 5",
+            "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and"
+        ],
+        "bbox": [
+            78,
+            114,
+            472,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "pattern recognition, pages 2414-2423, 2016. 2",
+            "[16] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. arXiv preprint arXiv:1406.2661, 2014. 2",
+            "[17] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 327-340, 2001. 2",
+            "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2017. 7",
+            "[19] Le Hui, Jia Yuan, Mingmei Cheng, Jin Xie, Xiaoya Zhang, and Jian Yang. Superpoint network for point cloud oversegmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5510-5519, 2021. 4",
+            "[20] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, 2016. 3",
+            "[21] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 2, 8",
+            "[22] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2",
+            "[23] 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, 2019.",
+            "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 1, 2",
+            "[25] Edwin H Land. The retina theory of color vision. Scientific american, 1977. 3",
+            "[26] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. arXiv preprint arXiv:2205.09343, 2022. 6, 1",
+            "[27] Zicheng Liao, Hugues Hoppe, David Forsyth, and Yizhou Yu. A subdivision-based representation for vector image editing. IEEE transactions on visualization and computer graphics, 2012. 2",
+            "[28] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. Advances in Neural Information Processing Systems, 34, 2021. 2",
+            "[29] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A multi-illumination dataset of indoor object appearance. In 2019 IEEE International Conference on Computer Vision (ICCV), Oct 2019. 2, 8",
+            "[30] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, Epic Games, Andreas Lehrmann, and AI Borealis. Learning physics-guided face relighting under directional light. 2020. 2"
+        ],
+        "bbox": [
+            503,
+            93,
+            893,
+            888
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "4239",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[31] Xingang Pan, Xudong Xu, Chen Change Loy, Christian Theobalt, and Bo Dai. A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2",
+            "[32] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2",
+            "[33] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2",
+            "[34] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 7",
+            "[35] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. IEEE Computer graphics and applications, 21(5):34-41, 2001. 2",
+            "[36] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 2",
+            "[37] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2",
+            "[38] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2",
+            "[39] Alon Shoshan, Nadav Bhonker, Igor Kviatkovsky, and Gerard Medioni. Gan-control: Explicitly controllable gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14083-14093, 2021. 2, 5, 7, 8",
+            "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. *ICLR*, 2015. 3",
+            "[41] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics, 2019. 2",
+            "[42] Feitong Tan, Sean Fanello, Abhimitra Meka, Sergio Orts-Escolano, Danhang Tang, Rohit Pandey, Jonathan Taylor, Ping Tan, and Yinda Zhang. Volux-gan: A generative model for 3d face synthesis with hdri relighting. arXiv preprint arXiv:2201.04873, 2022. 2",
+            "[43] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 2",
+            "[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2",
+            "[45] Andrey Voynov and Artem Babenko. Unsupervised discovery"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            891
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "of interpretable directions in the gan latent space. In International conference on machine learning, pages 9786-9796. PMLR, 2020. 1, 2",
+            "[46] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 2",
+            "[47] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. arXiv preprint arXiv: 2101.05278, 2021. 2",
+            "[48] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. International Journal of Computer Vision, 2020. 2, 5",
+            "[49] Ning Yu, Guilin Liu, Aysegul Dundar, Andrew Tao, Bryan Catanzaro, Larry S Davis, and Mario Fritz. Dual contrastive loss and attention for gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6731-6742, 2021. 3",
+            "[50] Amir R. Zamir, Alexander Sax, William B. Shen, Leonidas J. Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2018. 8",
+            "[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3",
+            "[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 2",
+            "[53] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 1, 2",
+            "[54] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, 2017. 2"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            633
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "4240",
+        "bbox": [
+            482,
+            944,
+            516,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2024/StyLitGAN_ Image-Based Relighting via Latent Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_model.json

@@ -0,0 +1,2112 @@
+[
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.107,
+                0.003,
+                0.182,
+                0.043
+            ],
+            "angle": 0,
+            "content": "CVF"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.237,
+                0.0,
+                0.812,
+                0.045
+            ],
+            "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": [
+                0.204,
+                0.131,
+                0.768,
+                0.154
+            ],
+            "angle": 0,
+            "content": "StyLitGAN: Image-based Relighting via Latent Control"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.275,
+                0.182,
+                0.696,
+                0.199
+            ],
+            "angle": 0,
+            "content": "Anand Bhattad James Soole D.A. Forsyth"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.323,
+                0.2,
+                0.649,
+                0.218
+            ],
+            "angle": 0,
+            "content": "University of Illinois Urbana-Champaign"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.243,
+                0.218,
+                0.726,
+                0.234
+            ],
+            "angle": 0,
+            "content": "https://anandbhattad.github.io/stylitgan/"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.135,
+                0.256,
+                0.228,
+                0.267
+            ],
+            "angle": 0,
+            "content": "Generated Image"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.246,
+                0.256,
+                0.842,
+                0.269
+            ],
+            "angle": 0,
+            "content": "Relit-1 \\((\\mathbf{w}^{+} + d_{1})\\) Relit-2 \\((\\mathbf{w}^{+} + d_{2})\\) Relit-3 \\((\\mathbf{w}^{+} + d_{3})\\) Relit-4 \\((\\mathbf{w}^{+} + d_{4})\\) Relit-5 \\((\\mathbf{w}^{+} + d_{5})\\)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.12,
+                0.268,
+                0.848,
+                0.55
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.075,
+                0.552,
+                0.895,
+                0.651
+            ],
+            "angle": 0,
+            "content": "Figure 1. StyLitGAN identifies directional vectors \\((d_i)\\) within StyleGAN's style space \\((\\mathcal{W}^{+})\\) which, when added to the \\(\\mathbf{w}^{+}\\)style code, effectively modify the lighting of generated images while preserving their geometry and albedo. This process eliminates the need for per-image search or model fine-tuning. The first column displays images generated from StyleGAN2; subsequent columns illustrate the same scene, each relit using a specific direction. These relighting directions \\((d_i)\\) are derived through a forward selection method, ensuring diversity and avoiding cherry-picking. The directional effects are consistent across different scenes: for instance, \\(d_1\\) activates an orange-tinged bedside lamp, \\(d_2\\) a less intense white-tinged lamp, \\(d_3\\) introduces strong directional light from the window, and so on, demonstrating diverse relighting capabilities of StyLitGAN."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.235,
+                0.662,
+                0.314,
+                0.679
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.075,
+                0.689,
+                0.475,
+                0.903
+            ],
+            "angle": 0,
+            "content": "We describe a novel method, StyLitGAN, for relighting and resurfacing images in the absence of labeled data. StyLitGAN generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects, without the need for paired or CGI data. StyLitGAN uses an intrinsic image method to decompose an image, followed by a search of the latent space of a pretrained StyleGAN to identify a set of directions. By prompting the model to fix one component (e.g., albedo) and vary another (e.g., shading), we generate relighted images by adding the identified directions to the latent style codes. Quantitative metrics of change in albedo and lighting diversity allow us to choose effective directions using a forward selection process. Qualitative evaluation confirms the effectiveness of our method."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.502,
+                0.664,
+                0.631,
+                0.68
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.689,
+                0.895,
+                0.825
+            ],
+            "angle": 0,
+            "content": "Scene appearance shifts dramatically with varying lighting conditions - a sunlit room takes on a different character as daylight fades, and interior spaces transform with the flick of a switch. Similarly, surface changes, like a wall's paint color, change not only the wall's appearance but also the overall image due to light reflection. Despite the impressive realism achieved by current generative models like StyleGAN [22-24], they fall short in dynamically controlling scene lighting, a key aspect of realistic image generation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.825,
+                0.895,
+                0.903
+            ],
+            "angle": 0,
+            "content": "In this work, we present StyLitGAN, a novel approach that extends the editing capabilities of StyleGAN [38, 45, 46, 53]. StyLitGAN uniquely manipulates style codes to selectively change lighting while preserving other image attributes like albedo and geometry. This selective editing"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.514,
+                0.957
+            ],
+            "angle": 0,
+            "content": "4231"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.092,
+                0.471,
+                0.137
+            ],
+            "angle": 0,
+            "content": "addresses a critical gap in current generative methods, which typically lack the precision to control individual scene components independently."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.137,
+                0.47,
+                0.272
+            ],
+            "angle": 0,
+            "content": "Our method, StyLitGAN, first uses StyleGAN to produce a set of images and then decomposes these generated images into albedo, diffuse shading, and glossy effects using an off-the-shelf, self-supervised network [14]. We then search for style code edits by prompting StyleGAN to produce images that (a) are diverse, but (b) have the same albedo (and so geometry and material) as the original generated images. Our search selects the most effective relighting directions in a data-driven manner."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.273,
+                0.471,
+                0.439
+            ],
+            "angle": 0,
+            "content": "Our approach generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects. Importantly, we observe that style code edits produce consistent effects across images. For instance, as seen in Fig. 1, adding the first and second directions tends to switch on bedside lamps (columns Relit-1 and Relit-2), while adding the fourth direction increases the light intensity from outside the window (column Relit-4). Since StyLitGAN can generate any image that a vanilla StyleGAN can, it also generate images that are out of distribution, one would expect FID scores to increase over StyleGAN; this happens."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.439,
+                0.47,
+                0.514
+            ],
+            "angle": 0,
+            "content": "The recent method GAN-control [39] controls lighting on face images using an attribute procedure. We show, in contrast to StyLitGAN, GAN-control fails on indoor scenes, likely because the attribute vocabulary is too easily subverted by the complex lighting effects in indoor scenes."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.515,
+                0.471,
+                0.636
+            ],
+            "angle": 0,
+            "content": "We demonstrate applications of StyLitGAN to standard vision problems. Using the Multi-Ilum dataset of [29], we show that predictions from a SOTA surface normal predictor [21] vary significantly when lighting is changed. Finetuning this normal predictor using StyLitGAN images suppresses this effect. The improvement is comparable with that obtained by finetuning with true multi-illuminant images (which are very difficult to obtain in quantity)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.649,
+                0.219,
+                0.665
+            ],
+            "angle": 0,
+            "content": "2. Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.674,
+                0.471,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Image Manipulation: A significant literature deals with manipulating and editing images [3, 10, 11, 15, 17, 27, 33, 35, 44, 54]. Editing procedures for generative image models [16] are important, because they demand compact image representations with useful, disentangled interpretations. StyleGAN [22-24] is currently de facto state-of-the-art for editing generated images, likely because its mapping of initial noise vectors to style codes which control entire feature layers produces latent spaces that are heavily disentangled and so easy to manipulate. Recent editing methods include [8, 36, 38, 45, 46, 53], with a survey in [47]. The architecture can be adapted to incorporate spatial priors for authoring novel and edited images [13, 28, 43]. In contrast to this literature, we show how to fix one physically meaningful image factor while changing another. Doing so is difficult"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.092,
+                0.894,
+                0.303
+            ],
+            "angle": 0,
+            "content": "because the latent spaces are not perfectly disentangled, and we must produce a diverse set of changes in only one factor. Relighting using StyleGAN: Relighting faces using StyleGAN can be achieved with Stylerig [43], but this method requires a 3D morphable face model. In contrast, StyLitGAN does not require a 3D model and can be extended to complex indoor scenes, which is not possible with Stylerig. Yang et al. [48] uses semantic label attributes to train a binary classifier to find latent space directions that represent indoor and natural lighting, but this method cannot produce diverse relighting effects. We also find Yang et al's relighting to change color or albedo. In contrast, StyLitGAN generates diverse realistic relighting effects without changing albedo and without requiring any labeled attributes."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.303,
+                0.894,
+                0.529
+            ],
+            "angle": 0,
+            "content": "StyleFlow [1] and GAN-control [39] require a parametric model to express lighting, such as spherical harmonics. These methods are limited to relighting faces and do not result in realistic relighting of rooms. Our experiments using GAN-control for rooms result in large geometry and albedo change. In contrast, StyLitGAN can produce relighted images without changing geometry or albedo. We also note that rooms are more challenging to relight than faces due to significant long-scale inter-reflection effects, diverse shadow patterns, stylized luminaires, stylized surface albedos, and surface brightnesses that are not a function of surface normal alone. These factors make it difficult to apply GAN control directly to rooms. Also, none of these methods can resurface or recolor rooms, though StyLitGAN can also edit color or materials while preserving the scene's lighting."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.53,
+                0.894,
+                0.681
+            ],
+            "angle": 0,
+            "content": "Other Face Relighting methods use carefully collected supervisory data from light-stages or parametric spherical harmonics [30, 32, 37, 41, 52]. ShadeGAN [31], Rendering with Style [6], and Volux-GAN [42] use a volumetric rendering approach to learn the 3D structure of the face and the illumination encoding. Volux-GAN [42] also requires image decomposition from [32] that is trained using carefully curated light-stage data. In comparison, we neither require any explicit 3D modeling of the scene nor labeled and curated data for training the image decomposition model."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.695,
+                0.609,
+                0.712
+            ],
+            "angle": 0,
+            "content": "3. Approach"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.72,
+                0.894,
+                0.885
+            ],
+            "angle": 0,
+            "content": "We follow convention and manipulate StyleGAN [24] by adjusting the \\(\\mathbf{w}^{+}\\) latent variables. We do not modify StyleGAN weights, but instead, seek a set of lighting directions \\(\\mathbf{d}_i\\) (same shape as \\(\\mathbf{w}^{+}\\)) which are independent of \\(\\mathbf{w}^{+}\\) and have desired effects on the generated image. We obtain these directions by constructing losses that capture the desired outcomes, and then search for directions that minimize these losses. We find all directions only once and use 2000 randomly generated images for this search. Once found, these lighting directions apply to all other generated images. Our search procedure only sees each image once."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.518,
+                0.886,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Fig. 2 summarizes our procedure and we call our model"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "4232"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.118,
+                0.088,
+                0.845,
+                0.253
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.26,
+                0.892,
+                0.345
+            ],
+            "angle": 0,
+            "content": "Figure 2. How StyLitGAN works: We generate an image from random Gaussian noise using a pretrained StyleGAN. We also generate novel relighted versions (16 in our case) of the same image using randomly initialized latent directions (d) that are added to \\(\\mathbf{w}^{+}\\)latent style codes. We train a classifier (F) that takes in all the pairs of relighted and original images and predicts the relighting direction applied to them. We apply a distinction loss and jointly update the latent directions and the classifier. Next, we generate the decomposition of these images from a pretrained decomposition model (D). We then apply losses that force StyLitGAN to find latent directions such that the albedo does not change (consistency loss), but the image does (diversity loss)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.351,
+                0.473,
+                0.576
+            ],
+            "angle": 0,
+            "content": "StyLitGAN. Our method consists of two stages. The first stage involves decomposing images using a pretrained model. The second stage jointly searches for directions and trains a classifier \\( \\mathrm{F} \\). The classifier \\( \\mathrm{F} \\) predicts the latent direction applied to image pairs. It's a classification task with a fixed number of jointly learned latent directions. While classifier \\( \\mathrm{F} \\) doesn't directly know whether latent directions relate to lighting, our use of image decomposition losses ensures these directions are lighting-related. The consistency loss maintains albedo, and the diversity loss ensures diverse shading changes, both properties expected when changing lighting conditions. Thus, while jointly updating the classifier \\( \\mathrm{F} \\) and latent directions, these losses ensure the discovered directions pertain to relighting. We now elaborate on our search for directions and losses in detail."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.58,
+                0.472,
+                0.656
+            ],
+            "angle": 0,
+            "content": "Base StyleGAN Models: We use baseline pretrained models from [49] that use a dual-contrastive loss to train StyleGAN for bedrooms, faces, and churches. We also use baseline pretrained StyleGAN2 models from [13] for conference rooms, kitchens, dining rooms, and living rooms."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.657,
+                0.472,
+                0.869
+            ],
+            "angle": 0,
+            "content": "Decomposition: We decompose images into albedo, shading, and gloss maps (gloss only when available) as \\( A \\times S + G \\), where \\( A \\) models albedo effects and \\( S \\) and \\( G \\) model shading and gloss effects respectively. We use the method of Forsyth and Rock [14] and its variant from Bhattad and Forsyth [3], which is easily adapted because it is self-supervised and uses only samples from statistical models derived from Land's Retinex theory [25]. By changing the statistical spatial model parameters, we can construct many decompositions using their approach. We evaluate many such decomposition models under several hyperparameter settings and create a large pool of relighting directions. We finalize our directions using a forward selection process that provides minimal albedo and geometry shift with a large relighting diversity (Section 4)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.871,
+                0.472,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Relighting a scene should produce a new, realistic image where the shading has changed but the albedo has"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.351,
+                0.895,
+                0.578
+            ],
+            "angle": 0,
+            "content": "not. Write \\( I(\\mathbf{w}^{+}) \\) for the image produced by StyleGAN given style codes \\( \\mathbf{w}^{+} \\), and \\( A(I) \\), \\( S(I) \\), and \\( G(I) \\) for the albedo, shading, and gloss respectively recovered from image \\( I \\). We search for multiple directions \\( \\mathbf{d}_i \\) such that: (a) \\( A(I(\\mathbf{w}^{+} + \\mathbf{d}_i)) \\) is very close to \\( A(I(\\mathbf{w}^{+})) \\) – so the image is a relighted version of \\( I(\\mathbf{w}^{+}) \\), a property we call persistent consistency; (b) the images produced by the different directions are linearly independent – relighting diversity; (c) which direction was used can be determined from the image, so that different directions have visibly distinct effects – distinctive relighting; and (d) the new shading field (map) is not strongly correlated to the albedo – independent relighting. Not every shading field can be paired with a given albedo, otherwise there would be nothing to do. We assume that edited \\( \\mathbf{w}^{+} \\) will result in realistic images [7]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.578,
+                0.894,
+                0.654
+            ],
+            "angle": 0,
+            "content": "Recoloring: Alternatively, we may wish to edit scenes where the colors or materials of objects have changed, but the lighting hasn't. Because shading conveys a great deal of information about shape, we can find these edits using modified losses by seeking consistency in the shading field."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.658,
+                0.895,
+                0.779
+            ],
+            "angle": 0,
+            "content": "Persistent Consistency: The albedo decomposition of both the relighted scene: \\( A_{R} = A(I(\\mathbf{w}^{+} + \\mathbf{d}_{i})) \\) and the original: \\( A_{O} = A(I(\\mathbf{w}^{+})) \\) must be the same; where R refers to relighted images and O refers to StyleGAN generated images. We use a Huber loss and a perceptual feature loss [20, 51] from a VGG feature extractor (Φ) [40] at various feature layers (j) to preserve persistent effects (geometry, appearance, and texture) in the scene."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.5,
+                0.788,
+                0.891,
+                0.851
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n s t}} \\left(A _ {O}, A _ {R}\\right) = \\left\\{ \\begin{array}{c} \\frac {1}{2} \\left[ A _ {O} - A _ {R} \\right] ^ {2} \\quad \\text {f o r} | A _ {O} - A _ {R} | \\leq \\delta , \\\\ \\delta \\left(| A _ {O} - A _ {R} | - \\delta / 2\\right) \\quad \\text {o t h e r w i s e .} \\end{array} \\right. (1) \\\\ \\mathcal {L} _ {p e r} \\left(A _ {O}, A _ {R}\\right) = \\left| \\left| \\Phi_ {j} \\left(A _ {O}\\right) - \\Phi_ {j} \\left(A _ {R}\\right) \\right| \\right| _ {2}. (2) \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.856,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Relighting Diversity: We want the set of relighted images produced by the directions to be diverse on a long scale so that regions that were in shadow in one image might be"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "4233"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.091,
+                0.47,
+                0.168
+            ],
+            "angle": 0,
+            "content": "bright in another. For each \\( S(\\mathbf{w}^{+} + \\mathbf{d}_{i}) \\), we stack the two shading and gloss: \\( S \\) and \\( G \\), and compute a smoothed and downsampled vector \\( \\mathbf{t}_{i} \\) from these maps. We then compute \\( \\mathcal{L}_{div}(S,G) \\) (diversity loss) which compels these \\( \\mathbf{t}_{i} \\) to be linearly independent and encourages diversity in relighting."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.19,
+                0.177,
+                0.469,
+                0.192
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\text {d i v}} (S, G) = - \\log \\det  N \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.097,
+                0.201,
+                0.371,
+                0.218
+            ],
+            "angle": 0,
+            "content": "where \\(i^{th}\\) & \\(j^{th}\\) component of \\(\\mathbf{N}\\) is \\(t_i^\\top t_j\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.222,
+                0.47,
+                0.329
+            ],
+            "angle": 0,
+            "content": "Distinctive Relighting: A network might try to cheat by making minimal changes to the image. Directions \\(\\mathbf{d}_i\\) should have the property that \\(\\mathbf{d}_i\\) is easy to impute from \\(I(\\mathbf{w}^{+} + \\mathbf{d}_{i})\\). Joint with the search for directions, we train a classifier to categorize the applied direction. This classifier accepts \\(I(\\mathbf{w}^{+})\\) and \\(I(\\mathbf{w}^{+} + \\mathbf{d}_{i})\\) and must predict \\(i\\). Its cross-entropy supplies our loss:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.153,
+                0.338,
+                0.469,
+                0.4
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\min  _ {l, F} \\mathcal {L} _ {\\text {d i s t}} (I (\\mathbf {w} ^ {+}), I (\\mathbf {w} ^ {+} + d _ {i})) \\\\ = - \\sum_ {i = 1} ^ {M} y _ {i} \\log F \\left(I \\left(\\mathbf {w} ^ {+}\\right), I \\left(\\mathbf {w} ^ {+} + d _ {i}\\right)\\right) \\tag {4} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.405,
+                0.47,
+                0.466
+            ],
+            "angle": 0,
+            "content": "Saturation Penalty: Our diversity loss might cheat and obtain high diversity by generating blocks of over-saturated or under-saturated pixels. So, we apply a saturation penalty over several pixels within a certain threshold."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.089,
+                0.477,
+                0.469,
+                0.564
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {s a t} = \\lambda_ {\\text {o v e r s a t}} \\left[ \\frac {1}{H \\times W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\max  (0, I _ {i, j} - s) ^ {2} \\right] \\tag {5} \\\\ + \\lambda_ {\\text {u n d e r s a t}} \\left[ \\frac {1}{H \\times W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\max  (0, s - I _ {i, j}) ^ {2} \\right] \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.576,
+                0.47,
+                0.68
+            ],
+            "angle": 0,
+            "content": "where \\(\\lambda_{oversat}\\) and \\(\\lambda_{undersat}\\) are the penalty weights for over-saturation and under-saturation respectively, \\(H\\) and \\(W\\) are the height and width of the images, \\(I_{i,j}\\) is the pixel intensity at pixel location \\((i,j)\\), and \\(s\\) is the saturation threshold (i.e., the maximum allowed pixel intensity). The penalty is computed as the mean squared difference between the pixel intensity and the saturation threshold."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.681,
+                0.471,
+                0.773
+            ],
+            "angle": 0,
+            "content": "Recoloring requires swapping albedo and shading components in all losses, except we do not use decorrelation loss while recoloring. Obtaining good results requires quite a careful choice of loss weights (\\(\\lambda\\) coefficients). We experiment with several \\(\\lambda\\) coefficients for both these edits (Section 4 and Supplementary)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.785,
+                0.363,
+                0.801
+            ],
+            "angle": 0,
+            "content": "4. Model and Directions Selection"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.81,
+                0.471,
+                0.901
+            ],
+            "angle": 0,
+            "content": "We prompt StyleGAN to find style code directions that: (a) do not change the albedo, and (b) strongly change the image. We use many image decomposition models to obtain directions across multiple hyperparameter settings. We have no particular reason to believe that a single model will give only good directions or all good directions. We then find"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.092,
+                0.894,
+                0.198
+            ],
+            "angle": 0,
+            "content": "a subset of admissible models. We must choose admissible models using a plot of albedo change versus diversity because there is no way to weigh these effects against one another. However, relatively few methods are admissible – see Figure 3. We then pool all directions from all of the admissible models and use forward selection to find a small set of polished directions in this pool."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.205,
+                0.894,
+                0.31
+            ],
+            "angle": 0,
+            "content": "Scoring Albedo Change: We use SuperPoints [19] to find 100 interest points in the original StyleGAN-generated image. Around each interest point, we form a \\(8 \\times 8\\) patch. We then compare these patches with patches in the same locations for multiple different relightings of that image. If the albedo in the image does not change, then each patch will have the same albedo but different lighting."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.312,
+                0.894,
+                0.54
+            ],
+            "angle": 0,
+            "content": "Given two color image patches \\(\\mathbf{p}\\) and \\(\\mathbf{q}\\), viewed under different lights, we must measure the difference between their albedos \\(d_{a}(\\mathbf{p},\\mathbf{q})\\). Write \\(p_{ij}\\) for the RGB vector at the \\(i\\), \\(j\\)'th location (\\(1 \\leq i \\leq M\\), \\(1 \\leq j \\leq N\\)) and write \\(p_{ij,k}\\) for the \\(k\\)'th RGB component at that location. The intensity of the light may change without the albedo changing, so this problem is homogeneous (i.e. for \\(\\lambda, \\mu > 0\\), \\(d_{a}(\\mathbf{p},\\mathbf{q}) = d_{a}(\\lambda \\mathbf{p},\\mu \\mathbf{q}))\\). Assume that the illumination intensity changes, but not color. The patches are small, so the illumination field on a patch can be modeled as a linear function, so there are albedos \\(\\mathbf{a}\\), \\(\\mathbf{b}\\) such that \\(p_{ij} = (p_{x}i + p_{y}j + p_{c})a_{ij}\\) and \\(q_{ij} = (q_{x}i + q_{y}j + q_{c})b_{ij}\\). If the two patches have similar albedo, there will be \\(p_{x}\\) etc. such that \\(p_{ij}' = (q_{x}i + q_{y}j + q_{c})p_{ij}\\) is the same as \\(q_{ij}' = (p_{x}i + p_{y}j + p_{c})q_{ij}\\). We measure the cosine distance"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.508,
+                0.552,
+                0.892,
+                0.598
+            ],
+            "angle": 0,
+            "content": "\\[\nd _ {a} (\\mathbf {p}, \\mathbf {q}) = 1 - \\max  _ {p _ {x}, \\dots q _ {c}} \\frac {\\sum_ {i j k} p _ {i j k} ^ {\\prime} q _ {i j k} ^ {\\prime}}{\\sqrt {\\sum_ {i j k} \\left(p _ {i j k} ^ {\\prime}\\right) ^ {2}} \\sqrt {\\sum_ {i j k} \\left(p _ {i j k} ^ {\\prime}\\right) ^ {2}}} \\tag {6}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.609,
+                0.892,
+                0.641
+            ],
+            "angle": 0,
+            "content": "The relevant maximum can be calculated by analogy with canonical correlation analysis (Supplementary)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.649,
+                0.892,
+                0.845
+            ],
+            "angle": 0,
+            "content": "Scoring Lighting Diversity: Illumination cone theory [2] yields that any non-negative linear combination of \\( k \\) shadings is a physically plausible shading. To determine if an image is new, we relax the non-negativity constraint and so must ensure that it cannot be expressed as a linear combination of existing images. In turn, we seek a measure of the linear independence of a set of images. This measure should: be large when there is a strong linear dependency; and not grow too fast when the images are scaled. Write \\( \\mathbf{x}_i \\) for the \\( i \\)'th image, and \\( \\mathcal{X} \\) for the matrix whose \\( i \\), \\( j \\)'th component is \\( \\mathbf{x}_i\\mathbf{x}_j \\). Then \\( -\\log \\det \\mathcal{X} \\) is very large when the \\( \\mathbf{x}_i \\) is close to linearly dependent, but does not scale too fast when the images are scaled."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.856,
+                0.894,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Decomposition Models Investigated: We searched 25 instances in total obtained with different hyperparameter settings from three families of decomposition. The first family"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "4234"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.122,
+                0.092,
+                0.422,
+                0.322
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.326,
+                0.473,
+                0.562
+            ],
+            "angle": 0,
+            "content": "Figure 3. Each model from StyLitGAN produces 16 directions. Models differ by choice of hyperparameters and intrinsic image decomposition. We evaluate models by albedo change and by diversity, averaged across a small fixed validation set of test scenes. As the figure shows, there is typically a payoff, but some models are not admissible. Figure 4 shows examples from some of the models considered here. We exclude inadmissible models, then pool all directions from all other models, and apply a forward selection procedure (section 4). This yields 16 strong relighting directions (the gold star). In our comparison with Yang et al. [48], we focus on the changes in albedo since their method identifies only a single relighting direction. This limits the scope for evaluating relighting diversity. Yang et al. [48] aggressively change scene albedo, while our StyLitGAN ensures only lighting changes. Additionally, we compare with GAN-control [39], which, while attempting relighting, often changes the scene layout, leading to large albedo change and increased diversity score due to layout variations."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.572,
+                0.472,
+                0.724
+            ],
+            "angle": 0,
+            "content": "is the SOTA unsupervised model of [14], which decomposes images into albedo and shading using example images drawn from statistical models. The second is a variant of that family that decomposes into albedo, shading, and gloss decomposition [3]. The third is an albedo, shading, and gloss decomposition that models fine edges in the albedo rather than the shading field. These models were chosen to represent a range of possible decompositions, but others could yield better results. The key point is that we can choose a model from a collection by a rational process."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.735,
+                0.473,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Selecting Directions: Our approach for selecting directions involves creating a scatter plot of 25 instances with various hyperparameters and image decomposition models. We find 16 directions for each instance in our final experiments, and the search for 16 directions takes about 14 minutes on an A40 GPU. We experimented with different numbers of directions of order \\(2^{n}\\) for \\(n = 2, 3, 4, 5, 6, 7\\) and found that 16 directions \\((n = 4)\\) strike a better balance between relighting diversity and albedo change. However, finding multiple directions is challenging because the search space is complex and high-dimensional, and we lack ground truth to supervise"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.511,
+                0.089,
+                0.88,
+                0.262
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.498,
+                0.264,
+                0.892,
+                0.362
+            ],
+            "angle": 0,
+            "content": "Figure 4. The bottom row shows scene relightings obtained using our final, forward selected, set of directions (star of Figure 3). For comparison, we also show scene relightings from different models obtained from StyLitGAN shown in Figure 3 (model numbers correspond to numbers on that figure). Note how most models are capable of producing some good directions, but not all directions from a given model may be good."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.373,
+                0.892,
+                0.403
+            ],
+            "angle": 0,
+            "content": "the search. Therefore, we apply a two-step process to find effective directions and filter out any bad directions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.404,
+                0.893,
+                0.524
+            ],
+            "angle": 0,
+            "content": "We first identify and discard inadmissible models that are located behind the Pareto frontier. We then select the top 10 admissible models based on their average albedo change when applied over a large set of fixed validation images. Our goal is to select the best relighting directions from these admissible models. To achieve this, a forward selection process is employed, which involves selecting a subset of directions from the set of admissible directions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.53,
+                0.894,
+                0.621
+            ],
+            "angle": 0,
+            "content": "Forward Selection Process: To select the best directions, we begin with all directions from the admissible models, resulting in 160 directions from 10 models. These directions are then filtered to remove \"bad\" directions that produce relighting similar to the original image or shading that does not vary across pixels, resulting in 108 directions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.621,
+                0.894,
+                0.727
+            ],
+            "angle": 0,
+            "content": "Next, we use a greedy process to select the best 16 directions from the remaining 108 directions. We evaluate each direction one at a time and add it to the pool if it provides a large diversity score while incurring a small penalty for large albedo change. This process continues until the desired number of directions are selected. The forward selection process is fast and efficient, taking less than a minute."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.727,
+                0.894,
+                0.787
+            ],
+            "angle": 0,
+            "content": "The resulting scores from the forward select 16 directions are marked with a star in golden color in Fig. 3. The directions obtained with this process are significantly better than individual models alone. A qualitative ablation is in Fig. 4."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.801,
+                0.633,
+                0.817
+            ],
+            "angle": 0,
+            "content": "5. Experiments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.825,
+                0.895,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Qualitative Evaluation: StyLitGAN produces realistic images that are out of distribution but known to exist for straightforward physical reasons. Because they're out of distribution, current quantitative evaluation tools do not apply. We evaluate realism qualitatively. Further, there is no direct"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "4235"
+        }
+    ],
+    [
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.084,
+                0.088,
+                0.179,
+                0.1
+            ],
+            "angle": 0,
+            "content": "Generated Image"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.192,
+                0.089,
+                0.273,
+                0.1
+            ],
+            "angle": 0,
+            "content": "Relit-1 \\((+d_{1})\\)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.293,
+                0.089,
+                0.376,
+                0.1
+            ],
+            "angle": 0,
+            "content": "Relit-2 \\((+d_{2})\\)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.394,
+                0.089,
+                0.476,
+                0.1
+            ],
+            "angle": 0,
+            "content": "Relit-3 \\((+d_{3})\\)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.497,
+                0.089,
+                0.578,
+                0.1
+            ],
+            "angle": 0,
+            "content": "Relit-4 \\((+d_{4})\\)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.597,
+                0.089,
+                0.679,
+                0.1
+            ],
+            "angle": 0,
+            "content": "Relit-5 \\((+d_{5})\\)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.699,
+                0.089,
+                0.781,
+                0.1
+            ],
+            "angle": 0,
+            "content": "Relit-6 \\((+d_{6})\\)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.801,
+                0.089,
+                0.882,
+                0.1
+            ],
+            "angle": 0,
+            "content": "Relit-7 \\((+d_{7})\\)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.081,
+                0.101,
+                0.894,
+                0.491
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.493,
+                0.893,
+                0.577
+            ],
+            "angle": 0,
+            "content": "Figure 5. First column images generated by the original StyleGAN. Other columns show images obtained from \\(\\mathbf{w}^{+} + \\mathbf{d}_{i}\\), our relighting directions added to the style codes \\((\\mathbf{w}^{+})\\) of the images in the first column. These directions have been chosen to fix albedo, but change shading. Note: each row shows the same scenes but with different illumination. Lighting varies aggressively, and the individual latent direction \\(\\mathbf{d}_{i}\\) has persistent semantics - each column corresponds to a type of illumination. For example, the second \\((d_{1})\\) and third \\((d_{2})\\) column switches on bedside lamps. It is worth noting the presence of soft shadows, cast shadows, inter-reflections, and glossy changes across relights. Another important observation to note is that all relightings are with respect to world coordinates, not camera coordinates."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.096,
+                0.582,
+                0.871,
+                0.744
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.747,
+                0.893,
+                0.817
+            ],
+            "angle": 0,
+            "content": "Figure 6. Resurfacing Generated Images. Instead of relighting images, we can generate resurfaced images by swapping our consistency and diversity loss. We apply diversity loss to change the albedo and consistency loss to maintain the shading and global illumination. The first column shows images generated by the original StyleGAN, and the other columns show images obtained from \\(\\mathbf{w}^{+} + \\mathbf{d}_{i}\\) for our resurfacing or recoloring directions. Each column shows the same scene as in the first column, but with varying surface colors and materials, while the individual latent direction \\(\\mathbf{d}_i\\) retains its semantics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.83,
+                0.473,
+                0.892
+            ],
+            "angle": 0,
+            "content": "comparable method. However, we show relighting comparisons to a recent SOTA method that is physically motivated and trained with CGI data [26] by using a SOTA inversion method [5] (Fig. 14 in Supplementary). For relighting, our"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.83,
+                0.895,
+                0.892
+            ],
+            "angle": 0,
+            "content": "method should generate images that: are clearly relightings of a scene; fix geometry and albedo but visibly change shading; and display complicated illumination effects, including soft shadows, cast shadows and gloss. For resurfacing, our"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.946,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "4236"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.1,
+                0.09,
+                0.872,
+                0.256
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.258,
+                0.894,
+                0.314
+            ],
+            "angle": 0,
+            "content": "Figure 7. Controllability: The first and last columns of the figure show relighted images generated using our Relit-1 and Relit-5 directions. The bottom section of the figure features a user-controllable slider that enables adjusting the weight of the relighting effects produced by these two directions. Moving the slider from left to right results in a seamless interpolation between the two lighting directions and provides precise control over the relighting of the generated images."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.096,
+                0.32,
+                0.877,
+                0.487
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.491,
+                0.893,
+                0.547
+            ],
+            "angle": 0,
+            "content": "Figure 8. Scaling Directions: The figure depicts the persistent and smooth effects of applying the direction at different scalar coefficients. We use our Relit-2 direction. A slider at the bottom allows the user to adjust the weight of the relighting direction, producing a seamless interpolation when increasing or decreasing the intensity of the chosen direction. The relighting effects range from a well-lit room with the bedside lamp off to weak external lighting with a bedside lamp on."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.561,
+                0.473,
+                0.727
+            ],
+            "angle": 0,
+            "content": "method should generate images that: are clearly images of the original layout, but with different materials or colors or changes in furniture; and display illumination effects that are consistent with these color changes. As Figure 6 shows, our method meets these goals. Figure 7 and Figure 8 show interpolation sequences for a relighting between two directions and scaling only one direction. Note that the lighting changes smoothly, as one would expect. Figure 15 in Supplementary shows that our relighting and recoloring directions are largely disentangled. An ablation for individual losses is also given in Supplementary Figure 13."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.728,
+                0.472,
+                0.818
+            ],
+            "angle": 0,
+            "content": "Quantitative Evaluation: Figure 3 shows how we can evaluate albedo and lighting shifts. In Table 1 we show we can generate image datasets with increased FID [18, 34] (clean-FID) from the base comparison set. This is strong evidence our method can produce a set of images that is a strict superset of those that the vanilla StyleGAN can produce."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.818,
+                0.471,
+                0.864
+            ],
+            "angle": 0,
+            "content": "Generality: We have applied our method to StyleGAN trained on Conference Room, Kitchen, Living Room, Dining Room, Church, and Face datasets (results in Figure 10)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.871,
+                0.473,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Comparison to GAN-control: GAN-control (GC) [39] represents lighting with a spherical harmonic predictor pre"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.498,
+                0.55,
+                0.892,
+                0.619
+            ],
+            "angle": 0,
+            "content": "Table 1. FID measures distribution shift and not realism. Our generated images are realistic and are out-of-distribution because of large illumination and color changes in the images. This results in large FID scores. \\( KDL \\) in the table is for kitchen, dining and living room which are jointly trained [13]."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.502,
+                0.619,
+                0.89,
+                0.687
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Type</td><td>Bedroom</td><td>KDL</td><td>Conference</td><td>Church</td><td>Faces</td></tr><tr><td>StyleGAN (SG)</td><td>5.01</td><td>5.86</td><td>9.35</td><td>3.80</td><td>5.02</td></tr><tr><td>SG + Relighting (RL)</td><td>14.23</td><td>6.87</td><td>10.48</td><td>12.12</td><td>37.87</td></tr><tr><td>SG + Resurfacing (RS)</td><td>17.03</td><td>9.41</td><td>10.63</td><td>18.60</td><td>34.06</td></tr><tr><td>SG + RL + RS</td><td>21.39</td><td>11.68</td><td>12.71</td><td>21.08</td><td>37.40</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.705,
+                0.895,
+                0.901
+            ],
+            "angle": 0,
+            "content": "trained on a parameterized 3D face reconstruction model [9]. This model does not apply to indoor lighting (among other problems, it predicts all points with the same normal have the same shading). We trained a GC model on the LSUN Bedroom dataset with 2 subspaces \\( z^k \\) - illumination corresponding to spherical harmonic coefficients, and other to represent all other structural information. The model was trained for 800 epochs. GC produces images whose structure varies wildly with any lighting changes, resulting in large albedo changes (GC in Figures 3, 9). The difficulty appears to be that the attribute predictor is easily subverted; if the lighting representation cannot produce an image that is (say) dark on the left side, the albedo is adjusted instead."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.482,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "4237"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.095,
+                0.089,
+                0.452,
+                0.154
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.077,
+                0.157,
+                0.47,
+                0.185
+            ],
+            "angle": 0,
+            "content": "Figure 9. GAN-control (GC) [39] cannot relight complex scenes like bedrooms; it completely changes the scene's layout."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.088,
+                0.195,
+                0.268,
+                0.332
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.089,
+                0.334,
+                0.266,
+                0.484
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.103,
+                0.486,
+                0.252,
+                0.497
+            ],
+            "angle": 0,
+            "content": "(c) Living Room Relighting"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.088,
+                0.498,
+                0.268,
+                0.635
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.092,
+                0.637,
+                0.264,
+                0.648
+            ],
+            "angle": 0,
+            "content": "(e) Relighting outdoor Churches"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.282,
+                0.195,
+                0.461,
+                0.332
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.31,
+                0.334,
+                0.431,
+                0.346
+            ],
+            "angle": 0,
+            "content": "(b) Kitchen Relighting"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.282,
+                0.347,
+                0.46,
+                0.483
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.296,
+                0.485,
+                0.446,
+                0.497
+            ],
+            "angle": 0,
+            "content": "(d) Dining Room Relighting"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.282,
+                0.498,
+                0.461,
+                0.635
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.312,
+                0.637,
+                0.428,
+                0.649
+            ],
+            "angle": 0,
+            "content": "(f) Portrait Relighting"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.077,
+                0.65,
+                0.47,
+                0.678
+            ],
+            "angle": 0,
+            "content": "Figure 10. StyLitGAN extends to finding relighting directions for StyleGANs trained on other datasets."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.693,
+                0.319,
+                0.71
+            ],
+            "angle": 0,
+            "content": "6. Downstream Applications"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.72,
+                0.473,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Lighting Variance in Surface Normal Prediction: The Multilum dataset [29] provides images of 1000 various indoor scenes, each under 25 lighting conditions, physically relit and captured. A SOTA normal predictor (Omnidata [12, 21]) applied to the test set produces surface normal predictions that vary significantly with changes of light in a fixed scene (Figure 11, purple bar). Finetuning Omnidata with the Multilum training dataset significantly reduces this variance (Figure 11, orange bar); but multiple lightings of a fixed scene are very hard to find. Finetuning with StyLitGAN relights produces comparable improvements; using seven distinct relights (Figure 11, pale orange bar) is slightly"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.552,
+                0.079,
+                0.837,
+                0.273
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.499,
+                0.277,
+                0.896,
+                0.403
+            ],
+            "angle": 0,
+            "content": "Figure 11. Normal variance to lighting reduces when fine-tuning on our relighting dataset. Purple boxplot shows normal variance under relighting for real test scenes from Multilum using the Omnidata normal predictor; orange shows the result of finetuning using Multilum training data; light orange and yellow show the result of finetuning using StyLitGAN images (7 and 25 per scene respectively). The measure is angular error in radians from the mean prediction of a scene for each relit image in the Multilum test set (30 scenes, 25 lightings each)."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.543,
+                0.411,
+                0.847,
+                0.586
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.499,
+                0.588,
+                0.894,
+                0.645
+            ],
+            "angle": 0,
+            "content": "Figure 12. A surface normal predictor is finetuned to increase prediction consistency for relit images of the same scene. General performance shown on Taskonomy test images parallels that of the original model. Mean across buildings is given in the last column."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.655,
+                0.895,
+                0.774
+            ],
+            "angle": 0,
+            "content": "worse than using 25 distinct relights (Figure 11, yellow bar). The resulting improvement is not at the cost of base accuracy. Figure 12 compares the accuracy of various finetuned models on the Taskonomy test set [50] (recall this involves thousands of frames each in 10 blocks; we show results by block). Note that finetuned methods mostly show slight accuracy improvements over the base model, but losses in some blocks result in means that match."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.775,
+                0.894,
+                0.807
+            ],
+            "angle": 0,
+            "content": "In our follow-up work, we also show that StyleGAN \"knows\" intrinsic images and can be easily extracted [4]."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.819,
+                0.652,
+                0.836
+            ],
+            "angle": 0,
+            "content": "Acknowledgment"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.84,
+                0.894,
+                0.902
+            ],
+            "angle": 0,
+            "content": "We thank Aniruddha Kembhavi, Derek Hoiem, Min Jin Chong and Shenlong Wang. This material is based upon work supported by the National Science Foundation under Grant No. 2106825 and by a gift from Boeing Corporation."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "4238"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.079,
+                0.09,
+                0.176,
+                0.107
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.115,
+                0.473,
+                0.183
+            ],
+            "angle": 0,
+            "content": "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.183,
+                0.472,
+                0.224
+            ],
+            "angle": 0,
+            "content": "[2] Peter N Belhumeur and David J Kriegman. What is the set of images of an object under all possible illumination conditions? International Journal of Computer Vision, 1998. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.224,
+                0.472,
+                0.277
+            ],
+            "angle": 0,
+            "content": "[3] Anand Bhattachad and D.A. Forsyth. Cut-and-paste object insertion by enabling deep image prior for reshading. In 2022 International Conference on 3D Vision (3DV). IEEE, 2022. 2, 3, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.278,
+                0.472,
+                0.33
+            ],
+            "angle": 0,
+            "content": "[4] Anand Bhattachad, Daniel McKee, Derek Hoiem, and David Forsyth. Stylegan knows normal, depth, albedo, and more. Advances in Neural Information Processing Systems, 36, 2024. 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.33,
+                0.472,
+                0.386
+            ],
+            "angle": 0,
+            "content": "[5] Anand Bhattachad, Viraj Shah, Derek Hoiem, and David A Forsyth. Make it so: Steering stylegan for any image inversion and editing. arXiv preprint arXiv:2304.14403, 2023. 6, 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.386,
+                0.472,
+                0.455
+            ],
+            "angle": 0,
+            "content": "[6] Prashanth Chandran, Sebastian Winberg, Gaspard Zoss, Jérémy Riviere, Markus Gross, Paulo Gotardo, and Derek Bradley. Rendering with style: combining traditional and neural approaches for high-quality face rendering. ACM Transactions on Graphics (ToG), 40(6):1-14, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.455,
+                0.472,
+                0.481
+            ],
+            "angle": 0,
+            "content": "[7] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. arXiv preprint arXiv:2112.11641, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.481,
+                0.472,
+                0.522
+            ],
+            "angle": 0,
+            "content": "[8] Min Jin Chong, Hsin-Ying Lee, and David Forsyth. Stylegan of all trades: Image manipulation with only pretrained stylegan. arXiv preprint arXiv:2111.01619, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.522,
+                0.472,
+                0.587
+            ],
+            "angle": 0,
+            "content": "[9] 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."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.079,
+                0.588,
+                0.472,
+                0.655
+            ],
+            "angle": 0,
+            "content": "[10] Aditya Deshpande, Jiajun Lu, Mao-Chuang Yeh, Min Jin Chong, and David Forsyth. Learning diverse image colorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6837-6845, 2017. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.079,
+                0.657,
+                0.472,
+                0.711
+            ],
+            "angle": 0,
+            "content": "[11] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 341-346, 2001. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.079,
+                0.711,
+                0.472,
+                0.779
+            ],
+            "angle": 0,
+            "content": "[12] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.079,
+                0.779,
+                0.472,
+                0.831
+            ],
+            "angle": 0,
+            "content": "[13] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A Efros. Blobgan: Spatially disentangled scene representations. arXiv preprint arXiv:2205.02837, 2022. 2, 3, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.079,
+                0.832,
+                0.472,
+                0.86
+            ],
+            "angle": 0,
+            "content": "[14] D.A. Forsyth and Jason J Rock. Intrinsic image decomposition using paradigms. TPAMI, 2022 in press. 2, 3, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.079,
+                0.86,
+                0.472,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.079,
+                0.115,
+                0.473,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.533,
+                0.094,
+                0.815,
+                0.106
+            ],
+            "angle": 0,
+            "content": "pattern recognition, pages 2414-2423, 2016. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.106,
+                0.893,
+                0.16
+            ],
+            "angle": 0,
+            "content": "[16] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. arXiv preprint arXiv:1406.2661, 2014. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.16,
+                0.894,
+                0.214
+            ],
+            "angle": 0,
+            "content": "[17] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 327-340, 2001. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.214,
+                0.894,
+                0.268
+            ],
+            "angle": 0,
+            "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2017. 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.268,
+                0.894,
+                0.322
+            ],
+            "angle": 0,
+            "content": "[19] Le Hui, Jia Yuan, Mingmei Cheng, Jin Xie, Xiaoya Zhang, and Jian Yang. Superpoint network for point cloud oversegmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5510-5519, 2021. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.322,
+                0.893,
+                0.363
+            ],
+            "angle": 0,
+            "content": "[20] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, 2016. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.363,
+                0.894,
+                0.429
+            ],
+            "angle": 0,
+            "content": "[21] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 2, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.43,
+                0.893,
+                0.485
+            ],
+            "angle": 0,
+            "content": "[22] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.484,
+                0.893,
+                0.539
+            ],
+            "angle": 0,
+            "content": "[23] 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, 2019."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.539,
+                0.893,
+                0.58
+            ],
+            "angle": 0,
+            "content": "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.58,
+                0.893,
+                0.605
+            ],
+            "angle": 0,
+            "content": "[25] Edwin H Land. The retina theory of color vision. Scientific american, 1977. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.605,
+                0.893,
+                0.673
+            ],
+            "angle": 0,
+            "content": "[26] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. arXiv preprint arXiv:2205.09343, 2022. 6, 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.673,
+                0.893,
+                0.727
+            ],
+            "angle": 0,
+            "content": "[27] Zicheng Liao, Hugues Hoppe, David Forsyth, and Yizhou Yu. A subdivision-based representation for vector image editing. IEEE transactions on visualization and computer graphics, 2012. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.727,
+                0.893,
+                0.782
+            ],
+            "angle": 0,
+            "content": "[28] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. Advances in Neural Information Processing Systems, 34, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.782,
+                0.893,
+                0.836
+            ],
+            "angle": 0,
+            "content": "[29] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A multi-illumination dataset of indoor object appearance. In 2019 IEEE International Conference on Computer Vision (ICCV), Oct 2019. 2, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.836,
+                0.893,
+                0.889
+            ],
+            "angle": 0,
+            "content": "[30] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, Epic Games, Andreas Lehrmann, and AI Borealis. Learning physics-guided face relighting under directional light. 2020. 2"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.504,
+                0.094,
+                0.894,
+                0.889
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.946,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "4239"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.092,
+                0.472,
+                0.16
+            ],
+            "angle": 0,
+            "content": "[31] Xingang Pan, Xudong Xu, Chen Change Loy, Christian Theobalt, and Bo Dai. A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.16,
+                0.472,
+                0.229
+            ],
+            "angle": 0,
+            "content": "[32] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.229,
+                0.472,
+                0.284
+            ],
+            "angle": 0,
+            "content": "[33] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.284,
+                0.472,
+                0.322
+            ],
+            "angle": 0,
+            "content": "[34] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.322,
+                0.472,
+                0.365
+            ],
+            "angle": 0,
+            "content": "[35] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. IEEE Computer graphics and applications, 21(5):34-41, 2001. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.365,
+                0.472,
+                0.432
+            ],
+            "angle": 0,
+            "content": "[36] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.432,
+                0.472,
+                0.486
+            ],
+            "angle": 0,
+            "content": "[37] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.486,
+                0.472,
+                0.54
+            ],
+            "angle": 0,
+            "content": "[38] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.54,
+                0.472,
+                0.595
+            ],
+            "angle": 0,
+            "content": "[39] Alon Shoshan, Nadav Bhonker, Igor Kviatkovsky, and Gerard Medioni. Gan-control: Explicitly controllable gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14083-14093, 2021. 2, 5, 7, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.595,
+                0.472,
+                0.633
+            ],
+            "angle": 0,
+            "content": "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. *ICLR*, 2015. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.634,
+                0.472,
+                0.689
+            ],
+            "angle": 0,
+            "content": "[41] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics, 2019. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.689,
+                0.472,
+                0.756
+            ],
+            "angle": 0,
+            "content": "[42] Feitong Tan, Sean Fanello, Abhimitra Meka, Sergio Orts-Escolano, Danhang Tang, Rohit Pandey, Jonathan Taylor, Ping Tan, and Yinda Zhang. Volux-gan: A generative model for 3d face synthesis with hdri relighting. arXiv preprint arXiv:2201.04873, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.756,
+                0.472,
+                0.839
+            ],
+            "angle": 0,
+            "content": "[43] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.839,
+                0.472,
+                0.879
+            ],
+            "angle": 0,
+            "content": "[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.879,
+                0.472,
+                0.892
+            ],
+            "angle": 0,
+            "content": "[45] Andrey Voynov and Artem Babenko. Unsupervised discovery"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.092,
+                0.472,
+                0.892
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.535,
+                0.093,
+                0.894,
+                0.133
+            ],
+            "angle": 0,
+            "content": "of interpretable directions in the gan latent space. In International conference on machine learning, pages 9786-9796. PMLR, 2020. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.133,
+                0.894,
+                0.201
+            ],
+            "angle": 0,
+            "content": "[46] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.201,
+                0.894,
+                0.242
+            ],
+            "angle": 0,
+            "content": "[47] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. arXiv preprint arXiv: 2101.05278, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.242,
+                0.894,
+                0.295
+            ],
+            "angle": 0,
+            "content": "[48] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. International Journal of Computer Vision, 2020. 2, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.295,
+                0.894,
+                0.363
+            ],
+            "angle": 0,
+            "content": "[49] Ning Yu, Guilin Liu, Aysegul Dundar, Andrew Tao, Bryan Catanzaro, Larry S Davis, and Mario Fritz. Dual contrastive loss and attention for gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6731-6742, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.363,
+                0.894,
+                0.431
+            ],
+            "angle": 0,
+            "content": "[50] Amir R. Zamir, Alexander Sax, William B. Shen, Leonidas J. Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2018. 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.431,
+                0.894,
+                0.472
+            ],
+            "angle": 0,
+            "content": "[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.472,
+                0.894,
+                0.525
+            ],
+            "angle": 0,
+            "content": "[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.525,
+                0.894,
+                0.58
+            ],
+            "angle": 0,
+            "content": "[53] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.58,
+                0.894,
+                0.635
+            ],
+            "angle": 0,
+            "content": "[54] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, 2017. 2"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.503,
+                0.093,
+                0.894,
+                0.635
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.517,
+                0.956
+            ],
+            "angle": 0,
+            "content": "4240"
+        }
+    ]
+]

2024/StyLitGAN_ Image-Based Relighting via Latent Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b612a2417e244bf3685c80b4e5e49b8134a122562ef469501182a252939d6fc4
+size 10342680

2024/StyLitGAN_ Image-Based Relighting via Latent Control/full.md

@@ -0,0 +1,280 @@
+# StyLitGAN: Image-based Relighting via Latent Control
+
+Anand Bhattad James Soole D.A. Forsyth
+
+University of Illinois Urbana-Champaign
+
+https://anandbhattad.github.io/stylitgan/
+
+Generated Image
+
+Relit-1  $(\mathbf{w}^{+} + d_{1})$  Relit-2  $(\mathbf{w}^{+} + d_{2})$  Relit-3  $(\mathbf{w}^{+} + d_{3})$  Relit-4  $(\mathbf{w}^{+} + d_{4})$  Relit-5  $(\mathbf{w}^{+} + d_{5})$
+
+![](images/ca50ceef1b4c4acb1b024de96b1a74ca80e39eb45597d829b2b3de7949c1e4ba.jpg)  
+Figure 1. StyLitGAN identifies directional vectors  $(d_i)$  within StyleGAN's style space  $(\mathcal{W}^{+})$  which, when added to the  $\mathbf{w}^{+}$ style code, effectively modify the lighting of generated images while preserving their geometry and albedo. This process eliminates the need for per-image search or model fine-tuning. The first column displays images generated from StyleGAN2; subsequent columns illustrate the same scene, each relit using a specific direction. These relighting directions  $(d_i)$  are derived through a forward selection method, ensuring diversity and avoiding cherry-picking. The directional effects are consistent across different scenes: for instance,  $d_1$  activates an orange-tinged bedside lamp,  $d_2$  a less intense white-tinged lamp,  $d_3$  introduces strong directional light from the window, and so on, demonstrating diverse relighting capabilities of StyLitGAN.
+
+# Abstract
+
+We describe a novel method, StyLitGAN, for relighting and resurfacing images in the absence of labeled data. StyLitGAN generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects, without the need for paired or CGI data. StyLitGAN uses an intrinsic image method to decompose an image, followed by a search of the latent space of a pretrained StyleGAN to identify a set of directions. By prompting the model to fix one component (e.g., albedo) and vary another (e.g., shading), we generate relighted images by adding the identified directions to the latent style codes. Quantitative metrics of change in albedo and lighting diversity allow us to choose effective directions using a forward selection process. Qualitative evaluation confirms the effectiveness of our method.
+
+# 1. Introduction
+
+Scene appearance shifts dramatically with varying lighting conditions - a sunlit room takes on a different character as daylight fades, and interior spaces transform with the flick of a switch. Similarly, surface changes, like a wall's paint color, change not only the wall's appearance but also the overall image due to light reflection. Despite the impressive realism achieved by current generative models like StyleGAN [22-24], they fall short in dynamically controlling scene lighting, a key aspect of realistic image generation.
+
+In this work, we present StyLitGAN, a novel approach that extends the editing capabilities of StyleGAN [38, 45, 46, 53]. StyLitGAN uniquely manipulates style codes to selectively change lighting while preserving other image attributes like albedo and geometry. This selective editing
+
+addresses a critical gap in current generative methods, which typically lack the precision to control individual scene components independently.
+
+Our method, StyLitGAN, first uses StyleGAN to produce a set of images and then decomposes these generated images into albedo, diffuse shading, and glossy effects using an off-the-shelf, self-supervised network [14]. We then search for style code edits by prompting StyleGAN to produce images that (a) are diverse, but (b) have the same albedo (and so geometry and material) as the original generated images. Our search selects the most effective relighting directions in a data-driven manner.
+
+Our approach generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects. Importantly, we observe that style code edits produce consistent effects across images. For instance, as seen in Fig. 1, adding the first and second directions tends to switch on bedside lamps (columns Relit-1 and Relit-2), while adding the fourth direction increases the light intensity from outside the window (column Relit-4). Since StyLitGAN can generate any image that a vanilla StyleGAN can, it also generate images that are out of distribution, one would expect FID scores to increase over StyleGAN; this happens.
+
+The recent method GAN-control [39] controls lighting on face images using an attribute procedure. We show, in contrast to StyLitGAN, GAN-control fails on indoor scenes, likely because the attribute vocabulary is too easily subverted by the complex lighting effects in indoor scenes.
+
+We demonstrate applications of StyLitGAN to standard vision problems. Using the Multi-Ilum dataset of [29], we show that predictions from a SOTA surface normal predictor [21] vary significantly when lighting is changed. Finetuning this normal predictor using StyLitGAN images suppresses this effect. The improvement is comparable with that obtained by finetuning with true multi-illuminant images (which are very difficult to obtain in quantity).
+
+# 2. Related Work
+
+Image Manipulation: A significant literature deals with manipulating and editing images [3, 10, 11, 15, 17, 27, 33, 35, 44, 54]. Editing procedures for generative image models [16] are important, because they demand compact image representations with useful, disentangled interpretations. StyleGAN [22-24] is currently de facto state-of-the-art for editing generated images, likely because its mapping of initial noise vectors to style codes which control entire feature layers produces latent spaces that are heavily disentangled and so easy to manipulate. Recent editing methods include [8, 36, 38, 45, 46, 53], with a survey in [47]. The architecture can be adapted to incorporate spatial priors for authoring novel and edited images [13, 28, 43]. In contrast to this literature, we show how to fix one physically meaningful image factor while changing another. Doing so is difficult
+
+because the latent spaces are not perfectly disentangled, and we must produce a diverse set of changes in only one factor. Relighting using StyleGAN: Relighting faces using StyleGAN can be achieved with Stylerig [43], but this method requires a 3D morphable face model. In contrast, StyLitGAN does not require a 3D model and can be extended to complex indoor scenes, which is not possible with Stylerig. Yang et al. [48] uses semantic label attributes to train a binary classifier to find latent space directions that represent indoor and natural lighting, but this method cannot produce diverse relighting effects. We also find Yang et al's relighting to change color or albedo. In contrast, StyLitGAN generates diverse realistic relighting effects without changing albedo and without requiring any labeled attributes.
+
+StyleFlow [1] and GAN-control [39] require a parametric model to express lighting, such as spherical harmonics. These methods are limited to relighting faces and do not result in realistic relighting of rooms. Our experiments using GAN-control for rooms result in large geometry and albedo change. In contrast, StyLitGAN can produce relighted images without changing geometry or albedo. We also note that rooms are more challenging to relight than faces due to significant long-scale inter-reflection effects, diverse shadow patterns, stylized luminaires, stylized surface albedos, and surface brightnesses that are not a function of surface normal alone. These factors make it difficult to apply GAN control directly to rooms. Also, none of these methods can resurface or recolor rooms, though StyLitGAN can also edit color or materials while preserving the scene's lighting.
+
+Other Face Relighting methods use carefully collected supervisory data from light-stages or parametric spherical harmonics [30, 32, 37, 41, 52]. ShadeGAN [31], Rendering with Style [6], and Volux-GAN [42] use a volumetric rendering approach to learn the 3D structure of the face and the illumination encoding. Volux-GAN [42] also requires image decomposition from [32] that is trained using carefully curated light-stage data. In comparison, we neither require any explicit 3D modeling of the scene nor labeled and curated data for training the image decomposition model.
+
+# 3. Approach
+
+We follow convention and manipulate StyleGAN [24] by adjusting the  $\mathbf{w}^{+}$  latent variables. We do not modify StyleGAN weights, but instead, seek a set of lighting directions  $\mathbf{d}_i$  (same shape as  $\mathbf{w}^{+}$ ) which are independent of  $\mathbf{w}^{+}$  and have desired effects on the generated image. We obtain these directions by constructing losses that capture the desired outcomes, and then search for directions that minimize these losses. We find all directions only once and use 2000 randomly generated images for this search. Once found, these lighting directions apply to all other generated images. Our search procedure only sees each image once.
+
+Fig. 2 summarizes our procedure and we call our model
+
+![](images/ef00eae8814e93b30e2a78df87c02ad12b5ad1ac95088f9a84c9e6b3b31bbaa6.jpg)  
+Figure 2. How StyLitGAN works: We generate an image from random Gaussian noise using a pretrained StyleGAN. We also generate novel relighted versions (16 in our case) of the same image using randomly initialized latent directions (d) that are added to  $\mathbf{w}^{+}$ latent style codes. We train a classifier (F) that takes in all the pairs of relighted and original images and predicts the relighting direction applied to them. We apply a distinction loss and jointly update the latent directions and the classifier. Next, we generate the decomposition of these images from a pretrained decomposition model (D). We then apply losses that force StyLitGAN to find latent directions such that the albedo does not change (consistency loss), but the image does (diversity loss).
+
+StyLitGAN. Our method consists of two stages. The first stage involves decomposing images using a pretrained model. The second stage jointly searches for directions and trains a classifier  $\mathrm{F}$ . The classifier  $\mathrm{F}$  predicts the latent direction applied to image pairs. It's a classification task with a fixed number of jointly learned latent directions. While classifier  $\mathrm{F}$  doesn't directly know whether latent directions relate to lighting, our use of image decomposition losses ensures these directions are lighting-related. The consistency loss maintains albedo, and the diversity loss ensures diverse shading changes, both properties expected when changing lighting conditions. Thus, while jointly updating the classifier  $\mathrm{F}$  and latent directions, these losses ensure the discovered directions pertain to relighting. We now elaborate on our search for directions and losses in detail.
+
+Base StyleGAN Models: We use baseline pretrained models from [49] that use a dual-contrastive loss to train StyleGAN for bedrooms, faces, and churches. We also use baseline pretrained StyleGAN2 models from [13] for conference rooms, kitchens, dining rooms, and living rooms.
+
+Decomposition: We decompose images into albedo, shading, and gloss maps (gloss only when available) as  $A \times S + G$ , where  $A$  models albedo effects and  $S$  and  $G$  model shading and gloss effects respectively. We use the method of Forsyth and Rock [14] and its variant from Bhattad and Forsyth [3], which is easily adapted because it is self-supervised and uses only samples from statistical models derived from Land's Retinex theory [25]. By changing the statistical spatial model parameters, we can construct many decompositions using their approach. We evaluate many such decomposition models under several hyperparameter settings and create a large pool of relighting directions. We finalize our directions using a forward selection process that provides minimal albedo and geometry shift with a large relighting diversity (Section 4).
+
+Relighting a scene should produce a new, realistic image where the shading has changed but the albedo has
+
+not. Write  $I(\mathbf{w}^{+})$  for the image produced by StyleGAN given style codes  $\mathbf{w}^{+}$ , and  $A(I)$ ,  $S(I)$ , and  $G(I)$  for the albedo, shading, and gloss respectively recovered from image  $I$ . We search for multiple directions  $\mathbf{d}_i$  such that: (a)  $A(I(\mathbf{w}^{+} + \mathbf{d}_i))$  is very close to  $A(I(\mathbf{w}^{+}))$  – so the image is a relighted version of  $I(\mathbf{w}^{+})$ , a property we call persistent consistency; (b) the images produced by the different directions are linearly independent – relighting diversity; (c) which direction was used can be determined from the image, so that different directions have visibly distinct effects – distinctive relighting; and (d) the new shading field (map) is not strongly correlated to the albedo – independent relighting. Not every shading field can be paired with a given albedo, otherwise there would be nothing to do. We assume that edited  $\mathbf{w}^{+}$  will result in realistic images [7].
+
+Recoloring: Alternatively, we may wish to edit scenes where the colors or materials of objects have changed, but the lighting hasn't. Because shading conveys a great deal of information about shape, we can find these edits using modified losses by seeking consistency in the shading field.
+
+Persistent Consistency: The albedo decomposition of both the relighted scene:  $A_{R} = A(I(\mathbf{w}^{+} + \mathbf{d}_{i}))$  and the original:  $A_{O} = A(I(\mathbf{w}^{+}))$  must be the same; where R refers to relighted images and O refers to StyleGAN generated images. We use a Huber loss and a perceptual feature loss [20, 51] from a VGG feature extractor (Φ) [40] at various feature layers (j) to preserve persistent effects (geometry, appearance, and texture) in the scene.
+
+$$
+\begin{array}{l} \mathcal {L} _ {\text {c o n s t}} \left(A _ {O}, A _ {R}\right) = \left\{ \begin{array}{c} \frac {1}{2} \left[ A _ {O} - A _ {R} \right] ^ {2} \quad \text {f o r} | A _ {O} - A _ {R} | \leq \delta , \\ \delta \left(| A _ {O} - A _ {R} | - \delta / 2\right) \quad \text {o t h e r w i s e .} \end{array} \right. (1) \\ \mathcal {L} _ {p e r} \left(A _ {O}, A _ {R}\right) = \left| \left| \Phi_ {j} \left(A _ {O}\right) - \Phi_ {j} \left(A _ {R}\right) \right| \right| _ {2}. (2) \\ \end{array}
+$$
+
+Relighting Diversity: We want the set of relighted images produced by the directions to be diverse on a long scale so that regions that were in shadow in one image might be
+
+bright in another. For each  $S(\mathbf{w}^{+} + \mathbf{d}_{i})$ , we stack the two shading and gloss:  $S$  and  $G$ , and compute a smoothed and downsampled vector  $\mathbf{t}_{i}$  from these maps. We then compute  $\mathcal{L}_{div}(S,G)$  (diversity loss) which compels these  $\mathbf{t}_{i}$  to be linearly independent and encourages diversity in relighting.
+
+$$
+\mathcal {L} _ {\text {d i v}} (S, G) = - \log \det  N \tag {3}
+$$
+
+where  $i^{th}$  &  $j^{th}$  component of  $\mathbf{N}$  is  $t_i^\top t_j$
+
+Distinctive Relighting: A network might try to cheat by making minimal changes to the image. Directions  $\mathbf{d}_i$  should have the property that  $\mathbf{d}_i$  is easy to impute from  $I(\mathbf{w}^{+} + \mathbf{d}_{i})$ . Joint with the search for directions, we train a classifier to categorize the applied direction. This classifier accepts  $I(\mathbf{w}^{+})$  and  $I(\mathbf{w}^{+} + \mathbf{d}_{i})$  and must predict  $i$ . Its cross-entropy supplies our loss:
+
+$$
+\begin{array}{l} \min  _ {l, F} \mathcal {L} _ {\text {d i s t}} (I (\mathbf {w} ^ {+}), I (\mathbf {w} ^ {+} + d _ {i})) \\ = - \sum_ {i = 1} ^ {M} y _ {i} \log F \left(I \left(\mathbf {w} ^ {+}\right), I \left(\mathbf {w} ^ {+} + d _ {i}\right)\right) \tag {4} \\ \end{array}
+$$
+
+Saturation Penalty: Our diversity loss might cheat and obtain high diversity by generating blocks of over-saturated or under-saturated pixels. So, we apply a saturation penalty over several pixels within a certain threshold.
+
+$$
+\begin{array}{l} \mathcal {L} _ {s a t} = \lambda_ {\text {o v e r s a t}} \left[ \frac {1}{H \times W} \sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} \max  (0, I _ {i, j} - s) ^ {2} \right] \tag {5} \\ + \lambda_ {\text {u n d e r s a t}} \left[ \frac {1}{H \times W} \sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} \max  (0, s - I _ {i, j}) ^ {2} \right] \\ \end{array}
+$$
+
+where  $\lambda_{oversat}$  and  $\lambda_{undersat}$  are the penalty weights for over-saturation and under-saturation respectively,  $H$  and  $W$  are the height and width of the images,  $I_{i,j}$  is the pixel intensity at pixel location  $(i,j)$ , and  $s$  is the saturation threshold (i.e., the maximum allowed pixel intensity). The penalty is computed as the mean squared difference between the pixel intensity and the saturation threshold.
+
+Recoloring requires swapping albedo and shading components in all losses, except we do not use decorrelation loss while recoloring. Obtaining good results requires quite a careful choice of loss weights ( $\lambda$  coefficients). We experiment with several  $\lambda$  coefficients for both these edits (Section 4 and Supplementary).
+
+# 4. Model and Directions Selection
+
+We prompt StyleGAN to find style code directions that: (a) do not change the albedo, and (b) strongly change the image. We use many image decomposition models to obtain directions across multiple hyperparameter settings. We have no particular reason to believe that a single model will give only good directions or all good directions. We then find
+
+a subset of admissible models. We must choose admissible models using a plot of albedo change versus diversity because there is no way to weigh these effects against one another. However, relatively few methods are admissible – see Figure 3. We then pool all directions from all of the admissible models and use forward selection to find a small set of polished directions in this pool.
+
+Scoring Albedo Change: We use SuperPoints [19] to find 100 interest points in the original StyleGAN-generated image. Around each interest point, we form a  $8 \times 8$  patch. We then compare these patches with patches in the same locations for multiple different relightings of that image. If the albedo in the image does not change, then each patch will have the same albedo but different lighting.
+
+Given two color image patches  $\mathbf{p}$  and  $\mathbf{q}$ , viewed under different lights, we must measure the difference between their albedos  $d_{a}(\mathbf{p},\mathbf{q})$ . Write  $p_{ij}$  for the RGB vector at the  $i$ ,  $j$ 'th location ( $1 \leq i \leq M$ ,  $1 \leq j \leq N$ ) and write  $p_{ij,k}$  for the  $k$ 'th RGB component at that location. The intensity of the light may change without the albedo changing, so this problem is homogeneous (i.e. for  $\lambda, \mu > 0$ ,  $d_{a}(\mathbf{p},\mathbf{q}) = d_{a}(\lambda \mathbf{p},\mu \mathbf{q}))$ . Assume that the illumination intensity changes, but not color. The patches are small, so the illumination field on a patch can be modeled as a linear function, so there are albedos  $\mathbf{a}$ ,  $\mathbf{b}$  such that  $p_{ij} = (p_{x}i + p_{y}j + p_{c})a_{ij}$  and  $q_{ij} = (q_{x}i + q_{y}j + q_{c})b_{ij}$ . If the two patches have similar albedo, there will be  $p_{x}$  etc. such that  $p_{ij}' = (q_{x}i + q_{y}j + q_{c})p_{ij}$  is the same as  $q_{ij}' = (p_{x}i + p_{y}j + p_{c})q_{ij}$ . We measure the cosine distance
+
+$$
+d _ {a} (\mathbf {p}, \mathbf {q}) = 1 - \max  _ {p _ {x}, \dots q _ {c}} \frac {\sum_ {i j k} p _ {i j k} ^ {\prime} q _ {i j k} ^ {\prime}}{\sqrt {\sum_ {i j k} \left(p _ {i j k} ^ {\prime}\right) ^ {2}} \sqrt {\sum_ {i j k} \left(p _ {i j k} ^ {\prime}\right) ^ {2}}} \tag {6}
+$$
+
+The relevant maximum can be calculated by analogy with canonical correlation analysis (Supplementary).
+
+Scoring Lighting Diversity: Illumination cone theory [2] yields that any non-negative linear combination of  $k$  shadings is a physically plausible shading. To determine if an image is new, we relax the non-negativity constraint and so must ensure that it cannot be expressed as a linear combination of existing images. In turn, we seek a measure of the linear independence of a set of images. This measure should: be large when there is a strong linear dependency; and not grow too fast when the images are scaled. Write  $\mathbf{x}_i$  for the  $i$ 'th image, and  $\mathcal{X}$  for the matrix whose  $i$ ,  $j$ 'th component is  $\mathbf{x}_i\mathbf{x}_j$ . Then  $-\log \det \mathcal{X}$  is very large when the  $\mathbf{x}_i$  is close to linearly dependent, but does not scale too fast when the images are scaled.
+
+Decomposition Models Investigated: We searched 25 instances in total obtained with different hyperparameter settings from three families of decomposition. The first family
+
+![](images/b5252d7045a48fc04cdc2dbabcad7e250e0ce1d2b8033904dcbac07430850133.jpg)  
+Figure 3. Each model from StyLitGAN produces 16 directions. Models differ by choice of hyperparameters and intrinsic image decomposition. We evaluate models by albedo change and by diversity, averaged across a small fixed validation set of test scenes. As the figure shows, there is typically a payoff, but some models are not admissible. Figure 4 shows examples from some of the models considered here. We exclude inadmissible models, then pool all directions from all other models, and apply a forward selection procedure (section 4). This yields 16 strong relighting directions (the gold star). In our comparison with Yang et al. [48], we focus on the changes in albedo since their method identifies only a single relighting direction. This limits the scope for evaluating relighting diversity. Yang et al. [48] aggressively change scene albedo, while our StyLitGAN ensures only lighting changes. Additionally, we compare with GAN-control [39], which, while attempting relighting, often changes the scene layout, leading to large albedo change and increased diversity score due to layout variations.
+
+is the SOTA unsupervised model of [14], which decomposes images into albedo and shading using example images drawn from statistical models. The second is a variant of that family that decomposes into albedo, shading, and gloss decomposition [3]. The third is an albedo, shading, and gloss decomposition that models fine edges in the albedo rather than the shading field. These models were chosen to represent a range of possible decompositions, but others could yield better results. The key point is that we can choose a model from a collection by a rational process.
+
+Selecting Directions: Our approach for selecting directions involves creating a scatter plot of 25 instances with various hyperparameters and image decomposition models. We find 16 directions for each instance in our final experiments, and the search for 16 directions takes about 14 minutes on an A40 GPU. We experimented with different numbers of directions of order  $2^{n}$  for  $n = 2, 3, 4, 5, 6, 7$  and found that 16 directions  $(n = 4)$  strike a better balance between relighting diversity and albedo change. However, finding multiple directions is challenging because the search space is complex and high-dimensional, and we lack ground truth to supervise
+
+![](images/7f5e3de54231b4937aaace146c50da9cea0c22c9600791c8d69cdae838a7e777.jpg)  
+Figure 4. The bottom row shows scene relightings obtained using our final, forward selected, set of directions (star of Figure 3). For comparison, we also show scene relightings from different models obtained from StyLitGAN shown in Figure 3 (model numbers correspond to numbers on that figure). Note how most models are capable of producing some good directions, but not all directions from a given model may be good.
+
+the search. Therefore, we apply a two-step process to find effective directions and filter out any bad directions.
+
+We first identify and discard inadmissible models that are located behind the Pareto frontier. We then select the top 10 admissible models based on their average albedo change when applied over a large set of fixed validation images. Our goal is to select the best relighting directions from these admissible models. To achieve this, a forward selection process is employed, which involves selecting a subset of directions from the set of admissible directions.
+
+Forward Selection Process: To select the best directions, we begin with all directions from the admissible models, resulting in 160 directions from 10 models. These directions are then filtered to remove "bad" directions that produce relighting similar to the original image or shading that does not vary across pixels, resulting in 108 directions.
+
+Next, we use a greedy process to select the best 16 directions from the remaining 108 directions. We evaluate each direction one at a time and add it to the pool if it provides a large diversity score while incurring a small penalty for large albedo change. This process continues until the desired number of directions are selected. The forward selection process is fast and efficient, taking less than a minute.
+
+The resulting scores from the forward select 16 directions are marked with a star in golden color in Fig. 3. The directions obtained with this process are significantly better than individual models alone. A qualitative ablation is in Fig. 4.
+
+# 5. Experiments
+
+Qualitative Evaluation: StyLitGAN produces realistic images that are out of distribution but known to exist for straightforward physical reasons. Because they're out of distribution, current quantitative evaluation tools do not apply. We evaluate realism qualitatively. Further, there is no direct
+
+![](images/7166d8d77f767b023bc10ba80437020a9bbcd67714d16b616b5c6ea92a4fd88f.jpg)  
+Generated Image  
+Relit-1  $(+d_{1})$  
+Relit-2  $(+d_{2})$  
+Relit-3  $(+d_{3})$  
+Relit-4  $(+d_{4})$  
+Relit-5  $(+d_{5})$  
+Relit-6  $(+d_{6})$  
+Relit-7  $(+d_{7})$  
+Figure 5. First column images generated by the original StyleGAN. Other columns show images obtained from  $\mathbf{w}^{+} + \mathbf{d}_{i}$ , our relighting directions added to the style codes  $(\mathbf{w}^{+})$  of the images in the first column. These directions have been chosen to fix albedo, but change shading. Note: each row shows the same scenes but with different illumination. Lighting varies aggressively, and the individual latent direction  $\mathbf{d}_{i}$  has persistent semantics - each column corresponds to a type of illumination. For example, the second  $(d_{1})$  and third  $(d_{2})$  column switches on bedside lamps. It is worth noting the presence of soft shadows, cast shadows, inter-reflections, and glossy changes across relights. Another important observation to note is that all relightings are with respect to world coordinates, not camera coordinates.
+
+![](images/d1bc21dc64fd1c6702e3e66077b3aaa869db1ebc164fc3a4c97c19c1a94e7b06.jpg)  
+Figure 6. Resurfacing Generated Images. Instead of relighting images, we can generate resurfaced images by swapping our consistency and diversity loss. We apply diversity loss to change the albedo and consistency loss to maintain the shading and global illumination. The first column shows images generated by the original StyleGAN, and the other columns show images obtained from  $\mathbf{w}^{+} + \mathbf{d}_{i}$  for our resurfacing or recoloring directions. Each column shows the same scene as in the first column, but with varying surface colors and materials, while the individual latent direction  $\mathbf{d}_i$  retains its semantics.
+
+comparable method. However, we show relighting comparisons to a recent SOTA method that is physically motivated and trained with CGI data [26] by using a SOTA inversion method [5] (Fig. 14 in Supplementary). For relighting, our
+
+method should generate images that: are clearly relightings of a scene; fix geometry and albedo but visibly change shading; and display complicated illumination effects, including soft shadows, cast shadows and gloss. For resurfacing, our
+
+![](images/2fd4b9a12e44e0c1910ee4e27fdfba0f92dd8940b5ac0a140ab0b22371286c64.jpg)  
+Figure 7. Controllability: The first and last columns of the figure show relighted images generated using our Relit-1 and Relit-5 directions. The bottom section of the figure features a user-controllable slider that enables adjusting the weight of the relighting effects produced by these two directions. Moving the slider from left to right results in a seamless interpolation between the two lighting directions and provides precise control over the relighting of the generated images.
+
+![](images/ea7eef6570ef1ec9bd1a3fc1856d73f98a7de1090b9b33827b3fb51fa5f55fae.jpg)  
+Figure 8. Scaling Directions: The figure depicts the persistent and smooth effects of applying the direction at different scalar coefficients. We use our Relit-2 direction. A slider at the bottom allows the user to adjust the weight of the relighting direction, producing a seamless interpolation when increasing or decreasing the intensity of the chosen direction. The relighting effects range from a well-lit room with the bedside lamp off to weak external lighting with a bedside lamp on.
+
+method should generate images that: are clearly images of the original layout, but with different materials or colors or changes in furniture; and display illumination effects that are consistent with these color changes. As Figure 6 shows, our method meets these goals. Figure 7 and Figure 8 show interpolation sequences for a relighting between two directions and scaling only one direction. Note that the lighting changes smoothly, as one would expect. Figure 15 in Supplementary shows that our relighting and recoloring directions are largely disentangled. An ablation for individual losses is also given in Supplementary Figure 13.
+
+Quantitative Evaluation: Figure 3 shows how we can evaluate albedo and lighting shifts. In Table 1 we show we can generate image datasets with increased FID [18, 34] (clean-FID) from the base comparison set. This is strong evidence our method can produce a set of images that is a strict superset of those that the vanilla StyleGAN can produce.
+
+Generality: We have applied our method to StyleGAN trained on Conference Room, Kitchen, Living Room, Dining Room, Church, and Face datasets (results in Figure 10).
+
+Comparison to GAN-control: GAN-control (GC) [39] represents lighting with a spherical harmonic predictor pre
+
+Table 1. FID measures distribution shift and not realism. Our generated images are realistic and are out-of-distribution because of large illumination and color changes in the images. This results in large FID scores.  $KDL$  in the table is for kitchen, dining and living room which are jointly trained [13].  
+
+<table><tr><td>Type</td><td>Bedroom</td><td>KDL</td><td>Conference</td><td>Church</td><td>Faces</td></tr><tr><td>StyleGAN (SG)</td><td>5.01</td><td>5.86</td><td>9.35</td><td>3.80</td><td>5.02</td></tr><tr><td>SG + Relighting (RL)</td><td>14.23</td><td>6.87</td><td>10.48</td><td>12.12</td><td>37.87</td></tr><tr><td>SG + Resurfacing (RS)</td><td>17.03</td><td>9.41</td><td>10.63</td><td>18.60</td><td>34.06</td></tr><tr><td>SG + RL + RS</td><td>21.39</td><td>11.68</td><td>12.71</td><td>21.08</td><td>37.40</td></tr></table>
+
+trained on a parameterized 3D face reconstruction model [9]. This model does not apply to indoor lighting (among other problems, it predicts all points with the same normal have the same shading). We trained a GC model on the LSUN Bedroom dataset with 2 subspaces  $z^k$  - illumination corresponding to spherical harmonic coefficients, and other to represent all other structural information. The model was trained for 800 epochs. GC produces images whose structure varies wildly with any lighting changes, resulting in large albedo changes (GC in Figures 3, 9). The difficulty appears to be that the attribute predictor is easily subverted; if the lighting representation cannot produce an image that is (say) dark on the left side, the albedo is adjusted instead.
+
+![](images/616e562904dbfc4791c0582ec40a42c8d2ed3c14537c46285ea6b0cee9bcadd9.jpg)  
+Figure 9. GAN-control (GC) [39] cannot relight complex scenes like bedrooms; it completely changes the scene's layout.
+
+![](images/136a1a252d87a1153f2c7be876b8d88a7fcd64fcff843b5203840378e6914d27.jpg)
+
+![](images/4ee88bae8989c9c8a895b5d15a0e25021316baa5246f7afedb1db4702bcb771b.jpg)  
+(c) Living Room Relighting
+
+![](images/41ebdd2d30f388edfd89eb5765d19b1696df76dcf09f9c25c465b7abb7791c2c.jpg)  
+(e) Relighting outdoor Churches  
+Figure 10. StyLitGAN extends to finding relighting directions for StyleGANs trained on other datasets.
+
+![](images/345544a3e5e24f28b193a84fae75a982292eb0e373c4b94c3494e93f648e7955.jpg)
+
+![](images/fdeb4c976808fe9c8c6d5bfed230d28b807fe438aab779b4a93459b9a4805f28.jpg)  
+(b) Kitchen Relighting
+
+![](images/86a2438cb584c617b683c965ca94228ad9a05d7ece2cf54a21e1e3ca5fad0ccf.jpg)  
+(d) Dining Room Relighting  
+(f) Portrait Relighting
+
+# 6. Downstream Applications
+
+Lighting Variance in Surface Normal Prediction: The Multilum dataset [29] provides images of 1000 various indoor scenes, each under 25 lighting conditions, physically relit and captured. A SOTA normal predictor (Omnidata [12, 21]) applied to the test set produces surface normal predictions that vary significantly with changes of light in a fixed scene (Figure 11, purple bar). Finetuning Omnidata with the Multilum training dataset significantly reduces this variance (Figure 11, orange bar); but multiple lightings of a fixed scene are very hard to find. Finetuning with StyLitGAN relights produces comparable improvements; using seven distinct relights (Figure 11, pale orange bar) is slightly
+
+![](images/b18315a4c356fb926ef337ffe428642de1853a27d7aba075bb8b6ab8715c80e0.jpg)  
+Figure 11. Normal variance to lighting reduces when fine-tuning on our relighting dataset. Purple boxplot shows normal variance under relighting for real test scenes from Multilum using the Omnidata normal predictor; orange shows the result of finetuning using Multilum training data; light orange and yellow show the result of finetuning using StyLitGAN images (7 and 25 per scene respectively). The measure is angular error in radians from the mean prediction of a scene for each relit image in the Multilum test set (30 scenes, 25 lightings each).
+
+![](images/648cf2ffc644a4acbef8476ef546b65980865d6af2bf979b6e9b5ecb93fd8ac4.jpg)  
+Figure 12. A surface normal predictor is finetuned to increase prediction consistency for relit images of the same scene. General performance shown on Taskonomy test images parallels that of the original model. Mean across buildings is given in the last column.
+
+worse than using 25 distinct relights (Figure 11, yellow bar). The resulting improvement is not at the cost of base accuracy. Figure 12 compares the accuracy of various finetuned models on the Taskonomy test set [50] (recall this involves thousands of frames each in 10 blocks; we show results by block). Note that finetuned methods mostly show slight accuracy improvements over the base model, but losses in some blocks result in means that match.
+
+In our follow-up work, we also show that StyleGAN "knows" intrinsic images and can be easily extracted [4].
+
+# Acknowledgment
+
+We thank Aniruddha Kembhavi, Derek Hoiem, Min Jin Chong and Shenlong Wang. This material is based upon work supported by the National Science Foundation under Grant No. 2106825 and by a gift from Boeing Corporation.
+
+# References
+
+[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2  
+[2] Peter N Belhumeur and David J Kriegman. What is the set of images of an object under all possible illumination conditions? International Journal of Computer Vision, 1998. 4  
+[3] Anand Bhattachad and D.A. Forsyth. Cut-and-paste object insertion by enabling deep image prior for reshading. In 2022 International Conference on 3D Vision (3DV). IEEE, 2022. 2, 3, 5  
+[4] Anand Bhattachad, Daniel McKee, Derek Hoiem, and David Forsyth. Stylegan knows normal, depth, albedo, and more. Advances in Neural Information Processing Systems, 36, 2024. 8  
+[5] Anand Bhattachad, Viraj Shah, Derek Hoiem, and David A Forsyth. Make it so: Steering stylegan for any image inversion and editing. arXiv preprint arXiv:2304.14403, 2023. 6, 1  
+[6] Prashanth Chandran, Sebastian Winberg, Gaspard Zoss, Jérémy Riviere, Markus Gross, Paulo Gotardo, and Derek Bradley. Rendering with style: combining traditional and neural approaches for high-quality face rendering. ACM Transactions on Graphics (ToG), 40(6):1-14, 2021. 2  
+[7] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. arXiv preprint arXiv:2112.11641, 2021. 3  
+[8] Min Jin Chong, Hsin-Ying Lee, and David Forsyth. Stylegan of all trades: Image manipulation with only pretrained stylegan. arXiv preprint arXiv:2111.01619, 2021. 2  
+[9] 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.  
+[10] Aditya Deshpande, Jiajun Lu, Mao-Chuang Yeh, Min Jin Chong, and David Forsyth. Learning diverse image colorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6837-6845, 2017. 2  
+[11] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 341-346, 2001. 2  
+[12] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 8  
+[13] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A Efros. Blobgan: Spatially disentangled scene representations. arXiv preprint arXiv:2205.02837, 2022. 2, 3, 7  
+[14] D.A. Forsyth and Jason J Rock. Intrinsic image decomposition using paradigms. TPAMI, 2022 in press. 2, 3, 5  
+[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and
+
+pattern recognition, pages 2414-2423, 2016. 2  
+[16] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. arXiv preprint arXiv:1406.2661, 2014. 2  
+[17] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 327-340, 2001. 2  
+[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2017. 7  
+[19] Le Hui, Jia Yuan, Mingmei Cheng, Jin Xie, Xiaoya Zhang, and Jian Yang. Superpoint network for point cloud oversegmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5510-5519, 2021. 4  
+[20] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, 2016. 3  
+[21] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 2, 8  
+[22] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2  
+[23] 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, 2019.  
+[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 1, 2  
+[25] Edwin H Land. The retina theory of color vision. Scientific american, 1977. 3  
+[26] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. arXiv preprint arXiv:2205.09343, 2022. 6, 1  
+[27] Zicheng Liao, Hugues Hoppe, David Forsyth, and Yizhou Yu. A subdivision-based representation for vector image editing. IEEE transactions on visualization and computer graphics, 2012. 2  
+[28] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. Advances in Neural Information Processing Systems, 34, 2021. 2  
+[29] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A multi-illumination dataset of indoor object appearance. In 2019 IEEE International Conference on Computer Vision (ICCV), Oct 2019. 2, 8  
+[30] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, Epic Games, Andreas Lehrmann, and AI Borealis. Learning physics-guided face relighting under directional light. 2020. 2
+
+[31] Xingang Pan, Xudong Xu, Chen Change Loy, Christian Theobalt, and Bo Dai. A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2  
+[32] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2  
+[33] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2  
+[34] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 7  
+[35] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. IEEE Computer graphics and applications, 21(5):34-41, 2001. 2  
+[36] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 2  
+[37] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2  
+[38] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2  
+[39] Alon Shoshan, Nadav Bhonker, Igor Kviatkovsky, and Gerard Medioni. Gan-control: Explicitly controllable gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14083-14093, 2021. 2, 5, 7, 8  
+[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. *ICLR*, 2015. 3  
+[41] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics, 2019. 2  
+[42] Feitong Tan, Sean Fanello, Abhimitra Meka, Sergio Orts-Escolano, Danhang Tang, Rohit Pandey, Jonathan Taylor, Ping Tan, and Yinda Zhang. Volux-gan: A generative model for 3d face synthesis with hdri relighting. arXiv preprint arXiv:2201.04873, 2022. 2  
+[43] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 2  
+[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2  
+[45] Andrey Voynov and Artem Babenko. Unsupervised discovery
+
+of interpretable directions in the gan latent space. In International conference on machine learning, pages 9786-9796. PMLR, 2020. 1, 2  
+[46] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 2  
+[47] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. arXiv preprint arXiv: 2101.05278, 2021. 2  
+[48] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. International Journal of Computer Vision, 2020. 2, 5  
+[49] Ning Yu, Guilin Liu, Aysegul Dundar, Andrew Tao, Bryan Catanzaro, Larry S Davis, and Mario Fritz. Dual contrastive loss and attention for gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6731-6742, 2021. 3  
+[50] Amir R. Zamir, Alexander Sax, William B. Shen, Leonidas J. Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2018. 8  
+[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3  
+[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 2  
+[53] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 1, 2  
+[54] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, 2017. 2

2024/StyLitGAN_ Image-Based Relighting via Latent Control/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76108c8f7346cfad3376e7288a5febac304c95a82d7f3ef8a7b80b74a8962f3f
+size 997540

2024/Style Aligned Image Generation via Shared Attention/0cae3168-e63f-4f71-908a-5bdeae8d8200_content_list.json

@@ -0,0 +1,2463 @@
+[
+    {
+        "type": "text",
+        "text": "Style Aligned Image Generation via Shared Attention",
+        "text_level": 1,
+        "bbox": [
+            214,
+            130,
+            756,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Amir Hertz\\*, Andrey Voynov\\*, Shlomi Fruchter\\*, and Daniel Cohen-Or\\*1,2  \n1 Google Research  \n2 Tel Aviv University",
+        "bbox": [
+            173,
+            179,
+            795,
+            234
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            268,
+            313,
+            284
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large-scale Text-to-Image (T2I) models have rapidly gained prominence across creative fields, generating visually compelling outputs from textual prompts. However, controlling these models to ensure consistent style remains challenging, with existing methods necessitating fine-tuning and manual intervention to disentangle content and style. In this paper, we introduce StyleAligned, a novel technique designed to establish style alignment among a series of generated images. By employing minimal 'attention sharing' during the diffusion process, our method maintains style consistency across images within T2I models. This approach allows for the creation of style-consistent images using a reference style through a straightforward inversion operation. Our method's evaluation across diverse styles and text prompts demonstrates high-quality synthesis and fidelity, underscoring its efficacy in achieving consistent style across various inputs.",
+        "bbox": [
+            75,
+            300,
+            473,
+            556
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            584,
+            209,
+            599
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large-scale Text-to-Image (T2I) generative models [37, 39, 45] have emerged as an essential tool across creative disciplines, such as art, graphic design, animation, architecture, gaming and more. These models show tremendous capabilities of translating an input text into an appealing visual result that is aligned with the input description.",
+        "bbox": [
+            75,
+            609,
+            468,
+            700
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "An envisioned application of T2I models revolves around the rendition of various concepts in a way that shares a consistent style and character, as though all were created by the same artist and method (see Fig. 1). While proficient in aligning with the textual description of the style, state-of-the-art T2I models often create images that diverge significantly in their interpretations of the same stylistic descriptor, as depicted in Fig. 2.",
+        "bbox": [
+            75,
+            700,
+            468,
+            820
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent methods mitigate this by fine-tuning the T2I model over a set of images that share the same style [13,49]. This optimization is computationally expensive and usually",
+        "bbox": [
+            75,
+            821,
+            468,
+            867
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/befe72dd08ea13e65a43000c1dc25189dcdb8c0c0bf181c36f354ee9aa00c776.jpg",
+        "image_caption": [
+            "Figure 1. Style aligned image set generation. By fusing the features of the toy train image (left) during the diffusion process, we can generate an image set of different content that shares the style."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            266,
+            893,
+            616
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "requires human input in order to find a plausible subset of images and texts that enables the disentanglement of content and style.",
+        "bbox": [
+            496,
+            696,
+            893,
+            742
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We introduce StyleAligned, a method that enables consistent style interpretation across a set of generated images (Fig. 1). Our method requires no optimization and can be applied to any attention-based text-to-image diffusion model. We show that adding minimal attention sharing operations along the diffusion process, from each generated image to the first one in a batch, leads to a style-consistent set. Moreover, using diffusion inversion, our method can be applied to generate style-consistent images given a reference style image, with no optimization or fine-tuning.",
+        "bbox": [
+            496,
+            742,
+            893,
+            893
+        ],
+        "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.",
+        "bbox": [
+            99,
+            875,
+            210,
+            886
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "Equal Advising.",
+        "bbox": [
+            99,
+            887,
+            197,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "4775",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/cca67eff1325ad439ed8bffe4784e7bbf26f6fce3d87078922c54c3404fd3b16.jpg",
+        "image_caption": [
+            "Figure 2. Standard text-to-image vs. StyleAligned set generation. Given style description of \"minimal origami\", standard text-to-image generation (top) results with an unaligned image set while our method (bottom) can generate variety of style aligned content."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            88,
+            468,
+            385
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We present our results over diverse styles and text prompts, demonstrating high-quality synthesis and fidelity to the prompts and reference style. We show diverse examples of generated images that share their style with a reference image that can possibly be a given input image. Importantly, our technique stands as a zero-shot solution, distinct from other personalization techniques, as it operates without any form of optimization or fine-tuning. For our code and more examples, please visit the project page style-aligned-gen.github.io",
+        "bbox": [
+            75,
+            491,
+            468,
+            643
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            656,
+            218,
+            672
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Text-to-image generation. Text conditioned image generative models [7, 31, 38] show unprecedented capabilities of generating high quality images from text descriptions. In particular, T2I diffusion models [35, 38, 46] are pushing the state of the art and they are quickly adopted for different generative visual tasks like inpainting [2, 44], imaged-to-image translation [55, 59], local image editing [9, 24], subject-driven image generation [42, 51] and more.",
+        "bbox": [
+            75,
+            681,
+            468,
+            801
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Attention control in diffusion models. Hertz et al. [17] have shown how cross and self-attention maps within the diffusion process determine the layout and content of the generated images. Moreover, they showed how the attention maps can be used for controlled image generation. Other studies have leveraged modifications in attention lay",
+        "bbox": [
+            75,
+            801,
+            468,
+            892
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "ers to enhance the fidelity or diversity of generated images [8, 34], or apply attention control for image editing [5, 12, 30, 32, 33, 53]. However, in contrast to prior approaches that primarily enable structure-preserving image editing, our method excels at generating images with diverse structures and content while maintaining a consistent style interpretation.",
+        "bbox": [
+            496,
+            90,
+            890,
+            196
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Style transfer. Transferring a style from a reference image to a target content image is well studied subject in computer graphics. Classic works [10, 11, 18, 26] rely on optimization of handcrafted features and texture resampling algorithms from an input texture image, combined with structure constrains of a content image. With the progress of deep learning research, another line of works utilizes deep neural priors for style transfer optimization using deep features of pre-trained networks [15, 52], or injecting attention features from a style image to a target one [1]. More related to our approach, Huang et al. [22] introduced a real time style transfer network based on Adaptive Instance Normalization layers (AdaIN) that are used to normalize deep features of a target image using deep features statistics of a reference style image. Follow-up works, employ the AdaIN layer for additional unsupervised learning tasks, like style-based image generation [25] and Image2Image translation [23, 29].",
+        "bbox": [
+            496,
+            196,
+            892,
+            468
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "T2I personalization To generalize T2I over new visual concepts, several works developed different optimization techniques over a small collection of input images that share the same concept [13, 16, 42, 56]. In instances where the collection shares a consistent style, the acquired concept becomes the style itself, affecting subsequent generations. Most close to our work is StyleDrop [49], a style personalization method that relies on fine-tuning of light weight adapter layers [20] at the end of each attention block in a non-autoregressive generative text-to-image transformer [7]. StyleDrop can generate a set of images in the same style of by training the adapter layers over a collection of images that share the same style. However, it struggles to generate a consistent image set of different content when training on a single image.",
+        "bbox": [
+            496,
+            468,
+            892,
+            694
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our method can generate a consistent image set without optimization phase and without relying on several images for training. To skip the training phase, recent works developed dedicated personalization encoders [14, 27, 47, 58, 59] that can directly inject new priors from a single input image to the T2I model. However, these methods encounter challenges to disentangle style from content as they focus on generating the same subject as in the input image.",
+        "bbox": [
+            496,
+            695,
+            890,
+            815
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3. Method overview",
+        "text_level": 1,
+        "bbox": [
+            500,
+            829,
+            668,
+            844
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In the following section we start with an overview of the T2I diffusion process, and in particular the self-attention mechanism Sec. 3.1. We continue by present",
+        "bbox": [
+            496,
+            854,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "4776",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/6bfec6c2924b31a525aed02432c3bbe38680437dfe55778583f7be18ac076611.jpg",
+        "image_caption": [
+            "Figure 3. Style Aligned Diffusion. Generation of images with a style aligned to the reference image on the left. In each diffusion denoising step all the images, except the reference, perform a shared self-attention with the reference image."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            76,
+            84,
+            872,
+            349
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/40d62b020e5800a0fb2d06cae575f19a45f0d7e49c04264b0542055e71d15eec.jpg",
+        "image_caption": [
+            "Figure 4. Shared attention layer. The target images attend to the reference image by applying AdaIN over their queries and keys using the reference queries and keys respectively. Then, we apply shared attention where the target features are updated by both the target values  $V_{t}$  and the reference values  $V_{r}$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            142,
+            410,
+            411,
+            720
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ing our attention-sharing operation within the self-attention layers that enable style aligned image set generation.",
+        "bbox": [
+            75,
+            825,
+            470,
+            858
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            500,
+            412,
+            640,
+            426
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Diffusion models [19, 48] are generative latent variable models that aim to model a distribution  $p_{\\theta}(x_0)$  that approximates the data distribution  $q(x_0)$  and are easy to sample from. Diffusion models are trained to reverse the diffusion \"forward process\":",
+        "bbox": [
+            498,
+            435,
+            892,
+            511
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nx _ {t} = \\sqrt {\\alpha_ {t}} x _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\epsilon \\sim N (0, I),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            557,
+            521,
+            831,
+            539
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $t \\in [0, \\infty)$  and the values of  $\\alpha_{t}$  are determined by a scheduler such that  $\\alpha_{0} = 1$  and  $\\lim_{t \\to \\infty} \\alpha_{t} = 0$ . During inference, we sample an image by gradually denoising an input noise image  $x_{T} \\sim \\mathcal{N}(0, I)$  via the reverse process:",
+        "bbox": [
+            498,
+            549,
+            893,
+            611
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nx _ {t - 1} = \\mu_ {t - 1} + \\sigma_ {t} z, z \\sim N (0, I),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            622,
+            815,
+            638
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where the value of  $\\sigma_{t}$  is determined by the sampler and  $\\mu_{t - 1}$  is given by",
+        "bbox": [
+            498,
+            648,
+            890,
+            680
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mu_ {t - 1} = \\frac {\\sqrt {\\alpha_ {t - 1}} x _ {t}}{\\sqrt {\\alpha_ {t}}} + \\left(\\sqrt {1 - \\alpha_ {t - 1}} - \\frac {\\sqrt {1 - \\alpha_ {t}}}{\\sqrt {\\alpha_ {t}}}\\right) \\epsilon_ {\\theta} (x _ {t}, t),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            500,
+            689,
+            890,
+            724
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\epsilon_{\\theta}(x_t,t)$  is the output of a diffusion model parameterized by  $\\theta$",
+        "bbox": [
+            498,
+            734,
+            890,
+            763
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Moreover, this process can be generalized for learning a marginal distribution using an additional input condition. That leads text-to-image diffusion models (T2I), where the output of the model  $\\epsilon_{\\theta}(x_t,t,y)$  is conditioned on a text prompt  $y$ .",
+        "bbox": [
+            498,
+            765,
+            890,
+            840
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Self-Attention in T2I diffusion models. State-of-the-art T2I diffusion models [4,35,46] employ a U-Net architecture [40] that consists of convolution layers and transformer attention blocks [54]. In these attention mechanisms, deep",
+        "bbox": [
+            498,
+            840,
+            893,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "4777",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/ba3e52212d6c2d820ad98b6bb7c6961b73723f21fa0dc4ed42a668f250f4f670.jpg",
+        "image_caption": [
+            "Figure 5. Ablation study - qualitative comparison. Each pair of rows shows two sets of images generated by the same set of prompts \"...in minimal flat design illustration\" using different configurations of our method, and each row in a pair uses a different seed. Sharing the self-attention between all images in the set (bottom) results with some diversity loss (style collapse across many seeds) and content leakage within each set (colors from one image leak to another). Disabling the queries keys AdALN operation results with less consistent image sets compared to our full method (top) which keeps on both diversity between different sets and consistency within each set."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            88,
+            472,
+            545
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "image features  $\\phi \\in \\mathbb{R}^{m\\times d_h}$  attend to each other via selfattention layers and to contextual text embedding via crossattention layers.",
+        "bbox": [
+            75,
+            734,
+            468,
+            781
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our work operates at the self-attention layers where deep features are being updated by attending to each other. First, the features are projected into queries  $Q \\in m \\times d_k$ , keys  $K \\in m \\times d_k$  and values  $V \\in m \\times d_h$  through learned linear layers. Then, the attention is computed by the scaled",
+        "bbox": [
+            76,
+            781,
+            470,
+            857
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "dot-product attention:",
+        "bbox": [
+            500,
+            90,
+            647,
+            104
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}} V\\right),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            545,
+            116,
+            844,
+            152
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $d_{k}$  is the dimension of  $Q$  and  $K$ . Intuitively, each image feature is updated by a weighted sum of  $V$ , where the weight depends on the correlation between the projected query  $q$  and the keys  $K$ . In practice, each self-attention layer consists of several attention heads, and then the residual is computed by concatenating and projecting the attention heads output back to the image feature space  $d_{h}$ :",
+        "bbox": [
+            496,
+            160,
+            893,
+            267
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\phi} = \\phi + \\text {M u l t i - H e a d - A t t e n t i o n} (\\phi).\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            277,
+            813,
+            295
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Style Aligned Image Set Generation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            306,
+            808,
+            321
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The goal of our method is to generate a set of images  $\\mathcal{I}_1\\ldots \\mathcal{I}_n$  that are aligned with an input set of text prompts  $y_{1}\\dots y_{n}$  and share a consistent style interpretation with each other. For example, see the garnered image set of toy objects in Fig. 3 that are style-aligned with each other and to the input text on top. A naive way to generate a style aligned image set of different content is to use the same style description in the text prompts. As can be seen at the bottom of Fig. 2, generating different images using a shared style description of \"in minimal origami style\" results in an unaligned set, since each image is unaware of the exact appearance of other images in the set during the generation process.",
+        "bbox": [
+            496,
+            329,
+            890,
+            525
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The key insight underlying our approach is the utilization of the self-attention mechanism to allow communication among various generated images. This is achieved by sharing attention layers across the generated images.",
+        "bbox": [
+            496,
+            526,
+            890,
+            585
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Formally, let  $Q_{i}$ ,  $K_{i}$ , and  $V_{i}$  be the queries, keys, and values, projected from deep features  $\\phi_{i}$  of  $\\mathcal{I}_{i}$  in the set, then, the attention update for  $\\phi_{i}$  is given by:",
+        "bbox": [
+            496,
+            585,
+            890,
+            632
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {A t t e n t i o n} \\left(Q _ {i}, K _ {1 \\dots n}, V _ {1 \\dots n}\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            598,
+            643,
+            890,
+            662
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\text {w h e r e} K _ {1 \\dots n} = \\left[ \\begin{array}{c} K _ {1} \\\\ K _ {2} \\\\ \\vdots \\\\ K _ {n} \\end{array} \\right] \\text {a n d} V _ {1 \\dots n} = \\left[ \\begin{array}{c} V _ {1} \\\\ V _ {2} \\\\ \\vdots \\\\ V _ {n} \\end{array} \\right]. \\text {H o w e v e r},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            500,
+            670,
+            890,
+            743
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "have noticed that by enabling full attention sharing, we may harm the quality of the generated set. As can be seen in Fig. 5 (bottom rows), full attention sharing results in content leakage among the images. For example, the unicorns got green paint from the garnered dino in the set. Moreover, full attention sharing results with less diverse sets of the same set of prompts, see the two sets in Fig. 5 in bottom rows compared to the sets above.",
+        "bbox": [
+            496,
+            746,
+            890,
+            866
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To restrict the content leakage and allow diverse sets, we share the attention to only one image in the generated set",
+        "bbox": [
+            500,
+            867,
+            890,
+            897
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "4778",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/925b1dcfbe4f36bc77e342a38b71d524bffa46834c75e5708976954f11d2d13e.jpg",
+        "image_caption": [
+            "Figure 6. Quantitative Comparison. We compare the results of the different methods (blue marks) and our ablation experiments (orange marks) in terms of text alignment (CLIP score) and set consistency (DINO embedding similarity)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            87,
+            454,
+            251
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/247a44494f02213f2b2b7c91ffa2bee923897c85a965aa4f01f500d88bcf6752.jpg",
+        "table_caption": [
+            "Table 1. User evaluation for style aligned image set generation. In each question, the user was asked to select between two image sets, Which is better in terms of style consistency and match to the text descriptions (see Sec. 4). We report the percentage of judgments in favor of StyleAligned over 800 answers (2400 in total)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>StyleDrop (unofficial MUSE)</td><td>StyleDrop (SDXL)</td><td>DreamBooth-LoRA (SDXL)</td></tr><tr><td>85.2 %</td><td>67.1 %</td><td>61.3%</td></tr></table>",
+        "bbox": [
+            96,
+            407,
+            444,
+            455
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "(the first in the batch). That is, target image features  $\\phi_t$  are attending to themselves and to the features of only one reference image in the set using Eq. 1. As can be seen in Fig. 5 (middle), sharing the attention to only one image in the set results in diverse sets that share a similar style. However, in that case, we have noticed that the style of different images is not well aligned. We suspect that this is due to low attention flow from the reference to the target image.",
+        "bbox": [
+            75,
+            469,
+            468,
+            589
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As illustrated in Fig. 4, to enable balanced attention reference, we normalize the queries  $Q_{t}$  and keys  $K_{t}$  of the target image using the queries  $Q_{r}$  and keys  $K_{r}$  of the reference image using the adaptive normalization operation (AdaIN) [22]:",
+        "bbox": [
+            75,
+            590,
+            468,
+            666
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {Q} _ {t} = \\operatorname {A d a I N} \\left(Q _ {t}, Q _ {r}\\right) \\quad \\hat {K} _ {t} = \\operatorname {A d a I N} \\left(K _ {t}, K _ {r}\\right),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            112,
+            676,
+            431,
+            695
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where the AdaIn operation is given by:",
+        "bbox": [
+            76,
+            705,
+            336,
+            720
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {A d a I N} (x, y) = \\sigma (y) \\left(\\frac {x - \\mu (x)}{\\sigma (x)}\\right) + \\mu_ {y},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            133,
+            732,
+            413,
+            765
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "and  $\\mu (x),\\sigma (x)\\in \\mathbb{R}^{d_k}$  are the mean and the standard deviation of queries and keys across different pixels. Finally, our shared attention is given by",
+        "bbox": [
+            75,
+            776,
+            468,
+            823
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {A t t e n t i o n} \\left(\\hat {Q} _ {t}, K _ {r t} ^ {T}, V _ {r t}\\right),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            187,
+            832,
+            357,
+            851
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $K_{rt} = \\left[ \\begin{array}{c}K_{r}\\\\ \\hat{K}_{t} \\end{array} \\right]$  and  $V_{rt} = \\left[ \\begin{array}{c}V_{r}\\\\ V_{t} \\end{array} \\right]$",
+        "bbox": [
+            76,
+            861,
+            326,
+            896
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Evaluations and Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            89,
+            769,
+            107
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We have implemented our method over Stable Diffusion XL (SDXL) [35] by applying our attention sharing overall 70 self-attention layers of the model. The generation of a four images set takes 29 seconds on a single A100 GPU. Notice that since the generation of the reference image is not influenced by other images in the batch, we can generate larger sets by fixing the prompt and seed of the reference image across the set generation.",
+        "bbox": [
+            496,
+            114,
+            890,
+            234
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For example, see the sets in Fig. 2 and 3.",
+        "bbox": [
+            517,
+            234,
+            792,
+            250
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Evaluation set. With the support of ChatGPT, we have generated 100 text prompts describing different image styles over four random objects. For example, “{A guitar, A hot air balloon, A sailboat, A mountain} in papercut art style.” For each style and set of objects, we use our method to generate a set of images. The full list of prompts is provided in the appendix.",
+        "bbox": [
+            496,
+            251,
+            890,
+            354
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Metrics. To verify that each generated image contains its specified object, we measure the CLIP similarity [36] between the image and the text description of the object. In addition, we evaluate the style consistency of each generated set, by measuring the pairwise average cosine similarity between DINO VIT-B/8 [6] embeddings of the generated images in each set. Following [41, 56], we used DINO embeddings instead of CLIP image embeddings for measuring image similarity, since CLIP was trained with class labels and therefore it might give a high score for different images in the set that have similar content but with a different style. On the other hand, DINO better distinguishes between different styles due to its self-supervised training.",
+        "bbox": [
+            496,
+            356,
+            890,
+            551
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            500,
+            560,
+            653,
+            575
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The quantitative results are summarized in Fig. 6, where the right-top place on the chart means better text similarity and style consistency, respectively. As a reference, we report the score obtained by generating the set of images using SDXL (T2I Reference) using the same seeds without any intervention. As can be seen, our method achieves a much higher style consistency score at the expense of text similarity. See qualitative comparison in Fig. 2",
+        "bbox": [
+            496,
+            583,
+            890,
+            703
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In addition, we compared our method to additional two variants of the shared attention as described in Sec. 3.2. The first variant uses full attention sharing (Full Attn. Share) where the keys and values are shared between each pair of images in the set. In the second variant (W.A. AdaIN) we omit the AdaIN operation over queries and keys. As expected, this Full Attn. Share variant, results with higher style consistency and lower text alignment. As can be seen in Fig. 5, Full Attn. Share harms the overall quality of the image sets and diversity across sets. Moreover, our method without the use of AdaIN results in much lower style consistency. Qualitative results can be seen in Fig. 5 and in the supplementary..",
+        "bbox": [
+            496,
+            704,
+            890,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "4779",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/90b4c6812a23bc64e0005d692fb5d98823a2a1d002f9f220cbd0b2c2e09b02ba.jpg",
+        "image_caption": [
+            "Reference image",
+            "A guitar"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            102,
+            200,
+            189
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/9182dc023ae4f02985381ae8250dd67f27a32d6b2304ba17eb545d23e84d78b2.jpg",
+        "image_caption": [
+            "StyleAligned",
+            "A hot air balloon"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            205,
+            103,
+            320,
+            189
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5aca3abfb4d09a6d3e5db3cba5ca2de4f2cc89a6531e18621a8d0d39d20800c7.jpg",
+        "image_caption": [
+            "A mountain"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            320,
+            103,
+            372,
+            189
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/c7948ce73d00ded74b3d54ff0bbc607a0b02b28b6b1d46b61271a0f2a65985d0.jpg",
+        "image_caption": [
+            "StyleDrop (SDXL)",
+            "A hot air balloon"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            375,
+            102,
+            488,
+            189
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/a5994e42c5975622b839fe5edb1ae76d3ccbb6ab0107ba916a23dc3bad6ba32b.jpg",
+        "image_caption": [
+            "A mountain"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            102,
+            542,
+            189
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5a0e47a4805c3521b0a200742e4d7a6efc924c7cb5a56dac468920f576d2922a.jpg",
+        "image_caption": [
+            "StyleDrop (MUSE unofficial)",
+            "A hot air balloon"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            547,
+            103,
+            658,
+            189
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/46b6a33a75f26a0b999c095607dae2c6762844b5d675645eb58f683a47db58dc.jpg",
+        "image_caption": [
+            "A mountain"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            658,
+            103,
+            712,
+            189
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/cf7b7ed8e76a13b03bed044b822d2266a9b6f5900b9b0d0dd4546a02b814c5f7.jpg",
+        "image_caption": [
+            "Dreambooth-LoRA (SDXL)",
+            "A hot air balloon"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            715,
+            103,
+            828,
+            189
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5df865f298975953ec22baa0012e6361e197e23bf7faab5934792aaa96698c2c.jpg",
+        "image_caption": [
+            "A mountain"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            831,
+            103,
+            883,
+            189
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/20e2aaaf0c1e6ab47123702fef3e6d7744bc6664e05e999232247491622d8124.jpg",
+        "image_caption": [
+            "A woman walking her dogs"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            198,
+            200,
+            282
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/2900abc58cb353417dbcacb0858899b10ac88bc6db44cd9e37deb49575279ad4.jpg",
+        "image_caption": [
+            "A friendly robot",
+            "Cherryblossom"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            205,
+            198,
+            320,
+            282
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/91f969d04a9a14bea047a577c124bb468e71870f92e46a74591a2accfd49efbf.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            320,
+            198,
+            372,
+            282
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/934d4e99253578a4646ad35ef4a76dc408e1ce411b4fae9ab69fc49920db8ff8.jpg",
+        "image_caption": [
+            "A friendly robot"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            375,
+            198,
+            488,
+            282
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/cddaf90b91fa69e42670e5a911d7462cb4857ff150b046a76dfa7ec3e37516d5.jpg",
+        "image_caption": [
+            "Cherryblossom"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            490,
+            198,
+            542,
+            282
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/2ecc2accac53d7b2fc69a9b04e6409aabeb20377e722c3e746a29ad3763ec90b.jpg",
+        "image_caption": [
+            "A robot"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            549,
+            198,
+            658,
+            282
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/eb1f356d8d9dbb4832264617ad7181e028801061eb592887032ae1c4e14a004b.jpg",
+        "image_caption": [
+            "Cherryblos"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            658,
+            198,
+            712,
+            282
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/4a444f5f41709352b61c2f29db3900fe49fb1982d985bb2a52e3d55197d629c5.jpg",
+        "image_caption": [
+            "A friendly robot",
+            "Cherryblossom"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            715,
+            198,
+            828,
+            282
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/610c9ba3314ca1226805491b5a9ce683756ab16a5b5f16aeebefaf025038cf13.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            828,
+            198,
+            883,
+            282
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/8840c5026e15cd714d61ba2d56d55ab5f9290a381cb36c9d92013488ec9d9cd7.jpg",
+        "image_caption": [
+            "A duck"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            294,
+            200,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/27d3adda1e9c1e8abb94cd5d3919471f7d4377e42f229c39b65589ea91c646c5.jpg",
+        "image_caption": [
+            "A ship"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            205,
+            294,
+            320,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5a84f2c0a5fbafa354a5977511fe2a587e3984e96500383873d23ee76b6ca112.jpg",
+        "image_caption": [
+            "A rocket"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            320,
+            294,
+            372,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/7f110362dbb4193ed2c122926456ca32cf0ebff2b338a2f79b2ba1a19ea9abe1.jpg",
+        "image_caption": [
+            "A ship"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            375,
+            294,
+            488,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/e719c9a00052dcead3a2edfa21f5306cf25728b1e523e32e19639193911e3f21.jpg",
+        "image_caption": [
+            "A rocket"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            490,
+            294,
+            542,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/520144230c3f6cda53ff73278bfc318dc00a4a7b4f25acc8f6aad269a22713e0.jpg",
+        "image_caption": [
+            "A ship"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            549,
+            294,
+            658,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/f30b62b042d2e1a7bc31ea8397e082c8f27c08c36af31b9f8e439353ddb0b106.jpg",
+        "image_caption": [
+            "A car"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            658,
+            294,
+            712,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/c120d4a94419eaaf18f09eb87f57ff0e51bc783cbfc3b7df039a20de156ed2fa.jpg",
+        "image_caption": [
+            "A ship"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            715,
+            294,
+            828,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/9d94bbdfcfdd7eeb1b3052b31dc2b9ffcd4e5c7236d8b171d9ae33c032af0cff.jpg",
+        "image_caption": [
+            "A rocket"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            828,
+            294,
+            883,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d7ddbb290dfd3012f5aa02c56956acb097eef5b8f80243a3deaff12c12b286c4.jpg",
+        "image_caption": [
+            "A cute bear"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            387,
+            200,
+            474
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/eb8f5046cf9749c732d4fd406a4560c5b3b793e5ae5ad161bdccdc6d44f34b1b.jpg",
+        "image_caption": [
+            "A moose"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            205,
+            388,
+            320,
+            474
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/101e1170f7064b5edc9f04d96143713764a19191953362df139e3913fa60e039.jpg",
+        "image_caption": [
+            "A baby pen"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            320,
+            388,
+            372,
+            474
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/0f3d2fa0c62c9e0e5edf09ed204fd59daf986792ce49f36c091cb0568c688e05.jpg",
+        "image_caption": [
+            "A moose",
+            "A baby penguin"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            375,
+            390,
+            488,
+            474
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/1e88d7841435824dd9ec845f5d2d1d4131c53799ec14308bbce9203f29322916.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            490,
+            390,
+            542,
+            474
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/066e32b3d3e53446c8e21b056da640992280644c433de7ec203d5b58ab4d1c94.jpg",
+        "image_caption": [
+            "A moose",
+            "A baby penguin"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            549,
+            390,
+            658,
+            474
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d493c3f5b38c916496d273246a57c83b152ee12941106b89e07e94b18f98934f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            658,
+            390,
+            712,
+            474
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/98bd1600692848c1d1e32c87ffbbaad7e8831f1325c1d591c3d4836fded90ab6.jpg",
+        "image_caption": [
+            "A moose",
+            "A baby penguin"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            715,
+            390,
+            828,
+            474
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/07dab4a6b25a0ac834d7e5208ec8f5ee690714b8ca83182f7753e0c5f5c91b88.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            828,
+            390,
+            883,
+            474
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/0e75922bafab63fd0dc43b2a236d62036d8df5d8480aa1a4c0b98594e1613c20.jpg",
+        "image_caption": [
+            "A camera"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            484,
+            200,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/fde80ce0ea8f158811b6028048d98b0700eaf3363d4247ad20510ff14bdce6ce.jpg",
+        "image_caption": [
+            "A fireman"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            205,
+            484,
+            320,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/3613566046215aec7374cc7ae12d282ad8e812100a330aa8355d1ff5806af360.jpg",
+        "image_caption": [
+            "A tall hill"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            320,
+            484,
+            372,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/687587c656d7c57e108c490ea6e0f9a58ca98ec07e812b9a7d6dd6c239aad61e.jpg",
+        "image_caption": [
+            "A fireman"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            375,
+            484,
+            488,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5f858526fbb5f99a5735067d56df198d3cd99d2fc790db7840d84b2fe9da5d37.jpg",
+        "image_caption": [
+            "A tall hill"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            490,
+            484,
+            542,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/a5ecc311b5dbeb6b2baedaef213a0d6b54334591474cb728be06ab329d4edeef.jpg",
+        "image_caption": [
+            "A fireman"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            549,
+            484,
+            658,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5f0710c2366220dd26304908de510520344952d40686136117fcf3bf7c907660.jpg",
+        "image_caption": [
+            "A tall hill"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            658,
+            484,
+            712,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/de633264efa8f257c089060d149edbe67c2102974cb50c9648219b00e7729e36.jpg",
+        "image_caption": [
+            "A fireman"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            715,
+            484,
+            828,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/848ac7555642c0cea82a2611ee7dab8c5f40775618c28cab35f1d537b3a2204a.jpg",
+        "image_caption": [
+            "A tall hill"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            828,
+            484,
+            883,
+            570
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/218c906898b3a5ae3af8a0c59cbfe3f2e6a876f559a6ed9dbed5f74cff59aca5.jpg",
+        "image_caption": [
+            "A saxophone"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            578,
+            200,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5a326baa1c71afd4fa92e7b945445d5df1a917fb68392dc96bde8df9060c0670.jpg",
+        "image_caption": [
+            "A sunflower"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            205,
+            578,
+            320,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/2e24b65414432bef1e741090cb6d689be11819594a2cdbe6f42fe4398c0cb96f.jpg",
+        "image_caption": [
+            "A compass"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            320,
+            578,
+            372,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/3bd7a2838367c664a9e337c39d2c1f17bc4b4d674a887d56572386b522ef23a5.jpg",
+        "image_caption": [
+            "A sunflower"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            375,
+            578,
+            488,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/a909ea9a7f355a26af4d2503b0d77e6884a4592309406b7b8a2b5cbf9376bc47.jpg",
+        "image_caption": [
+            "A compass"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            490,
+            578,
+            542,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/245f0efac9754680077bce0f2084c5a09d0700a318b35ddbc7df9e82dd856f22.jpg",
+        "image_caption": [
+            "A sunflower"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            549,
+            578,
+            658,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/1b207c7f42fabcce392a07367ce06138d3678cf13f55abd0b1abdda1ce539d27.jpg",
+        "image_caption": [
+            "A compass"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            658,
+            578,
+            712,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/0e392ae6870353c15e4c33cb5e500f566f4a5367ee5622f56e67fdfb217bce2f.jpg",
+        "image_caption": [
+            "A sunflower"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            715,
+            578,
+            828,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/2a20746fe4b8fc5b42a361f53ef12c4712aebf9df993f31740b2c8624cbe67e5.jpg",
+        "image_caption": [
+            "A compass"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            828,
+            578,
+            883,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/b266b1b05456e22c05ece494dafbad1df540c9dd01234ce23f0846dcfdab7187.jpg",
+        "image_caption": [
+            "A bottle of wine"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            675,
+            200,
+            760
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/0cc1655eb494e44e0a04cd133ad64293c8f42e2d1a50b974f0592bdb3ef10b6a.jpg",
+        "image_caption": [
+            "Scones"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            205,
+            675,
+            320,
+            760
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/c0febaa1d6f4af1489962cdf2c022e226c99f50d6390739b462f03a7e5894d83.jpg",
+        "image_caption": [
+            "A muffin"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            320,
+            675,
+            372,
+            760
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/a67f1819a5ac9832c842ffec3c5c2bea10ef041df432dc1197e1a531b7a6f866.jpg",
+        "image_caption": [
+            "Scones"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            375,
+            675,
+            488,
+            760
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/273b7d99f4027b80fb94d2a712e1526dccae55b952daf95ebaf79781277e4638.jpg",
+        "image_caption": [
+            "A muffin"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            490,
+            675,
+            542,
+            760
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/9729ac558578c2cc718793e23350acb7c4e090b725969a2dc8c12431c4a194f0.jpg",
+        "image_caption": [
+            "A fireman"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            549,
+            675,
+            658,
+            760
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/8fd9cff8bb4ccfef3049b2a2e4eb07b940bcfdadfe156730a753ba68f0cf99eb.jpg",
+        "image_caption": [
+            "A muffin"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            658,
+            675,
+            712,
+            760
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/3bc108ed9e3f17a44c0e1c114055d5844ede0e28a61f0b1c4879b10da89fb618.jpg",
+        "image_caption": [
+            "Scones"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            715,
+            675,
+            828,
+            760
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/e20646a086cffac4ae116315a50305af6784ccc6c59bd2082f4bdeee1ee46165.jpg",
+        "image_caption": [
+            "A muffin"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            828,
+            675,
+            883,
+            760
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/2fa74eae159321dd188194042693392198fcb61858fa854a7b9382491534aef9.jpg",
+        "image_caption": [
+            "A wise owl"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            770,
+            200,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/4f563e863ce68571d9fa9ceae9bb07ae33f3fb60e146a0dab8488c410f327a6e.jpg",
+        "image_caption": [
+            "Full moon",
+            "Figure 7. Qualitative comparison to personalization based methods."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            205,
+            770,
+            320,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/39fa8d35e7e3ba2427e87d43ffa4924cc93aae4c1d0d493c156b7c5d1acb4cdd.jpg",
+        "image_caption": [
+            "A book"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            320,
+            770,
+            372,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/2bb7b8d91f4af638659972a0ca014ec10d525a15fc9566d0869f376937cbc8a6.jpg",
+        "image_caption": [
+            "Full moon"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            375,
+            770,
+            488,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/bc6a9fbaf50c348093b27ffe01d00455c1a6a322d45a787eed6f99fd26ddc061.jpg",
+        "image_caption": [
+            "A book"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            490,
+            770,
+            542,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d5733e81801d09ddb91dd13aa468c0647906a7565c02d87e74e2a016c17152db.jpg",
+        "image_caption": [
+            "Full moon"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            549,
+            770,
+            658,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/3a71945b845f2970193a4628958143b266d4488bfa7f69cf7b4ced283b29b9c6.jpg",
+        "image_caption": [
+            "A book"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            658,
+            770,
+            712,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/ec4bb150faf2d7acfdb5c2bcb832dccc3d4f65619f949822f0f87ff929759a87.jpg",
+        "image_caption": [
+            "Full moon"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            715,
+            770,
+            828,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/991f0ea54f6cbd853d334702a4074807d20a62f8c656ffa058f21edf2e85dc5d.jpg",
+        "image_caption": [
+            "A book"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            828,
+            770,
+            883,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "4780",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/ada735eb1902ca79d5ab11a26e4fe25d02bb9659ddec19de02a6b75738006b8c.jpg",
+        "image_caption": [
+            "Figure 8. Varying level of attention sharing. By reducing the number of shared attention layers, i.e., allowing only self-attention in part of the layers, we can get more varied results (bottom rows) at the expense of style alignment (top row)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            89,
+            88,
+            468,
+            412
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2. Comparisons",
+        "text_level": 1,
+        "bbox": [
+            76,
+            491,
+            217,
+            507
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "For baselines, we compare our method to T2I personalization methods. We trained StyleDrop [49] and DreamBooth [41] over the first image in each set of our evaluation data, and use the trained personalized weights to generate the additional three images in each set. We use a public unofficial implementation of StyleDrop * (SDRP-unofficial) over non-regressive T2I model. Due to the large quality gap between the unofficial MUSE model † to the official one [7], we follow StyleDrop and implement an adapter model over SDXL (SDRP-SDXL) *, where we train a low rank linear layer after each Feed-Forward layer at the model's attention blocks. For training DreamBooth, we adapt the LoRA [21, 43] variant (DB-LoRA) over SDXL using the public huggingface-diffusers implementation ‡. We follow the hyperparameters tuning reported in [49] and train both SDRP-SDXL and DB-LoRA for 400 steps to prevent overfitting to the style training image.",
+        "bbox": [
+            75,
+            513,
+            468,
+            770
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As can be seen in the qualitative comparison, Fig. 7, the image sets generated by our method, are more consistent across style attributes like color palette, drawing style, and composition. Moreover, the personalization-based methods",
+        "bbox": [
+            75,
+            771,
+            470,
+            832
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "may leak the content of the training reference image (on the left) when generating the new images. For example, see the repeated woman and dogs in the results of DB-LoRA and SDRP-SDXL at the second row or the repeated owl at the bottom row. Similarly, because of the content leakage, these methods obtained lower text similarity scores and higher set consistency scores compared to our method.",
+        "bbox": [
+            500,
+            90,
+            890,
+            196
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We also apply two encoder-based personalization methods ELITE [57], IP-Adapter [59], and BLIP-Diffusion [27] over our evaluation set. These methods receive as input the first image in each set and use its embeddings to generate images with other content. Unlike the optimization-based techniques, these methods operate in a much faster feedforward diffusion loop, like our method. However, as can be seen in Fig. 6, their performance for style aligned image generation is poor compared to the other baselines. We argue that current encoder-based personalization techniques struggle to disentangle the content and the style of the input image. We supply qualitative results in the supplementary.",
+        "bbox": [
+            496,
+            196,
+            892,
+            378
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "User study. In addition to the automatic evaluation, we conducted a user study over the results of our method, StyleDrop (unofficial MUSE), StyleDrop (SDXL), and DreamBooth-LoRA (SDXL). In each question, we randomly sample one of the evaluation examples and show",
+        "bbox": [
+            498,
+            378,
+            890,
+            453
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/68ce1e879762b8e47f9f2caa41262a4af139d672c99f321df933fc63e9d4f341.jpg",
+        "image_caption": [
+            "Figure 9. Style aligned image generation to an input image. Given an input reference image (left column) and text description, we first apply DDIM inversion over the image to get the inverted diffusion trajectory  $x_{T}$ ,  $x_{T-1} \\ldots x_{0}$ . Then, starting from  $x_{T}$  and a new set of prompts, we apply our method to generate new content (right columns) with an aligned style to the input."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            501,
+            476,
+            890,
+            791
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "\\*github.com/aim-uofa/StyleDrop-PyTorch",
+        "bbox": [
+            99,
+            840,
+            401,
+            853
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "github.com/baaivision/MUSE-Pytorch",
+        "bbox": [
+            102,
+            854,
+            377,
+            864
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "\\*github.com/amirhertz/smlx_adapter",
+        "bbox": [
+            102,
+            864,
+            367,
+            876
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "$\\ddagger$ github.com/huggingface/diffusers",
+        "bbox": [
+            102,
+            877,
+            359,
+            888
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "4781",
+        "bbox": [
+            482,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "the user the 4 image set that resulted from our and another method (in a random order). The user had to choose which set is better in terms of style consistency, and text alignment. A print screen of the user study format is provided in the supplementary. Overall, we collected 2400 answers from 100 users using the Amazon Mechanical Turk service. The results are summarized in Tab. 1 where for each method, we report the percentage of judgments in our favor. As can be seen, most participants favored our method by a large margin. More information about our user study can be found in the supplementary.",
+        "bbox": [
+            75,
+            90,
+            472,
+            257
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.3. Additional Results",
+        "text_level": 1,
+        "bbox": [
+            76,
+            266,
+            256,
+            279
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Style alignment control. We provide means of control over the style alignment to the reference image by applying the shared attention over only part of the self-attention layers. As can be seen in Fig. 8, reducing the number of shared attention layers results with a more diverse image set, which still shares common attributes with the reference image.",
+        "bbox": [
+            75,
+            289,
+            468,
+            378
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "StyleAligned from an input image. To generate style-aligned images to an input image, we apply DDIM inversion [50] using a provided text caption. Then, we apply our method to generate new images in the style of the input using the inverted diffusion trajectory  $x_{T}, x_{T-1}, \\ldots, x_{0}$  for the reference image. Examples are shown in Fig. 9, where we use BLIP captioning [28] to get a caption for each input image. For example, we used the prompt \"A render of a house with a yellow roof\" for the DDIM inversion of the top example and replaced the word house with other objects to generate the style-aligned images of a car, a cat, and a cactus. Notice that this method does not require any optimization. However, DDIM inversion may fail [30] or results with an erroneous trajectory [24]. More results and analysis, are provided in the supplementary.",
+        "bbox": [
+            75,
+            378,
+            468,
+            604
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "StyleAligned with other methods. Since our method doesn't require training or optimization, it can be easily combined on top of other diffusion based methods to generate style-consistent image sets. Fig. 10 shows several such examples where we combine our method with ControlNet [60], DreamBooth [42] and MultiDiffusion [3]. More examples and details about the integration of StyleAligned with other methods can be found in supplementary.",
+        "bbox": [
+            75,
+            606,
+            468,
+            728
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusions",
+        "text_level": 1,
+        "bbox": [
+            76,
+            739,
+            204,
+            753
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We have presented StyleAligned, which addresses the challenge of achieving style-aligned image generation within the realm of large-scale Text-to-Image models. By introducing minimal attention sharing operations with AdaIN modulation during the diffusion process, our method successfully establishes style-consistency and visual coherence across generated images. The demonstrated efficacy of StyleAligned in producing high-quality, style-consistent images across diverse styles and textual prompts under",
+        "bbox": [
+            75,
+            763,
+            470,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/48eab7e15d523b557b2e445a5c7f8acebe22b8794caccc6a4756084a368b7c75.jpg",
+        "image_caption": [
+            "Figure 10. StyleAligned with other methods. On top, StyleAligned is combined with ControlNet to generate style-aligned images conditioned on depth maps. In the middle, our method combined with MultiDiffusion to generate panorama images that share multiple styles. On the bottom, style consistent and personalized content created by combining our method with pretrained personalized DreamBooth-LoRA models."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            85,
+            893,
+            529
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "scores its potential in creative domains and practical applications. Our results affirm StyleAligned capability to faithfully adhere to provided descriptions and reference styles while maintaining impressive synthesis quality.",
+        "bbox": [
+            496,
+            650,
+            890,
+            710
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In the future we would like to explore the scalability and adaptability of StyleAligned to have more control over the shape and appearance similarity among the generated images. Additionally, due to the limitation of current diffusion inversion methods, a promising direction is to leverage StyleAligned to assemble a style-aligned dataset which then can be used to train style condition text-to-image models.",
+        "bbox": [
+            496,
+            710,
+            890,
+            816
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            500,
+            830,
+            679,
+            847
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We thank Or Patashnik, Matan Cohen, Yael Pritch, and Yael Vinker for their valuable inputs that helped improve this work.",
+        "bbox": [
+            496,
+            853,
+            890,
+            895
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "4782",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            89,
+            173,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer, 2023. 2",
+            "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 42(4), jul 2023. 2",
+            "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 8",
+            "[4] James Betker, Gabriel Goh, Li Jing, TimBrooks, Jianfeng Wang, Linjie Li, LongOuyang, JuntangZhuang, JoyceLee, YufeiGuo, WesamManassra, PrafullaDhariwal, CaseyChu, YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 3",
+            "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560–22570, October 2023. 2",
+            "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5",
+            "[7] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming Yang, Kevin P. Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In International Conference on Machine Learning, 2023. 2, 7",
+            "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42:1 - 10, 2023. 2",
+            "[9] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In The Eleventh International Conference on Learning Representations, 2022. 2",
+            "[10] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 571-576. 2023. 2",
+            "[11] Alexei A Efros and Thomas K Leung. Texture synthesis by non-parametric sampling. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 1033-1038. IEEE, 1999. 2",
+            "[12] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 2",
+            "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2022. 1, 2",
+            "[14] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2",
+            "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2",
+            "[16] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris N. Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. ArXiv, abs/2303.11305, 2023. 2",
+            "[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2",
+            "[18] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 557-570. 2023. 2",
+            "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proc. NeurIPS, 2020. 3",
+            "[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2",
+            "[21] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 7",
+            "[22] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 5",
+            "[23] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 2",
+            "[24] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. arXiv preprint arXiv:2304.06140, 2023. 2, 8",
+            "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2",
+            "[26] Hochang Lee, Sanghyun Seo, Seungtaek Ryoo, and Kyunghyun Yoon. Directional texture transfer. In Proceedings of the 8th International Symposium on Non-Photorealistic Animation and Rendering, pages 43–48, 2010. 2",
+            "[27] Dongxu Li, Junnan Li, and Steven C. H. Hoi. BLIP-Diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. ArXiv, abs/2305.14720, 2023. 2, 7",
+            "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 8"
+        ],
+        "bbox": [
+            503,
+            92,
+            893,
+            885
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "4783",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[29] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In IEEE International Conference on Computer Vision (ICCV), 2019. 2",
+            "[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2, 8",
+            "[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 2",
+            "[32] Dong Huk Park*, Grace Luo*, Clayton Toste, Samaneh Azadi, Xihui Liu, Maka Karalashvili, Anna Rohrbach, and Trevor Darrell. Shape-guided diffusion with inside-outside attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2",
+            "[33] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2",
+            "[34] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. ArXiv, abs/2303.11306, 2023. 2",
+            "[35] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 3, 5",
+            "[36] 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 International Conference on Machine Learning, 2021. 5",
+            "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1",
+            "[38] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 2",
+            "[39] 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. 1",
+            "[40] 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,"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            893
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3",
+            "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 5, 7",
+            "[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 2, 8",
+            "[43] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2022.7",
+            "[44] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. ACM SIGGRAPH 2022 Conference Proceedings, 2021. 2",
+            "[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1",
+            "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2, 3",
+            "[47] Jing Shi, Wei Xiong, Zhe L. Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. ArXiv, abs/2304.03411, 2023. 2",
+            "[48] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3",
+            "[49] Kihyuk Sohn, Nataniel Ruiz, Kimin Lee, Daniel Castro Chin, Irina Blok, Huiwen Chang, Jarred Barber, Lu Jiang, Glenn Entis, Yuanzhen Li, et al. Styledrop: Text-to-image generation in any style. arXiv preprint arXiv:2306.00983, 2023. 1, 2, 7",
+            "[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 8",
+            "[51] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. SIGGRAPH 2023 Conference Proceedings, 2023. 2",
+            "[52] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 2"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            893
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "4784",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[53] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2",
+            "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3",
+            "[55] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. arXiv preprint arXiv:2211.13752, 2022. 2",
+            "[56] Andrey Voynov, Q. Chu, Daniel Cohen-Or, and Kfir Aberman. P+: Extended textual conditioning in text-to-image generation. ArXiv, abs/2303.09522, 2023. 2, 5",
+            "[57] Yuxiang Wei. Official implementation of ELITE. https://github.com/csyxwei/ELITE, 2023. 7",
+            "[58] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: Encoding visual concepts into textual embeddings for customized text-to-image generation. ArXiv, abs/2302.13848, 2023. 2",
+            "[59] Hu Ye, Jun Zhang, Siyi Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. ArXiv, abs/2308.06721, 2023. 2, 7",
+            "[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 8"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            527
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "4785",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2024/Style Aligned Image Generation via Shared Attention/0cae3168-e63f-4f71-908a-5bdeae8d8200_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fac94e9345423fc93f0ae9927fb2ed3c56b1c790779364354863699ed0e86eed
+size 9245834

2024/Style Aligned Image Generation via Shared Attention/full.md

@@ -0,0 +1,499 @@
+# Style Aligned Image Generation via Shared Attention
+
+Amir Hertz\*, Andrey Voynov\*, Shlomi Fruchter\*, and Daniel Cohen-Or\*1,2  
+1 Google Research  
+2 Tel Aviv University
+
+# Abstract
+
+Large-scale Text-to-Image (T2I) models have rapidly gained prominence across creative fields, generating visually compelling outputs from textual prompts. However, controlling these models to ensure consistent style remains challenging, with existing methods necessitating fine-tuning and manual intervention to disentangle content and style. In this paper, we introduce StyleAligned, a novel technique designed to establish style alignment among a series of generated images. By employing minimal 'attention sharing' during the diffusion process, our method maintains style consistency across images within T2I models. This approach allows for the creation of style-consistent images using a reference style through a straightforward inversion operation. Our method's evaluation across diverse styles and text prompts demonstrates high-quality synthesis and fidelity, underscoring its efficacy in achieving consistent style across various inputs.
+
+# 1. Introduction
+
+Large-scale Text-to-Image (T2I) generative models [37, 39, 45] have emerged as an essential tool across creative disciplines, such as art, graphic design, animation, architecture, gaming and more. These models show tremendous capabilities of translating an input text into an appealing visual result that is aligned with the input description.
+
+An envisioned application of T2I models revolves around the rendition of various concepts in a way that shares a consistent style and character, as though all were created by the same artist and method (see Fig. 1). While proficient in aligning with the textual description of the style, state-of-the-art T2I models often create images that diverge significantly in their interpretations of the same stylistic descriptor, as depicted in Fig. 2.
+
+Recent methods mitigate this by fine-tuning the T2I model over a set of images that share the same style [13,49]. This optimization is computationally expensive and usually
+
+![](images/befe72dd08ea13e65a43000c1dc25189dcdb8c0c0bf181c36f354ee9aa00c776.jpg)  
+Figure 1. Style aligned image set generation. By fusing the features of the toy train image (left) during the diffusion process, we can generate an image set of different content that shares the style.
+
+requires human input in order to find a plausible subset of images and texts that enables the disentanglement of content and style.
+
+We introduce StyleAligned, a method that enables consistent style interpretation across a set of generated images (Fig. 1). Our method requires no optimization and can be applied to any attention-based text-to-image diffusion model. We show that adding minimal attention sharing operations along the diffusion process, from each generated image to the first one in a batch, leads to a style-consistent set. Moreover, using diffusion inversion, our method can be applied to generate style-consistent images given a reference style image, with no optimization or fine-tuning.
+
+![](images/cca67eff1325ad439ed8bffe4784e7bbf26f6fce3d87078922c54c3404fd3b16.jpg)  
+Figure 2. Standard text-to-image vs. StyleAligned set generation. Given style description of "minimal origami", standard text-to-image generation (top) results with an unaligned image set while our method (bottom) can generate variety of style aligned content.
+
+We present our results over diverse styles and text prompts, demonstrating high-quality synthesis and fidelity to the prompts and reference style. We show diverse examples of generated images that share their style with a reference image that can possibly be a given input image. Importantly, our technique stands as a zero-shot solution, distinct from other personalization techniques, as it operates without any form of optimization or fine-tuning. For our code and more examples, please visit the project page style-aligned-gen.github.io
+
+# 2. Related Work
+
+Text-to-image generation. Text conditioned image generative models [7, 31, 38] show unprecedented capabilities of generating high quality images from text descriptions. In particular, T2I diffusion models [35, 38, 46] are pushing the state of the art and they are quickly adopted for different generative visual tasks like inpainting [2, 44], imaged-to-image translation [55, 59], local image editing [9, 24], subject-driven image generation [42, 51] and more.
+
+Attention control in diffusion models. Hertz et al. [17] have shown how cross and self-attention maps within the diffusion process determine the layout and content of the generated images. Moreover, they showed how the attention maps can be used for controlled image generation. Other studies have leveraged modifications in attention lay
+
+ers to enhance the fidelity or diversity of generated images [8, 34], or apply attention control for image editing [5, 12, 30, 32, 33, 53]. However, in contrast to prior approaches that primarily enable structure-preserving image editing, our method excels at generating images with diverse structures and content while maintaining a consistent style interpretation.
+
+Style transfer. Transferring a style from a reference image to a target content image is well studied subject in computer graphics. Classic works [10, 11, 18, 26] rely on optimization of handcrafted features and texture resampling algorithms from an input texture image, combined with structure constrains of a content image. With the progress of deep learning research, another line of works utilizes deep neural priors for style transfer optimization using deep features of pre-trained networks [15, 52], or injecting attention features from a style image to a target one [1]. More related to our approach, Huang et al. [22] introduced a real time style transfer network based on Adaptive Instance Normalization layers (AdaIN) that are used to normalize deep features of a target image using deep features statistics of a reference style image. Follow-up works, employ the AdaIN layer for additional unsupervised learning tasks, like style-based image generation [25] and Image2Image translation [23, 29].
+
+T2I personalization To generalize T2I over new visual concepts, several works developed different optimization techniques over a small collection of input images that share the same concept [13, 16, 42, 56]. In instances where the collection shares a consistent style, the acquired concept becomes the style itself, affecting subsequent generations. Most close to our work is StyleDrop [49], a style personalization method that relies on fine-tuning of light weight adapter layers [20] at the end of each attention block in a non-autoregressive generative text-to-image transformer [7]. StyleDrop can generate a set of images in the same style of by training the adapter layers over a collection of images that share the same style. However, it struggles to generate a consistent image set of different content when training on a single image.
+
+Our method can generate a consistent image set without optimization phase and without relying on several images for training. To skip the training phase, recent works developed dedicated personalization encoders [14, 27, 47, 58, 59] that can directly inject new priors from a single input image to the T2I model. However, these methods encounter challenges to disentangle style from content as they focus on generating the same subject as in the input image.
+
+# 3. Method overview
+
+In the following section we start with an overview of the T2I diffusion process, and in particular the self-attention mechanism Sec. 3.1. We continue by present
+
+![](images/6bfec6c2924b31a525aed02432c3bbe38680437dfe55778583f7be18ac076611.jpg)  
+Figure 3. Style Aligned Diffusion. Generation of images with a style aligned to the reference image on the left. In each diffusion denoising step all the images, except the reference, perform a shared self-attention with the reference image.
+
+![](images/40d62b020e5800a0fb2d06cae575f19a45f0d7e49c04264b0542055e71d15eec.jpg)  
+Figure 4. Shared attention layer. The target images attend to the reference image by applying AdaIN over their queries and keys using the reference queries and keys respectively. Then, we apply shared attention where the target features are updated by both the target values  $V_{t}$  and the reference values  $V_{r}$ .
+
+ing our attention-sharing operation within the self-attention layers that enable style aligned image set generation.
+
+# 3.1. Preliminaries
+
+Diffusion models [19, 48] are generative latent variable models that aim to model a distribution  $p_{\theta}(x_0)$  that approximates the data distribution  $q(x_0)$  and are easy to sample from. Diffusion models are trained to reverse the diffusion "forward process":
+
+$$
+x _ {t} = \sqrt {\alpha_ {t}} x _ {0} + \sqrt {1 - \alpha_ {t}} \epsilon , \epsilon \sim N (0, I),
+$$
+
+where  $t \in [0, \infty)$  and the values of  $\alpha_{t}$  are determined by a scheduler such that  $\alpha_{0} = 1$  and  $\lim_{t \to \infty} \alpha_{t} = 0$ . During inference, we sample an image by gradually denoising an input noise image  $x_{T} \sim \mathcal{N}(0, I)$  via the reverse process:
+
+$$
+x _ {t - 1} = \mu_ {t - 1} + \sigma_ {t} z, z \sim N (0, I),
+$$
+
+where the value of  $\sigma_{t}$  is determined by the sampler and  $\mu_{t - 1}$  is given by
+
+$$
+\mu_ {t - 1} = \frac {\sqrt {\alpha_ {t - 1}} x _ {t}}{\sqrt {\alpha_ {t}}} + \left(\sqrt {1 - \alpha_ {t - 1}} - \frac {\sqrt {1 - \alpha_ {t}}}{\sqrt {\alpha_ {t}}}\right) \epsilon_ {\theta} (x _ {t}, t),
+$$
+
+where  $\epsilon_{\theta}(x_t,t)$  is the output of a diffusion model parameterized by  $\theta$
+
+Moreover, this process can be generalized for learning a marginal distribution using an additional input condition. That leads text-to-image diffusion models (T2I), where the output of the model  $\epsilon_{\theta}(x_t,t,y)$  is conditioned on a text prompt  $y$ .
+
+Self-Attention in T2I diffusion models. State-of-the-art T2I diffusion models [4,35,46] employ a U-Net architecture [40] that consists of convolution layers and transformer attention blocks [54]. In these attention mechanisms, deep
+
+![](images/ba3e52212d6c2d820ad98b6bb7c6961b73723f21fa0dc4ed42a668f250f4f670.jpg)  
+Figure 5. Ablation study - qualitative comparison. Each pair of rows shows two sets of images generated by the same set of prompts "...in minimal flat design illustration" using different configurations of our method, and each row in a pair uses a different seed. Sharing the self-attention between all images in the set (bottom) results with some diversity loss (style collapse across many seeds) and content leakage within each set (colors from one image leak to another). Disabling the queries keys AdALN operation results with less consistent image sets compared to our full method (top) which keeps on both diversity between different sets and consistency within each set.
+
+image features  $\phi \in \mathbb{R}^{m\times d_h}$  attend to each other via selfattention layers and to contextual text embedding via crossattention layers.
+
+Our work operates at the self-attention layers where deep features are being updated by attending to each other. First, the features are projected into queries  $Q \in m \times d_k$ , keys  $K \in m \times d_k$  and values  $V \in m \times d_h$  through learned linear layers. Then, the attention is computed by the scaled
+
+dot-product attention:
+
+$$
+\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}} V\right),
+$$
+
+where  $d_{k}$  is the dimension of  $Q$  and  $K$ . Intuitively, each image feature is updated by a weighted sum of  $V$ , where the weight depends on the correlation between the projected query  $q$  and the keys  $K$ . In practice, each self-attention layer consists of several attention heads, and then the residual is computed by concatenating and projecting the attention heads output back to the image feature space  $d_{h}$ :
+
+$$
+\hat {\phi} = \phi + \text {M u l t i - H e a d - A t t e n t i o n} (\phi).
+$$
+
+# 3.2. Style Aligned Image Set Generation
+
+The goal of our method is to generate a set of images  $\mathcal{I}_1\ldots \mathcal{I}_n$  that are aligned with an input set of text prompts  $y_{1}\dots y_{n}$  and share a consistent style interpretation with each other. For example, see the garnered image set of toy objects in Fig. 3 that are style-aligned with each other and to the input text on top. A naive way to generate a style aligned image set of different content is to use the same style description in the text prompts. As can be seen at the bottom of Fig. 2, generating different images using a shared style description of "in minimal origami style" results in an unaligned set, since each image is unaware of the exact appearance of other images in the set during the generation process.
+
+The key insight underlying our approach is the utilization of the self-attention mechanism to allow communication among various generated images. This is achieved by sharing attention layers across the generated images.
+
+Formally, let  $Q_{i}$ ,  $K_{i}$ , and  $V_{i}$  be the queries, keys, and values, projected from deep features  $\phi_{i}$  of  $\mathcal{I}_{i}$  in the set, then, the attention update for  $\phi_{i}$  is given by:
+
+$$
+\operatorname {A t t e n t i o n} \left(Q _ {i}, K _ {1 \dots n}, V _ {1 \dots n}\right), \tag {1}
+$$
+
+$$
+\text {w h e r e} K _ {1 \dots n} = \left[ \begin{array}{c} K _ {1} \\ K _ {2} \\ \vdots \\ K _ {n} \end{array} \right] \text {a n d} V _ {1 \dots n} = \left[ \begin{array}{c} V _ {1} \\ V _ {2} \\ \vdots \\ V _ {n} \end{array} \right]. \text {H o w e v e r},
+$$
+
+have noticed that by enabling full attention sharing, we may harm the quality of the generated set. As can be seen in Fig. 5 (bottom rows), full attention sharing results in content leakage among the images. For example, the unicorns got green paint from the garnered dino in the set. Moreover, full attention sharing results with less diverse sets of the same set of prompts, see the two sets in Fig. 5 in bottom rows compared to the sets above.
+
+To restrict the content leakage and allow diverse sets, we share the attention to only one image in the generated set
+
+![](images/925b1dcfbe4f36bc77e342a38b71d524bffa46834c75e5708976954f11d2d13e.jpg)  
+Figure 6. Quantitative Comparison. We compare the results of the different methods (blue marks) and our ablation experiments (orange marks) in terms of text alignment (CLIP score) and set consistency (DINO embedding similarity).
+
+Table 1. User evaluation for style aligned image set generation. In each question, the user was asked to select between two image sets, Which is better in terms of style consistency and match to the text descriptions (see Sec. 4). We report the percentage of judgments in favor of StyleAligned over 800 answers (2400 in total).  
+
+<table><tr><td>StyleDrop (unofficial MUSE)</td><td>StyleDrop (SDXL)</td><td>DreamBooth-LoRA (SDXL)</td></tr><tr><td>85.2 %</td><td>67.1 %</td><td>61.3%</td></tr></table>
+
+(the first in the batch). That is, target image features  $\phi_t$  are attending to themselves and to the features of only one reference image in the set using Eq. 1. As can be seen in Fig. 5 (middle), sharing the attention to only one image in the set results in diverse sets that share a similar style. However, in that case, we have noticed that the style of different images is not well aligned. We suspect that this is due to low attention flow from the reference to the target image.
+
+As illustrated in Fig. 4, to enable balanced attention reference, we normalize the queries  $Q_{t}$  and keys  $K_{t}$  of the target image using the queries  $Q_{r}$  and keys  $K_{r}$  of the reference image using the adaptive normalization operation (AdaIN) [22]:
+
+$$
+\hat {Q} _ {t} = \operatorname {A d a I N} \left(Q _ {t}, Q _ {r}\right) \quad \hat {K} _ {t} = \operatorname {A d a I N} \left(K _ {t}, K _ {r}\right),
+$$
+
+where the AdaIn operation is given by:
+
+$$
+\operatorname {A d a I N} (x, y) = \sigma (y) \left(\frac {x - \mu (x)}{\sigma (x)}\right) + \mu_ {y},
+$$
+
+and  $\mu (x),\sigma (x)\in \mathbb{R}^{d_k}$  are the mean and the standard deviation of queries and keys across different pixels. Finally, our shared attention is given by
+
+$$
+\operatorname {A t t e n t i o n} \left(\hat {Q} _ {t}, K _ {r t} ^ {T}, V _ {r t}\right),
+$$
+
+where  $K_{rt} = \left[ \begin{array}{c}K_{r}\\ \hat{K}_{t} \end{array} \right]$  and  $V_{rt} = \left[ \begin{array}{c}V_{r}\\ V_{t} \end{array} \right]$
+
+# 4. Evaluations and Experiments
+
+We have implemented our method over Stable Diffusion XL (SDXL) [35] by applying our attention sharing overall 70 self-attention layers of the model. The generation of a four images set takes 29 seconds on a single A100 GPU. Notice that since the generation of the reference image is not influenced by other images in the batch, we can generate larger sets by fixing the prompt and seed of the reference image across the set generation.
+
+For example, see the sets in Fig. 2 and 3.
+
+Evaluation set. With the support of ChatGPT, we have generated 100 text prompts describing different image styles over four random objects. For example, “{A guitar, A hot air balloon, A sailboat, A mountain} in papercut art style.” For each style and set of objects, we use our method to generate a set of images. The full list of prompts is provided in the appendix.
+
+Metrics. To verify that each generated image contains its specified object, we measure the CLIP similarity [36] between the image and the text description of the object. In addition, we evaluate the style consistency of each generated set, by measuring the pairwise average cosine similarity between DINO VIT-B/8 [6] embeddings of the generated images in each set. Following [41, 56], we used DINO embeddings instead of CLIP image embeddings for measuring image similarity, since CLIP was trained with class labels and therefore it might give a high score for different images in the set that have similar content but with a different style. On the other hand, DINO better distinguishes between different styles due to its self-supervised training.
+
+# 4.1. Ablation Study
+
+The quantitative results are summarized in Fig. 6, where the right-top place on the chart means better text similarity and style consistency, respectively. As a reference, we report the score obtained by generating the set of images using SDXL (T2I Reference) using the same seeds without any intervention. As can be seen, our method achieves a much higher style consistency score at the expense of text similarity. See qualitative comparison in Fig. 2
+
+In addition, we compared our method to additional two variants of the shared attention as described in Sec. 3.2. The first variant uses full attention sharing (Full Attn. Share) where the keys and values are shared between each pair of images in the set. In the second variant (W.A. AdaIN) we omit the AdaIN operation over queries and keys. As expected, this Full Attn. Share variant, results with higher style consistency and lower text alignment. As can be seen in Fig. 5, Full Attn. Share harms the overall quality of the image sets and diversity across sets. Moreover, our method without the use of AdaIN results in much lower style consistency. Qualitative results can be seen in Fig. 5 and in the supplementary..
+
+![](images/90b4c6812a23bc64e0005d692fb5d98823a2a1d002f9f220cbd0b2c2e09b02ba.jpg)  
+Reference image  
+A guitar
+
+![](images/9182dc023ae4f02985381ae8250dd67f27a32d6b2304ba17eb545d23e84d78b2.jpg)  
+StyleAligned  
+A hot air balloon
+
+![](images/5aca3abfb4d09a6d3e5db3cba5ca2de4f2cc89a6531e18621a8d0d39d20800c7.jpg)  
+A mountain
+
+![](images/c7948ce73d00ded74b3d54ff0bbc607a0b02b28b6b1d46b61271a0f2a65985d0.jpg)  
+StyleDrop (SDXL)  
+A hot air balloon
+
+![](images/a5994e42c5975622b839fe5edb1ae76d3ccbb6ab0107ba916a23dc3bad6ba32b.jpg)  
+A mountain
+
+![](images/5a0e47a4805c3521b0a200742e4d7a6efc924c7cb5a56dac468920f576d2922a.jpg)  
+StyleDrop (MUSE unofficial)  
+A hot air balloon
+
+![](images/46b6a33a75f26a0b999c095607dae2c6762844b5d675645eb58f683a47db58dc.jpg)  
+A mountain
+
+![](images/cf7b7ed8e76a13b03bed044b822d2266a9b6f5900b9b0d0dd4546a02b814c5f7.jpg)  
+Dreambooth-LoRA (SDXL)  
+A hot air balloon
+
+![](images/5df865f298975953ec22baa0012e6361e197e23bf7faab5934792aaa96698c2c.jpg)  
+A mountain
+
+![](images/20e2aaaf0c1e6ab47123702fef3e6d7744bc6664e05e999232247491622d8124.jpg)  
+A woman walking her dogs
+
+![](images/2900abc58cb353417dbcacb0858899b10ac88bc6db44cd9e37deb49575279ad4.jpg)  
+A friendly robot  
+Cherryblossom
+
+![](images/91f969d04a9a14bea047a577c124bb468e71870f92e46a74591a2accfd49efbf.jpg)
+
+![](images/934d4e99253578a4646ad35ef4a76dc408e1ce411b4fae9ab69fc49920db8ff8.jpg)  
+A friendly robot
+
+![](images/cddaf90b91fa69e42670e5a911d7462cb4857ff150b046a76dfa7ec3e37516d5.jpg)  
+Cherryblossom
+
+![](images/2ecc2accac53d7b2fc69a9b04e6409aabeb20377e722c3e746a29ad3763ec90b.jpg)  
+A robot
+
+![](images/eb1f356d8d9dbb4832264617ad7181e028801061eb592887032ae1c4e14a004b.jpg)  
+Cherryblos
+
+![](images/4a444f5f41709352b61c2f29db3900fe49fb1982d985bb2a52e3d55197d629c5.jpg)  
+A friendly robot  
+Cherryblossom
+
+![](images/610c9ba3314ca1226805491b5a9ce683756ab16a5b5f16aeebefaf025038cf13.jpg)
+
+![](images/8840c5026e15cd714d61ba2d56d55ab5f9290a381cb36c9d92013488ec9d9cd7.jpg)  
+A duck
+
+![](images/27d3adda1e9c1e8abb94cd5d3919471f7d4377e42f229c39b65589ea91c646c5.jpg)  
+A ship
+
+![](images/5a84f2c0a5fbafa354a5977511fe2a587e3984e96500383873d23ee76b6ca112.jpg)  
+A rocket
+
+![](images/7f110362dbb4193ed2c122926456ca32cf0ebff2b338a2f79b2ba1a19ea9abe1.jpg)  
+A ship
+
+![](images/e719c9a00052dcead3a2edfa21f5306cf25728b1e523e32e19639193911e3f21.jpg)  
+A rocket
+
+![](images/520144230c3f6cda53ff73278bfc318dc00a4a7b4f25acc8f6aad269a22713e0.jpg)  
+A ship
+
+![](images/f30b62b042d2e1a7bc31ea8397e082c8f27c08c36af31b9f8e439353ddb0b106.jpg)  
+A car
+
+![](images/c120d4a94419eaaf18f09eb87f57ff0e51bc783cbfc3b7df039a20de156ed2fa.jpg)  
+A ship
+
+![](images/9d94bbdfcfdd7eeb1b3052b31dc2b9ffcd4e5c7236d8b171d9ae33c032af0cff.jpg)  
+A rocket
+
+![](images/d7ddbb290dfd3012f5aa02c56956acb097eef5b8f80243a3deaff12c12b286c4.jpg)  
+A cute bear
+
+![](images/eb8f5046cf9749c732d4fd406a4560c5b3b793e5ae5ad161bdccdc6d44f34b1b.jpg)  
+A moose
+
+![](images/101e1170f7064b5edc9f04d96143713764a19191953362df139e3913fa60e039.jpg)  
+A baby pen
+
+![](images/0f3d2fa0c62c9e0e5edf09ed204fd59daf986792ce49f36c091cb0568c688e05.jpg)  
+A moose  
+A baby penguin
+
+![](images/1e88d7841435824dd9ec845f5d2d1d4131c53799ec14308bbce9203f29322916.jpg)
+
+![](images/066e32b3d3e53446c8e21b056da640992280644c433de7ec203d5b58ab4d1c94.jpg)  
+A moose  
+A baby penguin
+
+![](images/d493c3f5b38c916496d273246a57c83b152ee12941106b89e07e94b18f98934f.jpg)
+
+![](images/98bd1600692848c1d1e32c87ffbbaad7e8831f1325c1d591c3d4836fded90ab6.jpg)  
+A moose  
+A baby penguin
+
+![](images/07dab4a6b25a0ac834d7e5208ec8f5ee690714b8ca83182f7753e0c5f5c91b88.jpg)
+
+![](images/0e75922bafab63fd0dc43b2a236d62036d8df5d8480aa1a4c0b98594e1613c20.jpg)  
+A camera
+
+![](images/fde80ce0ea8f158811b6028048d98b0700eaf3363d4247ad20510ff14bdce6ce.jpg)  
+A fireman
+
+![](images/3613566046215aec7374cc7ae12d282ad8e812100a330aa8355d1ff5806af360.jpg)  
+A tall hill
+
+![](images/687587c656d7c57e108c490ea6e0f9a58ca98ec07e812b9a7d6dd6c239aad61e.jpg)  
+A fireman
+
+![](images/5f858526fbb5f99a5735067d56df198d3cd99d2fc790db7840d84b2fe9da5d37.jpg)  
+A tall hill
+
+![](images/a5ecc311b5dbeb6b2baedaef213a0d6b54334591474cb728be06ab329d4edeef.jpg)  
+A fireman
+
+![](images/5f0710c2366220dd26304908de510520344952d40686136117fcf3bf7c907660.jpg)  
+A tall hill
+
+![](images/de633264efa8f257c089060d149edbe67c2102974cb50c9648219b00e7729e36.jpg)  
+A fireman
+
+![](images/848ac7555642c0cea82a2611ee7dab8c5f40775618c28cab35f1d537b3a2204a.jpg)  
+A tall hill
+
+![](images/218c906898b3a5ae3af8a0c59cbfe3f2e6a876f559a6ed9dbed5f74cff59aca5.jpg)  
+A saxophone
+
+![](images/5a326baa1c71afd4fa92e7b945445d5df1a917fb68392dc96bde8df9060c0670.jpg)  
+A sunflower
+
+![](images/2e24b65414432bef1e741090cb6d689be11819594a2cdbe6f42fe4398c0cb96f.jpg)  
+A compass
+
+![](images/3bd7a2838367c664a9e337c39d2c1f17bc4b4d674a887d56572386b522ef23a5.jpg)  
+A sunflower
+
+![](images/a909ea9a7f355a26af4d2503b0d77e6884a4592309406b7b8a2b5cbf9376bc47.jpg)  
+A compass
+
+![](images/245f0efac9754680077bce0f2084c5a09d0700a318b35ddbc7df9e82dd856f22.jpg)  
+A sunflower
+
+![](images/1b207c7f42fabcce392a07367ce06138d3678cf13f55abd0b1abdda1ce539d27.jpg)  
+A compass
+
+![](images/0e392ae6870353c15e4c33cb5e500f566f4a5367ee5622f56e67fdfb217bce2f.jpg)  
+A sunflower
+
+![](images/2a20746fe4b8fc5b42a361f53ef12c4712aebf9df993f31740b2c8624cbe67e5.jpg)  
+A compass
+
+![](images/b266b1b05456e22c05ece494dafbad1df540c9dd01234ce23f0846dcfdab7187.jpg)  
+A bottle of wine
+
+![](images/0cc1655eb494e44e0a04cd133ad64293c8f42e2d1a50b974f0592bdb3ef10b6a.jpg)  
+Scones
+
+![](images/c0febaa1d6f4af1489962cdf2c022e226c99f50d6390739b462f03a7e5894d83.jpg)  
+A muffin
+
+![](images/a67f1819a5ac9832c842ffec3c5c2bea10ef041df432dc1197e1a531b7a6f866.jpg)  
+Scones
+
+![](images/273b7d99f4027b80fb94d2a712e1526dccae55b952daf95ebaf79781277e4638.jpg)  
+A muffin
+
+![](images/9729ac558578c2cc718793e23350acb7c4e090b725969a2dc8c12431c4a194f0.jpg)  
+A fireman
+
+![](images/8fd9cff8bb4ccfef3049b2a2e4eb07b940bcfdadfe156730a753ba68f0cf99eb.jpg)  
+A muffin
+
+![](images/3bc108ed9e3f17a44c0e1c114055d5844ede0e28a61f0b1c4879b10da89fb618.jpg)  
+Scones
+
+![](images/e20646a086cffac4ae116315a50305af6784ccc6c59bd2082f4bdeee1ee46165.jpg)  
+A muffin
+
+![](images/2fa74eae159321dd188194042693392198fcb61858fa854a7b9382491534aef9.jpg)  
+A wise owl
+
+![](images/4f563e863ce68571d9fa9ceae9bb07ae33f3fb60e146a0dab8488c410f327a6e.jpg)  
+Full moon  
+Figure 7. Qualitative comparison to personalization based methods.
+
+![](images/39fa8d35e7e3ba2427e87d43ffa4924cc93aae4c1d0d493c156b7c5d1acb4cdd.jpg)  
+A book
+
+![](images/2bb7b8d91f4af638659972a0ca014ec10d525a15fc9566d0869f376937cbc8a6.jpg)  
+Full moon
+
+![](images/bc6a9fbaf50c348093b27ffe01d00455c1a6a322d45a787eed6f99fd26ddc061.jpg)  
+A book
+
+![](images/d5733e81801d09ddb91dd13aa468c0647906a7565c02d87e74e2a016c17152db.jpg)  
+Full moon
+
+![](images/3a71945b845f2970193a4628958143b266d4488bfa7f69cf7b4ced283b29b9c6.jpg)  
+A book
+
+![](images/ec4bb150faf2d7acfdb5c2bcb832dccc3d4f65619f949822f0f87ff929759a87.jpg)  
+Full moon
+
+![](images/991f0ea54f6cbd853d334702a4074807d20a62f8c656ffa058f21edf2e85dc5d.jpg)  
+A book
+
+![](images/ada735eb1902ca79d5ab11a26e4fe25d02bb9659ddec19de02a6b75738006b8c.jpg)  
+Figure 8. Varying level of attention sharing. By reducing the number of shared attention layers, i.e., allowing only self-attention in part of the layers, we can get more varied results (bottom rows) at the expense of style alignment (top row).
+
+# 4.2. Comparisons
+
+For baselines, we compare our method to T2I personalization methods. We trained StyleDrop [49] and DreamBooth [41] over the first image in each set of our evaluation data, and use the trained personalized weights to generate the additional three images in each set. We use a public unofficial implementation of StyleDrop * (SDRP-unofficial) over non-regressive T2I model. Due to the large quality gap between the unofficial MUSE model † to the official one [7], we follow StyleDrop and implement an adapter model over SDXL (SDRP-SDXL) *, where we train a low rank linear layer after each Feed-Forward layer at the model's attention blocks. For training DreamBooth, we adapt the LoRA [21, 43] variant (DB-LoRA) over SDXL using the public huggingface-diffusers implementation ‡. We follow the hyperparameters tuning reported in [49] and train both SDRP-SDXL and DB-LoRA for 400 steps to prevent overfitting to the style training image.
+
+As can be seen in the qualitative comparison, Fig. 7, the image sets generated by our method, are more consistent across style attributes like color palette, drawing style, and composition. Moreover, the personalization-based methods
+
+may leak the content of the training reference image (on the left) when generating the new images. For example, see the repeated woman and dogs in the results of DB-LoRA and SDRP-SDXL at the second row or the repeated owl at the bottom row. Similarly, because of the content leakage, these methods obtained lower text similarity scores and higher set consistency scores compared to our method.
+
+We also apply two encoder-based personalization methods ELITE [57], IP-Adapter [59], and BLIP-Diffusion [27] over our evaluation set. These methods receive as input the first image in each set and use its embeddings to generate images with other content. Unlike the optimization-based techniques, these methods operate in a much faster feedforward diffusion loop, like our method. However, as can be seen in Fig. 6, their performance for style aligned image generation is poor compared to the other baselines. We argue that current encoder-based personalization techniques struggle to disentangle the content and the style of the input image. We supply qualitative results in the supplementary.
+
+User study. In addition to the automatic evaluation, we conducted a user study over the results of our method, StyleDrop (unofficial MUSE), StyleDrop (SDXL), and DreamBooth-LoRA (SDXL). In each question, we randomly sample one of the evaluation examples and show
+
+![](images/68ce1e879762b8e47f9f2caa41262a4af139d672c99f321df933fc63e9d4f341.jpg)  
+Figure 9. Style aligned image generation to an input image. Given an input reference image (left column) and text description, we first apply DDIM inversion over the image to get the inverted diffusion trajectory  $x_{T}$ ,  $x_{T-1} \ldots x_{0}$ . Then, starting from  $x_{T}$  and a new set of prompts, we apply our method to generate new content (right columns) with an aligned style to the input.
+
+the user the 4 image set that resulted from our and another method (in a random order). The user had to choose which set is better in terms of style consistency, and text alignment. A print screen of the user study format is provided in the supplementary. Overall, we collected 2400 answers from 100 users using the Amazon Mechanical Turk service. The results are summarized in Tab. 1 where for each method, we report the percentage of judgments in our favor. As can be seen, most participants favored our method by a large margin. More information about our user study can be found in the supplementary.
+
+# 4.3. Additional Results
+
+Style alignment control. We provide means of control over the style alignment to the reference image by applying the shared attention over only part of the self-attention layers. As can be seen in Fig. 8, reducing the number of shared attention layers results with a more diverse image set, which still shares common attributes with the reference image.
+
+StyleAligned from an input image. To generate style-aligned images to an input image, we apply DDIM inversion [50] using a provided text caption. Then, we apply our method to generate new images in the style of the input using the inverted diffusion trajectory  $x_{T}, x_{T-1}, \ldots, x_{0}$  for the reference image. Examples are shown in Fig. 9, where we use BLIP captioning [28] to get a caption for each input image. For example, we used the prompt "A render of a house with a yellow roof" for the DDIM inversion of the top example and replaced the word house with other objects to generate the style-aligned images of a car, a cat, and a cactus. Notice that this method does not require any optimization. However, DDIM inversion may fail [30] or results with an erroneous trajectory [24]. More results and analysis, are provided in the supplementary.
+
+StyleAligned with other methods. Since our method doesn't require training or optimization, it can be easily combined on top of other diffusion based methods to generate style-consistent image sets. Fig. 10 shows several such examples where we combine our method with ControlNet [60], DreamBooth [42] and MultiDiffusion [3]. More examples and details about the integration of StyleAligned with other methods can be found in supplementary.
+
+# 5. Conclusions
+
+We have presented StyleAligned, which addresses the challenge of achieving style-aligned image generation within the realm of large-scale Text-to-Image models. By introducing minimal attention sharing operations with AdaIN modulation during the diffusion process, our method successfully establishes style-consistency and visual coherence across generated images. The demonstrated efficacy of StyleAligned in producing high-quality, style-consistent images across diverse styles and textual prompts under
+
+![](images/48eab7e15d523b557b2e445a5c7f8acebe22b8794caccc6a4756084a368b7c75.jpg)  
+Figure 10. StyleAligned with other methods. On top, StyleAligned is combined with ControlNet to generate style-aligned images conditioned on depth maps. In the middle, our method combined with MultiDiffusion to generate panorama images that share multiple styles. On the bottom, style consistent and personalized content created by combining our method with pretrained personalized DreamBooth-LoRA models.
+
+scores its potential in creative domains and practical applications. Our results affirm StyleAligned capability to faithfully adhere to provided descriptions and reference styles while maintaining impressive synthesis quality.
+
+In the future we would like to explore the scalability and adaptability of StyleAligned to have more control over the shape and appearance similarity among the generated images. Additionally, due to the limitation of current diffusion inversion methods, a promising direction is to leverage StyleAligned to assemble a style-aligned dataset which then can be used to train style condition text-to-image models.
+
+# 6. Acknowledgement
+
+We thank Or Patashnik, Matan Cohen, Yael Pritch, and Yael Vinker for their valuable inputs that helped improve this work.
+
+# References
+
+[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer, 2023. 2  
+[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 42(4), jul 2023. 2  
+[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 8  
+[4] James Betker, Gabriel Goh, Li Jing, TimBrooks, Jianfeng Wang, Linjie Li, LongOuyang, JuntangZhuang, JoyceLee, YufeiGuo, WesamManassra, PrafullaDhariwal, CaseyChu, YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 3  
+[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560–22570, October 2023. 2  
+[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5  
+[7] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming Yang, Kevin P. Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In International Conference on Machine Learning, 2023. 2, 7  
+[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42:1 - 10, 2023. 2  
+[9] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In The Eleventh International Conference on Learning Representations, 2022. 2  
+[10] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 571-576. 2023. 2  
+[11] Alexei A Efros and Thomas K Leung. Texture synthesis by non-parametric sampling. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 1033-1038. IEEE, 1999. 2  
+[12] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 2  
+[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2022. 1, 2  
+[14] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain
+
+tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2  
+[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2  
+[16] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris N. Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. ArXiv, abs/2303.11305, 2023. 2  
+[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2  
+[18] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 557-570. 2023. 2  
+[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proc. NeurIPS, 2020. 3  
+[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2  
+[21] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 7  
+[22] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 5  
+[23] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 2  
+[24] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. arXiv preprint arXiv:2304.06140, 2023. 2, 8  
+[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2  
+[26] Hochang Lee, Sanghyun Seo, Seungtaek Ryoo, and Kyunghyun Yoon. Directional texture transfer. In Proceedings of the 8th International Symposium on Non-Photorealistic Animation and Rendering, pages 43–48, 2010. 2  
+[27] Dongxu Li, Junnan Li, and Steven C. H. Hoi. BLIP-Diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. ArXiv, abs/2305.14720, 2023. 2, 7  
+[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 8
+
+[29] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In IEEE International Conference on Computer Vision (ICCV), 2019. 2  
+[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2, 8  
+[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 2  
+[32] Dong Huk Park*, Grace Luo*, Clayton Toste, Samaneh Azadi, Xihui Liu, Maka Karalashvili, Anna Rohrbach, and Trevor Darrell. Shape-guided diffusion with inside-outside attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2  
+[33] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2  
+[34] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. ArXiv, abs/2303.11306, 2023. 2  
+[35] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 3, 5  
+[36] 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 International Conference on Machine Learning, 2021. 5  
+[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1  
+[38] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 2  
+[39] 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. 1  
+[40] 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  
+[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 5, 7  
+[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 2, 8  
+[43] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2022.7  
+[44] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. ACM SIGGRAPH 2022 Conference Proceedings, 2021. 2  
+[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1  
+[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2, 3  
+[47] Jing Shi, Wei Xiong, Zhe L. Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. ArXiv, abs/2304.03411, 2023. 2  
+[48] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3  
+[49] Kihyuk Sohn, Nataniel Ruiz, Kimin Lee, Daniel Castro Chin, Irina Blok, Huiwen Chang, Jarred Barber, Lu Jiang, Glenn Entis, Yuanzhen Li, et al. Styledrop: Text-to-image generation in any style. arXiv preprint arXiv:2306.00983, 2023. 1, 2, 7  
+[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 8  
+[51] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. SIGGRAPH 2023 Conference Proceedings, 2023. 2  
+[52] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 2
+
+[53] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2  
+[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3  
+[55] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. arXiv preprint arXiv:2211.13752, 2022. 2  
+[56] Andrey Voynov, Q. Chu, Daniel Cohen-Or, and Kfir Aberman. P+: Extended textual conditioning in text-to-image generation. ArXiv, abs/2303.09522, 2023. 2, 5  
+[57] Yuxiang Wei. Official implementation of ELITE. https://github.com/csyxwei/ELITE, 2023. 7  
+[58] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: Encoding visual concepts into textual embeddings for customized text-to-image generation. ArXiv, abs/2302.13848, 2023. 2  
+[59] Hu Ye, Jun Zhang, Siyi Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. ArXiv, abs/2308.06721, 2023. 2, 7  
+[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 8

2024/Style Aligned Image Generation via Shared Attention/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fc5af7e8bacb904de7cfdcd2f59cacb00c2ad3e365d4925b98b8ec5d448a95f
+size 1404150

2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/1e340b0f-2934-49e8-9933-238d4fb891ff_content_list.json

@@ -0,0 +1,1743 @@
+[
+    {
+        "type": "text",
+        "text": "Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            107,
+            130,
+            861,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Woo-Jin Ahn<sup>1</sup> Geun-Yeong Yang<sup>1</sup> Hyun-Duck Choi<sup>2*</sup> Myo-Taeg Lim<sup>1*</sup>  \n<sup>1</sup>Korea University <sup>2</sup>Chonnam National University",
+        "bbox": [
+            183,
+            202,
+            790,
+            239
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}\\{wjahn, hggofficial, mlim\\} @korea.ac.kr$ $^{2}$ ducky.choi@jnu.ac.kr",
+        "bbox": [
+            218,
+            239,
+            750,
+            257
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            291,
+            313,
+            306
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Deep learning models for semantic segmentation often experience performance degradation when deployed to unseen target domains unidentified during the training phase. This is mainly due to variations in image texture (i.e. style) from different data sources. To tackle this challenge, existing domain generalized semantic segmentation (DGSS) methods attempt to remove style variations from the feature. However, these approaches struggle with the entanglement of style and content, which may lead to the unintentional removal of crucial content information, causing performance degradation. This study addresses this limitation by proposing BlindNet, a novel DGSS approach that blinds the style without external modules or datasets. The main idea behind our proposed approach is to alleviate the effect of style in the encoder whilst facilitating robust segmentation in the decoder. To achieve this, BlindNet comprises two key components: covariance alignment and semantic consistency contrastive learning. Specifically, the covariance alignment trains the encoder to uniformly recognize various styles and preserve the content information of the feature, rather than removing the style-sensitive factor. Meanwhile, semantic consistency contrastive learning enables the decoder to construct discriminative class embedding space and disentangles features that are vulnerable to misclassification. Through extensive experiments, our approach outperforms existing DGSS methods, exhibiting robustness and superior performance for semantic segmentation on unseen target domains. The code is available at https://github.com/root0yang/BlindNet.",
+        "bbox": [
+            75,
+            323,
+            472,
+            763
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            805,
+            209,
+            821
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Semantic segmentation, a technique that classifies each pixel in an image into predefined categories, has garnered significant attention due to its potential applications in vari-",
+        "bbox": [
+            75,
+            830,
+            468,
+            878
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/f1ff12b29489673484c490594c0b0065fa431db390eaa18e1fb3b77b2196fbd2.jpg",
+        "image_caption": [
+            "Figure 1. Comparison of semantic segmentation results between the baseline (DeepLabV3+ with ResNet50 backbone) and our BlindNet. Both models are trained on the source domain (GTAV [43]) and tested on the target domain (Cityscapes [8])."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            287,
+            893,
+            412
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ous fields. Particularly, it plays a crucial role in autonomous driving [1, 18] and robotic systems [34, 35]. Besides, with the advent of large datasets, deep neural networks have emerged as a trending approach for semantic segmentation tasks, achieving impressive results [3, 5, 44, 54]. However, there remains a major bottleneck detailing the meticulous and labor-intensive process of dataset labeling. More particularly, this process not only consumes time but also poses economic challenges [8, 46]. To address this challenge, synthetic datasets have emerged as a compelling alternative. Specifically, these datasets, generated using three-dimensional (3D) rendering techniques, offer vast amounts of easily accessible data, eliminating the need for manual labeling [43, 45]. However, a challenge arises when models trained on synthetic datasets are deployed in real-world scenarios. More precisely, a domain shift problem arises due to style factor discrepancies (e.g. texture, illumination, and image quality) between synthetic and real-world data, which affects the performance of the model, as shown in Fig. 1.",
+        "bbox": [
+            496,
+            506,
+            893,
+            808
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address the domain shift problem, domain adaptation semantic segmentation (DASS) has been introduced [15, 17, 22, 25, 51, 62]. Specifically, DASS aims to bridge the gap between source and target domains by aligning their data distributions. However, a significant limitation of DASS is its dependency on the target domain during train-",
+        "bbox": [
+            496,
+            810,
+            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,
+            810,
+            44
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding Authors",
+        "bbox": [
+            94,
+            886,
+            228,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "3616",
+        "bbox": [
+            482,
+            944,
+            516,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ing. For DASS to function effectively, target domain samples must be available during the training phase.",
+        "bbox": [
+            75,
+            90,
+            468,
+            121
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Meanwhile, domain generalized semantic segmentation (DGSS) has been introduced as an alternative approach to tackle the domain shift problem. Unlike DASS, DGSS is trained only with the source domain, aiming to extract domain-invariant features. To achieve this, two main techniques have been employed: domain randomization (DR) and feature normalization (FN).",
+        "bbox": [
+            75,
+            123,
+            467,
+            227
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "DR augments the training set by introducing variability, either by altering the image style [41, 60] or by modifying the feature representation [24, 53]. By exposing the model to a wider variety of styles via DR, the network is less likely to overfit to the specific styles present in the training data. Consequently, the robustness of the model is improved, making it more adept at generalizing to new, unseen domains. Nevertheless, a crucial limitation of DR is its significant dependence on auxiliary domains.",
+        "bbox": [
+            75,
+            229,
+            467,
+            364
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "FN methods, converse to DR, regularize the features to prevent the model from overfitting to the distinct styles or characteristics of the training data. This is achieved by removing domain-specific style information using feature statistics, such as instance normalization [38] or whitening transformation [7, 19, 39, 42]. While these approaches effectively remove style-related information, they simultaneously pose the challenge of removing semantic content because content and style information are entangled. Consequently, the model fails to capture the essential patterns or features required for accurate segmentation prediction.",
+        "bbox": [
+            75,
+            367,
+            467,
+            532
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address this problem, we propose BlindNet, a model that blinds the style within the encoder and improves the robustness of the decoder, without requiring auxiliary datasets or external modules. Specifically, our proposed BlindNet consists of two components: covariance alignment for the encoder and semantic consistency contrastive learning for the decoder. Precisely, the covariance alignment facilitates the generation of style-invariant features with the proposed covariance matching loss function (CML) and the cross-covariance loss function (CCL). Specifically, CML mitigates the effects of style variations, while CCL focuses on preserving content information, effectively addressing the prevalent content information loss observed in the FN method. To further improve the generalization ability, we develop semantic consistency contrastive learning, which consists of class-wise contrastive learning (CWCL) and semantic disentanglement contrastive learning (SDCL). Particularly, the CWCL constructs a discriminative class embedding space, while SDCL disentangles features of similar classes that often lead to prediction errors. Extensive experiments across various datasets demonstrate that the proposed BlindNet outperforms existing DGSS methods.",
+        "bbox": [
+            75,
+            535,
+            467,
+            866
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions are summarized as follows:",
+        "bbox": [
+            96,
+            868,
+            398,
+            882
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- We propose a covariance alignment within the encoder,",
+        "bbox": [
+            76,
+            885,
+            467,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "comprising CML and CCL. Specifically, the CML aims to mitigate the effects of style variations, while CCL ensures the preservation of content information, together facilitating the generation of style-agnostic features.",
+        "bbox": [
+            509,
+            90,
+            890,
+            151
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose semantic consistency contrastive learning within the decoder that comprises CWCL and SDCL, utilizing segmentation masks. Specifically, CWCL generates discriminative embeddings, whereas SDCL disentangles features of similar classes, enhancing the robustness of the model.",
+            "- Through extensive experiments, we demonstrate the superiority of our approach in DGSS, without the need to alter the network architecture or rely on external datasets."
+        ],
+        "bbox": [
+            500,
+            152,
+            890,
+            286
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Works",
+        "text_level": 1,
+        "bbox": [
+            500,
+            301,
+            648,
+            316
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Domain adaptation and generalization for semantic segmentation. Domain adaptation (DA) aims at minimizing the distribution discrepancy between different domains, enabling a model to generalize from a source to a target domain. For DASS, adversarial training and cross-domain self-training strategies are commonly used. Particularly, adversarial-based methods [15, 29] employ generative adversarial networks [11] to close the feature distribution gap between source and target domains. Meanwhile, cross-domain self-training methods [16, 37, 59, 62] generate pseudo-labels for target domain data using pre-trained models, and employ them as training data, thereby expanding the training data and reducing distribution differences between the source and target domain.",
+        "bbox": [
+            498,
+            325,
+            890,
+            537
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Domain generalization (DG) methods [26, 28, 30, 50, 61] aim to improve the generalization ability of the model without accessing the target domain during training. Since the difference in style of the image is the main cause of the disparity, most existing domain generalized semantic segmentation methods utilize the style information of the image for domain-invariant learning. Feature statistics (e.g. mean, variance, covariance, gram matrix, etc.) which are commonly used in style transfer [10, 21, 27, 57] are employed to capture the style information. Interestingly, existing DGSS methods can be separated into two parts: i) domain randomization to expand the distribution of style or ii) feature normalization to remove style.",
+        "bbox": [
+            496,
+            537,
+            890,
+            733
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "DR involves randomizing either the image or its features through stylization to learn domain-invariant features from various styles. For example, Peng et al. [42] extended the source domain data by stylizing images in the style of unreal paintings. Similarly, Yue et al. [60] and Huang et al. [22] attempted to enhance generalization by synthesizing images with diverse styles in the image space. In another study, Lee et al. [24] adopted ImageNet data [9] as wild data and performed randomization via synthesis in the feature space. Meanwhile, Wu et al. [53] diversified the trainable feature space by mixing the statistics of the feature and its color-",
+        "bbox": [
+            496,
+            734,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "3617",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "jittered feature with Ada-IN [21].",
+        "bbox": [
+            75,
+            90,
+            299,
+            104
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "FN methods aim to remove domain-specific styles from features, extracting only domain-invariant content. For instance, Pan et al. [39] first attempted the DGSS method, combining batch normalization (BN) [23] and instance normalization (IN) [48] in the network layer. While BN preserves the content information within discriminative features, IN focuses on removing domain-specific style information from features. In a study, Choi et al. [7] addressed the limitations of previous whitening transformation [6, 31] that can eliminate the content information from the feature. Specifically, they proposed an instance-selective whitening approach designed to remove covariance components that are sensitive to domain shifts. Peng et al. [41] developed semantic-aware normalization that performs on class-wise and semantic-aware whitening that aligns channels based on the prediction through group whitening transformation [6]. In a study, Xu et al. [55] introduced the prior guided attention module and guided feature whitening to re-calibrate the feature and remove domain-specific style effects, respectively. Unlike the FN methods that directly remove the style component, our work explores a covariance alignment method that mitigates the effect of the style's effect while preserving the content information. Our method achieves the DGSS without any additional modules or auxiliary datasets.",
+        "bbox": [
+            75,
+            109,
+            472,
+            488
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Contrastive Learning. Contrastive learning aims to learn representations by maximizing the similarity between positive pairs of samples while minimizing the similarity between negative pairs. In recent years, it has attracted significant attention for its effectiveness in learning discriminative representations across various tasks [2, 4, 12, 14]. Oord et al. [36] were the first to introduce the infoNCE loss, a type of contrastive loss function designed for self-contrastive learning. In a work, Park et al. [40] introduced patch-level contrastive learning for image translation, using co-located patches as positive pairs and spatially distant patches as negatives to maintain image context. For the semantic segmentation task, Wang et al. [52] introduced class-wise contrastive learning to aid the model in learning the embedding space of each class. Specifically, they sampled the classes existing in the images and applied contrastive learning based on the class label. For DGSS, Lee et al. [24] adopted contrastive learning to learn the ImageNet information in their model. Specifically, they set the ImageNet data as wild and applied the contrastive learning method by setting the wild-stylized feature and its closest wild content as positive samples. In another study, Yang et al. [56] developed multi-level contrastive learning, which designed instance prototypes and class prototypes for contrastive learning. Specifically, they sample each class's pixel features and apply contrastive learning with a transition-probability ability matrix. Unlike recent DGSS works that embed the origi",
+        "bbox": [
+            75,
+            493,
+            472,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "nal image, we define contrastive learning for the augmented image. Specifically, the proposed method builds a robust embedding space by preserving the semantic consistency of the feature representation across various domains.",
+        "bbox": [
+            496,
+            90,
+            893,
+            151
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Method",
+        "text_level": 1,
+        "bbox": [
+            498,
+            164,
+            591,
+            180
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The goal of the proposed method is to train a segmentation model  $\\varphi$  on a given source domain  $S$  and generalize well to the unseen target domain  $T$ . Precisely, the source domain  $S = \\{(x,y)\\}$  contains a paired image  $x \\in \\mathbb{R}^{H \\times W \\times 3}$  and segmentation label  $y \\in \\mathbb{R}^{H \\times W \\times C}$ , where  $H$ ,  $W$ , and  $C$  denote the height of the image, the width of the image, and the class number of the segmentation map, respectively. The model  $\\varphi$  takes an original  $x$  and its augmented counterpart  $x_{a}$ , which have the same content but different styles, and uses the feature information to enhance its generalization ability.",
+        "bbox": [
+            496,
+            190,
+            893,
+            354
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As shown in Fig. 2, our method leverages the feature information through two main approaches: covariance alignment and semantic consistency contrastive learning. Specifically, the covariance matching ensures that features, having different styles but the same content, contain similar information. Additionally, the semantic consistency contrastive learning embeds the generalized features into discriminative representation based on the segmentation label.",
+        "bbox": [
+            496,
+            356,
+            893,
+            478
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Covariance Alignment",
+        "text_level": 1,
+        "bbox": [
+            498,
+            486,
+            710,
+            502
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The domain shift in semantic segmentation results from changes in the visual characteristics of the image, known as style. This style information is typically detected in the shallow layers of networks [38], which are the encoders of the segmentation model. Based on this understanding, our method targets the encoder to handle style variations by employing the proposed covariance matching loss and cross-covariance loss function.",
+        "bbox": [
+            496,
+            508,
+            893,
+            630
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Covariance Matching Loss",
+        "text_level": 1,
+        "bbox": [
+            498,
+            648,
+            692,
+            664
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To train the network to uniformly recognize various styles without removing content information, we introduce the CML. Specifically, the loss aims to minimize the difference between covariance matrices derived from different styles of image features. Given an image pair  $(x,x_{a})$ , the features from  $i^{th}$  block of the encoder are represented as  $F^i\\in \\mathbb{R}^{(H^i\\times W^i)\\times C^i}$  and  $F_{a}^{i}\\in \\mathbb{R}^{(H^{i}\\times W^{i})\\times C^{i}}$ , respectively. Further, following the methodologies of [7, 41], we compute the covariance matrices using instance normalized features [48], which ensures consistent scaling across features. The feature maps are normalized and flattened into  $\\bar{F}$  and  $\\bar{F}_a\\in \\mathbb{R}^{(HW)\\times C}$ , which are given by:",
+        "bbox": [
+            496,
+            672,
+            893,
+            854
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\bar {F} = \\frac {(F - \\mu (F))}{\\sigma (F)}, \\bar {F} _ {a} = \\frac {(F _ {a} - \\mu (F _ {a}))}{\\sigma (F _ {a})}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            558,
+            864,
+            890,
+            898
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "3618",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/5fcbbb724d3720e68e61b5e9b5a91dabb76d3372b3e54fd809de935bffa75425.jpg",
+        "image_caption": [
+            "Figure 2. Overview of the proposed BlindNet. The network processes a pair of images - the original image  $x$  and its augmented counterpart  $x_{a}$ . It employs covariance alignment to treat encoder features and utilizes semantic consistency contrastive learning for the processing of decoder features."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            88,
+            890,
+            404
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mu (\\cdot)\\in \\mathbb{R}^{C}$  and  $\\sigma (\\cdot)\\in \\mathbb{R}^C$  denote the mean and standard deviation of the features. Utilizing the normalized features, we evaluate the covariance matrices for the original and augmented image features as:",
+        "bbox": [
+            75,
+            481,
+            468,
+            544
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\Sigma_ {x, x} ^ {i} = \\bar {F} ^ {i} ^ {T} \\cdot \\bar {F} ^ {i}, \\quad \\Sigma_ {x _ {a}, x _ {a}} ^ {i} = \\bar {F} _ {a} ^ {i} ^ {T} \\cdot \\bar {F} _ {a} ^ {i}. \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            553,
+            468,
+            575
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The CML is then formulated to align these covariance matrices, ensuring that the network maintains consistency in the presence of style variations. The CML is defined as:",
+        "bbox": [
+            75,
+            585,
+            468,
+            631
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {C M} = \\sum_ {i = 1} ^ {n _ {e}} \\| \\Sigma_ {x, x} ^ {i} - \\Sigma_ {x _ {a}, x _ {a}} ^ {i} \\| _ {2}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            165,
+            643,
+            468,
+            681
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $n_e$  denotes the number of blocks of the encoder.",
+        "bbox": [
+            76,
+            691,
+            437,
+            705
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Cross-covariance Loss",
+        "text_level": 1,
+        "bbox": [
+            76,
+            726,
+            236,
+            739
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "While the CML effectively aligns the internal distributions of features within the same image, it does not fully account for the direct correlations across paired images. To complement this, we introduce CCL, which aims to encode the consistent content information of an image pair  $(x,x_{a})$  by utilizing the cross-covariance of the image pair. Given the normalized feature pair  $(\\bar{F}^i,\\bar{F}_a^i)$ , the cross-covariance of the feature pair can be expressed as:",
+        "bbox": [
+            75,
+            748,
+            468,
+            869
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\Sigma_ {x, x _ {a}} ^ {i} = \\bar {F} ^ {i} ^ {T} \\cdot \\bar {F} _ {a} ^ {i}. \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            207,
+            880,
+            468,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The cross-covariance is expected to exhibit an identity matrix, as the feature pair should contain identical information. Nonetheless, the proposed CCL converges only the diagonal component of the covariance matrix to one. This is to prevent the drawbacks of the existing FN methods [7, 42] from removing content information. The CCL function is thus defined as:",
+        "bbox": [
+            498,
+            482,
+            890,
+            587
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {C C} = \\sum_ {i = 1} ^ {n _ {e}} \\| \\operatorname {d i a g} \\left(\\Sigma_ {x, x _ {a}} ^ {i}\\right) - \\mathbb {1} \\| _ {2}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            584,
+            592,
+            890,
+            631
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathrm{diag}(\\Sigma_c) \\in \\mathbb{R}^C$  denotes the column vector comprising diagonal elements of  $\\Sigma_{x,x_a}^i$  and  $\\mathbb{1} \\in \\mathbb{R}^C$  denotes the one vector.",
+        "bbox": [
+            498,
+            638,
+            890,
+            681
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Semantic Consistence Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            500,
+            691,
+            870,
+            707
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "While the encoder focuses on generating style-blinded features, the decoder aims to improve the robustness of the segmentation prediction against domain shifts. For the decoder, we employ a contrastive learning approach, which has demonstrated effectiveness in extracting discriminative features [4]. Specifically, we utilize the InfoNCE loss [36], which is formulated as:",
+        "bbox": [
+            498,
+            714,
+            890,
+            819
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {I N} (a, p, n) = - \\log \\left(\\frac {e ^ {(a \\cdot p / \\tau)}}{e ^ {(a \\cdot p / \\tau)} + \\sum_ {n} ^ {N ^ {-}} e ^ {(a \\cdot n / \\tau)}}\\right) \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            513,
+            823,
+            890,
+            864
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $a, p, n$ , and  $N^{-}$  denote anchor, positive sample, negative sample, and negative sample set, respectively.",
+        "bbox": [
+            498,
+            869,
+            890,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "3619",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/f9fb00733b6b0939d8e914bad891b9b9d71654c5baf4d18638084e017a3d791f.jpg",
+        "image_caption": [
+            "Figure 3. Illustration of semantic consistency contrastive learning: The mask  $M$  represents the error mask derived from the augmented segmentation map. CWCL conducts contrastive learning by sampling per segmentation class and SDCL conducts contrastive learning based on the  $M$ . Both methods share a projection head  $\\pi$  for the semantic representation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            98,
+            85,
+            872,
+            244
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To achieve consistent feature representation in DGSS across various styles, we introduce semantic consistency contrastive learning. Specifically, the anchor is derived from the augmented image  $x_{a}$ , while the positive sample is extracted from the corresponding pixel of the original image  $x$ . Our method consists of two main components based on the negative sample as shown in Fig. 3: class-wise contrastive learning and semantic disentanglement contrastive learning.",
+        "bbox": [
+            75,
+            321,
+            470,
+            459
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Class-wise Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            76,
+            491,
+            308,
+            507
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our CWCL aims to build a discriminative embedding space for each segmentation class using different classes of the original image as negatives. Given an image pair  $(x, x_{a})$ , the features from the  $j^{th}$  block of the decoder at pixel position  $(m, n)$  are denoted as  $F_{(m, n)}^{j}$  and  $F_{a, (m, n)}^{j} \\in \\mathbb{R}^{(1 \\times 1) \\times C^{j}}$ , where  $C^{j}$  indicates the channel length of the feature. As mentioned above, we take  $F_{a, (m, n)}^{j}$  as the anchor and  $F_{(m, n)}^{j}$  as the positive sample, since they represent the same content at the corresponding spatial location. To obtain the negative samples from  $F^{j}$ , we leverage the resized segmentation class label  $y^{j} \\in \\mathbb{R}^{(H^{j} \\times W^{j}) \\times C}$  to collect the different class samples. The samples are passed through the projection head, denoted as  $\\pi$ , resulting in the projected features  $\\tilde{F}$  and  $\\tilde{F}_{a}$ . We define our CWCL as:",
+        "bbox": [
+            75,
+            518,
+            470,
+            743
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {C W C L} = \\sum_ {j} ^ {n _ {d}} \\mathcal {L} _ {I N} \\left(\\tilde {F} _ {a, (m, n)} ^ {j}, \\tilde {F} _ {(m, n)} ^ {j}, \\tilde {F} _ {(p, q)} ^ {j}\\right) \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            114,
+            758,
+            468,
+            799
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $(p,q)\\in \\{(p,q)\\in P|y_{(p,q)}^j\\neq y_{(m,n)}^j\\}$",
+        "bbox": [
+            122,
+            801,
+            429,
+            824
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The set  $P$  represents all pixel positions in feature  $F^j$ , with dimensions  $H^j \\times W^j$  corresponding to the height and width. Additionally,  $n_d$  denotes the total number of blocks in the decoder.",
+        "bbox": [
+            75,
+            839,
+            470,
+            898
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/364e0bcd00ec452f09a91d3c336a3abe80fbdab1f326b463d00f6dc6623bce24.jpg",
+        "image_caption": [
+            "Figure 4. t-SNE [49] visualization comparing scenarios with and without  $\\mathcal{L}_{SCCL}$ . In (b), the application of SCCL results in a clear separation between the sidewalk (pink), the road (purple), and the building (gray)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            540,
+            318,
+            694,
+            431
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/1185b7eb568978d73c2f83c1c0f9c882983f91b10edd00fd456149b4d75b6763.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            697,
+            329,
+            852,
+            431
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Semantic Disentanglement Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            500,
+            522,
+            843,
+            537
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Domain shifts can lead to the entanglement of similar classes, causing the model to misclassify, as illustrated in Fig. 4. To mitigate this issue, we introduce the SDCL, specifically designed to disentangle the feature  $x_{a}$  that has been misclassified, making it closer to the correct class and far from the misclassified class to achieve effective disentanglement. To further ensure consistent feature space and capture the semantic meaning, we share the projection head  $\\pi$  used in the CWCL loss. Given the predicted segmentation map of the augmented image, represented as  $\\hat{y}_a = \\varphi (x_a)$ , we resize it to  $\\hat{y}_{a,(m,n)}^j\\in \\mathbb{R}^{(H^j\\times W^j)\\times C}$ . Similarly,  $y_{(m,n)}^j$  represents the ground truth segmentation map. Using these segmentation maps, we set the anchor at positions where  $\\hat{y}_{a,(m,n)}^j\\neq y_{(m,n)}^j$ . Negative samples are selected from the augmented image features corresponding to the anchor's misclassified class. The samples go through the projection head  $\\pi$ . Our SDCL loss is defined as follows:",
+        "bbox": [
+            496,
+            546,
+            890,
+            803
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {S D C L} = \\sum_ {j} ^ {n _ {d}} \\mathcal {L} _ {I N} \\left(\\tilde {F} _ {a, (m, n)} ^ {j}, \\tilde {F} _ {(m, n)} ^ {j}, \\tilde {F} _ {a, (r, s)} ^ {j}\\right) \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            524,
+            811,
+            890,
+            853
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $(r,s)\\in \\{(r,s)\\in P|y_{(r,s)}^j = \\hat{y}_{(m,n)}^j\\}$",
+        "bbox": [
+            544,
+            857,
+            844,
+            877
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Finally, combining the cross-entropy segmentation loss",
+        "bbox": [
+            519,
+            885,
+            890,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "3620",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/644a98889183e0ef6b889a9e2199e4b2bb590eeae1aa7f4a56ec1ca5acd5517c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Backbone</td><td rowspan=\"2\">Methods</td><td colspan=\"2\">External</td><td colspan=\"4\">Trained on GTAV (G)</td><td colspan=\"4\">Trained on Cityscapes (C)</td></tr><tr><td>Dataset</td><td>Module</td><td>C</td><td>B</td><td>M</td><td>S</td><td>B</td><td>M</td><td>S</td><td>G</td></tr><tr><td rowspan=\"10\">ResNet50 [13]</td><td>Baseline [3]</td><td>-</td><td>-</td><td>28.95</td><td>25.14</td><td>28.18</td><td>26.23</td><td>44.96</td><td>51.68</td><td>23.29</td><td>42.55</td></tr><tr><td>IBN-Net [38]</td><td>-</td><td>-</td><td>33.85</td><td>32.30</td><td>37.75</td><td>27.90</td><td>48.56</td><td>57.04</td><td>26.14</td><td>45.06</td></tr><tr><td>RobustNet [7]</td><td>-</td><td>-</td><td>37.31</td><td>35.20</td><td>40.33</td><td>28.30</td><td>50.73</td><td>58.64</td><td>26.20</td><td>45.00</td></tr><tr><td>SiamDoGe [53]</td><td>-</td><td>-</td><td>42.96</td><td>37.54</td><td>40.64</td><td>28.34</td><td>51.53</td><td>59.00</td><td>26.67</td><td>45.08</td></tr><tr><td>DIRL [55]</td><td>-</td><td>✓</td><td>41.04</td><td>39.15</td><td>41.60</td><td>-</td><td>51.80</td><td>-</td><td>26.50</td><td>46.52</td></tr><tr><td>WildNet [24]</td><td>✓</td><td>-</td><td>44.62</td><td>38.42</td><td>46.09</td><td>31.34</td><td>50.94</td><td>58.79</td><td>27.95</td><td>47.01</td></tr><tr><td>SANSAW [42]</td><td>-</td><td>✓</td><td>39.75</td><td>37.34</td><td>41.86</td><td>30.79</td><td>52.95</td><td>59.81</td><td>28.32</td><td>47.28</td></tr><tr><td>SPC [20]</td><td>-</td><td>✓</td><td>44.10</td><td>40.46</td><td>45.51</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DPCL [56]</td><td>-</td><td>✓</td><td>44.74</td><td>40.59</td><td>46.33</td><td>30.81</td><td>50.97</td><td>58.59</td><td>25.85</td><td>46.00</td></tr><tr><td>Ours</td><td>-</td><td>-</td><td>45.72</td><td>41.32</td><td>47.08</td><td>31.39</td><td>51.84</td><td>60.18</td><td>28.51</td><td>47.97</td></tr><tr><td rowspan=\"7\">ShuffleNetV2 [32]</td><td>Baseline [3]</td><td>-</td><td>-</td><td>25.56</td><td>22.17</td><td>28.60</td><td>23.33</td><td>36.84</td><td>43.13</td><td>21.56</td><td>36.95</td></tr><tr><td>IBN-Net [38]</td><td>-</td><td>-</td><td>27.10</td><td>31.82</td><td>34.89</td><td>25.56</td><td>41.89</td><td>46.35</td><td>22.99</td><td>40.91</td></tr><tr><td>RobustNet [7]</td><td>-</td><td>-</td><td>30.98</td><td>32.06</td><td>35.31</td><td>24.31</td><td>41.94</td><td>46.97</td><td>22.82</td><td>40.17</td></tr><tr><td>SiamDoGe [53]</td><td>-</td><td>-</td><td>34.40</td><td>34.23</td><td>35.87</td><td>21.95</td><td>42.61</td><td>47.48</td><td>23.13</td><td>40.93</td></tr><tr><td>DIRL [55]</td><td>-</td><td>✓</td><td>31.88</td><td>32.57</td><td>36.12</td><td>-</td><td>42.55</td><td>-</td><td>23.74</td><td>41.23</td></tr><tr><td>DPCL [56]</td><td>-</td><td>✓</td><td>36.66</td><td>34.35</td><td>39.92</td><td>22.66</td><td>43.90</td><td>48.95</td><td>22.47</td><td>41.07</td></tr><tr><td>Ours</td><td>-</td><td>-</td><td>38.56</td><td>34.51</td><td>40.11</td><td>25.64</td><td>44.22</td><td>49.69</td><td>23.54</td><td>41.10</td></tr></table>",
+        "bbox": [
+            78,
+            87,
+            895,
+            366
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Quantitative comparison of mIoU (%) between DGSS methods. External dataset denotes the necessity of an auxiliary dataset during training and External module denotes the requirement of an additional module during inference. G, C, B, M, and S denote GTAV, Cityscapes, BDD100K, Mapillary, and SYNTHIA, respectively. The best and second-best results are bolded and underlined, respectively.",
+        "bbox": [
+            75,
+            375,
+            892,
+            417
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "$\\mathcal{L}_{CE}$  with other loss components, the total is defined as:",
+        "bbox": [
+            76,
+            444,
+            447,
+            459
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {C E} + \\omega_ {1} \\mathcal {L} _ {C M} + \\omega_ {2} \\mathcal {L} _ {C C} \\tag {9} \\\\ + \\omega_ {3} \\mathcal {L} _ {C W C L} + \\omega_ {4} \\mathcal {L} _ {S D C L} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            135,
+            468,
+            468,
+            502
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\omega_{1},\\omega_{2},\\omega_{3}$  , and  $\\omega_{4}$  denote the weighting factor of each loss functions.",
+        "bbox": [
+            75,
+            512,
+            468,
+            542
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Experiment",
+        "text_level": 1,
+        "bbox": [
+            76,
+            555,
+            200,
+            571
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we describe the implementation details, the experimental setup for comparison with existing DGSS methods, and the ablation study conducted to further validate the effectiveness of our approach.",
+        "bbox": [
+            75,
+            580,
+            468,
+            642
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            76,
+            648,
+            294,
+            666
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We adopt DeepLabV3+ [3] for the segmentation architecture and use ResNet-50 [13], ShuffleNetV2 [32], and MobileNetV2 [47] as the backbone network of the segmentation network. The model is trained for 40K iterations with a batch size of 8 using the SGD optimizer, which has a momentum of 0.9 and a weight decay of 5e-4. We employ a polynomial learning rate schedule with an initial rate of 1e-2 and a power of 0.9. For the simulation of domain shift, we augment the image  $x_{a}$  using strong color jittering transformation similar to [7]. The weighting parameters of (9),  $\\omega_{1}$ ,  $\\omega_{2}$ ,  $\\omega_{3}$  and  $\\omega_{4}$ , are set as 0.2, 0.2, 0.3, and 0.3 respectively.",
+        "bbox": [
+            75,
+            672,
+            468,
+            839
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Datasets",
+        "text_level": 1,
+        "bbox": [
+            76,
+            847,
+            179,
+            861
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We use two synthetic datasets (GTA [43] and SYNTHIA [45]), and three real-world datasets (Cityscapes [8],",
+        "bbox": [
+            75,
+            869,
+            468,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "BDD-100K [58], and Mapillary [33]) for the experiment. All segmentation labels are evaluated based on 19 object categories.",
+        "bbox": [
+            496,
+            444,
+            890,
+            489
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "GTAV (G) [43] is a large-scale dataset generated from the Grand Theft Auto V (GTAV) game engine. It comprises 24,966 images, split into 12,403 for training, 6,382 for validation, and 6,181 for testing with a resolution of  $1914 \\times 1052$ .",
+        "bbox": [
+            496,
+            489,
+            892,
+            563
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "SYNTHIA (S) [45] is a virtual, photo-realistic urban scene dataset comprising 9,400 images with a resolution of  $960 \\times 720$ . Among these, 2,820 images are designated for evaluation.",
+        "bbox": [
+            496,
+            565,
+            893,
+            625
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Cityscapes (C) [8] is a large-scale urban scene dataset captured from 50 cities, primarily in Germany. Particularly, it contains 5,000 high-resolution images with a resolution of  $2048 \\times 1024$ . The dataset is divided into 2,975 images for training, 500 for validation, and 1,525 for testing.",
+        "bbox": [
+            496,
+            626,
+            893,
+            700
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "BDD-100K (B) [58] is another real-world urban scene dataset that contains more diverse 10000 urban driving scene images with a resolution of  $1280 \\times 720$ . Specifically, the validation split (1,000 images) is used for evaluation.",
+        "bbox": [
+            496,
+            700,
+            890,
+            761
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Mapillary (M) [33] contains 25,000 images with a minimum resolution of  $1920 \\times 1080$ , collected from various locations worldwide. Specifically, the validation split of 2,000 images is used for evaluation.",
+        "bbox": [
+            496,
+            761,
+            890,
+            821
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Comparison with DGSS methods",
+        "text_level": 1,
+        "bbox": [
+            498,
+            832,
+            790,
+            848
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We compare our methods with other state-of-the-art DGSS methods: Baseline (DeepLabV3+ [3] trained on the source domain), IBN-Net [38], RobustNet [7], SiamDoGe [53],",
+        "bbox": [
+            496,
+            854,
+            892,
+            901
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "3621",
+        "bbox": [
+            482,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/a8c1edacd555b9aeb544e8b1e27336634ab38752437a712ef806a02175809d01.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td rowspan=\"2\">External Module</td><td colspan=\"4\">Trained on GTAV (G)</td></tr><tr><td>C</td><td>B</td><td>M</td><td>Mean</td></tr><tr><td>Baseline [3]</td><td></td><td>25.94</td><td>25.73</td><td>26.45</td><td>26.04</td></tr><tr><td>IBN-Net [38]</td><td></td><td>30.14</td><td>27.66</td><td>27.07</td><td>28.29</td></tr><tr><td>RobustNet [7]</td><td></td><td>30.86</td><td>30.05</td><td>30.67</td><td>30.52</td></tr><tr><td>SiamDoGe [53]</td><td></td><td>34.15</td><td>34.50</td><td>32.34</td><td>33.67</td></tr><tr><td>DIRL [55]</td><td>✓</td><td>34.67</td><td>32.78</td><td>34.31</td><td>33.92</td></tr><tr><td>DPCL [56]</td><td>✓</td><td>37.57</td><td>35.45</td><td>40.30</td><td>37.77</td></tr><tr><td>Ours</td><td></td><td>37.66</td><td>36.10</td><td>40.40</td><td>38.05</td></tr></table>",
+        "bbox": [
+            78,
+            88,
+            467,
+            214
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "DIRL [55], WildNet [24], SANSAW [42], SPC [20], and DPCL [56]. To evaluate the generalization ability of the model on arbitrary unseen domains, we conduct the experiment on two scenarios: i) trained on GTAV, tested on Cityscapes, BDD-100K, and Mapillary, and ii) trained on Cityscapes, tested on BDD-100K, Mapillary, and SYNTHIA. The quantitative results are computed with mean intersection over union (mIoU). Additionally, we compared the method trained on the backbone of ResNet-50 [13], ShuffleNetV2 [32], and MoblieNet [47], pre-trained on ImageNet [9].",
+        "bbox": [
+            75,
+            280,
+            467,
+            446
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Quantitative and Qualitative Results",
+        "text_level": 1,
+        "bbox": [
+            76,
+            468,
+            336,
+            483
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 1 summarizes the quantitative results. Our method outperforms all other methods when trained on GTAV, using ResNet-50 as the backbone. When compared with FN methods that remove domain-specific styles, we demonstrate that our approach minimizes the loss of content information. We also show that our method effectively shows generalization ability when trained on Cityscapes. We further evaluate our methods with different backbones, showing the wide applicability of our method. When trained with ShuffleNetV2, our method achieves the first or second-best performance among unseen target domains. Table 2 shows the results of our method trained on GTAV with MobileNetV2, demonstrating the superiority of our method.",
+        "bbox": [
+            75,
+            492,
+            467,
+            686
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "For qualitative evaluation, we compare the visual result between DGSS methods and ours. As depicted in Fig. 5, our method demonstrates superior results compared to other approaches, particularly in its overall prediction accuracy. Notably, our proposed techniques enable distinct prediction of features such as on road and sidewalk, yielding clearer segmentation boundaries. Please refer to the supplementary material for more qualitative results.",
+        "bbox": [
+            75,
+            688,
+            467,
+            809
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Computational cost analysis",
+        "text_level": 1,
+        "bbox": [
+            76,
+            830,
+            277,
+            845
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To confirm that our approach does not incur additional computational overhead, we provide the number of parameters, GFLOPS, and average inference time of each method. As",
+        "bbox": [
+            75,
+            854,
+            467,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/cde9f4931872aa504e6d8ec0edfca433b12f3ceb3ea1a2ee335bce871e1abc73.jpg",
+        "table_caption": [
+            "Table 2. Quantitative comparison of mIoU  $(\\%)$  using MobileNetV2 [47] backbone trained on the GTAV dataset."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>External Module</td><td>Params (M)</td><td>GFLOPS</td><td>Time (ms)</td></tr><tr><td>Baseline [13]</td><td></td><td>45.08</td><td>277.77</td><td>10.01</td></tr><tr><td>SANSAW [42]</td><td>✓</td><td>25.63</td><td>421.86</td><td>68.96</td></tr><tr><td>SPC [20]</td><td>✓</td><td>45.22</td><td>286.09</td><td>12.24</td></tr><tr><td>DIRL [55]</td><td>✓</td><td>45.41</td><td>278.11</td><td>11.69</td></tr><tr><td>DPCL [56]</td><td>✓</td><td>56.46</td><td>1188.64</td><td>823.78</td></tr><tr><td>Ours</td><td></td><td>45.08</td><td>277.78</td><td>10.03</td></tr></table>",
+        "bbox": [
+            501,
+            88,
+            890,
+            196
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/31a215e998aaa44615ecc9f72d54f3cf3b11c9da13b663716cf47cf64d5fa051.jpg",
+        "table_caption": [
+            "Table 3. Computational cost comparison conducted using DeepLabV3+ with a ResNet-50 backbone on an NVIDIA Tesla V100 GPU with an image resolution of  $2048 \\times 1024$ . Inference time is averaged over 400 trials."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>LCM</td><td>CC</td><td>LCWCL</td><td>SDCL</td><td>C</td><td>B</td><td>M</td></tr><tr><td></td><td></td><td></td><td></td><td>28.95</td><td>25.14</td><td>28.18</td></tr><tr><td>✓</td><td></td><td></td><td></td><td>38.08</td><td>36.65</td><td>40.62</td></tr><tr><td>✓</td><td>✓</td><td></td><td></td><td>40.42</td><td>37.81</td><td>43.91</td></tr><tr><td></td><td></td><td>✓</td><td></td><td>42.03</td><td>38.27</td><td>44.02</td></tr><tr><td></td><td></td><td>✓</td><td>✓</td><td>43.16</td><td>38.59</td><td>45.38</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td></td><td>43.17</td><td>38.23</td><td>44.84</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>45.72</td><td>41.32</td><td>47.08</td></tr></table>",
+        "bbox": [
+            501,
+            277,
+            890,
+            392
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4. Ablation study on proposed losses. The experiments were conducted using DeepLabV3+ with ResNet-50 backbone, trained on the GTAV dataset. The losses are detailed in  $\\mathcal{L}_{CM}$ : (3),  $\\mathcal{L}_{CC}$ : (5),  $\\mathcal{L}_{CWCL}$ : (7),  $\\mathcal{L}_{SDCL}$ : (8)",
+        "bbox": [
+            496,
+            402,
+            890,
+            459
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "detailed in Table 3, our method operates comparably to baseline models by learning features intrinsically without adopting a separate module.",
+        "bbox": [
+            496,
+            487,
+            890,
+            532
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4. Ablation Studies",
+        "text_level": 1,
+        "bbox": [
+            500,
+            544,
+            663,
+            558
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this subsection, we conducted a series of ablation studies to demonstrate the individual contribution and effectiveness of each component within our method. Specifically, we investigate the impact of the following components:  $\\mathcal{L}_{CM}$ ,  $\\mathcal{L}_{CC}$ ,  $\\mathcal{L}_{CWCL}$ ,  $\\mathcal{L}_{SDCL}$ . For the study, we use a scenario where the DeepLabV3+ model with backbone ResNet-50 model is trained on GTA and tested on Cityscapes, BDD100K, and Mapillary.",
+        "bbox": [
+            496,
+            566,
+            890,
+            688
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4 presents the impact of various proposed losses on domain generalization performance. Specifically, the baseline model, trained solely with cross-entropy loss, exhibits suboptimal performance on target domains because of overfitting to the source domain. Conversely, the integration of any proposed loss mechanisms leads to a marked enhancement in performance. More specifically, the incorporation of the covariance alignment  $(\\mathcal{L}_{CM},\\mathcal{L}_{CC})$  shows its efficacy in preserving essential content information by correlating features of paired images. The differential impact of the semantic consistency constrastive learning  $(\\mathcal{L}_{CWCL},\\mathcal{L}_{SDCL})$  is also evident, as it significantly aids in disentangling features of similar classes, thereby constructing a more robust embedding space.",
+        "bbox": [
+            496,
+            689,
+            890,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "3622",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/5743df130013ec2edea812c12d2a7e375fd2bbd88be13afab8d1a8d9272aea9c.jpg",
+        "image_caption": [
+            "Target Data"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            88,
+            192,
+            155
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/337f71e74c8508b424ee737397e719b5030e129644c3e795c1a4ea71ba7d1e6a.jpg",
+        "image_caption": [
+            "Baseline"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            196,
+            88,
+            310,
+            155
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/49c2a2eee91c74f79070a020372836d10140ed59f2a9e5dc8645dd7751c88c18.jpg",
+        "image_caption": [
+            "RobustNet"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            312,
+            88,
+            424,
+            155
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/95e5f7abd34e2a68c19572c4e0fa1674385464d3501bd35dfd4aa3a6bd004b39.jpg",
+        "image_caption": [
+            "WildNet"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            428,
+            88,
+            542,
+            155
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ba681e1508b8da213ef145d65f63839dc55f7d2eace8e17b8c6c0de9e132f18f.jpg",
+        "image_caption": [
+            "SPC"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            544,
+            88,
+            656,
+            155
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/56a57369c29ab1ee852bc71832d6b55c8c7c94a5be7f39e7bf518f526eb5fc04.jpg",
+        "image_caption": [
+            "Ours"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            658,
+            88,
+            772,
+            155
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e75373901ae793412627ba1d0ec111765d7b497f114899c12c47858de9371ae6.jpg",
+        "image_caption": [
+            "Ground Truth"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            776,
+            88,
+            890,
+            155
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/97b7de89b1fade3976ea20663122c5a175bbbba4817f004ecaee8fedc9661f42.jpg",
+        "image_caption": [
+            "Figure 5. Qualitative comparison between DGSS methods trained on GTAV (G) and tested on unseen target domains of Cityscapes (C) using DeeplabV3+ with ResNet50 backbone.",
+            "Figure 6. t-SNE [49] visualization comparing the covariance with and without  $\\mathcal{L}_{CML}$ ."
+        ],
+        "image_footnote": [
+            "Baseline - Original  $x$",
+            "Baseline - Augmented  $x_{a}$",
+            "Ours - Original  $x$",
+            "Ours-Augmented  $x_{a}$"
+        ],
+        "bbox": [
+            117,
+            232,
+            310,
+            347
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Covariance Matching Loss. Fig. 6 presents t-SNE plots of covariances for original and augmented images, before and after the application of CML. The baseline network perceives original and augmented images differently from a style perspective. However, after applying CML, the distribution becomes more intermixed, indicating that our proposed CML effectively ensures similar recognition of different style images.",
+        "bbox": [
+            75,
+            414,
+            468,
+            535
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Calculation of CCL. Table 5a demonstrates that our proposed cross-covariance method, which converges the diagonal components to 1, yields superior performance. As mentioned before, removing non-diagonal components, which contain content information actually degrades performance.",
+        "bbox": [
+            75,
+            536,
+            468,
+            611
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Sampling number in CWCL. Table 5b and Table 5c show the impact of varying the number of classes sampled per image and the number of samples per class in CWCL, respectively. As shown in Table 5b, the performance improves with an increase in the diversity of classes sampled in CWCL. This suggests that contrasting a broader array of classes enhances the model's discriminative capability. Furthermore, Table 5c demonstrates that a balanced number of negative samples per class leads to optimal performance.",
+        "bbox": [
+            75,
+            612,
+            470,
+            748
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Projection Head for SDCL. The influence of different project head configurations on the SDCL is investigated. We experimented with three distinct approaches: individual projection head, copying the weights of CWCL's (stop gradient), and shared projection head of CWCL. As demonstrated in Table 5d, sharing the projection head yielded the most superior results. The results indicate that SDCL not only relies on the semantic information from CWCL for effective disentanglement of similar classes but also enhances the embedding space learned by CWCL.",
+        "bbox": [
+            75,
+            750,
+            470,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/8690ec98ec23f09059bc8e81f3d41c1140c4b223d47e8a5b7f326677aa559de0.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"4\">Cross-covariance loss</td></tr><tr><td>Method</td><td>C</td><td>B</td><td>M</td></tr><tr><td>Whitening</td><td>38.68</td><td>36.91</td><td>42.12</td></tr><tr><td>Diagonal</td><td>40.42</td><td>37.81</td><td>43.91</td></tr></table>",
+        "bbox": [
+            500,
+            232,
+            709,
+            289
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/3d78f9782c24c7a5e8b367c88b83157faf680aa15989be82e67c3294f27d90c3.jpg",
+        "table_caption": [
+            "(a)"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"4\"># of classes</td></tr><tr><td>#</td><td>C</td><td>B</td><td>M</td></tr><tr><td>10</td><td>45.57</td><td>38.88</td><td>46.37</td></tr><tr><td>15</td><td>45.72</td><td>41.32</td><td>47.08</td></tr></table>",
+        "bbox": [
+            712,
+            232,
+            880,
+            289
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/02ede73468dbdebefc060f9a495d9c8eb9276b5f007fc4eb3a4ddbc706d8d9e6.jpg",
+        "table_caption": [
+            "(b)"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"4\"># of negative samples</td></tr><tr><td>#</td><td>C</td><td>B</td><td>M</td></tr><tr><td>10</td><td>44.76</td><td>38.21</td><td>46.46</td></tr><tr><td>50</td><td>45.72</td><td>41.32</td><td>47.08</td></tr><tr><td>100</td><td>44.44</td><td>39.14</td><td>46.29</td></tr></table>",
+        "bbox": [
+            506,
+            309,
+            669,
+            373
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/865e936593eaf7e7c54b4fd622556dd46df755addabcc6363305b953c7327d9d.jpg",
+        "table_caption": [
+            "(c)",
+            "Projection Head of SDCL"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>MLP</td><td>C</td><td>B</td><td>M</td></tr><tr><td>Individual</td><td>44.91</td><td>38.45</td><td>46.31</td></tr><tr><td>Shared (SG)</td><td>44.03</td><td>38.15</td><td>46.64</td></tr><tr><td>Shared</td><td>45.72</td><td>41.32</td><td>47.08</td></tr></table>",
+        "bbox": [
+            679,
+            321,
+            883,
+            373
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "(d)",
+        "bbox": [
+            774,
+            375,
+            789,
+            383
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5. Ablation studies. (a) Calculation of  $\\mathcal{L}_{CC}$ . (b) Number of classes for  $\\mathcal{L}_{CWCL}$ . (c) Number of negative samples for  $\\mathcal{L}_{CWCL}$ . (d) Projection head of  $\\mathcal{L}_{SDCL}$ . \"SG\" indicates stop gradient.",
+        "bbox": [
+            498,
+            395,
+            892,
+            436
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            462,
+            617,
+            478
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a novel BlindNet with covariance alignment and semantic consistency contrastive learning. By introducing covariance alignment, our method effectively addresses style variations, ensuring the extraction of features that are consistent across different styles. Furthermore, with the proposed semantic consistency contrastive learning, we not only facilitate the extraction of discriminative features but also enhance the generalization capabilities of the model in semantic segmentation predictions. Comprehensive experimental results validate the effectiveness of our approach, demonstrating its ability to generalize across multiple unseen target domains without requiring auxiliary domains or additional modules. Our future work will be improving and stabilizing the covariance alignment method.",
+        "bbox": [
+            496,
+            488,
+            892,
+            700
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement. This work was supported in part by the Basic Science Research Program through National Research Foundation of Korea (NRF) (Grants No. NRF2022R1F1A1073543), the MSIT(Ministry of Science and ICT), Korea, under the ICAN(ICT Challenge and Advanced Network of HRD) support program(RS-2022-00156385) supervised by the IITP(Institute for Information & Communications Technology Planning & Evaluation), and Innovative Human Resource Development for Local Intellectualization program through the Institute of Information & Communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT)(IITP-2024-00156287).",
+        "bbox": [
+            496,
+            703,
+            893,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "3623",
+        "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] Florent Bartoccioni, Éloi Zablocki, Andrei Bursuc, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lara: Latents and rays for multi-camera bird's-eye-view semantic segmentation. In Conference on Robot Learning, pages 1663-1672. PMLR, 2023. 1",
+            "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 3",
+            "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 6, 7",
+            "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3, 4",
+            "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1",
+            "[6] Wonwooong Cho, Sungha Choi, David Keetae Park, Inkyu Shin, and Jaegul Choo. Image-to-image translation via group-wise deep whitening-and-coloring transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10639-10647, 2019. 3",
+            "[7] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11580-11590, 2021. 2, 3, 4, 6, 7",
+            "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 6",
+            "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 7",
+            "[10] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2",
+            "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2",
+            "[12] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3",
+            "[13] 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, 7",
+            "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3",
+            "[15] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 1, 2",
+            "[16] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022. 2",
+            "[17] Lukas Hoyer, Dengxin Dai, Qin Wang, Yuhua Chen, and Luc Van Gool. Improving semi-supervised and domain-adaptive semantic segmentation with self-supervised depth estimation. International Journal of Computer Vision, pages 1-27, 2023. 1",
+            "[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1",
+            "[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 2",
+            "[20] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3061-3071, 2023. 6, 7",
+            "[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 3",
+            "[22] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 1, 2",
+            "[23] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 3",
+            "[24] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In Proceedings of the IEEE/CVF"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "3624",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Conference on Computer Vision and Pattern Recognition, pages 9936-9946, 2022. 2, 3, 6, 7",
+            "[25] Guangrui Li, Guoliang Kang, Wu Liu, Yunchao Wei, and Yi Yang. Content-consistent matching for domain adaptive semantic segmentation. In European conference on computer vision, pages 440-456. Springer, 2020. 1",
+            "[26] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2021. 2",
+            "[27] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 2",
+            "[28] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European conference on computer vision (ECCV), pages 624-639, 2018. 2",
+            "[29] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6936-6945, 2019. 2",
+            "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2",
+            "[31] Ping Luo. Learning deep architectures via generalized whitened neural networks. In International Conference on Machine Learning, pages 2238-2246. PMLR, 2017. 3",
+            "[32] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 6, 7",
+            "[33] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017. 6",
+            "[34] David Nilsson, Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Embodied visual active learning for semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2373–2383, 2021. 1",
+            "[35] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4501-4508. IEEE, 2021. 1",
+            "[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 4",
+            "[37] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3764-3773, 2020. 2"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[38] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 2, 3, 6, 7",
+            "[39] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1863-1871, 2019. 2, 3",
+            "[40] Taesung Park, Alexei A Efros, Richard Zhang, and JunYan Zhu. Contrastive learning for unpaired image-to-image translation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 319-345. Springer, 2020. 3",
+            "[41] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2, 3",
+            "[42] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2594-2605, 2022. 2, 4, 6, 7",
+            "[43] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 102-118. Springer, 2016. 1, 6",
+            "[44] 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. 1",
+            "[45] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 1, 6",
+            "[46] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10765-10775, 2021. 1",
+            "[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6, 7",
+            "[48] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 3",
+            "[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5, 8"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "3625",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[50] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in neural information processing systems, 31, 2018. 2",
+            "[51] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 1",
+            "[52] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021. 3",
+            "[53] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In European Conference on Computer Vision, pages 603-620. Springer, 2022. 2, 6, 7",
+            "[54] 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. 1",
+            "[55] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2884-2892, 2022. 3, 6, 7",
+            "[56] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. arXiv preprint arXiv:2303.01906, 2023. 3, 6, 7",
+            "[57] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9036-9045, 2019. 2",
+            "[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 6",
+            "[59] Fei Yu, Mo Zhang, Hexin Dong, Sheng Hu, Bin Dong, and Li Zhang. Dast: Unsupervised domain adaptation in semantic segmentation based on discriminator attention and self-training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10754–10762, 2021. 2",
+            "[60] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2100-2110, 2019. 2",
+            "[61] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Do"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            900
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "main generalization with mixstyle. In International Conference on Learning Representations, 2020. 2",
+            "[62] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 1, 2"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            189
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "3626",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/1e340b0f-2934-49e8-9933-238d4fb891ff_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:210ab1c261d3c8d8d2f0c4875b0e27e76ea5891f0ba80c40572122e79e3059f8
+size 3461227

2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/full.md

@@ -0,0 +1,358 @@
+# Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning
+
+Woo-Jin Ahn<sup>1</sup> Geun-Yeong Yang<sup>1</sup> Hyun-Duck Choi<sup>2*</sup> Myo-Taeg Lim<sup>1*</sup>  
+<sup>1</sup>Korea University <sup>2</sup>Chonnam National University
+
+$^{1}\{wjahn, hggofficial, mlim\} @korea.ac.kr$ $^{2}$ ducky.choi@jnu.ac.kr
+
+# Abstract
+
+Deep learning models for semantic segmentation often experience performance degradation when deployed to unseen target domains unidentified during the training phase. This is mainly due to variations in image texture (i.e. style) from different data sources. To tackle this challenge, existing domain generalized semantic segmentation (DGSS) methods attempt to remove style variations from the feature. However, these approaches struggle with the entanglement of style and content, which may lead to the unintentional removal of crucial content information, causing performance degradation. This study addresses this limitation by proposing BlindNet, a novel DGSS approach that blinds the style without external modules or datasets. The main idea behind our proposed approach is to alleviate the effect of style in the encoder whilst facilitating robust segmentation in the decoder. To achieve this, BlindNet comprises two key components: covariance alignment and semantic consistency contrastive learning. Specifically, the covariance alignment trains the encoder to uniformly recognize various styles and preserve the content information of the feature, rather than removing the style-sensitive factor. Meanwhile, semantic consistency contrastive learning enables the decoder to construct discriminative class embedding space and disentangles features that are vulnerable to misclassification. Through extensive experiments, our approach outperforms existing DGSS methods, exhibiting robustness and superior performance for semantic segmentation on unseen target domains. The code is available at https://github.com/root0yang/BlindNet.
+
+# 1. Introduction
+
+Semantic segmentation, a technique that classifies each pixel in an image into predefined categories, has garnered significant attention due to its potential applications in vari-
+
+![](images/f1ff12b29489673484c490594c0b0065fa431db390eaa18e1fb3b77b2196fbd2.jpg)  
+Figure 1. Comparison of semantic segmentation results between the baseline (DeepLabV3+ with ResNet50 backbone) and our BlindNet. Both models are trained on the source domain (GTAV [43]) and tested on the target domain (Cityscapes [8]).
+
+ous fields. Particularly, it plays a crucial role in autonomous driving [1, 18] and robotic systems [34, 35]. Besides, with the advent of large datasets, deep neural networks have emerged as a trending approach for semantic segmentation tasks, achieving impressive results [3, 5, 44, 54]. However, there remains a major bottleneck detailing the meticulous and labor-intensive process of dataset labeling. More particularly, this process not only consumes time but also poses economic challenges [8, 46]. To address this challenge, synthetic datasets have emerged as a compelling alternative. Specifically, these datasets, generated using three-dimensional (3D) rendering techniques, offer vast amounts of easily accessible data, eliminating the need for manual labeling [43, 45]. However, a challenge arises when models trained on synthetic datasets are deployed in real-world scenarios. More precisely, a domain shift problem arises due to style factor discrepancies (e.g. texture, illumination, and image quality) between synthetic and real-world data, which affects the performance of the model, as shown in Fig. 1.
+
+To address the domain shift problem, domain adaptation semantic segmentation (DASS) has been introduced [15, 17, 22, 25, 51, 62]. Specifically, DASS aims to bridge the gap between source and target domains by aligning their data distributions. However, a significant limitation of DASS is its dependency on the target domain during train-
+
+ing. For DASS to function effectively, target domain samples must be available during the training phase.
+
+Meanwhile, domain generalized semantic segmentation (DGSS) has been introduced as an alternative approach to tackle the domain shift problem. Unlike DASS, DGSS is trained only with the source domain, aiming to extract domain-invariant features. To achieve this, two main techniques have been employed: domain randomization (DR) and feature normalization (FN).
+
+DR augments the training set by introducing variability, either by altering the image style [41, 60] or by modifying the feature representation [24, 53]. By exposing the model to a wider variety of styles via DR, the network is less likely to overfit to the specific styles present in the training data. Consequently, the robustness of the model is improved, making it more adept at generalizing to new, unseen domains. Nevertheless, a crucial limitation of DR is its significant dependence on auxiliary domains.
+
+FN methods, converse to DR, regularize the features to prevent the model from overfitting to the distinct styles or characteristics of the training data. This is achieved by removing domain-specific style information using feature statistics, such as instance normalization [38] or whitening transformation [7, 19, 39, 42]. While these approaches effectively remove style-related information, they simultaneously pose the challenge of removing semantic content because content and style information are entangled. Consequently, the model fails to capture the essential patterns or features required for accurate segmentation prediction.
+
+To address this problem, we propose BlindNet, a model that blinds the style within the encoder and improves the robustness of the decoder, without requiring auxiliary datasets or external modules. Specifically, our proposed BlindNet consists of two components: covariance alignment for the encoder and semantic consistency contrastive learning for the decoder. Precisely, the covariance alignment facilitates the generation of style-invariant features with the proposed covariance matching loss function (CML) and the cross-covariance loss function (CCL). Specifically, CML mitigates the effects of style variations, while CCL focuses on preserving content information, effectively addressing the prevalent content information loss observed in the FN method. To further improve the generalization ability, we develop semantic consistency contrastive learning, which consists of class-wise contrastive learning (CWCL) and semantic disentanglement contrastive learning (SDCL). Particularly, the CWCL constructs a discriminative class embedding space, while SDCL disentangles features of similar classes that often lead to prediction errors. Extensive experiments across various datasets demonstrate that the proposed BlindNet outperforms existing DGSS methods.
+
+Our contributions are summarized as follows:
+
+- We propose a covariance alignment within the encoder,
+
+comprising CML and CCL. Specifically, the CML aims to mitigate the effects of style variations, while CCL ensures the preservation of content information, together facilitating the generation of style-agnostic features.
+
+- We propose semantic consistency contrastive learning within the decoder that comprises CWCL and SDCL, utilizing segmentation masks. Specifically, CWCL generates discriminative embeddings, whereas SDCL disentangles features of similar classes, enhancing the robustness of the model.  
+- Through extensive experiments, we demonstrate the superiority of our approach in DGSS, without the need to alter the network architecture or rely on external datasets.
+
+# 2. Related Works
+
+Domain adaptation and generalization for semantic segmentation. Domain adaptation (DA) aims at minimizing the distribution discrepancy between different domains, enabling a model to generalize from a source to a target domain. For DASS, adversarial training and cross-domain self-training strategies are commonly used. Particularly, adversarial-based methods [15, 29] employ generative adversarial networks [11] to close the feature distribution gap between source and target domains. Meanwhile, cross-domain self-training methods [16, 37, 59, 62] generate pseudo-labels for target domain data using pre-trained models, and employ them as training data, thereby expanding the training data and reducing distribution differences between the source and target domain.
+
+Domain generalization (DG) methods [26, 28, 30, 50, 61] aim to improve the generalization ability of the model without accessing the target domain during training. Since the difference in style of the image is the main cause of the disparity, most existing domain generalized semantic segmentation methods utilize the style information of the image for domain-invariant learning. Feature statistics (e.g. mean, variance, covariance, gram matrix, etc.) which are commonly used in style transfer [10, 21, 27, 57] are employed to capture the style information. Interestingly, existing DGSS methods can be separated into two parts: i) domain randomization to expand the distribution of style or ii) feature normalization to remove style.
+
+DR involves randomizing either the image or its features through stylization to learn domain-invariant features from various styles. For example, Peng et al. [42] extended the source domain data by stylizing images in the style of unreal paintings. Similarly, Yue et al. [60] and Huang et al. [22] attempted to enhance generalization by synthesizing images with diverse styles in the image space. In another study, Lee et al. [24] adopted ImageNet data [9] as wild data and performed randomization via synthesis in the feature space. Meanwhile, Wu et al. [53] diversified the trainable feature space by mixing the statistics of the feature and its color-
+
+jittered feature with Ada-IN [21].
+
+FN methods aim to remove domain-specific styles from features, extracting only domain-invariant content. For instance, Pan et al. [39] first attempted the DGSS method, combining batch normalization (BN) [23] and instance normalization (IN) [48] in the network layer. While BN preserves the content information within discriminative features, IN focuses on removing domain-specific style information from features. In a study, Choi et al. [7] addressed the limitations of previous whitening transformation [6, 31] that can eliminate the content information from the feature. Specifically, they proposed an instance-selective whitening approach designed to remove covariance components that are sensitive to domain shifts. Peng et al. [41] developed semantic-aware normalization that performs on class-wise and semantic-aware whitening that aligns channels based on the prediction through group whitening transformation [6]. In a study, Xu et al. [55] introduced the prior guided attention module and guided feature whitening to re-calibrate the feature and remove domain-specific style effects, respectively. Unlike the FN methods that directly remove the style component, our work explores a covariance alignment method that mitigates the effect of the style's effect while preserving the content information. Our method achieves the DGSS without any additional modules or auxiliary datasets.
+
+Contrastive Learning. Contrastive learning aims to learn representations by maximizing the similarity between positive pairs of samples while minimizing the similarity between negative pairs. In recent years, it has attracted significant attention for its effectiveness in learning discriminative representations across various tasks [2, 4, 12, 14]. Oord et al. [36] were the first to introduce the infoNCE loss, a type of contrastive loss function designed for self-contrastive learning. In a work, Park et al. [40] introduced patch-level contrastive learning for image translation, using co-located patches as positive pairs and spatially distant patches as negatives to maintain image context. For the semantic segmentation task, Wang et al. [52] introduced class-wise contrastive learning to aid the model in learning the embedding space of each class. Specifically, they sampled the classes existing in the images and applied contrastive learning based on the class label. For DGSS, Lee et al. [24] adopted contrastive learning to learn the ImageNet information in their model. Specifically, they set the ImageNet data as wild and applied the contrastive learning method by setting the wild-stylized feature and its closest wild content as positive samples. In another study, Yang et al. [56] developed multi-level contrastive learning, which designed instance prototypes and class prototypes for contrastive learning. Specifically, they sample each class's pixel features and apply contrastive learning with a transition-probability ability matrix. Unlike recent DGSS works that embed the origi
+
+nal image, we define contrastive learning for the augmented image. Specifically, the proposed method builds a robust embedding space by preserving the semantic consistency of the feature representation across various domains.
+
+# 3. Method
+
+The goal of the proposed method is to train a segmentation model  $\varphi$  on a given source domain  $S$  and generalize well to the unseen target domain  $T$ . Precisely, the source domain  $S = \{(x,y)\}$  contains a paired image  $x \in \mathbb{R}^{H \times W \times 3}$  and segmentation label  $y \in \mathbb{R}^{H \times W \times C}$ , where  $H$ ,  $W$ , and  $C$  denote the height of the image, the width of the image, and the class number of the segmentation map, respectively. The model  $\varphi$  takes an original  $x$  and its augmented counterpart  $x_{a}$ , which have the same content but different styles, and uses the feature information to enhance its generalization ability.
+
+As shown in Fig. 2, our method leverages the feature information through two main approaches: covariance alignment and semantic consistency contrastive learning. Specifically, the covariance matching ensures that features, having different styles but the same content, contain similar information. Additionally, the semantic consistency contrastive learning embeds the generalized features into discriminative representation based on the segmentation label.
+
+# 3.1. Covariance Alignment
+
+The domain shift in semantic segmentation results from changes in the visual characteristics of the image, known as style. This style information is typically detected in the shallow layers of networks [38], which are the encoders of the segmentation model. Based on this understanding, our method targets the encoder to handle style variations by employing the proposed covariance matching loss and cross-covariance loss function.
+
+# Covariance Matching Loss
+
+To train the network to uniformly recognize various styles without removing content information, we introduce the CML. Specifically, the loss aims to minimize the difference between covariance matrices derived from different styles of image features. Given an image pair  $(x,x_{a})$ , the features from  $i^{th}$  block of the encoder are represented as  $F^i\in \mathbb{R}^{(H^i\times W^i)\times C^i}$  and  $F_{a}^{i}\in \mathbb{R}^{(H^{i}\times W^{i})\times C^{i}}$ , respectively. Further, following the methodologies of [7, 41], we compute the covariance matrices using instance normalized features [48], which ensures consistent scaling across features. The feature maps are normalized and flattened into  $\bar{F}$  and  $\bar{F}_a\in \mathbb{R}^{(HW)\times C}$ , which are given by:
+
+$$
+\bar {F} = \frac {(F - \mu (F))}{\sigma (F)}, \bar {F} _ {a} = \frac {(F _ {a} - \mu (F _ {a}))}{\sigma (F _ {a})}, \tag {1}
+$$
+
+![](images/5fcbbb724d3720e68e61b5e9b5a91dabb76d3372b3e54fd809de935bffa75425.jpg)  
+Figure 2. Overview of the proposed BlindNet. The network processes a pair of images - the original image  $x$  and its augmented counterpart  $x_{a}$ . It employs covariance alignment to treat encoder features and utilizes semantic consistency contrastive learning for the processing of decoder features.
+
+where  $\mu (\cdot)\in \mathbb{R}^{C}$  and  $\sigma (\cdot)\in \mathbb{R}^C$  denote the mean and standard deviation of the features. Utilizing the normalized features, we evaluate the covariance matrices for the original and augmented image features as:
+
+$$
+\Sigma_ {x, x} ^ {i} = \bar {F} ^ {i} ^ {T} \cdot \bar {F} ^ {i}, \quad \Sigma_ {x _ {a}, x _ {a}} ^ {i} = \bar {F} _ {a} ^ {i} ^ {T} \cdot \bar {F} _ {a} ^ {i}. \tag {2}
+$$
+
+The CML is then formulated to align these covariance matrices, ensuring that the network maintains consistency in the presence of style variations. The CML is defined as:
+
+$$
+\mathcal {L} _ {C M} = \sum_ {i = 1} ^ {n _ {e}} \| \Sigma_ {x, x} ^ {i} - \Sigma_ {x _ {a}, x _ {a}} ^ {i} \| _ {2}, \tag {3}
+$$
+
+where  $n_e$  denotes the number of blocks of the encoder.
+
+# Cross-covariance Loss
+
+While the CML effectively aligns the internal distributions of features within the same image, it does not fully account for the direct correlations across paired images. To complement this, we introduce CCL, which aims to encode the consistent content information of an image pair  $(x,x_{a})$  by utilizing the cross-covariance of the image pair. Given the normalized feature pair  $(\bar{F}^i,\bar{F}_a^i)$ , the cross-covariance of the feature pair can be expressed as:
+
+$$
+\Sigma_ {x, x _ {a}} ^ {i} = \bar {F} ^ {i} ^ {T} \cdot \bar {F} _ {a} ^ {i}. \tag {4}
+$$
+
+The cross-covariance is expected to exhibit an identity matrix, as the feature pair should contain identical information. Nonetheless, the proposed CCL converges only the diagonal component of the covariance matrix to one. This is to prevent the drawbacks of the existing FN methods [7, 42] from removing content information. The CCL function is thus defined as:
+
+$$
+\mathcal {L} _ {C C} = \sum_ {i = 1} ^ {n _ {e}} \| \operatorname {d i a g} \left(\Sigma_ {x, x _ {a}} ^ {i}\right) - \mathbb {1} \| _ {2}, \tag {5}
+$$
+
+where  $\mathrm{diag}(\Sigma_c) \in \mathbb{R}^C$  denotes the column vector comprising diagonal elements of  $\Sigma_{x,x_a}^i$  and  $\mathbb{1} \in \mathbb{R}^C$  denotes the one vector.
+
+# 3.2. Semantic Consistence Contrastive Learning
+
+While the encoder focuses on generating style-blinded features, the decoder aims to improve the robustness of the segmentation prediction against domain shifts. For the decoder, we employ a contrastive learning approach, which has demonstrated effectiveness in extracting discriminative features [4]. Specifically, we utilize the InfoNCE loss [36], which is formulated as:
+
+$$
+\mathcal {L} _ {I N} (a, p, n) = - \log \left(\frac {e ^ {(a \cdot p / \tau)}}{e ^ {(a \cdot p / \tau)} + \sum_ {n} ^ {N ^ {-}} e ^ {(a \cdot n / \tau)}}\right) \tag {6}
+$$
+
+where  $a, p, n$ , and  $N^{-}$  denote anchor, positive sample, negative sample, and negative sample set, respectively.
+
+![](images/f9fb00733b6b0939d8e914bad891b9b9d71654c5baf4d18638084e017a3d791f.jpg)  
+Figure 3. Illustration of semantic consistency contrastive learning: The mask  $M$  represents the error mask derived from the augmented segmentation map. CWCL conducts contrastive learning by sampling per segmentation class and SDCL conducts contrastive learning based on the  $M$ . Both methods share a projection head  $\pi$  for the semantic representation.
+
+To achieve consistent feature representation in DGSS across various styles, we introduce semantic consistency contrastive learning. Specifically, the anchor is derived from the augmented image  $x_{a}$ , while the positive sample is extracted from the corresponding pixel of the original image  $x$ . Our method consists of two main components based on the negative sample as shown in Fig. 3: class-wise contrastive learning and semantic disentanglement contrastive learning.
+
+# Class-wise Contrastive Learning
+
+Our CWCL aims to build a discriminative embedding space for each segmentation class using different classes of the original image as negatives. Given an image pair  $(x, x_{a})$ , the features from the  $j^{th}$  block of the decoder at pixel position  $(m, n)$  are denoted as  $F_{(m, n)}^{j}$  and  $F_{a, (m, n)}^{j} \in \mathbb{R}^{(1 \times 1) \times C^{j}}$ , where  $C^{j}$  indicates the channel length of the feature. As mentioned above, we take  $F_{a, (m, n)}^{j}$  as the anchor and  $F_{(m, n)}^{j}$  as the positive sample, since they represent the same content at the corresponding spatial location. To obtain the negative samples from  $F^{j}$ , we leverage the resized segmentation class label  $y^{j} \in \mathbb{R}^{(H^{j} \times W^{j}) \times C}$  to collect the different class samples. The samples are passed through the projection head, denoted as  $\pi$ , resulting in the projected features  $\tilde{F}$  and  $\tilde{F}_{a}$ . We define our CWCL as:
+
+$$
+\mathcal {L} _ {C W C L} = \sum_ {j} ^ {n _ {d}} \mathcal {L} _ {I N} \left(\tilde {F} _ {a, (m, n)} ^ {j}, \tilde {F} _ {(m, n)} ^ {j}, \tilde {F} _ {(p, q)} ^ {j}\right) \tag {7}
+$$
+
+where  $(p,q)\in \{(p,q)\in P|y_{(p,q)}^j\neq y_{(m,n)}^j\}$
+
+The set  $P$  represents all pixel positions in feature  $F^j$ , with dimensions  $H^j \times W^j$  corresponding to the height and width. Additionally,  $n_d$  denotes the total number of blocks in the decoder.
+
+![](images/364e0bcd00ec452f09a91d3c336a3abe80fbdab1f326b463d00f6dc6623bce24.jpg)  
+Figure 4. t-SNE [49] visualization comparing scenarios with and without  $\mathcal{L}_{SCCL}$ . In (b), the application of SCCL results in a clear separation between the sidewalk (pink), the road (purple), and the building (gray).
+
+![](images/1185b7eb568978d73c2f83c1c0f9c882983f91b10edd00fd456149b4d75b6763.jpg)
+
+# Semantic Disentanglement Contrastive Learning
+
+Domain shifts can lead to the entanglement of similar classes, causing the model to misclassify, as illustrated in Fig. 4. To mitigate this issue, we introduce the SDCL, specifically designed to disentangle the feature  $x_{a}$  that has been misclassified, making it closer to the correct class and far from the misclassified class to achieve effective disentanglement. To further ensure consistent feature space and capture the semantic meaning, we share the projection head  $\pi$  used in the CWCL loss. Given the predicted segmentation map of the augmented image, represented as  $\hat{y}_a = \varphi (x_a)$ , we resize it to  $\hat{y}_{a,(m,n)}^j\in \mathbb{R}^{(H^j\times W^j)\times C}$ . Similarly,  $y_{(m,n)}^j$  represents the ground truth segmentation map. Using these segmentation maps, we set the anchor at positions where  $\hat{y}_{a,(m,n)}^j\neq y_{(m,n)}^j$ . Negative samples are selected from the augmented image features corresponding to the anchor's misclassified class. The samples go through the projection head  $\pi$ . Our SDCL loss is defined as follows:
+
+$$
+\mathcal {L} _ {S D C L} = \sum_ {j} ^ {n _ {d}} \mathcal {L} _ {I N} \left(\tilde {F} _ {a, (m, n)} ^ {j}, \tilde {F} _ {(m, n)} ^ {j}, \tilde {F} _ {a, (r, s)} ^ {j}\right) \tag {8}
+$$
+
+where  $(r,s)\in \{(r,s)\in P|y_{(r,s)}^j = \hat{y}_{(m,n)}^j\}$
+
+Finally, combining the cross-entropy segmentation loss
+
+<table><tr><td rowspan="2">Backbone</td><td rowspan="2">Methods</td><td colspan="2">External</td><td colspan="4">Trained on GTAV (G)</td><td colspan="4">Trained on Cityscapes (C)</td></tr><tr><td>Dataset</td><td>Module</td><td>C</td><td>B</td><td>M</td><td>S</td><td>B</td><td>M</td><td>S</td><td>G</td></tr><tr><td rowspan="10">ResNet50 [13]</td><td>Baseline [3]</td><td>-</td><td>-</td><td>28.95</td><td>25.14</td><td>28.18</td><td>26.23</td><td>44.96</td><td>51.68</td><td>23.29</td><td>42.55</td></tr><tr><td>IBN-Net [38]</td><td>-</td><td>-</td><td>33.85</td><td>32.30</td><td>37.75</td><td>27.90</td><td>48.56</td><td>57.04</td><td>26.14</td><td>45.06</td></tr><tr><td>RobustNet [7]</td><td>-</td><td>-</td><td>37.31</td><td>35.20</td><td>40.33</td><td>28.30</td><td>50.73</td><td>58.64</td><td>26.20</td><td>45.00</td></tr><tr><td>SiamDoGe [53]</td><td>-</td><td>-</td><td>42.96</td><td>37.54</td><td>40.64</td><td>28.34</td><td>51.53</td><td>59.00</td><td>26.67</td><td>45.08</td></tr><tr><td>DIRL [55]</td><td>-</td><td>✓</td><td>41.04</td><td>39.15</td><td>41.60</td><td>-</td><td>51.80</td><td>-</td><td>26.50</td><td>46.52</td></tr><tr><td>WildNet [24]</td><td>✓</td><td>-</td><td>44.62</td><td>38.42</td><td>46.09</td><td>31.34</td><td>50.94</td><td>58.79</td><td>27.95</td><td>47.01</td></tr><tr><td>SANSAW [42]</td><td>-</td><td>✓</td><td>39.75</td><td>37.34</td><td>41.86</td><td>30.79</td><td>52.95</td><td>59.81</td><td>28.32</td><td>47.28</td></tr><tr><td>SPC [20]</td><td>-</td><td>✓</td><td>44.10</td><td>40.46</td><td>45.51</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DPCL [56]</td><td>-</td><td>✓</td><td>44.74</td><td>40.59</td><td>46.33</td><td>30.81</td><td>50.97</td><td>58.59</td><td>25.85</td><td>46.00</td></tr><tr><td>Ours</td><td>-</td><td>-</td><td>45.72</td><td>41.32</td><td>47.08</td><td>31.39</td><td>51.84</td><td>60.18</td><td>28.51</td><td>47.97</td></tr><tr><td rowspan="7">ShuffleNetV2 [32]</td><td>Baseline [3]</td><td>-</td><td>-</td><td>25.56</td><td>22.17</td><td>28.60</td><td>23.33</td><td>36.84</td><td>43.13</td><td>21.56</td><td>36.95</td></tr><tr><td>IBN-Net [38]</td><td>-</td><td>-</td><td>27.10</td><td>31.82</td><td>34.89</td><td>25.56</td><td>41.89</td><td>46.35</td><td>22.99</td><td>40.91</td></tr><tr><td>RobustNet [7]</td><td>-</td><td>-</td><td>30.98</td><td>32.06</td><td>35.31</td><td>24.31</td><td>41.94</td><td>46.97</td><td>22.82</td><td>40.17</td></tr><tr><td>SiamDoGe [53]</td><td>-</td><td>-</td><td>34.40</td><td>34.23</td><td>35.87</td><td>21.95</td><td>42.61</td><td>47.48</td><td>23.13</td><td>40.93</td></tr><tr><td>DIRL [55]</td><td>-</td><td>✓</td><td>31.88</td><td>32.57</td><td>36.12</td><td>-</td><td>42.55</td><td>-</td><td>23.74</td><td>41.23</td></tr><tr><td>DPCL [56]</td><td>-</td><td>✓</td><td>36.66</td><td>34.35</td><td>39.92</td><td>22.66</td><td>43.90</td><td>48.95</td><td>22.47</td><td>41.07</td></tr><tr><td>Ours</td><td>-</td><td>-</td><td>38.56</td><td>34.51</td><td>40.11</td><td>25.64</td><td>44.22</td><td>49.69</td><td>23.54</td><td>41.10</td></tr></table>
+
+Table 1. Quantitative comparison of mIoU (%) between DGSS methods. External dataset denotes the necessity of an auxiliary dataset during training and External module denotes the requirement of an additional module during inference. G, C, B, M, and S denote GTAV, Cityscapes, BDD100K, Mapillary, and SYNTHIA, respectively. The best and second-best results are bolded and underlined, respectively.
+
+$\mathcal{L}_{CE}$  with other loss components, the total is defined as:
+
+$$
+\begin{array}{l} \mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {C E} + \omega_ {1} \mathcal {L} _ {C M} + \omega_ {2} \mathcal {L} _ {C C} \tag {9} \\ + \omega_ {3} \mathcal {L} _ {C W C L} + \omega_ {4} \mathcal {L} _ {S D C L} \\ \end{array}
+$$
+
+where  $\omega_{1},\omega_{2},\omega_{3}$  , and  $\omega_{4}$  denote the weighting factor of each loss functions.
+
+# 4. Experiment
+
+In this section, we describe the implementation details, the experimental setup for comparison with existing DGSS methods, and the ablation study conducted to further validate the effectiveness of our approach.
+
+# 4.1. Implementation Details
+
+We adopt DeepLabV3+ [3] for the segmentation architecture and use ResNet-50 [13], ShuffleNetV2 [32], and MobileNetV2 [47] as the backbone network of the segmentation network. The model is trained for 40K iterations with a batch size of 8 using the SGD optimizer, which has a momentum of 0.9 and a weight decay of 5e-4. We employ a polynomial learning rate schedule with an initial rate of 1e-2 and a power of 0.9. For the simulation of domain shift, we augment the image  $x_{a}$  using strong color jittering transformation similar to [7]. The weighting parameters of (9),  $\omega_{1}$ ,  $\omega_{2}$ ,  $\omega_{3}$  and  $\omega_{4}$ , are set as 0.2, 0.2, 0.3, and 0.3 respectively.
+
+# 4.2. Datasets
+
+We use two synthetic datasets (GTA [43] and SYNTHIA [45]), and three real-world datasets (Cityscapes [8],
+
+BDD-100K [58], and Mapillary [33]) for the experiment. All segmentation labels are evaluated based on 19 object categories.
+
+GTAV (G) [43] is a large-scale dataset generated from the Grand Theft Auto V (GTAV) game engine. It comprises 24,966 images, split into 12,403 for training, 6,382 for validation, and 6,181 for testing with a resolution of  $1914 \times 1052$ .
+
+SYNTHIA (S) [45] is a virtual, photo-realistic urban scene dataset comprising 9,400 images with a resolution of  $960 \times 720$ . Among these, 2,820 images are designated for evaluation.
+
+Cityscapes (C) [8] is a large-scale urban scene dataset captured from 50 cities, primarily in Germany. Particularly, it contains 5,000 high-resolution images with a resolution of  $2048 \times 1024$ . The dataset is divided into 2,975 images for training, 500 for validation, and 1,525 for testing.
+
+BDD-100K (B) [58] is another real-world urban scene dataset that contains more diverse 10000 urban driving scene images with a resolution of  $1280 \times 720$ . Specifically, the validation split (1,000 images) is used for evaluation.
+
+Mapillary (M) [33] contains 25,000 images with a minimum resolution of  $1920 \times 1080$ , collected from various locations worldwide. Specifically, the validation split of 2,000 images is used for evaluation.
+
+# 4.3. Comparison with DGSS methods
+
+We compare our methods with other state-of-the-art DGSS methods: Baseline (DeepLabV3+ [3] trained on the source domain), IBN-Net [38], RobustNet [7], SiamDoGe [53],
+
+<table><tr><td rowspan="2">Methods</td><td rowspan="2">External Module</td><td colspan="4">Trained on GTAV (G)</td></tr><tr><td>C</td><td>B</td><td>M</td><td>Mean</td></tr><tr><td>Baseline [3]</td><td></td><td>25.94</td><td>25.73</td><td>26.45</td><td>26.04</td></tr><tr><td>IBN-Net [38]</td><td></td><td>30.14</td><td>27.66</td><td>27.07</td><td>28.29</td></tr><tr><td>RobustNet [7]</td><td></td><td>30.86</td><td>30.05</td><td>30.67</td><td>30.52</td></tr><tr><td>SiamDoGe [53]</td><td></td><td>34.15</td><td>34.50</td><td>32.34</td><td>33.67</td></tr><tr><td>DIRL [55]</td><td>✓</td><td>34.67</td><td>32.78</td><td>34.31</td><td>33.92</td></tr><tr><td>DPCL [56]</td><td>✓</td><td>37.57</td><td>35.45</td><td>40.30</td><td>37.77</td></tr><tr><td>Ours</td><td></td><td>37.66</td><td>36.10</td><td>40.40</td><td>38.05</td></tr></table>
+
+DIRL [55], WildNet [24], SANSAW [42], SPC [20], and DPCL [56]. To evaluate the generalization ability of the model on arbitrary unseen domains, we conduct the experiment on two scenarios: i) trained on GTAV, tested on Cityscapes, BDD-100K, and Mapillary, and ii) trained on Cityscapes, tested on BDD-100K, Mapillary, and SYNTHIA. The quantitative results are computed with mean intersection over union (mIoU). Additionally, we compared the method trained on the backbone of ResNet-50 [13], ShuffleNetV2 [32], and MoblieNet [47], pre-trained on ImageNet [9].
+
+# Quantitative and Qualitative Results
+
+Table 1 summarizes the quantitative results. Our method outperforms all other methods when trained on GTAV, using ResNet-50 as the backbone. When compared with FN methods that remove domain-specific styles, we demonstrate that our approach minimizes the loss of content information. We also show that our method effectively shows generalization ability when trained on Cityscapes. We further evaluate our methods with different backbones, showing the wide applicability of our method. When trained with ShuffleNetV2, our method achieves the first or second-best performance among unseen target domains. Table 2 shows the results of our method trained on GTAV with MobileNetV2, demonstrating the superiority of our method.
+
+For qualitative evaluation, we compare the visual result between DGSS methods and ours. As depicted in Fig. 5, our method demonstrates superior results compared to other approaches, particularly in its overall prediction accuracy. Notably, our proposed techniques enable distinct prediction of features such as on road and sidewalk, yielding clearer segmentation boundaries. Please refer to the supplementary material for more qualitative results.
+
+# Computational cost analysis
+
+To confirm that our approach does not incur additional computational overhead, we provide the number of parameters, GFLOPS, and average inference time of each method. As
+
+Table 2. Quantitative comparison of mIoU  $(\%)$  using MobileNetV2 [47] backbone trained on the GTAV dataset.  
+
+<table><tr><td>Methods</td><td>External Module</td><td>Params (M)</td><td>GFLOPS</td><td>Time (ms)</td></tr><tr><td>Baseline [13]</td><td></td><td>45.08</td><td>277.77</td><td>10.01</td></tr><tr><td>SANSAW [42]</td><td>✓</td><td>25.63</td><td>421.86</td><td>68.96</td></tr><tr><td>SPC [20]</td><td>✓</td><td>45.22</td><td>286.09</td><td>12.24</td></tr><tr><td>DIRL [55]</td><td>✓</td><td>45.41</td><td>278.11</td><td>11.69</td></tr><tr><td>DPCL [56]</td><td>✓</td><td>56.46</td><td>1188.64</td><td>823.78</td></tr><tr><td>Ours</td><td></td><td>45.08</td><td>277.78</td><td>10.03</td></tr></table>
+
+Table 3. Computational cost comparison conducted using DeepLabV3+ with a ResNet-50 backbone on an NVIDIA Tesla V100 GPU with an image resolution of  $2048 \times 1024$ . Inference time is averaged over 400 trials.  
+
+<table><tr><td>LCM</td><td>CC</td><td>LCWCL</td><td>SDCL</td><td>C</td><td>B</td><td>M</td></tr><tr><td></td><td></td><td></td><td></td><td>28.95</td><td>25.14</td><td>28.18</td></tr><tr><td>✓</td><td></td><td></td><td></td><td>38.08</td><td>36.65</td><td>40.62</td></tr><tr><td>✓</td><td>✓</td><td></td><td></td><td>40.42</td><td>37.81</td><td>43.91</td></tr><tr><td></td><td></td><td>✓</td><td></td><td>42.03</td><td>38.27</td><td>44.02</td></tr><tr><td></td><td></td><td>✓</td><td>✓</td><td>43.16</td><td>38.59</td><td>45.38</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td></td><td>43.17</td><td>38.23</td><td>44.84</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>45.72</td><td>41.32</td><td>47.08</td></tr></table>
+
+Table 4. Ablation study on proposed losses. The experiments were conducted using DeepLabV3+ with ResNet-50 backbone, trained on the GTAV dataset. The losses are detailed in  $\mathcal{L}_{CM}$ : (3),  $\mathcal{L}_{CC}$ : (5),  $\mathcal{L}_{CWCL}$ : (7),  $\mathcal{L}_{SDCL}$ : (8)
+
+detailed in Table 3, our method operates comparably to baseline models by learning features intrinsically without adopting a separate module.
+
+# 4.4. Ablation Studies
+
+In this subsection, we conducted a series of ablation studies to demonstrate the individual contribution and effectiveness of each component within our method. Specifically, we investigate the impact of the following components:  $\mathcal{L}_{CM}$ ,  $\mathcal{L}_{CC}$ ,  $\mathcal{L}_{CWCL}$ ,  $\mathcal{L}_{SDCL}$ . For the study, we use a scenario where the DeepLabV3+ model with backbone ResNet-50 model is trained on GTA and tested on Cityscapes, BDD100K, and Mapillary.
+
+Table 4 presents the impact of various proposed losses on domain generalization performance. Specifically, the baseline model, trained solely with cross-entropy loss, exhibits suboptimal performance on target domains because of overfitting to the source domain. Conversely, the integration of any proposed loss mechanisms leads to a marked enhancement in performance. More specifically, the incorporation of the covariance alignment  $(\mathcal{L}_{CM},\mathcal{L}_{CC})$  shows its efficacy in preserving essential content information by correlating features of paired images. The differential impact of the semantic consistency constrastive learning  $(\mathcal{L}_{CWCL},\mathcal{L}_{SDCL})$  is also evident, as it significantly aids in disentangling features of similar classes, thereby constructing a more robust embedding space.
+
+![](images/5743df130013ec2edea812c12d2a7e375fd2bbd88be13afab8d1a8d9272aea9c.jpg)  
+Target Data
+
+![](images/337f71e74c8508b424ee737397e719b5030e129644c3e795c1a4ea71ba7d1e6a.jpg)  
+Baseline
+
+![](images/49c2a2eee91c74f79070a020372836d10140ed59f2a9e5dc8645dd7751c88c18.jpg)  
+RobustNet
+
+![](images/95e5f7abd34e2a68c19572c4e0fa1674385464d3501bd35dfd4aa3a6bd004b39.jpg)  
+WildNet
+
+![](images/ba681e1508b8da213ef145d65f63839dc55f7d2eace8e17b8c6c0de9e132f18f.jpg)  
+SPC
+
+![](images/56a57369c29ab1ee852bc71832d6b55c8c7c94a5be7f39e7bf518f526eb5fc04.jpg)  
+Ours
+
+![](images/e75373901ae793412627ba1d0ec111765d7b497f114899c12c47858de9371ae6.jpg)  
+Ground Truth
+
+Figure 5. Qualitative comparison between DGSS methods trained on GTAV (G) and tested on unseen target domains of Cityscapes (C) using DeeplabV3+ with ResNet50 backbone.  
+Figure 6. t-SNE [49] visualization comparing the covariance with and without  $\mathcal{L}_{CML}$ .  
+![](images/97b7de89b1fade3976ea20663122c5a175bbbba4817f004ecaee8fedc9661f42.jpg)  
+Baseline - Original  $x$  
+Baseline - Augmented  $x_{a}$  
+Ours - Original  $x$  
+Ours-Augmented  $x_{a}$
+
+Covariance Matching Loss. Fig. 6 presents t-SNE plots of covariances for original and augmented images, before and after the application of CML. The baseline network perceives original and augmented images differently from a style perspective. However, after applying CML, the distribution becomes more intermixed, indicating that our proposed CML effectively ensures similar recognition of different style images.
+
+Calculation of CCL. Table 5a demonstrates that our proposed cross-covariance method, which converges the diagonal components to 1, yields superior performance. As mentioned before, removing non-diagonal components, which contain content information actually degrades performance.
+
+Sampling number in CWCL. Table 5b and Table 5c show the impact of varying the number of classes sampled per image and the number of samples per class in CWCL, respectively. As shown in Table 5b, the performance improves with an increase in the diversity of classes sampled in CWCL. This suggests that contrasting a broader array of classes enhances the model's discriminative capability. Furthermore, Table 5c demonstrates that a balanced number of negative samples per class leads to optimal performance.
+
+Projection Head for SDCL. The influence of different project head configurations on the SDCL is investigated. We experimented with three distinct approaches: individual projection head, copying the weights of CWCL's (stop gradient), and shared projection head of CWCL. As demonstrated in Table 5d, sharing the projection head yielded the most superior results. The results indicate that SDCL not only relies on the semantic information from CWCL for effective disentanglement of similar classes but also enhances the embedding space learned by CWCL.
+
+<table><tr><td colspan="4">Cross-covariance loss</td></tr><tr><td>Method</td><td>C</td><td>B</td><td>M</td></tr><tr><td>Whitening</td><td>38.68</td><td>36.91</td><td>42.12</td></tr><tr><td>Diagonal</td><td>40.42</td><td>37.81</td><td>43.91</td></tr></table>
+
+(a)  
+
+<table><tr><td colspan="4"># of classes</td></tr><tr><td>#</td><td>C</td><td>B</td><td>M</td></tr><tr><td>10</td><td>45.57</td><td>38.88</td><td>46.37</td></tr><tr><td>15</td><td>45.72</td><td>41.32</td><td>47.08</td></tr></table>
+
+(b)  
+
+<table><tr><td colspan="4"># of negative samples</td></tr><tr><td>#</td><td>C</td><td>B</td><td>M</td></tr><tr><td>10</td><td>44.76</td><td>38.21</td><td>46.46</td></tr><tr><td>50</td><td>45.72</td><td>41.32</td><td>47.08</td></tr><tr><td>100</td><td>44.44</td><td>39.14</td><td>46.29</td></tr></table>
+
+(c)  
+Projection Head of SDCL  
+
+<table><tr><td>MLP</td><td>C</td><td>B</td><td>M</td></tr><tr><td>Individual</td><td>44.91</td><td>38.45</td><td>46.31</td></tr><tr><td>Shared (SG)</td><td>44.03</td><td>38.15</td><td>46.64</td></tr><tr><td>Shared</td><td>45.72</td><td>41.32</td><td>47.08</td></tr></table>
+
+(d)
+
+Table 5. Ablation studies. (a) Calculation of  $\mathcal{L}_{CC}$ . (b) Number of classes for  $\mathcal{L}_{CWCL}$ . (c) Number of negative samples for  $\mathcal{L}_{CWCL}$ . (d) Projection head of  $\mathcal{L}_{SDCL}$ . "SG" indicates stop gradient.
+
+# 5. Conclusion
+
+In this paper, we propose a novel BlindNet with covariance alignment and semantic consistency contrastive learning. By introducing covariance alignment, our method effectively addresses style variations, ensuring the extraction of features that are consistent across different styles. Furthermore, with the proposed semantic consistency contrastive learning, we not only facilitate the extraction of discriminative features but also enhance the generalization capabilities of the model in semantic segmentation predictions. Comprehensive experimental results validate the effectiveness of our approach, demonstrating its ability to generalize across multiple unseen target domains without requiring auxiliary domains or additional modules. Our future work will be improving and stabilizing the covariance alignment method.
+
+Acknowledgement. This work was supported in part by the Basic Science Research Program through National Research Foundation of Korea (NRF) (Grants No. NRF2022R1F1A1073543), the MSIT(Ministry of Science and ICT), Korea, under the ICAN(ICT Challenge and Advanced Network of HRD) support program(RS-2022-00156385) supervised by the IITP(Institute for Information & Communications Technology Planning & Evaluation), and Innovative Human Resource Development for Local Intellectualization program through the Institute of Information & Communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT)(IITP-2024-00156287).
+
+# References
+
+[1] Florent Bartoccioni, Éloi Zablocki, Andrei Bursuc, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lara: Latents and rays for multi-camera bird's-eye-view semantic segmentation. In Conference on Robot Learning, pages 1663-1672. PMLR, 2023. 1  
+[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 3  
+[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 6, 7  
+[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3, 4  
+[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1  
+[6] Wonwooong Cho, Sungha Choi, David Keetae Park, Inkyu Shin, and Jaegul Choo. Image-to-image translation via group-wise deep whitening-and-coloring transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10639-10647, 2019. 3  
+[7] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11580-11590, 2021. 2, 3, 4, 6, 7  
+[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 6  
+[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 7  
+[10] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2  
+[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2  
+[12] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach
+
+to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3  
+[13] 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, 7  
+[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3  
+[15] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 1, 2  
+[16] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022. 2  
+[17] Lukas Hoyer, Dengxin Dai, Qin Wang, Yuhua Chen, and Luc Van Gool. Improving semi-supervised and domain-adaptive semantic segmentation with self-supervised depth estimation. International Journal of Computer Vision, pages 1-27, 2023. 1  
+[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1  
+[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 2  
+[20] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3061-3071, 2023. 6, 7  
+[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 3  
+[22] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 1, 2  
+[23] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 3  
+[24] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In Proceedings of the IEEE/CVF
+
+Conference on Computer Vision and Pattern Recognition, pages 9936-9946, 2022. 2, 3, 6, 7  
+[25] Guangrui Li, Guoliang Kang, Wu Liu, Yunchao Wei, and Yi Yang. Content-consistent matching for domain adaptive semantic segmentation. In European conference on computer vision, pages 440-456. Springer, 2020. 1  
+[26] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2021. 2  
+[27] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 2  
+[28] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European conference on computer vision (ECCV), pages 624-639, 2018. 2  
+[29] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6936-6945, 2019. 2  
+[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2  
+[31] Ping Luo. Learning deep architectures via generalized whitened neural networks. In International Conference on Machine Learning, pages 2238-2246. PMLR, 2017. 3  
+[32] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 6, 7  
+[33] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017. 6  
+[34] David Nilsson, Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Embodied visual active learning for semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2373–2383, 2021. 1  
+[35] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4501-4508. IEEE, 2021. 1  
+[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 4  
+[37] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3764-3773, 2020. 2
+
+[38] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 2, 3, 6, 7  
+[39] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1863-1871, 2019. 2, 3  
+[40] Taesung Park, Alexei A Efros, Richard Zhang, and JunYan Zhu. Contrastive learning for unpaired image-to-image translation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 319-345. Springer, 2020. 3  
+[41] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2, 3  
+[42] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2594-2605, 2022. 2, 4, 6, 7  
+[43] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 102-118. Springer, 2016. 1, 6  
+[44] 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. 1  
+[45] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 1, 6  
+[46] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10765-10775, 2021. 1  
+[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6, 7  
+[48] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 3  
+[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5, 8
+
+[50] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in neural information processing systems, 31, 2018. 2  
+[51] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 1  
+[52] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021. 3  
+[53] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In European Conference on Computer Vision, pages 603-620. Springer, 2022. 2, 6, 7  
+[54] 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. 1  
+[55] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2884-2892, 2022. 3, 6, 7  
+[56] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. arXiv preprint arXiv:2303.01906, 2023. 3, 6, 7  
+[57] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9036-9045, 2019. 2  
+[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 6  
+[59] Fei Yu, Mo Zhang, Hexin Dong, Sheng Hu, Bin Dong, and Li Zhang. Dast: Unsupervised domain adaptation in semantic segmentation based on discriminator attention and self-training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10754–10762, 2021. 2  
+[60] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2100-2110, 2019. 2  
+[61] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Do
+
+main generalization with mixstyle. In International Conference on Learning Representations, 2020. 2  
+[62] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 1, 2

2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0993954cb406f0074349f9ac5f0d37a6051081431de193a4bab76cf489098ed8
+size 578235

2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/43978830-73bd-4db5-8b8c-57888ce93c6e_content_list.json

@@ -0,0 +1,1665 @@
+[
+    {
+        "type": "text",
+        "text": "Style Injection in Diffusion: A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer",
+        "text_level": 1,
+        "bbox": [
+            176,
+            130,
+            794,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jiwoo Chung*, Sangeek Hyun*, Jae-Pil Heo† Sungkyunkwan University",
+        "bbox": [
+            305,
+            202,
+            661,
+            239
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{wldn0202, hsi1032, jaepilheo}@g.skku.edu",
+        "bbox": [
+            303,
+            241,
+            665,
+            255
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            291,
+            313,
+            306
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite the impressive generative capabilities of diffusion models, existing diffusion model-based style transfer methods require inference-stage optimization (e.g. fine-tuning or textual inversion of style) which is time-consuming, or fails to leverage the generative ability of large-scale diffusion models. To address these issues, we introduce a novel artistic style transfer method based on a pre-trained large-scale diffusion model without any optimization. Specifically, we manipulate the features of self-attention layers as the way the cross-attention mechanism works; in the generation process, substituting the key and value of content with those of style image. This approach provides several desirable characteristics for style transfer including 1) preservation of content by transferring similar styles into similar image patches and 2) transfer of style based on similarity of local texture (e.g. edge) between content and style images. Furthermore, we introduce query preservation and attention temperature scaling to mitigate the issue of disruption of original content, and initial latent Adaptive Instance Normalization (AdaIN) to deal with the disharmonious color (failure to transfer the colors of style). Our experimental results demonstrate that our proposed method surpasses state-of-the-art methods in both conventional and diffusion-based style transfer baselines. Codes are available at github.com/jiwoogit/StyleID.",
+        "bbox": [
+            75,
+            323,
+            473,
+            700
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            734,
+            209,
+            751
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent advances in Diffusion Models (DMs) have led to breakthroughs in various generative applications such as text-to-image synthesis [30, 33, 35] and image or video editing [2, 4, 6, 14, 19, 41, 48]. One of these efforts is also applied to the task of style transfer [10, 18, 45, 47, 53]; given style and content images, modifying the style of the content image to possess the given style.",
+        "bbox": [
+            75,
+            761,
+            468,
+            867
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/64e6849128fd581115042f13f5ea0b408d3614d988eeffcf75b7b26502729014.jpg",
+        "image_caption": [
+            "(a) General usage of self-attention in DM"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            295,
+            888,
+            359
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/8ede24d878fbeafce1263d42f2ba33274a2148939c7c478f43bbeff71c501e9c.jpg",
+        "image_caption": [
+            "(b) Our adaptation of self-attention for style transfer",
+            "Figure 1. Manipulation of self-attention features for style transfer. (a) General self-attention (SA) deploys the query, key, and value features from a single image in both the training and inference phases. (b) At inference phase, we suggest that manipulating features of self-attention of pre-trained large-scale DM is an effective way to transfer the styles; injection of key and value of styles into SA of contents is a proper way for transferring styles. As a result, style-injected content  $z_{t - 1}^{c}$  would maintain contents while modifying its style to resemble the target style."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            377,
+            888,
+            508
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "General approaches for diffusion model-based style transfer leverage the generative capability of pre-trained DM. Some of these works focus on explicit disentangling style and content for interpretable and controllable style transfer [45], or inversion of the style image into the textual latent space of a large-scale text-to-image DM [53]. However, these methods additionally require gradient-based optimization for fine-tuning and textual inversion [34] for each style image, which is time-consuming. Without this issue, DiffStyle [18] introduces training-free style transfer, but they are known to be hardly applicable to Latent Diffusion Model [33] which is widely adopted for training large-scale text-to-image DM such as Stable Diffusion [33], hindering the users from taking advantage of the prominent generative ability of large-scale models.",
+        "bbox": [
+            496,
+            674,
+            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_footnote",
+        "text": "* Equal contribution",
+        "bbox": [
+            101,
+            875,
+            212,
+            887
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "† Corresponding author",
+        "bbox": [
+            102,
+            887,
+            228,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "8795",
+        "bbox": [
+            482,
+            944,
+            514,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/a5ef20a21d0b05940f620e5cd1ca924f81d6dc810fb5c5f55b2724dc2eda0ae7.jpg",
+        "image_caption": [
+            "(b) query-key similarity between patches of similar texture",
+            "Figure 2. Desirable attributes of self-attention (SA) for style transfer. (a) Visualization of query by PCA shows that query features well-reflect similarities among patches. That is, style transfer employing SA can preserve the original content, as content patches with similarities tend to receive similar attention scores from a corresponding style image patch. (b) We visualize a similarity map between the blue box (edge) query of the content image, and key of the style image. Thanks to the features representation of large-scale DM encompassing texture and semantics, a query exhibits higher similarity to keys that share a similar style, such as edges."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            88,
+            468,
+            228
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we focus on extending the training-free style transfer to its application on large-scale pre-trained DM. We start from the observation of recent advances in image-to-image translation based on large-scale DM; they uncover the image editing capability of attention layers. Notably, Plug-and-play [41] shows that the residual block and the attention map of self-attention (SA) determine the spatial layout of generated images. Also, Prompt-to-Prompt [14] locally edits the image by replacing key and value of cross-attention (CA) obtained from text prompt, while keeping their original attention maps. That is, all these works suggest that 1) attention maps determine the spatial layout and 2) key and value of CA adjust the content to fill.",
+        "bbox": [
+            75,
+            400,
+            468,
+            609
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Inspired by the aforementioned methods, we newly argue that manipulating the SA layer is an effective way to transfer the styles (Fig. 1). Specifically, similar to CA, we substitute the key and value of SA and observe that the generated images are still visually plausible and naturally incorporate the elements of the substituted image into the original image. This observation motivates us to propose a style transfer technique based on self-attention, which combines the styles (textures) of a specific image with the content (semantics and spatial layout) of different images. Furthermore, we highlight that SA layer has desirable characteristics for style transfer. First, as shown in Fig. 2 (a), in SA-based style transfer, the content image patches (query) that share semantic similarities engage with a similar style (key), thereby maintaining the relationship among these content image patches after the transfer. Next, thanks to the powerful feature representation of large-scale DM [49], each patch of the query reveals higher similarity to keys which has similar texture and semantics. For instance, in Fig. 2 (b),",
+        "bbox": [
+            75,
+            613,
+            468,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "we can observe that the query feature of content within the blue box exhibits a high similarity to the key features of style with similar edge texture. This encourages the model to transfer style based on the similarity of local texture (e.g. edge) between content and style.",
+        "bbox": [
+            496,
+            90,
+            890,
+            167
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As a result, our method aims to transfer the textures of the style image to the content images by manipulating self-attention features of pre-trained large-scale DM without any optimization. To this end, we first propose an attention-based style injection method. The basic idea of it is substituting the content's key and value of SA with those of the style image, especially layers in the latter part of decoder which are relevant to the local textures. As mentioned in above paragraph, exchanged styles are well aligned with the content and texture of original image, exploiting the similarity-based attention mechanism. With the proposed style injection, we observe that the local texture patterns are successfully transferred, but there still are remaining problems such as disruption of original content and disharmonious colors. To handle these problems, we additionally propose the following techniques; query preservation, attention temperature scaling, and initial latent AdaIN. Query preservation makes the reverse diffusion process to retain the spatial structure of original content by preserving the query of the content image in the SA. Attention temperature scaling also aims to keep the structure of content by dealing with the blurred self-attention map introduced from the substitution of key. Lastly, initial latent AdaIN corrects inharmonious color problem, referring that the color distribution of style images is not properly transferred, by modulating the statistics of initial noise in the diffusion model.",
+        "bbox": [
+            496,
+            167,
+            890,
+            559
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our main contributions are summarized as follows:",
+        "bbox": [
+            517,
+            561,
+            859,
+            575
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose a style transfer method exploiting the large-scale pre-trained DM by simple manipulation of the features in self-attention; substituting key and value of content with those of styles without any requirements of optimization or supervision (e.g. text).",
+            "- We further improve the naive approach for style transfer to properly adapt the styles by proposing three components; query preservation, attention temperature scaling, and initial latent AdaIN.",
+            "- Extensive experiments on the style transfer dataset validate the proposed method significantly outperforms previous methods and achieves state-of-the-art performance."
+        ],
+        "bbox": [
+            496,
+            579,
+            890,
+            758
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            773,
+            640,
+            789
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Diffusion Model-based Neural Style Transfer",
+        "text_level": 1,
+        "bbox": [
+            500,
+            800,
+            879,
+            815
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Neural style transfer [11, 24-26, 29, 31, 43, 44, 52] is an example-guided image generation task that transfers the style of one image onto another while retaining the content of the original. In the realm of diffusion models, neural style transfer has evolved by leveraging the generative capability",
+        "bbox": [
+            496,
+            824,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "8796",
+        "bbox": [
+            480,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "of pre-trained diffusion models. For instance, InST [53] introduced a textual inversion-based approach, aiming to map a given style into corresponding textual embeddings. StyleDiffusion [45] aimed to disentangle style and content by introducing CLIP-based style disentanglement loss for fine-tuning DM for style transfer. Also, several approaches utilize the text input as a style condition or for determining the content to synthesize [10, 47].",
+        "bbox": [
+            75,
+            90,
+            468,
+            210
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Conversely, DiffStyle [18] proposed a training-free style transfer method that leverages  $h$ -space [23] and adjusts skip connections for effectively conveying style and content information, respectively. However, when DiffStyle is applied to Stable Diffusion [33, 42], their behavior is quite different from typical style-transfer methods; not only textures but also semantics such as spatial layout are also changed.",
+        "bbox": [
+            75,
+            214,
+            468,
+            321
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To address these limitations, we propose a novel algorithm that harmoniously merges style and content features within the self-attention layers of Stable Diffusion without any optimization process.",
+        "bbox": [
+            75,
+            323,
+            468,
+            385
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2. Attention-based Image Editing in DM",
+        "text_level": 1,
+        "bbox": [
+            76,
+            402,
+            403,
+            420
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Following the remarkable advances achieved by pre-trained text-to-image DMs [32, 42], there have been numerous image editing works [2, 6, 19, 37] utilizing these DMs. Notably, Prompt-to-Prompt [14] proposed text-based local image editing by manipulating the cross-attention map. Specifically, they observe that cross-attention largely contributes to modeling the relation between the spatial layout of the image to each word in the prompt. Hence, they substitute the original words and cross-attention map with desirable ones, obtaining edited images matched with text conditions. Subsequently, Plug-and-play [41] introduces text-guided image-to-image translation method. They found that the spatial features (i.e. feature from residual block) and self-attention map determine the spatial layout of the synthesized image. Thus, while generating a new image with the given text condition, they guide the diffusion model with features and attention map from the original image for preserving the original spatial layout. Recently, MasaCtrl [3] proposes mutual self-attention control for consistent image editing using text prompts. In detail, they retain the source image's key and value of the self-attention layers, while conditioning the model with desired text prompts.",
+        "bbox": [
+            75,
+            429,
+            468,
+            762
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Along with these works, we recognize the potential of attention maps in representing spatial information. However, different from the aforementioned methods concentrating on exploiting textual condition, we focus on conditioning by style and content images composed of two images from distinct styles. By combining the features in self-attention layers of both style and content images with precise adjustment of statistics in intermediate representations, we transfer the texture of the content image to the given style.",
+        "bbox": [
+            75,
+            765,
+            468,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Background",
+        "text_level": 1,
+        "bbox": [
+            500,
+            89,
+            627,
+            107
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Latent Diffusion Model (LDM) [33] is a type of diffusion model trained in the low dimensional latent space to focus on semantic bits of data and reduce computation costs. Given an image  $x \\in \\mathbb{R}^{H \\times W \\times 3}$ , the encoder  $\\mathcal{E}$  encodes  $x$  into the latent representation  $z \\in \\mathbb{R}^{h \\times w \\times c}$  and decoder reconstructs the image from the latent.",
+        "bbox": [
+            496,
+            116,
+            890,
+            207
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "With the pretrained encoder, they encode the entire images in the dataset and train a diffusion model on latent space  $z$ , by predicting noise  $\\epsilon$  from the noised version of latent  $z_{t}$  at time step  $t$ . The corresponding training objective is",
+        "bbox": [
+            496,
+            209,
+            890,
+            282
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {\\mathrm {L D M}} = \\mathbb {E} _ {z, \\epsilon , t} [ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, y) \\| _ {2} ^ {2} ], \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            578,
+            289,
+            890,
+            308
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\epsilon \\in \\mathcal{N}(0,1)$  is a noise,  $t$  is the number of time steps which uniformly sampled from  $\\{1,\\dots,T\\}$ ,  $y$  is a condition, and  $\\epsilon_{\\theta}$  is a neural network which predicts the noise added to the  $z$ .",
+        "bbox": [
+            496,
+            319,
+            890,
+            378
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In our work, we utilize Stable Diffusion (SD) [33] which is the only publicized large-scale pre-trained DM. In the case of SD,  $y$  is a text, and  $\\epsilon_{\\theta}$  is a U-Net architecture in which a block for each resolution comprises a residual block, self-attention block (SA), and cross-attention block (CA), sequentially. Among these modules, we focus on the SA block to transfer the styles, as discussed in Sec. 1. Given a feature  $\\phi$  after the residual block, the self-attention block performs as follows:",
+        "bbox": [
+            496,
+            381,
+            890,
+            517
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nQ = W _ {Q} (\\phi), K = W _ {K} (\\phi), V = W _ {V} (\\phi),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            532,
+            841,
+            550
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\phi_ {\\text {o u t}} = \\operatorname {A t t n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) \\cdot V, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            529,
+            551,
+            890,
+            585
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $d$  denotes the dimension of the projected query, and  $W_{(\\cdot)}$  is a projection layer. Note that, we don't use any text conditions, so the variable  $y$  is always an empty text prompt  $(\\ldots)$ .",
+        "bbox": [
+            500,
+            599,
+            890,
+            661
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4. Method",
+        "text_level": 1,
+        "bbox": [
+            500,
+            676,
+            589,
+            691
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we aim to solve artistic style transfer by leveraging the generative capability of a pre-trained large-scale text-to-image diffusion model. Briefly, artistic style transfer is the task of modifying the style of a given content image  $I^{c}$  to that of style image  $I^{s}$ . Then, the stylized image  $I^{cs}$  would maintain the semantic content of  $I^{c}$  while its style (such as texture) is transferred from  $I^{s}$ . For simplicity, we skip the explanations about the encoding and decoding process of the autoencoder in the LDM. Instead, we focus on elaborating the proposed method in the aspect of the diffusion process. Thus, in the following sections, we regard the content, style, and stylized images same as their encoded counterparts  $z_{0}^{c}, z_{0}^{s}$ , and  $z_{0}^{cs}$ .",
+        "bbox": [
+            496,
+            704,
+            890,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "8797",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/d557a0ed5ba8b09332a9a9cb85c7f7e3dad6ad5ddc3172013a934b5322b0b4d1.jpg",
+        "image_caption": [
+            "Figure 3. Overall framework. (Left) Illustration for the proposed style transfer method. We first invert content image  $z_0^c$  and style image  $z_0^s$  into the latent noise space as  $z_T^c$  and  $z_T^s$ , respectively. Then, we initialize the initial noise of stylized image  $z_T^{cs}$  from initial latent AdaIN (Sec. 4.3) which combines the content and style noise,  $z_T^c$  and  $z_T^s$ . While performing the reverse diffusion process with  $z_T^{cs}$ , we inject the information of content and style by attention-based style injection (Sec. 4.1) and attention temperature scaling (Sec. 4.2). (Right) Detailed explanation of style injection and initial noise AdaIN. Style injection is basically the manipulation of self-attention (SA) layer during the reverse diffusion process. Specifically, at time step  $t$ , we substitute the key  $(K_t^{cs})$  and value  $(V_t^{cs})$  in SA of stylized image with those of style features,  $K_t^s$  and  $V_t^s$ , from identical timestep  $t$ . At the same time, we preserve the content information by blending the query of content  $Q_t^c$  and query of stylized image  $Q_t^{cs}$ . Finally, we scale the magnitude of the attention map to deal with the magnitude decrease that the substitution of feature leads to. Initial latent AdaIN produces the initial noise  $z_T^{cs}$  by combining style noise  $z_T^s$  and content noise  $z_T^s$ . Specifically, we modify the channel statistics of  $z_T^c$  to resemble the statistics of  $z_T^s$  and regard it as  $z_T^{cs}$ . We observe this operation enables us to keep the spatial layout of content image while well-reflecting the color tones of a given style image."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            83,
+            92,
+            580,
+            361
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/25341082c1bb4fa13bb877698f1b3db935c09e0b80ac2ef541c5291fe7f10777.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            583,
+            92,
+            883,
+            359
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1. Attention-based Style Injection",
+        "text_level": 1,
+        "bbox": [
+            76,
+            532,
+            352,
+            547
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We start from the observation in previous image-to-image translation methods, especially Prompt-to-Prompt [14]. The key idea of their method is changing the text condition for cross-attention (CA) while keeping the attention map. Since the attention map affects the spatial layout of output, substituted text conditions determine what to draw in the generated image, and these conditions are actually key and value in CA. Inspired by them, we manipulate the features in self-attention layer as like cross-attention, regarding the features from style image  $I^{s}$  as the condition. Specifically, in the generation process, we substitute the key and value of content image with those of style for transferring the texture of style image into the content image.",
+        "bbox": [
+            75,
+            559,
+            468,
+            755
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To this end, we first obtain the latent for content and style images with DDIM inversion [39], and then collect the SA features of style image over the DDIM inversion process. Specifically, for pre-defined timesteps  $t = \\{0,\\dots,T\\}$ , style and content images  $z_0^c$  and  $z_0^s$  are inverted from image  $(t = 0)$  to gaussian noise  $(t = T)$ . During DDIM inversion, we also collect query features of content  $(Q_t^c)$  and key and value features of style  $(K_t^s,V_t^s)$  at every time steps.",
+        "bbox": [
+            75,
+            760,
+            468,
+            881
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "After that, we initialize stylized latent noise  $z_T^{cs}$  by copy-",
+        "bbox": [
+            96,
+            885,
+            468,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ing content latent noise  $z_{T}^{c}$ . Then, we transfer the target style to the stylized latent by injecting the key  $K_{t}^{s}$  and value  $V_{t}^{s}$  collected from the style into SA layer, instead of the original key  $K_{t}^{cs}$  and value  $V_{t}^{cs}$ , when performing the entire reverse process of stylized latent  $z_{t}^{cs}$ . However, only applying this substitution can lead to content disruption, since the content of stylized latent would be progressively changed as attended value changes. Hence, we propose query preservation to maintain original content. Simply, we blend query of stylized latent  $Q_{t}^{cs}$  and that of content  $Q_{t}^{c}$  for the entire reverse process. These style injection and query preservation processes at time step  $t$  are expressed as follows:",
+        "bbox": [
+            496,
+            532,
+            890,
+            714
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\tilde {Q} _ {t} ^ {c s} = \\gamma \\times Q _ {t} ^ {c} + (1 - \\gamma) \\times Q _ {t} ^ {c s}, \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            584,
+            728,
+            890,
+            746
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\phi_ {\\text {o u t}} ^ {\\mathrm {c s}} = \\operatorname {A t t n} \\left(\\tilde {Q} _ {t} ^ {c s}, K _ {t} ^ {s}, V _ {t} ^ {s}\\right), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            604,
+            765,
+            890,
+            782
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\gamma$  is degree of blending in range of [0, 1]. In addition, we apply these operations on the latter layers of decoder (7-  $12^{\\mathrm{th}}$  decoder layers in SD) relevant to local textures. We also highlight that the proposed method can adjust the degree of style transfer by changing query preservation ratio  $\\gamma$ . Specifically, higher  $\\gamma$  maintains more content, while lower  $\\gamma$  strengthens effects of style transfer.",
+        "bbox": [
+            496,
+            794,
+            890,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "8798",
+        "bbox": [
+            480,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/3dfeb95f4b4a72fd4f26c9fa31b3133f0dc69f21163d9f1bc6c7575f44d06a02.jpg",
+        "image_caption": [
+            "(a) Std. of attention map",
+            "(b) Ratio of std. of attention map"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            90,
+            294,
+            189
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/ac32451fdf99d903083809667e55399f7e6415ded3047660913ea5c999dfe958.jpg",
+        "image_caption": [
+            "Figure 4. Visualization of the standard deviation of attention map before softmax. (a) Attention-based style injection reduces the standard deviation of self-attention map. Original denotes SA maps from the generation process without style injection. We use both style and content images for generation. (b) We compute the ratio between attention maps w/ and w/o style injection. For the std of original image, we use averaged std. of content and style."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            294,
+            90,
+            464,
+            189
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2. Attention Temperature Scaling",
+        "text_level": 1,
+        "bbox": [
+            76,
+            320,
+            352,
+            335
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Attention map is computed by scaled dot-product between query and key features. During training, query and key features in the SA layer originate from an identical image. However, if we substitute the key features with those of style images, the magnitude of similarity would be overall lowered as style and content are highly likely to be irrelevant. Thus, the computed attention map can be blurred or smoothed, and it would further make output images unsharp, which is detrimental to capturing both content and style information.",
+        "bbox": [
+            75,
+            343,
+            468,
+            492
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To quantify this issue, we measure the standard deviation of attention map, while ablating the attention-based style injection. In detail, we calculate the attention map before applying softmax, which is scaled-dot product between query and key. As shown in Fig. 4 (a), we validate that this style injection tends to lower the standard deviation of the attention map over the entire timesteps. That is, attention maps after softmax with style injection would be overly smooth.",
+        "bbox": [
+            75,
+            494,
+            468,
+            614
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To rectify the attention map sharper, we introduce an attention temperature scaling parameter. In detail, we multiply the attention map before softmax by a constant temperature scaling parameter  $\\tau$  larger than 1. Thus, the attention map after softmax would be sharper than its original values. The modified attention process is represented as follows:",
+        "bbox": [
+            75,
+            614,
+            468,
+            705
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {A t t n} _ {\\tau} \\left(\\tilde {Q _ {t} ^ {c s}}, K _ {t} ^ {s}, V _ {t} ^ {s}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {\\tau \\tilde {Q _ {t} ^ {c s}} \\left(K _ {t} ^ {s}\\right) ^ {T}}{\\sqrt {d}}\\right) \\cdot V _ {t} ^ {s}, \\tau > 1. \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            76,
+            715,
+            467,
+            762
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use  $\\tau = 1.5$  as a default setting, which is the average ratio over entire timesteps. As reported in Fig. 4 (b), we confirm that it effectively calibrates the standard deviation of attention map similar to its original values.",
+        "bbox": [
+            75,
+            763,
+            468,
+            823
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3. Initial Latent AdaIN",
+        "text_level": 1,
+        "bbox": [
+            76,
+            832,
+            272,
+            845
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In artistic style transfer, the color tone generally takes up a significant portion of the style information. In this context, we observe that the style transfer only with attention-based",
+        "bbox": [
+            75,
+            854,
+            470,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/8cfdad2fb268773913895461a5819b09eacd13147ba70e421af6bead6f86e5ba.jpg",
+        "image_caption": [
+            "(a) Generated results only w/ style injection  $(\\mathrm{K}^s,V^s)$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            89,
+            746,
+            180
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/dde322ca136330611fdd146218e09cf6e5cd3e362c095d788c8b9605543f7f15.jpg",
+        "image_caption": [
+            "njection  $(\\mathrm{K}^s,V^s)$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            746,
+            90,
+            867,
+            180
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/7c4e3c4344e80896acccab14266518ab7f61d2f69ac574ece9e3eb52bc3f048a.jpg",
+        "image_caption": [
+            "(b) Generated results only w/ style injection  $(\\mathrm{Q}^s,\\mathrm{K}^s,V^s)$",
+            "Figure 5. Generated results only w/ style injection. (a) We observe that generated images only with attention-based style injection do not harmonize with the given style in the aspect of color tone. (b) To identify the effects of every feature in SA on color tones, we additionally include query in the style injection process. However, color tones still resemble those of content, concluding features in self-attention have less effect on the color tones."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            199,
+            746,
+            291
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/8c8d4174f83169458e3f8dc0035282fc426522048821c800fcd49fd4a8947b3f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            746,
+            199,
+            867,
+            290
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "style injection often fails in terms of capturing the color tone of the given style. As shown in Fig. 5 (a), textures and local patterns are successfully transferred to the content image while the color tone of the content image still remains. Furthermore, even with injecting the query, key, and value of styles, the resulting images still preserve the color tone of the content, as shown in Fig. 5 (b).",
+        "bbox": [
+            496,
+            422,
+            893,
+            527
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As substituting the self-attention features has less effects to color tone, we analyze the other vital part of DM; initial latent noise. One of the recent discoveries in DM is that the DM struggles to synthesize purely white or black images [13]. Instead, they tend to generate images of median color as the initial noise is sampled from zero mean and unit variance. Thus, we hypothesize the statistics of initial noise largely affect the colors and brightness of generated images.",
+        "bbox": [
+            496,
+            527,
+            892,
+            648
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Based on this assumption, we attempt to use the initial latent of style  $z_{T}^{s}$  for the style transfer process. However, if we simply start to generate the image from style latent  $z_{T}^{s}$ , the structural information of synthesized results also follows the style image and loses the structure of the content. To harness valuable information in both initial latents, we consider that the tone information is intricately connected with the channel statistics of the initial latent, following the principle underlying Style Loss [11] and AdaIN [17]. Thus, we employ AdaIN to modulate the initial latent for effective tone information transfer, represented as:",
+        "bbox": [
+            496,
+            648,
+            893,
+            814
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nz _ {T} ^ {c s} = \\sigma \\left(z _ {T} ^ {s}\\right) \\left(\\frac {z _ {T} ^ {c} - \\mu \\left(z _ {T} ^ {c}\\right)}{\\sigma \\left(z _ {T} ^ {c}\\right)}\\right) + \\mu \\left(z _ {T} ^ {s}\\right), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            563,
+            825,
+            890,
+            859
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mu (\\cdot),\\sigma (\\cdot)$  denote channel-wise mean and standard deviation, respectively. Based on this, the initial latent  $z_{T}^{cs}$",
+        "bbox": [
+            498,
+            869,
+            892,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "8799",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/338e8deaf2ebc4d404dadea79f5e31bf92496db89fb348e14e2423eaae02ebae.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Metric</td><td>Ours</td><td>AesPA-Net</td><td>CAST</td><td>StyTR2</td><td>EFDM</td><td>MAST</td><td>AdaAttn</td><td>ArtFlow</td><td>AdaConv</td><td>AdaIN</td><td>DiffuseIT</td><td>InST</td><td>DiffStyle</td></tr><tr><td>ArtFID ↓</td><td>28.801</td><td>31.420</td><td>34.685</td><td>30.720</td><td>34.605</td><td>31.282</td><td>30.350</td><td>34.630</td><td>31.856</td><td>30.933</td><td>40.721</td><td>40.633</td><td>41.464</td></tr><tr><td>FID ↓</td><td>18.131</td><td>19.760</td><td>20.395</td><td>18.890</td><td>20.062</td><td>18.199</td><td>18.658</td><td>21.252</td><td>19.022</td><td>18.242</td><td>23.065</td><td>21.571</td><td>20.903</td></tr><tr><td>LPIPS ↓</td><td>0.5055</td><td>0.5135</td><td>0.6212</td><td>0.5445</td><td>0.6430</td><td>0.6293</td><td>0.5439</td><td>0.5562</td><td>0.5910</td><td>0.6076</td><td>0.6921</td><td>0.8002</td><td>0.8931</td></tr><tr><td>CFSD ↓</td><td>0.2281</td><td>0.2464</td><td>0.2918</td><td>0.3011</td><td>0.3346</td><td>0.3043</td><td>0.2862</td><td>0.2920</td><td>0.3600</td><td>0.3155</td><td>0.3428</td><td>0.6759</td><td>0.2819</td></tr></table>",
+        "bbox": [
+            78,
+            89,
+            888,
+            171
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Quantitative comparison with conventional  $\\left( {{3}^{\\text{rd }} - {11}^{\\text{th }}\\text{columns}}\\right)$  and diffusion model baselines  $\\left( {{12}^{\\text{th }} - {14}^{\\text{th }}\\text{columns}}\\right)$",
+        "bbox": [
+            124,
+            174,
+            841,
+            188
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "preserves content information from  $z_T^c$  while aligning the channel-wise mean and standard deviation with  $z_T^s$ .",
+        "bbox": [
+            75,
+            200,
+            468,
+            233
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            244,
+            209,
+            262
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1. Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            76,
+            270,
+            282,
+            286
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We conduct all experiments in Stable Diffusion 1.4 pretrained on LAION dataset [36] and adopt DDIM sampling [39] with a total 50 timesteps ( $t = \\{1, \\dots, 50\\}$ ). For default settings for hyperparameters, we use  $\\gamma = 0.75$  and  $\\tau = 1.5$ , if they are not mentioned separately.",
+        "bbox": [
+            75,
+            292,
+            470,
+            369
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2. Evaluation Protocol",
+        "text_level": 1,
+        "bbox": [
+            76,
+            378,
+            266,
+            393
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Conventional style transfer methods typically utilize Style Loss [11] as both training objective and evaluation metric, so their results tend to overfit the Style Loss. Thus, for a fair comparison, we employ a recently proposed metric, ArtFID [46] which evaluates overall style transfer performances with consideration of both content and style preservation and also is known as strongly coinciding with human judgment. Specifically, ArtFID is computed as  $(\\text{ArtFID} = (1 + \\text{LPIPS}) \\cdot (1 + \\text{FID}))$ . LPIPS [50] measures content fidelity between the stylized image and the corresponding content image, and FID [15] assesses the style fidelity between the stylized image and the corresponding style image.",
+        "bbox": [
+            75,
+            402,
+            468,
+            585
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Dataset. Our evaluations employ content images from MS-COCO [27] dataset and style images from WikiArt [40] dataset. All input images are center-cropped to  $512 \\times 512$  resolution. Also, for quantitative comparison, we randomly selected 20 content and 40 style images from each dataset, yielding 800 stylized images as StyTR² [9] has done.",
+        "bbox": [
+            75,
+            590,
+            468,
+            681
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Content Feature Structural Distance (CFSD). In the style transfer evaluation, the assessment of content fidelity often relies on the LPIPS distance. However, since LPIPS utilizes the feature space of AlexNet [20] pre-trained for classification task on ImageNet [7], which is known as texture-biased [12]. Thus, the style information of the images can affect the LPIPS score. To mitigate this style influence, we additionally introduce Content Feature Structural Distance (CFSD) which is a distance measure that only considers the spatial correlation between image patches.",
+        "bbox": [
+            75,
+            688,
+            468,
+            839
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In detail, we first define the correlation map between image patch features as follows. For a given image  $I$ , we obtain feature maps  $F \\in \\mathbb{R}^{hw \\times c}$ , which is the output feature of conv3 in VGG19 [38]. Then, we calculate the patch sim",
+        "bbox": [
+            75,
+            839,
+            468,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The similarity map  $M = F \\times F^T$ ,  $M \\in \\mathbb{R}^{hw \\times hw}$ , which is a similarity map between every pair of features in  $F$ . After that, for computing the distance between two patch similarity maps, we model the similarity between a single patch and the others as a probability distribution by applying softmax operation. Finally, the correlation map is represented as  $S = [\\text{softmax}(M_i)]_{i=1}^{hw}$ ,  $S \\in \\mathbb{R}^{hw \\times hw}$ , where  $M_i \\in \\mathbb{R}^{1 \\times hw}$  is a similarity map between  $i^{\\text{th}}$  patch and the other patches.",
+        "bbox": [
+            496,
+            200,
+            890,
+            321
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Then, CFSD is defined as KL-divergence between two correlation maps. In our case, we compute CFSD between the correlation map of the content  $(S^c)$  and stylized images  $(S^{cs})$  as follows:",
+        "bbox": [
+            496,
+            321,
+            890,
+            382
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathrm {C F S D} = \\frac {1}{h w} \\sum_ {i = 1} ^ {h w} D _ {\\mathrm {K L}} \\left(S _ {i} ^ {c} | | S _ {i} ^ {c s}\\right), \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            584,
+            393,
+            890,
+            435
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.3.Quantitative Comparison",
+        "text_level": 1,
+        "bbox": [
+            498,
+            454,
+            732,
+            470
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our proposed method through comparison with twelve state-of-the-art methods, including nine conventional style transfer methods (AesPA-Net [16], CAST [52], StyTR $^2$  [9], EFDM [51], MAST [8], AdaAttN [28], ArtFlow [1], AdaConv [5], AdaIN [17]) and three diffusion-based style transfer methods (DiffuseIT [22], InST [53], DiffStyle [18]), which have a style image as input. We employ the publicly available implementations of all baselines, using their recommended configurations.",
+        "bbox": [
+            496,
+            478,
+            890,
+            614
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Comparison with Conventional Style Transfer. As shown in Tab. 1, our method largely surpasses the conventional style transfer methods in terms of ArtFID, which is known as coinciding the human preference. In addition, the proposed method records the lowest FID, which denotes that stylized images highly resemble the target styles. For content fidelity metrics, ours shows superior scores in both CFSD and LPIPS. We point out that ours achieves much lower CFSD compared to other methods, which is the metric to only consider the spatial correlation.",
+        "bbox": [
+            496,
+            621,
+            890,
+            771
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In addition, we also emphasize that the proposed method can arbitrarily adjust the degree of style transfer by changing the  $\\gamma$ , and the proposed method significantly surpasses all the other methods in terms of FID (style), when we match the value of LPIPS (content) (Fig. 10).",
+        "bbox": [
+            496,
+            772,
+            890,
+            849
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Comparison with Diffusion-based Style Transfer. Our method demonstrates the best performance in terms of LPIPS, FID, and their combination (ArtFID) with a large",
+        "bbox": [
+            496,
+            854,
+            893,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "8800",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Content Style Ours AesPA-Net CAST StyTR2 EFDM MAST AdaAttN AdaIN DiffuseIT InST DiffStyle",
+        "bbox": [
+            83,
+            87,
+            885,
+            101
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/860ff3210846250b16d879d171e60f15c4647f18d2abd94bffd6a5026d2fa2ad.jpg",
+        "image_caption": [
+            "Figure 6. Qualitative comparison with conventional  $(4^{\\mathrm{th}} - 10^{\\mathrm{th}}$  columns) and diffusion model baselines  $(11^{\\mathrm{th}} - 13^{\\mathrm{th}}$  columns)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            101,
+            890,
+            294
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2838e8ad98d425397cce2a5d39e7ef935a30c31d0c61a70bce8bd7758a67ed21.jpg",
+        "image_caption": [
+            "Figure 7. Qualitative comparison with best ArtFID (AdaAttN) and most recently proposed baselines (AesPA-Net) with additional zooming details"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            94,
+            318,
+            450,
+            449
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/a075caf6dea678e94b4921075cb4d81fd5903e126d48e55406b153bbf8e7277b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Metric</td><td>DiffuseIT</td><td>InST</td><td>DiffStyle</td><td>Ours</td></tr><tr><td>Time (sec)</td><td>792.8</td><td>816.9</td><td>355.9</td><td>12.4</td></tr></table>",
+        "bbox": [
+            133,
+            498,
+            410,
+            537
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Comparison of inference time of diffusion-based methods for style transferring a given style and content pair",
+        "bbox": [
+            75,
+            542,
+            467,
+            570
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "margin, as shown in Tab. 1. One significant factor for diffusion models is their running times, since they require several steps to synthesize a single image and it requires inevitable time cost. Hence, we measure the inference time for a pair of content and style images on a single Titan RTX GPU, as shown in Tab. 2. Our method requires a total of 12.4 seconds, with 8.2 seconds for DDIM inversions and 4.2 seconds for sampling costs. As reported, we validate the proposed method significantly faster than other methods, even exploiting large-scale DM. This faster speed comes from the fact that the proposed methods can use the much smaller steps of DDIM inversion, because we additionally utilize the features collected during inversion steps, largely reducing the necessity for perfect inversion of content and style.",
+        "bbox": [
+            75,
+            579,
+            467,
+            790
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.4. Qualitative Comparison",
+        "text_level": 1,
+        "bbox": [
+            76,
+            801,
+            297,
+            816
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Comparison with Conventional Style Transfer. As shown in Fig. 6, we observe that our method tends to highly preserve the structural information of the content image, while also transferring the style well. For instance, as shown in the third row, ours retains the structure of the",
+        "bbox": [
+            75,
+            825,
+            467,
+            898
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "bridge, but the baselines struggle to preserve structure or transfer the style. We also provide the qualitative comparison with zooming details in Fig. 7 and Supplementary.",
+        "bbox": [
+            498,
+            320,
+            890,
+            364
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Comparison with Diffusion-based Style Transfer. We also compare our method with recent diffusion-based style transfer baselines [18, 23, 53]. As shown in Fig. 6, we observe the proposed technique transfers the style to the content well. On the other hand, baselines often lose the structure of content or fail to transfer the style, when an arbitrary content style pair is given. For instance, DiffuseIT and DiffStyle suffer from generating shape and visually plausible images or drop the original content. Differently, InST synthesizes the realistic images, while struggling to transfer style (1 $^{\\text{st}}$  row) or change content of image (2 $^{\\text{nd}}$ , 3 $^{\\text{rd}}$  rows).",
+        "bbox": [
+            498,
+            372,
+            890,
+            539
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.5. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            500,
+            551,
+            651,
+            568
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To validate the effectiveness of the proposed components, we conduct ablation studies in both quantitative and qualitative ways. As shown in Fig. 8 and Tab. 3, style injection is significant for guiding the style and content of given images (Config. B). Besides, initial latent AdaIN has a large portion of transferring the color tone of style (Config. D). Attention temperature scaling is in charge of enhancement of quality in synthesized results such as sharpening details and resolving blurriness. For instance, this scaling jointly reduces the FID and LPIPS (Config.  $\\mathrm{A}^*$  vs. C in Tab. 3). For more detailed analysis, we provide quantitative metrics with the style-content trade-off, while changing the attention scaling parameter  $\\tau$  in Fig. 10 (b). As reported, attention scaling effectively reduces both FID and LPIPS, proving its effects on the preservation of content and capability of style transfer ( $\\tau = 1.0$  vs.  $\\tau = 1.5$ ).",
+        "bbox": [
+            496,
+            575,
+            890,
+            816
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.6. Additional Analysis",
+        "text_level": 1,
+        "bbox": [
+            500,
+            830,
+            687,
+            845
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Content-Style Trade-Off. Our proposed method offers flexible control of the trade-off relation between content and style fidelity by adjusting the parameter  $\\gamma$ , as discussed in",
+        "bbox": [
+            498,
+            854,
+            890,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "8801",
+        "bbox": [
+            482,
+            944,
+            513,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/5b1b936a45b01d2bdd11131aae08a19b15765285713166401ef2ee2900e905a9.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Configuration</td><td>ArtFID</td><td>FID</td><td>LPIPS</td></tr><tr><td>A</td><td>Ours (γ = 0.75, default)</td><td>28.80</td><td>18.13</td><td>0.505</td></tr><tr><td>A*</td><td>Ours (γ = 0.6)</td><td>27.97</td><td>17.21</td><td>0.535</td></tr><tr><td>B</td><td>- Style Inject. (Sec. 4.1)</td><td>43.72</td><td>27.13</td><td>0.554</td></tr><tr><td>C</td><td>- Attn. Scaling (Sec. 4.2)</td><td>29.02</td><td>17.81</td><td>0.542</td></tr><tr><td>D</td><td>- Init. AdaIN (Sec. 4.3)</td><td>29.26</td><td>20.05</td><td>0.390</td></tr></table>",
+        "bbox": [
+            78,
+            88,
+            475,
+            195
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Ablation study on proposed components",
+        "bbox": [
+            125,
+            200,
+            419,
+            214
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/b89ab7b4bcfdd501d4f0c4ece99c1f74adbac5ab3e9aafff8a18fd52f34f13c6.jpg",
+        "image_caption": [
+            "Figure 8. Qualitative comparison with ablation studies"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            84,
+            222,
+            459,
+            363
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/4cf5a63c61f5638d6fffff5dc9acbbc0b67f136c825e93b8776c146c9759bf5c.jpg",
+        "image_caption": [
+            "Figure 9. Visualization of effects of query preservation ratio  $\\gamma$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            76,
+            386,
+            460,
+            549
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Sec. 4.1. In detail, we compute the FID and LPIPS while varying  $\\gamma$  within the range [0.3, 1] with a step size of 0.1. As shown in Fig. 10 (a), our method surpasses baseline methods across all ranges of content and style fidelity. This result implies the proposed method significantly outperforms the other methods, when we match style or content metric to the compared model by adjusting the  $\\gamma$  of ours. Note that, dotted lines refer to our model reported in Tab. 1.",
+        "bbox": [
+            75,
+            585,
+            468,
+            707
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We also visualize the effects of the style-content trade-off by synthesizing images by adjusting  $\\gamma$ . As shown in Fig. 9, the lower  $\\gamma$  highly reflects the style while losing the content of the given image, and vice versa. This characteristic of the proposed method suggests that the users can adjust the degree of style by following their personal preferences.",
+        "bbox": [
+            75,
+            709,
+            468,
+            801
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Study on the value of  $\\tau$ . We observe that the gradual increase of  $\\tau$  enhances the performance of style transfer, although its effects on enhancement become smaller as  $\\tau$  goes larger, as shown in Fig. 10 (b). This result implies that the attention temperature scaling effectively works with a simple modification of the magnitude of the attention map.",
+        "bbox": [
+            75,
+            809,
+            470,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/0574ed7e223feb7c0b3309243ecdcabf39ec3c7cba7af2ebe674151a020c830e.jpg",
+        "image_caption": [
+            "Figure 10. Style-content trade-offs"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            491,
+            92,
+            736,
+            244
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/d13886104303c34f63fe8e4d7a7cc55500466e7e10422acf2cfaad9f948dd77d.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            738,
+            92,
+            883,
+            244
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6f76cbcee65f8d474e662ff237a8efaf946791fe47816741a406344a17535b5e.jpg",
+        "image_caption": [
+            "Figure 11. Comparison with text-guided style transfer methods"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            516,
+            268,
+            879,
+            359
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Comparison with text-guided style transfer. We additionally compare the proposed method with the style transfer methods [21, 41] which are conditioned on the textual inputs. As text-guided methods tend to modify the style largely, we use  $\\gamma = 0.3$  for this experiment. Since the text condition hardly contains all the information in the style image such as texture and color tones, the transferred results less resemble the target style, as shown in Fig. 11. Differently, we validate that the proposed method successfully transfers the style with high fidelity.",
+        "bbox": [
+            496,
+            385,
+            890,
+            537
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            545,
+            617,
+            560
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our work addresses the challenges associated with diffusion model-based style transfer methods, which often require time-consuming optimization steps or struggle to leverage the generative potential of large-scale diffusion models. To this end, we propose the method of adapting the pre-trained large-scale diffusion model on style transfer in a training-free way. Our method focuses on manipulating the features of self-attention layers, akin to the cross-attention mechanism, by substituting the key and value during the content generation process with those of the style. Furthermore, we propose the query preservation and attention temperature scaling to mitigate the issue of disruption of content, and initial latent AdaIN to handle the disharmonious color (failure to transfer the colors of style). Experimental results show the superiority of our proposed method over state-of-the-art techniques in previous baselines.",
+        "bbox": [
+            496,
+            571,
+            890,
+            813
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            821,
+            658,
+            838
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This work was supported in part by MSIT/IITP (No. 2022-0-00680, 2019-0-00421, 2020-0-01821, 2021-0-02068), and MSIT&KNPA/KIPoT (Police Lab 2.0, No. 210121M06).",
+        "bbox": [
+            496,
+            839,
+            890,
+            898
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8802",
+        "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] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 862-871, 2021. 6",
+            "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 1, 3",
+            "[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560-22570, 2023. 3",
+            "[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 1",
+            "[5] Prashanth Chandran, Gaspard Zoss, Paulo Gotardo, Markus Gross, and Derek Bradley. Adaptive convolutions for structure-aware style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7972-7981, 2021. 6",
+            "[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR 2023 (Eleventh International Conference on Learning Representations), 2023. 1, 3",
+            "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6",
+            "[8] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM international conference on multimedia, pages 2719-2727, 2020. 6",
+            "[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 6",
+            "[10] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Susstrunk, and Radhakrishna Achanta. Diffusion in style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2251-2261, 2023. 1, 3",
+            "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 5, 6",
+            "[12] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 6"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[13] Nocholas Guttenberg. Diffusion with offset noise. https://www.crosslabs.org/blog/diffusion-with-offset-noise, 2023.5",
+            "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 4",
+            "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6",
+            "[16] Kibeom Hong, Seogkyu Jeon, Junsoo Lee, Namhyuk Ahn, Kunhee Kim, Pilhyeon Lee, Daesik Kim, Youngjung Uh, and Hyeran Byun. Aespa-net: Aesthetic pattern-aware style transfer networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22758-22767, 2023. 6",
+            "[17] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5, 6",
+            "[18] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-free style transfer emerges from h-space in diffusion models. arXiv preprint arXiv:2303.15403, 2023. 1, 3, 6, 7",
+            "[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Conference on Computer Vision and Pattern Recognition 2023, 2023. 1, 3",
+            "[20] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 6",
+            "[21] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 8",
+            "[22] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264, 2022. 6",
+            "[23] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 3, 7",
+            "[24] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 2",
+            "[25] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017.",
+            "[26] Yijun Li, Ming-Yu Liu, Xueting Li, Ming-Hsuan Yang, and Jan Kautz. A closed-form solution to photorealistic image"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            902
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "8803",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "stylization. In Proceedings of the European conference on computer vision (ECCV), pages 453-468, 2018. 2",
+            "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6",
+            "[28] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 6",
+            "[29] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5952-5961, 2019. 2",
+            "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1",
+            "[31] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880–5888, 2019. 2",
+            "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. 3",
+            "[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. 1, 3",
+            "[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1",
+            "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1",
+            "[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6",
+            "[37] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 3"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6",
+            "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 4, 6",
+            "[40] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 6",
+            "[41] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2, 3, 8",
+            "[42] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022.3",
+            "[43] Huan Wang, Yijun Li, Yuehai Wang, Haoji Hu, and Ming-Hsuan Yang. Collaborative distillation for ultra-resolution universal style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1860-1869, 2020. 2",
+            "[44] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7789-7798, 2020. 2",
+            "[45] Zhizhong Wang, Lei Zhao, and Wei Xing. *Style diffusions: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 1, 3",
+            "[46] Matthias Wright and Björn Ommer. Artfid: Quantitative evaluation of neural style transfer. In DAGM German Conference on Pattern Recognition, pages 560-576. Springer, 2022. 6",
+            "[47] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. arXiv preprint arXiv:2303.08622, 2023. 1, 3",
+            "[48] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 1",
+            "[49] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. arXiv preprint arXiv:2305.15347, 2023. 2",
+            "[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6",
+            "[51] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            901
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "8804",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8035-8045, 2022. 6",
+            "[52] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 2, 6",
+            "[53] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10146-10156, 2023. 1, 3, 6, 7"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            271
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "8805",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/43978830-73bd-4db5-8b8c-57888ce93c6e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d41ef29d6d3fd5ebfa0e6f37a8297ed18bec92b91b1c7a7846ed5433fb5681da
+size 7797603

2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/full.md

@@ -0,0 +1,326 @@
+# Style Injection in Diffusion: A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer
+
+Jiwoo Chung*, Sangeek Hyun*, Jae-Pil Heo† Sungkyunkwan University
+
+{wldn0202, hsi1032, jaepilheo}@g.skku.edu
+
+# Abstract
+
+Despite the impressive generative capabilities of diffusion models, existing diffusion model-based style transfer methods require inference-stage optimization (e.g. fine-tuning or textual inversion of style) which is time-consuming, or fails to leverage the generative ability of large-scale diffusion models. To address these issues, we introduce a novel artistic style transfer method based on a pre-trained large-scale diffusion model without any optimization. Specifically, we manipulate the features of self-attention layers as the way the cross-attention mechanism works; in the generation process, substituting the key and value of content with those of style image. This approach provides several desirable characteristics for style transfer including 1) preservation of content by transferring similar styles into similar image patches and 2) transfer of style based on similarity of local texture (e.g. edge) between content and style images. Furthermore, we introduce query preservation and attention temperature scaling to mitigate the issue of disruption of original content, and initial latent Adaptive Instance Normalization (AdaIN) to deal with the disharmonious color (failure to transfer the colors of style). Our experimental results demonstrate that our proposed method surpasses state-of-the-art methods in both conventional and diffusion-based style transfer baselines. Codes are available at github.com/jiwoogit/StyleID.
+
+# 1. Introduction
+
+Recent advances in Diffusion Models (DMs) have led to breakthroughs in various generative applications such as text-to-image synthesis [30, 33, 35] and image or video editing [2, 4, 6, 14, 19, 41, 48]. One of these efforts is also applied to the task of style transfer [10, 18, 45, 47, 53]; given style and content images, modifying the style of the content image to possess the given style.
+
+![](images/64e6849128fd581115042f13f5ea0b408d3614d988eeffcf75b7b26502729014.jpg)  
+(a) General usage of self-attention in DM
+
+![](images/8ede24d878fbeafce1263d42f2ba33274a2148939c7c478f43bbeff71c501e9c.jpg)  
+(b) Our adaptation of self-attention for style transfer  
+Figure 1. Manipulation of self-attention features for style transfer. (a) General self-attention (SA) deploys the query, key, and value features from a single image in both the training and inference phases. (b) At inference phase, we suggest that manipulating features of self-attention of pre-trained large-scale DM is an effective way to transfer the styles; injection of key and value of styles into SA of contents is a proper way for transferring styles. As a result, style-injected content  $z_{t - 1}^{c}$  would maintain contents while modifying its style to resemble the target style.
+
+General approaches for diffusion model-based style transfer leverage the generative capability of pre-trained DM. Some of these works focus on explicit disentangling style and content for interpretable and controllable style transfer [45], or inversion of the style image into the textual latent space of a large-scale text-to-image DM [53]. However, these methods additionally require gradient-based optimization for fine-tuning and textual inversion [34] for each style image, which is time-consuming. Without this issue, DiffStyle [18] introduces training-free style transfer, but they are known to be hardly applicable to Latent Diffusion Model [33] which is widely adopted for training large-scale text-to-image DM such as Stable Diffusion [33], hindering the users from taking advantage of the prominent generative ability of large-scale models.
+
+![](images/a5ef20a21d0b05940f620e5cd1ca924f81d6dc810fb5c5f55b2724dc2eda0ae7.jpg)  
+(b) query-key similarity between patches of similar texture  
+Figure 2. Desirable attributes of self-attention (SA) for style transfer. (a) Visualization of query by PCA shows that query features well-reflect similarities among patches. That is, style transfer employing SA can preserve the original content, as content patches with similarities tend to receive similar attention scores from a corresponding style image patch. (b) We visualize a similarity map between the blue box (edge) query of the content image, and key of the style image. Thanks to the features representation of large-scale DM encompassing texture and semantics, a query exhibits higher similarity to keys that share a similar style, such as edges.
+
+In this paper, we focus on extending the training-free style transfer to its application on large-scale pre-trained DM. We start from the observation of recent advances in image-to-image translation based on large-scale DM; they uncover the image editing capability of attention layers. Notably, Plug-and-play [41] shows that the residual block and the attention map of self-attention (SA) determine the spatial layout of generated images. Also, Prompt-to-Prompt [14] locally edits the image by replacing key and value of cross-attention (CA) obtained from text prompt, while keeping their original attention maps. That is, all these works suggest that 1) attention maps determine the spatial layout and 2) key and value of CA adjust the content to fill.
+
+Inspired by the aforementioned methods, we newly argue that manipulating the SA layer is an effective way to transfer the styles (Fig. 1). Specifically, similar to CA, we substitute the key and value of SA and observe that the generated images are still visually plausible and naturally incorporate the elements of the substituted image into the original image. This observation motivates us to propose a style transfer technique based on self-attention, which combines the styles (textures) of a specific image with the content (semantics and spatial layout) of different images. Furthermore, we highlight that SA layer has desirable characteristics for style transfer. First, as shown in Fig. 2 (a), in SA-based style transfer, the content image patches (query) that share semantic similarities engage with a similar style (key), thereby maintaining the relationship among these content image patches after the transfer. Next, thanks to the powerful feature representation of large-scale DM [49], each patch of the query reveals higher similarity to keys which has similar texture and semantics. For instance, in Fig. 2 (b),
+
+we can observe that the query feature of content within the blue box exhibits a high similarity to the key features of style with similar edge texture. This encourages the model to transfer style based on the similarity of local texture (e.g. edge) between content and style.
+
+As a result, our method aims to transfer the textures of the style image to the content images by manipulating self-attention features of pre-trained large-scale DM without any optimization. To this end, we first propose an attention-based style injection method. The basic idea of it is substituting the content's key and value of SA with those of the style image, especially layers in the latter part of decoder which are relevant to the local textures. As mentioned in above paragraph, exchanged styles are well aligned with the content and texture of original image, exploiting the similarity-based attention mechanism. With the proposed style injection, we observe that the local texture patterns are successfully transferred, but there still are remaining problems such as disruption of original content and disharmonious colors. To handle these problems, we additionally propose the following techniques; query preservation, attention temperature scaling, and initial latent AdaIN. Query preservation makes the reverse diffusion process to retain the spatial structure of original content by preserving the query of the content image in the SA. Attention temperature scaling also aims to keep the structure of content by dealing with the blurred self-attention map introduced from the substitution of key. Lastly, initial latent AdaIN corrects inharmonious color problem, referring that the color distribution of style images is not properly transferred, by modulating the statistics of initial noise in the diffusion model.
+
+Our main contributions are summarized as follows:
+
+- We propose a style transfer method exploiting the large-scale pre-trained DM by simple manipulation of the features in self-attention; substituting key and value of content with those of styles without any requirements of optimization or supervision (e.g. text).  
+- We further improve the naive approach for style transfer to properly adapt the styles by proposing three components; query preservation, attention temperature scaling, and initial latent AdaIN.  
+- Extensive experiments on the style transfer dataset validate the proposed method significantly outperforms previous methods and achieves state-of-the-art performance.
+
+# 2. Related Work
+
+# 2.1. Diffusion Model-based Neural Style Transfer
+
+Neural style transfer [11, 24-26, 29, 31, 43, 44, 52] is an example-guided image generation task that transfers the style of one image onto another while retaining the content of the original. In the realm of diffusion models, neural style transfer has evolved by leveraging the generative capability
+
+of pre-trained diffusion models. For instance, InST [53] introduced a textual inversion-based approach, aiming to map a given style into corresponding textual embeddings. StyleDiffusion [45] aimed to disentangle style and content by introducing CLIP-based style disentanglement loss for fine-tuning DM for style transfer. Also, several approaches utilize the text input as a style condition or for determining the content to synthesize [10, 47].
+
+Conversely, DiffStyle [18] proposed a training-free style transfer method that leverages  $h$ -space [23] and adjusts skip connections for effectively conveying style and content information, respectively. However, when DiffStyle is applied to Stable Diffusion [33, 42], their behavior is quite different from typical style-transfer methods; not only textures but also semantics such as spatial layout are also changed.
+
+To address these limitations, we propose a novel algorithm that harmoniously merges style and content features within the self-attention layers of Stable Diffusion without any optimization process.
+
+# 2.2. Attention-based Image Editing in DM
+
+Following the remarkable advances achieved by pre-trained text-to-image DMs [32, 42], there have been numerous image editing works [2, 6, 19, 37] utilizing these DMs. Notably, Prompt-to-Prompt [14] proposed text-based local image editing by manipulating the cross-attention map. Specifically, they observe that cross-attention largely contributes to modeling the relation between the spatial layout of the image to each word in the prompt. Hence, they substitute the original words and cross-attention map with desirable ones, obtaining edited images matched with text conditions. Subsequently, Plug-and-play [41] introduces text-guided image-to-image translation method. They found that the spatial features (i.e. feature from residual block) and self-attention map determine the spatial layout of the synthesized image. Thus, while generating a new image with the given text condition, they guide the diffusion model with features and attention map from the original image for preserving the original spatial layout. Recently, MasaCtrl [3] proposes mutual self-attention control for consistent image editing using text prompts. In detail, they retain the source image's key and value of the self-attention layers, while conditioning the model with desired text prompts.
+
+Along with these works, we recognize the potential of attention maps in representing spatial information. However, different from the aforementioned methods concentrating on exploiting textual condition, we focus on conditioning by style and content images composed of two images from distinct styles. By combining the features in self-attention layers of both style and content images with precise adjustment of statistics in intermediate representations, we transfer the texture of the content image to the given style.
+
+# 3. Background
+
+Latent Diffusion Model (LDM) [33] is a type of diffusion model trained in the low dimensional latent space to focus on semantic bits of data and reduce computation costs. Given an image  $x \in \mathbb{R}^{H \times W \times 3}$ , the encoder  $\mathcal{E}$  encodes  $x$  into the latent representation  $z \in \mathbb{R}^{h \times w \times c}$  and decoder reconstructs the image from the latent.
+
+With the pretrained encoder, they encode the entire images in the dataset and train a diffusion model on latent space  $z$ , by predicting noise  $\epsilon$  from the noised version of latent  $z_{t}$  at time step  $t$ . The corresponding training objective is
+
+$$
+L _ {\mathrm {L D M}} = \mathbb {E} _ {z, \epsilon , t} [ \| \epsilon - \epsilon_ {\theta} (z _ {t}, t, y) \| _ {2} ^ {2} ], \tag {1}
+$$
+
+where  $\epsilon \in \mathcal{N}(0,1)$  is a noise,  $t$  is the number of time steps which uniformly sampled from  $\{1,\dots,T\}$ ,  $y$  is a condition, and  $\epsilon_{\theta}$  is a neural network which predicts the noise added to the  $z$ .
+
+In our work, we utilize Stable Diffusion (SD) [33] which is the only publicized large-scale pre-trained DM. In the case of SD,  $y$  is a text, and  $\epsilon_{\theta}$  is a U-Net architecture in which a block for each resolution comprises a residual block, self-attention block (SA), and cross-attention block (CA), sequentially. Among these modules, we focus on the SA block to transfer the styles, as discussed in Sec. 1. Given a feature  $\phi$  after the residual block, the self-attention block performs as follows:
+
+$$
+Q = W _ {Q} (\phi), K = W _ {K} (\phi), V = W _ {V} (\phi),
+$$
+
+$$
+\phi_ {\text {o u t}} = \operatorname {A t t n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) \cdot V, \tag {2}
+$$
+
+where  $d$  denotes the dimension of the projected query, and  $W_{(\cdot)}$  is a projection layer. Note that, we don't use any text conditions, so the variable  $y$  is always an empty text prompt  $(\ldots)$ .
+
+# 4. Method
+
+In this paper, we aim to solve artistic style transfer by leveraging the generative capability of a pre-trained large-scale text-to-image diffusion model. Briefly, artistic style transfer is the task of modifying the style of a given content image  $I^{c}$  to that of style image  $I^{s}$ . Then, the stylized image  $I^{cs}$  would maintain the semantic content of  $I^{c}$  while its style (such as texture) is transferred from  $I^{s}$ . For simplicity, we skip the explanations about the encoding and decoding process of the autoencoder in the LDM. Instead, we focus on elaborating the proposed method in the aspect of the diffusion process. Thus, in the following sections, we regard the content, style, and stylized images same as their encoded counterparts  $z_{0}^{c}, z_{0}^{s}$ , and  $z_{0}^{cs}$ .
+
+![](images/d557a0ed5ba8b09332a9a9cb85c7f7e3dad6ad5ddc3172013a934b5322b0b4d1.jpg)  
+Figure 3. Overall framework. (Left) Illustration for the proposed style transfer method. We first invert content image  $z_0^c$  and style image  $z_0^s$  into the latent noise space as  $z_T^c$  and  $z_T^s$ , respectively. Then, we initialize the initial noise of stylized image  $z_T^{cs}$  from initial latent AdaIN (Sec. 4.3) which combines the content and style noise,  $z_T^c$  and  $z_T^s$ . While performing the reverse diffusion process with  $z_T^{cs}$ , we inject the information of content and style by attention-based style injection (Sec. 4.1) and attention temperature scaling (Sec. 4.2). (Right) Detailed explanation of style injection and initial noise AdaIN. Style injection is basically the manipulation of self-attention (SA) layer during the reverse diffusion process. Specifically, at time step  $t$ , we substitute the key  $(K_t^{cs})$  and value  $(V_t^{cs})$  in SA of stylized image with those of style features,  $K_t^s$  and  $V_t^s$ , from identical timestep  $t$ . At the same time, we preserve the content information by blending the query of content  $Q_t^c$  and query of stylized image  $Q_t^{cs}$ . Finally, we scale the magnitude of the attention map to deal with the magnitude decrease that the substitution of feature leads to. Initial latent AdaIN produces the initial noise  $z_T^{cs}$  by combining style noise  $z_T^s$  and content noise  $z_T^s$ . Specifically, we modify the channel statistics of  $z_T^c$  to resemble the statistics of  $z_T^s$  and regard it as  $z_T^{cs}$ . We observe this operation enables us to keep the spatial layout of content image while well-reflecting the color tones of a given style image.
+
+![](images/25341082c1bb4fa13bb877698f1b3db935c09e0b80ac2ef541c5291fe7f10777.jpg)
+
+# 4.1. Attention-based Style Injection
+
+We start from the observation in previous image-to-image translation methods, especially Prompt-to-Prompt [14]. The key idea of their method is changing the text condition for cross-attention (CA) while keeping the attention map. Since the attention map affects the spatial layout of output, substituted text conditions determine what to draw in the generated image, and these conditions are actually key and value in CA. Inspired by them, we manipulate the features in self-attention layer as like cross-attention, regarding the features from style image  $I^{s}$  as the condition. Specifically, in the generation process, we substitute the key and value of content image with those of style for transferring the texture of style image into the content image.
+
+To this end, we first obtain the latent for content and style images with DDIM inversion [39], and then collect the SA features of style image over the DDIM inversion process. Specifically, for pre-defined timesteps  $t = \{0,\dots,T\}$ , style and content images  $z_0^c$  and  $z_0^s$  are inverted from image  $(t = 0)$  to gaussian noise  $(t = T)$ . During DDIM inversion, we also collect query features of content  $(Q_t^c)$  and key and value features of style  $(K_t^s,V_t^s)$  at every time steps.
+
+After that, we initialize stylized latent noise  $z_T^{cs}$  by copy-
+
+ing content latent noise  $z_{T}^{c}$ . Then, we transfer the target style to the stylized latent by injecting the key  $K_{t}^{s}$  and value  $V_{t}^{s}$  collected from the style into SA layer, instead of the original key  $K_{t}^{cs}$  and value  $V_{t}^{cs}$ , when performing the entire reverse process of stylized latent  $z_{t}^{cs}$ . However, only applying this substitution can lead to content disruption, since the content of stylized latent would be progressively changed as attended value changes. Hence, we propose query preservation to maintain original content. Simply, we blend query of stylized latent  $Q_{t}^{cs}$  and that of content  $Q_{t}^{c}$  for the entire reverse process. These style injection and query preservation processes at time step  $t$  are expressed as follows:
+
+$$
+\tilde {Q} _ {t} ^ {c s} = \gamma \times Q _ {t} ^ {c} + (1 - \gamma) \times Q _ {t} ^ {c s}, \tag {3}
+$$
+
+$$
+\phi_ {\text {o u t}} ^ {\mathrm {c s}} = \operatorname {A t t n} \left(\tilde {Q} _ {t} ^ {c s}, K _ {t} ^ {s}, V _ {t} ^ {s}\right), \tag {4}
+$$
+
+where  $\gamma$  is degree of blending in range of [0, 1]. In addition, we apply these operations on the latter layers of decoder (7-  $12^{\mathrm{th}}$  decoder layers in SD) relevant to local textures. We also highlight that the proposed method can adjust the degree of style transfer by changing query preservation ratio  $\gamma$ . Specifically, higher  $\gamma$  maintains more content, while lower  $\gamma$  strengthens effects of style transfer.
+
+![](images/3dfeb95f4b4a72fd4f26c9fa31b3133f0dc69f21163d9f1bc6c7575f44d06a02.jpg)  
+(a) Std. of attention map  
+(b) Ratio of std. of attention map
+
+![](images/ac32451fdf99d903083809667e55399f7e6415ded3047660913ea5c999dfe958.jpg)  
+Figure 4. Visualization of the standard deviation of attention map before softmax. (a) Attention-based style injection reduces the standard deviation of self-attention map. Original denotes SA maps from the generation process without style injection. We use both style and content images for generation. (b) We compute the ratio between attention maps w/ and w/o style injection. For the std of original image, we use averaged std. of content and style.
+
+# 4.2. Attention Temperature Scaling
+
+Attention map is computed by scaled dot-product between query and key features. During training, query and key features in the SA layer originate from an identical image. However, if we substitute the key features with those of style images, the magnitude of similarity would be overall lowered as style and content are highly likely to be irrelevant. Thus, the computed attention map can be blurred or smoothed, and it would further make output images unsharp, which is detrimental to capturing both content and style information.
+
+To quantify this issue, we measure the standard deviation of attention map, while ablating the attention-based style injection. In detail, we calculate the attention map before applying softmax, which is scaled-dot product between query and key. As shown in Fig. 4 (a), we validate that this style injection tends to lower the standard deviation of the attention map over the entire timesteps. That is, attention maps after softmax with style injection would be overly smooth.
+
+To rectify the attention map sharper, we introduce an attention temperature scaling parameter. In detail, we multiply the attention map before softmax by a constant temperature scaling parameter  $\tau$  larger than 1. Thus, the attention map after softmax would be sharper than its original values. The modified attention process is represented as follows:
+
+$$
+\operatorname {A t t n} _ {\tau} \left(\tilde {Q _ {t} ^ {c s}}, K _ {t} ^ {s}, V _ {t} ^ {s}\right) = \operatorname {s o f t m a x} \left(\frac {\tau \tilde {Q _ {t} ^ {c s}} \left(K _ {t} ^ {s}\right) ^ {T}}{\sqrt {d}}\right) \cdot V _ {t} ^ {s}, \tau > 1. \tag {5}
+$$
+
+We use  $\tau = 1.5$  as a default setting, which is the average ratio over entire timesteps. As reported in Fig. 4 (b), we confirm that it effectively calibrates the standard deviation of attention map similar to its original values.
+
+# 4.3. Initial Latent AdaIN
+
+In artistic style transfer, the color tone generally takes up a significant portion of the style information. In this context, we observe that the style transfer only with attention-based
+
+![](images/8cfdad2fb268773913895461a5819b09eacd13147ba70e421af6bead6f86e5ba.jpg)  
+(a) Generated results only w/ style injection  $(\mathrm{K}^s,V^s)$
+
+![](images/dde322ca136330611fdd146218e09cf6e5cd3e362c095d788c8b9605543f7f15.jpg)  
+njection  $(\mathrm{K}^s,V^s)$
+
+![](images/7c4e3c4344e80896acccab14266518ab7f61d2f69ac574ece9e3eb52bc3f048a.jpg)  
+(b) Generated results only w/ style injection  $(\mathrm{Q}^s,\mathrm{K}^s,V^s)$  
+Figure 5. Generated results only w/ style injection. (a) We observe that generated images only with attention-based style injection do not harmonize with the given style in the aspect of color tone. (b) To identify the effects of every feature in SA on color tones, we additionally include query in the style injection process. However, color tones still resemble those of content, concluding features in self-attention have less effect on the color tones.
+
+![](images/8c8d4174f83169458e3f8dc0035282fc426522048821c800fcd49fd4a8947b3f.jpg)
+
+style injection often fails in terms of capturing the color tone of the given style. As shown in Fig. 5 (a), textures and local patterns are successfully transferred to the content image while the color tone of the content image still remains. Furthermore, even with injecting the query, key, and value of styles, the resulting images still preserve the color tone of the content, as shown in Fig. 5 (b).
+
+As substituting the self-attention features has less effects to color tone, we analyze the other vital part of DM; initial latent noise. One of the recent discoveries in DM is that the DM struggles to synthesize purely white or black images [13]. Instead, they tend to generate images of median color as the initial noise is sampled from zero mean and unit variance. Thus, we hypothesize the statistics of initial noise largely affect the colors and brightness of generated images.
+
+Based on this assumption, we attempt to use the initial latent of style  $z_{T}^{s}$  for the style transfer process. However, if we simply start to generate the image from style latent  $z_{T}^{s}$ , the structural information of synthesized results also follows the style image and loses the structure of the content. To harness valuable information in both initial latents, we consider that the tone information is intricately connected with the channel statistics of the initial latent, following the principle underlying Style Loss [11] and AdaIN [17]. Thus, we employ AdaIN to modulate the initial latent for effective tone information transfer, represented as:
+
+$$
+z _ {T} ^ {c s} = \sigma \left(z _ {T} ^ {s}\right) \left(\frac {z _ {T} ^ {c} - \mu \left(z _ {T} ^ {c}\right)}{\sigma \left(z _ {T} ^ {c}\right)}\right) + \mu \left(z _ {T} ^ {s}\right), \tag {6}
+$$
+
+where  $\mu (\cdot),\sigma (\cdot)$  denote channel-wise mean and standard deviation, respectively. Based on this, the initial latent  $z_{T}^{cs}$
+
+<table><tr><td>Metric</td><td>Ours</td><td>AesPA-Net</td><td>CAST</td><td>StyTR2</td><td>EFDM</td><td>MAST</td><td>AdaAttn</td><td>ArtFlow</td><td>AdaConv</td><td>AdaIN</td><td>DiffuseIT</td><td>InST</td><td>DiffStyle</td></tr><tr><td>ArtFID ↓</td><td>28.801</td><td>31.420</td><td>34.685</td><td>30.720</td><td>34.605</td><td>31.282</td><td>30.350</td><td>34.630</td><td>31.856</td><td>30.933</td><td>40.721</td><td>40.633</td><td>41.464</td></tr><tr><td>FID ↓</td><td>18.131</td><td>19.760</td><td>20.395</td><td>18.890</td><td>20.062</td><td>18.199</td><td>18.658</td><td>21.252</td><td>19.022</td><td>18.242</td><td>23.065</td><td>21.571</td><td>20.903</td></tr><tr><td>LPIPS ↓</td><td>0.5055</td><td>0.5135</td><td>0.6212</td><td>0.5445</td><td>0.6430</td><td>0.6293</td><td>0.5439</td><td>0.5562</td><td>0.5910</td><td>0.6076</td><td>0.6921</td><td>0.8002</td><td>0.8931</td></tr><tr><td>CFSD ↓</td><td>0.2281</td><td>0.2464</td><td>0.2918</td><td>0.3011</td><td>0.3346</td><td>0.3043</td><td>0.2862</td><td>0.2920</td><td>0.3600</td><td>0.3155</td><td>0.3428</td><td>0.6759</td><td>0.2819</td></tr></table>
+
+Table 1. Quantitative comparison with conventional  $\left( {{3}^{\text{rd }} - {11}^{\text{th }}\text{columns}}\right)$  and diffusion model baselines  $\left( {{12}^{\text{th }} - {14}^{\text{th }}\text{columns}}\right)$
+
+preserves content information from  $z_T^c$  while aligning the channel-wise mean and standard deviation with  $z_T^s$ .
+
+# 5. Experiments
+
+# 5.1. Experimental Settings
+
+We conduct all experiments in Stable Diffusion 1.4 pretrained on LAION dataset [36] and adopt DDIM sampling [39] with a total 50 timesteps ( $t = \{1, \dots, 50\}$ ). For default settings for hyperparameters, we use  $\gamma = 0.75$  and  $\tau = 1.5$ , if they are not mentioned separately.
+
+# 5.2. Evaluation Protocol
+
+Conventional style transfer methods typically utilize Style Loss [11] as both training objective and evaluation metric, so their results tend to overfit the Style Loss. Thus, for a fair comparison, we employ a recently proposed metric, ArtFID [46] which evaluates overall style transfer performances with consideration of both content and style preservation and also is known as strongly coinciding with human judgment. Specifically, ArtFID is computed as  $(\text{ArtFID} = (1 + \text{LPIPS}) \cdot (1 + \text{FID}))$ . LPIPS [50] measures content fidelity between the stylized image and the corresponding content image, and FID [15] assesses the style fidelity between the stylized image and the corresponding style image.
+
+Dataset. Our evaluations employ content images from MS-COCO [27] dataset and style images from WikiArt [40] dataset. All input images are center-cropped to  $512 \times 512$  resolution. Also, for quantitative comparison, we randomly selected 20 content and 40 style images from each dataset, yielding 800 stylized images as StyTR² [9] has done.
+
+Content Feature Structural Distance (CFSD). In the style transfer evaluation, the assessment of content fidelity often relies on the LPIPS distance. However, since LPIPS utilizes the feature space of AlexNet [20] pre-trained for classification task on ImageNet [7], which is known as texture-biased [12]. Thus, the style information of the images can affect the LPIPS score. To mitigate this style influence, we additionally introduce Content Feature Structural Distance (CFSD) which is a distance measure that only considers the spatial correlation between image patches.
+
+In detail, we first define the correlation map between image patch features as follows. For a given image  $I$ , we obtain feature maps  $F \in \mathbb{R}^{hw \times c}$ , which is the output feature of conv3 in VGG19 [38]. Then, we calculate the patch sim
+
+The similarity map  $M = F \times F^T$ ,  $M \in \mathbb{R}^{hw \times hw}$ , which is a similarity map between every pair of features in  $F$ . After that, for computing the distance between two patch similarity maps, we model the similarity between a single patch and the others as a probability distribution by applying softmax operation. Finally, the correlation map is represented as  $S = [\text{softmax}(M_i)]_{i=1}^{hw}$ ,  $S \in \mathbb{R}^{hw \times hw}$ , where  $M_i \in \mathbb{R}^{1 \times hw}$  is a similarity map between  $i^{\text{th}}$  patch and the other patches.
+
+Then, CFSD is defined as KL-divergence between two correlation maps. In our case, we compute CFSD between the correlation map of the content  $(S^c)$  and stylized images  $(S^{cs})$  as follows:
+
+$$
+\mathrm {C F S D} = \frac {1}{h w} \sum_ {i = 1} ^ {h w} D _ {\mathrm {K L}} \left(S _ {i} ^ {c} | | S _ {i} ^ {c s}\right), \tag {7}
+$$
+
+# 5.3.Quantitative Comparison
+
+We evaluate our proposed method through comparison with twelve state-of-the-art methods, including nine conventional style transfer methods (AesPA-Net [16], CAST [52], StyTR $^2$  [9], EFDM [51], MAST [8], AdaAttN [28], ArtFlow [1], AdaConv [5], AdaIN [17]) and three diffusion-based style transfer methods (DiffuseIT [22], InST [53], DiffStyle [18]), which have a style image as input. We employ the publicly available implementations of all baselines, using their recommended configurations.
+
+Comparison with Conventional Style Transfer. As shown in Tab. 1, our method largely surpasses the conventional style transfer methods in terms of ArtFID, which is known as coinciding the human preference. In addition, the proposed method records the lowest FID, which denotes that stylized images highly resemble the target styles. For content fidelity metrics, ours shows superior scores in both CFSD and LPIPS. We point out that ours achieves much lower CFSD compared to other methods, which is the metric to only consider the spatial correlation.
+
+In addition, we also emphasize that the proposed method can arbitrarily adjust the degree of style transfer by changing the  $\gamma$ , and the proposed method significantly surpasses all the other methods in terms of FID (style), when we match the value of LPIPS (content) (Fig. 10).
+
+Comparison with Diffusion-based Style Transfer. Our method demonstrates the best performance in terms of LPIPS, FID, and their combination (ArtFID) with a large
+
+Content Style Ours AesPA-Net CAST StyTR2 EFDM MAST AdaAttN AdaIN DiffuseIT InST DiffStyle
+
+![](images/860ff3210846250b16d879d171e60f15c4647f18d2abd94bffd6a5026d2fa2ad.jpg)  
+Figure 6. Qualitative comparison with conventional  $(4^{\mathrm{th}} - 10^{\mathrm{th}}$  columns) and diffusion model baselines  $(11^{\mathrm{th}} - 13^{\mathrm{th}}$  columns)
+
+![](images/2838e8ad98d425397cce2a5d39e7ef935a30c31d0c61a70bce8bd7758a67ed21.jpg)  
+Figure 7. Qualitative comparison with best ArtFID (AdaAttN) and most recently proposed baselines (AesPA-Net) with additional zooming details
+
+<table><tr><td>Metric</td><td>DiffuseIT</td><td>InST</td><td>DiffStyle</td><td>Ours</td></tr><tr><td>Time (sec)</td><td>792.8</td><td>816.9</td><td>355.9</td><td>12.4</td></tr></table>
+
+Table 2. Comparison of inference time of diffusion-based methods for style transferring a given style and content pair
+
+margin, as shown in Tab. 1. One significant factor for diffusion models is their running times, since they require several steps to synthesize a single image and it requires inevitable time cost. Hence, we measure the inference time for a pair of content and style images on a single Titan RTX GPU, as shown in Tab. 2. Our method requires a total of 12.4 seconds, with 8.2 seconds for DDIM inversions and 4.2 seconds for sampling costs. As reported, we validate the proposed method significantly faster than other methods, even exploiting large-scale DM. This faster speed comes from the fact that the proposed methods can use the much smaller steps of DDIM inversion, because we additionally utilize the features collected during inversion steps, largely reducing the necessity for perfect inversion of content and style.
+
+# 5.4. Qualitative Comparison
+
+Comparison with Conventional Style Transfer. As shown in Fig. 6, we observe that our method tends to highly preserve the structural information of the content image, while also transferring the style well. For instance, as shown in the third row, ours retains the structure of the
+
+bridge, but the baselines struggle to preserve structure or transfer the style. We also provide the qualitative comparison with zooming details in Fig. 7 and Supplementary.
+
+Comparison with Diffusion-based Style Transfer. We also compare our method with recent diffusion-based style transfer baselines [18, 23, 53]. As shown in Fig. 6, we observe the proposed technique transfers the style to the content well. On the other hand, baselines often lose the structure of content or fail to transfer the style, when an arbitrary content style pair is given. For instance, DiffuseIT and DiffStyle suffer from generating shape and visually plausible images or drop the original content. Differently, InST synthesizes the realistic images, while struggling to transfer style (1 $^{\text{st}}$  row) or change content of image (2 $^{\text{nd}}$ , 3 $^{\text{rd}}$  rows).
+
+# 5.5. Ablation Study
+
+To validate the effectiveness of the proposed components, we conduct ablation studies in both quantitative and qualitative ways. As shown in Fig. 8 and Tab. 3, style injection is significant for guiding the style and content of given images (Config. B). Besides, initial latent AdaIN has a large portion of transferring the color tone of style (Config. D). Attention temperature scaling is in charge of enhancement of quality in synthesized results such as sharpening details and resolving blurriness. For instance, this scaling jointly reduces the FID and LPIPS (Config.  $\mathrm{A}^*$  vs. C in Tab. 3). For more detailed analysis, we provide quantitative metrics with the style-content trade-off, while changing the attention scaling parameter  $\tau$  in Fig. 10 (b). As reported, attention scaling effectively reduces both FID and LPIPS, proving its effects on the preservation of content and capability of style transfer ( $\tau = 1.0$  vs.  $\tau = 1.5$ ).
+
+# 5.6. Additional Analysis
+
+Content-Style Trade-Off. Our proposed method offers flexible control of the trade-off relation between content and style fidelity by adjusting the parameter  $\gamma$ , as discussed in
+
+<table><tr><td></td><td>Configuration</td><td>ArtFID</td><td>FID</td><td>LPIPS</td></tr><tr><td>A</td><td>Ours (γ = 0.75, default)</td><td>28.80</td><td>18.13</td><td>0.505</td></tr><tr><td>A*</td><td>Ours (γ = 0.6)</td><td>27.97</td><td>17.21</td><td>0.535</td></tr><tr><td>B</td><td>- Style Inject. (Sec. 4.1)</td><td>43.72</td><td>27.13</td><td>0.554</td></tr><tr><td>C</td><td>- Attn. Scaling (Sec. 4.2)</td><td>29.02</td><td>17.81</td><td>0.542</td></tr><tr><td>D</td><td>- Init. AdaIN (Sec. 4.3)</td><td>29.26</td><td>20.05</td><td>0.390</td></tr></table>
+
+Table 3. Ablation study on proposed components
+
+![](images/b89ab7b4bcfdd501d4f0c4ece99c1f74adbac5ab3e9aafff8a18fd52f34f13c6.jpg)  
+Figure 8. Qualitative comparison with ablation studies
+
+![](images/4cf5a63c61f5638d6fffff5dc9acbbc0b67f136c825e93b8776c146c9759bf5c.jpg)  
+Figure 9. Visualization of effects of query preservation ratio  $\gamma$
+
+Sec. 4.1. In detail, we compute the FID and LPIPS while varying  $\gamma$  within the range [0.3, 1] with a step size of 0.1. As shown in Fig. 10 (a), our method surpasses baseline methods across all ranges of content and style fidelity. This result implies the proposed method significantly outperforms the other methods, when we match style or content metric to the compared model by adjusting the  $\gamma$  of ours. Note that, dotted lines refer to our model reported in Tab. 1.
+
+We also visualize the effects of the style-content trade-off by synthesizing images by adjusting  $\gamma$ . As shown in Fig. 9, the lower  $\gamma$  highly reflects the style while losing the content of the given image, and vice versa. This characteristic of the proposed method suggests that the users can adjust the degree of style by following their personal preferences.
+
+Study on the value of  $\tau$ . We observe that the gradual increase of  $\tau$  enhances the performance of style transfer, although its effects on enhancement become smaller as  $\tau$  goes larger, as shown in Fig. 10 (b). This result implies that the attention temperature scaling effectively works with a simple modification of the magnitude of the attention map.
+
+![](images/0574ed7e223feb7c0b3309243ecdcabf39ec3c7cba7af2ebe674151a020c830e.jpg)  
+Figure 10. Style-content trade-offs
+
+![](images/d13886104303c34f63fe8e4d7a7cc55500466e7e10422acf2cfaad9f948dd77d.jpg)
+
+![](images/6f76cbcee65f8d474e662ff237a8efaf946791fe47816741a406344a17535b5e.jpg)  
+Figure 11. Comparison with text-guided style transfer methods
+
+Comparison with text-guided style transfer. We additionally compare the proposed method with the style transfer methods [21, 41] which are conditioned on the textual inputs. As text-guided methods tend to modify the style largely, we use  $\gamma = 0.3$  for this experiment. Since the text condition hardly contains all the information in the style image such as texture and color tones, the transferred results less resemble the target style, as shown in Fig. 11. Differently, we validate that the proposed method successfully transfers the style with high fidelity.
+
+# 6. Conclusion
+
+Our work addresses the challenges associated with diffusion model-based style transfer methods, which often require time-consuming optimization steps or struggle to leverage the generative potential of large-scale diffusion models. To this end, we propose the method of adapting the pre-trained large-scale diffusion model on style transfer in a training-free way. Our method focuses on manipulating the features of self-attention layers, akin to the cross-attention mechanism, by substituting the key and value during the content generation process with those of the style. Furthermore, we propose the query preservation and attention temperature scaling to mitigate the issue of disruption of content, and initial latent AdaIN to handle the disharmonious color (failure to transfer the colors of style). Experimental results show the superiority of our proposed method over state-of-the-art techniques in previous baselines.
+
+# Acknowledgments
+
+This work was supported in part by MSIT/IITP (No. 2022-0-00680, 2019-0-00421, 2020-0-01821, 2021-0-02068), and MSIT&KNPA/KIPoT (Police Lab 2.0, No. 210121M06).
+
+# References
+
+[1] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 862-871, 2021. 6  
+[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 1, 3  
+[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560-22570, 2023. 3  
+[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 1  
+[5] Prashanth Chandran, Gaspard Zoss, Paulo Gotardo, Markus Gross, and Derek Bradley. Adaptive convolutions for structure-aware style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7972-7981, 2021. 6  
+[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR 2023 (Eleventh International Conference on Learning Representations), 2023. 1, 3  
+[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6  
+[8] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM international conference on multimedia, pages 2719-2727, 2020. 6  
+[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 6  
+[10] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Susstrunk, and Radhakrishna Achanta. Diffusion in style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2251-2261, 2023. 1, 3  
+[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 5, 6  
+[12] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 6
+
+[13] Nocholas Guttenberg. Diffusion with offset noise. https://www.crosslabs.org/blog/diffusion-with-offset-noise, 2023.5  
+[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 4  
+[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6  
+[16] Kibeom Hong, Seogkyu Jeon, Junsoo Lee, Namhyuk Ahn, Kunhee Kim, Pilhyeon Lee, Daesik Kim, Youngjung Uh, and Hyeran Byun. Aespa-net: Aesthetic pattern-aware style transfer networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22758-22767, 2023. 6  
+[17] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5, 6  
+[18] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-free style transfer emerges from h-space in diffusion models. arXiv preprint arXiv:2303.15403, 2023. 1, 3, 6, 7  
+[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Conference on Computer Vision and Pattern Recognition 2023, 2023. 1, 3  
+[20] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 6  
+[21] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 8  
+[22] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264, 2022. 6  
+[23] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 3, 7  
+[24] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 2  
+[25] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017.  
+[26] Yijun Li, Ming-Yu Liu, Xueting Li, Ming-Hsuan Yang, and Jan Kautz. A closed-form solution to photorealistic image
+
+stylization. In Proceedings of the European conference on computer vision (ECCV), pages 453-468, 2018. 2  
+[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6  
+[28] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 6  
+[29] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5952-5961, 2019. 2  
+[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1  
+[31] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880–5888, 2019. 2  
+[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. 3  
+[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. 1, 3  
+[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1  
+[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1  
+[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6  
+[37] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 3
+
+[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6  
+[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 4, 6  
+[40] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 6  
+[41] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2, 3, 8  
+[42] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022.3  
+[43] Huan Wang, Yijun Li, Yuehai Wang, Haoji Hu, and Ming-Hsuan Yang. Collaborative distillation for ultra-resolution universal style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1860-1869, 2020. 2  
+[44] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7789-7798, 2020. 2  
+[45] Zhizhong Wang, Lei Zhao, and Wei Xing. *Style diffusions: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 1, 3  
+[46] Matthias Wright and Björn Ommer. Artfid: Quantitative evaluation of neural style transfer. In DAGM German Conference on Pattern Recognition, pages 560-576. Springer, 2022. 6  
+[47] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. arXiv preprint arXiv:2303.08622, 2023. 1, 3  
+[48] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 1  
+[49] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. arXiv preprint arXiv:2305.15347, 2023. 2  
+[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6  
+[51] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of
+
+the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8035-8045, 2022. 6  
+[52] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 2, 6  
+[53] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10146-10156, 2023. 1, 3, 6, 7

2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:adb39747c2d5a9d5ab7721b345788144334ce4e33206e10f4ddea5fad134f18d
+size 843834

2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_content_list.json

@@ -0,0 +1,1594 @@
+[
+    {
+        "type": "text",
+        "text": "StyleCineGAN: Landscape Cinemagraph Generation using a Pre-trained StyleGAN",
+        "text_level": 1,
+        "bbox": [
+            215,
+            130,
+            754,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jongwoo Choi",
+        "bbox": [
+            158,
+            203,
+            279,
+            220
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Kwanggyoon Seo",
+        "bbox": [
+            318,
+            203,
+            462,
+            220
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Amirsaman Ashtari",
+        "bbox": [
+            501,
+            203,
+            660,
+            220
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Junyong Noh",
+        "bbox": [
+            699,
+            203,
+            807,
+            220
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Visual Media Lab, KAIST",
+        "bbox": [
+            379,
+            222,
+            589,
+            238
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{jeolpyeoni0,skg1023,a.s.ashtari,junyongnoh}@kaist.ac.kr",
+        "bbox": [
+            235,
+            241,
+            728,
+            257
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Figure 1. Given a landscape image, StyleCineGAN generates a seamless cinematograph at  $1024 \\times 1024$  resolution. This figure contains video clips, thus consider viewing it using Adobe Reader. The same results are also included in the supplementary video.",
+        "bbox": [
+            75,
+            465,
+            893,
+            496
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            508,
+            313,
+            523
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We propose a method that can generate cinematographs automatically from a still landscape image using a pretrained StyleGAN. Inspired by the success of recent unconditional video generation, we leverage a powerful pretrained image generator to synthesize high-quality cinematographs. Unlike previous approaches that mainly utilize the latent space of a pre-trained StyleGAN, our approach utilizes its deep feature space for both GAN inversion and cinematograph generation. Specifically, we propose multi-scale deep feature warping (MSDFW), which warps the intermediate features of a pre-trained StyleGAN at different resolutions. By using MSDFW, the generated cinematographs are of high resolution and exhibit plausible looping animation. We demonstrate the superiority of our method through user studies and quantitative comparisons with state-of-the-art cinematograph generation methods and a video generation method that uses a pre-trained StyleGAN.",
+        "bbox": [
+            75,
+            527,
+            472,
+            785
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            801,
+            209,
+            816
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Cinemagraph is a unique form of media that combines a still image and video. While most of the scene remains still, subtle and repeated movements in only a small region effectively highlight a particular object or an important event that occurred at the time of capture. Because this mixture",
+        "bbox": [
+            75,
+            824,
+            470,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "of still and moving elements in one scene creates an interesting eye-catching effect, cinematographs have recently attracted increasing attention in social media.",
+        "bbox": [
+            496,
+            508,
+            890,
+            555
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite the growing popularity of cinematograph, its creation heavily relies on a manual process. To create a motion in part of the input image, one typically utilizes an image manipulation tool to stretch, scale, rotate, or shift the part of the image in a physically plausible and aesthetically pleasing way. This process is often time-consuming and requires a high level of skill in the use of image editing tools. Therefore, creating a cinematograph has generally been considered to be a personal project of professionals so far.",
+        "bbox": [
+            496,
+            561,
+            892,
+            698
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To allow ordinary users to create a cinematograph with a single image, automatic methods have been proposed. One line of research uses a reference video as guidance [14, 25, 33]. These methods are capable of producing realistic motions in the target scene similar to a reference video. Recent methods obviate the need for a reference video, by training deep generative models [4, 6, 9, 15, 17-19, 23]. These methods decompose the task into two processes of learning motion and spatial information individually from separate datasets. The decomposition effectively reduces the complexity of simultaneously learning both temporal and spatial information, and improves the perceptual quality of the generated videos. However, these models",
+        "bbox": [
+            496,
+            703,
+            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,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "7872",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "must be trained from scratch, which sometimes requires several days if not weeks on a modern GPU. Moreover, these models are not specifically designed to generate high-resolution of  $1024 \\times 1024$  cinemagrams, because of the significant requirements in memory and processing power.",
+        "bbox": [
+            75,
+            90,
+            467,
+            166
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose the first approach to high-quality one-shot landscape cinematograph generation based on a pre-trained StyleGAN [12]. By using the learned image prior of StyleGAN, our method removes the need for training a large model from scratch and systematically improves the resolution of the generated cinematographs to  $1024 \\times 1024$ . Moreover, our method enables a user to easily edit the style and appearance of the generated cinematograph by leveraging the properties of StyleGAN for image stylization and editing.",
+        "bbox": [
+            75,
+            166,
+            467,
+            316
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our method is inspired by recent unconditional video generation methods [7, 34] which allow to navigate the latent space of a pre-trained image generator, to synthesize a high-quality temporally coherent video. Unlike these methods that utilize the latent codes of StyleGAN, we opt to use the deep features that are generated by convolution operations in each layer of StyleGAN. We use these deep features for two reasons. First, we observed that highly detailed landscape images cannot be reconstructed accurately from the latent codes using GAN inversion methods because these latent codes are low-dimensional [30]. Second, a plausible motion that preserves the content cannot be created by only navigating the latent space, because this space is highly semantic-condensed and lacks explicit spatial prior [38].",
+        "bbox": [
+            75,
+            316,
+            467,
+            542
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To solve these problems, for the first time in the cinematograph generation domain, we utilize the deep features of a pre-trained StyleGAN. These deep features preserve spatial information and encode both high-level semantic and low-level style appearances across high and low-level convolutional layers. To produce cinematographs from these deep features, we propose a multi-scale deep feature warping (MSDFW) to apply the motions generated using a motion generator to the deep feature space of StyleGAN.",
+        "bbox": [
+            75,
+            542,
+            467,
+            676
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We demonstrate the effectiveness and advantages of our method by comparing it to state-of-the-art methods in cinematagraph generation [6, 9, 17, 19] and to an unconditional video generation method that uses a pre-trained StyleGAN [34], using various metrics. We also performed a user study to verify the effectiveness of our method for creating visually pleasing cinematagraph effects. Both qualitative and quantitative results confirm that our method substantially outperforms all existing baselines.",
+        "bbox": [
+            75,
+            678,
+            467,
+            811
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            824,
+            218,
+            839
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1. Cinemagraph Generation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            849,
+            310,
+            864
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "One of the early studies for cinemagraph generation used a procedural approach to decompose each object into layers",
+        "bbox": [
+            75,
+            869,
+            467,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "and apply time-varying stochastic motion [5]. This method can handle a diverse range of images such as photos and paintings by relying on manual interaction from the user. Another approach is to use a reference video to animate the given image [20-22, 25]. These methods rely on a statistical motion analysis of videos to transfer their periodic motion property to the desired image. Another method distills a dynamic NeRF to render looping 3D video textures from the representative motion found in the reference video [14]. Without any reference video as guidance, Halperin et al. [8] proposed a framework to animate arbitrary objects with periodic patterns in a simple motion direction.",
+        "bbox": [
+            496,
+            90,
+            890,
+            271
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recent approaches use the capacity of deep learning to automatically create cinemagrams from a single image. One study trained an image-based renderer to generate water animation, utilizing the water simulation results [32]. Another research first predicts a sequence of normal maps, then generates the corresponding RGB images that show a looping animation of a garment as if it is blown in the wind [4]. Another line of research trains a generator to produce a motion field [6, 9, 15, 18, 19]. Our work is similar to these deep learning-based approaches in that a motion generator is utilized to synthesize the cinemagrams. The difference is that our method leverages the deep features of a pre-trained image generator and thus can generate plausible high-quality results without fine-tuning or training a separate synthesis network. At the same time, our method systematically improves the resolution of the generated cinemagrams to  $1024 \\times 1024$ .",
+        "bbox": [
+            496,
+            272,
+            890,
+            527
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2. Unconditional Video Generation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            532,
+            784,
+            547
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The task of video generation is known to be difficult because the generation process has to take into account both spatial and temporal information. To ease the problem, previous approaches [27, 36] decompose video generation into content and motion generation. These methods first predict the latent trajectory for motion, then generate a video from the set of predicted latent codes using the image generator. Instead of training the generation model from scratch, MoCoGAN-HD [34] and StyleVideoGAN [7] leverage a pre-trained image generator model, StyleGAN [12]. These methods use the property of well-constructed latent space to morph one frame toward the next frames [10]. Utilizing the image generation capability of a pre-trained generator, MoCoGAN-HD and StyleVideoGAN only need to train a motion trajectory generator, which greatly reduces the training time. These methods can generate high-quality videos in multiple domains such as faces, cars, and even sky. However, the content details are not well preserved, which has hindered the methods' application for cinematograph generation. Our method builds upon a similar concept of using a pre-trained StyleGAN. We leverage the deep features instead of the latent codes of StyleGAN to generate looping and content-preserving cinematographs.",
+        "bbox": [
+            496,
+            551,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "7873",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3. Video Synthesis using pre-trained StyleGAN",
+        "text_level": 1,
+        "bbox": [
+            76,
+            90,
+            454,
+            107
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "A pre-trained generative model, StyleGAN [11-13], has been actively employed for downstream image and video editing applications. For GAN-based applications, one of the essential steps is to project the desired images or videos to the latent space of the pre-trained GAN model, which is known as GAN inversion. The latent codes are obtained using either an optimization technique [1, 2], a trained encoder [26, 35], or a hybrid of both approaches [16]. For video editing applications, previous methods [3, 29, 37, 42] have utilized a pre-trained image encoder to project all of the frames to the latent space. While most of the methods operate in  $\\mathcal{W}+$  space [2], we opt to use both  $w^{+}$  and deep features [24, 43, 44] of the pre-trained StyleGAN to accurately reconstruct the original image and synthesize motion.",
+        "bbox": [
+            75,
+            114,
+            472,
+            327
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Methods",
+        "text_level": 1,
+        "bbox": [
+            76,
+            343,
+            176,
+            359
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "An overview of our method is shown in Figure 2. Given a landscape image  $I$ , our method generates a seamlessly looping cinemagraph  $V = \\{\\hat{I}_0, \\dots, \\hat{I}_N\\}$  using a pre-trained StyleGAN. At the core of our method, we use a pre-trained generator without additional fine-tuning. The overall process is described below.",
+        "bbox": [
+            75,
+            369,
+            468,
+            460
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(A) First, we project the landscape image into both the latent space and the feature space of StyleGAN. To this end, we train an encoder that outputs both latent codes  $w^{+}$  and intermediate features  $D^{10}$  of StyleGAN (Sec. 3.1).",
+            "(B) In addition, we predict a mask  $S$  to divide the image into static and dynamic regions (Sec. 3.2).",
+            "(C) Next, we use a motion generator that accepts the landscape image  $I$  to synthesize a motion field  $M$ , which defines the position of each pixel in the future frames (Sec. 3.3).",
+            "(D) Lastly, we generate the final cinematograph frames using the pre-trained StyleGAN with deep feature warping (DFW) operation added in between the original layers. (Sec. 3.4)."
+        ],
+        "bbox": [
+            60,
+            465,
+            468,
+            675
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the following sections, we will describe the details of each process.",
+        "bbox": [
+            75,
+            681,
+            468,
+            712
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. GAN Inversion",
+        "text_level": 1,
+        "bbox": [
+            76,
+            724,
+            230,
+            739
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The first step for generating a cinematograph is to project the input image to the latent space of a pre-trained StyleGAN. This process is necessary because StyleGAN is an unconditional generator that cannot take a conditioning input. Many previous methods have used  $\\mathcal{W} +$  [2] latent space, which is an extended space of native latent space  $\\mathcal{W}$ . However, we observed that  $w^{+} \\in \\mathcal{W}+$  is not expressive enough to reconstruct the original high-frequency details of the landscape images. Therefore, we chose to use the generated deep features  $D^{10}$  of StyleGAN as well as  $w^{+}$ . The use of  $D^{10}$",
+        "bbox": [
+            75,
+            750,
+            472,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "enables us to recover details in the original input as shown in Figure 8.",
+        "bbox": [
+            496,
+            90,
+            890,
+            121
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To project an image  $I$  to both the latent space and feature space and obtain  $w^{+}$  and  $D^{10}$ , we train an encoder  $E$  similarly to Yao et al. [43], in which the encoder takes  $I$  as input and predicts  $(w^{+}, D^{10})$ . Additional training details can be found in the supplementary material.",
+        "bbox": [
+            496,
+            121,
+            893,
+            198
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Mask Prediction",
+        "text_level": 1,
+        "bbox": [
+            498,
+            208,
+            663,
+            222
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We improve the quality of GAN inversion described in Sec. 3.1 by training an additional classifier that predicts a mask to separate the static and dynamic regions. Using the mask, the structure of the static regions can be preserved. To this end, we train a multi-layer-perceptron (MLP) classifier that accepts the deep features of StyleGAN as its input and outputs the mask that specifies each dynamic region in the image, as performed in DatasetGAN [46].",
+        "bbox": [
+            496,
+            229,
+            890,
+            351
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To train the classifier, we manually annotate 32 segmentation masks  $S$  of the selected images. We then follow the same procedure as DatasetGAN but using the deep features  $D^{i}$  where  $i \\in \\{10,11,\\dots,18\\}$ . We resize all these deep features followed by concatenating them in the channel dimension to construct the input feature  $D^{*}$ . In the end, paired data  $(D^{*},S)$  is constructed. An ensemble of 10 MLP classifiers is trained with a cross-entropy loss. To further improve the performance of the previous work, at inference, we also refine the predicted mask. For additional details, please refer to the supplementary material.",
+        "bbox": [
+            496,
+            353,
+            892,
+            518
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3. Motion Generation",
+        "text_level": 1,
+        "bbox": [
+            498,
+            529,
+            684,
+            542
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To generate motion from a static landscape image  $I$ , we use an Eulerian motion field  $M$  which represents the immediate velocity of each pixel. To predict  $M$  given  $I$ , an imaged-to-image translation network [39] is trained as a motion generator. We train the network with paired data  $(I, M)$  as performed in the previous method [9]. Adding controls in motion generation is also possible, either by using text prompts [19] or drawing arrows [18].",
+        "bbox": [
+            496,
+            551,
+            890,
+            672
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "While the trained network works well to a certain degree, we observed that the predicted motion field  $M$  does not align with the boundaries between the static and dynamic regions and often contains motion in static regions. Thus, we further refine  $M$  by multiplying it with the segmentation mask  $S$ , which effectively removes most errors. The refined  $M$  is then used to simulate the motion of a pixel at time  $t$  using the Euler integration as follows:",
+        "bbox": [
+            496,
+            674,
+            890,
+            794
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} x _ {t + 1} = x _ {t} + M \\left(x _ {t}\\right), \\quad M \\left(x _ {t}\\right) = F _ {t \\rightarrow t + 1} \\left(x _ {t}\\right), \\\\ F _ {0 \\rightarrow t} (x _ {0}) = F _ {0 \\rightarrow t - 1} (x _ {0}) + M \\left(x _ {0} + F _ {0 \\rightarrow t - 1} (x _ {0})\\right), \\tag {1} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            511,
+            806,
+            890,
+            843
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $x_0$  is the source pixel position at time 0,  $F_{0\\rightarrow t}$  is the displacement field that defines the displacement of the pixel position  $x_0$  to the pixel position  $x_t$  at time  $t$ .",
+        "bbox": [
+            496,
+            854,
+            890,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "7874",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/4a67e97121df12534c46aef569f41f801346f128f3bc0015d37c5f820bb34afe.jpg",
+        "image_caption": [
+            "Motion Generation"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            89,
+            524,
+            311
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/db2dd263e1cfeda263e26d7f00f69f00cf198b488f3b2ce313b548f3c3422577.jpg",
+        "image_caption": [
+            "Cinemagraph Generation",
+            "Figure 2. Overview of StyleCineGAN. Given an input landscape image  $I$ , our goal is to generate a cienemagraph using a fixed pre-trained StyleGAN  $G$ . We project the image into both latent codes  $w^{+}$  and deep features  $D^{10}$  of  $G$ . Using the deep features  $D^{*}$ , a mask predictor predicts a segmentation mask  $S$ . To animate the input image, we use a motion generator to predict the motion field  $M$  from  $I$ .  $M$  is refined using  $S$ . Through Euler integration,  $M$  produces the future and past displacement fields  $F_{0\\rightarrow t}$  and  $F_{N\\rightarrow t}$ . To synthesize cienemagraph frames, we add a DFW layer in between the layers of  $G$ . DFW refers to Eqns. 2 and 3. This modification enables the intermediate features of  $G$  to be warped according to  $F_{0\\rightarrow t}$  and  $F_{N\\rightarrow t}$  using a joint splatting method at different resolutions, specifically for the StyleGAN layers indexed with  $i\\in [10,12,14,16,18]$ . The warped deep features are used to synthesize frames  $\\hat{I}_t$  resulting in the final cienemagraph video."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            531,
+            89,
+            890,
+            309
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4. Cinemagraph Generation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            452,
+            310,
+            468
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "After acquiring the displacement field  $F_{0\\rightarrow t}$  (Sec. 3.3), latent codes  $w^{+}$ , and deep features  $D^{10}$  of StyleGAN (Sec. 3.1), we feed these data into the pre-trained StyleGAN with our DFW layer added. We apply forward warping (or splatting) to the original content to generate a cinematograph. Here, the application of forward warping directly to an RGB image often results in images with tearing artifacts. To reduce the artifacts, we warp the deep features  $D_0^i$  of StyleGAN in different resolutions or scales:",
+        "bbox": [
+            75,
+            476,
+            468,
+            609
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nD _ {t} ^ {i} = \\operatorname {W a r p} \\left(D _ {0} ^ {i}, F _ {0 \\rightarrow t}\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            187,
+            617,
+            468,
+            636
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\text{Warp}$  is the forward warping function and  $D_{t}^{i}$  is the warped deep feature at time  $t$  and scale  $i$ . We observed that warping a single deep feature (e.g., only  $D_{0}^{10}$ ) results in blurry textures in the generated cinemagrams. Therefore, we opt to warp the multi-scale deep features; we call this operation a MSDFW. Specifically, we apply warping to the deep features  $D_{0}^{i}$  where  $i \\in [10, 12, 14, 16, 18]$ . The deep feature  $D_{0}^{10}$  is acquired using GAN inversion, and the consequent deep features  $D_{0}^{12}, D_{0}^{14}, \\ldots$ , and  $D_{0}^{18}$  are subsequently generated using both  $D_{0}^{10}$  and  $w^{+}$ , as shown in the rightmost column of Figure 2.",
+        "bbox": [
+            75,
+            643,
+            468,
+            809
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Looping video  $V$  with frame length  $N + 1$  is synthesized using the joint splating technique [9] using the future and past displacement fields  $F_{0\\rightarrow t}$  and  $F_{N\\rightarrow t}$ , generated through Euler integration. The displacement fields are computed with  $t$  times of integration on  $M$  for  $F_{0\\rightarrow t}$ , and  $N - t$  times of integration on  $-M$  for  $F_{N\\rightarrow t}$ . The multi-scale fea",
+        "bbox": [
+            75,
+            810,
+            468,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ture  $D_0^i$  is warped in both directions and is composited to form  $D_t^i$ . Specifically,  $D_t^i$  is computed as a weighted sum of  $Warp(D_0^i,F_{0\\to t})$  and  $Warp(D_0^i,F_{N\\to t})$  as follows:",
+        "bbox": [
+            498,
+            452,
+            890,
+            500
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} D _ {t} ^ {i} \\left(x ^ {\\prime}\\right) = \\frac {\\sum_ {x \\in \\chi} \\alpha_ {t} \\cdot W a r p \\left(D _ {0} ^ {i} (x) , F _ {0 \\rightarrow t}\\right)}{\\sum_ {x \\in \\chi} \\alpha_ {t}} \\\\ + \\frac {\\sum_ {x \\in \\chi} \\left(1 - \\alpha_ {t}\\right) \\cdot W a r p \\left(D _ {0} ^ {i} (x) , F _ {N \\rightarrow t}\\right)}{\\sum_ {x \\in \\chi} \\left(1 - \\alpha_ {t}\\right)}, \\tag {3} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            514,
+            513,
+            890,
+            590
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $x \\in \\chi$  is a set of pixels being mapped to the same destination pixel  $x'$ , and  $\\alpha_{t}$  is the looping weight defined as  $(1 - \\frac{t}{N})$ .",
+        "bbox": [
+            498,
+            602,
+            890,
+            648
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "With the above DFW module, we can generate a video frame given the predicted deep features  $D^{10}$ , latent code  $w^{+}$ , and mask  $S$ :",
+        "bbox": [
+            498,
+            648,
+            890,
+            691
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\hat {I} _ {t} = S \\odot \\left(G _ {\\text {w a r p}} \\left(D _ {0} ^ {1 0}, w ^ {+}, F _ {0 \\rightarrow t}, F _ {N \\rightarrow t}\\right)\\right) \\\\ + (\\mathbf {1} - S) \\odot I, \\tag {4} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            514,
+            700,
+            890,
+            736
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\odot$  is an element-wise multiplication and  $G_{\\text{warp}}$  is a fixed pre-trained StyleGAN that incorporates our DFW module.",
+        "bbox": [
+            498,
+            750,
+            890,
+            792
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Style Interpolation In addition to motion generation, the change of appearance style is an additional feature often observed in a landscape cinematograph. This can be achieved by interpolating the latent code with the target latent code as follows:",
+        "bbox": [
+            496,
+            797,
+            890,
+            869
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nw _ {s} ^ {+} = w ^ {+} \\cdot (1 - \\beta) + w _ {t} ^ {+} \\cdot \\beta , \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            593,
+            883,
+            890,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "7875",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $w_{s}^{+}$  is the interpolated latent code,  $w_{t}^{+}$  is the latent code of the target style, and  $\\beta$  is the interpolation weight. This is possible because the later layers of the StyleGAN only modify the color of the synthesized image while the original structure is preserved.",
+        "bbox": [
+            75,
+            236,
+            468,
+            311
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "With modified Equation 4, we prevent the visual mismatch between the static and dynamic regions by also reflecting the changes to the static regions as follows:",
+        "bbox": [
+            76,
+            313,
+            468,
+            359
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\hat {I} _ {t} = S \\odot \\left(G _ {\\text {w a r p}} \\left(D _ {0} ^ {1 0}, w _ {s} ^ {+}, F _ {0 \\rightarrow t}, F _ {N \\rightarrow t}\\right)\\right) \\tag {6} \\\\ + (\\mathbf {1} - S) \\odot (I + \\Delta I _ {s}), \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            129,
+            371,
+            468,
+            409
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\Delta I _ {s} = G \\left(D _ {0} ^ {1 0}, w _ {s} ^ {+}\\right) - G \\left(D _ {0} ^ {1 0}, w ^ {+}\\right), \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            163,
+            411,
+            468,
+            430
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\Delta I_{s}$  is the color difference between the two images generated from latent codes  $w_{s}^{+}$  and  $w^{+}$ .",
+        "bbox": [
+            76,
+            443,
+            468,
+            474
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            489,
+            209,
+            507
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we compare our method with state-of-the-art landscape cinematograph generation methods [6, 9, 17, 19] (Sec. 4.1) and with an unconditional video generation method [34] (Sec. 4.2). In addition, we show the importance of the components used in our method through ablation studies (Sec. 4.3). To observe more various results, we recommend readers see the supplementary video. Example results are presented in Figures 1 and 3.",
+        "bbox": [
+            75,
+            516,
+            468,
+            638
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1. Comparisons with Cinemagraph Generation Methods",
+        "text_level": 1,
+        "bbox": [
+            76,
+            648,
+            468,
+            679
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We compared our method with state-of-the-art cinemagraph generation methods: Animating Landscape (AL) [6], Deep Landscape (DL) [17], Eulerian Motion Field (EMF) [9], and Text2Cinemagraph (T2C) [19]. AL is a learning-based method that uses backward warping to generate cinemagraph frames. DL trains a style-based generator to synthesize images and dynamic motion in a scene. EMF and T2C train an encoder-decoder architecture paired with a motion generator to produce looping cinemagrams. We used official implementations of AL, DL, and T2C, and faithfully reproduced EMF based on the provided training details and hyper-parameters. In the following, we will examine the qualitative difference, evaluate the results using two metrics, and report the results of human perceptual studies.",
+        "bbox": [
+            75,
+            689,
+            470,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/0969637ceb9ca37a33559ba2e6c9207ff7047175487bea4ca60a64b79dfa815f.jpg",
+        "image_caption": [
+            "Figure 3. Generated cinematagraph results. This figure contains video clips, thus consider viewing it using Adobe Reader. The first two are cinematographs without appearance change, and the last two are cinematographs with appearance change."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            90,
+            869,
+            169
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/39a0213e700ef9bc56ba9bc73a6eeddde37538bac4a9f16f4257e2c9eadcc04f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            171,
+            867,
+            251
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/2b2dff4bbec86926ff16f3a9189aec4cf3d1734c21855e584c6339d7e0200f38.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            253,
+            867,
+            333
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/fd20f8b3e221aecb6ce65ca529f33fe1dbf2ad802f07ce17154bbfe3efe03578.jpg",
+        "image_caption": [
+            "Figure 4. Qualitative comparison with state-of-the-art cinematograph generation methods. Please refer to the supplementary video for more examples."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            335,
+            867,
+            415
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/491143a31de31d72610b9141bf036fd9cfeed33cf9e9ae62ad2c04a6ea78a207.jpg",
+        "table_caption": [
+            "Table 1. Quantitative comparison with the state-of-the-art cinematograph generation methods. We compared our method with AL, EMF, DL, and T2C. The best scores are bolded."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Static Consistency</td><td colspan=\"4\">Motion Quality</td></tr><tr><td>LPIPS↓</td><td>MS-SSIM↑</td><td>RMSE↓</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>RMSE↓</td><td>FID↓</td></tr><tr><td>AL</td><td>0.0477</td><td>0.9524</td><td>8.8956</td><td>0.0617</td><td>0.6819</td><td>25.1089</td><td>53.6893</td></tr><tr><td>EMF</td><td>0.0103</td><td>0.9723</td><td>6.0440</td><td>0.0533</td><td>0.7102</td><td>22.8932</td><td>45.8035</td></tr><tr><td>DL</td><td>0.0071</td><td>0.9931</td><td>2.1044</td><td>0.0504</td><td>0.7062</td><td>21.8011</td><td>41.6374</td></tr><tr><td>T2C</td><td>0.0063</td><td>0.9773</td><td>4.8785</td><td>0.0159</td><td>0.7224</td><td>21.2784</td><td>40.1186</td></tr><tr><td>Ours</td><td>0.0062</td><td>0.9962</td><td>1.9430</td><td>0.0131</td><td>0.7237</td><td>20.8948</td><td>39.2113</td></tr></table>",
+        "bbox": [
+            509,
+            523,
+            880,
+            589
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Qualitative Comparisons Figure 4 shows visual results from AL, EMF, DL, T2C, and our method. AL exhibits stretching artifacts (row 1) due to recurrent backward warping, whereas our method uses forward warping that prevents the appearance of such distortions. EMF struggles with precise reconstruction (row 2), because its encoder-decoder architecture is trained from scratch on a low-quality video dataset. In contrast, our method leverages a StyleGAN model pre-trained on high-quality images, resulting in cinemagrams with high perceptual quality. DL often loses fine details in textures (row 3) as it performs GAN inversion using the latent codes of StyleGAN, and fine-tunes the generator. Our method, utilizing deep features for GAN inversion, retains original texture details. T2C encounters difficulties in preserving high-frequency details, and accurate motion segmentation between static and dynamic regions (row 4). In contrast, our method utilizes a pre-trained StyleGAN for cinematograph generation, and uses its features for mask prediction, preserving both fine and structural details of the source image.",
+        "bbox": [
+            496,
+            606,
+            890,
+            909
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "7876",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Quantitative Comparisons The evaluation was performed considering two aspects: (1) static consistency and (2) motion quality. Static consistency measures the consistency over time of non-movable regions such as buildings and mountains. Motion quality refers to both the image quality of the animated regions and the animation plausibility. For a fair comparison, we normalized the speed of motion in the generated results for all methods to match the average speed in the test dataset. AL, EMF, DL, and T2C were provided with the first frame  $I_0$  of the test video to generate  $512 \\times 512$  videos. Our method generates  $1024 \\times 1024$  results, thus we downsampled it to the size of  $512 \\times 512$ .",
+        "bbox": [
+            75,
+            90,
+            468,
+            271
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For a quantitative evaluation, we used 224 test videos from the Sky Time-Lapse dataset [41]. To measure the static consistency, we computed LPIPS [45], Root-Mean-Squared Error (RMSE), and MS-SSIM [40] between generated frame  $\\hat{I}_n$  and the input image  $I_0$  with the sky masked out. To measure the motion quality, we used the same evaluation metrics along with Frechet inception distance (FID) [28] between  $\\hat{I}_n$  and the ground-truth frame  $I_n$  with the static parts masked out.",
+        "bbox": [
+            75,
+            272,
+            467,
+            407
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1 reveals the quantitative evaluation results. For both static consistency and motion quality, our method outperforms the other approaches. Our cinematographs represent static consistency better because use of the mask improves the accuracy in the detection of static regions. In addition, the quality of texture details are improved in the generated frames due to GAN inversion and the MSDFW based on the deep features of StyleGAN. The use of our motion generator and DFW leads to improved motion quality for each scene, compared with that of the results of previous approaches.",
+        "bbox": [
+            75,
+            407,
+            467,
+            574
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "User Study We conducted a user study with 17 participants to subjectively evaluate our method in comparison with previous cinematograph generation methods. We did not target any specific demographic groups, as the evaluation of cinematographs should not be dependent on particular demographic distributions. We conducted two evaluations: score evaluation and side-by-side comparison. Both evaluations assessed static consistency and motion quality. In the score evaluation, participants rated the video presented with a reference image on a 1-to-5 scale, with \"very bad\" being 1 and \"very good\" being 5. In the side-by-side comparison, participants selected the preferred cinematograph from two videos generated using different methods, presented with a reference image. For both evaluations, we used ten samples randomly chosen from the LHQ [31] dataset. To eliminate bias, distinct samples were used for each evaluation, and the positions of the cinematographs were randomized in all questions.",
+        "bbox": [
+            75,
+            582,
+            467,
+            852
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Tables 2 and 3 summarize the statistics of the user studies. For both score evaluation and human preference, our method outperforms the previous approaches by a substan",
+        "bbox": [
+            76,
+            854,
+            467,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/f8057d5bc126095eab08959849b67219fc35cf69206b4c6e6da21bc3815500da.jpg",
+        "table_caption": [
+            "Table 2. Human perceptual study results for score evaluation. The best scores are bolded."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Static Consistency</td><td>Motion Quality</td></tr><tr><td>AL</td><td>1.59 ± 0.20</td><td>1.76 ± 0.45</td></tr><tr><td>EMF</td><td>2.35 ± 0.89</td><td>2.24 ± 0.99</td></tr><tr><td>DL</td><td>3.35 ± 0.83</td><td>3.25 ± 0.67</td></tr><tr><td>T2C</td><td>3.75 ± 0.69</td><td>2.96 ± 0.66</td></tr><tr><td>Ours</td><td>4.37 ± 0.18</td><td>3.86 ± 0.53</td></tr></table>",
+        "bbox": [
+            560,
+            119,
+            831,
+            202
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/86e87569558fbeb09657d161c9158875cc390d45ad0ca0494cf9e3109fc09b26.jpg",
+        "table_caption": [
+            "Table 3. Human perceptual study results for side-by-side comparison. The percentage of ours chosen against each competing method is reported."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">Human Preference (Ours %)</td></tr><tr><td>Static Consistency</td><td>Motion Quality</td></tr><tr><td>vs AL</td><td>99.35%</td><td>92.81%</td></tr><tr><td>vs EMF</td><td>96.08%</td><td>86.93%</td></tr><tr><td>vs DL</td><td>94.77%</td><td>84.97%</td></tr><tr><td>vs T2C</td><td>77.78%</td><td>73.86%</td></tr></table>",
+        "bbox": [
+            557,
+            261,
+            828,
+            343
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "tial margin. We also performed statistical analysis (Kruskal-Wallis H-test) on the resulting evaluation scores. The results showed that our method achieved significantly higher scores in all pairs of comparisons with every previous method  $(p < 0.001$  in post-hoc analysis). Both user study results reveal the advantages of our method for generating cinematographs with high image quality and plausible animation.",
+        "bbox": [
+            496,
+            354,
+            890,
+            460
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Comparisons with Video Generation Methods",
+        "text_level": 1,
+        "bbox": [
+            498,
+            472,
+            885,
+            488
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We compared our approach with MoCoGAN-HD [34], a video generation method that predicts the trajectory within the latent space of a pre-trained StyleGAN. For content generation, we used the same pre-trained StyleGAN as that used by MoCoGAN-HD, which was trained on the Sky Time-lapse [41] dataset to generate  $128 \\times 128$  images. The comparison was made in an unconditional manner using 300 randomly sampled latent codes, at a resolution of  $128 \\times 128$ . We compared the first 60 frames of the generated videos with those of MoCoGAN-HD, as our method produces looping videos.",
+        "bbox": [
+            496,
+            496,
+            890,
+            662
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Qualitative Comparisons Figure 5 presents a qualitative comparison. As shown in the first and third rows, the videos generated using our method exhibit static consistencies, with the clouds moving while the ground remains static. In contrast, as shown in the second and fourth rows, the frames generated by MoCoGAN-HD deviate significantly from the original image, making the method unsuitable for cinematograph generation.",
+        "bbox": [
+            496,
+            667,
+            890,
+            787
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Quantitative Comparisons We compared the content preservation ability of both methods by measuring LPIPS [45], RMSE, and MS-SSIM [40] between the first frame  $\\hat{I}_0$  and the subsequent frames  $\\hat{I}_n$ . Because the main content to be preserved in kinemagrams is the static regions, we defined content preservation as static consistency. We masked out the sky for all image frames according to the",
+        "bbox": [
+            496,
+            794,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "7877",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/6019b077d7cbd2e52e32a8f09f5c75e1aed65c3732c400e006c828ad8c966ef1.jpg",
+        "image_caption": [
+            "Figure 5. Qualitative comparison with the state-of-the-art video generation method, MoCoGAN-HD [34]. For more examples from this comparison, please refer to the supplementary video."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            106,
+            87,
+            436,
+            347
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/a794bf00fcc35105d17098df0a2401efecfbcc2e2b8133156a66ab6a93226d33.jpg",
+        "image_caption": [
+            "Figure 6. Quantitative comparison of content preservation with state-of-the-art video generation method MoCoGAN-HD [34]."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            96,
+            401,
+            447,
+            489
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "segmentation mask and compared only the static parts. We used 219 videos in which the scene could be horizontally divided into static and animated regions. LPIPS, RMSE, and MS-SSIM were all computed for each pair of  $\\hat{I}_0$  and  $\\hat{I}_n$ , and were averaged over the number of samples. Figure 6 shows that MoCoGAN-HD exhibits significant divergence in terms of LPIPS, RMSE, and MS-SSIM over time. This indicates that MoCoGAN-HD is unable to generate motion that preserves the content of the original image. In contrast, our method exhibits a small diverging trend, confirming its superiority in preserving the content of the original image over time, by utilizing deep features.",
+        "bbox": [
+            75,
+            532,
+            470,
+            715
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3. Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            76,
+            728,
+            230,
+            744
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To demonstrate the effectiveness of our design choices for the proposed method, we conducted a series of ablation studies. Specifically, we focused on the effectiveness of our warping and GAN inversion methods. To evaluate warping, we compared the image frames generated with and without the use of 1) forward warping, 2) DFW, 3) MSDFW, and 4) segmentation mask. The first frames of 224 test videos from the Sky Time-Lapse [41] dataset were given as input to generate  $1024 \\times 1024$  videos, and we used the first 60 frames for the evaluation. For the evaluation of GAN inversion, we",
+        "bbox": [
+            75,
+            750,
+            468,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/eb5b3cdaa79bcc779d32b3d97b9387f720c1d68cd27af3559c9ae022d5cdd0ab.jpg",
+        "image_caption": [
+            "Figure 7. Results of qualitative comparison in ablation study. Please see the supplementary video for the animated results."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            88,
+            692,
+            325
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/1ef88055d4ef5ca0a0e9822c91ff7031dd6c9f05db397b92441a12a59e9fd8f4.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            696,
+            89,
+            864,
+            325
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/a7f42b19cf715c06c7d4e413353f86ecd73d80170520fac7dc3477bb3184a3f1.jpg",
+        "image_caption": [
+            "Figure 8. Qualitative comparison results on GAN inversion."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            527,
+            362,
+            859,
+            474
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "compared images reconstructed with and without the use of the deep features. A total of 256 images, 128 from the Sky Time-Lapse and 128 from the Eulerian [9] dataset were provided to generate  $1024 \\times 1024$  reconstructed images.",
+        "bbox": [
+            496,
+            500,
+            890,
+            561
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Forward Warping We compared the image frames generated with and without forward warping (FW). For the case without FW, we warped the deep features of StyleGAN using backward warping. The first row in Figure 7 shows the qualitative results of this comparison. As revealed in the figure, without FW, the results usually contain unrealistically stretched textures. For quantitative comparisons, we computed LPIPS, MS-SSIM, and RMSE between  $I_{n}$  and  $\\hat{I}_n$ . Table 4 shows that excluding forward warping resulted in significant degradation in terms of MS-SSIM. This indicates that the stretched textures resulted in huge structural distortions in the generated images.",
+        "bbox": [
+            496,
+            571,
+            892,
+            755
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Deep Feature Warping We then compared frames generated with and without DFW. For the case without DFW, we directly warped RGB images using predicted motion fields for comparison. The second row in Figure 7 presents the qualitative comparison. As shown in the figure, the frame generated without DFW contained tearing artifacts in the animated regions. Table 4 shows an increase in the LPIPS score, which indicates that excluding DFW degraded the perceptual quality by introducing tearing artifacts.",
+        "bbox": [
+            496,
+            763,
+            890,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "7878",
+        "bbox": [
+            482,
+            944,
+            514,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/0f1926ecc6f88fc688d361c20efbe8fa56288085c64664b12f1383dad8e827db.jpg",
+        "table_caption": [
+            "Table 4. Results of quantitative evaluation in ablation study. The best scores are bolded."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>RMSE↓</td></tr><tr><td>Ours - Full</td><td>0.0511</td><td>0.7165</td><td>21.8188</td></tr><tr><td>Ours - w/o FW</td><td>0.0524</td><td>0.6853</td><td>22.9816</td></tr><tr><td>Ours - w/o DFW</td><td>0.0537</td><td>0.6908</td><td>22.4993</td></tr><tr><td>Ours - w/o MSDFW</td><td>0.0629</td><td>0.6946</td><td>22.3227</td></tr><tr><td>Ours - w/o Mask</td><td>0.0564</td><td>0.6980</td><td>22.6213</td></tr></table>",
+        "bbox": [
+            117,
+            121,
+            429,
+            204
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/1bb14de5cb3824060f668bd40cc7402e33027587a02326286e5d4f0fb2c41a93.jpg",
+        "table_caption": [
+            "Table 5. Quantitative evaluation results on GAN inversion. The best scores are bolded."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>w+</td><td>D10</td><td>FID↓</td><td>LPIPS↓</td><td>RMSE↓</td><td>MS-SSIM↑</td></tr><tr><td>✓</td><td>-</td><td>40.8276</td><td>0.0153</td><td>18.5901</td><td>0.7891</td></tr><tr><td>✓</td><td>✓</td><td>11.8231</td><td>0.0019</td><td>5.6410</td><td>0.9907</td></tr></table>",
+        "bbox": [
+            114,
+            243,
+            426,
+            282
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Multi-scale Deep Feature Warping Image frames generated with and without MSDFW were also compared. For the case without MSDFW, only a single deep feature  $D^{10}$  was warped and propagated through the next blocks of StyleGAN. The results of the qualitative comparison are shown in the third row of Figure 7. As revealed in the figure, excluding MSDFW resulted in blurry textures in the dynamic region. The quantitative comparison reported in Table 4 shows a significant increase in the LPIPS score. This indicates that excluding MSDFW degraded the perceptual quality, especially the texture details.",
+        "bbox": [
+            75,
+            295,
+            467,
+            460
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Segmentation Mask We compared the image frames generated with and without a segmentation mask. For the case without a mask, we used the initially predicted motion field for warping. The fourth row in Figure 7 shows the qualitative result of this comparison. As shown in the figure, excluding the mask resulted in erroneous movement in the static regions. The quantitative comparison reported in Table 4 reveals a significant degradation of MS-SSIM, which indicates the structural distortion caused by the erroneous motion.",
+        "bbox": [
+            75,
+            465,
+            467,
+            614
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Deep Feature Inversion To evaluate the effectiveness of GAN inversion, we compared the images reconstructed with and without the use of the deep features  $D^{10}$ . For the case without deep feature inversion, only the latent codes  $w^{+}$  of a pre-trained StyleGAN were used to reconstruct the input landscape image. Figure 8 shows a qualitative result of this comparison. As revealed in the figure, using only  $w^{+}$  failed to accurately reconstruct the details of the original images. For quantitative comparisons, FID, LPIPS, RMSE, and MS-SSIM were computed between  $I$  and  $\\hat{I}$ . Table 5 reveals significant improvements in the perceptual quality of the results when reconstructed using the deep features  $D^{10}$ .",
+        "bbox": [
+            75,
+            621,
+            467,
+            801
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Limitations and Future Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            818,
+            349,
+            834
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our mask predictor and motion generator generally performed well for landscape images. However, automatic prediction cannot be accurate for all images because most landscape images contain inherent ambiguity in their motion",
+        "bbox": [
+            75,
+            839,
+            467,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/0a3178b499e93c16913a8a8349e6d82741ef8a510389777e8ab9dc7d67285d5f.jpg",
+        "image_caption": [
+            "(a)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            535,
+            90,
+            692,
+            151
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/99266baccdda3e0ba47175c59d0df00a990c5ee099f3af6ed316f18498209f98.jpg",
+        "image_caption": [
+            "(b)",
+            "Figure 9. Limitations of StyleCineGAN: (a) automatic prediction of motion cannot be accurate for all images, and (b) the motion of a very thin structured object is hard to be isolated."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            718,
+            90,
+            795,
+            150
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/66837910d83aaf923e84f8e453c57f1e4b218fed97a2f2fc7b454e387d80de0e.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            797,
+            90,
+            872,
+            150
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "directions except for some obvious cases (e.g., waterfalls). The failure case is illustrated in Figure 9 (a). User-defined motion hints can be used to resolve such ambiguities and provide further control capability during the generation process [18, 19]. In addition, it is hard for our method to isolate the motion of a very thin structured object placed within the animated region as shown in Figure 9 (b). This is because our method performs warping of features at multiple resolutions, in which low-resolution features cannot spatially represent the thin structures.",
+        "bbox": [
+            496,
+            205,
+            890,
+            354
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this work, we mainly focused on animating the landscape images, particularly skies and fluids, while putting other types of animations outside the scope. In future developments, we would like to expand the capabilities to include other forms of motion. Investigating the rotating hands of a clock, the playing arm of a guitarist moving up and down, and a flag or the wings of a bird fluttering for cinematograph generation would be a very interesting direction to pursue.",
+        "bbox": [
+            496,
+            356,
+            890,
+            478
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            489,
+            617,
+            505
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We proposed the first approach that leverages a pre-trained StyleGAN for high-quality one-shot landscape cinematograph generation. In contrast to previous studies, our method does not require training a large image generator from scratch, and also systematically improves the resolution of the generated cinematographs to  $1024 \\times 1024$ . At the core of our method, we utilized the deep features of a pretrained StyleGAN, because those features can help preserve spatial information and encode both high-level semantic and low-level style appearances. Using our MSDFW approach, we applied the predicted motion to the deep feature space of StyleGAN. Both qualitative and quantitative results confirm that our method substantially outperforms existing baselines.",
+        "bbox": [
+            496,
+            512,
+            890,
+            723
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: Development of Universal Fashion Creation Platform Technology for Avatar Personality Expression, Project Number: RS-2023-00228331, Contribution Rate:  $90\\%$  ) and Korea Institute for Advancement of Technology(KIAT) grant funded by the Korea Government(MOTIE) (Project Name: The Competency Development Program for Industry Specialist, Project Number: P0012746, Contribution Rate:  $10\\%$ ).",
+        "bbox": [
+            496,
+            734,
+            890,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "7879",
+        "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] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 3",
+            "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? CoRR, abs/1911.11544, 2019.3",
+            "[3] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. arXiv preprint arXiv:2201.13433, 2022. 3",
+            "[4] Hugo Bertiche, Niloy J. Mitra, Kuldeep Kulkarni, ChunHao Paul Huang, Tuanfeng Y. Wang, Meysam Madadi, Sergio Escalera, and Duygu Ceylan. Blowing in the wind: Cyclenet for human cinematographs from still images, 2023. 1, 2",
+            "[5] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. In ACM SIGGRAPH 2005 Papers, pages 853-860. 2005. 2",
+            "[6] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM Transactions on Graphics (Proc. of SIGGRAPH ASIA 2019), 38(6):175:1-175:19, 2019. 1, 2, 5",
+            "[7] Gereon Fox, Ayush Tewari, Mohamed Elgharib, and Christian Theobalt. Stylevideogan: A temporal generative model using a pretrained stylegan, 2021. 2",
+            "[8] Tavi Halperin, Hanit Hakim, Orestis Vantzos, Gershon Hochman, Netai Benaim, Lior Sassy, Michael Kupchik, Ofir Bibi, and Ohad Fried. Endless loops: detecting and animating periodic patterns in still images. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 2",
+            "[9] Aleksander Holynski, Brian L. Curless, Steven M. Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5810-5819, 2021. 1, 2, 3, 4, 5, 7",
+            "[10] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the \"steerability\" of generative adversarial networks. In International Conference on Learning Representations, 2020. 2",
+            "[11] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 3",
+            "[12] 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. 2",
+            "[13] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 3",
+            "[14] Alexander Kristoffersen. Loopnerf: Exploring temporal compression for 3d video textures. Master's thesis, EECS Department, University of California, Berkeley, 2023. 1, 2",
+            "[15] Xingyi Li, Zhiguo Cao, Huiqiang Sun, Jianming Zhang, Ke Xian, and Guosheng Lin. 3d cinematography from a single"
+        ],
+        "bbox": [
+            78,
+            114,
+            468,
+            901
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4595-4605, 2023. 1, 2",
+            "[16] Ji Lin, Richard Zhang, Frieder Ganz, Song Han, and JunYan Zhu. Anycost gans for interactive image synthesis and editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3",
+            "[17] Elizaveta Logacheva, Roman Suvorov, Oleg Khomenko, Anton Mashikhin, and Victor Lempitsky. Deeplandscape: Adversarial modeling of landscape videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 5",
+            "[18] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 8",
+            "[19] Aniruddha Mahapatra, Aliaksandr Siarohin, Hsin-Ying Lee, Sergey Tulyakov, and Jun-Yan Zhu. Text-guided synthesis of eulerian cinemagrams. 2023. 1, 2, 3, 5, 8",
+            "[20] Makoto Okabe, Ken Anjyo, Takeo Igarashi, and Hans-Peter Seidel. Animating pictures of fluid using video examples. In Computer Graphics Forum, pages 677-686. Wiley Online Library, 2009. 2",
+            "[21] Makoto Okabe, Ken Anjyor, and Rikio Onai. Creating fluid animation from a single image using video database. In Computer Graphics Forum, pages 1973-1982. Wiley Online Library, 2011.",
+            "[22] Makoto Okabe, Yoshinori Dobashi, and Ken Anjyo. Animating pictures of water scenes using video retrieval. The Visual Computer, 34(3):347-358, 2018. 2",
+            "[23] Ehsan Pajouheshgar, Yitao Xu, Tong Zhang, and Sabine Süssstrunk. Dynca: Real-time dynamic texture synthesis using neural cellular automata. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1",
+            "[24] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, Jun-Yan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3",
+            "[25] Ekta Prashnani, Maneli Noorkami, Daniel Vaquero, and Pradeep Sen. A phase-based approach for animating images using video examples. In Computer Graphics Forum, pages 303-311. Wiley Online Library, 2017. 1, 2",
+            "[26] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3",
+            "[27] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan, 2020. 2",
+            "[28] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6"
+        ],
+        "bbox": [
+            503,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "7880",
+        "bbox": [
+            482,
+            945,
+            514,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[29] Kwanggyoon Seo, Seoung Wug Oh, Jingwan Lu, Joon-Young Lee, Seonghyeon Kim, and Junyong Noh. Styleportraitvideo: Editing portrait videos with expression optimization. 41(7), 2022. 3",
+            "[30] Claude E. Shannon. Coding Theorems for a Discrete Source With a Fidelity Criterion Institute of Radio Engineers, International Convention Record, vol. 7, 1959., pages 325-350. 1993. 2",
+            "[31] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. arXiv preprint arXiv:2104.06954, 2021. 6",
+            "[32] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V. Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 2",
+            "[33] Matthew Tesfaldet, Marcus A. Brubaker, and Konstantinos G. Derpanis. Two-stream convolutional networks for dynamic texture synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1",
+            "[34] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. In International Conference on Learning Representations, 2021. 2, 5, 6, 7",
+            "[35] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 3",
+            "[36] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1526-1535, 2018. 2",
+            "[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 3",
+            "[38] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2",
+            "[39] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3",
+            "[40] Zhou Wang, Eero P. Simoncelli, and Alan Conrad Bovik. Multiscale structural similarity for image quality assessment. The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, 2:1398-1402 Vol.2, 2003. 6",
+            "[41] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6, 7"
+        ],
+        "bbox": [
+            78,
+            90,
+            470,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[42] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. pages 357-374. Springer, 2022. 3",
+            "[43] Xu Yao, Alasdair Newson, Yann Gousseau, and Pierre Hellier. A style-based gan encoder for high fidelity reconstruction of images and videos. European conference on computer vision, 2022. 3",
+            "[44] 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. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 85-101. Springer, 2022. 3",
+            "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6",
+            "[46] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            415
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "7881",
+        "bbox": [
+            482,
+            945,
+            513,
+            955
+        ],
+        "page_idx": 9
+    }
+]

2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_model.json

@@ -0,0 +1,2200 @@
+[
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.107,
+                0.003,
+                0.182,
+                0.043
+            ],
+            "angle": 0,
+            "content": "CVF"
+        },
+        {
+            "type": "header",
+            "bbox": [
+                0.237,
+                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": [
+                0.217,
+                0.131,
+                0.756,
+                0.177
+            ],
+            "angle": 0,
+            "content": "StyleCineGAN: Landscape Cinemagraph Generation using a Pre-trained StyleGAN"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.16,
+                0.204,
+                0.28,
+                0.222
+            ],
+            "angle": 0,
+            "content": "Jongwoo Choi"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.319,
+                0.204,
+                0.463,
+                0.222
+            ],
+            "angle": 0,
+            "content": "Kwanggyoon Seo"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.502,
+                0.204,
+                0.661,
+                0.222
+            ],
+            "angle": 0,
+            "content": "Amirsaman Ashtari"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.7,
+                0.204,
+                0.808,
+                0.222
+            ],
+            "angle": 0,
+            "content": "Junyong Noh"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.38,
+                0.223,
+                0.591,
+                0.239
+            ],
+            "angle": 0,
+            "content": "Visual Media Lab, KAIST"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.236,
+                0.242,
+                0.73,
+                0.258
+            ],
+            "angle": 0,
+            "content": "{jeolpyeoni0,skg1023,a.s.ashtari,junyongnoh}@kaist.ac.kr"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.467,
+                0.895,
+                0.497
+            ],
+            "angle": 0,
+            "content": "Figure 1. Given a landscape image, StyleCineGAN generates a seamless cinematograph at \\(1024 \\times 1024\\) resolution. This figure contains video clips, thus consider viewing it using Adobe Reader. The same results are also included in the supplementary video."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.235,
+                0.509,
+                0.314,
+                0.525
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.529,
+                0.473,
+                0.786
+            ],
+            "angle": 0,
+            "content": "We propose a method that can generate cinematographs automatically from a still landscape image using a pretrained StyleGAN. Inspired by the success of recent unconditional video generation, we leverage a powerful pretrained image generator to synthesize high-quality cinematographs. Unlike previous approaches that mainly utilize the latent space of a pre-trained StyleGAN, our approach utilizes its deep feature space for both GAN inversion and cinematograph generation. Specifically, we propose multi-scale deep feature warping (MSDFW), which warps the intermediate features of a pre-trained StyleGAN at different resolutions. By using MSDFW, the generated cinematographs are of high resolution and exhibit plausible looping animation. We demonstrate the superiority of our method through user studies and quantitative comparisons with state-of-the-art cinematograph generation methods and a video generation method that uses a pre-trained StyleGAN."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.802,
+                0.21,
+                0.818
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.825,
+                0.471,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Cinemagraph is a unique form of media that combines a still image and video. While most of the scene remains still, subtle and repeated movements in only a small region effectively highlight a particular object or an important event that occurred at the time of capture. Because this mixture"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.51,
+                0.892,
+                0.556
+            ],
+            "angle": 0,
+            "content": "of still and moving elements in one scene creates an interesting eye-catching effect, cinematographs have recently attracted increasing attention in social media."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.562,
+                0.893,
+                0.699
+            ],
+            "angle": 0,
+            "content": "Despite the growing popularity of cinematograph, its creation heavily relies on a manual process. To create a motion in part of the input image, one typically utilizes an image manipulation tool to stretch, scale, rotate, or shift the part of the image in a physically plausible and aesthetically pleasing way. This process is often time-consuming and requires a high level of skill in the use of image editing tools. Therefore, creating a cinematograph has generally been considered to be a personal project of professionals so far."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.704,
+                0.895,
+                0.902
+            ],
+            "angle": 0,
+            "content": "To allow ordinary users to create a cinematograph with a single image, automatic methods have been proposed. One line of research uses a reference video as guidance [14, 25, 33]. These methods are capable of producing realistic motions in the target scene similar to a reference video. Recent methods obviate the need for a reference video, by training deep generative models [4, 6, 9, 15, 17-19, 23]. These methods decompose the task into two processes of learning motion and spatial information individually from separate datasets. The decomposition effectively reduces the complexity of simultaneously learning both temporal and spatial information, and improves the perceptual quality of the generated videos. However, these models"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "7872"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.092,
+                0.468,
+                0.167
+            ],
+            "angle": 0,
+            "content": "must be trained from scratch, which sometimes requires several days if not weeks on a modern GPU. Moreover, these models are not specifically designed to generate high-resolution of \\(1024 \\times 1024\\) cinemagrams, because of the significant requirements in memory and processing power."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.167,
+                0.468,
+                0.317
+            ],
+            "angle": 0,
+            "content": "In this paper, we propose the first approach to high-quality one-shot landscape cinematograph generation based on a pre-trained StyleGAN [12]. By using the learned image prior of StyleGAN, our method removes the need for training a large model from scratch and systematically improves the resolution of the generated cinematographs to \\(1024 \\times 1024\\). Moreover, our method enables a user to easily edit the style and appearance of the generated cinematograph by leveraging the properties of StyleGAN for image stylization and editing."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.317,
+                0.468,
+                0.543
+            ],
+            "angle": 0,
+            "content": "Our method is inspired by recent unconditional video generation methods [7, 34] which allow to navigate the latent space of a pre-trained image generator, to synthesize a high-quality temporally coherent video. Unlike these methods that utilize the latent codes of StyleGAN, we opt to use the deep features that are generated by convolution operations in each layer of StyleGAN. We use these deep features for two reasons. First, we observed that highly detailed landscape images cannot be reconstructed accurately from the latent codes using GAN inversion methods because these latent codes are low-dimensional [30]. Second, a plausible motion that preserves the content cannot be created by only navigating the latent space, because this space is highly semantic-condensed and lacks explicit spatial prior [38]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.543,
+                0.468,
+                0.678
+            ],
+            "angle": 0,
+            "content": "To solve these problems, for the first time in the cinematograph generation domain, we utilize the deep features of a pre-trained StyleGAN. These deep features preserve spatial information and encode both high-level semantic and low-level style appearances across high and low-level convolutional layers. To produce cinematographs from these deep features, we propose a multi-scale deep feature warping (MSDFW) to apply the motions generated using a motion generator to the deep feature space of StyleGAN."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.679,
+                0.468,
+                0.813
+            ],
+            "angle": 0,
+            "content": "We demonstrate the effectiveness and advantages of our method by comparing it to state-of-the-art methods in cinematagraph generation [6, 9, 17, 19] and to an unconditional video generation method that uses a pre-trained StyleGAN [34], using various metrics. We also performed a user study to verify the effectiveness of our method for creating visually pleasing cinematagraph effects. Both qualitative and quantitative results confirm that our method substantially outperforms all existing baselines."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.825,
+                0.219,
+                0.84
+            ],
+            "angle": 0,
+            "content": "2. Related Work"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.85,
+                0.311,
+                0.866
+            ],
+            "angle": 0,
+            "content": "2.1. Cinemagraph Generation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.871,
+                0.468,
+                0.901
+            ],
+            "angle": 0,
+            "content": "One of the early studies for cinemagraph generation used a procedural approach to decompose each object into layers"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.092,
+                0.892,
+                0.272
+            ],
+            "angle": 0,
+            "content": "and apply time-varying stochastic motion [5]. This method can handle a diverse range of images such as photos and paintings by relying on manual interaction from the user. Another approach is to use a reference video to animate the given image [20-22, 25]. These methods rely on a statistical motion analysis of videos to transfer their periodic motion property to the desired image. Another method distills a dynamic NeRF to render looping 3D video textures from the representative motion found in the reference video [14]. Without any reference video as guidance, Halperin et al. [8] proposed a framework to animate arbitrary objects with periodic patterns in a simple motion direction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.273,
+                0.892,
+                0.529
+            ],
+            "angle": 0,
+            "content": "Recent approaches use the capacity of deep learning to automatically create cinemagrams from a single image. One study trained an image-based renderer to generate water animation, utilizing the water simulation results [32]. Another research first predicts a sequence of normal maps, then generates the corresponding RGB images that show a looping animation of a garment as if it is blown in the wind [4]. Another line of research trains a generator to produce a motion field [6, 9, 15, 18, 19]. Our work is similar to these deep learning-based approaches in that a motion generator is utilized to synthesize the cinemagrams. The difference is that our method leverages the deep features of a pre-trained image generator and thus can generate plausible high-quality results without fine-tuning or training a separate synthesis network. At the same time, our method systematically improves the resolution of the generated cinemagrams to \\(1024 \\times 1024\\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.534,
+                0.785,
+                0.549
+            ],
+            "angle": 0,
+            "content": "2.2. Unconditional Video Generation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.553,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "The task of video generation is known to be difficult because the generation process has to take into account both spatial and temporal information. To ease the problem, previous approaches [27, 36] decompose video generation into content and motion generation. These methods first predict the latent trajectory for motion, then generate a video from the set of predicted latent codes using the image generator. Instead of training the generation model from scratch, MoCoGAN-HD [34] and StyleVideoGAN [7] leverage a pre-trained image generator model, StyleGAN [12]. These methods use the property of well-constructed latent space to morph one frame toward the next frames [10]. Utilizing the image generation capability of a pre-trained generator, MoCoGAN-HD and StyleVideoGAN only need to train a motion trajectory generator, which greatly reduces the training time. These methods can generate high-quality videos in multiple domains such as faces, cars, and even sky. However, the content details are not well preserved, which has hindered the methods' application for cinematograph generation. Our method builds upon a similar concept of using a pre-trained StyleGAN. We leverage the deep features instead of the latent codes of StyleGAN to generate looping and content-preserving cinematographs."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "7873"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.091,
+                0.455,
+                0.108
+            ],
+            "angle": 0,
+            "content": "2.3. Video Synthesis using pre-trained StyleGAN"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.115,
+                0.473,
+                0.328
+            ],
+            "angle": 0,
+            "content": "A pre-trained generative model, StyleGAN [11-13], has been actively employed for downstream image and video editing applications. For GAN-based applications, one of the essential steps is to project the desired images or videos to the latent space of the pre-trained GAN model, which is known as GAN inversion. The latent codes are obtained using either an optimization technique [1, 2], a trained encoder [26, 35], or a hybrid of both approaches [16]. For video editing applications, previous methods [3, 29, 37, 42] have utilized a pre-trained image encoder to project all of the frames to the latent space. While most of the methods operate in \\(\\mathcal{W}+\\) space [2], we opt to use both \\(w^{+}\\) and deep features [24, 43, 44] of the pre-trained StyleGAN to accurately reconstruct the original image and synthesize motion."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.344,
+                0.178,
+                0.36
+            ],
+            "angle": 0,
+            "content": "3. Methods"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.37,
+                0.47,
+                0.461
+            ],
+            "angle": 0,
+            "content": "An overview of our method is shown in Figure 2. Given a landscape image \\(I\\), our method generates a seamlessly looping cinemagraph \\(V = \\{\\hat{I}_0, \\dots, \\hat{I}_N\\}\\) using a pre-trained StyleGAN. At the core of our method, we use a pre-trained generator without additional fine-tuning. The overall process is described below."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.061,
+                0.466,
+                0.47,
+                0.526
+            ],
+            "angle": 0,
+            "content": "(A) First, we project the landscape image into both the latent space and the feature space of StyleGAN. To this end, we train an encoder that outputs both latent codes \\( w^{+} \\) and intermediate features \\( D^{10} \\) of StyleGAN (Sec. 3.1)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.061,
+                0.527,
+                0.469,
+                0.556
+            ],
+            "angle": 0,
+            "content": "(B) In addition, we predict a mask \\( S \\) to divide the image into static and dynamic regions (Sec. 3.2)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.062,
+                0.557,
+                0.469,
+                0.616
+            ],
+            "angle": 0,
+            "content": "(C) Next, we use a motion generator that accepts the landscape image \\(I\\) to synthesize a motion field \\(M\\), which defines the position of each pixel in the future frames (Sec. 3.3)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.061,
+                0.618,
+                0.469,
+                0.676
+            ],
+            "angle": 0,
+            "content": "(D) Lastly, we generate the final cinematograph frames using the pre-trained StyleGAN with deep feature warping (DFW) operation added in between the original layers. (Sec. 3.4)."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.061,
+                0.466,
+                0.47,
+                0.676
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.682,
+                0.47,
+                0.713
+            ],
+            "angle": 0,
+            "content": "In the following sections, we will describe the details of each process."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.726,
+                0.231,
+                0.74
+            ],
+            "angle": 0,
+            "content": "3.1. GAN Inversion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.75,
+                0.473,
+                0.902
+            ],
+            "angle": 0,
+            "content": "The first step for generating a cinematograph is to project the input image to the latent space of a pre-trained StyleGAN. This process is necessary because StyleGAN is an unconditional generator that cannot take a conditioning input. Many previous methods have used \\(\\mathcal{W} +\\) [2] latent space, which is an extended space of native latent space \\(\\mathcal{W}\\). However, we observed that \\(w^{+} \\in \\mathcal{W}+\\) is not expressive enough to reconstruct the original high-frequency details of the landscape images. Therefore, we chose to use the generated deep features \\(D^{10}\\) of StyleGAN as well as \\(w^{+}\\). The use of \\(D^{10}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.092,
+                0.892,
+                0.122
+            ],
+            "angle": 0,
+            "content": "enables us to recover details in the original input as shown in Figure 8."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.122,
+                0.894,
+                0.199
+            ],
+            "angle": 0,
+            "content": "To project an image \\(I\\) to both the latent space and feature space and obtain \\(w^{+}\\) and \\(D^{10}\\), we train an encoder \\(E\\) similarly to Yao et al. [43], in which the encoder takes \\(I\\) as input and predicts \\((w^{+}, D^{10})\\). Additional training details can be found in the supplementary material."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.209,
+                0.665,
+                0.223
+            ],
+            "angle": 0,
+            "content": "3.2. Mask Prediction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.231,
+                0.892,
+                0.352
+            ],
+            "angle": 0,
+            "content": "We improve the quality of GAN inversion described in Sec. 3.1 by training an additional classifier that predicts a mask to separate the static and dynamic regions. Using the mask, the structure of the static regions can be preserved. To this end, we train a multi-layer-perceptron (MLP) classifier that accepts the deep features of StyleGAN as its input and outputs the mask that specifies each dynamic region in the image, as performed in DatasetGAN [46]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.354,
+                0.893,
+                0.52
+            ],
+            "angle": 0,
+            "content": "To train the classifier, we manually annotate 32 segmentation masks \\( S \\) of the selected images. We then follow the same procedure as DatasetGAN but using the deep features \\( D^{i} \\) where \\( i \\in \\{10,11,\\dots,18\\} \\). We resize all these deep features followed by concatenating them in the channel dimension to construct the input feature \\( D^{*} \\). In the end, paired data \\( (D^{*},S) \\) is constructed. An ensemble of 10 MLP classifiers is trained with a cross-entropy loss. To further improve the performance of the previous work, at inference, we also refine the predicted mask. For additional details, please refer to the supplementary material."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.53,
+                0.685,
+                0.544
+            ],
+            "angle": 0,
+            "content": "3.3. Motion Generation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.553,
+                0.892,
+                0.674
+            ],
+            "angle": 0,
+            "content": "To generate motion from a static landscape image \\( I \\), we use an Eulerian motion field \\( M \\) which represents the immediate velocity of each pixel. To predict \\( M \\) given \\( I \\), an imaged-to-image translation network [39] is trained as a motion generator. We train the network with paired data \\( (I, M) \\) as performed in the previous method [9]. Adding controls in motion generation is also possible, either by using text prompts [19] or drawing arrows [18]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.675,
+                0.892,
+                0.795
+            ],
+            "angle": 0,
+            "content": "While the trained network works well to a certain degree, we observed that the predicted motion field \\(M\\) does not align with the boundaries between the static and dynamic regions and often contains motion in static regions. Thus, we further refine \\(M\\) by multiplying it with the segmentation mask \\(S\\), which effectively removes most errors. The refined \\(M\\) is then used to simulate the motion of a pixel at time \\(t\\) using the Euler integration as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.512,
+                0.808,
+                0.892,
+                0.844
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} x _ {t + 1} = x _ {t} + M \\left(x _ {t}\\right), \\quad M \\left(x _ {t}\\right) = F _ {t \\rightarrow t + 1} \\left(x _ {t}\\right), \\\\ F _ {0 \\rightarrow t} (x _ {0}) = F _ {0 \\rightarrow t - 1} (x _ {0}) + M \\left(x _ {0} + F _ {0 \\rightarrow t - 1} (x _ {0})\\right), \\tag {1} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.856,
+                0.892,
+                0.902
+            ],
+            "angle": 0,
+            "content": "where \\( x_0 \\) is the source pixel position at time 0, \\( F_{0\\rightarrow t} \\) is the displacement field that defines the displacement of the pixel position \\( x_0 \\) to the pixel position \\( x_t \\) at time \\( t \\)."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "7874"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.082,
+                0.09,
+                0.526,
+                0.312
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.298,
+                0.316,
+                0.387,
+                0.326
+            ],
+            "angle": 0,
+            "content": "Motion Generation"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.532,
+                0.09,
+                0.891,
+                0.31
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.547,
+                0.316,
+                0.666,
+                0.327
+            ],
+            "angle": 0,
+            "content": "Cinemagraph Generation"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.33,
+                0.892,
+                0.44
+            ],
+            "angle": 0,
+            "content": "Figure 2. Overview of StyleCineGAN. Given an input landscape image \\( I \\), our goal is to generate a cienemagraph using a fixed pre-trained StyleGAN \\( G \\). We project the image into both latent codes \\( w^{+} \\) and deep features \\( D^{10} \\) of \\( G \\). Using the deep features \\( D^{*} \\), a mask predictor predicts a segmentation mask \\( S \\). To animate the input image, we use a motion generator to predict the motion field \\( M \\) from \\( I \\). \\( M \\) is refined using \\( S \\). Through Euler integration, \\( M \\) produces the future and past displacement fields \\( F_{0\\rightarrow t} \\) and \\( F_{N\\rightarrow t} \\). To synthesize cienemagraph frames, we add a DFW layer in between the layers of \\( G \\). DFW refers to Eqns. 2 and 3. This modification enables the intermediate features of \\( G \\) to be warped according to \\( F_{0\\rightarrow t} \\) and \\( F_{N\\rightarrow t} \\) using a joint splatting method at different resolutions, specifically for the StyleGAN layers indexed with \\( i\\in [10,12,14,16,18] \\). The warped deep features are used to synthesize frames \\( \\hat{I}_t \\) resulting in the final cienemagraph video."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.453,
+                0.312,
+                0.469
+            ],
+            "angle": 0,
+            "content": "3.4. Cinemagraph Generation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.477,
+                0.47,
+                0.611
+            ],
+            "angle": 0,
+            "content": "After acquiring the displacement field \\( F_{0\\rightarrow t} \\) (Sec. 3.3), latent codes \\( w^{+} \\), and deep features \\( D^{10} \\) of StyleGAN (Sec. 3.1), we feed these data into the pre-trained StyleGAN with our DFW layer added. We apply forward warping (or splatting) to the original content to generate a cinematograph. Here, the application of forward warping directly to an RGB image often results in images with tearing artifacts. To reduce the artifacts, we warp the deep features \\( D_0^i \\) of StyleGAN in different resolutions or scales:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.188,
+                0.618,
+                0.469,
+                0.637
+            ],
+            "angle": 0,
+            "content": "\\[\nD _ {t} ^ {i} = \\operatorname {W a r p} \\left(D _ {0} ^ {i}, F _ {0 \\rightarrow t}\\right), \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.644,
+                0.469,
+                0.81
+            ],
+            "angle": 0,
+            "content": "where \\( \\text{Warp} \\) is the forward warping function and \\( D_{t}^{i} \\) is the warped deep feature at time \\( t \\) and scale \\( i \\). We observed that warping a single deep feature (e.g., only \\( D_{0}^{10} \\)) results in blurry textures in the generated cinemagrams. Therefore, we opt to warp the multi-scale deep features; we call this operation a MSDFW. Specifically, we apply warping to the deep features \\( D_{0}^{i} \\) where \\( i \\in [10, 12, 14, 16, 18] \\). The deep feature \\( D_{0}^{10} \\) is acquired using GAN inversion, and the consequent deep features \\( D_{0}^{12}, D_{0}^{14}, \\ldots \\), and \\( D_{0}^{18} \\) are subsequently generated using both \\( D_{0}^{10} \\) and \\( w^{+} \\), as shown in the rightmost column of Figure 2."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.811,
+                0.469,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Looping video \\(V\\) with frame length \\(N + 1\\) is synthesized using the joint splating technique [9] using the future and past displacement fields \\(F_{0\\rightarrow t}\\) and \\(F_{N\\rightarrow t}\\), generated through Euler integration. The displacement fields are computed with \\(t\\) times of integration on \\(M\\) for \\(F_{0\\rightarrow t}\\), and \\(N - t\\) times of integration on \\(-M\\) for \\(F_{N\\rightarrow t}\\). The multi-scale fea"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.453,
+                0.892,
+                0.501
+            ],
+            "angle": 0,
+            "content": "ture \\(D_0^i\\) is warped in both directions and is composited to form \\(D_t^i\\). Specifically, \\(D_t^i\\) is computed as a weighted sum of \\(Warp(D_0^i,F_{0\\to t})\\) and \\(Warp(D_0^i,F_{N\\to t})\\) as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.516,
+                0.514,
+                0.891,
+                0.592
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} D _ {t} ^ {i} \\left(x ^ {\\prime}\\right) = \\frac {\\sum_ {x \\in \\chi} \\alpha_ {t} \\cdot W a r p \\left(D _ {0} ^ {i} (x) , F _ {0 \\rightarrow t}\\right)}{\\sum_ {x \\in \\chi} \\alpha_ {t}} \\\\ + \\frac {\\sum_ {x \\in \\chi} \\left(1 - \\alpha_ {t}\\right) \\cdot W a r p \\left(D _ {0} ^ {i} (x) , F _ {N \\rightarrow t}\\right)}{\\sum_ {x \\in \\chi} \\left(1 - \\alpha_ {t}\\right)}, \\tag {3} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.603,
+                0.891,
+                0.649
+            ],
+            "angle": 0,
+            "content": "where \\( x \\in \\chi \\) is a set of pixels being mapped to the same destination pixel \\( x' \\), and \\( \\alpha_{t} \\) is the looping weight defined as \\( (1 - \\frac{t}{N}) \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.649,
+                0.891,
+                0.692
+            ],
+            "angle": 0,
+            "content": "With the above DFW module, we can generate a video frame given the predicted deep features \\(D^{10}\\), latent code \\(w^{+}\\), and mask \\(S\\):"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.515,
+                0.701,
+                0.891,
+                0.737
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\hat {I} _ {t} = S \\odot \\left(G _ {\\text {w a r p}} \\left(D _ {0} ^ {1 0}, w ^ {+}, F _ {0 \\rightarrow t}, F _ {N \\rightarrow t}\\right)\\right) \\\\ + (\\mathbf {1} - S) \\odot I, \\tag {4} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.499,
+                0.75,
+                0.891,
+                0.793
+            ],
+            "angle": 0,
+            "content": "where \\(\\odot\\) is an element-wise multiplication and \\(G_{\\text{warp}}\\) is a fixed pre-trained StyleGAN that incorporates our DFW module."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.798,
+                0.891,
+                0.871
+            ],
+            "angle": 0,
+            "content": "Style Interpolation In addition to motion generation, the change of appearance style is an additional feature often observed in a landscape cinematograph. This can be achieved by interpolating the latent code with the target latent code as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.594,
+                0.885,
+                0.891,
+                0.902
+            ],
+            "angle": 0,
+            "content": "\\[\nw _ {s} ^ {+} = w ^ {+} \\cdot (1 - \\beta) + w _ {t} ^ {+} \\cdot \\beta , \\tag {5}\n\\]"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.515,
+                0.957
+            ],
+            "angle": 0,
+            "content": "7875"
+        }
+    ],
+    [
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.077,
+                0.17,
+                0.473,
+                0.228
+            ],
+            "angle": 0,
+            "content": "Figure 3. Generated cinematagraph results. This figure contains video clips, thus consider viewing it using Adobe Reader. The first two are cinematographs without appearance change, and the last two are cinematographs with appearance change."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.237,
+                0.47,
+                0.313
+            ],
+            "angle": 0,
+            "content": "where \\( w_{s}^{+} \\) is the interpolated latent code, \\( w_{t}^{+} \\) is the latent code of the target style, and \\( \\beta \\) is the interpolation weight. This is possible because the later layers of the StyleGAN only modify the color of the synthesized image while the original structure is preserved."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.314,
+                0.47,
+                0.36
+            ],
+            "angle": 0,
+            "content": "With modified Equation 4, we prevent the visual mismatch between the static and dynamic regions by also reflecting the changes to the static regions as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.13,
+                0.372,
+                0.47,
+                0.41
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\hat {I} _ {t} = S \\odot \\left(G _ {\\text {w a r p}} \\left(D _ {0} ^ {1 0}, w _ {s} ^ {+}, F _ {0 \\rightarrow t}, F _ {N \\rightarrow t}\\right)\\right) \\tag {6} \\\\ + (\\mathbf {1} - S) \\odot (I + \\Delta I _ {s}), \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.165,
+                0.412,
+                0.469,
+                0.431
+            ],
+            "angle": 0,
+            "content": "\\[\n\\Delta I _ {s} = G \\left(D _ {0} ^ {1 0}, w _ {s} ^ {+}\\right) - G \\left(D _ {0} ^ {1 0}, w ^ {+}\\right), \\tag {7}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.444,
+                0.47,
+                0.475
+            ],
+            "angle": 0,
+            "content": "where \\(\\Delta I_{s}\\) is the color difference between the two images generated from latent codes \\(w_{s}^{+}\\) and \\(w^{+}\\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.491,
+                0.21,
+                0.508
+            ],
+            "angle": 0,
+            "content": "4. Experiments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.517,
+                0.47,
+                0.639
+            ],
+            "angle": 0,
+            "content": "In this section, we compare our method with state-of-the-art landscape cinematograph generation methods [6, 9, 17, 19] (Sec. 4.1) and with an unconditional video generation method [34] (Sec. 4.2). In addition, we show the importance of the components used in our method through ablation studies (Sec. 4.3). To observe more various results, we recommend readers see the supplementary video. Example results are presented in Figures 1 and 3."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.65,
+                0.47,
+                0.68
+            ],
+            "angle": 0,
+            "content": "4.1. Comparisons with Cinemagraph Generation Methods"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.69,
+                0.471,
+                0.902
+            ],
+            "angle": 0,
+            "content": "We compared our method with state-of-the-art cinemagraph generation methods: Animating Landscape (AL) [6], Deep Landscape (DL) [17], Eulerian Motion Field (EMF) [9], and Text2Cinemagraph (T2C) [19]. AL is a learning-based method that uses backward warping to generate cinemagraph frames. DL trains a style-based generator to synthesize images and dynamic motion in a scene. EMF and T2C train an encoder-decoder architecture paired with a motion generator to produce looping cinemagrams. We used official implementations of AL, DL, and T2C, and faithfully reproduced EMF based on the provided training details and hyper-parameters. In the following, we will examine the qualitative difference, evaluate the results using two metrics, and report the results of human perceptual studies."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.522,
+                0.092,
+                0.87,
+                0.17
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.523,
+                0.172,
+                0.869,
+                0.252
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.523,
+                0.254,
+                0.868,
+                0.334
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.523,
+                0.337,
+                0.868,
+                0.416
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.499,
+                0.42,
+                0.892,
+                0.462
+            ],
+            "angle": 0,
+            "content": "Figure 4. Qualitative comparison with state-of-the-art cinematograph generation methods. Please refer to the supplementary video for more examples."
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.5,
+                0.472,
+                0.892,
+                0.514
+            ],
+            "angle": 0,
+            "content": "Table 1. Quantitative comparison with the state-of-the-art cinematograph generation methods. We compared our method with AL, EMF, DL, and T2C. The best scores are bolded."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.511,
+                0.525,
+                0.882,
+                0.59
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Static Consistency</td><td colspan=\"4\">Motion Quality</td></tr><tr><td>LPIPS↓</td><td>MS-SSIM↑</td><td>RMSE↓</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>RMSE↓</td><td>FID↓</td></tr><tr><td>AL</td><td>0.0477</td><td>0.9524</td><td>8.8956</td><td>0.0617</td><td>0.6819</td><td>25.1089</td><td>53.6893</td></tr><tr><td>EMF</td><td>0.0103</td><td>0.9723</td><td>6.0440</td><td>0.0533</td><td>0.7102</td><td>22.8932</td><td>45.8035</td></tr><tr><td>DL</td><td>0.0071</td><td>0.9931</td><td>2.1044</td><td>0.0504</td><td>0.7062</td><td>21.8011</td><td>41.6374</td></tr><tr><td>T2C</td><td>0.0063</td><td>0.9773</td><td>4.8785</td><td>0.0159</td><td>0.7224</td><td>21.2784</td><td>40.1186</td></tr><tr><td>Ours</td><td>0.0062</td><td>0.9962</td><td>1.9430</td><td>0.0131</td><td>0.7237</td><td>20.8948</td><td>39.2113</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.607,
+                0.892,
+                0.91
+            ],
+            "angle": 0,
+            "content": "Qualitative Comparisons Figure 4 shows visual results from AL, EMF, DL, T2C, and our method. AL exhibits stretching artifacts (row 1) due to recurrent backward warping, whereas our method uses forward warping that prevents the appearance of such distortions. EMF struggles with precise reconstruction (row 2), because its encoder-decoder architecture is trained from scratch on a low-quality video dataset. In contrast, our method leverages a StyleGAN model pre-trained on high-quality images, resulting in cinemagrams with high perceptual quality. DL often loses fine details in textures (row 3) as it performs GAN inversion using the latent codes of StyleGAN, and fine-tunes the generator. Our method, utilizing deep features for GAN inversion, retains original texture details. T2C encounters difficulties in preserving high-frequency details, and accurate motion segmentation between static and dynamic regions (row 4). In contrast, our method utilizes a pre-trained StyleGAN for cinematograph generation, and uses its features for mask prediction, preserving both fine and structural details of the source image."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "7876"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.092,
+                0.47,
+                0.272
+            ],
+            "angle": 0,
+            "content": "Quantitative Comparisons The evaluation was performed considering two aspects: (1) static consistency and (2) motion quality. Static consistency measures the consistency over time of non-movable regions such as buildings and mountains. Motion quality refers to both the image quality of the animated regions and the animation plausibility. For a fair comparison, we normalized the speed of motion in the generated results for all methods to match the average speed in the test dataset. AL, EMF, DL, and T2C were provided with the first frame \\( I_0 \\) of the test video to generate \\( 512 \\times 512 \\) videos. Our method generates \\( 1024 \\times 1024 \\) results, thus we downsampled it to the size of \\( 512 \\times 512 \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.273,
+                0.468,
+                0.408
+            ],
+            "angle": 0,
+            "content": "For a quantitative evaluation, we used 224 test videos from the Sky Time-Lapse dataset [41]. To measure the static consistency, we computed LPIPS [45], Root-Mean-Squared Error (RMSE), and MS-SSIM [40] between generated frame \\(\\hat{I}_n\\) and the input image \\(I_0\\) with the sky masked out. To measure the motion quality, we used the same evaluation metrics along with Frechet inception distance (FID) [28] between \\(\\hat{I}_n\\) and the ground-truth frame \\(I_n\\) with the static parts masked out."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.409,
+                0.468,
+                0.575
+            ],
+            "angle": 0,
+            "content": "Table 1 reveals the quantitative evaluation results. For both static consistency and motion quality, our method outperforms the other approaches. Our cinematographs represent static consistency better because use of the mask improves the accuracy in the detection of static regions. In addition, the quality of texture details are improved in the generated frames due to GAN inversion and the MSDFW based on the deep features of StyleGAN. The use of our motion generator and DFW leads to improved motion quality for each scene, compared with that of the results of previous approaches."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.583,
+                0.468,
+                0.853
+            ],
+            "angle": 0,
+            "content": "User Study We conducted a user study with 17 participants to subjectively evaluate our method in comparison with previous cinematograph generation methods. We did not target any specific demographic groups, as the evaluation of cinematographs should not be dependent on particular demographic distributions. We conducted two evaluations: score evaluation and side-by-side comparison. Both evaluations assessed static consistency and motion quality. In the score evaluation, participants rated the video presented with a reference image on a 1-to-5 scale, with \"very bad\" being 1 and \"very good\" being 5. In the side-by-side comparison, participants selected the preferred cinematograph from two videos generated using different methods, presented with a reference image. For both evaluations, we used ten samples randomly chosen from the LHQ [31] dataset. To eliminate bias, distinct samples were used for each evaluation, and the positions of the cinematographs were randomized in all questions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.077,
+                0.856,
+                0.468,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Tables 2 and 3 summarize the statistics of the user studies. For both score evaluation and human preference, our method outperforms the previous approaches by a substan"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.499,
+                0.09,
+                0.892,
+                0.116
+            ],
+            "angle": 0,
+            "content": "Table 2. Human perceptual study results for score evaluation. The best scores are bolded."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.561,
+                0.121,
+                0.832,
+                0.203
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Method</td><td>Static Consistency</td><td>Motion Quality</td></tr><tr><td>AL</td><td>1.59 ± 0.20</td><td>1.76 ± 0.45</td></tr><tr><td>EMF</td><td>2.35 ± 0.89</td><td>2.24 ± 0.99</td></tr><tr><td>DL</td><td>3.35 ± 0.83</td><td>3.25 ± 0.67</td></tr><tr><td>T2C</td><td>3.75 ± 0.69</td><td>2.96 ± 0.66</td></tr><tr><td>Ours</td><td>4.37 ± 0.18</td><td>3.86 ± 0.53</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.499,
+                0.217,
+                0.892,
+                0.258
+            ],
+            "angle": 0,
+            "content": "Table 3. Human perceptual study results for side-by-side comparison. The percentage of ours chosen against each competing method is reported."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.558,
+                0.262,
+                0.83,
+                0.344
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">Human Preference (Ours %)</td></tr><tr><td>Static Consistency</td><td>Motion Quality</td></tr><tr><td>vs AL</td><td>99.35%</td><td>92.81%</td></tr><tr><td>vs EMF</td><td>96.08%</td><td>86.93%</td></tr><tr><td>vs DL</td><td>94.77%</td><td>84.97%</td></tr><tr><td>vs T2C</td><td>77.78%</td><td>73.86%</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.355,
+                0.892,
+                0.461
+            ],
+            "angle": 0,
+            "content": "tial margin. We also performed statistical analysis (Kruskal-Wallis H-test) on the resulting evaluation scores. The results showed that our method achieved significantly higher scores in all pairs of comparisons with every previous method \\((p < 0.001\\) in post-hoc analysis). Both user study results reveal the advantages of our method for generating cinematographs with high image quality and plausible animation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.473,
+                0.887,
+                0.489
+            ],
+            "angle": 0,
+            "content": "4.2. Comparisons with Video Generation Methods"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.497,
+                0.892,
+                0.663
+            ],
+            "angle": 0,
+            "content": "We compared our approach with MoCoGAN-HD [34], a video generation method that predicts the trajectory within the latent space of a pre-trained StyleGAN. For content generation, we used the same pre-trained StyleGAN as that used by MoCoGAN-HD, which was trained on the Sky Time-lapse [41] dataset to generate \\(128 \\times 128\\) images. The comparison was made in an unconditional manner using 300 randomly sampled latent codes, at a resolution of \\(128 \\times 128\\). We compared the first 60 frames of the generated videos with those of MoCoGAN-HD, as our method produces looping videos."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.669,
+                0.892,
+                0.789
+            ],
+            "angle": 0,
+            "content": "Qualitative Comparisons Figure 5 presents a qualitative comparison. As shown in the first and third rows, the videos generated using our method exhibit static consistencies, with the clouds moving while the ground remains static. In contrast, as shown in the second and fourth rows, the frames generated by MoCoGAN-HD deviate significantly from the original image, making the method unsuitable for cinematograph generation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.795,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Quantitative Comparisons We compared the content preservation ability of both methods by measuring LPIPS [45], RMSE, and MS-SSIM [40] between the first frame \\(\\hat{I}_0\\) and the subsequent frames \\(\\hat{I}_n\\). Because the main content to be preserved in kinemagrams is the static regions, we defined content preservation as static consistency. We masked out the sky for all image frames according to the"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "7877"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.107,
+                0.088,
+                0.437,
+                0.348
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.077,
+                0.352,
+                0.47,
+                0.395
+            ],
+            "angle": 0,
+            "content": "Figure 5. Qualitative comparison with the state-of-the-art video generation method, MoCoGAN-HD [34]. For more examples from this comparison, please refer to the supplementary video."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.097,
+                0.402,
+                0.449,
+                0.49
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.077,
+                0.495,
+                0.47,
+                0.523
+            ],
+            "angle": 0,
+            "content": "Figure 6. Quantitative comparison of content preservation with state-of-the-art video generation method MoCoGAN-HD [34]."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.534,
+                0.471,
+                0.716
+            ],
+            "angle": 0,
+            "content": "segmentation mask and compared only the static parts. We used 219 videos in which the scene could be horizontally divided into static and animated regions. LPIPS, RMSE, and MS-SSIM were all computed for each pair of \\(\\hat{I}_0\\) and \\(\\hat{I}_n\\), and were averaged over the number of samples. Figure 6 shows that MoCoGAN-HD exhibits significant divergence in terms of LPIPS, RMSE, and MS-SSIM over time. This indicates that MoCoGAN-HD is unable to generate motion that preserves the content of the original image. In contrast, our method exhibits a small diverging trend, confirming its superiority in preserving the content of the original image over time, by utilizing deep features."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.729,
+                0.231,
+                0.745
+            ],
+            "angle": 0,
+            "content": "4.3. Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.75,
+                0.47,
+                0.901
+            ],
+            "angle": 0,
+            "content": "To demonstrate the effectiveness of our design choices for the proposed method, we conducted a series of ablation studies. Specifically, we focused on the effectiveness of our warping and GAN inversion methods. To evaluate warping, we compared the image frames generated with and without the use of 1) forward warping, 2) DFW, 3) MSDFW, and 4) segmentation mask. The first frames of 224 test videos from the Sky Time-Lapse [41] dataset were given as input to generate \\(1024 \\times 1024\\) videos, and we used the first 60 frames for the evaluation. For the evaluation of GAN inversion, we"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.526,
+                0.089,
+                0.693,
+                0.326
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.697,
+                0.09,
+                0.865,
+                0.326
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.5,
+                0.33,
+                0.892,
+                0.358
+            ],
+            "angle": 0,
+            "content": "Figure 7. Results of qualitative comparison in ablation study. Please see the supplementary video for the animated results."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.529,
+                0.363,
+                0.86,
+                0.476
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.517,
+                0.481,
+                0.874,
+                0.495
+            ],
+            "angle": 0,
+            "content": "Figure 8. Qualitative comparison results on GAN inversion."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.501,
+                0.892,
+                0.563
+            ],
+            "angle": 0,
+            "content": "compared images reconstructed with and without the use of the deep features. A total of 256 images, 128 from the Sky Time-Lapse and 128 from the Eulerian [9] dataset were provided to generate \\(1024 \\times 1024\\) reconstructed images."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.573,
+                0.893,
+                0.756
+            ],
+            "angle": 0,
+            "content": "Forward Warping We compared the image frames generated with and without forward warping (FW). For the case without FW, we warped the deep features of StyleGAN using backward warping. The first row in Figure 7 shows the qualitative results of this comparison. As revealed in the figure, without FW, the results usually contain unrealistically stretched textures. For quantitative comparisons, we computed LPIPS, MS-SSIM, and RMSE between \\( I_{n} \\) and \\( \\hat{I}_n \\). Table 4 shows that excluding forward warping resulted in significant degradation in terms of MS-SSIM. This indicates that the stretched textures resulted in huge structural distortions in the generated images."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.765,
+                0.892,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Deep Feature Warping We then compared frames generated with and without DFW. For the case without DFW, we directly warped RGB images using predicted motion fields for comparison. The second row in Figure 7 presents the qualitative comparison. As shown in the figure, the frame generated without DFW contained tearing artifacts in the animated regions. Table 4 shows an increase in the LPIPS score, which indicates that excluding DFW degraded the perceptual quality by introducing tearing artifacts."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "7878"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.077,
+                0.09,
+                0.47,
+                0.117
+            ],
+            "angle": 0,
+            "content": "Table 4. Results of quantitative evaluation in ablation study. The best scores are bolded."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.118,
+                0.122,
+                0.43,
+                0.205
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Method</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>RMSE↓</td></tr><tr><td>Ours - Full</td><td>0.0511</td><td>0.7165</td><td>21.8188</td></tr><tr><td>Ours - w/o FW</td><td>0.0524</td><td>0.6853</td><td>22.9816</td></tr><tr><td>Ours - w/o DFW</td><td>0.0537</td><td>0.6908</td><td>22.4993</td></tr><tr><td>Ours - w/o MSDFW</td><td>0.0629</td><td>0.6946</td><td>22.3227</td></tr><tr><td>Ours - w/o Mask</td><td>0.0564</td><td>0.6980</td><td>22.6213</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.077,
+                0.213,
+                0.468,
+                0.24
+            ],
+            "angle": 0,
+            "content": "Table 5. Quantitative evaluation results on GAN inversion. The best scores are bolded."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.115,
+                0.244,
+                0.428,
+                0.284
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>w+</td><td>D10</td><td>FID↓</td><td>LPIPS↓</td><td>RMSE↓</td><td>MS-SSIM↑</td></tr><tr><td>✓</td><td>-</td><td>40.8276</td><td>0.0153</td><td>18.5901</td><td>0.7891</td></tr><tr><td>✓</td><td>✓</td><td>11.8231</td><td>0.0019</td><td>5.6410</td><td>0.9907</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.296,
+                0.468,
+                0.462
+            ],
+            "angle": 0,
+            "content": "Multi-scale Deep Feature Warping Image frames generated with and without MSDFW were also compared. For the case without MSDFW, only a single deep feature \\(D^{10}\\) was warped and propagated through the next blocks of StyleGAN. The results of the qualitative comparison are shown in the third row of Figure 7. As revealed in the figure, excluding MSDFW resulted in blurry textures in the dynamic region. The quantitative comparison reported in Table 4 shows a significant increase in the LPIPS score. This indicates that excluding MSDFW degraded the perceptual quality, especially the texture details."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.467,
+                0.468,
+                0.616
+            ],
+            "angle": 0,
+            "content": "Segmentation Mask We compared the image frames generated with and without a segmentation mask. For the case without a mask, we used the initially predicted motion field for warping. The fourth row in Figure 7 shows the qualitative result of this comparison. As shown in the figure, excluding the mask resulted in erroneous movement in the static regions. The quantitative comparison reported in Table 4 reveals a significant degradation of MS-SSIM, which indicates the structural distortion caused by the erroneous motion."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.622,
+                0.468,
+                0.803
+            ],
+            "angle": 0,
+            "content": "Deep Feature Inversion To evaluate the effectiveness of GAN inversion, we compared the images reconstructed with and without the use of the deep features \\(D^{10}\\). For the case without deep feature inversion, only the latent codes \\(w^{+}\\) of a pre-trained StyleGAN were used to reconstruct the input landscape image. Figure 8 shows a qualitative result of this comparison. As revealed in the figure, using only \\(w^{+}\\) failed to accurately reconstruct the details of the original images. For quantitative comparisons, FID, LPIPS, RMSE, and MS-SSIM were computed between \\(I\\) and \\(\\hat{I}\\). Table 5 reveals significant improvements in the perceptual quality of the results when reconstructed using the deep features \\(D^{10}\\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.819,
+                0.35,
+                0.835
+            ],
+            "angle": 0,
+            "content": "5. Limitations and Future Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.84,
+                0.468,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Our mask predictor and motion generator generally performed well for landscape images. However, automatic prediction cannot be accurate for all images because most landscape images contain inherent ambiguity in their motion"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.517,
+                0.117,
+                0.532,
+                0.127
+            ],
+            "angle": 0,
+            "content": "(a)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.536,
+                0.092,
+                0.693,
+                0.152
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.701,
+                0.117,
+                0.718,
+                0.127
+            ],
+            "angle": 0,
+            "content": "(b)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.72,
+                0.092,
+                0.796,
+                0.151
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.798,
+                0.092,
+                0.874,
+                0.151
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.499,
+                0.156,
+                0.892,
+                0.198
+            ],
+            "angle": 0,
+            "content": "Figure 9. Limitations of StyleCineGAN: (a) automatic prediction of motion cannot be accurate for all images, and (b) the motion of a very thin structured object is hard to be isolated."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.206,
+                0.892,
+                0.356
+            ],
+            "angle": 0,
+            "content": "directions except for some obvious cases (e.g., waterfalls). The failure case is illustrated in Figure 9 (a). User-defined motion hints can be used to resolve such ambiguities and provide further control capability during the generation process [18, 19]. In addition, it is hard for our method to isolate the motion of a very thin structured object placed within the animated region as shown in Figure 9 (b). This is because our method performs warping of features at multiple resolutions, in which low-resolution features cannot spatially represent the thin structures."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.357,
+                0.892,
+                0.479
+            ],
+            "angle": 0,
+            "content": "In this work, we mainly focused on animating the landscape images, particularly skies and fluids, while putting other types of animations outside the scope. In future developments, we would like to expand the capabilities to include other forms of motion. Investigating the rotating hands of a clock, the playing arm of a guitarist moving up and down, and a flag or the wings of a bird fluttering for cinematograph generation would be a very interesting direction to pursue."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.491,
+                0.619,
+                0.506
+            ],
+            "angle": 0,
+            "content": "6. Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.513,
+                0.892,
+                0.724
+            ],
+            "angle": 0,
+            "content": "We proposed the first approach that leverages a pre-trained StyleGAN for high-quality one-shot landscape cinematograph generation. In contrast to previous studies, our method does not require training a large image generator from scratch, and also systematically improves the resolution of the generated cinematographs to \\(1024 \\times 1024\\). At the core of our method, we utilized the deep features of a pretrained StyleGAN, because those features can help preserve spatial information and encode both high-level semantic and low-level style appearances. Using our MSDFW approach, we applied the predicted motion to the deep feature space of StyleGAN. Both qualitative and quantitative results confirm that our method substantially outperforms existing baselines."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.735,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Acknowledgements This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: Development of Universal Fashion Creation Platform Technology for Avatar Personality Expression, Project Number: RS-2023-00228331, Contribution Rate: \\(90\\%\\) ) and Korea Institute for Advancement of Technology(KIAT) grant funded by the Korea Government(MOTIE) (Project Name: The Competency Development Program for Industry Specialist, Project Number: P0012746, Contribution Rate: \\(10\\%\\))."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.945,
+                0.516,
+                0.957
+            ],
+            "angle": 0,
+            "content": "7879"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.08,
+                0.09,
+                0.174,
+                0.105
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.115,
+                0.47,
+                0.171
+            ],
+            "angle": 0,
+            "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.172,
+                0.469,
+                0.212
+            ],
+            "angle": 0,
+            "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? CoRR, abs/1911.11544, 2019.3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.213,
+                0.469,
+                0.269
+            ],
+            "angle": 0,
+            "content": "[3] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. arXiv preprint arXiv:2201.13433, 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.27,
+                0.469,
+                0.337
+            ],
+            "angle": 0,
+            "content": "[4] Hugo Bertiche, Niloy J. Mitra, Kuldeep Kulkarni, ChunHao Paul Huang, Tuanfeng Y. Wang, Meysam Madadi, Sergio Escalera, and Duygu Ceylan. Blowing in the wind: Cyclenet for human cinematographs from still images, 2023. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.34,
+                0.469,
+                0.396
+            ],
+            "angle": 0,
+            "content": "[5] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. In ACM SIGGRAPH 2005 Papers, pages 853-860. 2005. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.397,
+                0.469,
+                0.464
+            ],
+            "angle": 0,
+            "content": "[6] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM Transactions on Graphics (Proc. of SIGGRAPH ASIA 2019), 38(6):175:1-175:19, 2019. 1, 2, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.466,
+                0.469,
+                0.507
+            ],
+            "angle": 0,
+            "content": "[7] Gereon Fox, Ayush Tewari, Mohamed Elgharib, and Christian Theobalt. Stylevideogan: A temporal generative model using a pretrained stylegan, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.508,
+                0.469,
+                0.577
+            ],
+            "angle": 0,
+            "content": "[8] Tavi Halperin, Hanit Hakim, Orestis Vantzos, Gershon Hochman, Netai Benaim, Lior Sassy, Michael Kupchik, Ofir Bibi, and Ohad Fried. Endless loops: detecting and animating periodic patterns in still images. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.578,
+                0.469,
+                0.646
+            ],
+            "angle": 0,
+            "content": "[9] Aleksander Holynski, Brian L. Curless, Steven M. Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5810-5819, 2021. 1, 2, 3, 4, 5, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.649,
+                0.469,
+                0.689
+            ],
+            "angle": 0,
+            "content": "[10] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the \"steerability\" of generative adversarial networks. In International Conference on Learning Representations, 2020. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.69,
+                0.469,
+                0.732
+            ],
+            "angle": 0,
+            "content": "[11] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.733,
+                0.469,
+                0.788
+            ],
+            "angle": 0,
+            "content": "[12] 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. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.789,
+                0.469,
+                0.83
+            ],
+            "angle": 0,
+            "content": "[13] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.831,
+                0.469,
+                0.872
+            ],
+            "angle": 0,
+            "content": "[14] Alexander Kristoffersen. Loopnerf: Exploring temporal compression for 3d video textures. Master's thesis, EECS Department, University of California, Berkeley, 2023. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.873,
+                0.469,
+                0.902
+            ],
+            "angle": 0,
+            "content": "[15] Xingyi Li, Zhiguo Cao, Huiqiang Sun, Jianming Zhang, Ke Xian, and Guosheng Lin. 3d cinematography from a single"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.115,
+                0.47,
+                0.902
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.535,
+                0.093,
+                0.892,
+                0.134
+            ],
+            "angle": 0,
+            "content": "image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4595-4605, 2023. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.137,
+                0.892,
+                0.193
+            ],
+            "angle": 0,
+            "content": "[16] Ji Lin, Richard Zhang, Frieder Ganz, Song Han, and JunYan Zhu. Anycost gans for interactive image synthesis and editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.195,
+                0.892,
+                0.263
+            ],
+            "angle": 0,
+            "content": "[17] Elizaveta Logacheva, Roman Suvorov, Oleg Khomenko, Anton Mashikhin, and Victor Lempitsky. Deeplandscape: Adversarial modeling of landscape videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 5"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.266,
+                0.892,
+                0.322
+            ],
+            "angle": 0,
+            "content": "[18] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.324,
+                0.892,
+                0.366
+            ],
+            "angle": 0,
+            "content": "[19] Aniruddha Mahapatra, Aliaksandr Siarohin, Hsin-Ying Lee, Sergey Tulyakov, and Jun-Yan Zhu. Text-guided synthesis of eulerian cinemagrams. 2023. 1, 2, 3, 5, 8"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.368,
+                0.892,
+                0.424
+            ],
+            "angle": 0,
+            "content": "[20] Makoto Okabe, Ken Anjyo, Takeo Igarashi, and Hans-Peter Seidel. Animating pictures of fluid using video examples. In Computer Graphics Forum, pages 677-686. Wiley Online Library, 2009. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.426,
+                0.892,
+                0.482
+            ],
+            "angle": 0,
+            "content": "[21] Makoto Okabe, Ken Anjyor, and Rikio Onai. Creating fluid animation from a single image using video database. In Computer Graphics Forum, pages 1973-1982. Wiley Online Library, 2011."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.484,
+                0.892,
+                0.526
+            ],
+            "angle": 0,
+            "content": "[22] Makoto Okabe, Yoshinori Dobashi, and Ken Anjyo. Animating pictures of water scenes using video retrieval. The Visual Computer, 34(3):347-358, 2018. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.528,
+                0.892,
+                0.598
+            ],
+            "angle": 0,
+            "content": "[23] Ehsan Pajouheshgar, Yitao Xu, Tong Zhang, and Sabine Süssstrunk. Dynca: Real-time dynamic texture synthesis using neural cellular automata. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.6,
+                0.892,
+                0.669
+            ],
+            "angle": 0,
+            "content": "[24] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, Jun-Yan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.671,
+                0.892,
+                0.728
+            ],
+            "angle": 0,
+            "content": "[25] Ekta Prashnani, Maneli Noorkami, Daniel Vaquero, and Pradeep Sen. A phase-based approach for animating images using video examples. In Computer Graphics Forum, pages 303-311. Wiley Online Library, 2017. 1, 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.729,
+                0.892,
+                0.799
+            ],
+            "angle": 0,
+            "content": "[26] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.801,
+                0.892,
+                0.857
+            ],
+            "angle": 0,
+            "content": "[27] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan, 2020. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.859,
+                0.892,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[28] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.504,
+                0.093,
+                0.892,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.946,
+                0.516,
+                0.956
+            ],
+            "angle": 0,
+            "content": "7880"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.092,
+                0.47,
+                0.147
+            ],
+            "angle": 0,
+            "content": "[29] Kwanggyoon Seo, Seoung Wug Oh, Jingwan Lu, Joon-Young Lee, Seonghyeon Kim, and Junyong Noh. Styleportraitvideo: Editing portrait videos with expression optimization. 41(7), 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.151,
+                0.472,
+                0.207
+            ],
+            "angle": 0,
+            "content": "[30] Claude E. Shannon. Coding Theorems for a Discrete Source With a Fidelity Criterion Institute of Radio Engineers, International Convention Record, vol. 7, 1959., pages 325-350. 1993. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.209,
+                0.472,
+                0.251
+            ],
+            "angle": 0,
+            "content": "[31] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. arXiv preprint arXiv:2104.06954, 2021. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.253,
+                0.472,
+                0.323
+            ],
+            "angle": 0,
+            "content": "[32] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V. Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.325,
+                0.472,
+                0.381
+            ],
+            "angle": 0,
+            "content": "[33] Matthew Tesfaldet, Marcus A. Brubaker, and Konstantinos G. Derpanis. Two-stream convolutional networks for dynamic texture synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.383,
+                0.472,
+                0.453
+            ],
+            "angle": 0,
+            "content": "[34] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. In International Conference on Learning Representations, 2021. 2, 5, 6, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.455,
+                0.472,
+                0.51
+            ],
+            "angle": 0,
+            "content": "[35] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.513,
+                0.472,
+                0.569
+            ],
+            "angle": 0,
+            "content": "[36] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1526-1535, 2018. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.572,
+                0.472,
+                0.642
+            ],
+            "angle": 0,
+            "content": "[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.644,
+                0.472,
+                0.699
+            ],
+            "angle": 0,
+            "content": "[38] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.702,
+                0.472,
+                0.771
+            ],
+            "angle": 0,
+            "content": "[39] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.773,
+                0.472,
+                0.829
+            ],
+            "angle": 0,
+            "content": "[40] Zhou Wang, Eero P. Simoncelli, and Alan Conrad Bovik. Multiscale structural similarity for image quality assessment. The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, 2:1398-1402 Vol.2, 2003. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.832,
+                0.472,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[41] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6, 7"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.092,
+                0.472,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.093,
+                0.892,
+                0.133
+            ],
+            "angle": 0,
+            "content": "[42] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. pages 357-374. Springer, 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.136,
+                0.892,
+                0.19
+            ],
+            "angle": 0,
+            "content": "[43] Xu Yao, Alasdair Newson, Yann Gousseau, and Pierre Hellier. A style-based gan encoder for high fidelity reconstruction of images and videos. European conference on computer vision, 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.193,
+                0.892,
+                0.289
+            ],
+            "angle": 0,
+            "content": "[44] 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. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 85-101. Springer, 2022. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.291,
+                0.892,
+                0.331
+            ],
+            "angle": 0,
+            "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.333,
+                0.892,
+                0.416
+            ],
+            "angle": 0,
+            "content": "[46] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.503,
+                0.093,
+                0.892,
+                0.416
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.946,
+                0.514,
+                0.956
+            ],
+            "angle": 0,
+            "content": "7881"
+        }
+    ]
+]

2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2cb5e8c2101a58160e1e26c2b8a997e4486434d4d60529675d9d1e3546906096
+size 9571833

2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/full.md

@@ -0,0 +1,317 @@
+# StyleCineGAN: Landscape Cinemagraph Generation using a Pre-trained StyleGAN
+
+Jongwoo Choi
+
+Kwanggyoon Seo
+
+Amirsaman Ashtari
+
+Junyong Noh
+
+Visual Media Lab, KAIST
+
+{jeolpyeoni0,skg1023,a.s.ashtari,junyongnoh}@kaist.ac.kr
+
+Figure 1. Given a landscape image, StyleCineGAN generates a seamless cinematograph at  $1024 \times 1024$  resolution. This figure contains video clips, thus consider viewing it using Adobe Reader. The same results are also included in the supplementary video.
+
+# Abstract
+
+We propose a method that can generate cinematographs automatically from a still landscape image using a pretrained StyleGAN. Inspired by the success of recent unconditional video generation, we leverage a powerful pretrained image generator to synthesize high-quality cinematographs. Unlike previous approaches that mainly utilize the latent space of a pre-trained StyleGAN, our approach utilizes its deep feature space for both GAN inversion and cinematograph generation. Specifically, we propose multi-scale deep feature warping (MSDFW), which warps the intermediate features of a pre-trained StyleGAN at different resolutions. By using MSDFW, the generated cinematographs are of high resolution and exhibit plausible looping animation. We demonstrate the superiority of our method through user studies and quantitative comparisons with state-of-the-art cinematograph generation methods and a video generation method that uses a pre-trained StyleGAN.
+
+# 1. Introduction
+
+Cinemagraph is a unique form of media that combines a still image and video. While most of the scene remains still, subtle and repeated movements in only a small region effectively highlight a particular object or an important event that occurred at the time of capture. Because this mixture
+
+of still and moving elements in one scene creates an interesting eye-catching effect, cinematographs have recently attracted increasing attention in social media.
+
+Despite the growing popularity of cinematograph, its creation heavily relies on a manual process. To create a motion in part of the input image, one typically utilizes an image manipulation tool to stretch, scale, rotate, or shift the part of the image in a physically plausible and aesthetically pleasing way. This process is often time-consuming and requires a high level of skill in the use of image editing tools. Therefore, creating a cinematograph has generally been considered to be a personal project of professionals so far.
+
+To allow ordinary users to create a cinematograph with a single image, automatic methods have been proposed. One line of research uses a reference video as guidance [14, 25, 33]. These methods are capable of producing realistic motions in the target scene similar to a reference video. Recent methods obviate the need for a reference video, by training deep generative models [4, 6, 9, 15, 17-19, 23]. These methods decompose the task into two processes of learning motion and spatial information individually from separate datasets. The decomposition effectively reduces the complexity of simultaneously learning both temporal and spatial information, and improves the perceptual quality of the generated videos. However, these models
+
+must be trained from scratch, which sometimes requires several days if not weeks on a modern GPU. Moreover, these models are not specifically designed to generate high-resolution of  $1024 \times 1024$  cinemagrams, because of the significant requirements in memory and processing power.
+
+In this paper, we propose the first approach to high-quality one-shot landscape cinematograph generation based on a pre-trained StyleGAN [12]. By using the learned image prior of StyleGAN, our method removes the need for training a large model from scratch and systematically improves the resolution of the generated cinematographs to  $1024 \times 1024$ . Moreover, our method enables a user to easily edit the style and appearance of the generated cinematograph by leveraging the properties of StyleGAN for image stylization and editing.
+
+Our method is inspired by recent unconditional video generation methods [7, 34] which allow to navigate the latent space of a pre-trained image generator, to synthesize a high-quality temporally coherent video. Unlike these methods that utilize the latent codes of StyleGAN, we opt to use the deep features that are generated by convolution operations in each layer of StyleGAN. We use these deep features for two reasons. First, we observed that highly detailed landscape images cannot be reconstructed accurately from the latent codes using GAN inversion methods because these latent codes are low-dimensional [30]. Second, a plausible motion that preserves the content cannot be created by only navigating the latent space, because this space is highly semantic-condensed and lacks explicit spatial prior [38].
+
+To solve these problems, for the first time in the cinematograph generation domain, we utilize the deep features of a pre-trained StyleGAN. These deep features preserve spatial information and encode both high-level semantic and low-level style appearances across high and low-level convolutional layers. To produce cinematographs from these deep features, we propose a multi-scale deep feature warping (MSDFW) to apply the motions generated using a motion generator to the deep feature space of StyleGAN.
+
+We demonstrate the effectiveness and advantages of our method by comparing it to state-of-the-art methods in cinematagraph generation [6, 9, 17, 19] and to an unconditional video generation method that uses a pre-trained StyleGAN [34], using various metrics. We also performed a user study to verify the effectiveness of our method for creating visually pleasing cinematagraph effects. Both qualitative and quantitative results confirm that our method substantially outperforms all existing baselines.
+
+# 2. Related Work
+
+# 2.1. Cinemagraph Generation
+
+One of the early studies for cinemagraph generation used a procedural approach to decompose each object into layers
+
+and apply time-varying stochastic motion [5]. This method can handle a diverse range of images such as photos and paintings by relying on manual interaction from the user. Another approach is to use a reference video to animate the given image [20-22, 25]. These methods rely on a statistical motion analysis of videos to transfer their periodic motion property to the desired image. Another method distills a dynamic NeRF to render looping 3D video textures from the representative motion found in the reference video [14]. Without any reference video as guidance, Halperin et al. [8] proposed a framework to animate arbitrary objects with periodic patterns in a simple motion direction.
+
+Recent approaches use the capacity of deep learning to automatically create cinemagrams from a single image. One study trained an image-based renderer to generate water animation, utilizing the water simulation results [32]. Another research first predicts a sequence of normal maps, then generates the corresponding RGB images that show a looping animation of a garment as if it is blown in the wind [4]. Another line of research trains a generator to produce a motion field [6, 9, 15, 18, 19]. Our work is similar to these deep learning-based approaches in that a motion generator is utilized to synthesize the cinemagrams. The difference is that our method leverages the deep features of a pre-trained image generator and thus can generate plausible high-quality results without fine-tuning or training a separate synthesis network. At the same time, our method systematically improves the resolution of the generated cinemagrams to  $1024 \times 1024$ .
+
+# 2.2. Unconditional Video Generation
+
+The task of video generation is known to be difficult because the generation process has to take into account both spatial and temporal information. To ease the problem, previous approaches [27, 36] decompose video generation into content and motion generation. These methods first predict the latent trajectory for motion, then generate a video from the set of predicted latent codes using the image generator. Instead of training the generation model from scratch, MoCoGAN-HD [34] and StyleVideoGAN [7] leverage a pre-trained image generator model, StyleGAN [12]. These methods use the property of well-constructed latent space to morph one frame toward the next frames [10]. Utilizing the image generation capability of a pre-trained generator, MoCoGAN-HD and StyleVideoGAN only need to train a motion trajectory generator, which greatly reduces the training time. These methods can generate high-quality videos in multiple domains such as faces, cars, and even sky. However, the content details are not well preserved, which has hindered the methods' application for cinematograph generation. Our method builds upon a similar concept of using a pre-trained StyleGAN. We leverage the deep features instead of the latent codes of StyleGAN to generate looping and content-preserving cinematographs.
+
+# 2.3. Video Synthesis using pre-trained StyleGAN
+
+A pre-trained generative model, StyleGAN [11-13], has been actively employed for downstream image and video editing applications. For GAN-based applications, one of the essential steps is to project the desired images or videos to the latent space of the pre-trained GAN model, which is known as GAN inversion. The latent codes are obtained using either an optimization technique [1, 2], a trained encoder [26, 35], or a hybrid of both approaches [16]. For video editing applications, previous methods [3, 29, 37, 42] have utilized a pre-trained image encoder to project all of the frames to the latent space. While most of the methods operate in  $\mathcal{W}+$  space [2], we opt to use both  $w^{+}$  and deep features [24, 43, 44] of the pre-trained StyleGAN to accurately reconstruct the original image and synthesize motion.
+
+# 3. Methods
+
+An overview of our method is shown in Figure 2. Given a landscape image  $I$ , our method generates a seamlessly looping cinemagraph  $V = \{\hat{I}_0, \dots, \hat{I}_N\}$  using a pre-trained StyleGAN. At the core of our method, we use a pre-trained generator without additional fine-tuning. The overall process is described below.
+
+(A) First, we project the landscape image into both the latent space and the feature space of StyleGAN. To this end, we train an encoder that outputs both latent codes  $w^{+}$  and intermediate features  $D^{10}$  of StyleGAN (Sec. 3.1).  
+(B) In addition, we predict a mask  $S$  to divide the image into static and dynamic regions (Sec. 3.2).  
+(C) Next, we use a motion generator that accepts the landscape image  $I$  to synthesize a motion field  $M$ , which defines the position of each pixel in the future frames (Sec. 3.3).  
+(D) Lastly, we generate the final cinematograph frames using the pre-trained StyleGAN with deep feature warping (DFW) operation added in between the original layers. (Sec. 3.4).
+
+In the following sections, we will describe the details of each process.
+
+# 3.1. GAN Inversion
+
+The first step for generating a cinematograph is to project the input image to the latent space of a pre-trained StyleGAN. This process is necessary because StyleGAN is an unconditional generator that cannot take a conditioning input. Many previous methods have used  $\mathcal{W} +$  [2] latent space, which is an extended space of native latent space  $\mathcal{W}$ . However, we observed that  $w^{+} \in \mathcal{W}+$  is not expressive enough to reconstruct the original high-frequency details of the landscape images. Therefore, we chose to use the generated deep features  $D^{10}$  of StyleGAN as well as  $w^{+}$ . The use of  $D^{10}$
+
+enables us to recover details in the original input as shown in Figure 8.
+
+To project an image  $I$  to both the latent space and feature space and obtain  $w^{+}$  and  $D^{10}$ , we train an encoder  $E$  similarly to Yao et al. [43], in which the encoder takes  $I$  as input and predicts  $(w^{+}, D^{10})$ . Additional training details can be found in the supplementary material.
+
+# 3.2. Mask Prediction
+
+We improve the quality of GAN inversion described in Sec. 3.1 by training an additional classifier that predicts a mask to separate the static and dynamic regions. Using the mask, the structure of the static regions can be preserved. To this end, we train a multi-layer-perceptron (MLP) classifier that accepts the deep features of StyleGAN as its input and outputs the mask that specifies each dynamic region in the image, as performed in DatasetGAN [46].
+
+To train the classifier, we manually annotate 32 segmentation masks  $S$  of the selected images. We then follow the same procedure as DatasetGAN but using the deep features  $D^{i}$  where  $i \in \{10,11,\dots,18\}$ . We resize all these deep features followed by concatenating them in the channel dimension to construct the input feature  $D^{*}$ . In the end, paired data  $(D^{*},S)$  is constructed. An ensemble of 10 MLP classifiers is trained with a cross-entropy loss. To further improve the performance of the previous work, at inference, we also refine the predicted mask. For additional details, please refer to the supplementary material.
+
+# 3.3. Motion Generation
+
+To generate motion from a static landscape image  $I$ , we use an Eulerian motion field  $M$  which represents the immediate velocity of each pixel. To predict  $M$  given  $I$ , an imaged-to-image translation network [39] is trained as a motion generator. We train the network with paired data  $(I, M)$  as performed in the previous method [9]. Adding controls in motion generation is also possible, either by using text prompts [19] or drawing arrows [18].
+
+While the trained network works well to a certain degree, we observed that the predicted motion field  $M$  does not align with the boundaries between the static and dynamic regions and often contains motion in static regions. Thus, we further refine  $M$  by multiplying it with the segmentation mask  $S$ , which effectively removes most errors. The refined  $M$  is then used to simulate the motion of a pixel at time  $t$  using the Euler integration as follows:
+
+$$
+\begin{array}{l} x _ {t + 1} = x _ {t} + M \left(x _ {t}\right), \quad M \left(x _ {t}\right) = F _ {t \rightarrow t + 1} \left(x _ {t}\right), \\ F _ {0 \rightarrow t} (x _ {0}) = F _ {0 \rightarrow t - 1} (x _ {0}) + M \left(x _ {0} + F _ {0 \rightarrow t - 1} (x _ {0})\right), \tag {1} \\ \end{array}
+$$
+
+where  $x_0$  is the source pixel position at time 0,  $F_{0\rightarrow t}$  is the displacement field that defines the displacement of the pixel position  $x_0$  to the pixel position  $x_t$  at time  $t$ .
+
+![](images/4a67e97121df12534c46aef569f41f801346f128f3bc0015d37c5f820bb34afe.jpg)  
+Motion Generation
+
+![](images/db2dd263e1cfeda263e26d7f00f69f00cf198b488f3b2ce313b548f3c3422577.jpg)  
+Cinemagraph Generation  
+Figure 2. Overview of StyleCineGAN. Given an input landscape image  $I$ , our goal is to generate a cienemagraph using a fixed pre-trained StyleGAN  $G$ . We project the image into both latent codes  $w^{+}$  and deep features  $D^{10}$  of  $G$ . Using the deep features  $D^{*}$ , a mask predictor predicts a segmentation mask  $S$ . To animate the input image, we use a motion generator to predict the motion field  $M$  from  $I$ .  $M$  is refined using  $S$ . Through Euler integration,  $M$  produces the future and past displacement fields  $F_{0\rightarrow t}$  and  $F_{N\rightarrow t}$ . To synthesize cienemagraph frames, we add a DFW layer in between the layers of  $G$ . DFW refers to Eqns. 2 and 3. This modification enables the intermediate features of  $G$  to be warped according to  $F_{0\rightarrow t}$  and  $F_{N\rightarrow t}$  using a joint splatting method at different resolutions, specifically for the StyleGAN layers indexed with  $i\in [10,12,14,16,18]$ . The warped deep features are used to synthesize frames  $\hat{I}_t$  resulting in the final cienemagraph video.
+
+# 3.4. Cinemagraph Generation
+
+After acquiring the displacement field  $F_{0\rightarrow t}$  (Sec. 3.3), latent codes  $w^{+}$ , and deep features  $D^{10}$  of StyleGAN (Sec. 3.1), we feed these data into the pre-trained StyleGAN with our DFW layer added. We apply forward warping (or splatting) to the original content to generate a cinematograph. Here, the application of forward warping directly to an RGB image often results in images with tearing artifacts. To reduce the artifacts, we warp the deep features  $D_0^i$  of StyleGAN in different resolutions or scales:
+
+$$
+D _ {t} ^ {i} = \operatorname {W a r p} \left(D _ {0} ^ {i}, F _ {0 \rightarrow t}\right), \tag {2}
+$$
+
+where  $\text{Warp}$  is the forward warping function and  $D_{t}^{i}$  is the warped deep feature at time  $t$  and scale  $i$ . We observed that warping a single deep feature (e.g., only  $D_{0}^{10}$ ) results in blurry textures in the generated cinemagrams. Therefore, we opt to warp the multi-scale deep features; we call this operation a MSDFW. Specifically, we apply warping to the deep features  $D_{0}^{i}$  where  $i \in [10, 12, 14, 16, 18]$ . The deep feature  $D_{0}^{10}$  is acquired using GAN inversion, and the consequent deep features  $D_{0}^{12}, D_{0}^{14}, \ldots$ , and  $D_{0}^{18}$  are subsequently generated using both  $D_{0}^{10}$  and  $w^{+}$ , as shown in the rightmost column of Figure 2.
+
+Looping video  $V$  with frame length  $N + 1$  is synthesized using the joint splating technique [9] using the future and past displacement fields  $F_{0\rightarrow t}$  and  $F_{N\rightarrow t}$ , generated through Euler integration. The displacement fields are computed with  $t$  times of integration on  $M$  for  $F_{0\rightarrow t}$ , and  $N - t$  times of integration on  $-M$  for  $F_{N\rightarrow t}$ . The multi-scale fea
+
+ture  $D_0^i$  is warped in both directions and is composited to form  $D_t^i$ . Specifically,  $D_t^i$  is computed as a weighted sum of  $Warp(D_0^i,F_{0\to t})$  and  $Warp(D_0^i,F_{N\to t})$  as follows:
+
+$$
+\begin{array}{l} D _ {t} ^ {i} \left(x ^ {\prime}\right) = \frac {\sum_ {x \in \chi} \alpha_ {t} \cdot W a r p \left(D _ {0} ^ {i} (x) , F _ {0 \rightarrow t}\right)}{\sum_ {x \in \chi} \alpha_ {t}} \\ + \frac {\sum_ {x \in \chi} \left(1 - \alpha_ {t}\right) \cdot W a r p \left(D _ {0} ^ {i} (x) , F _ {N \rightarrow t}\right)}{\sum_ {x \in \chi} \left(1 - \alpha_ {t}\right)}, \tag {3} \\ \end{array}
+$$
+
+where  $x \in \chi$  is a set of pixels being mapped to the same destination pixel  $x'$ , and  $\alpha_{t}$  is the looping weight defined as  $(1 - \frac{t}{N})$ .
+
+With the above DFW module, we can generate a video frame given the predicted deep features  $D^{10}$ , latent code  $w^{+}$ , and mask  $S$ :
+
+$$
+\begin{array}{l} \hat {I} _ {t} = S \odot \left(G _ {\text {w a r p}} \left(D _ {0} ^ {1 0}, w ^ {+}, F _ {0 \rightarrow t}, F _ {N \rightarrow t}\right)\right) \\ + (\mathbf {1} - S) \odot I, \tag {4} \\ \end{array}
+$$
+
+where  $\odot$  is an element-wise multiplication and  $G_{\text{warp}}$  is a fixed pre-trained StyleGAN that incorporates our DFW module.
+
+Style Interpolation In addition to motion generation, the change of appearance style is an additional feature often observed in a landscape cinematograph. This can be achieved by interpolating the latent code with the target latent code as follows:
+
+$$
+w _ {s} ^ {+} = w ^ {+} \cdot (1 - \beta) + w _ {t} ^ {+} \cdot \beta , \tag {5}
+$$
+
+where  $w_{s}^{+}$  is the interpolated latent code,  $w_{t}^{+}$  is the latent code of the target style, and  $\beta$  is the interpolation weight. This is possible because the later layers of the StyleGAN only modify the color of the synthesized image while the original structure is preserved.
+
+With modified Equation 4, we prevent the visual mismatch between the static and dynamic regions by also reflecting the changes to the static regions as follows:
+
+$$
+\begin{array}{l} \hat {I} _ {t} = S \odot \left(G _ {\text {w a r p}} \left(D _ {0} ^ {1 0}, w _ {s} ^ {+}, F _ {0 \rightarrow t}, F _ {N \rightarrow t}\right)\right) \tag {6} \\ + (\mathbf {1} - S) \odot (I + \Delta I _ {s}), \\ \end{array}
+$$
+
+$$
+\Delta I _ {s} = G \left(D _ {0} ^ {1 0}, w _ {s} ^ {+}\right) - G \left(D _ {0} ^ {1 0}, w ^ {+}\right), \tag {7}
+$$
+
+where  $\Delta I_{s}$  is the color difference between the two images generated from latent codes  $w_{s}^{+}$  and  $w^{+}$ .
+
+# 4. Experiments
+
+In this section, we compare our method with state-of-the-art landscape cinematograph generation methods [6, 9, 17, 19] (Sec. 4.1) and with an unconditional video generation method [34] (Sec. 4.2). In addition, we show the importance of the components used in our method through ablation studies (Sec. 4.3). To observe more various results, we recommend readers see the supplementary video. Example results are presented in Figures 1 and 3.
+
+# 4.1. Comparisons with Cinemagraph Generation Methods
+
+We compared our method with state-of-the-art cinemagraph generation methods: Animating Landscape (AL) [6], Deep Landscape (DL) [17], Eulerian Motion Field (EMF) [9], and Text2Cinemagraph (T2C) [19]. AL is a learning-based method that uses backward warping to generate cinemagraph frames. DL trains a style-based generator to synthesize images and dynamic motion in a scene. EMF and T2C train an encoder-decoder architecture paired with a motion generator to produce looping cinemagrams. We used official implementations of AL, DL, and T2C, and faithfully reproduced EMF based on the provided training details and hyper-parameters. In the following, we will examine the qualitative difference, evaluate the results using two metrics, and report the results of human perceptual studies.
+
+![](images/0969637ceb9ca37a33559ba2e6c9207ff7047175487bea4ca60a64b79dfa815f.jpg)  
+Figure 3. Generated cinematagraph results. This figure contains video clips, thus consider viewing it using Adobe Reader. The first two are cinematographs without appearance change, and the last two are cinematographs with appearance change.
+
+![](images/39a0213e700ef9bc56ba9bc73a6eeddde37538bac4a9f16f4257e2c9eadcc04f.jpg)
+
+![](images/2b2dff4bbec86926ff16f3a9189aec4cf3d1734c21855e584c6339d7e0200f38.jpg)
+
+![](images/fd20f8b3e221aecb6ce65ca529f33fe1dbf2ad802f07ce17154bbfe3efe03578.jpg)  
+Figure 4. Qualitative comparison with state-of-the-art cinematograph generation methods. Please refer to the supplementary video for more examples.
+
+Table 1. Quantitative comparison with the state-of-the-art cinematograph generation methods. We compared our method with AL, EMF, DL, and T2C. The best scores are bolded.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Static Consistency</td><td colspan="4">Motion Quality</td></tr><tr><td>LPIPS↓</td><td>MS-SSIM↑</td><td>RMSE↓</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>RMSE↓</td><td>FID↓</td></tr><tr><td>AL</td><td>0.0477</td><td>0.9524</td><td>8.8956</td><td>0.0617</td><td>0.6819</td><td>25.1089</td><td>53.6893</td></tr><tr><td>EMF</td><td>0.0103</td><td>0.9723</td><td>6.0440</td><td>0.0533</td><td>0.7102</td><td>22.8932</td><td>45.8035</td></tr><tr><td>DL</td><td>0.0071</td><td>0.9931</td><td>2.1044</td><td>0.0504</td><td>0.7062</td><td>21.8011</td><td>41.6374</td></tr><tr><td>T2C</td><td>0.0063</td><td>0.9773</td><td>4.8785</td><td>0.0159</td><td>0.7224</td><td>21.2784</td><td>40.1186</td></tr><tr><td>Ours</td><td>0.0062</td><td>0.9962</td><td>1.9430</td><td>0.0131</td><td>0.7237</td><td>20.8948</td><td>39.2113</td></tr></table>
+
+Qualitative Comparisons Figure 4 shows visual results from AL, EMF, DL, T2C, and our method. AL exhibits stretching artifacts (row 1) due to recurrent backward warping, whereas our method uses forward warping that prevents the appearance of such distortions. EMF struggles with precise reconstruction (row 2), because its encoder-decoder architecture is trained from scratch on a low-quality video dataset. In contrast, our method leverages a StyleGAN model pre-trained on high-quality images, resulting in cinemagrams with high perceptual quality. DL often loses fine details in textures (row 3) as it performs GAN inversion using the latent codes of StyleGAN, and fine-tunes the generator. Our method, utilizing deep features for GAN inversion, retains original texture details. T2C encounters difficulties in preserving high-frequency details, and accurate motion segmentation between static and dynamic regions (row 4). In contrast, our method utilizes a pre-trained StyleGAN for cinematograph generation, and uses its features for mask prediction, preserving both fine and structural details of the source image.
+
+Quantitative Comparisons The evaluation was performed considering two aspects: (1) static consistency and (2) motion quality. Static consistency measures the consistency over time of non-movable regions such as buildings and mountains. Motion quality refers to both the image quality of the animated regions and the animation plausibility. For a fair comparison, we normalized the speed of motion in the generated results for all methods to match the average speed in the test dataset. AL, EMF, DL, and T2C were provided with the first frame  $I_0$  of the test video to generate  $512 \times 512$  videos. Our method generates  $1024 \times 1024$  results, thus we downsampled it to the size of  $512 \times 512$ .
+
+For a quantitative evaluation, we used 224 test videos from the Sky Time-Lapse dataset [41]. To measure the static consistency, we computed LPIPS [45], Root-Mean-Squared Error (RMSE), and MS-SSIM [40] between generated frame  $\hat{I}_n$  and the input image  $I_0$  with the sky masked out. To measure the motion quality, we used the same evaluation metrics along with Frechet inception distance (FID) [28] between  $\hat{I}_n$  and the ground-truth frame  $I_n$  with the static parts masked out.
+
+Table 1 reveals the quantitative evaluation results. For both static consistency and motion quality, our method outperforms the other approaches. Our cinematographs represent static consistency better because use of the mask improves the accuracy in the detection of static regions. In addition, the quality of texture details are improved in the generated frames due to GAN inversion and the MSDFW based on the deep features of StyleGAN. The use of our motion generator and DFW leads to improved motion quality for each scene, compared with that of the results of previous approaches.
+
+User Study We conducted a user study with 17 participants to subjectively evaluate our method in comparison with previous cinematograph generation methods. We did not target any specific demographic groups, as the evaluation of cinematographs should not be dependent on particular demographic distributions. We conducted two evaluations: score evaluation and side-by-side comparison. Both evaluations assessed static consistency and motion quality. In the score evaluation, participants rated the video presented with a reference image on a 1-to-5 scale, with "very bad" being 1 and "very good" being 5. In the side-by-side comparison, participants selected the preferred cinematograph from two videos generated using different methods, presented with a reference image. For both evaluations, we used ten samples randomly chosen from the LHQ [31] dataset. To eliminate bias, distinct samples were used for each evaluation, and the positions of the cinematographs were randomized in all questions.
+
+Tables 2 and 3 summarize the statistics of the user studies. For both score evaluation and human preference, our method outperforms the previous approaches by a substan
+
+Table 2. Human perceptual study results for score evaluation. The best scores are bolded.  
+
+<table><tr><td>Method</td><td>Static Consistency</td><td>Motion Quality</td></tr><tr><td>AL</td><td>1.59 ± 0.20</td><td>1.76 ± 0.45</td></tr><tr><td>EMF</td><td>2.35 ± 0.89</td><td>2.24 ± 0.99</td></tr><tr><td>DL</td><td>3.35 ± 0.83</td><td>3.25 ± 0.67</td></tr><tr><td>T2C</td><td>3.75 ± 0.69</td><td>2.96 ± 0.66</td></tr><tr><td>Ours</td><td>4.37 ± 0.18</td><td>3.86 ± 0.53</td></tr></table>
+
+Table 3. Human perceptual study results for side-by-side comparison. The percentage of ours chosen against each competing method is reported.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Human Preference (Ours %)</td></tr><tr><td>Static Consistency</td><td>Motion Quality</td></tr><tr><td>vs AL</td><td>99.35%</td><td>92.81%</td></tr><tr><td>vs EMF</td><td>96.08%</td><td>86.93%</td></tr><tr><td>vs DL</td><td>94.77%</td><td>84.97%</td></tr><tr><td>vs T2C</td><td>77.78%</td><td>73.86%</td></tr></table>
+
+tial margin. We also performed statistical analysis (Kruskal-Wallis H-test) on the resulting evaluation scores. The results showed that our method achieved significantly higher scores in all pairs of comparisons with every previous method  $(p < 0.001$  in post-hoc analysis). Both user study results reveal the advantages of our method for generating cinematographs with high image quality and plausible animation.
+
+# 4.2. Comparisons with Video Generation Methods
+
+We compared our approach with MoCoGAN-HD [34], a video generation method that predicts the trajectory within the latent space of a pre-trained StyleGAN. For content generation, we used the same pre-trained StyleGAN as that used by MoCoGAN-HD, which was trained on the Sky Time-lapse [41] dataset to generate  $128 \times 128$  images. The comparison was made in an unconditional manner using 300 randomly sampled latent codes, at a resolution of  $128 \times 128$ . We compared the first 60 frames of the generated videos with those of MoCoGAN-HD, as our method produces looping videos.
+
+Qualitative Comparisons Figure 5 presents a qualitative comparison. As shown in the first and third rows, the videos generated using our method exhibit static consistencies, with the clouds moving while the ground remains static. In contrast, as shown in the second and fourth rows, the frames generated by MoCoGAN-HD deviate significantly from the original image, making the method unsuitable for cinematograph generation.
+
+Quantitative Comparisons We compared the content preservation ability of both methods by measuring LPIPS [45], RMSE, and MS-SSIM [40] between the first frame  $\hat{I}_0$  and the subsequent frames  $\hat{I}_n$ . Because the main content to be preserved in kinemagrams is the static regions, we defined content preservation as static consistency. We masked out the sky for all image frames according to the
+
+![](images/6019b077d7cbd2e52e32a8f09f5c75e1aed65c3732c400e006c828ad8c966ef1.jpg)  
+Figure 5. Qualitative comparison with the state-of-the-art video generation method, MoCoGAN-HD [34]. For more examples from this comparison, please refer to the supplementary video.
+
+![](images/a794bf00fcc35105d17098df0a2401efecfbcc2e2b8133156a66ab6a93226d33.jpg)  
+Figure 6. Quantitative comparison of content preservation with state-of-the-art video generation method MoCoGAN-HD [34].
+
+segmentation mask and compared only the static parts. We used 219 videos in which the scene could be horizontally divided into static and animated regions. LPIPS, RMSE, and MS-SSIM were all computed for each pair of  $\hat{I}_0$  and  $\hat{I}_n$ , and were averaged over the number of samples. Figure 6 shows that MoCoGAN-HD exhibits significant divergence in terms of LPIPS, RMSE, and MS-SSIM over time. This indicates that MoCoGAN-HD is unable to generate motion that preserves the content of the original image. In contrast, our method exhibits a small diverging trend, confirming its superiority in preserving the content of the original image over time, by utilizing deep features.
+
+# 4.3. Ablation Study
+
+To demonstrate the effectiveness of our design choices for the proposed method, we conducted a series of ablation studies. Specifically, we focused on the effectiveness of our warping and GAN inversion methods. To evaluate warping, we compared the image frames generated with and without the use of 1) forward warping, 2) DFW, 3) MSDFW, and 4) segmentation mask. The first frames of 224 test videos from the Sky Time-Lapse [41] dataset were given as input to generate  $1024 \times 1024$  videos, and we used the first 60 frames for the evaluation. For the evaluation of GAN inversion, we
+
+![](images/eb5b3cdaa79bcc779d32b3d97b9387f720c1d68cd27af3559c9ae022d5cdd0ab.jpg)  
+Figure 7. Results of qualitative comparison in ablation study. Please see the supplementary video for the animated results.
+
+![](images/1ef88055d4ef5ca0a0e9822c91ff7031dd6c9f05db397b92441a12a59e9fd8f4.jpg)
+
+![](images/a7f42b19cf715c06c7d4e413353f86ecd73d80170520fac7dc3477bb3184a3f1.jpg)  
+Figure 8. Qualitative comparison results on GAN inversion.
+
+compared images reconstructed with and without the use of the deep features. A total of 256 images, 128 from the Sky Time-Lapse and 128 from the Eulerian [9] dataset were provided to generate  $1024 \times 1024$  reconstructed images.
+
+Forward Warping We compared the image frames generated with and without forward warping (FW). For the case without FW, we warped the deep features of StyleGAN using backward warping. The first row in Figure 7 shows the qualitative results of this comparison. As revealed in the figure, without FW, the results usually contain unrealistically stretched textures. For quantitative comparisons, we computed LPIPS, MS-SSIM, and RMSE between  $I_{n}$  and  $\hat{I}_n$ . Table 4 shows that excluding forward warping resulted in significant degradation in terms of MS-SSIM. This indicates that the stretched textures resulted in huge structural distortions in the generated images.
+
+Deep Feature Warping We then compared frames generated with and without DFW. For the case without DFW, we directly warped RGB images using predicted motion fields for comparison. The second row in Figure 7 presents the qualitative comparison. As shown in the figure, the frame generated without DFW contained tearing artifacts in the animated regions. Table 4 shows an increase in the LPIPS score, which indicates that excluding DFW degraded the perceptual quality by introducing tearing artifacts.
+
+Table 4. Results of quantitative evaluation in ablation study. The best scores are bolded.  
+
+<table><tr><td>Method</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>RMSE↓</td></tr><tr><td>Ours - Full</td><td>0.0511</td><td>0.7165</td><td>21.8188</td></tr><tr><td>Ours - w/o FW</td><td>0.0524</td><td>0.6853</td><td>22.9816</td></tr><tr><td>Ours - w/o DFW</td><td>0.0537</td><td>0.6908</td><td>22.4993</td></tr><tr><td>Ours - w/o MSDFW</td><td>0.0629</td><td>0.6946</td><td>22.3227</td></tr><tr><td>Ours - w/o Mask</td><td>0.0564</td><td>0.6980</td><td>22.6213</td></tr></table>
+
+Table 5. Quantitative evaluation results on GAN inversion. The best scores are bolded.  
+
+<table><tr><td>w+</td><td>D10</td><td>FID↓</td><td>LPIPS↓</td><td>RMSE↓</td><td>MS-SSIM↑</td></tr><tr><td>✓</td><td>-</td><td>40.8276</td><td>0.0153</td><td>18.5901</td><td>0.7891</td></tr><tr><td>✓</td><td>✓</td><td>11.8231</td><td>0.0019</td><td>5.6410</td><td>0.9907</td></tr></table>
+
+Multi-scale Deep Feature Warping Image frames generated with and without MSDFW were also compared. For the case without MSDFW, only a single deep feature  $D^{10}$  was warped and propagated through the next blocks of StyleGAN. The results of the qualitative comparison are shown in the third row of Figure 7. As revealed in the figure, excluding MSDFW resulted in blurry textures in the dynamic region. The quantitative comparison reported in Table 4 shows a significant increase in the LPIPS score. This indicates that excluding MSDFW degraded the perceptual quality, especially the texture details.
+
+Segmentation Mask We compared the image frames generated with and without a segmentation mask. For the case without a mask, we used the initially predicted motion field for warping. The fourth row in Figure 7 shows the qualitative result of this comparison. As shown in the figure, excluding the mask resulted in erroneous movement in the static regions. The quantitative comparison reported in Table 4 reveals a significant degradation of MS-SSIM, which indicates the structural distortion caused by the erroneous motion.
+
+Deep Feature Inversion To evaluate the effectiveness of GAN inversion, we compared the images reconstructed with and without the use of the deep features  $D^{10}$ . For the case without deep feature inversion, only the latent codes  $w^{+}$  of a pre-trained StyleGAN were used to reconstruct the input landscape image. Figure 8 shows a qualitative result of this comparison. As revealed in the figure, using only  $w^{+}$  failed to accurately reconstruct the details of the original images. For quantitative comparisons, FID, LPIPS, RMSE, and MS-SSIM were computed between  $I$  and  $\hat{I}$ . Table 5 reveals significant improvements in the perceptual quality of the results when reconstructed using the deep features  $D^{10}$ .
+
+# 5. Limitations and Future Work
+
+Our mask predictor and motion generator generally performed well for landscape images. However, automatic prediction cannot be accurate for all images because most landscape images contain inherent ambiguity in their motion
+
+![](images/0a3178b499e93c16913a8a8349e6d82741ef8a510389777e8ab9dc7d67285d5f.jpg)  
+(a)
+
+![](images/99266baccdda3e0ba47175c59d0df00a990c5ee099f3af6ed316f18498209f98.jpg)  
+(b)  
+Figure 9. Limitations of StyleCineGAN: (a) automatic prediction of motion cannot be accurate for all images, and (b) the motion of a very thin structured object is hard to be isolated.
+
+![](images/66837910d83aaf923e84f8e453c57f1e4b218fed97a2f2fc7b454e387d80de0e.jpg)
+
+directions except for some obvious cases (e.g., waterfalls). The failure case is illustrated in Figure 9 (a). User-defined motion hints can be used to resolve such ambiguities and provide further control capability during the generation process [18, 19]. In addition, it is hard for our method to isolate the motion of a very thin structured object placed within the animated region as shown in Figure 9 (b). This is because our method performs warping of features at multiple resolutions, in which low-resolution features cannot spatially represent the thin structures.
+
+In this work, we mainly focused on animating the landscape images, particularly skies and fluids, while putting other types of animations outside the scope. In future developments, we would like to expand the capabilities to include other forms of motion. Investigating the rotating hands of a clock, the playing arm of a guitarist moving up and down, and a flag or the wings of a bird fluttering for cinematograph generation would be a very interesting direction to pursue.
+
+# 6. Conclusion
+
+We proposed the first approach that leverages a pre-trained StyleGAN for high-quality one-shot landscape cinematograph generation. In contrast to previous studies, our method does not require training a large image generator from scratch, and also systematically improves the resolution of the generated cinematographs to  $1024 \times 1024$ . At the core of our method, we utilized the deep features of a pretrained StyleGAN, because those features can help preserve spatial information and encode both high-level semantic and low-level style appearances. Using our MSDFW approach, we applied the predicted motion to the deep feature space of StyleGAN. Both qualitative and quantitative results confirm that our method substantially outperforms existing baselines.
+
+Acknowledgements This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: Development of Universal Fashion Creation Platform Technology for Avatar Personality Expression, Project Number: RS-2023-00228331, Contribution Rate:  $90\%$  ) and Korea Institute for Advancement of Technology(KIAT) grant funded by the Korea Government(MOTIE) (Project Name: The Competency Development Program for Industry Specialist, Project Number: P0012746, Contribution Rate:  $10\%$ ).
+
+# References
+
+[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 3  
+[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? CoRR, abs/1911.11544, 2019.3  
+[3] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. arXiv preprint arXiv:2201.13433, 2022. 3  
+[4] Hugo Bertiche, Niloy J. Mitra, Kuldeep Kulkarni, ChunHao Paul Huang, Tuanfeng Y. Wang, Meysam Madadi, Sergio Escalera, and Duygu Ceylan. Blowing in the wind: Cyclenet for human cinematographs from still images, 2023. 1, 2  
+[5] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. In ACM SIGGRAPH 2005 Papers, pages 853-860. 2005. 2  
+[6] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM Transactions on Graphics (Proc. of SIGGRAPH ASIA 2019), 38(6):175:1-175:19, 2019. 1, 2, 5  
+[7] Gereon Fox, Ayush Tewari, Mohamed Elgharib, and Christian Theobalt. Stylevideogan: A temporal generative model using a pretrained stylegan, 2021. 2  
+[8] Tavi Halperin, Hanit Hakim, Orestis Vantzos, Gershon Hochman, Netai Benaim, Lior Sassy, Michael Kupchik, Ofir Bibi, and Ohad Fried. Endless loops: detecting and animating periodic patterns in still images. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 2  
+[9] Aleksander Holynski, Brian L. Curless, Steven M. Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5810-5819, 2021. 1, 2, 3, 4, 5, 7  
+[10] Ali Jahanian, Lucy Chai, and Phillip Isola. "On the "steerability" of generative adversarial networks. In International Conference on Learning Representations, 2020. 2  
+[11] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 3  
+[12] 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. 2  
+[13] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 3  
+[14] Alexander Kristoffersen. Loopnerf: Exploring temporal compression for 3d video textures. Master's thesis, EECS Department, University of California, Berkeley, 2023. 1, 2  
+[15] Xingyi Li, Zhiguo Cao, Huiqiang Sun, Jianming Zhang, Ke Xian, and Guosheng Lin. 3d cinematography from a single
+
+image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4595-4605, 2023. 1, 2  
+[16] Ji Lin, Richard Zhang, Frieder Ganz, Song Han, and JunYan Zhu. Anycost gans for interactive image synthesis and editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3  
+[17] Elizaveta Logacheva, Roman Suvorov, Oleg Khomenko, Anton Mashikhin, and Victor Lempitsky. Deeplandscape: Adversarial modeling of landscape videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 5  
+[18] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 8  
+[19] Aniruddha Mahapatra, Aliaksandr Siarohin, Hsin-Ying Lee, Sergey Tulyakov, and Jun-Yan Zhu. Text-guided synthesis of eulerian cinemagrams. 2023. 1, 2, 3, 5, 8  
+[20] Makoto Okabe, Ken Anjyo, Takeo Igarashi, and Hans-Peter Seidel. Animating pictures of fluid using video examples. In Computer Graphics Forum, pages 677-686. Wiley Online Library, 2009. 2  
+[21] Makoto Okabe, Ken Anjyor, and Rikio Onai. Creating fluid animation from a single image using video database. In Computer Graphics Forum, pages 1973-1982. Wiley Online Library, 2011.  
+[22] Makoto Okabe, Yoshinori Dobashi, and Ken Anjyo. Animating pictures of water scenes using video retrieval. The Visual Computer, 34(3):347-358, 2018. 2  
+[23] Ehsan Pajouheshgar, Yitao Xu, Tong Zhang, and Sabine Süssstrunk. Dynca: Real-time dynamic texture synthesis using neural cellular automata. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1  
+[24] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, Jun-Yan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3  
+[25] Ekta Prashnani, Maneli Noorkami, Daniel Vaquero, and Pradeep Sen. A phase-based approach for animating images using video examples. In Computer Graphics Forum, pages 303-311. Wiley Online Library, 2017. 1, 2  
+[26] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3  
+[27] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan, 2020. 2  
+[28] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6
+
+[29] Kwanggyoon Seo, Seoung Wug Oh, Jingwan Lu, Joon-Young Lee, Seonghyeon Kim, and Junyong Noh. Styleportraitvideo: Editing portrait videos with expression optimization. 41(7), 2022. 3  
+[30] Claude E. Shannon. Coding Theorems for a Discrete Source With a Fidelity Criterion Institute of Radio Engineers, International Convention Record, vol. 7, 1959., pages 325-350. 1993. 2  
+[31] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. arXiv preprint arXiv:2104.06954, 2021. 6  
+[32] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V. Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 2  
+[33] Matthew Tesfaldet, Marcus A. Brubaker, and Konstantinos G. Derpanis. Two-stream convolutional networks for dynamic texture synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1  
+[34] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. In International Conference on Learning Representations, 2021. 2, 5, 6, 7  
+[35] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 3  
+[36] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1526-1535, 2018. 2  
+[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 3  
+[38] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2  
+[39] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3  
+[40] Zhou Wang, Eero P. Simoncelli, and Alan Conrad Bovik. Multiscale structural similarity for image quality assessment. The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, 2:1398-1402 Vol.2, 2003. 6  
+[41] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6, 7
+
+[42] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. pages 357-374. Springer, 2022. 3  
+[43] Xu Yao, Alasdair Newson, Yann Gousseau, and Pierre Hellier. A style-based gan encoder for high fidelity reconstruction of images and videos. European conference on computer vision, 2022. 3  
+[44] 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. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 85-101. Springer, 2022. 3  
+[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6  
+[46] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3

2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccb6bc7c92ecd16747151c7ff7851f45159be93ade091329d03ffd6efec9af4f
+size 430388

2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/8c37e125-7381-4a71-8f2b-015c3d8a066a_content_list.json

@@ -0,0 +1,1812 @@
+[
+    {
+        "type": "text",
+        "text": "SuGaR: Surface-Aligned Gaussian Splitting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering",
+        "text_level": 1,
+        "bbox": [
+            135,
+            130,
+            833,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Antoine Guédon Vincent Lepetit  \nLIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France  \nhttps://anttwo.github.io/sugar/",
+        "bbox": [
+            243,
+            203,
+            725,
+            256
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/6673930ff670e4525dde86ea891322dc4edf5ae835eac2957ca13ae2f137f242.jpg",
+        "image_caption": [
+            "Figure 1. We introduce a method that extracts accurate and editable meshes from 3D Gaussian Splitting representations within minutes on a single GPU. The meshes can be edited, animated, composited, etc. with very realistic Gaussian Splitting rendering, offering new possibilities for Computer Graphics. Note for example that we changed the posture of the robot between the captured scene on the bottom left and the composited scene on the right. The supplementary material provides more examples, including a video illustrating our results."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            78,
+            267,
+            202,
+            429
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/1e71b34782521a295dba22014fb33764264018e29489fec3a41da095d26a740b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            217,
+            268,
+            504,
+            429
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/10648b5890419237a4fb6653f09ebb64111dc67816d3d828c0fbbb2c4b6de2cf.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            268,
+            890,
+            429
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            510,
+            313,
+            525
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We propose a method to allow precise and extremely fast mesh extraction from 3D Gaussian Splatting [15]. Gaussian Splatting has recently become very popular as it yields realistic rendering while being significantly faster to train than NeRFs. It is however challenging to extract a mesh from the millions of tiny 3D Gaussians as these Gaussians tend to be unorganized after optimization and no method has been proposed so far. Our first key contribution is a regularization term that encourages the Gaussians to align well with the surface of the scene. We then introduce a method that exploits this alignment to extract a mesh from the Gaussians using Poisson reconstruction, which is fast, scalable, and preserves details, in contrast to the Marching Cubes algorithm usually applied to extract meshes from Neural SDFs. Finally, we introduce an optional refinement strategy that binds Gaussians to the surface of the mesh, and jointly optimizes these Gaussians and the mesh through Gaussian splatting rendering. This enables easy editing, sculpting, animating, and relighting of the Gaussians by manipulating the mesh instead of the Gaussians themselves. Retrieving such an editable mesh for realistic rendering is done within minutes with our method, compared to hours with the state-of-the-art method on SDFs, while providing a better rendering quality.",
+        "bbox": [
+            73,
+            531,
+            472,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            501,
+            510,
+            630,
+            525
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "After NeRFs [22], 3D Gaussian Splatting [15] has recently become very popular for capturing a 3D scene and rendering it from novel points of view. 3D Gaussian Splatting optimizes the positions, orientations, appearances (represented as spherical harmonics), and alpha blending of many tiny 3D Gaussians on the basis of a set of training images of the scene to capture the scene geometry and appearance. Because rendering the Gaussians is much faster than rendering a neural field, 3D Gaussian Splatting is much faster than NeRFs and can capture a scene in a few minutes.",
+        "bbox": [
+            496,
+            542,
+            890,
+            690
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "While the Gaussians allow very realistic renderings of the scene, it is still however challenging to extract the surface of the scene from them: As shown in Figure 3, after optimization by 3D Gaussian Splatting, the Gaussians do not take an ordered structure in general and do not correspond well to the actual surface of the scene. In addition to the surface itself, it is also often desirable to represent the scene as a mesh, which remains the representation of choice in many pipelines: A mesh-based representation allows for powerful tools for editing, sculpting, animating, and relighting the scene. Because the Gaussians after Gaussian Splatting are unstructured, it is very challenging to extract a mesh from them. Note that this is also challenging with NeRFs albeit for different reasons.",
+        "bbox": [
+            496,
+            695,
+            892,
+            898
+        ],
+        "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,
+            44
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "5354",
+        "bbox": [
+            482,
+            943,
+            514,
+            955
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/965e4b706f98280301b6f99dcf08dc1d3cc1bda348d4aebf93eb5600c76ab7b7.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            89,
+            272,
+            170
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/63849b2eb6568bab393e1eb9db3e9606a0f34ca11c35142170ea5710115c3897.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            287,
+            89,
+            473,
+            170
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/7f0927d0b0e6405cd930616ebc9dd362849d6074393daefa14f02ea6d4638a45.jpg",
+        "image_caption": [
+            "Figure 2. Our algorithm can extract a highly detailed mesh from any 3D Gaussian Splatting scene [15] within minutes on a single GPU (top: Renderings of our meshes without texture, bottom: Renderings of the meshes with bound Gaussians)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            86,
+            172,
+            272,
+            253
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/414eed21c26dc56e63563eb1defede617594decabb74654cbaab4632ce6bb7ca.jpg",
+        "image_caption": [
+            "Figure 3. Extracting a mesh from Gaussians. Without regularization, the Gaussians have no special arrangement after optimization, which makes extracting a mesh very difficult. Without our regularization term, Marching Cubes fail to extract an acceptable mesh. With our regularization term, Marching Cubes recover an extremely noisy mesh even with a very fine 3D grid. Our scalable extraction method obtains a mesh even without our regularization term. Still, the mesh is noisy. By contrast, our full method succeeds in reconstructing an accurate mesh very efficiently."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            287,
+            172,
+            473,
+            253
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we first propose a regularization term that encourages the Gaussians to be well distributed over the scene surface so that the Gaussians capture much better the scene geometry, as shown in Figure 3. Our approach is to derive a volume density from the Gaussians under the assumption that the Gaussians are flat and well distributed over the scene surface. By minimizing the difference between this density and the actual one computed from the Gaussians during optimization, we encourage the 3D Gaussians to represent well the surface geometry.",
+        "bbox": [
+            75,
+            351,
+            467,
+            498
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Thanks to this regularization term, it becomes easier to extract a mesh from the Gaussians. In fact, since we introduce a density function to evaluate our regularization term, a natural approach would be to extract level sets of this density function. However, Gaussian Splatting performs densification in order to capture details of the scene with high fidelity, which results in a drastic increase in the number of Gaussians. Real scenes typically end up with one or several millions of 3D Gaussians with different scales and rotations, the majority of them being extremely small in order to reproduce texture and details in the scene. This results in a density function that is close to zero almost everywhere, and the Marching Cubes algorithm [21] fails to extract proper level sets of such a sparse density function even with a fine voxel grid, as also shown in Figure 3.",
+        "bbox": [
+            75,
+            500,
+            467,
+            719
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Instead, we introduce a method that very efficiently samples points on the visible part of a level set of the density function, allowing us to run the Poisson reconstruction algorithm [14] on these points to obtain a triangle mesh. This approach is scalable, by contrast with the Marching Cubes algorithm for example, and reconstructs a surface mesh within minutes on a single GPU, compared to other state of the art methods relying on Neural SDFs for extracting meshes from radiance fields, that require at least 24 hours on one GPU [20, 36, 38, 39] and rely on multiple GPUs to speed up the process [26].",
+        "bbox": [
+            75,
+            722,
+            467,
+            883
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As illustrated in Figures 2 and 4, our method produces",
+        "bbox": [
+            96,
+            885,
+            467,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/fab143a17035a4621c16f71a04dc960390996ce00a5ee09eef27a7f69788960d.jpg",
+        "image_caption": [
+            "without our regularization term"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            514,
+            101,
+            632,
+            165
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/62c6b2f08a971899191a5b562162f4a4a0d84c2fac41173d6dd4fdd4d21db07f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            637,
+            101,
+            756,
+            165
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/ad83aca945c2ef13508a04bdafb0e6aa0808eddaf7faabbba997f79174034c89.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            758,
+            101,
+            875,
+            165
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/4015513f1d9b4cc0061301924259fc7103dea1c2a531de2b078dd3f682cdbdbe.jpg",
+        "image_caption": [
+            "with our regularization term",
+            "zoom on Gaussians on a planar surface"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            514,
+            181,
+            632,
+            244
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/f9718805da1261a9b074b2a8f184814cf14f0acfe7762c7ef3b4057ab334615d.jpg",
+        "image_caption": [
+            "mesh with Marching Cubes"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            637,
+            181,
+            756,
+            244
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/38688d6186e8082e57213c850b4fedc5c366bbb99ad6bf40978505ff29e8c3eb.jpg",
+        "image_caption": [
+            "mesh with our extraction method"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            758,
+            181,
+            874,
+            244
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "high quality meshes. The challenge is in efficiently identifying points lying on the level set. To do this, we rely on the Gaussians depth maps seen from the training viewpoints. These depth maps can be obtained by extending the Gaussian Splatting rasterizer, and we show how to accurately sample points on the level set starting from these depth maps.",
+        "bbox": [
+            496,
+            436,
+            890,
+            539
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Finally, after extracting this mesh, we propose an optional refinement strategy that jointly optimizes the mesh and a set of 3D Gaussians through Gaussian splatting rendering only. This optimization enables high-quality rendering of the mesh using Gaussian splatting rendering rather than traditional textured mesh rendering. This results in higher performance in terms of rendering quality than other radiance field models relying on an underlying mesh at inference [6, 26, 39]. As shown in Figure 1, this makes possible the use of traditional mesh-editing tools for editing a Gaussian Splatting representation of a scene, offering endless possibilities for Computer Graphics.",
+        "bbox": [
+            496,
+            541,
+            890,
+            717
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To summarize, our contributions are:",
+        "bbox": [
+            517,
+            719,
+            764,
+            732
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- a regularization term that makes the Gaussians capture accurately the geometry of the scene;",
+            "- an efficient algorithm that extracts an accurate mesh from the Gaussians within minutes;",
+            "- a method to bind the Gaussians to the mesh, resulting in a more accurate mesh, higher rendering quality than state of the art methods using a mesh for Novel View Synthesis [6, 26, 39], and allowing editing the scene in many different ways."
+        ],
+        "bbox": [
+            500,
+            736,
+            890,
+            867
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We call our approach SuGaR. In the remainder of the paper, we discuss related work, give a brief overview of vanilla",
+        "bbox": [
+            500,
+            869,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "5355",
+        "bbox": [
+            482,
+            943,
+            514,
+            954
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3D Gaussian Splatting, describe SuGaR, and compare it to the state of the art.",
+        "bbox": [
+            76,
+            90,
+            468,
+            119
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            76,
+            137,
+            218,
+            152
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Image-based rendering (IBR) methods rely on a set of two-dimensional images of a scene to generate a representation of the scene and render novel views. The very first novel-view synthesis approaches were based on light fields [19], and developed the concept of volume rendering for novel views. Their work emphasized the importance of efficiently traversing volumetric data to produce realistic images.",
+        "bbox": [
+            75,
+            164,
+            467,
+            268
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Various scene representations have been proposed since, such as triangle meshes, point clouds, voxel grids, multiplane images, or neural implicit functions.",
+        "bbox": [
+            76,
+            270,
+            467,
+            314
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Traditional mesh-based IBR methods. Structure-from-motion (SfM) [32] and subsequent multi-view stereo (MVS) [10] allow for 3D reconstruction of surfaces, leading to the development of several view synthesis algorithms relying on triangle meshes as the primary 3D representation of scenes. Such algorithms consider textured triangles or warp and blend captured images on the mesh surface to generate novel views [4, 12, 37]. [29, 30] consider deep learning-based mesh representations for better view synthesis, bridging the gap between traditional graphics and modern machine learning techniques. While these mesh-based methods take advantage of existing graphics hardware and software for efficient rendering, they struggle with the capture of accurate geometry and appearance in complex regions.",
+        "bbox": [
+            75,
+            337,
+            467,
+            561
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Volumetric IBR methods. Volumetric methods use voxel grids, multiplane images, or neural networks to represent scenes as continuous volumetric functions of density and color. Recently, Neural Radiance Fields (NeRF) [22] introduced a novel scene representation based on a continuous volumetric function parameterized by a multilayer perceptron (MLP). NeRF produces photorealistic renderings with fine details and view-dependent effects, achieved through volumetric ray tracing. However, the original NeRF is computationally expensive and memory intensive.",
+        "bbox": [
+            75,
+            585,
+            467,
+            733
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To address these challenges, several works have improved NeRF's performance and scalability. These methods leverage discretized or sparse volumetric representations like voxel grids and hash tables as ways to store learnable features acting as positional encodings for 3D points [5, 13, 23, 34, 41], hierarchical sampling strategies [2, 11, 28, 40], or low-rank approximations [5]. However, they still rely on volumetric ray marching, which is incompatible with standard graphics hardware and software designed for rendering polygonal surfaces. Recent works have proposed modifying the NeRF's representation",
+        "bbox": [
+            75,
+            736,
+            467,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "of geometry and emitted radiance to allow for better reconstruction of specular materials [35] or relighting the scene through an explicit decomposition into material and lighting properties [3, 18, 33, 43].",
+        "bbox": [
+            498,
+            90,
+            890,
+            151
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Hybrid IBR methods. Some methods build on differentiable rendering to combine the advantages of mesh-based and volumetric methods, and allow for surface reconstruction as well as better editability. They use a hybrid volume-surface representation, which enables high-quality meshes suitable for downstream graphics applications while efficiently modeling view-dependent appearance. In particular, some works optimize neural signed distance functions (SDF) by training neural radiance fields in which the density is derived as a differentiable transformation of the SDF [7, 8, 20, 24, 36, 38]. A triangle mesh can finally be reconstructed from the SDF by applying the Marching Cubes algorithm [21]. However, most of these methods do not target real-time rendering.",
+        "bbox": [
+            496,
+            172,
+            890,
+            381
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Alternatively, other approaches \"bake\" the rendering capacity of an optimized NeRF or neural SDF into a much efficient structure relying on an underlying triangle mesh [6] that could benefit from the traditional triangle rasterization pipeline. In particular, the recent BakedSDF [39] reconstructs high quality meshes by optimizing a full neural SDF model, baking it into a high-resolution triangle mesh that combines mesh rendering for interpolating features and deep learning to translate these features into images, and finally optimizes a view-dependent appearance model.",
+        "bbox": [
+            496,
+            383,
+            890,
+            532
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "However, even though it achieves real-time rendering and produces impressive meshes of the surface of the scene, this model demands training a full neural SDF with an architecture identical to Mip-NeRF360 [1], which necessitates 48 hours of training.",
+        "bbox": [
+            496,
+            532,
+            890,
+            607
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Similarly, the recent method NeRFMeshing [26] proposes to also bake any NeRF model into a mesh structure, achieving real-time rendering. However, the meshing performed in this method lowers the quality of the rendering and results in a PSNR much lower than our method. Additionally, this method still requires training a full NeRF model beforehand, and needs approximately an hour of training on 8 V100 NVIDIA GPUs to allow for mesh training and extraction.",
+        "bbox": [
+            496,
+            608,
+            890,
+            742
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our method is much faster at retrieving a 3D mesh from 3D Gaussian Splatting, which is itself much faster than NeRFs. As our experiments show, our rendering done by bounding Gaussians to the mesh results in higher quality than previous solutions based on meshes.",
+        "bbox": [
+            496,
+            743,
+            890,
+            818
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Point-based IBR methods. Alternatively, point-based representations for radiance field excel at modeling thin geometry and leverage fast point rasterization pipelines to render images using  $\\alpha$ -blending rather than ray-marching [17,",
+        "bbox": [
+            496,
+            840,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "5356",
+        "bbox": [
+            482,
+            943,
+            514,
+            954
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "31]. In particular, the very recent 3D Gaussian Splitting model [15] allows for optimizing and rendering scenes with speed and quality never seen before.",
+        "bbox": [
+            75,
+            90,
+            468,
+            135
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3. 3D Gaussian Splitting",
+        "text_level": 1,
+        "bbox": [
+            76,
+            148,
+            290,
+            166
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For the sake of completeness, we briefly describe the original 3D Gaussian Splitting method here. The scene is represented as a (large) set of Gaussians, where each Gaussian  $g$  is represented by its mean  $\\mu_{g}$  and its covariance  $\\Sigma_{g}$  is parameterized by a scaling vector  $s_g\\in \\mathbb{R}^3$  and a quaternion  $q_{g}\\in \\mathbb{R}^{4}$  encoding the rotation of the Gaussian. In addition, each Gaussian is associated with its opacity  $\\alpha_{g}\\in [0,1]$  and a set of spherical harmonics coordinates describing the colors emitted by the Gaussian for all directions.",
+        "bbox": [
+            75,
+            172,
+            468,
+            306
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "An image of a set of Gaussians can be rendered from a given viewpoint thanks to a rasterizer. This rasterizer splats the 3D Gaussians into 2D Gaussians parallel to the image plane for rendering, which results in an extremely fast rendering process. This is the key component that makes 3D Gaussian Splating much faster than NeRFs, as it is much faster than the ray-marching compositing required in the optimization of NeRFs.",
+        "bbox": [
+            75,
+            308,
+            468,
+            426
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Given a set of images, the set of Gaussians is initialized from the point cloud produced by SfM [32]. The Gaussians' parameters (means, quaternions, scaling vectors, but also opacities and spherical harmonics parameters) are optimized to make the renderings of the Gaussians match the input images. During optimization, more Gaussians are added to better fit the scene's geometry. As a consequence, Gaussian Splatting generally produces scenes with millions of Gaussians that can be extremely small.",
+        "bbox": [
+            75,
+            428,
+            468,
+            561
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4. Method",
+        "text_level": 1,
+        "bbox": [
+            76,
+            574,
+            166,
+            589
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We present our SuGaR in this section:",
+        "bbox": [
+            76,
+            599,
+            331,
+            614
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- First, we detail our loss term that enforces the alignment of the 3D Gaussians with the surface of the scene during the optimization of Gaussian Splatting.",
+            "- We then detail our method that exploits this alignment for extracting a highly detailed mesh from the Gaussians within minutes on a single GPU.",
+            "- Finally, we describe our optional refinement strategy that jointly optimizes the mesh and 3D Gaussians located on the surface of the mesh using Gaussian Splatting rendering. This strategy results in a new set of Gaussians bound to an editable mesh."
+        ],
+        "bbox": [
+            76,
+            614,
+            468,
+            777
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1. Aligning the Gaussians with the Surface",
+        "text_level": 1,
+        "bbox": [
+            75,
+            787,
+            419,
+            803
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As discussed in the introduction, to facilitate the creation of a mesh from the Gaussians, we introduce a regularization term into the Gaussian Splitting optimization that encourages the Gaussians to be aligned with the surface of the scene and well distributed over this surface. Our approach is to derive an SDF from the Gaussians under the assumption",
+        "bbox": [
+            75,
+            811,
+            468,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "that the Gaussians have the desired properties. By minimizing the difference between this SDF and the actual SDF computed for the Gaussians, we encourage the Gaussians to have these properties.",
+        "bbox": [
+            496,
+            90,
+            890,
+            151
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For a given Gaussian Splitting scene, we start by considering the corresponding density function  $d:\\mathbb{R}^3\\to \\mathbb{R}_+$  computed as the sum of the Gaussian values weighted by their alpha-blending coefficients at any space location  $p$  ..",
+        "bbox": [
+            496,
+            162,
+            890,
+            222
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nd (p) = \\sum_ {g} \\alpha_ {g} \\exp \\left(- \\frac {1}{2} \\left(p - \\mu_ {g}\\right) ^ {T} \\Sigma_ {g} ^ {- 1} \\left(p - \\mu_ {g}\\right)\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            511,
+            232,
+            890,
+            268
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where the  $\\mu_g, \\Sigma_g$ , and  $\\alpha_g$  are the centers, covariances, and alpha-blending coefficients of the Gaussians, respectively. Let us consider what this density function becomes if the Gaussians are well distributed and aligned with the surface.",
+        "bbox": [
+            496,
+            281,
+            890,
+            340
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "First, in such scenario, the Gaussians would have limited overlap with their neighbors. As illustrated in Figure 3 (top-left), this is not the case in general. Then, for any point  $p \\in \\mathbb{R}^3$  close to the surface of the scene, the Gaussian  $g^{*}$  closest to the point  $p$  is likely to contribute much more than others to the density value  $d(p)$ . We could then approximate the Gaussian density at  $p$  by:",
+        "bbox": [
+            496,
+            351,
+            890,
+            455
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\alpha_ {g ^ {*}} \\exp \\left(- \\frac {1}{2} (p - \\mu_ {g ^ {*}}) ^ {T} \\Sigma_ {g ^ {*}} ^ {- 1} (p - \\mu_ {g ^ {*}})\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            549,
+            467,
+            890,
+            501
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where the \"closest Gaussian\"  $g^{*}$  is taken as the Gaussian with the largest contribution at point  $p$ :",
+        "bbox": [
+            496,
+            510,
+            890,
+            541
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ng ^ {*} = \\arg \\min  _ {g} \\left\\{\\left(p - \\mu_ {g}\\right) ^ {T} \\Sigma_ {g} ^ {- 1} \\left(p - \\mu_ {g}\\right) \\right\\}. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            553,
+            550,
+            890,
+            575
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Eq. (2) thus considers that the contribution of the closest Gaussian  $g^{*}$  to the density at  $p$  is much higher than the contribution of the other Gaussians. This will help us encourage the Gaussians to be well spread.",
+        "bbox": [
+            496,
+            585,
+            890,
+            645
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We also would like the 3D Gaussians to be flat, as they would then be aligned more closely with the surface of the mesh. Consequently, every Gaussian  $g$  would have one of its three scaling factors close to 0 and:",
+        "bbox": [
+            496,
+            656,
+            890,
+            715
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left(p - \\mu_ {g}\\right) ^ {T} \\Sigma_ {g} ^ {- 1} \\left(p - \\mu_ {g}\\right) \\approx \\frac {1}{s _ {g} ^ {2}} \\langle p - \\mu_ {g}, n _ {g} \\rangle^ {2}, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            545,
+            726,
+            890,
+            760
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $s_g$  the smallest scaling factor of the Gaussian and  $n_g$  the direction of the corresponding axis. Moreover, because we want Gaussians to describe the true surface of the scene, we need to avoid semi-transparent Gaussians. Therefore, we want Gaussians to be either opaque or fully transparent, in which case we can drop them for rendering. Consequently, we want to have  $\\alpha_g = 1$  for any Gaussian  $g$ .",
+        "bbox": [
+            496,
+            770,
+            890,
+            875
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In such scenario, the density of the Gaussians could fi-",
+        "bbox": [
+            517,
+            885,
+            890,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "5357",
+        "bbox": [
+            482,
+            943,
+            514,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/a56292c9d827b58edba578199500579430f2a6afc2342e78c183f2cabb75b7e6.jpg",
+        "image_caption": [
+            "(a) Mesh & Gaussians"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            88,
+            210,
+            263
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/0fcc44c7ccb311ac823d6d50fb0fc3578e6568e81d27799a90f6d5a308775a37.jpg",
+        "image_caption": [
+            "(b) Mesh (No Texture)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            212,
+            88,
+            344,
+            263
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/d0edac57d5ee016f3498c03b58541f242fc507dbc35d2b68fac779908d8555be.jpg",
+        "image_caption": [
+            "(c) Mesh normals"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            346,
+            88,
+            478,
+            263
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/c917145fa695713ee2573a050eaa62a4f8dc031b2f14f7df00ab8096b2a22b0e.jpg",
+        "image_caption": [
+            "(a) Mesh & Gaussians",
+            "Figure 4. Examples of (a) renderings and (b) reconstructed meshes with SuGaR. The (c) normal maps help visualize the geometry."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            491,
+            88,
+            622,
+            263
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/3cb31b2b075cc43fa4311d547582d177406bfe27cb552cf805412325a9c0c952.jpg",
+        "image_caption": [
+            "(b) Mesh (No Texture)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            625,
+            88,
+            756,
+            263
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/4a714642ab921d433fa94db5a38240ef13762090f89437b638b27ee572ab763f.jpg",
+        "image_caption": [
+            "(c) Mesh normals"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            759,
+            88,
+            890,
+            263
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "nally be approximated by density  $\\bar{d} (p)$  with:",
+        "bbox": [
+            76,
+            330,
+            372,
+            348
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\bar {d} (p) = \\exp \\left(- \\frac {1}{2 s _ {g ^ {*}} ^ {2}} \\langle p - \\mu_ {g ^ {*}}, n _ {g ^ {*}} \\rangle^ {2}\\right). \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            137,
+            357,
+            468,
+            393
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "A first strategy to enforce our regularization is to add term  $|d(p) - \\bar{d}(p)|$  to the optimization loss. While this approach works well to align Gaussians with the surface, we noticed that computing a slightly different loss relying on an SDF rather than on density further increases the alignment of Gaussians with the surface of the scene. For a given flat Gaussian, i.e.,  $s_g = 0$ , considering level sets is meaningless since all level sets would degenerate toward the plane passing through the center of the Gaussian  $\\mu_g$  with normal  $n_g$ . The distance between point  $p$  and the true surface of the scene would be approximately  $|\\langle p - \\mu_{g'}, n_{g'} \\rangle|$ , the distance from  $p$  to this plane. Consequently, the zero-crossings of the Signed Distance Function",
+        "bbox": [
+            75,
+            412,
+            468,
+            606
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\bar {f} (p) = \\pm s _ {g *} \\sqrt {- 2 \\log (\\bar {d} (p))} \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            171,
+            618,
+            468,
+            643
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "corresponds to the surface of the scene. More generally, we define",
+        "bbox": [
+            75,
+            652,
+            468,
+            680
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf (p) = \\pm s _ {g *} \\sqrt {- 2 \\log (d (p))} \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            174,
+            681,
+            468,
+            700
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "as the \"ideal\" distance function associated with the density function  $d$ . This distance function corresponds to the true surface of the scene in an ideal scenario where  $d = \\bar{d}$ . We therefore take our regularization term  $\\mathcal{R}$  as",
+        "bbox": [
+            75,
+            705,
+            468,
+            765
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {R} = \\frac {1}{| \\mathcal {P} |} \\sum_ {p \\in \\mathcal {P}} \\left| | \\hat {f} (p) | - | f (p) | \\right|, \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            161,
+            775,
+            468,
+            813
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "by sampling 3D points  $p$  and summing the differences at these points between the ideal SDF  $f(p)$  and an estimate  $\\hat{f}(p)$  of the SDF of the surface created by the current Gaussians.  $\\mathcal{P}$  refers to the set of sampled points. During optimization, we simply want the zero-level sets of both the",
+        "bbox": [
+            75,
+            823,
+            468,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/0fda687334d3dc4b93586b3ae1e96bd344c0ae51d878b5b833c5a193d5922480.jpg",
+        "image_caption": [
+            "Figure 5. Efficiently estimating  $\\hat{f}(p)$  of the SDF of the surface generated from Gaussians. We render depth maps of the Gaussians, sample points  $p$  in the viewpoint according to the distribution of the Gaussians. Value  $\\hat{f}(p)$  is taken as the 3D distance between  $p$  and the intersection between the line of sight for  $p$  and the depth map."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            588,
+            349,
+            799,
+            476
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "SDF estimator  $\\hat{f}$  and the ideal SDF  $f$  to align. Since a zero-level set is entirely determined by an unsigned distance function, we actually do not need to compute the sign of the SDFs and use absolute values.",
+        "bbox": [
+            496,
+            603,
+            890,
+            662
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Computing efficiently  $\\hat{f}(p)$  is a priori challenging. To do so, we propose to use the depth maps of the Gaussians from the viewpoints used for training—these depth maps can be rendered efficiently by extending the splatting rasterizer. Then, as shown in Figure 5, for a point  $p$  visible from a training viewpoint,  $\\hat{f}(p)$  is the difference between the depth of  $p$  and the depth in the corresponding depth map at the projection of  $p$ . Moreover, we sample points  $p$  following the distribution of the Gaussians:",
+        "bbox": [
+            496,
+            667,
+            890,
+            801
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np \\sim \\prod_ {g} \\mathcal {N} \\left(.; \\mu_ {g}, \\Sigma_ {g}\\right), \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            617,
+            811,
+            890,
+            844
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "with  $\\mathcal{N}(:,\\mu_g,\\Sigma_g)$  the Gaussian distribution of mean  $\\mu_{g}$  and covariance  $\\Sigma_{g}$  as these points are likely to correspond to a high gradient for  $\\mathcal{R}$ .",
+        "bbox": [
+            496,
+            854,
+            890,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "5358",
+        "bbox": [
+            482,
+            943,
+            514,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/7793310e8a1663a1bc425bb1121acfd80c404fd4cb0d52af2380bd5663ebf64a.jpg",
+        "image_caption": [
+            "Figure 6. Sampling points on a level set for Poisson reconstruction. Left: We sample points on the depth maps of the Gaussians and refine the point locations to move the points on the level set. Right: Comparison between the extracted mesh without (left) and with (right) our refinement step. Since splatted depth maps are not exact, using directly the depth points for reconstruction usually results in a large amount of noise and missing details."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            81,
+            93,
+            251,
+            224
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/d262f28448b4d3af275cf2b66155ba011da0fa7967794a9bd0b45cb786fba190.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            272,
+            92,
+            370,
+            222
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/2f706608e69c018da35d9debfa62abf0fe7d326101fcafab0e862679a53f0601.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            370,
+            92,
+            467,
+            222
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We also add a regularization term to encourage the normals of SDF  $f$  and the normals of SDF  $\\bar{f}$  to be similar:",
+        "bbox": [
+            76,
+            359,
+            468,
+            390
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {R} _ {\\text {N o r m}} = \\frac {1}{| \\mathcal {P} |} \\sum_ {p \\in \\mathcal {P}} \\left\\| \\frac {\\nabla f (p)}{\\| \\nabla f (p) \\| _ {2}} - n _ {g ^ {*}} \\right\\| _ {2} ^ {2}. \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            133,
+            400,
+            468,
+            441
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2. Efficient Mesh Extraction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            454,
+            310,
+            468
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To create a mesh from the Gaussians obtained after optimization using our regularization terms in Eq. (8) and Eq. (10), we sample 3D points on a level set of the density computed from the Gaussians. The level set depends on a level parameter  $\\lambda$ . Then, we obtain a mesh by simply running a Poisson reconstruction [14] on these points. Note that we can also easily assign the points with the normals of the SDF, which improves the mesh quality.",
+        "bbox": [
+            75,
+            479,
+            468,
+            598
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The challenge is in efficiently identifying points lying on the level set. For this, as shown in Figure 6, we again rely on the depth maps of the Gaussians as seen from the training viewpoints. We first randomly sample pixels from each depth map. For each pixel  $m$ , we sample its line of sight to find a 3D point on the level set. Formally, we sample  $n$  points  $p + t_i v$ , where  $p$  is the 3D point in the depth map that reprojects on pixel  $m$ ,  $v$  is the direction of the line of sight, and  $t_i \\in [-3\\sigma_g(v), 3\\sigma_g(v)]$  where  $\\sigma_g(v)$  is the standard deviation of the 3D Gaussian  $g$  in the direction of the camera. The interval  $[-3\\sigma_g(v), 3\\sigma_g(v)]$  is the confidence interval for the 99.7 confidence level of the 1D Gaussian function of  $t$  along the ray.",
+        "bbox": [
+            75,
+            599,
+            468,
+            794
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Then, we compute the density values  $d_{i} = d(p + t_{i}v)$  from Eq. (1) of these sampled points. If there exist  $i,j$  such that  $d_{i} < \\lambda < d_{j}$ , then there is a level set point located in this range. If so, we use linear interpolation to compute the coefficient  $t^{*}$  such that  $p + t^{*}v$  is the level set point closest to the camera, verifying  $d(p + t^{*}v) = \\lambda$ . We also compute the normals of the surface at points  $\\hat{p}$ , which we naturally",
+        "bbox": [
+            75,
+            796,
+            468,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5e1246092c67462f639fac33a618dd8fa91f5355a55569554ece5a95bc25dd3d.jpg",
+        "image_caption": [
+            "Figure 7. Joint refinement of mesh and Gaussians. Left: We bind Gaussians to the triangles of the mesh. Depending on the number of triangles in the scene, we bind a different number of Gaussians per triangle, with predefined barycentric coordinates. Right: Mesh before and after joint refinement."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            540,
+            88,
+            874,
+            241
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "define as the normalized analytical gradient of the density  $\\frac{\\nabla d(\\hat{p})}{\\|\\nabla d(\\hat{p})\\|_2}$ .",
+        "bbox": [
+            500,
+            361,
+            888,
+            397
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Finally, we apply Poisson reconstruction to reconstruct a surface mesh from the level set points and their normals.",
+        "bbox": [
+            500,
+            401,
+            890,
+            430
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3. Binding New 3D Gaussians to the Mesh",
+        "text_level": 1,
+        "bbox": [
+            500,
+            460,
+            836,
+            477
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Once we have extracted a first mesh, we can refine this mesh by binding new Gaussians to the mesh triangles and optimize the Gaussians and the mesh jointly using the Gaussian Splatting rasterizer. This enables the edition of the Gaussian splatting scene with popular mesh editing tools while keeping high-quality rendering thanks to the Gaussians.",
+        "bbox": [
+            498,
+            491,
+            890,
+            580
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Given the initial mesh, we instantiate new 3D Gaussians on the mesh. More exactly, we associate a set of  $n$  thin 3D Gaussians to each triangle of the mesh, sampled on the surface of the triangle, as illustrated in Figure 7. To do so, we slightly modify the structure of the original 3D Gaussian Splitting model.",
+        "bbox": [
+            498,
+            587,
+            890,
+            676
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We explicitly compute the means of the Gaussians from the mesh vertices using predefined barycentric coordinates in the corresponding triangles during optimization. Also, the Gaussians have only 2 learnable scaling factors instead of 3 and only 1 learnable 2D rotation encoded with a complex number rather than a quaternion, to keep the Gaussians flat and aligned with the mesh triangles. More details about this parameterisation are given in the supplementary material. Like the original model, we also optimize an opacity value and a set of spherical harmonics for every Gaussian to encode the color emitted in all directions.",
+        "bbox": [
+            496,
+            684,
+            890,
+            848
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Figure 7 shows an example of a mesh before and after refinement. Figure 1 and the supplementary material give examples of what can be done by editing the mesh.",
+        "bbox": [
+            498,
+            854,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "5359",
+        "bbox": [
+            482,
+            943,
+            514,
+            954
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5. Experiments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            90,
+            209,
+            107
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1. Implementation details",
+        "text_level": 1,
+        "bbox": [
+            76,
+            114,
+            290,
+            130
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "All our models are optimized on a single GPU Nvidia Tesla V100 SXM2 32 Go.",
+        "bbox": [
+            76,
+            137,
+            468,
+            167
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Regularization. For all scenes, we start by optimizing a Gaussian Splatting with no regularization for 7,000 iterations to let the 3D Gaussians position themselves without any additional constraint. Then, we perform 2,000 iterations with an additional entropy loss on the opacities  $\\alpha_{g}$  of the Gaussians, as a way to enforce them to become binary.",
+        "bbox": [
+            76,
+            186,
+            467,
+            276
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Finally, we remove Gaussians with opacity values under 0.5 and perform 6,000 iterations with the regularization term introduced in Subsection 4.1, which makes a total of 15,000 iterations. To compute the density values of points from a Gaussian  $g$ , we sum only the Gaussian functions from the 16 nearest Gaussians of  $g$  and update the list of nearest neighbors every 500 iterations. Optimization takes between 15 and 45 minutes depending on the scene.",
+        "bbox": [
+            75,
+            277,
+            467,
+            396
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Mesh extraction. For all experiments except the ablation presented in Table 2, we extract the  $\\lambda$ -level set of the density function for  $\\lambda = 0.3$ . We perform Poisson reconstruction with depth 10 and apply mesh simplification using quadric error metrics [9] to decrease the resolution of the meshes. Mesh extraction generally takes between 5 and 10 minutes depending on the scene.",
+        "bbox": [
+            75,
+            414,
+            467,
+            518
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Joint refinement. We jointly refine the mesh and the bound 3D Gaussians for either 2,000, 7,000 or 15,000 iterations. Depending on the number of iterations, the duration of refinement goes from a few minutes to an hour.",
+        "bbox": [
+            76,
+            537,
+            467,
+            598
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2. Real-Time Rendering of Real Scenes",
+        "text_level": 1,
+        "bbox": [
+            76,
+            607,
+            393,
+            625
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "For evaluating our model, we follow the approach from the original 3D Gaussian Splatting paper [15] and compare the performance of several variations of our method SuGaR after refinement on real 3D scenes from 3 different datasets: Mip-NeRF360 [1], DeepBlending [12] and Tanks&Temples [16]. We call R-SuGaR-NK a refined SuGaR model optimized for  $N$  iterations during refinement.",
+        "bbox": [
+            76,
+            631,
+            467,
+            734
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Following [15], we select the same sets of 2 scenes from Tanks&Temples (Truck and Train) and 2 scenes from DeepBlending (Playroom and Dr. Johnson). However, due to licensing issues and the unavailability of the scenes Flowers and Treehill, we perform the evaluation of all methods only on 7 scenes from Mip-NeRF360 instead of the full set of 9 scenes.",
+        "bbox": [
+            76,
+            736,
+            467,
+            838
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We compute the standard metrics PSNR, SSIM and LPIPS [44] to evaluate the quality of SuGaR's rendering using our extracted meshes and their bound surface Gaussians. Note that [6, 26, 39] also do not use plain textured",
+        "bbox": [
+            76,
+            840,
+            467,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "mesh rendering. We compare to several baselines, some of them focusing only on Novel View Synthesis [2, 15, 23, 41] and others relying on a reconstructed mesh [6, 26, 39], just like SuGaR. Results on the Mip-NeRF360 dataset are given in Table 1. Results on Tanks&Temple and DeepBlending are similar and can be found in the supplementary material.",
+        "bbox": [
+            496,
+            90,
+            890,
+            180
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Even though SuGaR focuses on aligning 3D Gaussians for reconstructing a high quality mesh during the first stage of its optimization, it significantly outperforms the state of the art methods for Novel View Synthesis using a mesh and reaches better performance than several famous models that focus only on rendering, such as I-NGP [23] and Plenoxels [41]. This performance is remarkable as SuGaR is able to extract a mesh significantly faster than other methods.",
+        "bbox": [
+            496,
+            181,
+            890,
+            301
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Moreover, SuGaR even reaches performance similar to state-of-the-art models for rendering quality [2, 15] on some of the scenes used for evaluation. Two main reasons explain this performance. First, the mesh extracted after the first stage of optimization serves as an excellent initialization for positioning Gaussians when starting the refinement phase. Then, the Gaussians constrained to remain on the surface during refinement greatly increase the rendering quality as they play the role of an efficient texturing tool and help reconstructing very fine details missing in the extracted mesh. Additional qualitative results are available in Figure 4.",
+        "bbox": [
+            496,
+            301,
+            890,
+            467
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3. Mesh Extraction",
+        "text_level": 1,
+        "bbox": [
+            500,
+            479,
+            663,
+            494
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To demonstrate the ability of our mesh extraction method for reconstructing high-quality meshes that are well-suited for view synthesis, we compare different mesh extraction algorithms. In particular, we optimize several variations of SuGaR by following the exact same pipeline as our standard model, except for the mesh extraction process: We either extract the mesh using a very fine marching cubes algorithm [21], by applying Poisson reconstruction [14] using the centers of the 3D Gaussians as the surface point cloud, or by applying our mesh extraction method on different level sets. Quantitative results are available in Table 2 and show the clear superiority of our approach for meshing 3D Gaussians. Figure 3 also illustrates how the marching cubes algorithm fails in this context.",
+        "bbox": [
+            496,
+            503,
+            890,
+            713
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.4. Mesh Rendering Ablation",
+        "text_level": 1,
+        "bbox": [
+            500,
+            726,
+            735,
+            742
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3 provides additional results to quantify how various parameters impact rendering performance. In particular, we evaluate how the resolution of the mesh extraction, i.e., the number of triangles, modifies the rendering quality. For fair comparison, we increase the number of surface-aligned Gaussians per triangle when we decrease the number of triangles. Results show that increasing the number of vertices increases the quality of rendering with surface Gaussians, but meshes with less triangles are already able to reach state of the art results.",
+        "bbox": [
+            496,
+            750,
+            890,
+            898
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "5360",
+        "bbox": [
+            482,
+            943,
+            514,
+            954
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/671fe6d63757a6e2852b35c099121a5965e771f4c7633e1e4ba7462c5e273efb.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"3\">Indoor scenes</td><td colspan=\"3\">Outdoor scenes</td><td colspan=\"3\">Average on all scenes</td></tr><tr><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td></tr><tr><td colspan=\"10\">No mesh (except SuGaR)</td></tr><tr><td>Plenoxels [42]</td><td>24.83</td><td>0.766</td><td>0.426</td><td>22.02</td><td>0.542</td><td>0.465</td><td>23.62</td><td>0.670</td><td>0.443</td></tr><tr><td>INGP-Base [23]</td><td>28.65</td><td>0.840</td><td>0.281</td><td>23.47</td><td>0.571</td><td>0.416</td><td>26.43</td><td>0.725</td><td>0.339</td></tr><tr><td>INGP-Big [23]</td><td>29.14</td><td>0.863</td><td>0.242</td><td>23.57</td><td>0.602</td><td>0.375</td><td>26.75</td><td>0.751</td><td>0.299</td></tr><tr><td>Mip-NeRF360 [2]</td><td>31.58</td><td>0.914</td><td>0.182</td><td>25.79</td><td>0.746</td><td>0.247</td><td>29.09</td><td>0.842</td><td>0.210</td></tr><tr><td>3DGS [15]</td><td>30.41</td><td>0.920</td><td>0.189</td><td>26.40</td><td>0.805</td><td>0.173</td><td>28.69</td><td>0.870</td><td>0.182</td></tr><tr><td>R-SuGaR-15K (Ours)</td><td>29.43</td><td>0.910</td><td>0.216</td><td>24.40</td><td>0.699</td><td>0.301</td><td>27.27</td><td>0.820</td><td>0.253</td></tr><tr><td colspan=\"10\">With mesh</td></tr><tr><td>Mobile-NeRF [6]</td><td>-</td><td>-</td><td>-</td><td>21.95</td><td>0.470</td><td>0.470</td><td>-</td><td>-</td><td>-</td></tr><tr><td>NeRFMeshing [26]</td><td>23.83</td><td>-</td><td>-</td><td>22.23</td><td>-</td><td>-</td><td>23.15</td><td>-</td><td>-</td></tr><tr><td>BakedSDF [39]</td><td>27.06</td><td>0.836</td><td>0.258</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>R-SuGaR-2K (Ours)</td><td>26.29</td><td>0.872</td><td>0.262</td><td>22.97</td><td>0.648</td><td>0.360</td><td>24.87</td><td>0.776</td><td>0.304</td></tr><tr><td>R-SuGaR-7K (Ours)</td><td>28.73</td><td>0.904</td><td>0.226</td><td>24.16</td><td>0.691</td><td>0.313</td><td>26.77</td><td>0.813</td><td>0.263</td></tr><tr><td>R-SuGaR-15K (Ours)</td><td>29.43</td><td>0.910</td><td>0.216</td><td>24.40</td><td>0.699</td><td>0.301</td><td>27.27</td><td>0.820</td><td>0.253</td></tr></table>",
+        "bbox": [
+            107,
+            88,
+            857,
+            351
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/64196f0b1efc41c3b79491cb9c860df5e66b20cb9461243b7e45ba98b9b2501a.jpg",
+        "table_caption": [
+            "Table 1. Quantitative evaluation of rendering quality on the Mip-NeRF360 dataset [2]. SuGaR is best among the methods that recover a mesh, and still performs well compared to NeRF methods and vanilla 3D Gaussian Splatting."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Extraction method</td><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td></tr><tr><td>Marching Cubes [21]</td><td>23.91</td><td>0.703</td><td>0.392</td></tr><tr><td>Poisson (centers) [14]</td><td>23.76</td><td>0.756</td><td>0.340</td></tr><tr><td>Ours (Surface level 0.1)</td><td>24.62</td><td>0.765</td><td>0.313</td></tr><tr><td>Ours (Surface level 0.3)</td><td>24.87</td><td>0.776</td><td>0.304</td></tr><tr><td>Ours (Surface level 0.5)</td><td>24.91</td><td>0.777</td><td>0.304</td></tr></table>",
+        "bbox": [
+            91,
+            414,
+            450,
+            513
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/1d39f4556690dd0181417cf11f5b127f9c425cf3c7ba4eda4b8cdf0933ee8866.jpg",
+        "table_caption": [
+            "Table 2. Ablation for different mesh extraction methods on the Mip-NeRF360 dataset [2] after applying our regularization term. For 'Poisson (centers)', we apply Poisson reconstruction [14] using as surface points the centers of the 3D Gaussians. For fair comparison, we calibrate the methods to enforce all extracted meshes to have approximately 1,000,000 vertices."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td></tr><tr><td>1M vertices (3DGS)</td><td>24.51</td><td>0.768</td><td>0.295</td></tr><tr><td>1M vertices (UV)</td><td>21.24</td><td>0.609</td><td>0.478</td></tr><tr><td>200K vertices (3DGS)</td><td>24.24</td><td>0.757</td><td>0.300</td></tr><tr><td>200K vertices (UV)</td><td>21.44</td><td>0.656</td><td>0.419</td></tr></table>",
+        "bbox": [
+            96,
+            625,
+            444,
+            717
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Comparison between surface-aligned 3D Gaussians and an optimized traditional UV texture on the Mip-NeRF360 dataset [2]. For fair comparison, we only use the diffuse spherical harmonics component when rendering images with SuGaR. Using 3D Gaussians bound to the mesh greatly improves rendering quality, even though it contains less parameters than the UV texture.",
+        "bbox": [
+            75,
+            727,
+            468,
+            811
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Then, we illustrate the benefits of using Gaussians aligned on the surface as a texturing tool for rendering meshes. To this end, we also optimize traditional UV textures on our meshes using differentiable mesh rendering",
+        "bbox": [
+            75,
+            840,
+            470,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "with traditional triangle rasterization. Even though rendering with surface-aligned Gaussians provides better performance, rendering our meshes with traditional UV textures still produces satisfying results, which further illustrates the quality of our extracted meshes. Qualitative comparisons are provided in the supplementary material.",
+        "bbox": [
+            498,
+            416,
+            890,
+            508
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            531,
+            619,
+            546
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We proposed a very fast algorithm to obtain an accurate 3D triangle mesh for a scene via Gaussian Splatting. Moreover, by combining meshing and Gaussian Splatting, we make possible intuitive manipulation of the captured scenes and realistic rendering, offering new possibilities for creators. SuGaR does not come without limitations: Gaussians do tend to \"cheat\" on the geometry and depth by creating cavities to reproduce specular effects, instead of relying on spherical harmonics. SuGaR's regularization mitigates this issue, but Gaussians can still distort the surface in regions with intense specularity. In addition, SuGaR's assumption that the scene can be represented as a surface complicates the rendering of volumetric effects or fuzzy materials.",
+        "bbox": [
+            496,
+            559,
+            890,
+            757
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements.",
+        "text_level": 1,
+        "bbox": [
+            500,
+            781,
+            669,
+            797
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This project was funded in part by the European Union (ERC Advanced Grant explorer Funding ID #101097259). This work was granted access to the HPC resources of IDRIS under the allocation 2023-AD011013387R1 made by GENCI. We thank George Drettakis and Elliot Vincent for inspiring discussions and valuable feedback.",
+        "bbox": [
+            496,
+            810,
+            893,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "5361",
+        "bbox": [
+            482,
+            943,
+            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] Jonathan T. Barron. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In International Conference on Computer Vision, 2021. 3, 7",
+            "[2] Jonathan T. Barron. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 7, 8, 2, 4",
+            "[3] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural Reflectance Decomposition from Image Collections. In International Conference on Computer Vision, 2021. 3",
+            "[4] Chris Buehler, Michael Bosse, Leonard Mcmillan, Steven Gortler, and Michael Cohen. Unstructured Lumigraph Rendering. In ACM SIGGRAPH, 2001. 3",
+            "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial Radiance Fields. In European Conference on Computer Vision, 2022. 3",
+            "[6] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. MobileNeRF: Exploiting the Polygon Rasterization Pipeline for Efficient Neural Field Rendering on Mobile Architectures. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 7, 8",
+            "[7] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. NeuMesh: Learning Disentangled Neural Mesh-Based Implicit Field for Geometry and Texture Editing. In European Conference on Computer Vision, 2022. 3",
+            "[8] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving Neural Implicit Surfaces Geometry with Patch Warping. In Conference on Computer Vision and Pattern Recognition, 2022. 3",
+            "[9] Michael Garland and Paul S. Heckbert. Surface Simplification Using Quadric Error Metrics. In ACM SIGGRAPH, 1997. 7",
+            "[10] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven Seitz. Multi-View Stereo for Community Photo Collections. In International Conference on Computer Vision, 2007. 3",
+            "[11] Peter Hedman and Pratul P. Srinivasan. Baking Neural Radiance Fields for Real-Time View Synthesis. In International Conference on Computer Vision, 2021. 3",
+            "[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. In ACM SIGGRAPH, 2018. 3, 7, 2, 4",
+            "[13] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. ReLU Fields: The Little Non-Linearity That Could. In ACM SIGGRAPH, 2022. 3",
+            "[14] Michael M. Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson Surface Reconstruction. In Eurographics, 2006. 2, 6, 7, 8",
+            "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. In ACM SIGGRAPH, 2023. 1, 2, 4, 7, 8"
+        ],
+        "bbox": [
+            78,
+            114,
+            468,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[16] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. In ACM SIGGRAPH, 2017. 7, 2, 4",
+            "[17] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-Based Neural Rendering with Per-View Optimization. In Computer Graphics Forum, 2021. 3",
+            "[18] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: Neural Rendering of Objects from Online Image Collections. In ACM SIGGRAPH, 2022. 3",
+            "[19] Marc Levoy and Pat Hanrahan. Light Field Rendering. In ACM SIGGRAPH, 1996. 3",
+            "[20] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-Fidelity Neural Surface Reconstruction. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 3",
+            "[21] William E. Lorensen and Harvey E. Cline. Marching Cubes: A High Resolution 3D Surface Construction Algorithm. In ACM SIGGRAPH, 1987. 2, 3, 7, 8",
+            "[22] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, 2020. 1, 3",
+            "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. In ACM SIGGRAPH, 2022. 3, 7, 8, 2",
+            "[24] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying Neural Implicit Surfaces and Radiance Fields for Multi-View Reconstruction. In International Conference on Computer Vision, 2021. 3",
+            "[25] 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. 2",
+            "[26] Marie-Julie Rakotosaona, Fabian Manhardt, Diego Martin Arroyo, Michael Niemeyer, Abhijit Kundu, and Federico Tombari. NeRFMeshing: Distilling Neural Radiance Fields into Geometrically-Accurate 3D Meshes. In DV, 2023. 2, 3, 7, 8",
+            "[27] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D Deep Learning with PyTorch3D. In arXiv Preprint, 2020. 2",
+            "[28] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding Up Neural Radiance Fields with Thousands of Tiny MLPs. In International Conference on Computer Vision, 2021. 3",
+            "[29] Gernot Riegler and Vladlen Koltun. Free View Synthesis. In European Conference on Computer Vision, 2020. 3"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "5362",
+        "bbox": [
+            482,
+            943,
+            514,
+            954
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[30] Gernot Riegler and Vladlen Koltun. Stable View Synthesis. In Conference on Computer Vision and Pattern Recognition, 2021. 3",
+            "[31] Darius Rückert, Linus Franke, and Marc Stamminger. ADOP: Approximate Differentiable One-Pixel Point Rendering. In ACM SIGGRAPH, 2022. 4",
+            "[32] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo Tourism: Exploring Photo Collections in 3D. In ACM SIGGRAPH, 2006. 3, 4",
+            "[33] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural Reflectance and Visibility Fields for Relighting and View Synthesis. In Conference on Computer Vision and Pattern Recognition, 2021. 3",
+            "[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct Voxel Grid Optimization: Super-Fast Convergence for Radiance Fields Reconstruction. In Conference on Computer Vision and Pattern Recognition, 2022. 3",
+            "[35] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields. In Conference on Computer Vision and Pattern Recognition, 2022. 3",
+            "[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-View Reconstruction. In Advances in Neural Information Processing Systems, 2021. 2, 3",
+            "[37] Daniel N. Wood, Daniel I. Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H. Salesin, and Werner Stuetzle. Surface Light Fields for 3D Photography. In ACM SIGGRAPH, 2000. 3",
+            "[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume Rendering of Neural Implicit Surfaces. In Advances in Neural Information Processing Systems, 2021. 2, 3",
+            "[39] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, and Jonathan T. Barron. BakedSDF: Meshing Neural SDFs for Real-Time View Synthesis. In ACM SIGGRAPH, 2023. 2, 3, 7, 8",
+            "[40] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees For Real-Time Rendering of Neural Radiance Fields. In International Conference on Computer Vision, 2021. 3",
+            "[41] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 7",
+            "[42] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2022. 8, 2",
+            "[43] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse Rendering with Spherical Gaussians for Physics-Based Material Editing and Relighting. In Conference on Computer Vision and Pattern Recognition, 2021. 3",
+            "[44] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            900
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Deep Features as a Perceptual Metric. In Conference on Computer Vision and Pattern Recognition, 2018. 7, 2, 4",
+        "bbox": [
+            535,
+            92,
+            890,
+            119
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "5363",
+        "bbox": [
+            482,
+            943,
+            514,
+            954
+        ],
+        "page_idx": 9
+    }
+]

2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/8c37e125-7381-4a71-8f2b-015c3d8a066a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95cb88bfe88ab982bb3b46b3764cc4aff0d828ccde2be668f05551ea56ca40f3
+size 6804928

2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/full.md

@@ -0,0 +1,360 @@
+# SuGaR: Surface-Aligned Gaussian Splitting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering
+
+Antoine Guédon Vincent Lepetit  
+LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France  
+https://anttwo.github.io/sugar/
+
+![](images/6673930ff670e4525dde86ea891322dc4edf5ae835eac2957ca13ae2f137f242.jpg)  
+Figure 1. We introduce a method that extracts accurate and editable meshes from 3D Gaussian Splitting representations within minutes on a single GPU. The meshes can be edited, animated, composited, etc. with very realistic Gaussian Splitting rendering, offering new possibilities for Computer Graphics. Note for example that we changed the posture of the robot between the captured scene on the bottom left and the composited scene on the right. The supplementary material provides more examples, including a video illustrating our results.
+
+![](images/1e71b34782521a295dba22014fb33764264018e29489fec3a41da095d26a740b.jpg)
+
+![](images/10648b5890419237a4fb6653f09ebb64111dc67816d3d828c0fbbb2c4b6de2cf.jpg)
+
+# Abstract
+
+We propose a method to allow precise and extremely fast mesh extraction from 3D Gaussian Splatting [15]. Gaussian Splatting has recently become very popular as it yields realistic rendering while being significantly faster to train than NeRFs. It is however challenging to extract a mesh from the millions of tiny 3D Gaussians as these Gaussians tend to be unorganized after optimization and no method has been proposed so far. Our first key contribution is a regularization term that encourages the Gaussians to align well with the surface of the scene. We then introduce a method that exploits this alignment to extract a mesh from the Gaussians using Poisson reconstruction, which is fast, scalable, and preserves details, in contrast to the Marching Cubes algorithm usually applied to extract meshes from Neural SDFs. Finally, we introduce an optional refinement strategy that binds Gaussians to the surface of the mesh, and jointly optimizes these Gaussians and the mesh through Gaussian splatting rendering. This enables easy editing, sculpting, animating, and relighting of the Gaussians by manipulating the mesh instead of the Gaussians themselves. Retrieving such an editable mesh for realistic rendering is done within minutes with our method, compared to hours with the state-of-the-art method on SDFs, while providing a better rendering quality.
+
+# 1. Introduction
+
+After NeRFs [22], 3D Gaussian Splatting [15] has recently become very popular for capturing a 3D scene and rendering it from novel points of view. 3D Gaussian Splatting optimizes the positions, orientations, appearances (represented as spherical harmonics), and alpha blending of many tiny 3D Gaussians on the basis of a set of training images of the scene to capture the scene geometry and appearance. Because rendering the Gaussians is much faster than rendering a neural field, 3D Gaussian Splatting is much faster than NeRFs and can capture a scene in a few minutes.
+
+While the Gaussians allow very realistic renderings of the scene, it is still however challenging to extract the surface of the scene from them: As shown in Figure 3, after optimization by 3D Gaussian Splatting, the Gaussians do not take an ordered structure in general and do not correspond well to the actual surface of the scene. In addition to the surface itself, it is also often desirable to represent the scene as a mesh, which remains the representation of choice in many pipelines: A mesh-based representation allows for powerful tools for editing, sculpting, animating, and relighting the scene. Because the Gaussians after Gaussian Splatting are unstructured, it is very challenging to extract a mesh from them. Note that this is also challenging with NeRFs albeit for different reasons.
+
+![](images/965e4b706f98280301b6f99dcf08dc1d3cc1bda348d4aebf93eb5600c76ab7b7.jpg)
+
+![](images/63849b2eb6568bab393e1eb9db3e9606a0f34ca11c35142170ea5710115c3897.jpg)
+
+![](images/7f0927d0b0e6405cd930616ebc9dd362849d6074393daefa14f02ea6d4638a45.jpg)  
+Figure 2. Our algorithm can extract a highly detailed mesh from any 3D Gaussian Splatting scene [15] within minutes on a single GPU (top: Renderings of our meshes without texture, bottom: Renderings of the meshes with bound Gaussians).
+
+![](images/414eed21c26dc56e63563eb1defede617594decabb74654cbaab4632ce6bb7ca.jpg)  
+Figure 3. Extracting a mesh from Gaussians. Without regularization, the Gaussians have no special arrangement after optimization, which makes extracting a mesh very difficult. Without our regularization term, Marching Cubes fail to extract an acceptable mesh. With our regularization term, Marching Cubes recover an extremely noisy mesh even with a very fine 3D grid. Our scalable extraction method obtains a mesh even without our regularization term. Still, the mesh is noisy. By contrast, our full method succeeds in reconstructing an accurate mesh very efficiently.
+
+In this paper, we first propose a regularization term that encourages the Gaussians to be well distributed over the scene surface so that the Gaussians capture much better the scene geometry, as shown in Figure 3. Our approach is to derive a volume density from the Gaussians under the assumption that the Gaussians are flat and well distributed over the scene surface. By minimizing the difference between this density and the actual one computed from the Gaussians during optimization, we encourage the 3D Gaussians to represent well the surface geometry.
+
+Thanks to this regularization term, it becomes easier to extract a mesh from the Gaussians. In fact, since we introduce a density function to evaluate our regularization term, a natural approach would be to extract level sets of this density function. However, Gaussian Splatting performs densification in order to capture details of the scene with high fidelity, which results in a drastic increase in the number of Gaussians. Real scenes typically end up with one or several millions of 3D Gaussians with different scales and rotations, the majority of them being extremely small in order to reproduce texture and details in the scene. This results in a density function that is close to zero almost everywhere, and the Marching Cubes algorithm [21] fails to extract proper level sets of such a sparse density function even with a fine voxel grid, as also shown in Figure 3.
+
+Instead, we introduce a method that very efficiently samples points on the visible part of a level set of the density function, allowing us to run the Poisson reconstruction algorithm [14] on these points to obtain a triangle mesh. This approach is scalable, by contrast with the Marching Cubes algorithm for example, and reconstructs a surface mesh within minutes on a single GPU, compared to other state of the art methods relying on Neural SDFs for extracting meshes from radiance fields, that require at least 24 hours on one GPU [20, 36, 38, 39] and rely on multiple GPUs to speed up the process [26].
+
+As illustrated in Figures 2 and 4, our method produces
+
+![](images/fab143a17035a4621c16f71a04dc960390996ce00a5ee09eef27a7f69788960d.jpg)  
+without our regularization term
+
+![](images/62c6b2f08a971899191a5b562162f4a4a0d84c2fac41173d6dd4fdd4d21db07f.jpg)
+
+![](images/ad83aca945c2ef13508a04bdafb0e6aa0808eddaf7faabbba997f79174034c89.jpg)
+
+![](images/4015513f1d9b4cc0061301924259fc7103dea1c2a531de2b078dd3f682cdbdbe.jpg)  
+with our regularization term  
+zoom on Gaussians on a planar surface
+
+![](images/f9718805da1261a9b074b2a8f184814cf14f0acfe7762c7ef3b4057ab334615d.jpg)  
+mesh with Marching Cubes
+
+![](images/38688d6186e8082e57213c850b4fedc5c366bbb99ad6bf40978505ff29e8c3eb.jpg)  
+mesh with our extraction method
+
+high quality meshes. The challenge is in efficiently identifying points lying on the level set. To do this, we rely on the Gaussians depth maps seen from the training viewpoints. These depth maps can be obtained by extending the Gaussian Splatting rasterizer, and we show how to accurately sample points on the level set starting from these depth maps.
+
+Finally, after extracting this mesh, we propose an optional refinement strategy that jointly optimizes the mesh and a set of 3D Gaussians through Gaussian splatting rendering only. This optimization enables high-quality rendering of the mesh using Gaussian splatting rendering rather than traditional textured mesh rendering. This results in higher performance in terms of rendering quality than other radiance field models relying on an underlying mesh at inference [6, 26, 39]. As shown in Figure 1, this makes possible the use of traditional mesh-editing tools for editing a Gaussian Splatting representation of a scene, offering endless possibilities for Computer Graphics.
+
+To summarize, our contributions are:
+
+- a regularization term that makes the Gaussians capture accurately the geometry of the scene;  
+- an efficient algorithm that extracts an accurate mesh from the Gaussians within minutes;  
+- a method to bind the Gaussians to the mesh, resulting in a more accurate mesh, higher rendering quality than state of the art methods using a mesh for Novel View Synthesis [6, 26, 39], and allowing editing the scene in many different ways.
+
+We call our approach SuGaR. In the remainder of the paper, we discuss related work, give a brief overview of vanilla
+
+3D Gaussian Splatting, describe SuGaR, and compare it to the state of the art.
+
+# 2. Related Work
+
+Image-based rendering (IBR) methods rely on a set of two-dimensional images of a scene to generate a representation of the scene and render novel views. The very first novel-view synthesis approaches were based on light fields [19], and developed the concept of volume rendering for novel views. Their work emphasized the importance of efficiently traversing volumetric data to produce realistic images.
+
+Various scene representations have been proposed since, such as triangle meshes, point clouds, voxel grids, multiplane images, or neural implicit functions.
+
+Traditional mesh-based IBR methods. Structure-from-motion (SfM) [32] and subsequent multi-view stereo (MVS) [10] allow for 3D reconstruction of surfaces, leading to the development of several view synthesis algorithms relying on triangle meshes as the primary 3D representation of scenes. Such algorithms consider textured triangles or warp and blend captured images on the mesh surface to generate novel views [4, 12, 37]. [29, 30] consider deep learning-based mesh representations for better view synthesis, bridging the gap between traditional graphics and modern machine learning techniques. While these mesh-based methods take advantage of existing graphics hardware and software for efficient rendering, they struggle with the capture of accurate geometry and appearance in complex regions.
+
+Volumetric IBR methods. Volumetric methods use voxel grids, multiplane images, or neural networks to represent scenes as continuous volumetric functions of density and color. Recently, Neural Radiance Fields (NeRF) [22] introduced a novel scene representation based on a continuous volumetric function parameterized by a multilayer perceptron (MLP). NeRF produces photorealistic renderings with fine details and view-dependent effects, achieved through volumetric ray tracing. However, the original NeRF is computationally expensive and memory intensive.
+
+To address these challenges, several works have improved NeRF's performance and scalability. These methods leverage discretized or sparse volumetric representations like voxel grids and hash tables as ways to store learnable features acting as positional encodings for 3D points [5, 13, 23, 34, 41], hierarchical sampling strategies [2, 11, 28, 40], or low-rank approximations [5]. However, they still rely on volumetric ray marching, which is incompatible with standard graphics hardware and software designed for rendering polygonal surfaces. Recent works have proposed modifying the NeRF's representation
+
+of geometry and emitted radiance to allow for better reconstruction of specular materials [35] or relighting the scene through an explicit decomposition into material and lighting properties [3, 18, 33, 43].
+
+Hybrid IBR methods. Some methods build on differentiable rendering to combine the advantages of mesh-based and volumetric methods, and allow for surface reconstruction as well as better editability. They use a hybrid volume-surface representation, which enables high-quality meshes suitable for downstream graphics applications while efficiently modeling view-dependent appearance. In particular, some works optimize neural signed distance functions (SDF) by training neural radiance fields in which the density is derived as a differentiable transformation of the SDF [7, 8, 20, 24, 36, 38]. A triangle mesh can finally be reconstructed from the SDF by applying the Marching Cubes algorithm [21]. However, most of these methods do not target real-time rendering.
+
+Alternatively, other approaches "bake" the rendering capacity of an optimized NeRF or neural SDF into a much efficient structure relying on an underlying triangle mesh [6] that could benefit from the traditional triangle rasterization pipeline. In particular, the recent BakedSDF [39] reconstructs high quality meshes by optimizing a full neural SDF model, baking it into a high-resolution triangle mesh that combines mesh rendering for interpolating features and deep learning to translate these features into images, and finally optimizes a view-dependent appearance model.
+
+However, even though it achieves real-time rendering and produces impressive meshes of the surface of the scene, this model demands training a full neural SDF with an architecture identical to Mip-NeRF360 [1], which necessitates 48 hours of training.
+
+Similarly, the recent method NeRFMeshing [26] proposes to also bake any NeRF model into a mesh structure, achieving real-time rendering. However, the meshing performed in this method lowers the quality of the rendering and results in a PSNR much lower than our method. Additionally, this method still requires training a full NeRF model beforehand, and needs approximately an hour of training on 8 V100 NVIDIA GPUs to allow for mesh training and extraction.
+
+Our method is much faster at retrieving a 3D mesh from 3D Gaussian Splatting, which is itself much faster than NeRFs. As our experiments show, our rendering done by bounding Gaussians to the mesh results in higher quality than previous solutions based on meshes.
+
+Point-based IBR methods. Alternatively, point-based representations for radiance field excel at modeling thin geometry and leverage fast point rasterization pipelines to render images using  $\alpha$ -blending rather than ray-marching [17,
+
+31]. In particular, the very recent 3D Gaussian Splitting model [15] allows for optimizing and rendering scenes with speed and quality never seen before.
+
+# 3. 3D Gaussian Splitting
+
+For the sake of completeness, we briefly describe the original 3D Gaussian Splitting method here. The scene is represented as a (large) set of Gaussians, where each Gaussian  $g$  is represented by its mean  $\mu_{g}$  and its covariance  $\Sigma_{g}$  is parameterized by a scaling vector  $s_g\in \mathbb{R}^3$  and a quaternion  $q_{g}\in \mathbb{R}^{4}$  encoding the rotation of the Gaussian. In addition, each Gaussian is associated with its opacity  $\alpha_{g}\in [0,1]$  and a set of spherical harmonics coordinates describing the colors emitted by the Gaussian for all directions.
+
+An image of a set of Gaussians can be rendered from a given viewpoint thanks to a rasterizer. This rasterizer splats the 3D Gaussians into 2D Gaussians parallel to the image plane for rendering, which results in an extremely fast rendering process. This is the key component that makes 3D Gaussian Splating much faster than NeRFs, as it is much faster than the ray-marching compositing required in the optimization of NeRFs.
+
+Given a set of images, the set of Gaussians is initialized from the point cloud produced by SfM [32]. The Gaussians' parameters (means, quaternions, scaling vectors, but also opacities and spherical harmonics parameters) are optimized to make the renderings of the Gaussians match the input images. During optimization, more Gaussians are added to better fit the scene's geometry. As a consequence, Gaussian Splatting generally produces scenes with millions of Gaussians that can be extremely small.
+
+# 4. Method
+
+We present our SuGaR in this section:
+
+- First, we detail our loss term that enforces the alignment of the 3D Gaussians with the surface of the scene during the optimization of Gaussian Splatting.  
+- We then detail our method that exploits this alignment for extracting a highly detailed mesh from the Gaussians within minutes on a single GPU.  
+- Finally, we describe our optional refinement strategy that jointly optimizes the mesh and 3D Gaussians located on the surface of the mesh using Gaussian Splatting rendering. This strategy results in a new set of Gaussians bound to an editable mesh.
+
+# 4.1. Aligning the Gaussians with the Surface
+
+As discussed in the introduction, to facilitate the creation of a mesh from the Gaussians, we introduce a regularization term into the Gaussian Splitting optimization that encourages the Gaussians to be aligned with the surface of the scene and well distributed over this surface. Our approach is to derive an SDF from the Gaussians under the assumption
+
+that the Gaussians have the desired properties. By minimizing the difference between this SDF and the actual SDF computed for the Gaussians, we encourage the Gaussians to have these properties.
+
+For a given Gaussian Splitting scene, we start by considering the corresponding density function  $d:\mathbb{R}^3\to \mathbb{R}_+$  computed as the sum of the Gaussian values weighted by their alpha-blending coefficients at any space location  $p$  ..
+
+$$
+d (p) = \sum_ {g} \alpha_ {g} \exp \left(- \frac {1}{2} \left(p - \mu_ {g}\right) ^ {T} \Sigma_ {g} ^ {- 1} \left(p - \mu_ {g}\right)\right), \tag {1}
+$$
+
+where the  $\mu_g, \Sigma_g$ , and  $\alpha_g$  are the centers, covariances, and alpha-blending coefficients of the Gaussians, respectively. Let us consider what this density function becomes if the Gaussians are well distributed and aligned with the surface.
+
+First, in such scenario, the Gaussians would have limited overlap with their neighbors. As illustrated in Figure 3 (top-left), this is not the case in general. Then, for any point  $p \in \mathbb{R}^3$  close to the surface of the scene, the Gaussian  $g^{*}$  closest to the point  $p$  is likely to contribute much more than others to the density value  $d(p)$ . We could then approximate the Gaussian density at  $p$  by:
+
+$$
+\alpha_ {g ^ {*}} \exp \left(- \frac {1}{2} (p - \mu_ {g ^ {*}}) ^ {T} \Sigma_ {g ^ {*}} ^ {- 1} (p - \mu_ {g ^ {*}})\right), \tag {2}
+$$
+
+where the "closest Gaussian"  $g^{*}$  is taken as the Gaussian with the largest contribution at point  $p$ :
+
+$$
+g ^ {*} = \arg \min  _ {g} \left\{\left(p - \mu_ {g}\right) ^ {T} \Sigma_ {g} ^ {- 1} \left(p - \mu_ {g}\right) \right\}. \tag {3}
+$$
+
+Eq. (2) thus considers that the contribution of the closest Gaussian  $g^{*}$  to the density at  $p$  is much higher than the contribution of the other Gaussians. This will help us encourage the Gaussians to be well spread.
+
+We also would like the 3D Gaussians to be flat, as they would then be aligned more closely with the surface of the mesh. Consequently, every Gaussian  $g$  would have one of its three scaling factors close to 0 and:
+
+$$
+\left(p - \mu_ {g}\right) ^ {T} \Sigma_ {g} ^ {- 1} \left(p - \mu_ {g}\right) \approx \frac {1}{s _ {g} ^ {2}} \langle p - \mu_ {g}, n _ {g} \rangle^ {2}, \tag {4}
+$$
+
+where  $s_g$  the smallest scaling factor of the Gaussian and  $n_g$  the direction of the corresponding axis. Moreover, because we want Gaussians to describe the true surface of the scene, we need to avoid semi-transparent Gaussians. Therefore, we want Gaussians to be either opaque or fully transparent, in which case we can drop them for rendering. Consequently, we want to have  $\alpha_g = 1$  for any Gaussian  $g$ .
+
+In such scenario, the density of the Gaussians could fi-
+
+![](images/a56292c9d827b58edba578199500579430f2a6afc2342e78c183f2cabb75b7e6.jpg)  
+(a) Mesh & Gaussians
+
+![](images/0fcc44c7ccb311ac823d6d50fb0fc3578e6568e81d27799a90f6d5a308775a37.jpg)  
+(b) Mesh (No Texture)
+
+![](images/d0edac57d5ee016f3498c03b58541f242fc507dbc35d2b68fac779908d8555be.jpg)  
+(c) Mesh normals
+
+![](images/c917145fa695713ee2573a050eaa62a4f8dc031b2f14f7df00ab8096b2a22b0e.jpg)  
+(a) Mesh & Gaussians  
+Figure 4. Examples of (a) renderings and (b) reconstructed meshes with SuGaR. The (c) normal maps help visualize the geometry.
+
+![](images/3cb31b2b075cc43fa4311d547582d177406bfe27cb552cf805412325a9c0c952.jpg)  
+(b) Mesh (No Texture)
+
+![](images/4a714642ab921d433fa94db5a38240ef13762090f89437b638b27ee572ab763f.jpg)  
+(c) Mesh normals
+
+nally be approximated by density  $\bar{d} (p)$  with:
+
+$$
+\bar {d} (p) = \exp \left(- \frac {1}{2 s _ {g ^ {*}} ^ {2}} \langle p - \mu_ {g ^ {*}}, n _ {g ^ {*}} \rangle^ {2}\right). \tag {5}
+$$
+
+A first strategy to enforce our regularization is to add term  $|d(p) - \bar{d}(p)|$  to the optimization loss. While this approach works well to align Gaussians with the surface, we noticed that computing a slightly different loss relying on an SDF rather than on density further increases the alignment of Gaussians with the surface of the scene. For a given flat Gaussian, i.e.,  $s_g = 0$ , considering level sets is meaningless since all level sets would degenerate toward the plane passing through the center of the Gaussian  $\mu_g$  with normal  $n_g$ . The distance between point  $p$  and the true surface of the scene would be approximately  $|\langle p - \mu_{g'}, n_{g'} \rangle|$ , the distance from  $p$  to this plane. Consequently, the zero-crossings of the Signed Distance Function
+
+$$
+\bar {f} (p) = \pm s _ {g *} \sqrt {- 2 \log (\bar {d} (p))} \tag {6}
+$$
+
+corresponds to the surface of the scene. More generally, we define
+
+$$
+f (p) = \pm s _ {g *} \sqrt {- 2 \log (d (p))} \tag {7}
+$$
+
+as the "ideal" distance function associated with the density function  $d$ . This distance function corresponds to the true surface of the scene in an ideal scenario where  $d = \bar{d}$ . We therefore take our regularization term  $\mathcal{R}$  as
+
+$$
+\mathcal {R} = \frac {1}{| \mathcal {P} |} \sum_ {p \in \mathcal {P}} \left| | \hat {f} (p) | - | f (p) | \right|, \tag {8}
+$$
+
+by sampling 3D points  $p$  and summing the differences at these points between the ideal SDF  $f(p)$  and an estimate  $\hat{f}(p)$  of the SDF of the surface created by the current Gaussians.  $\mathcal{P}$  refers to the set of sampled points. During optimization, we simply want the zero-level sets of both the
+
+![](images/0fda687334d3dc4b93586b3ae1e96bd344c0ae51d878b5b833c5a193d5922480.jpg)  
+Figure 5. Efficiently estimating  $\hat{f}(p)$  of the SDF of the surface generated from Gaussians. We render depth maps of the Gaussians, sample points  $p$  in the viewpoint according to the distribution of the Gaussians. Value  $\hat{f}(p)$  is taken as the 3D distance between  $p$  and the intersection between the line of sight for  $p$  and the depth map.
+
+SDF estimator  $\hat{f}$  and the ideal SDF  $f$  to align. Since a zero-level set is entirely determined by an unsigned distance function, we actually do not need to compute the sign of the SDFs and use absolute values.
+
+Computing efficiently  $\hat{f}(p)$  is a priori challenging. To do so, we propose to use the depth maps of the Gaussians from the viewpoints used for training—these depth maps can be rendered efficiently by extending the splatting rasterizer. Then, as shown in Figure 5, for a point  $p$  visible from a training viewpoint,  $\hat{f}(p)$  is the difference between the depth of  $p$  and the depth in the corresponding depth map at the projection of  $p$ . Moreover, we sample points  $p$  following the distribution of the Gaussians:
+
+$$
+p \sim \prod_ {g} \mathcal {N} \left(.; \mu_ {g}, \Sigma_ {g}\right), \tag {9}
+$$
+
+with  $\mathcal{N}(:,\mu_g,\Sigma_g)$  the Gaussian distribution of mean  $\mu_{g}$  and covariance  $\Sigma_{g}$  as these points are likely to correspond to a high gradient for  $\mathcal{R}$ .
+
+![](images/7793310e8a1663a1bc425bb1121acfd80c404fd4cb0d52af2380bd5663ebf64a.jpg)  
+Figure 6. Sampling points on a level set for Poisson reconstruction. Left: We sample points on the depth maps of the Gaussians and refine the point locations to move the points on the level set. Right: Comparison between the extracted mesh without (left) and with (right) our refinement step. Since splatted depth maps are not exact, using directly the depth points for reconstruction usually results in a large amount of noise and missing details.
+
+![](images/d262f28448b4d3af275cf2b66155ba011da0fa7967794a9bd0b45cb786fba190.jpg)
+
+![](images/2f706608e69c018da35d9debfa62abf0fe7d326101fcafab0e862679a53f0601.jpg)
+
+We also add a regularization term to encourage the normals of SDF  $f$  and the normals of SDF  $\bar{f}$  to be similar:
+
+$$
+\mathcal {R} _ {\text {N o r m}} = \frac {1}{| \mathcal {P} |} \sum_ {p \in \mathcal {P}} \left\| \frac {\nabla f (p)}{\| \nabla f (p) \| _ {2}} - n _ {g ^ {*}} \right\| _ {2} ^ {2}. \tag {10}
+$$
+
+# 4.2. Efficient Mesh Extraction
+
+To create a mesh from the Gaussians obtained after optimization using our regularization terms in Eq. (8) and Eq. (10), we sample 3D points on a level set of the density computed from the Gaussians. The level set depends on a level parameter  $\lambda$ . Then, we obtain a mesh by simply running a Poisson reconstruction [14] on these points. Note that we can also easily assign the points with the normals of the SDF, which improves the mesh quality.
+
+The challenge is in efficiently identifying points lying on the level set. For this, as shown in Figure 6, we again rely on the depth maps of the Gaussians as seen from the training viewpoints. We first randomly sample pixels from each depth map. For each pixel  $m$ , we sample its line of sight to find a 3D point on the level set. Formally, we sample  $n$  points  $p + t_i v$ , where  $p$  is the 3D point in the depth map that reprojects on pixel  $m$ ,  $v$  is the direction of the line of sight, and  $t_i \in [-3\sigma_g(v), 3\sigma_g(v)]$  where  $\sigma_g(v)$  is the standard deviation of the 3D Gaussian  $g$  in the direction of the camera. The interval  $[-3\sigma_g(v), 3\sigma_g(v)]$  is the confidence interval for the 99.7 confidence level of the 1D Gaussian function of  $t$  along the ray.
+
+Then, we compute the density values  $d_{i} = d(p + t_{i}v)$  from Eq. (1) of these sampled points. If there exist  $i,j$  such that  $d_{i} < \lambda < d_{j}$ , then there is a level set point located in this range. If so, we use linear interpolation to compute the coefficient  $t^{*}$  such that  $p + t^{*}v$  is the level set point closest to the camera, verifying  $d(p + t^{*}v) = \lambda$ . We also compute the normals of the surface at points  $\hat{p}$ , which we naturally
+
+![](images/5e1246092c67462f639fac33a618dd8fa91f5355a55569554ece5a95bc25dd3d.jpg)  
+Figure 7. Joint refinement of mesh and Gaussians. Left: We bind Gaussians to the triangles of the mesh. Depending on the number of triangles in the scene, we bind a different number of Gaussians per triangle, with predefined barycentric coordinates. Right: Mesh before and after joint refinement.
+
+define as the normalized analytical gradient of the density  $\frac{\nabla d(\hat{p})}{\|\nabla d(\hat{p})\|_2}$ .
+
+Finally, we apply Poisson reconstruction to reconstruct a surface mesh from the level set points and their normals.
+
+# 4.3. Binding New 3D Gaussians to the Mesh
+
+Once we have extracted a first mesh, we can refine this mesh by binding new Gaussians to the mesh triangles and optimize the Gaussians and the mesh jointly using the Gaussian Splatting rasterizer. This enables the edition of the Gaussian splatting scene with popular mesh editing tools while keeping high-quality rendering thanks to the Gaussians.
+
+Given the initial mesh, we instantiate new 3D Gaussians on the mesh. More exactly, we associate a set of  $n$  thin 3D Gaussians to each triangle of the mesh, sampled on the surface of the triangle, as illustrated in Figure 7. To do so, we slightly modify the structure of the original 3D Gaussian Splitting model.
+
+We explicitly compute the means of the Gaussians from the mesh vertices using predefined barycentric coordinates in the corresponding triangles during optimization. Also, the Gaussians have only 2 learnable scaling factors instead of 3 and only 1 learnable 2D rotation encoded with a complex number rather than a quaternion, to keep the Gaussians flat and aligned with the mesh triangles. More details about this parameterisation are given in the supplementary material. Like the original model, we also optimize an opacity value and a set of spherical harmonics for every Gaussian to encode the color emitted in all directions.
+
+Figure 7 shows an example of a mesh before and after refinement. Figure 1 and the supplementary material give examples of what can be done by editing the mesh.
+
+# 5. Experiments
+
+# 5.1. Implementation details
+
+All our models are optimized on a single GPU Nvidia Tesla V100 SXM2 32 Go.
+
+Regularization. For all scenes, we start by optimizing a Gaussian Splatting with no regularization for 7,000 iterations to let the 3D Gaussians position themselves without any additional constraint. Then, we perform 2,000 iterations with an additional entropy loss on the opacities  $\alpha_{g}$  of the Gaussians, as a way to enforce them to become binary.
+
+Finally, we remove Gaussians with opacity values under 0.5 and perform 6,000 iterations with the regularization term introduced in Subsection 4.1, which makes a total of 15,000 iterations. To compute the density values of points from a Gaussian  $g$ , we sum only the Gaussian functions from the 16 nearest Gaussians of  $g$  and update the list of nearest neighbors every 500 iterations. Optimization takes between 15 and 45 minutes depending on the scene.
+
+Mesh extraction. For all experiments except the ablation presented in Table 2, we extract the  $\lambda$ -level set of the density function for  $\lambda = 0.3$ . We perform Poisson reconstruction with depth 10 and apply mesh simplification using quadric error metrics [9] to decrease the resolution of the meshes. Mesh extraction generally takes between 5 and 10 minutes depending on the scene.
+
+Joint refinement. We jointly refine the mesh and the bound 3D Gaussians for either 2,000, 7,000 or 15,000 iterations. Depending on the number of iterations, the duration of refinement goes from a few minutes to an hour.
+
+# 5.2. Real-Time Rendering of Real Scenes
+
+For evaluating our model, we follow the approach from the original 3D Gaussian Splatting paper [15] and compare the performance of several variations of our method SuGaR after refinement on real 3D scenes from 3 different datasets: Mip-NeRF360 [1], DeepBlending [12] and Tanks&Temples [16]. We call R-SuGaR-NK a refined SuGaR model optimized for  $N$  iterations during refinement.
+
+Following [15], we select the same sets of 2 scenes from Tanks&Temples (Truck and Train) and 2 scenes from DeepBlending (Playroom and Dr. Johnson). However, due to licensing issues and the unavailability of the scenes Flowers and Treehill, we perform the evaluation of all methods only on 7 scenes from Mip-NeRF360 instead of the full set of 9 scenes.
+
+We compute the standard metrics PSNR, SSIM and LPIPS [44] to evaluate the quality of SuGaR's rendering using our extracted meshes and their bound surface Gaussians. Note that [6, 26, 39] also do not use plain textured
+
+mesh rendering. We compare to several baselines, some of them focusing only on Novel View Synthesis [2, 15, 23, 41] and others relying on a reconstructed mesh [6, 26, 39], just like SuGaR. Results on the Mip-NeRF360 dataset are given in Table 1. Results on Tanks&Temple and DeepBlending are similar and can be found in the supplementary material.
+
+Even though SuGaR focuses on aligning 3D Gaussians for reconstructing a high quality mesh during the first stage of its optimization, it significantly outperforms the state of the art methods for Novel View Synthesis using a mesh and reaches better performance than several famous models that focus only on rendering, such as I-NGP [23] and Plenoxels [41]. This performance is remarkable as SuGaR is able to extract a mesh significantly faster than other methods.
+
+Moreover, SuGaR even reaches performance similar to state-of-the-art models for rendering quality [2, 15] on some of the scenes used for evaluation. Two main reasons explain this performance. First, the mesh extracted after the first stage of optimization serves as an excellent initialization for positioning Gaussians when starting the refinement phase. Then, the Gaussians constrained to remain on the surface during refinement greatly increase the rendering quality as they play the role of an efficient texturing tool and help reconstructing very fine details missing in the extracted mesh. Additional qualitative results are available in Figure 4.
+
+# 5.3. Mesh Extraction
+
+To demonstrate the ability of our mesh extraction method for reconstructing high-quality meshes that are well-suited for view synthesis, we compare different mesh extraction algorithms. In particular, we optimize several variations of SuGaR by following the exact same pipeline as our standard model, except for the mesh extraction process: We either extract the mesh using a very fine marching cubes algorithm [21], by applying Poisson reconstruction [14] using the centers of the 3D Gaussians as the surface point cloud, or by applying our mesh extraction method on different level sets. Quantitative results are available in Table 2 and show the clear superiority of our approach for meshing 3D Gaussians. Figure 3 also illustrates how the marching cubes algorithm fails in this context.
+
+# 5.4. Mesh Rendering Ablation
+
+Table 3 provides additional results to quantify how various parameters impact rendering performance. In particular, we evaluate how the resolution of the mesh extraction, i.e., the number of triangles, modifies the rendering quality. For fair comparison, we increase the number of surface-aligned Gaussians per triangle when we decrease the number of triangles. Results show that increasing the number of vertices increases the quality of rendering with surface Gaussians, but meshes with less triangles are already able to reach state of the art results.
+
+<table><tr><td rowspan="2"></td><td colspan="3">Indoor scenes</td><td colspan="3">Outdoor scenes</td><td colspan="3">Average on all scenes</td></tr><tr><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td></tr><tr><td colspan="10">No mesh (except SuGaR)</td></tr><tr><td>Plenoxels [42]</td><td>24.83</td><td>0.766</td><td>0.426</td><td>22.02</td><td>0.542</td><td>0.465</td><td>23.62</td><td>0.670</td><td>0.443</td></tr><tr><td>INGP-Base [23]</td><td>28.65</td><td>0.840</td><td>0.281</td><td>23.47</td><td>0.571</td><td>0.416</td><td>26.43</td><td>0.725</td><td>0.339</td></tr><tr><td>INGP-Big [23]</td><td>29.14</td><td>0.863</td><td>0.242</td><td>23.57</td><td>0.602</td><td>0.375</td><td>26.75</td><td>0.751</td><td>0.299</td></tr><tr><td>Mip-NeRF360 [2]</td><td>31.58</td><td>0.914</td><td>0.182</td><td>25.79</td><td>0.746</td><td>0.247</td><td>29.09</td><td>0.842</td><td>0.210</td></tr><tr><td>3DGS [15]</td><td>30.41</td><td>0.920</td><td>0.189</td><td>26.40</td><td>0.805</td><td>0.173</td><td>28.69</td><td>0.870</td><td>0.182</td></tr><tr><td>R-SuGaR-15K (Ours)</td><td>29.43</td><td>0.910</td><td>0.216</td><td>24.40</td><td>0.699</td><td>0.301</td><td>27.27</td><td>0.820</td><td>0.253</td></tr><tr><td colspan="10">With mesh</td></tr><tr><td>Mobile-NeRF [6]</td><td>-</td><td>-</td><td>-</td><td>21.95</td><td>0.470</td><td>0.470</td><td>-</td><td>-</td><td>-</td></tr><tr><td>NeRFMeshing [26]</td><td>23.83</td><td>-</td><td>-</td><td>22.23</td><td>-</td><td>-</td><td>23.15</td><td>-</td><td>-</td></tr><tr><td>BakedSDF [39]</td><td>27.06</td><td>0.836</td><td>0.258</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>R-SuGaR-2K (Ours)</td><td>26.29</td><td>0.872</td><td>0.262</td><td>22.97</td><td>0.648</td><td>0.360</td><td>24.87</td><td>0.776</td><td>0.304</td></tr><tr><td>R-SuGaR-7K (Ours)</td><td>28.73</td><td>0.904</td><td>0.226</td><td>24.16</td><td>0.691</td><td>0.313</td><td>26.77</td><td>0.813</td><td>0.263</td></tr><tr><td>R-SuGaR-15K (Ours)</td><td>29.43</td><td>0.910</td><td>0.216</td><td>24.40</td><td>0.699</td><td>0.301</td><td>27.27</td><td>0.820</td><td>0.253</td></tr></table>
+
+Table 1. Quantitative evaluation of rendering quality on the Mip-NeRF360 dataset [2]. SuGaR is best among the methods that recover a mesh, and still performs well compared to NeRF methods and vanilla 3D Gaussian Splatting.  
+
+<table><tr><td>Extraction method</td><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td></tr><tr><td>Marching Cubes [21]</td><td>23.91</td><td>0.703</td><td>0.392</td></tr><tr><td>Poisson (centers) [14]</td><td>23.76</td><td>0.756</td><td>0.340</td></tr><tr><td>Ours (Surface level 0.1)</td><td>24.62</td><td>0.765</td><td>0.313</td></tr><tr><td>Ours (Surface level 0.3)</td><td>24.87</td><td>0.776</td><td>0.304</td></tr><tr><td>Ours (Surface level 0.5)</td><td>24.91</td><td>0.777</td><td>0.304</td></tr></table>
+
+Table 2. Ablation for different mesh extraction methods on the Mip-NeRF360 dataset [2] after applying our regularization term. For 'Poisson (centers)', we apply Poisson reconstruction [14] using as surface points the centers of the 3D Gaussians. For fair comparison, we calibrate the methods to enforce all extracted meshes to have approximately 1,000,000 vertices.  
+
+<table><tr><td></td><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td></tr><tr><td>1M vertices (3DGS)</td><td>24.51</td><td>0.768</td><td>0.295</td></tr><tr><td>1M vertices (UV)</td><td>21.24</td><td>0.609</td><td>0.478</td></tr><tr><td>200K vertices (3DGS)</td><td>24.24</td><td>0.757</td><td>0.300</td></tr><tr><td>200K vertices (UV)</td><td>21.44</td><td>0.656</td><td>0.419</td></tr></table>
+
+Table 3. Comparison between surface-aligned 3D Gaussians and an optimized traditional UV texture on the Mip-NeRF360 dataset [2]. For fair comparison, we only use the diffuse spherical harmonics component when rendering images with SuGaR. Using 3D Gaussians bound to the mesh greatly improves rendering quality, even though it contains less parameters than the UV texture.
+
+Then, we illustrate the benefits of using Gaussians aligned on the surface as a texturing tool for rendering meshes. To this end, we also optimize traditional UV textures on our meshes using differentiable mesh rendering
+
+with traditional triangle rasterization. Even though rendering with surface-aligned Gaussians provides better performance, rendering our meshes with traditional UV textures still produces satisfying results, which further illustrates the quality of our extracted meshes. Qualitative comparisons are provided in the supplementary material.
+
+# 6. Conclusion
+
+We proposed a very fast algorithm to obtain an accurate 3D triangle mesh for a scene via Gaussian Splatting. Moreover, by combining meshing and Gaussian Splatting, we make possible intuitive manipulation of the captured scenes and realistic rendering, offering new possibilities for creators. SuGaR does not come without limitations: Gaussians do tend to "cheat" on the geometry and depth by creating cavities to reproduce specular effects, instead of relying on spherical harmonics. SuGaR's regularization mitigates this issue, but Gaussians can still distort the surface in regions with intense specularity. In addition, SuGaR's assumption that the scene can be represented as a surface complicates the rendering of volumetric effects or fuzzy materials.
+
+# Acknowledgements.
+
+This project was funded in part by the European Union (ERC Advanced Grant explorer Funding ID #101097259). This work was granted access to the HPC resources of IDRIS under the allocation 2023-AD011013387R1 made by GENCI. We thank George Drettakis and Elliot Vincent for inspiring discussions and valuable feedback.
+
+# References
+
+[1] Jonathan T. Barron. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In International Conference on Computer Vision, 2021. 3, 7  
+[2] Jonathan T. Barron. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 7, 8, 2, 4  
+[3] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural Reflectance Decomposition from Image Collections. In International Conference on Computer Vision, 2021. 3  
+[4] Chris Buehler, Michael Bosse, Leonard Mcmillan, Steven Gortler, and Michael Cohen. Unstructured Lumigraph Rendering. In ACM SIGGRAPH, 2001. 3  
+[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial Radiance Fields. In European Conference on Computer Vision, 2022. 3  
+[6] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. MobileNeRF: Exploiting the Polygon Rasterization Pipeline for Efficient Neural Field Rendering on Mobile Architectures. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 7, 8  
+[7] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. NeuMesh: Learning Disentangled Neural Mesh-Based Implicit Field for Geometry and Texture Editing. In European Conference on Computer Vision, 2022. 3  
+[8] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving Neural Implicit Surfaces Geometry with Patch Warping. In Conference on Computer Vision and Pattern Recognition, 2022. 3  
+[9] Michael Garland and Paul S. Heckbert. Surface Simplification Using Quadric Error Metrics. In ACM SIGGRAPH, 1997. 7  
+[10] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven Seitz. Multi-View Stereo for Community Photo Collections. In International Conference on Computer Vision, 2007. 3  
+[11] Peter Hedman and Pratul P. Srinivasan. Baking Neural Radiance Fields for Real-Time View Synthesis. In International Conference on Computer Vision, 2021. 3  
+[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. In ACM SIGGRAPH, 2018. 3, 7, 2, 4  
+[13] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. ReLU Fields: The Little Non-Linearity That Could. In ACM SIGGRAPH, 2022. 3  
+[14] Michael M. Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson Surface Reconstruction. In Eurographics, 2006. 2, 6, 7, 8  
+[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. In ACM SIGGRAPH, 2023. 1, 2, 4, 7, 8
+
+[16] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. In ACM SIGGRAPH, 2017. 7, 2, 4  
+[17] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-Based Neural Rendering with Per-View Optimization. In Computer Graphics Forum, 2021. 3  
+[18] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: Neural Rendering of Objects from Online Image Collections. In ACM SIGGRAPH, 2022. 3  
+[19] Marc Levoy and Pat Hanrahan. Light Field Rendering. In ACM SIGGRAPH, 1996. 3  
+[20] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-Fidelity Neural Surface Reconstruction. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 3  
+[21] William E. Lorensen and Harvey E. Cline. Marching Cubes: A High Resolution 3D Surface Construction Algorithm. In ACM SIGGRAPH, 1987. 2, 3, 7, 8  
+[22] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, 2020. 1, 3  
+[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. In ACM SIGGRAPH, 2022. 3, 7, 8, 2  
+[24] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying Neural Implicit Surfaces and Radiance Fields for Multi-View Reconstruction. In International Conference on Computer Vision, 2021. 3  
+[25] 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. 2  
+[26] Marie-Julie Rakotosaona, Fabian Manhardt, Diego Martin Arroyo, Michael Niemeyer, Abhijit Kundu, and Federico Tombari. NeRFMeshing: Distilling Neural Radiance Fields into Geometrically-Accurate 3D Meshes. In DV, 2023. 2, 3, 7, 8  
+[27] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D Deep Learning with PyTorch3D. In arXiv Preprint, 2020. 2  
+[28] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding Up Neural Radiance Fields with Thousands of Tiny MLPs. In International Conference on Computer Vision, 2021. 3  
+[29] Gernot Riegler and Vladlen Koltun. Free View Synthesis. In European Conference on Computer Vision, 2020. 3
+
+[30] Gernot Riegler and Vladlen Koltun. Stable View Synthesis. In Conference on Computer Vision and Pattern Recognition, 2021. 3  
+[31] Darius Rückert, Linus Franke, and Marc Stamminger. ADOP: Approximate Differentiable One-Pixel Point Rendering. In ACM SIGGRAPH, 2022. 4  
+[32] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo Tourism: Exploring Photo Collections in 3D. In ACM SIGGRAPH, 2006. 3, 4  
+[33] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural Reflectance and Visibility Fields for Relighting and View Synthesis. In Conference on Computer Vision and Pattern Recognition, 2021. 3  
+[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct Voxel Grid Optimization: Super-Fast Convergence for Radiance Fields Reconstruction. In Conference on Computer Vision and Pattern Recognition, 2022. 3  
+[35] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields. In Conference on Computer Vision and Pattern Recognition, 2022. 3  
+[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-View Reconstruction. In Advances in Neural Information Processing Systems, 2021. 2, 3  
+[37] Daniel N. Wood, Daniel I. Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H. Salesin, and Werner Stuetzle. Surface Light Fields for 3D Photography. In ACM SIGGRAPH, 2000. 3  
+[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume Rendering of Neural Implicit Surfaces. In Advances in Neural Information Processing Systems, 2021. 2, 3  
+[39] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, and Jonathan T. Barron. BakedSDF: Meshing Neural SDFs for Real-Time View Synthesis. In ACM SIGGRAPH, 2023. 2, 3, 7, 8  
+[40] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees For Real-Time Rendering of Neural Radiance Fields. In International Conference on Computer Vision, 2021. 3  
+[41] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 7  
+[42] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2022. 8, 2  
+[43] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse Rendering with Spherical Gaussians for Physics-Based Material Editing and Relighting. In Conference on Computer Vision and Pattern Recognition, 2021. 3  
+[44] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of
+
+Deep Features as a Perceptual Metric. In Conference on Computer Vision and Pattern Recognition, 2018. 7, 2, 4

2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a598b088d88250ed975ce16a174df3f2d3c36b25502354a51e5947f94661f440
+size 563285

2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_content_list.json

@@ -0,0 +1,1294 @@
+[
+    {
+        "type": "text",
+        "text": "SubT-MRS Dataset: Pushing SLAM Towards All-weather Environments https://superodometry.com/datasets",
+        "text_level": 1,
+        "bbox": [
+            117,
+            128,
+            851,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Shibo Zhao\\*, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, Yi Du, Shaoshu Su, Can Xu, John Keller, Jay Karhade, Lucas Nogueira, Sourojit Saha, Ji Zhang, Wenshan Wang, Chen Wang, Sebastian Scherer",
+        "bbox": [
+            84,
+            203,
+            895,
+            257
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Carnegie Mellon University  $^{2}$ University at Buffalo",
+        "bbox": [
+            289,
+            261,
+            687,
+            279
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            314,
+            313,
+            329
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Simultaneous localization and mapping (SLAM) is a fundamental task for numerous applications such as autonomous navigation and exploration. Despite many SLAM datasets have been released, current SLAM solutions still struggle to have sustained and resilient performance. One major issue is the absence of high-quality datasets including diverse all-weather conditions and a reliable metric for assessing robustness. This limitation significantly restricts the scalability and generalizability of SLAM technologies, impacting their development, validation, and deployment. To address this problem, we present SubT-MRS, an extremely challenging real-world dataset designed to push SLAM towards all-weather environments to pursue the most robust SLAM performance. It contains multi-degraded environments including over 30 diverse scenes such as structureless corridors, varying lighting conditions, and perceptual obscurants like smoke and dust; multimodal sensors such as LiDAR, fisheye camera, IMU, and thermal camera; and multiple locomotions like aerial, legged, and wheeled robots. We developed accuracy and robustness evaluation tracks for SLAM and introduced novel robustness metrics. Comprehensive studies are performed, revealing new observations, challenges, and opportunities for future research.",
+        "bbox": [
+            75,
+            345,
+            473,
+            694
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            76,
+            718,
+            209,
+            734
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Simultaneous Localization and Mapping (SLAM) is essential in robotics since it provides foundational perception and spatial awareness, and enables machines to understand and interact with the physical world in real-time. Therefore, it has a wide range of applications such as autonomous driving and space exploration. Despite significant advancements in both geometric [5, 34, 41] and data-driven SLAM methods [43, 48], existing solutions remain fragile in challenging conditions. One main reason is that existing SLAM",
+        "bbox": [
+            75,
+            739,
+            468,
+            877
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "algorithms are often developed and evaluated with datasets from controlled environments [3, 12, 18, 21, 33, 35, 39, 54]. For example, KITTI [18, 21] dataset is mostly collected in sunny weather and EuROC-MAV dataset [3] is collected in small well-lit rooms. These datasets, unfortunately, fail to capture the challenges in real-world scenarios, hindering the development of robust SLAM solutions.",
+        "bbox": [
+            500,
+            315,
+            893,
+            421
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To bridge this gap and push SLAM towards all-weather environments, we present an extremely challenging dataset, SubT-MRS, including scenarios featuring various sensor degradation, aggressive locomotions, and extreme-weather conditions. The SubT-MRS dataset comprises 3 years of data from the DARPA Subterranean (SubT) Challenge [1] (2019-2021) and extends with an additional 2 years of diverse environments (2022-2023), containing mixed indoor and outdoor settings, including long corridors, off-road scenario, tunnels, caves, deserts, forests, and bushlands. Cumulatively, this forms a 5-year dataset encompassing over 500 hours and 100 kilometers of terrain subjected to multimodal sensors including LiDAR, fisheye cameras, depth cameras, thermal cameras, and IMU; heterogeneous platforms including RC cars, legged robots, aerial robots, and wheeled robots; and extreme obscurant conditions such as dense fog, dust, smoke, and heavy snow.",
+        "bbox": [
+            496,
+            426,
+            893,
+            684
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Additionally, we find there is no well-established metric to evaluate the robustness of a SLAM system. Existing evaluation metrics such as absolute trajectory error (ATE) [3] are not representative of the actual performance in robotic applications. We argue that to ensure the safety control of a robot, SLAM system evaluation should not only focus on poses but also its velocity. Taking robot localization as an example, while a momentary pose error spike may only slightly undermine the overall ATE, however, it can lead to a catastrophic crash in an aerial robotic platform. To effectively gauge the performance of a SLAM algorithm, we introduce a new robustness metric to evaluate the reliability of the SLAM system, particularly examining smoothness and accuracy of velocity estimation.",
+        "bbox": [
+            496,
+            688,
+            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,
+            810,
+            46
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author: shiboz@andrew.cmu.edu",
+        "bbox": [
+            94,
+            886,
+            359,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "22647",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/b3d861d86989a475f5b89f18a2bceb82e3f55d46da24ddf4bfe5e6348b1bff74.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            89,
+            893,
+            292
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/3e2278a11c657ec21d4a1b1d9bd38063d647760e8f90618979c79422d7d0a8b0.jpg",
+        "image_caption": [
+            "SubT-Tunnel",
+            "SubT-Urban",
+            "SubT-Cavern",
+            "Multi Floors"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            292,
+            893,
+            333
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/4ff2e7e8ea5727e51f0d28853e2951db92306ff0c8c78c86400d2dbee3f76ece.jpg",
+        "image_caption": [
+            "Dark Long Corridor",
+            "Dense Fog",
+            "Laurel Cavern",
+            "Textureless Corridor",
+            "Darkness Stairs with Jerk Motion",
+            "Dense Dust and Smoke",
+            "Offroad with Aggressive Motion",
+            "Snow with Slipping Motion",
+            "Figure 1. Dense reconstruction from the SubT-MRS dataset, achieved through collaboration with diverse robots equipped with multimodal sensors. Colors represent different challenging environments (tunnels, caves, urban, confined spaces) captured by various robot types (aerial, legged, wheeled). The bottom section showcases a gallery with diverse visual, LiDAR, and mixed degradations."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            345,
+            893,
+            388
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Lastly, we perform extensive experiments using proposed degradation datasets to benchmark visual and LiDAR SLAM algorithms. These experiments identify the limitations of the existing SLAM systems and evaluate their robustness across the degradation. Our contributions include:",
+        "bbox": [
+            75,
+            479,
+            468,
+            556
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- All-weather Environments To push SLAM toward all-weather environments, we introduce SubT-MRS, an extremely challenging dataset. Spanning five years, it comprises over 500 hours and 100 kilometers of accurately measured trajectories. To the best of our knowledge, SubT-MRS is the first real-world dataset that specifically addresses failure scenarios of SLAM by incorporating a variety of degraded conditions, multiple robotic platforms, and diverse sets of multimodal sensors.",
+            "- Robustness Metric To evaluate the actual performance running on robots, we propose a robustness metric, which, to the best of our knowledge, is the first metric evaluating the reliability, safety, and resilience of SLAM.",
+            "- SLAM Challenge We provide a comprehensive benchmark<sup>1</sup> based on the SubT-MRS dataset by conducting ICCV'23 SLAM Challenge. Our evaluation reveals the limitations of state-of-the-art visual and LiDAR SLAM solutions. Furthermore, we conduct extensive experiments to verify robustness metric against various sensor degradation and length of trajectories."
+        ],
+        "bbox": [
+            76,
+            556,
+            468,
+            859
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related Work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            478,
+            640,
+            494
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Multimodal Sensors Multimodal sensor datasets are crucial for the development of robust SLAM systems, as single sensor modalities are not comprehensive for all scenarios. Existing datasets typically focus on a limited range of modalities, such as monocular or stereo cameras [3, 49], LiDAR [19], event cameras [29], and depth cameras [42]. The KITTI dataset [19], although covering most modalities, is geared towards on-road scenarios and lacks thermal cameras, limiting its applicability to simple, controlled environments. Conversely, the ViViD++ dataset [23] incorporates thermal images in outdoor settings but is deficient in hardware synchronization, posing challenges for SLAM system development. The Weichen dataset [10] offers thermal images in indoor environments with accurate ground truth poses but is confined to motion capture room settings. In contrast, our SubT-MRS dataset delivers a comprehensive suite of time-synchronized multimodal sensors, including LiDAR, monocular cameras, thermal cameras, depth cameras, and IMU, along with centimeter-level ground truth, catering to a diverse range of research needs.",
+        "bbox": [
+            496,
+            503,
+            890,
+            805
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Heterogeneous Platforms Developing a robust SLAM system necessitates versatility in handling different motion patterns and various robotic platforms. However, most existing datasets are tailored for single-robot scenarios. While there are a few multi-robot datasets, they typically involve homogeneous platforms [9, 15, 24, 44], posing challenges",
+        "bbox": [
+            496,
+            809,
+            890,
+            901
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>SubT-MRS dataset is completely real-world data. Simulated data in the benchmark is to ensure evaluation completeness",
+        "bbox": [
+            76,
+            875,
+            468,
+            898
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "22648",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/01f3a26eaab9dae558cba6309166aa16b55eafc0d3c44dd7989cb8309dc889dd.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"4\">Multi-Spectrum</td><td colspan=\"6\">Multi-Degradation</td><td colspan=\"4\">Multi-Robot</td></tr><tr><td>Camera</td><td>IMU</td><td>LiDAR/Depth</td><td>Thermal</td><td>Illumination</td><td>Snow Smoke</td><td>Structureless</td><td>SubT</td><td>Aggressive Motion</td><td>Vehicle</td><td>Drone</td><td>Legged</td><td>Handheld</td><td></td></tr><tr><td>EuRoC MAV [3]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td></td></tr><tr><td>PennCOsYVIO [33]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>TUM VIO [39]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>UZH-FPV [12]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td></td></tr><tr><td>College Dataset [52]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>RobotCar [28]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>UMA VI [54]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>UMich [6]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>KITTI [19]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>Virtual KITTI [16]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>DARPA SubT [37]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>Hilti SLAM [21]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>M3ED [7]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td></td></tr><tr><td>TartanAir [47]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td></td></tr><tr><td>SubT-MRS (Ours)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td></td></tr></table>",
+        "bbox": [
+            80,
+            90,
+            890,
+            281
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Comparison of SLAM datasets on multi-sensors, multi robots, and multi degradation.",
+        "bbox": [
+            204,
+            287,
+            759,
+            301
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "in evaluating SLAM performance across varied robotic platforms. The AirMuseum dataset [13] includes drones and ground robots but lacks data from legged robots. As a comparison, SubT-MRS dataset features diverse robotic platforms including legged robots, aerial robots, and wheeled robots, operating in varied environments and sensor setups.",
+        "bbox": [
+            75,
+            328,
+            467,
+            419
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Extreme Environmental Conditions Developing and testing SLAM systems in extreme environmental conditions is crucial for mitigating potential real-world failures. However, most existing SLAM datasets have been primarily limited to single, controlled environments. The TUM-VI [39] and UMA-VI [54] datasets, being indoor-outdoor visual-inertial datasets, pose challenges due to varying illumination and low-texture environments. The EuRoC MAV [3] and UZH-FPV drone racing [12] datasets, popular in the SLAM community, offer data on aggressive drone movements but usually in consistently lit conditions. The KITTI dataset [19], a staple in autonomous driving research for its outdoor LiDAR-Visual-Inertial data. The Hilti SLAM dataset [21] includes both indoor and outdoor LiDAR-visual-inertial datasets with dynamic lighting and confined spaces. However, both of them focus more on accuracy than on a variety of environmental degradations. TartanAir [47] covers most of the challenging environments but it is a simulation dataset that will pose the sim-to-real gap.",
+        "bbox": [
+            75,
+            425,
+            467,
+            712
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In contrast, the SubT-MRS dataset provides 30 varied scenes, encompassing a wide array of environments. These include SubT tunnels, urban areas, caverns, multifloor structures, dark corridors, and foggy conditions. It also features textureless surfaces, dusty and smoky environments, off-road areas with aggressive motion, and snowy terrains prone to slippage. We include challenging lighting conditions such as extreme darkness and overexposure, along with geometrically challenging scenarios like featureless corridors, staircases, and self-similar cave layouts. Additionally, the dataset incorporates perceptual challenges posed by smoke, fog, dust, and various weather conditions.",
+        "bbox": [
+            75,
+            719,
+            467,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. SubT-MRS Dataset",
+        "text_level": 1,
+        "bbox": [
+            498,
+            325,
+            686,
+            342
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We next present the SubT-MRS dataset from three aspects including its collecting environments and settings, ground truth collection, and the new robustness metric. A detailed comparison with other datasets is listed in Table 1.",
+        "bbox": [
+            496,
+            348,
+            890,
+            406
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. All-weather Environments",
+        "text_level": 1,
+        "bbox": [
+            498,
+            407,
+            738,
+            422
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As mentioned in Sec. 2, the SubT-MRS dataset distinguishes itself through its multimodal sensor setups, heterogeneous robotic platforms, and extreme environmental conditions. This section emphasizes this comprehensive nature pushing SLAM towards all-weather environments.",
+        "bbox": [
+            496,
+            425,
+            890,
+            500
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1.1. Multimodal Sensors",
+        "text_level": 1,
+        "bbox": [
+            498,
+            503,
+            702,
+            518
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Multimodal sensors provide a wealth of information for SLAM systems operating in all-weather environments. This diversity improves scene understanding and strengthens the system's perception ability. Therefore, We incorporate diverse sensors to ensure system robustness.",
+        "bbox": [
+            496,
+            521,
+            890,
+            595
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Sensor Pack We embedded 4 Leopard Imaging RGB fisheye cameras, 1 Velodyne puck, 1 Epson M-G365 IMU, 1 FLIR Boson thermal camera, and NVIDIA Jetson AGX Xavier as a sensor pack shown in Figure 2.",
+        "bbox": [
+            496,
+            597,
+            890,
+            657
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Time Synchronization To ensure the overall consistency of the fused data, we meticulously synchronize the sensors' time using the 'pulse per second (PPS)' technique. As illustrated in Figure 2, the IMU, LiDAR, and thermal camera are directly synchronized with the CPU clock, while the four RGB camera are synchronized using an FPGA board. Consequently, we can effectively manage the time synchronization gap between each pair of sensors not to exceed 3ms.",
+        "bbox": [
+            496,
+            657,
+            890,
+            779
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Fisheye and Thermal Camera Calibration Camera calibration plays an essential role in the efficiency of a SLAM system. For calibrating the fisheye cameras, we employed the open-source toolkit Kalibr [36], focusing on the intrinsic and extrinsic parameters. Specifically, we use the radial-tangential distortion model to rectify the omnidirectional fisheye camera model. Calibrating thermal cameras, however, poses unique challenges, especially in gathering high-",
+        "bbox": [
+            496,
+            779,
+            890,
+            900
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "22649",
+        "bbox": [
+            478,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/4a5dcfa2e9f9e3cf0bef64f230e655cdef81aa0eee5146db6cfaca25eed1ea7d.jpg",
+        "image_caption": [
+            "Figure 2. An overview of the sensor pack used in SubT-MRS dataset. It is equipped with a Xavier processing unit with hardware time synchronization for multimodal sensors including LiDAR, fisheye cameras, thermal cameras, depth cameras (option), and an IMU."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            88,
+            851,
+            256
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "quality thermal data. To tackle this, we set up a  $7 \\times 9$  chessboard, heated by direct sunlight, to generate high-contrast thermal data, ensuring accuracy in the calibration process.",
+        "bbox": [
+            75,
+            321,
+            468,
+            366
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "IMU Calibration and Extrinsic Calibration For Lidar-IMU extrinsic calibration, we utilize the CAD model to obtain the calibration parameters. In the case of Camera-IMU extrinsic calibration, we employ the Kalibr toolbox [36] to estimate the extrinsic matrix. To estimate the sensors' bias and the random walk noise of the gyroscope and accelerometer, we collected static data from the IMU and calibrated it using an Allan variance-based IMU tool [14, 17].",
+        "bbox": [
+            75,
+            366,
+            468,
+            487
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1.2. Multi-degraded Environments",
+        "text_level": 1,
+        "bbox": [
+            76,
+            489,
+            361,
+            506
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "SubT-MRS includes multiple challenging environments including visually degraded environments, geometrically degraded environments, and their combination.",
+        "bbox": [
+            75,
+            507,
+            468,
+            551
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Visual Degradation Poor-quality visual features can significantly hinder the performance of feature extraction processes and disrupt the accuracy of feature matching. Such issues arise in low-light conditions with inconsistent brightness or image noise introduced by air obscurants. SubT-MRS encompasses these challenges and provides a wide range of visual degradation. This includes environments with limited lighting, such as hospital interiors and caves (Figure 3 A-F), as well as smoky or dusty conditions that cause visual obstruction (Figure 3 M-N), and snowy areas with reduced visibility (Figure 3 H, K, and L).",
+        "bbox": [
+            75,
+            551,
+            468,
+            717
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Geometric Degradation Lack of geometrical features poses significant challenges to LiDAR odometry. The root of this issue often lies in the limited sensing capabilities of LiDAR sensors and the constraints due to their mechanical installation. SubT-MRS captures a variety of environments that exemplify such challenges. It includes long, featureless corridors (Figure 3 E) and staircases (Figure 3 C and G). These scenarios illustrate various forms of geometric degradation, contributing to the improvement of LiDAR odometry systems in challenging conditions.",
+        "bbox": [
+            75,
+            719,
+            468,
+            869
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Mixed Degradation A mixed visual and geometric degradation can further hinder the performance of SLAM sys",
+        "bbox": [
+            75,
+            869,
+            468,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "tems. Examples in SubT-MRS include long, dimly lit corridors (Figure 3 A, E, and F), poorly illuminated staircases (Figure 3 C), and environments affected by snowy weather (Figure 3 H, K, and L). In these settings, both LiDAR odometry and visual odometry are prone to failure due to the compounded effects of mixed degradation. The inclusion of such scenarios in SubT-MRS is crucial for evaluating and improving the resilience of multimodal SLAM algorithms.",
+        "bbox": [
+            496,
+            321,
+            890,
+            443
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1.3. Heterogeneous Robot Platforms",
+        "text_level": 1,
+        "bbox": [
+            498,
+            446,
+            795,
+            462
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Most of current datasets focus on single-robot systems, limiting multi-robot SLAM development as shown in Table 1. To address this, we employed diverse robot platforms, including aerial, legged, and wheeled robots, navigating through various environments from urban campuses to medical facilities and natural terrains like caves. This diversity offers a range of scenarios to enhance SLAM algorithms for effective operation in challenging conditions.",
+        "bbox": [
+            496,
+            464,
+            890,
+            584
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Extrinsic Calibration for Multiple Robots To ensure the multi-robot system shares a common coordinate system, we perform the extrinsic calibration process [38]. It has 2 steps: First, each robot runs Super Odometry [53] to generate its local map using LiDAR data and share it with other robots through a wireless network; Second, the remaining robots identify overlapping regions between their local maps and estimate the extrinsic parameters using GICP [40].",
+        "bbox": [
+            496,
+            585,
+            890,
+            708
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2. Ground Truth",
+        "text_level": 1,
+        "bbox": [
+            500,
+            714,
+            645,
+            729
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Ground Truth Map As depicted in Figure 4, we utilized the high-precision FARO Focus 3D S120 3D scanner for creating ground-truth maps. This cutting-edge 3D scanner can measure distances up to  $120\\mathrm{m}$ , with a maximum measurement rate of  $976\\mathrm{K}$  points per second. The accuracy of the ground truth map is noteworthy, maintaining a range error within  $\\pm 2\\mathrm{mm}$ . We have developed ground truth models for diverse environments, including subterranean areas like urban areas  $(350\\mathrm{m} \\times 350\\mathrm{m})$ , caves  $(150\\mathrm{m} \\times 200\\mathrm{m})$  and tunnels  $(100 \\times 200\\mathrm{m})$ , as well as indoor-outdoor spaces  $(200\\mathrm{m} \\times 200\\mathrm{m})$  within the CMU campus. To reduce drift, a",
+        "bbox": [
+            496,
+            734,
+            890,
+            900
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "22650",
+        "bbox": [
+            478,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/e99cad20d18ec8f0048688cceaa90f07e28c3d2bab5ffea665868a82026408ba.jpg",
+        "image_caption": [
+            "Figure 3. The SubT-MRS datasets were collected across diverse seasons, capturing environments with perceptual challenges such as poor illumination, darkness, and water puddles, where visual sensors falter. They also include geometrically complex areas like long featureless corridors and steep multi-floor structures, challenging LiDAR systems with potential drift. Moreover, these datasets cover conditions with airborne obscurants like dust, fog, snow, and smoke in tough environments, including caves, deserts, long tunnels, and off-road areas."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            80,
+            88,
+            890,
+            410
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/3c0d36f7d01aeff96cc49b14894396b1d84ab5655b90321852f2810c4bddc139.jpg",
+        "image_caption": [
+            "Figure 4. The SubT-MRS dataset facilitates the generation of high-precision ground truth maps and trajectories. Figure (a) shows the ground truth trajectory in multi-floor settings. Figure (b) displays ground truth maps for indoor and outdoor areas, encompassing long corridors, multi-floor structures, and open spaces. Figure (c) features photo-realistic scans based on our ground truth maps in cave environments."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            491,
+            851,
+            694
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "loop closure algorithm was applied to correct the poses. Notably,  $96\\%$  of the scans achieved a position uncertainty of less than  $2\\mathrm{mm}$ . This ground-truth map is crucial for SLAM development where precision is of utmost importance.",
+        "bbox": [
+            75,
+            773,
+            470,
+            835
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Ground Truth Trajectory Ground truth trajectories for all sequences are generated based on the ground truth map. In this process, we establish point-to-point, point-to-plane, and point-to-line correspondences[40] between the Ground",
+        "bbox": [
+            75,
+            839,
+            470,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Truth map and the current LiDAR scan. Our estimated ground truth trajectory is a comprehensive fusion of various constraints derived from the ground truth map, visual odometry, LiDAR odometry, and IMU measurements. This approach ensures the robustness of our solution across diverse environmental conditions. To verify accuracy, we reconstruct LiDAR maps from ground truth trajectory and compare it with our ground truth maps.",
+        "bbox": [
+            496,
+            773,
+            893,
+            897
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "22651",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/fdec7445e1994ac9e22a65403478b80a8d9b0efe3e082ff2af469ac640c8c455.jpg",
+        "table_caption": [
+            "Table 2. SLAM Challenge Results (Blue shadings indicate rankings of ATE and Robustness Metric; L: LiDAR I: IMU C: Camera)"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>#</td><td>Team</td><td>Method</td><td>Odometry Type</td><td>Device</td><td>RealTime (s)</td><td>CPU/GPU (%)</td><td>RAM (GB)</td><td>ATE↓</td><td>\\( {R}_{v} \\uparrow /{R}_{w} \\uparrow \\)</td><td>Sensors L I C</td></tr><tr><td>1</td><td>Liu et al</td><td>FAST-LIO2 [51], HBA [27]</td><td>Filter</td><td>Intel i7-9700K</td><td>51.310</td><td>98.667 / 0</td><td>4.052</td><td>0.588</td><td>0.517/0.770</td><td>✓✓</td></tr><tr><td>2</td><td>Yibin et al</td><td>LIO-EKF [45]</td><td>Filter</td><td>Intel i7-10700</td><td>0.006</td><td>52.167 / 0</td><td>0.072</td><td>4.313</td><td>0.441/0.574</td><td>✓✓</td></tr><tr><td>3</td><td>Weitong et al</td><td>FAST-LIO[2], Pose Graph[11]</td><td>Filter</td><td>Intel Xeon(R)E3-1240v5</td><td>0.125</td><td>22.63 / 0</td><td>4.305</td><td>0.663</td><td>0.473/0.747</td><td>✓✓</td></tr><tr><td>4</td><td>Kim et al</td><td>FAST-LIO2[50], Point-LIO[20], Quatro[25]</td><td>Filter</td><td>Intel i5-12500</td><td>0.268</td><td>101.108 / 0</td><td>55.64</td><td>3.825</td><td>0.479/0.615</td><td>✓✓</td></tr><tr><td>5</td><td>Zhong et al</td><td>DLO[8], Scan-Context++[22]</td><td>SW Opt</td><td>AMD Ryzen 9 5900x</td><td>0.027</td><td>13.289 / 0</td><td>1.174</td><td>1.209</td><td>0.276/0.486</td><td>✓✓</td></tr><tr><td>1</td><td>Peng et al</td><td>DVI-SLAM [32]</td><td>Learning</td><td>Intel i9-12900</td><td>183.233</td><td>- / 149</td><td>11 (4)</td><td>0.547</td><td>0.473/0.788</td><td>✓✓</td></tr><tr><td>2</td><td>Jiang et al</td><td>LET-NET[26], VINS-Mono[34]</td><td>Hybrid</td><td>Intel i5-9400</td><td>0.064</td><td>40.35 / 0</td><td>4.337</td><td>1.093</td><td>0.078/0.322</td><td>✓✓</td></tr><tr><td>3</td><td>Thien et al</td><td>VR-SLAM[31]</td><td>SW Opt</td><td>Intel i9-12900</td><td>0.142</td><td>176.44 / 0</td><td>9.111</td><td>3.037</td><td>0.083/0.372</td><td>✓✓</td></tr><tr><td>4</td><td>Li et al</td><td>ORB-SLAM3[4]</td><td>SW Opt.</td><td>Intel i7-10700</td><td>0.019</td><td>65.028 / 0</td><td>0.386</td><td>8.975</td><td>0.163/0.474</td><td>✓✓</td></tr></table>",
+        "bbox": [
+            80,
+            114,
+            903,
+            260
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/2ed6115a3035f29a5745994d8f0ae28f3eb62609250d5bd3571af50b6022515a.jpg",
+        "table_caption": [
+            "Table 3. Accuracy Performance on Geometric Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Team</td><td colspan=\"6\">Geometric Degradation (Real World)</td><td colspan=\"3\">Simulation</td><td colspan=\"3\">Mix Degradation</td><td rowspan=\"2\">Average</td></tr><tr><td>Urban</td><td>Tunnel</td><td>Cave</td><td>Nuclear_1</td><td>Nuclear_2</td><td>Laurel_Caverns</td><td>Factory</td><td>Ocean</td><td>Sewerage</td><td>Long Corridor</td><td>Multi Floor</td><td>Block Lidar</td></tr><tr><td>Liu et al1</td><td>0.307</td><td>0.095</td><td>0.629</td><td>0.122</td><td>0.235</td><td>0.260</td><td>0.889</td><td>0.757</td><td>0.978</td><td>1.454</td><td>0.401</td><td>0.934</td><td>0.588</td></tr><tr><td>Weitong et al2</td><td>0.26</td><td>0.096</td><td>0.617</td><td>0.120</td><td>0.222</td><td>0.402</td><td>0.998</td><td>0.770</td><td>1.586</td><td>1.254</td><td>0.577</td><td>1.056</td><td>0.663</td></tr><tr><td>Kim et al3</td><td>0.331</td><td>0.092</td><td>0.787</td><td>0.123</td><td>0.270</td><td>0.279</td><td>10.628</td><td>22.425</td><td>7.147</td><td>2.100</td><td>0.650</td><td>1.068</td><td>3.825</td></tr><tr><td>Yibin et al4</td><td>1.060</td><td>0.220</td><td>0.750</td><td>0.470</td><td>0.620</td><td>9.140</td><td>4.920</td><td>0.280</td><td>24.460</td><td>2.990</td><td>5.500</td><td>1.340</td><td>4.312</td></tr><tr><td>Zhong et al5</td><td>1.205</td><td>0.695</td><td>-</td><td>1.175</td><td>1.72</td><td>2.08</td><td>0.889</td><td>0.778</td><td>1.13</td><td>-</td><td>-</td><td>-</td><td>1.209*</td></tr><tr><td>Average</td><td>0.633</td><td>0.240</td><td>0.696</td><td>0.402</td><td>0.6134</td><td>2.432</td><td>3.665</td><td>5.002</td><td>7.060</td><td>1.950</td><td>1.782</td><td>1.099</td><td></td></tr></table>",
+        "bbox": [
+            81,
+            299,
+            883,
+            412
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.3. Robustness Metric",
+        "text_level": 1,
+        "bbox": [
+            76,
+            436,
+            256,
+            450
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As discussed in Sec. 1, existing evaluation metrics such as absolute trajectory error (ATE) [3] have limitations in evaluating the SLAM's robustness in real-world applications. ATE primarily focuses on the accuracy of trajectory and does not consider the completeness of trajectory (recall). Also, it can not effectively capture velocity changes that have a direct impact on the robot's safety. For example, a pose spike error might not be immediately noticeable in ATE evaluation, it may result in a crash in an aerial robot control. To address these gaps, we introduce a new robustness metric based on estimated velocity. That is because velocity estimation from SLAM is crucial for the robot's control system which directly impacts stability, affecting the robot's safety. The new robustness metric is the area under the curve (AUC) of the F-1 score:",
+        "bbox": [
+            75,
+            454,
+            472,
+            680
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nF _ {1} (e) = \\frac {2 P (e <   T) R (e <   T)}{P (e <   T) + R (e <   T)}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            158,
+            691,
+            468,
+            724
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where the precision  $P$  quantifies the precision of the estimated velocity as a percentage: how closely the estimated velocity points lie to the ground truth point and the recall rate  $R$  quantifies the estimated velocity's completeness: to what extent all the ground-truth points are covered. A high F-1 score can only be achieved by the velocity estimation that is both accurate and complete throughout the entire run. Specifically, an estimated error  $e$  is regarded as an inlier, if it is smaller than a threshold  $T$ . To scale the threshold  $T$  within the range of [0, 1], we apply exponential mapping  $\\exp(-10T)$  when calculating robustness metric.",
+        "bbox": [
+            75,
+            734,
+            472,
+            902
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Position Robustness  $R_{p}$  & Rotation Robustness  $R_{r}$  The AUC of F-1 score can be defined using both linear velocity and angular velocity, which can reflect the robustness of position and rotation estimation, respectively. Specifically, we define  $R_{p} = \\mathrm{AUC}(F_{1}(\\mathrm{v}_{\\mathrm{e}}))$  and  $R_{r} = \\mathrm{AUC}(F_{1}(\\omega_{\\mathrm{e}}))$ , where  $v_{e}$  and  $\\omega_{e}$  are the estimated error of linear velocity and rotation velocity, respectively. Note that not all SLAM solutions can output the velocity at the desired frequency. To address this, we use B-splines [30, 46] to obtain smooth trajectories, compute derivatives for smooth trajectories, and derive estimated linear velocity and angular velocity. For more details, please refer to the supplementary.",
+        "bbox": [
+            496,
+            436,
+            890,
+            618
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4. Open SLAM Challenge Results",
+        "text_level": 1,
+        "bbox": [
+            498,
+            628,
+            785,
+            645
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we will present our works from two perspectives: accuracy evaluation and robustness evaluation.",
+        "bbox": [
+            496,
+            648,
+            890,
+            679
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Accuracy Evaluation",
+        "text_level": 1,
+        "bbox": [
+            498,
+            686,
+            697,
+            700
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The results of the ICCV SLAM challenge underscore the necessity for advancements in system robustness. From 29 submissions, we identified 5 winners in the LiDAR category and 4 in the visual category. However, in the sensor fusion track, which addresses both visual and geometric degradation, no submissions met the criteria for success. This result reveals the existing SLAM systems still have lots of space to improve. Sequence characteristics will be detailed in the supplementary material. Table 2 shows the results for both the LiDAR and visual tracks. Unfortunately, there are no current solutions that can balance high accuracy and real-time performance in challenging environments. Since several environments lack geometric features or visual features,",
+        "bbox": [
+            496,
+            703,
+            890,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "22652",
+        "bbox": [
+            478,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/811b48b0e0fffb482362b1dad99129a7b483227a851a403173d18320f4a4f3dc.jpg",
+        "table_caption": [
+            "Table 4. Accuracy Performance on Visual Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Team</td><td colspan=\"6\">Visual Degradation (Real World)</td><td colspan=\"3\">Simulation</td><td rowspan=\"2\">Average</td></tr><tr><td>Lowlight 1</td><td>Lowlight 2</td><td>Over Exposure</td><td>Flash Light</td><td>Smoke Room</td><td>Outdoor Night</td><td>End of World</td><td>Moon</td><td>Western Desert</td></tr><tr><td>Peng et al1</td><td>1.063</td><td>1.637</td><td>0.503</td><td>0.44</td><td>0.153</td><td>0.827</td><td>0.038</td><td>0.195</td><td>0.070</td><td>0.547</td></tr><tr><td>Thien et al2</td><td>1.081</td><td>2.054</td><td>1.733</td><td>1.054</td><td>10.532</td><td>7.692</td><td>0.753</td><td>1.228</td><td>1.209</td><td>3.037</td></tr><tr><td>Jiang et al3</td><td>1.019</td><td>1.126</td><td>1.911</td><td>2.341</td><td>3.757</td><td>11.821</td><td>2.154</td><td>0.604</td><td>4.010</td><td>3.193</td></tr><tr><td>Li et al4</td><td>5.768</td><td>7.834</td><td>1.757</td><td>1.295</td><td>5.370</td><td>10.766</td><td>-</td><td>30.07</td><td>-</td><td>8.98*</td></tr><tr><td>Average</td><td>2.232</td><td>3.163</td><td>1.476</td><td>1.282</td><td>4.953</td><td>7.776</td><td>0.982</td><td>8.024</td><td>1.763</td><td></td></tr></table>",
+        "bbox": [
+            119,
+            114,
+            846,
+            213
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7b8e8b8ab874deecdf5e7d6a75971d02e8e9b765ff81b297d02a906c91599ca0.jpg",
+        "image_caption": [
+            "Figure 5. From left to right, it shows robustness metric  $R_{p}$  and  $R_{r}$  for LiDAR and visual sequences respectively. Note: This is a summary of results for all sequences, with weights based on the trajectory length. The area under the curve (AUC) represents the robustness  $(R_{p}, R_{r})$ . The x-axis shows velocity thresholds for classifying estimated velocities as inliers and the y-axis is F-1 score."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            138,
+            224,
+            316,
+            359
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/55cfc6031a312bd26283cd3f30215cbe3a40100a153c8fd52f3f5b533cf7c593.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            318,
+            226,
+            482,
+            359
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/1ddcbb645616be2d30870b196461b2b337a9d4108afc101e9f4d046bbd9ca7d8.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            483,
+            226,
+            658,
+            359
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/ddea1daf4391967666e91c1708573d8df339b08b3c6cbb9141b9eeb3029f5647.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            661,
+            226,
+            831,
+            359
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "the algorithms required more processing times to achieve reasonably high accuracy. Liu et al. at the University of Hong Kong, who achieved the highest accuracy, recorded an Absolute Trajectory Error (ATE) of  $0.588\\mathrm{m}$ , but each iteration took 51.3 seconds. In the visual track, Peng et al. from the Samsung Research Center attained a leading ATE of  $0.547\\mathrm{m}$ , but each iteration required 183 seconds.",
+        "bbox": [
+            75,
+            429,
+            470,
+            535
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "LiDAR Track Discussion To enhance robustness in all-weather environments, a pivotal question arises: what are the limits of current LiDAR SLAM algorithms? Table 3 presents a summary of the ATE/RPE errors observed in real-world geometrically degraded, simulated, and mixed degradation scenarios across the top five teams. In real-world geometrically degraded environments, we observed the first limitation: existing LiDAR solutions struggle in confined spaces such as caves. Almost all algorithms perform worse in Cave (Figure 3S) and Laureal_Cavern environments (Figure 3N) with average ATE 0.696 and 2.432 meters respectively. One reason is that cave environments is the most confined space which may lack sufficient geometric features, compared with other environments like tunnel (Figure 3J) and Urban (Figure 3B) and Nuclear scenes (Figure 3U).",
+        "bbox": [
+            75,
+            536,
+            468,
+            763
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In simulated environments, we encountered the second limitation: existing methods are tailored for fixed motion patterns like vehicles and struggle with unpredictable motion patterns. Since our simulation sequence is derived from TartanAir [47], featuring aggressive and random motion patterns, it might not follow the usual velocity distribution domain and has not been thoroughly tested. As shown in Table 3 (simulation), it resulted in significantly higher ATE errors  $(5.239\\mathrm{m})$  compared to real-world's  $(0.836\\mathrm{m})$ .",
+        "bbox": [
+            75,
+            763,
+            470,
+            902
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In mixed degradation environments, the third limitation we found is that existing methods cannot actively select the most informative measurements to adapt to new environments. For example, in the Long Corridor sequence (Figure 3 A) with low lighting and a featureless environment, the scenario is intricately designed for collaborative usage of LiDAR and Visual sensors. Algorithms are expected to utilize visual information in geometrically degraded environments while disregarding it in darkness. Similarly, the Block LiDAR sequences (Figure 3 V) aim to simulate sensor drop scenarios frequently encountered in robotic applications. These sequences involve alternating periods of LiDAR and Visual data loss, challenging algorithms to promptly detect sensor failures and switch to the other modality for SLAM. However, the average ATE error  $(1.61\\mathrm{m})$  in Table 3 (mixed degradation) is much higher than in geometric degradation scenarios  $(0.836\\mathrm{m})$ , suggesting the above limitations. Details are in the supplementary.",
+        "bbox": [
+            496,
+            429,
+            890,
+            700
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Visual Track Discussion To improve the robustness, a pivotal question arises: what are the limits of current visual SLAM? To what extent does image quality impact a visual SLAM system? Table 4 shows a summary of ATE errors from real-world and simulation scenarios from the awarded 4 teams. In real-world scenarios, the first limitation is the lack of anti-noise capability in current methods, especially in low-quality image settings. We assessed visual odometry accuracy in visually degraded environments, including low lighting, sunlight overexposure, flashing lights, smoke-filled rooms, and nighttime outdoor settings (Figure 3). Trajectories in conditions like the Smoke Room and Outdoor Night exhibit significantly higher average ATE er",
+        "bbox": [
+            496,
+            703,
+            892,
+            901
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "22653",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/673a090df4bcba9895606a892977e11654621b21cdf8b4bf2da0799f720b5c65.jpg",
+        "table_caption": [
+            "Table 5. Robustness Performance on Geometric Degradation. Red numbers are robustness ranking. Larger values indicate better robustness."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\" colspan=\"2\">Team</td><td colspan=\"6\">Geometric Degradation (Real World)</td><td colspan=\"3\">Simulation</td><td colspan=\"3\">Mix Degradation</td><td rowspan=\"2\">Average</td></tr><tr><td>Urban</td><td>Tunnel</td><td>Cave</td><td>Nuclear_1</td><td>Nuclear_2</td><td>Laurel_Caverns</td><td>Factory</td><td>Ocean</td><td>Sewerage</td><td>Long Corridor</td><td>Multi Floor</td><td>Block Lidar</td></tr><tr><td>Liu et al1</td><td rowspan=\"5\">Rp↑</td><td>0.811</td><td>0.865</td><td>0.736</td><td>0.747</td><td>0.504</td><td>0.816</td><td>0.157</td><td>0.135</td><td>0.135</td><td>0.396</td><td>0.529</td><td>0.371</td><td>0.516</td></tr><tr><td>Weitong et al2</td><td>0.773</td><td>0.838</td><td>0.690</td><td>0.871</td><td>0.739</td><td>0.152</td><td>0.153</td><td>0.149</td><td>0.081</td><td>0.407</td><td>0.456</td><td>0.345</td><td>0.471</td></tr><tr><td>Kim et al4</td><td>0.747</td><td>0.865</td><td>0.737</td><td>0.870</td><td>0.689</td><td>0.777</td><td>0.014</td><td>0.001</td><td>0.010</td><td>0.410</td><td>0.285</td><td>0.345</td><td>0.479</td></tr><tr><td>Yibin et al3</td><td>0.650</td><td>0.783</td><td>0.557</td><td>0.721</td><td>0.490</td><td>0.481</td><td>0.219</td><td>0.560</td><td>0.019</td><td>0.175</td><td>0.260</td><td>0.370</td><td>0.440</td></tr><tr><td>Zhong et al5</td><td>0.567</td><td>0.683</td><td>0.204</td><td>0.680</td><td>0.426</td><td>0.447</td><td>0.103</td><td>0.084</td><td>0.122</td><td>0</td><td>0</td><td>0</td><td>0.276</td></tr><tr><td>Liu et al1</td><td rowspan=\"5\">Rr↑</td><td>0.888</td><td>0.893</td><td>0.816</td><td>0.857</td><td>0.778</td><td>0.861</td><td>0.692</td><td>0.719</td><td>0.652</td><td>0.753</td><td>0.689</td><td>0.643</td><td>0.770</td></tr><tr><td>Weitong et al2</td><td>0.886</td><td>0.892</td><td>0.811</td><td>0.893</td><td>0.840</td><td>0.598</td><td>0.688</td><td>0.720</td><td>0.514</td><td>0.766</td><td>0.706</td><td>0.644</td><td>0.746</td></tr><tr><td>Kim et al4</td><td>0.680</td><td>0.893</td><td>0.816</td><td>0.893</td><td>0.803</td><td>0.808</td><td>0.198</td><td>0.125</td><td>0.300</td><td>0.802</td><td>0.422</td><td>0.642</td><td>0.615</td></tr><tr><td>Yibin et al3</td><td>0.624</td><td>0.707</td><td>0.462</td><td>0.733</td><td>0.540</td><td>0.467</td><td>0.862</td><td>0.962</td><td>0.496</td><td>0.477</td><td>0.342</td><td>0.213</td><td>0.573</td></tr><tr><td>Zhong et al5</td><td>0.731</td><td>0.745</td><td>0.437</td><td>0.757</td><td>0.603</td><td>0.523</td><td>0.688</td><td>0.704</td><td>0.643</td><td>0</td><td>0</td><td>0</td><td>0.485</td></tr><tr><td colspan=\"2\">AverageRp</td><td>0.710</td><td>0.807</td><td>0.585</td><td>0.778</td><td>0.570</td><td>0.535</td><td>0.130</td><td>0.186</td><td>0.073</td><td>0.278</td><td>0.306</td><td>0.286</td><td></td></tr><tr><td colspan=\"2\">Average Rr</td><td>0.762</td><td>0.826</td><td>0.668</td><td>0.827</td><td>0.713</td><td>0.651</td><td>0.626</td><td>0.646</td><td>0.521</td><td>0.560</td><td>0.432</td><td>0.428</td><td></td></tr></table>",
+        "bbox": [
+            112,
+            111,
+            854,
+            277
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/0554e414eda0c184fb5086e9c578d4cb0574e8ba3786ffd0e627be43dfee6a5e.jpg",
+        "table_caption": [
+            "Table 6. Robustness Performance on Visual Degradation. Red numbers are robustness ranking. Larger values indicate better robustness."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\" colspan=\"2\">Team</td><td colspan=\"6\">Visual Degradation (Real World)</td><td colspan=\"3\">Simulation</td><td rowspan=\"2\">Average</td></tr><tr><td>Lowlight 1</td><td>Lowlight 2</td><td>Over Exposure</td><td>Flash Light</td><td>Smoke Room</td><td>Outdoor Night</td><td>End of World</td><td>Moon</td><td>Western Desert</td></tr><tr><td>Peng et al1</td><td></td><td>0.357</td><td>0.227</td><td>0.264</td><td>0.203</td><td>0.536</td><td>0.270</td><td>0.699</td><td>0.893</td><td>0.806</td><td>0.472</td></tr><tr><td>Thien et al3</td><td></td><td>0.045</td><td>0.070</td><td>0.240</td><td>0.156</td><td>0.131</td><td>0.075</td><td>0</td><td>0.031</td><td>0</td><td>0.083</td></tr><tr><td>Jiang et al4</td><td>Rp↑</td><td>0.046</td><td>0.039</td><td>0.194</td><td>0.088</td><td>0.242</td><td>0.095</td><td>0</td><td>0</td><td>0</td><td>0.078</td></tr><tr><td>Li et al2</td><td></td><td>0.342</td><td>0.187</td><td>0.257</td><td>0.208</td><td>0.322</td><td>0.142</td><td>0</td><td>0.006</td><td>0</td><td>0.162</td></tr><tr><td>Peng et al1</td><td></td><td>0.641</td><td>0.581</td><td>0.744</td><td>0.650</td><td>0.878</td><td>0.670</td><td>0.975</td><td>0.975</td><td>0.974</td><td>0.787</td></tr><tr><td>Thien et al3</td><td></td><td>0.413</td><td>0.445</td><td>0.610</td><td>0.269</td><td>0.474</td><td>0.487</td><td>0.177</td><td>0.315</td><td>0</td><td>0.354</td></tr><tr><td>Jiang et al4</td><td>Rr↑</td><td>0.453</td><td>0.452</td><td>0.619</td><td>0.252</td><td>0.542</td><td>0.577</td><td>0.002</td><td>0</td><td>0.157</td><td>0.339</td></tr><tr><td>Li et al2</td><td></td><td>0.642</td><td>0.574</td><td>0.657</td><td>0.651</td><td>0.773</td><td>0.660</td><td>0.0</td><td>0.305</td><td>0</td><td>0.473</td></tr><tr><td colspan=\"2\">AverageRp</td><td>0.198</td><td>0.131</td><td>0.239</td><td>0.164</td><td>0.308</td><td>0.146</td><td>0.175</td><td>0.232</td><td>0.202</td><td></td></tr><tr><td colspan=\"2\">Average Rr</td><td>0.537</td><td>0.513</td><td>0.658</td><td>0.456</td><td>0.667</td><td>0.598</td><td>0.288</td><td>0.399</td><td>0.283</td><td></td></tr></table>",
+        "bbox": [
+            106,
+            311,
+            854,
+            470
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "rors (4.95m, 7.76m respectively). Smoke and night scenes challenge feature extraction and tracking and introduce sensor noise, emphasizing the need for robust anti-noise algorithms. The second limitation is that existing methods struggle to overcome aggressive motion. Even in simulations with relatively good image quality, most methods still show significant average ATE errors (8.024m) on the Moon sequence (Figure 3 Y). See more details in supplementary.",
+        "bbox": [
+            75,
+            484,
+            468,
+            607
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.2. Robustness Evaluation",
+        "text_level": 1,
+        "bbox": [
+            76,
+            611,
+            289,
+            625
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The F-1 robustness curve provides a detailed evaluation under different error tolerances, while  $R_{p}$  and  $R_{r}$  provide an overall measure of robustness that is not dependent on a specific decision threshold. This suggests the flexibility of our new robustness metric. Figure 5 clearly shows the robustness performance of all teams regarding position and rotation, highlighting the differences in robustness between the methods. The values of  $R_{p}$  and  $R_{r}$ , which are the AUC under F-1 curves, are displayed in the brackets of the legends. It reveals that Liu et al. and Peng et al. are the most robust solutions for LiDAR and visual tracks, respectively.",
+        "bbox": [
+            75,
+            628,
+            468,
+            794
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Is the robustness metric robust? We did extensive experiments on our robustness metric against diverse sensor degradation settings. Table 5 and 6 present a summary of robustness performance in real-world, simulated, and mixed degradation scenarios for both LiDAR and visual sequences. In Table 5, we observe that the average values of  $R_{p}$  and  $R_{r}$  for the mixed degradation environment (0.290,",
+        "bbox": [
+            75,
+            795,
+            470,
+            901
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "0.473) are significantly lower than geometric degradation environments (0.664, 0.754). This observation suggests that the majority of SLAM algorithms exhibit reduced robustness in mixed degradation environments as compared to geometrically degraded ones. This aligns with our expectations and verifies the effectiveness of the robustness metric.",
+        "bbox": [
+            496,
+            484,
+            890,
+            575
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our robustness ranking, indicated by small red numbers, differs from those on ATE ranking shown in Table 3 and 4. This is because our F-1 score-based metrics  $R_{p}$  and  $R_{r}$  jointly consider precision and recall rate to provide a balanced evaluation of SLAM performance, considering both metrics across the full spectrum of thresholds. In contrast, the ATE focuses solely on precision, neglecting the trajectory's completeness (recall). We also evaluate our metric on various synthetic trajectories in the supplementary.",
+        "bbox": [
+            496,
+            580,
+            890,
+            717
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            739,
+            617,
+            753
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We introduce SubT-MRS, a comprehensive SLAM dataset with various sensor data, locomotion patterns, and over 30 degradation types in simulation and real-world settings to push SLAM towards all-weather environments. Additionally, we introduce a new robustness metric to evaluate the reliability of SLAM systems, enhancing robot safety control. 29 teams have tested SubT-MRS in the organized SLAM challenge and we expect that it will serve as a critical benchmark for future SLAM development.",
+        "bbox": [
+            496,
+            763,
+            890,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "22654",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            80,
+            90,
+            173,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] https://www.darpa.mil/program/darpa-subterranean-challenge. 1",
+            "[2] Chunge Bai, Tao Xiao, Yajie Chen, Haoqian Wang, Fang Zhang, and Xiang Gao. Faster-lio: Lightweight tightly coupled lidar-inertial odometry using parallel sparse incremental voxels. IEEE Robotics and Automation Letters, 7(2):4861-4868, 2022. 6",
+            "[3] Michael Burri, Janosch Nikolic, Pascal Gohl, Thomas Schneider, Joern Rehder, Sammy Omari, Markus W Achtelik, and Roland Siegwart. The EuRoC micro aerial vehicle datasets. The International Journal of Robotics Research, 2016. 1, 2, 3, 6",
+            "[4] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 6",
+            "[5] Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel, and Juan D. Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6): 1874-1890, 2021. 1",
+            "[6] Nicholas Carlevaris-Bianco, Arash K Ushani, and Ryan M Eustice. University of michigan north campus long-term vision and lidar dataset. The International Journal of Robotics Research, 35(9):1023-1035, 2016. 3",
+            "[7] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J Taylor, and Kostas Daniilidis. M3ed: Multi-robot, multi-sensor, multi-environment event dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015–4022, 2023. 3",
+            "[8] Kenny Chen, Brett T. Lopez, Ali-akbar Agha-mohammadi, and Ankur Mehta. Direct lidar odometry: Fast localization with dense point clouds. IEEE Robotics and Automation Letters, 7(2):2000–2007, 2022. 6",
+            "[9] Long Chen, Yuchen Li, Chao Huang, Bai Li, Yang Xing, Daxin Tian, Li Li, Zhongxu Hu, Xiaoxiang Na, Zixuan Li, et al. Milestones in autonomous driving and intelligent vehicles: Survey of surveys. IEEE Transactions on Intelligent Vehicles, 8(2):1046-1056, 2022. 2",
+            "[10] Weichen Dai, Yu Zhang, Shenzhou Chen, Donglei Sun, and Da Kong. A multi-spectral dataset for evaluating motion estimation systems. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5560-5566. IEEE, 2021. 2",
+            "[11] Frank Dellaert, Michael Kaess, et al. Factor graphs for robot perception. Foundations and Trends® in Robotics, 6(1-2): 1-139, 2017. 6",
+            "[12] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the UZH-FPV drone racing dataset. In IEEE Int. Conf. Robot. Autom. (ICRA), 2019. 1, 3",
+            "[13] Rodolphe Dubois, Alexandre Eudes, and Vincent Frémont. Airmuseum: a heterogeneous multi-robot dataset for stereovisual and inertial simultaneous localization and mapping. In"
+        ],
+        "bbox": [
+            78,
+            109,
+            470,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "2020 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI), pages 166-172, 2020. 3",
+            "[14] Naser El-Sheimi, Haiying Hou, and Xiaoji Niu. Analysis and modeling of inertial sensors using allan variance. IEEE Transactions on Instrumentation and Measurement, 57(1): 140-149, 2008. 4",
+            "[15] Dapeng Feng, Yuhua Qi, Shipeng Zhong, Zhiqiang Chen, Yudu Jiao, Qiming Chen, Tao Jiang, and Hongbo Chen. S3e: A large-scale multimodal dataset for collaborative slam. arXiv preprint arXiv:2210.13723, 2022. 2",
+            "[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4340-4349, 2016. 3",
+            "[17] Wenliang Gao. A ros package tool to analyze the imu performance, 2018. 4",
+            "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI Vision Benchmark Suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 1",
+            "[19] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 3",
+            "[20] Dongjiao He, Wei Xu, Nan Chen, Fanze Kong, Chongjian Yuan, and Fu Zhang. Point-lio: Robust high-bandwidth light detection and ranging inertial odometry. Advanced Intelligent Systems, page 2200459, 2023. 6",
+            "[21] Michael Helmberger, Kristian Morin, Beda Berner, Nitish Kumar, Giovanni Cioffi, and Davide Scaramuzza. The hilti slam challenge dataset. IEEE Robotics and Automation Letters, 7(3):7518-7525, 2022. 1, 3",
+            "[22] Giseop Kim and Ayoung Kim. Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4802-4809. IEEE, 2018. 6",
+            "[23] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. Vivid++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, 2022. 2",
+            "[24] Keith YK Leung, Yoni Halpern, Timothy D Barfoot, and Hugh HT Liu. The utias multi-robot cooperative localization and mapping dataset. The International Journal of Robotics Research, 30(8):969–974, 2011. 2",
+            "[25] Hyungtae Lim, Suyong Yeon, Suyong Ryu, Yonghan Lee, Youngji Kim, Jaeseong Yun, Euigon Jung, Donghwan Lee, and Hyun Myung. A single correspondence is enough: Robust global registration to avoid degeneracy in urban environments. page Accepted. To appear, 2022. 6",
+            "[26] Yicheng Lin, Shuo Wang, Yunlong Jiang, and Bin Han. Breaking of brightness consistency in optical flow with a lightweight cnn network, 2023. 6",
+            "[27] Zheng Liu, Xiyuan Liu, and Fu Zhang. Efficient and consistent bundle adjustment on lidar point clouds. IEEE Transactions on Robotics, 2023. 6"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "22655",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[28] Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year,  $1000\\mathrm{km}$ : The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3-15, 2017. 3",
+            "[29] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research, 36(2):142-149, 2017. 2",
+            "[30] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Transactions on Robotics, 34(6): 1425-1440, 2018. 6",
+            "[31] Thien Hoang Nguyen, Shenghai Yuan, and Lihua Xie. Vrslam: A visual-range simultaneous localization and mapping system using monocular camera and ultra-wideband sensors, 2023. 6",
+            "[32] Xiongfeng Peng, Zhihua Liu, Weiming Li, Ping Tan, SoonYong Cho, and Qiang Wang. Dvi-slam: A dual visual inertial slam network, 2023. 6",
+            "[33] Bernd Pfrommer, Nitin Sanket, Kostas Daniilidis, and Jonas Cleveland. Penncosyvio: A challenging visual inertial odometry benchmark. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3847-3854. IEEE, 2017. 1, 3",
+            "[34] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics, 34(4):1004-1020, 2018. 1, 6",
+            "[35] Milad Ramezani, Yiduo Wang, Marco Camurri, David Wisth, Matias Mattamala, and Maurice Fallon. The newer college dataset: Handheld lidar, inertial and vision with ground truth. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4353-4360. IEEE, 2020. 1",
+            "[36] Joern Rehder, Janosch Nikolic, Thomas Schneider, Timo Hinzmann, and Roland Siegwart. Extending kalibr: Calibrating the extrinsic of multiple imus and of individual axes. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 4304-4311. IEEE, 2016. 3, 4",
+            "[37] John G Rogers, Arthur Schang, Carlos Nieto-Granda, John Ware, John Carter, Jonathan Fink, and Ethan Stump. The darpa subt urban circuit mapping dataset and evaluation metric. In Experimental Robotics: The 17th International Symposium, pages 391–401. Springer, 2021. 3",
+            "[38] Sebastian Scherer, Vasu Agrawal, Graeme Best, Chao Cao, Katarina Cujic, R Darnley, R DeBortoli, E Dexheimer, B Drozd, R Garg, et al. Resilient and modular subterranean exploration with a team of roving and flying robots. Field Robotics, 2022. 4",
+            "[39] David Schubert, Thore Goll, Nikolaus Demmel, Vladyslav Usenko, Jörg Stuckler, and Daniel Cremers. The tum vi benchmark for evaluating visual-inertial odometry. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1680-1687. IEEE, 2018. 1, 3",
+            "[40] Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-icp. In Robotics: science and systems, page 435. Seattle, WA, 2009. 4, 5"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[41] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Rus Daniela. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5135-5142. IEEE, 2020. 1",
+            "[42] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In Proc. of the International Conference on Intelligent Robot Systems (IROS), 2012. 2",
+            "[43] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1",
+            "[44] Yulun Tian, Yun Chang, Long Quang, Arthur Schang, Carlos Nieto-Granda, Jonathan P How, and Luca Carlone. Resilient and distributed multi-robot visual slam: Datasets, experiments, and lessons learned. arXiv preprint arXiv:2304.04362, 2023. 2",
+            "[45] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters, 8(2):1029–1036, 2023. 6",
+            "[46] Chen Wang, Dasong Gao, Kuan Xu, Junyi Geng, Yaoyu Hu, Yuheng Qiu, Bowen Li, Fan Yang, Brady Moon, Abhinav Pandey, Aryan, Jiahe Xu, Tianhao Wu, Haonan He, Daning Huang, Zhongqiang Ren, Shibo Zhao, Taimeng Fu, Pranay Reddy, Xiao Lin, Wenshan Wang, Jingnan Shi, Rajat Talak, Kun Cao, Yi Du, Han Wang, Huai Yu, Shanzhao Wang, Siyu Chen, Ananth Kashyap, Rohan Bandaru, Karthik Dantu, Jiajun Wu, Lihua Xie, Luca Carlone, Marco Hutter, and Sebastian Scherer. PyPose: A library for robot learning with physics-based optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6",
+            "[47] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual SLAM. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 3, 7",
+            "[48] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. Tartanvo: A generalizable learning-based vo. In Conference on Robot Learning, pages 1761-1772. PMLR, 2021. 1",
+            "[49] P. Wenzel, R. Wang, N. Yang, Q. Cheng, Q. Khan, L. von Stumberg, N. Zeller, and D. Cremers. 4Seasons: A cross-season dataset for multi-weather SLAM in autonomous driving. In Proceedings of the German Conference on Pattern Recognition (GCPR), 2020. 2",
+            "[50] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. FAST-LIO2: fast direct lidar-inertial odometry. IEEE Transactions on Robotics, abs/2107.06829, 2022. 6",
+            "[51] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. Fast-lio2: Fast direct lidar-inertial odometry. IEEE Transactions on Robotics, 38(4):2053–2073, 2022. 6",
+            "[52] Lintong Zhang, Marco Camurri, and Maurice Fallon. Multicamera lidar inertial extension to the newer college dataset. arXiv preprint arXiv:2112.08854, 2021. 3",
+            "[53] Shibo Zhao, Hengrui Zhang, Peng Wang, Lucas Nogueira, and Sebastian Scherer. Super odometry: Imu-centric"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "22656",
+        "bbox": [
+            478,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "lidar-visual-inertial estimator for challenging environments. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8729-8736. IEEE, 2021. 4",
+            "[54] David Zuniga-Noël, Alberto Jaenal, Ruben Gomez-Ojeda, and Javier Gonzalez-Jimenez. The uma-vi dataset: Visual-inertial odometry in low-textured and dynamic illumination environments. The International Journal of Robotics Research, 39(9):1052-1060, 2020. 1, 3"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            218
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "22657",
+        "bbox": [
+            478,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 10
+    }
+]

2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_model.json

@@ -0,0 +1,2048 @@
+[
+    [
+        {
+            "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": [
+                0.119,
+                0.13,
+                0.852,
+                0.176
+            ],
+            "angle": 0,
+            "content": "SubT-MRS Dataset: Pushing SLAM Towards All-weather Environments https://superodometry.com/datasets"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.086,
+                0.204,
+                0.897,
+                0.258
+            ],
+            "angle": 0,
+            "content": "Shibo Zhao\\*, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, Yi Du, Shaoshu Su, Can Xu, John Keller, Jay Karhade, Lucas Nogueira, Sourojit Saha, Ji Zhang, Wenshan Wang, Chen Wang, Sebastian Scherer"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.29,
+                0.262,
+                0.688,
+                0.28
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)Carnegie Mellon University \\(^{2}\\)University at Buffalo"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.235,
+                0.315,
+                0.314,
+                0.33
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.347,
+                0.474,
+                0.695
+            ],
+            "angle": 0,
+            "content": "Simultaneous localization and mapping (SLAM) is a fundamental task for numerous applications such as autonomous navigation and exploration. Despite many SLAM datasets have been released, current SLAM solutions still struggle to have sustained and resilient performance. One major issue is the absence of high-quality datasets including diverse all-weather conditions and a reliable metric for assessing robustness. This limitation significantly restricts the scalability and generalizability of SLAM technologies, impacting their development, validation, and deployment. To address this problem, we present SubT-MRS, an extremely challenging real-world dataset designed to push SLAM towards all-weather environments to pursue the most robust SLAM performance. It contains multi-degraded environments including over 30 diverse scenes such as structureless corridors, varying lighting conditions, and perceptual obscurants like smoke and dust; multimodal sensors such as LiDAR, fisheye camera, IMU, and thermal camera; and multiple locomotions like aerial, legged, and wheeled robots. We developed accuracy and robustness evaluation tracks for SLAM and introduced novel robustness metrics. Comprehensive studies are performed, revealing new observations, challenges, and opportunities for future research."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.078,
+                0.719,
+                0.21,
+                0.735
+            ],
+            "angle": 0,
+            "content": "1. Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.74,
+                0.47,
+                0.878
+            ],
+            "angle": 0,
+            "content": "Simultaneous Localization and Mapping (SLAM) is essential in robotics since it provides foundational perception and spatial awareness, and enables machines to understand and interact with the physical world in real-time. Therefore, it has a wide range of applications such as autonomous driving and space exploration. Despite significant advancements in both geometric [5, 34, 41] and data-driven SLAM methods [43, 48], existing solutions remain fragile in challenging conditions. One main reason is that existing SLAM"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.5,
+                0.316,
+                0.895,
+                0.422
+            ],
+            "angle": 0,
+            "content": "algorithms are often developed and evaluated with datasets from controlled environments [3, 12, 18, 21, 33, 35, 39, 54]. For example, KITTI [18, 21] dataset is mostly collected in sunny weather and EuROC-MAV dataset [3] is collected in small well-lit rooms. These datasets, unfortunately, fail to capture the challenges in real-world scenarios, hindering the development of robust SLAM solutions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.427,
+                0.895,
+                0.685
+            ],
+            "angle": 0,
+            "content": "To bridge this gap and push SLAM towards all-weather environments, we present an extremely challenging dataset, SubT-MRS, including scenarios featuring various sensor degradation, aggressive locomotions, and extreme-weather conditions. The SubT-MRS dataset comprises 3 years of data from the DARPA Subterranean (SubT) Challenge [1] (2019-2021) and extends with an additional 2 years of diverse environments (2022-2023), containing mixed indoor and outdoor settings, including long corridors, off-road scenario, tunnels, caves, deserts, forests, and bushlands. Cumulatively, this forms a 5-year dataset encompassing over 500 hours and 100 kilometers of terrain subjected to multimodal sensors including LiDAR, fisheye cameras, depth cameras, thermal cameras, and IMU; heterogeneous platforms including RC cars, legged robots, aerial robots, and wheeled robots; and extreme obscurant conditions such as dense fog, dust, smoke, and heavy snow."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.689,
+                0.895,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Additionally, we find there is no well-established metric to evaluate the robustness of a SLAM system. Existing evaluation metrics such as absolute trajectory error (ATE) [3] are not representative of the actual performance in robotic applications. We argue that to ensure the safety control of a robot, SLAM system evaluation should not only focus on poses but also its velocity. Taking robot localization as an example, while a momentary pose error spike may only slightly undermine the overall ATE, however, it can lead to a catastrophic crash in an aerial robotic platform. To effectively gauge the performance of a SLAM algorithm, we introduce a new robustness metric to evaluate the reliability of the SLAM system, particularly examining smoothness and accuracy of velocity estimation."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.096,
+                0.887,
+                0.361,
+                0.901
+            ],
+            "angle": 0,
+            "content": "*Corresponding author: shiboz@andrew.cmu.edu"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "22647"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.081,
+                0.09,
+                0.894,
+                0.293
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.081,
+                0.293,
+                0.894,
+                0.334
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.124,
+                0.336,
+                0.175,
+                0.344
+            ],
+            "angle": 0,
+            "content": "SubT-Tunnel"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.252,
+                0.336,
+                0.301,
+                0.344
+            ],
+            "angle": 0,
+            "content": "SubT-Urban"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.388,
+                0.336,
+                0.439,
+                0.344
+            ],
+            "angle": 0,
+            "content": "SubT-Cavern"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.525,
+                0.336,
+                0.574,
+                0.344
+            ],
+            "angle": 0,
+            "content": "Multi Floors"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.653,
+                0.336,
+                0.73,
+                0.345
+            ],
+            "angle": 0,
+            "content": "Dark Long Corridor"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.801,
+                0.336,
+                0.843,
+                0.345
+            ],
+            "angle": 0,
+            "content": "Dense Fog"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.081,
+                0.347,
+                0.894,
+                0.39
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.119,
+                0.392,
+                0.174,
+                0.4
+            ],
+            "angle": 0,
+            "content": "Laurel Cavern"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.235,
+                0.392,
+                0.314,
+                0.401
+            ],
+            "angle": 0,
+            "content": "Textureless Corridor"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.355,
+                0.392,
+                0.48,
+                0.4
+            ],
+            "angle": 0,
+            "content": "Darkness Stairs with Jerk Motion"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.484,
+                0.392,
+                0.596,
+                0.401
+            ],
+            "angle": 0,
+            "content": "Dense Dust and Smoke"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.625,
+                0.392,
+                0.751,
+                0.401
+            ],
+            "angle": 0,
+            "content": "Offroad with Aggressive Motion"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.771,
+                0.392,
+                0.875,
+                0.401
+            ],
+            "angle": 0,
+            "content": "Snow with Slipping Motion"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.412,
+                0.893,
+                0.456
+            ],
+            "angle": 0,
+            "content": "Figure 1. Dense reconstruction from the SubT-MRS dataset, achieved through collaboration with diverse robots equipped with multimodal sensors. Colors represent different challenging environments (tunnels, caves, urban, confined spaces) captured by various robot types (aerial, legged, wheeled). The bottom section showcases a gallery with diverse visual, LiDAR, and mixed degradations."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.481,
+                0.469,
+                0.557
+            ],
+            "angle": 0,
+            "content": "Lastly, we perform extensive experiments using proposed degradation datasets to benchmark visual and LiDAR SLAM algorithms. These experiments identify the limitations of the existing SLAM systems and evaluate their robustness across the degradation. Our contributions include:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.078,
+                0.558,
+                0.469,
+                0.693
+            ],
+            "angle": 0,
+            "content": "- All-weather Environments To push SLAM toward all-weather environments, we introduce SubT-MRS, an extremely challenging dataset. Spanning five years, it comprises over 500 hours and 100 kilometers of accurately measured trajectories. To the best of our knowledge, SubT-MRS is the first real-world dataset that specifically addresses failure scenarios of SLAM by incorporating a variety of degraded conditions, multiple robotic platforms, and diverse sets of multimodal sensors."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.078,
+                0.694,
+                0.469,
+                0.754
+            ],
+            "angle": 0,
+            "content": "- Robustness Metric To evaluate the actual performance running on robots, we propose a robustness metric, which, to the best of our knowledge, is the first metric evaluating the reliability, safety, and resilience of SLAM."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.078,
+                0.755,
+                0.469,
+                0.86
+            ],
+            "angle": 0,
+            "content": "- SLAM Challenge We provide a comprehensive benchmark<sup>1</sup> based on the SubT-MRS dataset by conducting ICCV'23 SLAM Challenge. Our evaluation reveals the limitations of state-of-the-art visual and LiDAR SLAM solutions. Furthermore, we conduct extensive experiments to verify robustness metric against various sensor degradation and length of trajectories."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.078,
+                0.558,
+                0.469,
+                0.86
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.479,
+                0.642,
+                0.495
+            ],
+            "angle": 0,
+            "content": "2. Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.504,
+                0.892,
+                0.806
+            ],
+            "angle": 0,
+            "content": "Multimodal Sensors Multimodal sensor datasets are crucial for the development of robust SLAM systems, as single sensor modalities are not comprehensive for all scenarios. Existing datasets typically focus on a limited range of modalities, such as monocular or stereo cameras [3, 49], LiDAR [19], event cameras [29], and depth cameras [42]. The KITTI dataset [19], although covering most modalities, is geared towards on-road scenarios and lacks thermal cameras, limiting its applicability to simple, controlled environments. Conversely, the ViViD++ dataset [23] incorporates thermal images in outdoor settings but is deficient in hardware synchronization, posing challenges for SLAM system development. The Weichen dataset [10] offers thermal images in indoor environments with accurate ground truth poses but is confined to motion capture room settings. In contrast, our SubT-MRS dataset delivers a comprehensive suite of time-synchronized multimodal sensors, including LiDAR, monocular cameras, thermal cameras, depth cameras, and IMU, along with centimeter-level ground truth, catering to a diverse range of research needs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.81,
+                0.892,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Heterogeneous Platforms Developing a robust SLAM system necessitates versatility in handling different motion patterns and various robotic platforms. However, most existing datasets are tailored for single-robot scenarios. While there are a few multi-robot datasets, they typically involve homogeneous platforms [9, 15, 24, 44], posing challenges"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.077,
+                0.875,
+                0.469,
+                0.9
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>SubT-MRS dataset is completely real-world data. Simulated data in the benchmark is to ensure evaluation completeness"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22648"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.081,
+                0.091,
+                0.891,
+                0.282
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"4\">Multi-Spectrum</td><td colspan=\"6\">Multi-Degradation</td><td colspan=\"4\">Multi-Robot</td></tr><tr><td>Camera</td><td>IMU</td><td>LiDAR/Depth</td><td>Thermal</td><td>Illumination</td><td>Snow Smoke</td><td>Structureless</td><td>SubT</td><td>Aggressive Motion</td><td>Vehicle</td><td>Drone</td><td>Legged</td><td>Handheld</td><td></td></tr><tr><td>EuRoC MAV [3]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td></td></tr><tr><td>PennCOsYVIO [33]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>TUM VIO [39]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>UZH-FPV [12]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td></td></tr><tr><td>College Dataset [52]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>RobotCar [28]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>UMA VI [54]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>UMich [6]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>KITTI [19]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>Virtual KITTI [16]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>DARPA SubT [37]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>Hilti SLAM [21]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>M3ED [7]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td></td></tr><tr><td>TartanAir [47]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td></td></tr><tr><td>SubT-MRS (Ours)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td></td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.205,
+                0.288,
+                0.761,
+                0.302
+            ],
+            "angle": 0,
+            "content": "Table 1. Comparison of SLAM datasets on multi-sensors, multi robots, and multi degradation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.329,
+                0.468,
+                0.42
+            ],
+            "angle": 0,
+            "content": "in evaluating SLAM performance across varied robotic platforms. The AirMuseum dataset [13] includes drones and ground robots but lacks data from legged robots. As a comparison, SubT-MRS dataset features diverse robotic platforms including legged robots, aerial robots, and wheeled robots, operating in varied environments and sensor setups."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.426,
+                0.468,
+                0.713
+            ],
+            "angle": 0,
+            "content": "Extreme Environmental Conditions Developing and testing SLAM systems in extreme environmental conditions is crucial for mitigating potential real-world failures. However, most existing SLAM datasets have been primarily limited to single, controlled environments. The TUM-VI [39] and UMA-VI [54] datasets, being indoor-outdoor visual-inertial datasets, pose challenges due to varying illumination and low-texture environments. The EuRoC MAV [3] and UZH-FPV drone racing [12] datasets, popular in the SLAM community, offer data on aggressive drone movements but usually in consistently lit conditions. The KITTI dataset [19], a staple in autonomous driving research for its outdoor LiDAR-Visual-Inertial data. The Hilti SLAM dataset [21] includes both indoor and outdoor LiDAR-visual-inertial datasets with dynamic lighting and confined spaces. However, both of them focus more on accuracy than on a variety of environmental degradations. TartanAir [47] covers most of the challenging environments but it is a simulation dataset that will pose the sim-to-real gap."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.72,
+                0.468,
+                0.901
+            ],
+            "angle": 0,
+            "content": "In contrast, the SubT-MRS dataset provides 30 varied scenes, encompassing a wide array of environments. These include SubT tunnels, urban areas, caverns, multifloor structures, dark corridors, and foggy conditions. It also features textureless surfaces, dusty and smoky environments, off-road areas with aggressive motion, and snowy terrains prone to slippage. We include challenging lighting conditions such as extreme darkness and overexposure, along with geometrically challenging scenarios like featureless corridors, staircases, and self-similar cave layouts. Additionally, the dataset incorporates perceptual challenges posed by smoke, fog, dust, and various weather conditions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.327,
+                0.687,
+                0.343
+            ],
+            "angle": 0,
+            "content": "3. SubT-MRS Dataset"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.349,
+                0.892,
+                0.407
+            ],
+            "angle": 0,
+            "content": "We next present the SubT-MRS dataset from three aspects including its collecting environments and settings, ground truth collection, and the new robustness metric. A detailed comparison with other datasets is listed in Table 1."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.408,
+                0.74,
+                0.423
+            ],
+            "angle": 0,
+            "content": "3.1. All-weather Environments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.426,
+                0.892,
+                0.501
+            ],
+            "angle": 0,
+            "content": "As mentioned in Sec. 2, the SubT-MRS dataset distinguishes itself through its multimodal sensor setups, heterogeneous robotic platforms, and extreme environmental conditions. This section emphasizes this comprehensive nature pushing SLAM towards all-weather environments."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.505,
+                0.704,
+                0.519
+            ],
+            "angle": 0,
+            "content": "3.1.1. Multimodal Sensors"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.522,
+                0.892,
+                0.597
+            ],
+            "angle": 0,
+            "content": "Multimodal sensors provide a wealth of information for SLAM systems operating in all-weather environments. This diversity improves scene understanding and strengthens the system's perception ability. Therefore, We incorporate diverse sensors to ensure system robustness."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.598,
+                0.892,
+                0.658
+            ],
+            "angle": 0,
+            "content": "Sensor Pack We embedded 4 Leopard Imaging RGB fisheye cameras, 1 Velodyne puck, 1 Epson M-G365 IMU, 1 FLIR Boson thermal camera, and NVIDIA Jetson AGX Xavier as a sensor pack shown in Figure 2."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.659,
+                0.892,
+                0.78
+            ],
+            "angle": 0,
+            "content": "Time Synchronization To ensure the overall consistency of the fused data, we meticulously synchronize the sensors' time using the 'pulse per second (PPS)' technique. As illustrated in Figure 2, the IMU, LiDAR, and thermal camera are directly synchronized with the CPU clock, while the four RGB camera are synchronized using an FPGA board. Consequently, we can effectively manage the time synchronization gap between each pair of sensors not to exceed 3ms."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.78,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Fisheye and Thermal Camera Calibration Camera calibration plays an essential role in the efficiency of a SLAM system. For calibrating the fisheye cameras, we employed the open-source toolkit Kalibr [36], focusing on the intrinsic and extrinsic parameters. Specifically, we use the radial-tangential distortion model to rectify the omnidirectional fisheye camera model. Calibrating thermal cameras, however, poses unique challenges, especially in gathering high-"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.945,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22649"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.119,
+                0.089,
+                0.852,
+                0.257
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.267,
+                0.895,
+                0.297
+            ],
+            "angle": 0,
+            "content": "Figure 2. An overview of the sensor pack used in SubT-MRS dataset. It is equipped with a Xavier processing unit with hardware time synchronization for multimodal sensors including LiDAR, fisheye cameras, thermal cameras, depth cameras (option), and an IMU."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.322,
+                0.47,
+                0.367
+            ],
+            "angle": 0,
+            "content": "quality thermal data. To tackle this, we set up a \\(7 \\times 9\\) chessboard, heated by direct sunlight, to generate high-contrast thermal data, ensuring accuracy in the calibration process."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.367,
+                0.47,
+                0.488
+            ],
+            "angle": 0,
+            "content": "IMU Calibration and Extrinsic Calibration For Lidar-IMU extrinsic calibration, we utilize the CAD model to obtain the calibration parameters. In the case of Camera-IMU extrinsic calibration, we employ the Kalibr toolbox [36] to estimate the extrinsic matrix. To estimate the sensors' bias and the random walk noise of the gyroscope and accelerometer, we collected static data from the IMU and calibrated it using an Allan variance-based IMU tool [14, 17]."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.491,
+                0.362,
+                0.507
+            ],
+            "angle": 0,
+            "content": "3.1.2. Multi-degraded Environments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.508,
+                0.47,
+                0.553
+            ],
+            "angle": 0,
+            "content": "SubT-MRS includes multiple challenging environments including visually degraded environments, geometrically degraded environments, and their combination."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.553,
+                0.47,
+                0.718
+            ],
+            "angle": 0,
+            "content": "Visual Degradation Poor-quality visual features can significantly hinder the performance of feature extraction processes and disrupt the accuracy of feature matching. Such issues arise in low-light conditions with inconsistent brightness or image noise introduced by air obscurants. SubT-MRS encompasses these challenges and provides a wide range of visual degradation. This includes environments with limited lighting, such as hospital interiors and caves (Figure 3 A-F), as well as smoky or dusty conditions that cause visual obstruction (Figure 3 M-N), and snowy areas with reduced visibility (Figure 3 H, K, and L)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.72,
+                0.47,
+                0.87
+            ],
+            "angle": 0,
+            "content": "Geometric Degradation Lack of geometrical features poses significant challenges to LiDAR odometry. The root of this issue often lies in the limited sensing capabilities of LiDAR sensors and the constraints due to their mechanical installation. SubT-MRS captures a variety of environments that exemplify such challenges. It includes long, featureless corridors (Figure 3 E) and staircases (Figure 3 C and G). These scenarios illustrate various forms of geometric degradation, contributing to the improvement of LiDAR odometry systems in challenging conditions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.871,
+                0.469,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Mixed Degradation A mixed visual and geometric degradation can further hinder the performance of SLAM sys"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.322,
+                0.892,
+                0.444
+            ],
+            "angle": 0,
+            "content": "tems. Examples in SubT-MRS include long, dimly lit corridors (Figure 3 A, E, and F), poorly illuminated staircases (Figure 3 C), and environments affected by snowy weather (Figure 3 H, K, and L). In these settings, both LiDAR odometry and visual odometry are prone to failure due to the compounded effects of mixed degradation. The inclusion of such scenarios in SubT-MRS is crucial for evaluating and improving the resilience of multimodal SLAM algorithms."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.447,
+                0.796,
+                0.463
+            ],
+            "angle": 0,
+            "content": "3.1.3. Heterogeneous Robot Platforms"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.465,
+                0.892,
+                0.585
+            ],
+            "angle": 0,
+            "content": "Most of current datasets focus on single-robot systems, limiting multi-robot SLAM development as shown in Table 1. To address this, we employed diverse robot platforms, including aerial, legged, and wheeled robots, navigating through various environments from urban campuses to medical facilities and natural terrains like caves. This diversity offers a range of scenarios to enhance SLAM algorithms for effective operation in challenging conditions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.587,
+                0.892,
+                0.709
+            ],
+            "angle": 0,
+            "content": "Extrinsic Calibration for Multiple Robots To ensure the multi-robot system shares a common coordinate system, we perform the extrinsic calibration process [38]. It has 2 steps: First, each robot runs Super Odometry [53] to generate its local map using LiDAR data and share it with other robots through a wireless network; Second, the remaining robots identify overlapping regions between their local maps and estimate the extrinsic parameters using GICP [40]."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.715,
+                0.646,
+                0.73
+            ],
+            "angle": 0,
+            "content": "3.2. Ground Truth"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.735,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "Ground Truth Map As depicted in Figure 4, we utilized the high-precision FARO Focus 3D S120 3D scanner for creating ground-truth maps. This cutting-edge 3D scanner can measure distances up to \\(120\\mathrm{m}\\), with a maximum measurement rate of \\(976\\mathrm{K}\\) points per second. The accuracy of the ground truth map is noteworthy, maintaining a range error within \\(\\pm 2\\mathrm{mm}\\). We have developed ground truth models for diverse environments, including subterranean areas like urban areas \\((350\\mathrm{m} \\times 350\\mathrm{m})\\), caves \\((150\\mathrm{m} \\times 200\\mathrm{m})\\) and tunnels \\((100 \\times 200\\mathrm{m})\\), as well as indoor-outdoor spaces \\((200\\mathrm{m} \\times 200\\mathrm{m})\\) within the CMU campus. To reduce drift, a"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.945,
+                0.521,
+                0.958
+            ],
+            "angle": 0,
+            "content": "22650"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.081,
+                0.089,
+                0.891,
+                0.411
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.421,
+                0.896,
+                0.478
+            ],
+            "angle": 0,
+            "content": "Figure 3. The SubT-MRS datasets were collected across diverse seasons, capturing environments with perceptual challenges such as poor illumination, darkness, and water puddles, where visual sensors falter. They also include geometrically complex areas like long featureless corridors and steep multi-floor structures, challenging LiDAR systems with potential drift. Moreover, these datasets cover conditions with airborne obscurants like dust, fog, snow, and smoke in tough environments, including caves, deserts, long tunnels, and off-road areas."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.119,
+                0.492,
+                0.852,
+                0.695
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.707,
+                0.895,
+                0.751
+            ],
+            "angle": 0,
+            "content": "Figure 4. The SubT-MRS dataset facilitates the generation of high-precision ground truth maps and trajectories. Figure (a) shows the ground truth trajectory in multi-floor settings. Figure (b) displays ground truth maps for indoor and outdoor areas, encompassing long corridors, multi-floor structures, and open spaces. Figure (c) features photo-realistic scans based on our ground truth maps in cave environments."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.775,
+                0.471,
+                0.837
+            ],
+            "angle": 0,
+            "content": "loop closure algorithm was applied to correct the poses. Notably, \\(96\\%\\) of the scans achieved a position uncertainty of less than \\(2\\mathrm{mm}\\). This ground-truth map is crucial for SLAM development where precision is of utmost importance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.84,
+                0.471,
+                0.903
+            ],
+            "angle": 0,
+            "content": "Ground Truth Trajectory Ground truth trajectories for all sequences are generated based on the ground truth map. In this process, we establish point-to-point, point-to-plane, and point-to-line correspondences[40] between the Ground"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.775,
+                0.895,
+                0.898
+            ],
+            "angle": 0,
+            "content": "Truth map and the current LiDAR scan. Our estimated ground truth trajectory is a comprehensive fusion of various constraints derived from the ground truth map, visual odometry, LiDAR odometry, and IMU measurements. This approach ensures the robustness of our solution across diverse environmental conditions. To verify accuracy, we reconstruct LiDAR maps from ground truth trajectory and compare it with our ground truth maps."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.946,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22651"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.1,
+                0.09,
+                0.872,
+                0.104
+            ],
+            "angle": 0,
+            "content": "Table 2. SLAM Challenge Results (Blue shadings indicate rankings of ATE and Robustness Metric; L: LiDAR I: IMU C: Camera)"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.081,
+                0.115,
+                0.905,
+                0.261
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>#</td><td>Team</td><td>Method</td><td>Odometry Type</td><td>Device</td><td>RealTime (s)</td><td>CPU/GPU (%)</td><td>RAM (GB)</td><td>ATE↓</td><td>\\( {R}_{v} \\uparrow /{R}_{w} \\uparrow \\)</td><td>Sensors L I C</td></tr><tr><td>1</td><td>Liu et al</td><td>FAST-LIO2 [51], HBA [27]</td><td>Filter</td><td>Intel i7-9700K</td><td>51.310</td><td>98.667 / 0</td><td>4.052</td><td>0.588</td><td>0.517/0.770</td><td>✓✓</td></tr><tr><td>2</td><td>Yibin et al</td><td>LIO-EKF [45]</td><td>Filter</td><td>Intel i7-10700</td><td>0.006</td><td>52.167 / 0</td><td>0.072</td><td>4.313</td><td>0.441/0.574</td><td>✓✓</td></tr><tr><td>3</td><td>Weitong et al</td><td>FAST-LIO[2], Pose Graph[11]</td><td>Filter</td><td>Intel Xeon(R)E3-1240v5</td><td>0.125</td><td>22.63 / 0</td><td>4.305</td><td>0.663</td><td>0.473/0.747</td><td>✓✓</td></tr><tr><td>4</td><td>Kim et al</td><td>FAST-LIO2[50], Point-LIO[20], Quatro[25]</td><td>Filter</td><td>Intel i5-12500</td><td>0.268</td><td>101.108 / 0</td><td>55.64</td><td>3.825</td><td>0.479/0.615</td><td>✓✓</td></tr><tr><td>5</td><td>Zhong et al</td><td>DLO[8], Scan-Context++[22]</td><td>SW Opt</td><td>AMD Ryzen 9 5900x</td><td>0.027</td><td>13.289 / 0</td><td>1.174</td><td>1.209</td><td>0.276/0.486</td><td>✓✓</td></tr><tr><td>1</td><td>Peng et al</td><td>DVI-SLAM [32]</td><td>Learning</td><td>Intel i9-12900</td><td>183.233</td><td>- / 149</td><td>11 (4)</td><td>0.547</td><td>0.473/0.788</td><td>✓✓</td></tr><tr><td>2</td><td>Jiang et al</td><td>LET-NET[26], VINS-Mono[34]</td><td>Hybrid</td><td>Intel i5-9400</td><td>0.064</td><td>40.35 / 0</td><td>4.337</td><td>1.093</td><td>0.078/0.322</td><td>✓✓</td></tr><tr><td>3</td><td>Thien et al</td><td>VR-SLAM[31]</td><td>SW Opt</td><td>Intel i9-12900</td><td>0.142</td><td>176.44 / 0</td><td>9.111</td><td>3.037</td><td>0.083/0.372</td><td>✓✓</td></tr><tr><td>4</td><td>Li et al</td><td>ORB-SLAM3[4]</td><td>SW Opt.</td><td>Intel i7-10700</td><td>0.019</td><td>65.028 / 0</td><td>0.386</td><td>8.975</td><td>0.163/0.474</td><td>✓✓</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.091,
+                0.274,
+                0.877,
+                0.289
+            ],
+            "angle": 0,
+            "content": "Table 3. Accuracy Performance on Geometric Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.082,
+                0.3,
+                0.885,
+                0.413
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Team</td><td colspan=\"6\">Geometric Degradation (Real World)</td><td colspan=\"3\">Simulation</td><td colspan=\"3\">Mix Degradation</td><td rowspan=\"2\">Average</td></tr><tr><td>Urban</td><td>Tunnel</td><td>Cave</td><td>Nuclear_1</td><td>Nuclear_2</td><td>Laurel_Caverns</td><td>Factory</td><td>Ocean</td><td>Sewerage</td><td>Long Corridor</td><td>Multi Floor</td><td>Block Lidar</td></tr><tr><td>Liu et al1</td><td>0.307</td><td>0.095</td><td>0.629</td><td>0.122</td><td>0.235</td><td>0.260</td><td>0.889</td><td>0.757</td><td>0.978</td><td>1.454</td><td>0.401</td><td>0.934</td><td>0.588</td></tr><tr><td>Weitong et al2</td><td>0.26</td><td>0.096</td><td>0.617</td><td>0.120</td><td>0.222</td><td>0.402</td><td>0.998</td><td>0.770</td><td>1.586</td><td>1.254</td><td>0.577</td><td>1.056</td><td>0.663</td></tr><tr><td>Kim et al3</td><td>0.331</td><td>0.092</td><td>0.787</td><td>0.123</td><td>0.270</td><td>0.279</td><td>10.628</td><td>22.425</td><td>7.147</td><td>2.100</td><td>0.650</td><td>1.068</td><td>3.825</td></tr><tr><td>Yibin et al4</td><td>1.060</td><td>0.220</td><td>0.750</td><td>0.470</td><td>0.620</td><td>9.140</td><td>4.920</td><td>0.280</td><td>24.460</td><td>2.990</td><td>5.500</td><td>1.340</td><td>4.312</td></tr><tr><td>Zhong et al5</td><td>1.205</td><td>0.695</td><td>-</td><td>1.175</td><td>1.72</td><td>2.08</td><td>0.889</td><td>0.778</td><td>1.13</td><td>-</td><td>-</td><td>-</td><td>1.209*</td></tr><tr><td>Average</td><td>0.633</td><td>0.240</td><td>0.696</td><td>0.402</td><td>0.6134</td><td>2.432</td><td>3.665</td><td>5.002</td><td>7.060</td><td>1.950</td><td>1.782</td><td>1.099</td><td></td></tr></table>"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.437,
+                0.258,
+                0.451
+            ],
+            "angle": 0,
+            "content": "3.3. Robustness Metric"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.455,
+                0.473,
+                0.681
+            ],
+            "angle": 0,
+            "content": "As discussed in Sec. 1, existing evaluation metrics such as absolute trajectory error (ATE) [3] have limitations in evaluating the SLAM's robustness in real-world applications. ATE primarily focuses on the accuracy of trajectory and does not consider the completeness of trajectory (recall). Also, it can not effectively capture velocity changes that have a direct impact on the robot's safety. For example, a pose spike error might not be immediately noticeable in ATE evaluation, it may result in a crash in an aerial robot control. To address these gaps, we introduce a new robustness metric based on estimated velocity. That is because velocity estimation from SLAM is crucial for the robot's control system which directly impacts stability, affecting the robot's safety. The new robustness metric is the area under the curve (AUC) of the F-1 score:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.16,
+                0.692,
+                0.47,
+                0.725
+            ],
+            "angle": 0,
+            "content": "\\[\nF _ {1} (e) = \\frac {2 P (e <   T) R (e <   T)}{P (e <   T) + R (e <   T)}, \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.735,
+                0.473,
+                0.903
+            ],
+            "angle": 0,
+            "content": "where the precision \\( P \\) quantifies the precision of the estimated velocity as a percentage: how closely the estimated velocity points lie to the ground truth point and the recall rate \\( R \\) quantifies the estimated velocity's completeness: to what extent all the ground-truth points are covered. A high F-1 score can only be achieved by the velocity estimation that is both accurate and complete throughout the entire run. Specifically, an estimated error \\( e \\) is regarded as an inlier, if it is smaller than a threshold \\( T \\). To scale the threshold \\( T \\) within the range of [0, 1], we apply exponential mapping \\( \\exp(-10T) \\) when calculating robustness metric."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.437,
+                0.892,
+                0.619
+            ],
+            "angle": 0,
+            "content": "Position Robustness \\( R_{p} \\) & Rotation Robustness \\( R_{r} \\) The AUC of F-1 score can be defined using both linear velocity and angular velocity, which can reflect the robustness of position and rotation estimation, respectively. Specifically, we define \\( R_{p} = \\mathrm{AUC}(F_{1}(\\mathrm{v}_{\\mathrm{e}})) \\) and \\( R_{r} = \\mathrm{AUC}(F_{1}(\\omega_{\\mathrm{e}})) \\), where \\( v_{e} \\) and \\( \\omega_{e} \\) are the estimated error of linear velocity and rotation velocity, respectively. Note that not all SLAM solutions can output the velocity at the desired frequency. To address this, we use B-splines [30, 46] to obtain smooth trajectories, compute derivatives for smooth trajectories, and derive estimated linear velocity and angular velocity. For more details, please refer to the supplementary."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.629,
+                0.787,
+                0.646
+            ],
+            "angle": 0,
+            "content": "4. Open SLAM Challenge Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.65,
+                0.892,
+                0.68
+            ],
+            "angle": 0,
+            "content": "In this section, we will present our works from two perspectives: accuracy evaluation and robustness evaluation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.499,
+                0.687,
+                0.698,
+                0.702
+            ],
+            "angle": 0,
+            "content": "4.1. Accuracy Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.704,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "The results of the ICCV SLAM challenge underscore the necessity for advancements in system robustness. From 29 submissions, we identified 5 winners in the LiDAR category and 4 in the visual category. However, in the sensor fusion track, which addresses both visual and geometric degradation, no submissions met the criteria for success. This result reveals the existing SLAM systems still have lots of space to improve. Sequence characteristics will be detailed in the supplementary material. Table 2 shows the results for both the LiDAR and visual tracks. Unfortunately, there are no current solutions that can balance high accuracy and real-time performance in challenging environments. Since several environments lack geometric features or visual features,"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.946,
+                0.521,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22652"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.104,
+                0.09,
+                0.864,
+                0.104
+            ],
+            "angle": 0,
+            "content": "Table 4. Accuracy Performance on Visual Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.12,
+                0.115,
+                0.847,
+                0.214
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Team</td><td colspan=\"6\">Visual Degradation (Real World)</td><td colspan=\"3\">Simulation</td><td rowspan=\"2\">Average</td></tr><tr><td>Lowlight 1</td><td>Lowlight 2</td><td>Over Exposure</td><td>Flash Light</td><td>Smoke Room</td><td>Outdoor Night</td><td>End of World</td><td>Moon</td><td>Western Desert</td></tr><tr><td>Peng et al1</td><td>1.063</td><td>1.637</td><td>0.503</td><td>0.44</td><td>0.153</td><td>0.827</td><td>0.038</td><td>0.195</td><td>0.070</td><td>0.547</td></tr><tr><td>Thien et al2</td><td>1.081</td><td>2.054</td><td>1.733</td><td>1.054</td><td>10.532</td><td>7.692</td><td>0.753</td><td>1.228</td><td>1.209</td><td>3.037</td></tr><tr><td>Jiang et al3</td><td>1.019</td><td>1.126</td><td>1.911</td><td>2.341</td><td>3.757</td><td>11.821</td><td>2.154</td><td>0.604</td><td>4.010</td><td>3.193</td></tr><tr><td>Li et al4</td><td>5.768</td><td>7.834</td><td>1.757</td><td>1.295</td><td>5.370</td><td>10.766</td><td>-</td><td>30.07</td><td>-</td><td>8.98*</td></tr><tr><td>Average</td><td>2.232</td><td>3.163</td><td>1.476</td><td>1.282</td><td>4.953</td><td>7.776</td><td>0.982</td><td>8.024</td><td>1.763</td><td></td></tr></table>"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.139,
+                0.226,
+                0.318,
+                0.361
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.319,
+                0.227,
+                0.483,
+                0.361
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.484,
+                0.227,
+                0.659,
+                0.361
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.662,
+                0.227,
+                0.832,
+                0.361
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.076,
+                0.372,
+                0.895,
+                0.415
+            ],
+            "angle": 0,
+            "content": "Figure 5. From left to right, it shows robustness metric \\( R_{p} \\) and \\( R_{r} \\) for LiDAR and visual sequences respectively. Note: This is a summary of results for all sequences, with weights based on the trajectory length. The area under the curve (AUC) represents the robustness \\( (R_{p}, R_{r}) \\). The x-axis shows velocity thresholds for classifying estimated velocities as inliers and the y-axis is F-1 score."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.43,
+                0.471,
+                0.536
+            ],
+            "angle": 0,
+            "content": "the algorithms required more processing times to achieve reasonably high accuracy. Liu et al. at the University of Hong Kong, who achieved the highest accuracy, recorded an Absolute Trajectory Error (ATE) of \\(0.588\\mathrm{m}\\), but each iteration took 51.3 seconds. In the visual track, Peng et al. from the Samsung Research Center attained a leading ATE of \\(0.547\\mathrm{m}\\), but each iteration required 183 seconds."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.537,
+                0.47,
+                0.765
+            ],
+            "angle": 0,
+            "content": "LiDAR Track Discussion To enhance robustness in all-weather environments, a pivotal question arises: what are the limits of current LiDAR SLAM algorithms? Table 3 presents a summary of the ATE/RPE errors observed in real-world geometrically degraded, simulated, and mixed degradation scenarios across the top five teams. In real-world geometrically degraded environments, we observed the first limitation: existing LiDAR solutions struggle in confined spaces such as caves. Almost all algorithms perform worse in Cave (Figure 3S) and Laureal_Cavern environments (Figure 3N) with average ATE 0.696 and 2.432 meters respectively. One reason is that cave environments is the most confined space which may lack sufficient geometric features, compared with other environments like tunnel (Figure 3J) and Urban (Figure 3B) and Nuclear scenes (Figure 3U)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.765,
+                0.471,
+                0.903
+            ],
+            "angle": 0,
+            "content": "In simulated environments, we encountered the second limitation: existing methods are tailored for fixed motion patterns like vehicles and struggle with unpredictable motion patterns. Since our simulation sequence is derived from TartanAir [47], featuring aggressive and random motion patterns, it might not follow the usual velocity distribution domain and has not been thoroughly tested. As shown in Table 3 (simulation), it resulted in significantly higher ATE errors \\((5.239\\mathrm{m})\\) compared to real-world's \\((0.836\\mathrm{m})\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.43,
+                0.892,
+                0.702
+            ],
+            "angle": 0,
+            "content": "In mixed degradation environments, the third limitation we found is that existing methods cannot actively select the most informative measurements to adapt to new environments. For example, in the Long Corridor sequence (Figure 3 A) with low lighting and a featureless environment, the scenario is intricately designed for collaborative usage of LiDAR and Visual sensors. Algorithms are expected to utilize visual information in geometrically degraded environments while disregarding it in darkness. Similarly, the Block LiDAR sequences (Figure 3 V) aim to simulate sensor drop scenarios frequently encountered in robotic applications. These sequences involve alternating periods of LiDAR and Visual data loss, challenging algorithms to promptly detect sensor failures and switch to the other modality for SLAM. However, the average ATE error \\((1.61\\mathrm{m})\\) in Table 3 (mixed degradation) is much higher than in geometric degradation scenarios \\((0.836\\mathrm{m})\\), suggesting the above limitations. Details are in the supplementary."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.704,
+                0.893,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Visual Track Discussion To improve the robustness, a pivotal question arises: what are the limits of current visual SLAM? To what extent does image quality impact a visual SLAM system? Table 4 shows a summary of ATE errors from real-world and simulation scenarios from the awarded 4 teams. In real-world scenarios, the first limitation is the lack of anti-noise capability in current methods, especially in low-quality image settings. We assessed visual odometry accuracy in visually degraded environments, including low lighting, sunlight overexposure, flashing lights, smoke-filled rooms, and nighttime outdoor settings (Figure 3). Trajectories in conditions like the Smoke Room and Outdoor Night exhibit significantly higher average ATE er"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.479,
+                0.946,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22653"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.078,
+                0.09,
+                0.892,
+                0.104
+            ],
+            "angle": 0,
+            "content": "Table 5. Robustness Performance on Geometric Degradation. Red numbers are robustness ranking. Larger values indicate better robustness."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.113,
+                0.112,
+                0.855,
+                0.279
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\" colspan=\"2\">Team</td><td colspan=\"6\">Geometric Degradation (Real World)</td><td colspan=\"3\">Simulation</td><td colspan=\"3\">Mix Degradation</td><td rowspan=\"2\">Average</td></tr><tr><td>Urban</td><td>Tunnel</td><td>Cave</td><td>Nuclear_1</td><td>Nuclear_2</td><td>Laurel_Caverns</td><td>Factory</td><td>Ocean</td><td>Sewerage</td><td>Long Corridor</td><td>Multi Floor</td><td>Block Lidar</td></tr><tr><td>Liu et al1</td><td rowspan=\"5\">Rp↑</td><td>0.811</td><td>0.865</td><td>0.736</td><td>0.747</td><td>0.504</td><td>0.816</td><td>0.157</td><td>0.135</td><td>0.135</td><td>0.396</td><td>0.529</td><td>0.371</td><td>0.516</td></tr><tr><td>Weitong et al2</td><td>0.773</td><td>0.838</td><td>0.690</td><td>0.871</td><td>0.739</td><td>0.152</td><td>0.153</td><td>0.149</td><td>0.081</td><td>0.407</td><td>0.456</td><td>0.345</td><td>0.471</td></tr><tr><td>Kim et al4</td><td>0.747</td><td>0.865</td><td>0.737</td><td>0.870</td><td>0.689</td><td>0.777</td><td>0.014</td><td>0.001</td><td>0.010</td><td>0.410</td><td>0.285</td><td>0.345</td><td>0.479</td></tr><tr><td>Yibin et al3</td><td>0.650</td><td>0.783</td><td>0.557</td><td>0.721</td><td>0.490</td><td>0.481</td><td>0.219</td><td>0.560</td><td>0.019</td><td>0.175</td><td>0.260</td><td>0.370</td><td>0.440</td></tr><tr><td>Zhong et al5</td><td>0.567</td><td>0.683</td><td>0.204</td><td>0.680</td><td>0.426</td><td>0.447</td><td>0.103</td><td>0.084</td><td>0.122</td><td>0</td><td>0</td><td>0</td><td>0.276</td></tr><tr><td>Liu et al1</td><td rowspan=\"5\">Rr↑</td><td>0.888</td><td>0.893</td><td>0.816</td><td>0.857</td><td>0.778</td><td>0.861</td><td>0.692</td><td>0.719</td><td>0.652</td><td>0.753</td><td>0.689</td><td>0.643</td><td>0.770</td></tr><tr><td>Weitong et al2</td><td>0.886</td><td>0.892</td><td>0.811</td><td>0.893</td><td>0.840</td><td>0.598</td><td>0.688</td><td>0.720</td><td>0.514</td><td>0.766</td><td>0.706</td><td>0.644</td><td>0.746</td></tr><tr><td>Kim et al4</td><td>0.680</td><td>0.893</td><td>0.816</td><td>0.893</td><td>0.803</td><td>0.808</td><td>0.198</td><td>0.125</td><td>0.300</td><td>0.802</td><td>0.422</td><td>0.642</td><td>0.615</td></tr><tr><td>Yibin et al3</td><td>0.624</td><td>0.707</td><td>0.462</td><td>0.733</td><td>0.540</td><td>0.467</td><td>0.862</td><td>0.962</td><td>0.496</td><td>0.477</td><td>0.342</td><td>0.213</td><td>0.573</td></tr><tr><td>Zhong et al5</td><td>0.731</td><td>0.745</td><td>0.437</td><td>0.757</td><td>0.603</td><td>0.523</td><td>0.688</td><td>0.704</td><td>0.643</td><td>0</td><td>0</td><td>0</td><td>0.485</td></tr><tr><td colspan=\"2\">AverageRp</td><td>0.710</td><td>0.807</td><td>0.585</td><td>0.778</td><td>0.570</td><td>0.535</td><td>0.130</td><td>0.186</td><td>0.073</td><td>0.278</td><td>0.306</td><td>0.286</td><td></td></tr><tr><td colspan=\"2\">Average Rr</td><td>0.762</td><td>0.826</td><td>0.668</td><td>0.827</td><td>0.713</td><td>0.651</td><td>0.626</td><td>0.646</td><td>0.521</td><td>0.560</td><td>0.432</td><td>0.428</td><td></td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.084,
+                0.292,
+                0.885,
+                0.306
+            ],
+            "angle": 0,
+            "content": "Table 6. Robustness Performance on Visual Degradation. Red numbers are robustness ranking. Larger values indicate better robustness."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.107,
+                0.313,
+                0.855,
+                0.472
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\" colspan=\"2\">Team</td><td colspan=\"6\">Visual Degradation (Real World)</td><td colspan=\"3\">Simulation</td><td rowspan=\"2\">Average</td></tr><tr><td>Lowlight 1</td><td>Lowlight 2</td><td>Over Exposure</td><td>Flash Light</td><td>Smoke Room</td><td>Outdoor Night</td><td>End of World</td><td>Moon</td><td>Western Desert</td></tr><tr><td>Peng et al1</td><td></td><td>0.357</td><td>0.227</td><td>0.264</td><td>0.203</td><td>0.536</td><td>0.270</td><td>0.699</td><td>0.893</td><td>0.806</td><td>0.472</td></tr><tr><td>Thien et al3</td><td></td><td>0.045</td><td>0.070</td><td>0.240</td><td>0.156</td><td>0.131</td><td>0.075</td><td>0</td><td>0.031</td><td>0</td><td>0.083</td></tr><tr><td>Jiang et al4</td><td>Rp↑</td><td>0.046</td><td>0.039</td><td>0.194</td><td>0.088</td><td>0.242</td><td>0.095</td><td>0</td><td>0</td><td>0</td><td>0.078</td></tr><tr><td>Li et al2</td><td></td><td>0.342</td><td>0.187</td><td>0.257</td><td>0.208</td><td>0.322</td><td>0.142</td><td>0</td><td>0.006</td><td>0</td><td>0.162</td></tr><tr><td>Peng et al1</td><td></td><td>0.641</td><td>0.581</td><td>0.744</td><td>0.650</td><td>0.878</td><td>0.670</td><td>0.975</td><td>0.975</td><td>0.974</td><td>0.787</td></tr><tr><td>Thien et al3</td><td></td><td>0.413</td><td>0.445</td><td>0.610</td><td>0.269</td><td>0.474</td><td>0.487</td><td>0.177</td><td>0.315</td><td>0</td><td>0.354</td></tr><tr><td>Jiang et al4</td><td>Rr↑</td><td>0.453</td><td>0.452</td><td>0.619</td><td>0.252</td><td>0.542</td><td>0.577</td><td>0.002</td><td>0</td><td>0.157</td><td>0.339</td></tr><tr><td>Li et al2</td><td></td><td>0.642</td><td>0.574</td><td>0.657</td><td>0.651</td><td>0.773</td><td>0.660</td><td>0.0</td><td>0.305</td><td>0</td><td>0.473</td></tr><tr><td colspan=\"2\">AverageRp</td><td>0.198</td><td>0.131</td><td>0.239</td><td>0.164</td><td>0.308</td><td>0.146</td><td>0.175</td><td>0.232</td><td>0.202</td><td></td></tr><tr><td colspan=\"2\">Average Rr</td><td>0.537</td><td>0.513</td><td>0.658</td><td>0.456</td><td>0.667</td><td>0.598</td><td>0.288</td><td>0.399</td><td>0.283</td><td></td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.486,
+                0.47,
+                0.608
+            ],
+            "angle": 0,
+            "content": "rors (4.95m, 7.76m respectively). Smoke and night scenes challenge feature extraction and tracking and introduce sensor noise, emphasizing the need for robust anti-noise algorithms. The second limitation is that existing methods struggle to overcome aggressive motion. Even in simulations with relatively good image quality, most methods still show significant average ATE errors (8.024m) on the Moon sequence (Figure 3 Y). See more details in supplementary."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.077,
+                0.612,
+                0.29,
+                0.626
+            ],
+            "angle": 0,
+            "content": "4.2. Robustness Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.629,
+                0.47,
+                0.795
+            ],
+            "angle": 0,
+            "content": "The F-1 robustness curve provides a detailed evaluation under different error tolerances, while \\( R_{p} \\) and \\( R_{r} \\) provide an overall measure of robustness that is not dependent on a specific decision threshold. This suggests the flexibility of our new robustness metric. Figure 5 clearly shows the robustness performance of all teams regarding position and rotation, highlighting the differences in robustness between the methods. The values of \\( R_{p} \\) and \\( R_{r} \\), which are the AUC under F-1 curves, are displayed in the brackets of the legends. It reveals that Liu et al. and Peng et al. are the most robust solutions for LiDAR and visual tracks, respectively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.076,
+                0.796,
+                0.471,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Is the robustness metric robust? We did extensive experiments on our robustness metric against diverse sensor degradation settings. Table 5 and 6 present a summary of robustness performance in real-world, simulated, and mixed degradation scenarios for both LiDAR and visual sequences. In Table 5, we observe that the average values of \\( R_{p} \\) and \\( R_{r} \\) for the mixed degradation environment (0.290,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.486,
+                0.892,
+                0.577
+            ],
+            "angle": 0,
+            "content": "0.473) are significantly lower than geometric degradation environments (0.664, 0.754). This observation suggests that the majority of SLAM algorithms exhibit reduced robustness in mixed degradation environments as compared to geometrically degraded ones. This aligns with our expectations and verifies the effectiveness of the robustness metric."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.581,
+                0.892,
+                0.718
+            ],
+            "angle": 0,
+            "content": "Our robustness ranking, indicated by small red numbers, differs from those on ATE ranking shown in Table 3 and 4. This is because our F-1 score-based metrics \\( R_{p} \\) and \\( R_{r} \\) jointly consider precision and recall rate to provide a balanced evaluation of SLAM performance, considering both metrics across the full spectrum of thresholds. In contrast, the ATE focuses solely on precision, neglecting the trajectory's completeness (recall). We also evaluate our metric on various synthetic trajectories in the supplementary."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.5,
+                0.74,
+                0.619,
+                0.755
+            ],
+            "angle": 0,
+            "content": "5. Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.498,
+                0.765,
+                0.892,
+                0.901
+            ],
+            "angle": 0,
+            "content": "We introduce SubT-MRS, a comprehensive SLAM dataset with various sensor data, locomotion patterns, and over 30 degradation types in simulation and real-world settings to push SLAM towards all-weather environments. Additionally, we introduce a new robustness metric to evaluate the reliability of SLAM systems, enhancing robot safety control. 29 teams have tested SubT-MRS in the organized SLAM challenge and we expect that it will serve as a critical benchmark for future SLAM development."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22654"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.081,
+                0.091,
+                0.174,
+                0.105
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.11,
+                0.422,
+                0.138
+            ],
+            "angle": 0,
+            "content": "[1] https://www.darpa.mil/program/darpa-subterranean-challenge. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.14,
+                0.47,
+                0.207
+            ],
+            "angle": 0,
+            "content": "[2] Chunge Bai, Tao Xiao, Yajie Chen, Haoqian Wang, Fang Zhang, and Xiang Gao. Faster-lio: Lightweight tightly coupled lidar-inertial odometry using parallel sparse incremental voxels. IEEE Robotics and Automation Letters, 7(2):4861-4868, 2022. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.21,
+                0.471,
+                0.278
+            ],
+            "angle": 0,
+            "content": "[3] Michael Burri, Janosch Nikolic, Pascal Gohl, Thomas Schneider, Joern Rehder, Sammy Omari, Markus W Achtelik, and Roland Siegwart. The EuRoC micro aerial vehicle datasets. The International Journal of Robotics Research, 2016. 1, 2, 3, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.281,
+                0.471,
+                0.348
+            ],
+            "angle": 0,
+            "content": "[4] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.351,
+                0.469,
+                0.419
+            ],
+            "angle": 0,
+            "content": "[5] Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel, and Juan D. Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6): 1874-1890, 2021. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.421,
+                0.469,
+                0.476
+            ],
+            "angle": 0,
+            "content": "[6] Nicholas Carlevaris-Bianco, Arash K Ushani, and Ryan M Eustice. University of michigan north campus long-term vision and lidar dataset. The International Journal of Robotics Research, 35(9):1023-1035, 2016. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.478,
+                0.469,
+                0.561
+            ],
+            "angle": 0,
+            "content": "[7] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J Taylor, and Kostas Daniilidis. M3ed: Multi-robot, multi-sensor, multi-environment event dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015–4022, 2023. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.563,
+                0.469,
+                0.617
+            ],
+            "angle": 0,
+            "content": "[8] Kenny Chen, Brett T. Lopez, Ali-akbar Agha-mohammadi, and Ankur Mehta. Direct lidar odometry: Fast localization with dense point clouds. IEEE Robotics and Automation Letters, 7(2):2000–2007, 2022. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.087,
+                0.619,
+                0.469,
+                0.687
+            ],
+            "angle": 0,
+            "content": "[9] Long Chen, Yuchen Li, Chao Huang, Bai Li, Yang Xing, Daxin Tian, Li Li, Zhongxu Hu, Xiaoxiang Na, Zixuan Li, et al. Milestones in autonomous driving and intelligent vehicles: Survey of surveys. IEEE Transactions on Intelligent Vehicles, 8(2):1046-1056, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.689,
+                0.469,
+                0.757
+            ],
+            "angle": 0,
+            "content": "[10] Weichen Dai, Yu Zhang, Shenzhou Chen, Donglei Sun, and Da Kong. A multi-spectral dataset for evaluating motion estimation systems. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5560-5566. IEEE, 2021. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.76,
+                0.469,
+                0.8
+            ],
+            "angle": 0,
+            "content": "[11] Frank Dellaert, Michael Kaess, et al. Factor graphs for robot perception. Foundations and Trends® in Robotics, 6(1-2): 1-139, 2017. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.803,
+                0.469,
+                0.857
+            ],
+            "angle": 0,
+            "content": "[12] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the UZH-FPV drone racing dataset. In IEEE Int. Conf. Robot. Autom. (ICRA), 2019. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.86,
+                0.469,
+                0.901
+            ],
+            "angle": 0,
+            "content": "[13] Rodolphe Dubois, Alexandre Eudes, and Vincent Frémont. Airmuseum: a heterogeneous multi-robot dataset for stereovisual and inertial simultaneous localization and mapping. In"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.11,
+                0.471,
+                0.901
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.535,
+                0.093,
+                0.892,
+                0.133
+            ],
+            "angle": 0,
+            "content": "2020 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI), pages 166-172, 2020. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.136,
+                0.892,
+                0.189
+            ],
+            "angle": 0,
+            "content": "[14] Naser El-Sheimi, Haiying Hou, and Xiaoji Niu. Analysis and modeling of inertial sensors using allan variance. IEEE Transactions on Instrumentation and Measurement, 57(1): 140-149, 2008. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.193,
+                0.892,
+                0.247
+            ],
+            "angle": 0,
+            "content": "[15] Dapeng Feng, Yuhua Qi, Shipeng Zhong, Zhiqiang Chen, Yudu Jiao, Qiming Chen, Tao Jiang, and Hongbo Chen. S3e: A large-scale multimodal dataset for collaborative slam. arXiv preprint arXiv:2210.13723, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.25,
+                0.892,
+                0.304
+            ],
+            "angle": 0,
+            "content": "[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4340-4349, 2016. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.307,
+                0.892,
+                0.332
+            ],
+            "angle": 0,
+            "content": "[17] Wenliang Gao. A ros package tool to analyze the imu performance, 2018. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.335,
+                0.892,
+                0.389
+            ],
+            "angle": 0,
+            "content": "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI Vision Benchmark Suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.392,
+                0.892,
+                0.445
+            ],
+            "angle": 0,
+            "content": "[19] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.449,
+                0.892,
+                0.503
+            ],
+            "angle": 0,
+            "content": "[20] Dongjiao He, Wei Xu, Nan Chen, Fanze Kong, Chongjian Yuan, and Fu Zhang. Point-lio: Robust high-bandwidth light detection and ranging inertial odometry. Advanced Intelligent Systems, page 2200459, 2023. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.506,
+                0.892,
+                0.559
+            ],
+            "angle": 0,
+            "content": "[21] Michael Helmberger, Kristian Morin, Beda Berner, Nitish Kumar, Giovanni Cioffi, and Davide Scaramuzza. The hilti slam challenge dataset. IEEE Robotics and Automation Letters, 7(3):7518-7525, 2022. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.563,
+                0.892,
+                0.629
+            ],
+            "angle": 0,
+            "content": "[22] Giseop Kim and Ayoung Kim. Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4802-4809. IEEE, 2018. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.633,
+                0.892,
+                0.687
+            ],
+            "angle": 0,
+            "content": "[23] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. Vivid++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, 2022. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.69,
+                0.892,
+                0.743
+            ],
+            "angle": 0,
+            "content": "[24] Keith YK Leung, Yoni Halpern, Timothy D Barfoot, and Hugh HT Liu. The utias multi-robot cooperative localization and mapping dataset. The International Journal of Robotics Research, 30(8):969–974, 2011. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.747,
+                0.892,
+                0.815
+            ],
+            "angle": 0,
+            "content": "[25] Hyungtae Lim, Suyong Yeon, Suyong Ryu, Yonghan Lee, Youngji Kim, Jaeseong Yun, Euigon Jung, Donghwan Lee, and Hyun Myung. A single correspondence is enough: Robust global registration to avoid degeneracy in urban environments. page Accepted. To appear, 2022. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.817,
+                0.892,
+                0.857
+            ],
+            "angle": 0,
+            "content": "[26] Yicheng Lin, Shuo Wang, Yunlong Jiang, and Bin Han. Breaking of brightness consistency in optical flow with a lightweight cnn network, 2023. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.504,
+                0.86,
+                0.892,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[27] Zheng Liu, Xiyuan Liu, and Fu Zhang. Efficient and consistent bundle adjustment on lidar point clouds. IEEE Transactions on Robotics, 2023. 6"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.503,
+                0.093,
+                0.892,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.518,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22655"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.092,
+                0.47,
+                0.146
+            ],
+            "angle": 0,
+            "content": "[28] Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year, \\(1000\\mathrm{km}\\): The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3-15, 2017. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.148,
+                0.47,
+                0.217
+            ],
+            "angle": 0,
+            "content": "[29] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research, 36(2):142-149, 2017. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.218,
+                0.469,
+                0.272
+            ],
+            "angle": 0,
+            "content": "[30] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Transactions on Robotics, 34(6): 1425-1440, 2018. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.274,
+                0.469,
+                0.327
+            ],
+            "angle": 0,
+            "content": "[31] Thien Hoang Nguyen, Shenghai Yuan, and Lihua Xie. Vrslam: A visual-range simultaneous localization and mapping system using monocular camera and ultra-wideband sensors, 2023. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.33,
+                0.469,
+                0.37
+            ],
+            "angle": 0,
+            "content": "[32] Xiongfeng Peng, Zhihua Liu, Weiming Li, Ping Tan, SoonYong Cho, and Qiang Wang. Dvi-slam: A dual visual inertial slam network, 2023. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.372,
+                0.469,
+                0.439
+            ],
+            "angle": 0,
+            "content": "[33] Bernd Pfrommer, Nitin Sanket, Kostas Daniilidis, and Jonas Cleveland. Penncosyvio: A challenging visual inertial odometry benchmark. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3847-3854. IEEE, 2017. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.442,
+                0.469,
+                0.495
+            ],
+            "angle": 0,
+            "content": "[34] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics, 34(4):1004-1020, 2018. 1, 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.497,
+                0.469,
+                0.579
+            ],
+            "angle": 0,
+            "content": "[35] Milad Ramezani, Yiduo Wang, Marco Camurri, David Wisth, Matias Mattamala, and Maurice Fallon. The newer college dataset: Handheld lidar, inertial and vision with ground truth. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4353-4360. IEEE, 2020. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.581,
+                0.469,
+                0.649
+            ],
+            "angle": 0,
+            "content": "[36] Joern Rehder, Janosch Nikolic, Thomas Schneider, Timo Hinzmann, and Roland Siegwart. Extending kalibr: Calibrating the extrinsic of multiple imus and of individual axes. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 4304-4311. IEEE, 2016. 3, 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.651,
+                0.469,
+                0.719
+            ],
+            "angle": 0,
+            "content": "[37] John G Rogers, Arthur Schang, Carlos Nieto-Granda, John Ware, John Carter, Jonathan Fink, and Ethan Stump. The darpa subt urban circuit mapping dataset and evaluation metric. In Experimental Robotics: The 17th International Symposium, pages 391–401. Springer, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.721,
+                0.469,
+                0.787
+            ],
+            "angle": 0,
+            "content": "[38] Sebastian Scherer, Vasu Agrawal, Graeme Best, Chao Cao, Katarina Cujic, R Darnley, R DeBortoli, E Dexheimer, B Drozd, R Garg, et al. Resilient and modular subterranean exploration with a team of roving and flying robots. Field Robotics, 2022. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.79,
+                0.469,
+                0.858
+            ],
+            "angle": 0,
+            "content": "[39] David Schubert, Thore Goll, Nikolaus Demmel, Vladyslav Usenko, Jörg Stuckler, and Daniel Cremers. The tum vi benchmark for evaluating visual-inertial odometry. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1680-1687. IEEE, 2018. 1, 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.86,
+                0.469,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[40] Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-icp. In Robotics: science and systems, page 435. Seattle, WA, 2009. 4, 5"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.092,
+                0.47,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.093,
+                0.892,
+                0.161
+            ],
+            "angle": 0,
+            "content": "[41] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Rus Daniela. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5135-5142. IEEE, 2020. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.163,
+                0.892,
+                0.217
+            ],
+            "angle": 0,
+            "content": "[42] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In Proc. of the International Conference on Intelligent Robot Systems (IROS), 2012. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.219,
+                0.892,
+                0.272
+            ],
+            "angle": 0,
+            "content": "[43] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.274,
+                0.892,
+                0.342
+            ],
+            "angle": 0,
+            "content": "[44] Yulun Tian, Yun Chang, Long Quang, Arthur Schang, Carlos Nieto-Granda, Jonathan P How, and Luca Carlone. Resilient and distributed multi-robot visual slam: Datasets, experiments, and lessons learned. arXiv preprint arXiv:2304.04362, 2023. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.344,
+                0.892,
+                0.412
+            ],
+            "angle": 0,
+            "content": "[45] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters, 8(2):1029–1036, 2023. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.414,
+                0.892,
+                0.565
+            ],
+            "angle": 0,
+            "content": "[46] Chen Wang, Dasong Gao, Kuan Xu, Junyi Geng, Yaoyu Hu, Yuheng Qiu, Bowen Li, Fan Yang, Brady Moon, Abhinav Pandey, Aryan, Jiahe Xu, Tianhao Wu, Haonan He, Daning Huang, Zhongqiang Ren, Shibo Zhao, Taimeng Fu, Pranay Reddy, Xiao Lin, Wenshan Wang, Jingnan Shi, Rajat Talak, Kun Cao, Yi Du, Han Wang, Huai Yu, Shanzhao Wang, Siyu Chen, Ananth Kashyap, Rohan Bandaru, Karthik Dantu, Jiajun Wu, Lihua Xie, Luca Carlone, Marco Hutter, and Sebastian Scherer. PyPose: A library for robot learning with physics-based optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.567,
+                0.892,
+                0.635
+            ],
+            "angle": 0,
+            "content": "[47] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual SLAM. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 3, 7"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.637,
+                0.892,
+                0.677
+            ],
+            "angle": 0,
+            "content": "[48] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. Tartanvo: A generalizable learning-based vo. In Conference on Robot Learning, pages 1761-1772. PMLR, 2021. 1"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.679,
+                0.892,
+                0.746
+            ],
+            "angle": 0,
+            "content": "[49] P. Wenzel, R. Wang, N. Yang, Q. Cheng, Q. Khan, L. von Stumberg, N. Zeller, and D. Cremers. 4Seasons: A cross-season dataset for multi-weather SLAM in autonomous driving. In Proceedings of the German Conference on Pattern Recognition (GCPR), 2020. 2"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.748,
+                0.892,
+                0.787
+            ],
+            "angle": 0,
+            "content": "[50] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. FAST-LIO2: fast direct lidar-inertial odometry. IEEE Transactions on Robotics, abs/2107.06829, 2022. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.79,
+                0.892,
+                0.83
+            ],
+            "angle": 0,
+            "content": "[51] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. Fast-lio2: Fast direct lidar-inertial odometry. IEEE Transactions on Robotics, 38(4):2053–2073, 2022. 6"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.832,
+                0.892,
+                0.872
+            ],
+            "angle": 0,
+            "content": "[52] Lintong Zhang, Marco Camurri, and Maurice Fallon. Multicamera lidar inertial extension to the newer college dataset. arXiv preprint arXiv:2112.08854, 2021. 3"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.503,
+                0.874,
+                0.892,
+                0.9
+            ],
+            "angle": 0,
+            "content": "[53] Shibo Zhao, Hengrui Zhang, Peng Wang, Lucas Nogueira, and Sebastian Scherer. Super odometry: Imu-centric"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.503,
+                0.093,
+                0.892,
+                0.9
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.52,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22656"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.11,
+                0.092,
+                0.47,
+                0.146
+            ],
+            "angle": 0,
+            "content": "lidar-visual-inertial estimator for challenging environments. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8729-8736. IEEE, 2021. 4"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.08,
+                0.149,
+                0.47,
+                0.219
+            ],
+            "angle": 0,
+            "content": "[54] David Zuniga-Noël, Alberto Jaenal, Ruben Gomez-Ojeda, and Javier Gonzalez-Jimenez. The uma-vi dataset: Visual-inertial odometry in low-textured and dynamic illumination environments. The International Journal of Robotics Research, 39(9):1052-1060, 2020. 1, 3"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.08,
+                0.092,
+                0.47,
+                0.219
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.946,
+                0.519,
+                0.957
+            ],
+            "angle": 0,
+            "content": "22657"
+        }
+    ]
+]

2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2e2ca39b1212cdf2f8344c703f529f689000dd620f0e1a806abab261cd57466
+size 2825928

2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/full.md

@@ -0,0 +1,257 @@
+# SubT-MRS Dataset: Pushing SLAM Towards All-weather Environments https://superodometry.com/datasets
+
+Shibo Zhao\*, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, Yi Du, Shaoshu Su, Can Xu, John Keller, Jay Karhade, Lucas Nogueira, Sourojit Saha, Ji Zhang, Wenshan Wang, Chen Wang, Sebastian Scherer
+
+$^{1}$ Carnegie Mellon University  $^{2}$ University at Buffalo
+
+# Abstract
+
+Simultaneous localization and mapping (SLAM) is a fundamental task for numerous applications such as autonomous navigation and exploration. Despite many SLAM datasets have been released, current SLAM solutions still struggle to have sustained and resilient performance. One major issue is the absence of high-quality datasets including diverse all-weather conditions and a reliable metric for assessing robustness. This limitation significantly restricts the scalability and generalizability of SLAM technologies, impacting their development, validation, and deployment. To address this problem, we present SubT-MRS, an extremely challenging real-world dataset designed to push SLAM towards all-weather environments to pursue the most robust SLAM performance. It contains multi-degraded environments including over 30 diverse scenes such as structureless corridors, varying lighting conditions, and perceptual obscurants like smoke and dust; multimodal sensors such as LiDAR, fisheye camera, IMU, and thermal camera; and multiple locomotions like aerial, legged, and wheeled robots. We developed accuracy and robustness evaluation tracks for SLAM and introduced novel robustness metrics. Comprehensive studies are performed, revealing new observations, challenges, and opportunities for future research.
+
+# 1. Introduction
+
+Simultaneous Localization and Mapping (SLAM) is essential in robotics since it provides foundational perception and spatial awareness, and enables machines to understand and interact with the physical world in real-time. Therefore, it has a wide range of applications such as autonomous driving and space exploration. Despite significant advancements in both geometric [5, 34, 41] and data-driven SLAM methods [43, 48], existing solutions remain fragile in challenging conditions. One main reason is that existing SLAM
+
+algorithms are often developed and evaluated with datasets from controlled environments [3, 12, 18, 21, 33, 35, 39, 54]. For example, KITTI [18, 21] dataset is mostly collected in sunny weather and EuROC-MAV dataset [3] is collected in small well-lit rooms. These datasets, unfortunately, fail to capture the challenges in real-world scenarios, hindering the development of robust SLAM solutions.
+
+To bridge this gap and push SLAM towards all-weather environments, we present an extremely challenging dataset, SubT-MRS, including scenarios featuring various sensor degradation, aggressive locomotions, and extreme-weather conditions. The SubT-MRS dataset comprises 3 years of data from the DARPA Subterranean (SubT) Challenge [1] (2019-2021) and extends with an additional 2 years of diverse environments (2022-2023), containing mixed indoor and outdoor settings, including long corridors, off-road scenario, tunnels, caves, deserts, forests, and bushlands. Cumulatively, this forms a 5-year dataset encompassing over 500 hours and 100 kilometers of terrain subjected to multimodal sensors including LiDAR, fisheye cameras, depth cameras, thermal cameras, and IMU; heterogeneous platforms including RC cars, legged robots, aerial robots, and wheeled robots; and extreme obscurant conditions such as dense fog, dust, smoke, and heavy snow.
+
+Additionally, we find there is no well-established metric to evaluate the robustness of a SLAM system. Existing evaluation metrics such as absolute trajectory error (ATE) [3] are not representative of the actual performance in robotic applications. We argue that to ensure the safety control of a robot, SLAM system evaluation should not only focus on poses but also its velocity. Taking robot localization as an example, while a momentary pose error spike may only slightly undermine the overall ATE, however, it can lead to a catastrophic crash in an aerial robotic platform. To effectively gauge the performance of a SLAM algorithm, we introduce a new robustness metric to evaluate the reliability of the SLAM system, particularly examining smoothness and accuracy of velocity estimation.
+
+![](images/b3d861d86989a475f5b89f18a2bceb82e3f55d46da24ddf4bfe5e6348b1bff74.jpg)
+
+![](images/3e2278a11c657ec21d4a1b1d9bd38063d647760e8f90618979c79422d7d0a8b0.jpg)  
+SubT-Tunnel  
+SubT-Urban  
+SubT-Cavern  
+Multi Floors
+
+![](images/4ff2e7e8ea5727e51f0d28853e2951db92306ff0c8c78c86400d2dbee3f76ece.jpg)  
+Dark Long Corridor  
+Dense Fog  
+Laurel Cavern  
+Textureless Corridor  
+Darkness Stairs with Jerk Motion  
+Dense Dust and Smoke  
+Offroad with Aggressive Motion  
+Snow with Slipping Motion  
+Figure 1. Dense reconstruction from the SubT-MRS dataset, achieved through collaboration with diverse robots equipped with multimodal sensors. Colors represent different challenging environments (tunnels, caves, urban, confined spaces) captured by various robot types (aerial, legged, wheeled). The bottom section showcases a gallery with diverse visual, LiDAR, and mixed degradations.
+
+Lastly, we perform extensive experiments using proposed degradation datasets to benchmark visual and LiDAR SLAM algorithms. These experiments identify the limitations of the existing SLAM systems and evaluate their robustness across the degradation. Our contributions include:
+
+- All-weather Environments To push SLAM toward all-weather environments, we introduce SubT-MRS, an extremely challenging dataset. Spanning five years, it comprises over 500 hours and 100 kilometers of accurately measured trajectories. To the best of our knowledge, SubT-MRS is the first real-world dataset that specifically addresses failure scenarios of SLAM by incorporating a variety of degraded conditions, multiple robotic platforms, and diverse sets of multimodal sensors.  
+- Robustness Metric To evaluate the actual performance running on robots, we propose a robustness metric, which, to the best of our knowledge, is the first metric evaluating the reliability, safety, and resilience of SLAM.  
+- SLAM Challenge We provide a comprehensive benchmark<sup>1</sup> based on the SubT-MRS dataset by conducting ICCV'23 SLAM Challenge. Our evaluation reveals the limitations of state-of-the-art visual and LiDAR SLAM solutions. Furthermore, we conduct extensive experiments to verify robustness metric against various sensor degradation and length of trajectories.
+
+# 2. Related Work
+
+Multimodal Sensors Multimodal sensor datasets are crucial for the development of robust SLAM systems, as single sensor modalities are not comprehensive for all scenarios. Existing datasets typically focus on a limited range of modalities, such as monocular or stereo cameras [3, 49], LiDAR [19], event cameras [29], and depth cameras [42]. The KITTI dataset [19], although covering most modalities, is geared towards on-road scenarios and lacks thermal cameras, limiting its applicability to simple, controlled environments. Conversely, the ViViD++ dataset [23] incorporates thermal images in outdoor settings but is deficient in hardware synchronization, posing challenges for SLAM system development. The Weichen dataset [10] offers thermal images in indoor environments with accurate ground truth poses but is confined to motion capture room settings. In contrast, our SubT-MRS dataset delivers a comprehensive suite of time-synchronized multimodal sensors, including LiDAR, monocular cameras, thermal cameras, depth cameras, and IMU, along with centimeter-level ground truth, catering to a diverse range of research needs.
+
+Heterogeneous Platforms Developing a robust SLAM system necessitates versatility in handling different motion patterns and various robotic platforms. However, most existing datasets are tailored for single-robot scenarios. While there are a few multi-robot datasets, they typically involve homogeneous platforms [9, 15, 24, 44], posing challenges
+
+<table><tr><td rowspan="2">Dataset</td><td colspan="4">Multi-Spectrum</td><td colspan="6">Multi-Degradation</td><td colspan="4">Multi-Robot</td></tr><tr><td>Camera</td><td>IMU</td><td>LiDAR/Depth</td><td>Thermal</td><td>Illumination</td><td>Snow Smoke</td><td>Structureless</td><td>SubT</td><td>Aggressive Motion</td><td>Vehicle</td><td>Drone</td><td>Legged</td><td>Handheld</td><td></td></tr><tr><td>EuRoC MAV [3]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td></td></tr><tr><td>PennCOsYVIO [33]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>TUM VIO [39]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>UZH-FPV [12]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td></td></tr><tr><td>College Dataset [52]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>RobotCar [28]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>UMA VI [54]</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>UMich [6]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>KITTI [19]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>Virtual KITTI [16]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>DARPA SubT [37]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td></td></tr><tr><td>Hilti SLAM [21]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td></td></tr><tr><td>M3ED [7]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>X</td><td>X</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td></td></tr><tr><td>TartanAir [47]</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>X</td><td></td></tr><tr><td>SubT-MRS (Ours)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td></td></tr></table>
+
+Table 1. Comparison of SLAM datasets on multi-sensors, multi robots, and multi degradation.
+
+in evaluating SLAM performance across varied robotic platforms. The AirMuseum dataset [13] includes drones and ground robots but lacks data from legged robots. As a comparison, SubT-MRS dataset features diverse robotic platforms including legged robots, aerial robots, and wheeled robots, operating in varied environments and sensor setups.
+
+Extreme Environmental Conditions Developing and testing SLAM systems in extreme environmental conditions is crucial for mitigating potential real-world failures. However, most existing SLAM datasets have been primarily limited to single, controlled environments. The TUM-VI [39] and UMA-VI [54] datasets, being indoor-outdoor visual-inertial datasets, pose challenges due to varying illumination and low-texture environments. The EuRoC MAV [3] and UZH-FPV drone racing [12] datasets, popular in the SLAM community, offer data on aggressive drone movements but usually in consistently lit conditions. The KITTI dataset [19], a staple in autonomous driving research for its outdoor LiDAR-Visual-Inertial data. The Hilti SLAM dataset [21] includes both indoor and outdoor LiDAR-visual-inertial datasets with dynamic lighting and confined spaces. However, both of them focus more on accuracy than on a variety of environmental degradations. TartanAir [47] covers most of the challenging environments but it is a simulation dataset that will pose the sim-to-real gap.
+
+In contrast, the SubT-MRS dataset provides 30 varied scenes, encompassing a wide array of environments. These include SubT tunnels, urban areas, caverns, multifloor structures, dark corridors, and foggy conditions. It also features textureless surfaces, dusty and smoky environments, off-road areas with aggressive motion, and snowy terrains prone to slippage. We include challenging lighting conditions such as extreme darkness and overexposure, along with geometrically challenging scenarios like featureless corridors, staircases, and self-similar cave layouts. Additionally, the dataset incorporates perceptual challenges posed by smoke, fog, dust, and various weather conditions.
+
+# 3. SubT-MRS Dataset
+
+We next present the SubT-MRS dataset from three aspects including its collecting environments and settings, ground truth collection, and the new robustness metric. A detailed comparison with other datasets is listed in Table 1.
+
+# 3.1. All-weather Environments
+
+As mentioned in Sec. 2, the SubT-MRS dataset distinguishes itself through its multimodal sensor setups, heterogeneous robotic platforms, and extreme environmental conditions. This section emphasizes this comprehensive nature pushing SLAM towards all-weather environments.
+
+# 3.1.1. Multimodal Sensors
+
+Multimodal sensors provide a wealth of information for SLAM systems operating in all-weather environments. This diversity improves scene understanding and strengthens the system's perception ability. Therefore, We incorporate diverse sensors to ensure system robustness.
+
+Sensor Pack We embedded 4 Leopard Imaging RGB fisheye cameras, 1 Velodyne puck, 1 Epson M-G365 IMU, 1 FLIR Boson thermal camera, and NVIDIA Jetson AGX Xavier as a sensor pack shown in Figure 2.
+
+Time Synchronization To ensure the overall consistency of the fused data, we meticulously synchronize the sensors' time using the 'pulse per second (PPS)' technique. As illustrated in Figure 2, the IMU, LiDAR, and thermal camera are directly synchronized with the CPU clock, while the four RGB camera are synchronized using an FPGA board. Consequently, we can effectively manage the time synchronization gap between each pair of sensors not to exceed 3ms.
+
+Fisheye and Thermal Camera Calibration Camera calibration plays an essential role in the efficiency of a SLAM system. For calibrating the fisheye cameras, we employed the open-source toolkit Kalibr [36], focusing on the intrinsic and extrinsic parameters. Specifically, we use the radial-tangential distortion model to rectify the omnidirectional fisheye camera model. Calibrating thermal cameras, however, poses unique challenges, especially in gathering high-
+
+![](images/4a5dcfa2e9f9e3cf0bef64f230e655cdef81aa0eee5146db6cfaca25eed1ea7d.jpg)  
+Figure 2. An overview of the sensor pack used in SubT-MRS dataset. It is equipped with a Xavier processing unit with hardware time synchronization for multimodal sensors including LiDAR, fisheye cameras, thermal cameras, depth cameras (option), and an IMU.
+
+quality thermal data. To tackle this, we set up a  $7 \times 9$  chessboard, heated by direct sunlight, to generate high-contrast thermal data, ensuring accuracy in the calibration process.
+
+IMU Calibration and Extrinsic Calibration For Lidar-IMU extrinsic calibration, we utilize the CAD model to obtain the calibration parameters. In the case of Camera-IMU extrinsic calibration, we employ the Kalibr toolbox [36] to estimate the extrinsic matrix. To estimate the sensors' bias and the random walk noise of the gyroscope and accelerometer, we collected static data from the IMU and calibrated it using an Allan variance-based IMU tool [14, 17].
+
+# 3.1.2. Multi-degraded Environments
+
+SubT-MRS includes multiple challenging environments including visually degraded environments, geometrically degraded environments, and their combination.
+
+Visual Degradation Poor-quality visual features can significantly hinder the performance of feature extraction processes and disrupt the accuracy of feature matching. Such issues arise in low-light conditions with inconsistent brightness or image noise introduced by air obscurants. SubT-MRS encompasses these challenges and provides a wide range of visual degradation. This includes environments with limited lighting, such as hospital interiors and caves (Figure 3 A-F), as well as smoky or dusty conditions that cause visual obstruction (Figure 3 M-N), and snowy areas with reduced visibility (Figure 3 H, K, and L).
+
+Geometric Degradation Lack of geometrical features poses significant challenges to LiDAR odometry. The root of this issue often lies in the limited sensing capabilities of LiDAR sensors and the constraints due to their mechanical installation. SubT-MRS captures a variety of environments that exemplify such challenges. It includes long, featureless corridors (Figure 3 E) and staircases (Figure 3 C and G). These scenarios illustrate various forms of geometric degradation, contributing to the improvement of LiDAR odometry systems in challenging conditions.
+
+Mixed Degradation A mixed visual and geometric degradation can further hinder the performance of SLAM sys
+
+tems. Examples in SubT-MRS include long, dimly lit corridors (Figure 3 A, E, and F), poorly illuminated staircases (Figure 3 C), and environments affected by snowy weather (Figure 3 H, K, and L). In these settings, both LiDAR odometry and visual odometry are prone to failure due to the compounded effects of mixed degradation. The inclusion of such scenarios in SubT-MRS is crucial for evaluating and improving the resilience of multimodal SLAM algorithms.
+
+# 3.1.3. Heterogeneous Robot Platforms
+
+Most of current datasets focus on single-robot systems, limiting multi-robot SLAM development as shown in Table 1. To address this, we employed diverse robot platforms, including aerial, legged, and wheeled robots, navigating through various environments from urban campuses to medical facilities and natural terrains like caves. This diversity offers a range of scenarios to enhance SLAM algorithms for effective operation in challenging conditions.
+
+Extrinsic Calibration for Multiple Robots To ensure the multi-robot system shares a common coordinate system, we perform the extrinsic calibration process [38]. It has 2 steps: First, each robot runs Super Odometry [53] to generate its local map using LiDAR data and share it with other robots through a wireless network; Second, the remaining robots identify overlapping regions between their local maps and estimate the extrinsic parameters using GICP [40].
+
+# 3.2. Ground Truth
+
+Ground Truth Map As depicted in Figure 4, we utilized the high-precision FARO Focus 3D S120 3D scanner for creating ground-truth maps. This cutting-edge 3D scanner can measure distances up to  $120\mathrm{m}$ , with a maximum measurement rate of  $976\mathrm{K}$  points per second. The accuracy of the ground truth map is noteworthy, maintaining a range error within  $\pm 2\mathrm{mm}$ . We have developed ground truth models for diverse environments, including subterranean areas like urban areas  $(350\mathrm{m} \times 350\mathrm{m})$ , caves  $(150\mathrm{m} \times 200\mathrm{m})$  and tunnels  $(100 \times 200\mathrm{m})$ , as well as indoor-outdoor spaces  $(200\mathrm{m} \times 200\mathrm{m})$  within the CMU campus. To reduce drift, a
+
+![](images/e99cad20d18ec8f0048688cceaa90f07e28c3d2bab5ffea665868a82026408ba.jpg)  
+Figure 3. The SubT-MRS datasets were collected across diverse seasons, capturing environments with perceptual challenges such as poor illumination, darkness, and water puddles, where visual sensors falter. They also include geometrically complex areas like long featureless corridors and steep multi-floor structures, challenging LiDAR systems with potential drift. Moreover, these datasets cover conditions with airborne obscurants like dust, fog, snow, and smoke in tough environments, including caves, deserts, long tunnels, and off-road areas.
+
+![](images/3c0d36f7d01aeff96cc49b14894396b1d84ab5655b90321852f2810c4bddc139.jpg)  
+Figure 4. The SubT-MRS dataset facilitates the generation of high-precision ground truth maps and trajectories. Figure (a) shows the ground truth trajectory in multi-floor settings. Figure (b) displays ground truth maps for indoor and outdoor areas, encompassing long corridors, multi-floor structures, and open spaces. Figure (c) features photo-realistic scans based on our ground truth maps in cave environments.
+
+loop closure algorithm was applied to correct the poses. Notably,  $96\%$  of the scans achieved a position uncertainty of less than  $2\mathrm{mm}$ . This ground-truth map is crucial for SLAM development where precision is of utmost importance.
+
+Ground Truth Trajectory Ground truth trajectories for all sequences are generated based on the ground truth map. In this process, we establish point-to-point, point-to-plane, and point-to-line correspondences[40] between the Ground
+
+Truth map and the current LiDAR scan. Our estimated ground truth trajectory is a comprehensive fusion of various constraints derived from the ground truth map, visual odometry, LiDAR odometry, and IMU measurements. This approach ensures the robustness of our solution across diverse environmental conditions. To verify accuracy, we reconstruct LiDAR maps from ground truth trajectory and compare it with our ground truth maps.
+
+Table 2. SLAM Challenge Results (Blue shadings indicate rankings of ATE and Robustness Metric; L: LiDAR I: IMU C: Camera)  
+
+<table><tr><td>#</td><td>Team</td><td>Method</td><td>Odometry Type</td><td>Device</td><td>RealTime (s)</td><td>CPU/GPU (%)</td><td>RAM (GB)</td><td>ATE↓</td><td>\( {R}_{v} \uparrow /{R}_{w} \uparrow \)</td><td>Sensors L I C</td></tr><tr><td>1</td><td>Liu et al</td><td>FAST-LIO2 [51], HBA [27]</td><td>Filter</td><td>Intel i7-9700K</td><td>51.310</td><td>98.667 / 0</td><td>4.052</td><td>0.588</td><td>0.517/0.770</td><td>✓✓</td></tr><tr><td>2</td><td>Yibin et al</td><td>LIO-EKF [45]</td><td>Filter</td><td>Intel i7-10700</td><td>0.006</td><td>52.167 / 0</td><td>0.072</td><td>4.313</td><td>0.441/0.574</td><td>✓✓</td></tr><tr><td>3</td><td>Weitong et al</td><td>FAST-LIO[2], Pose Graph[11]</td><td>Filter</td><td>Intel Xeon(R)E3-1240v5</td><td>0.125</td><td>22.63 / 0</td><td>4.305</td><td>0.663</td><td>0.473/0.747</td><td>✓✓</td></tr><tr><td>4</td><td>Kim et al</td><td>FAST-LIO2[50], Point-LIO[20], Quatro[25]</td><td>Filter</td><td>Intel i5-12500</td><td>0.268</td><td>101.108 / 0</td><td>55.64</td><td>3.825</td><td>0.479/0.615</td><td>✓✓</td></tr><tr><td>5</td><td>Zhong et al</td><td>DLO[8], Scan-Context++[22]</td><td>SW Opt</td><td>AMD Ryzen 9 5900x</td><td>0.027</td><td>13.289 / 0</td><td>1.174</td><td>1.209</td><td>0.276/0.486</td><td>✓✓</td></tr><tr><td>1</td><td>Peng et al</td><td>DVI-SLAM [32]</td><td>Learning</td><td>Intel i9-12900</td><td>183.233</td><td>- / 149</td><td>11 (4)</td><td>0.547</td><td>0.473/0.788</td><td>✓✓</td></tr><tr><td>2</td><td>Jiang et al</td><td>LET-NET[26], VINS-Mono[34]</td><td>Hybrid</td><td>Intel i5-9400</td><td>0.064</td><td>40.35 / 0</td><td>4.337</td><td>1.093</td><td>0.078/0.322</td><td>✓✓</td></tr><tr><td>3</td><td>Thien et al</td><td>VR-SLAM[31]</td><td>SW Opt</td><td>Intel i9-12900</td><td>0.142</td><td>176.44 / 0</td><td>9.111</td><td>3.037</td><td>0.083/0.372</td><td>✓✓</td></tr><tr><td>4</td><td>Li et al</td><td>ORB-SLAM3[4]</td><td>SW Opt.</td><td>Intel i7-10700</td><td>0.019</td><td>65.028 / 0</td><td>0.386</td><td>8.975</td><td>0.163/0.474</td><td>✓✓</td></tr></table>
+
+Table 3. Accuracy Performance on Geometric Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions.  
+
+<table><tr><td rowspan="2">Team</td><td colspan="6">Geometric Degradation (Real World)</td><td colspan="3">Simulation</td><td colspan="3">Mix Degradation</td><td rowspan="2">Average</td></tr><tr><td>Urban</td><td>Tunnel</td><td>Cave</td><td>Nuclear_1</td><td>Nuclear_2</td><td>Laurel_Caverns</td><td>Factory</td><td>Ocean</td><td>Sewerage</td><td>Long Corridor</td><td>Multi Floor</td><td>Block Lidar</td></tr><tr><td>Liu et al1</td><td>0.307</td><td>0.095</td><td>0.629</td><td>0.122</td><td>0.235</td><td>0.260</td><td>0.889</td><td>0.757</td><td>0.978</td><td>1.454</td><td>0.401</td><td>0.934</td><td>0.588</td></tr><tr><td>Weitong et al2</td><td>0.26</td><td>0.096</td><td>0.617</td><td>0.120</td><td>0.222</td><td>0.402</td><td>0.998</td><td>0.770</td><td>1.586</td><td>1.254</td><td>0.577</td><td>1.056</td><td>0.663</td></tr><tr><td>Kim et al3</td><td>0.331</td><td>0.092</td><td>0.787</td><td>0.123</td><td>0.270</td><td>0.279</td><td>10.628</td><td>22.425</td><td>7.147</td><td>2.100</td><td>0.650</td><td>1.068</td><td>3.825</td></tr><tr><td>Yibin et al4</td><td>1.060</td><td>0.220</td><td>0.750</td><td>0.470</td><td>0.620</td><td>9.140</td><td>4.920</td><td>0.280</td><td>24.460</td><td>2.990</td><td>5.500</td><td>1.340</td><td>4.312</td></tr><tr><td>Zhong et al5</td><td>1.205</td><td>0.695</td><td>-</td><td>1.175</td><td>1.72</td><td>2.08</td><td>0.889</td><td>0.778</td><td>1.13</td><td>-</td><td>-</td><td>-</td><td>1.209*</td></tr><tr><td>Average</td><td>0.633</td><td>0.240</td><td>0.696</td><td>0.402</td><td>0.6134</td><td>2.432</td><td>3.665</td><td>5.002</td><td>7.060</td><td>1.950</td><td>1.782</td><td>1.099</td><td></td></tr></table>
+
+# 3.3. Robustness Metric
+
+As discussed in Sec. 1, existing evaluation metrics such as absolute trajectory error (ATE) [3] have limitations in evaluating the SLAM's robustness in real-world applications. ATE primarily focuses on the accuracy of trajectory and does not consider the completeness of trajectory (recall). Also, it can not effectively capture velocity changes that have a direct impact on the robot's safety. For example, a pose spike error might not be immediately noticeable in ATE evaluation, it may result in a crash in an aerial robot control. To address these gaps, we introduce a new robustness metric based on estimated velocity. That is because velocity estimation from SLAM is crucial for the robot's control system which directly impacts stability, affecting the robot's safety. The new robustness metric is the area under the curve (AUC) of the F-1 score:
+
+$$
+F _ {1} (e) = \frac {2 P (e <   T) R (e <   T)}{P (e <   T) + R (e <   T)}, \tag {1}
+$$
+
+where the precision  $P$  quantifies the precision of the estimated velocity as a percentage: how closely the estimated velocity points lie to the ground truth point and the recall rate  $R$  quantifies the estimated velocity's completeness: to what extent all the ground-truth points are covered. A high F-1 score can only be achieved by the velocity estimation that is both accurate and complete throughout the entire run. Specifically, an estimated error  $e$  is regarded as an inlier, if it is smaller than a threshold  $T$ . To scale the threshold  $T$  within the range of [0, 1], we apply exponential mapping  $\exp(-10T)$  when calculating robustness metric.
+
+Position Robustness  $R_{p}$  & Rotation Robustness  $R_{r}$  The AUC of F-1 score can be defined using both linear velocity and angular velocity, which can reflect the robustness of position and rotation estimation, respectively. Specifically, we define  $R_{p} = \mathrm{AUC}(F_{1}(\mathrm{v}_{\mathrm{e}}))$  and  $R_{r} = \mathrm{AUC}(F_{1}(\omega_{\mathrm{e}}))$ , where  $v_{e}$  and  $\omega_{e}$  are the estimated error of linear velocity and rotation velocity, respectively. Note that not all SLAM solutions can output the velocity at the desired frequency. To address this, we use B-splines [30, 46] to obtain smooth trajectories, compute derivatives for smooth trajectories, and derive estimated linear velocity and angular velocity. For more details, please refer to the supplementary.
+
+# 4. Open SLAM Challenge Results
+
+In this section, we will present our works from two perspectives: accuracy evaluation and robustness evaluation.
+
+# 4.1. Accuracy Evaluation
+
+The results of the ICCV SLAM challenge underscore the necessity for advancements in system robustness. From 29 submissions, we identified 5 winners in the LiDAR category and 4 in the visual category. However, in the sensor fusion track, which addresses both visual and geometric degradation, no submissions met the criteria for success. This result reveals the existing SLAM systems still have lots of space to improve. Sequence characteristics will be detailed in the supplementary material. Table 2 shows the results for both the LiDAR and visual tracks. Unfortunately, there are no current solutions that can balance high accuracy and real-time performance in challenging environments. Since several environments lack geometric features or visual features,
+
+Table 4. Accuracy Performance on Visual Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions.  
+
+<table><tr><td rowspan="2">Team</td><td colspan="6">Visual Degradation (Real World)</td><td colspan="3">Simulation</td><td rowspan="2">Average</td></tr><tr><td>Lowlight 1</td><td>Lowlight 2</td><td>Over Exposure</td><td>Flash Light</td><td>Smoke Room</td><td>Outdoor Night</td><td>End of World</td><td>Moon</td><td>Western Desert</td></tr><tr><td>Peng et al1</td><td>1.063</td><td>1.637</td><td>0.503</td><td>0.44</td><td>0.153</td><td>0.827</td><td>0.038</td><td>0.195</td><td>0.070</td><td>0.547</td></tr><tr><td>Thien et al2</td><td>1.081</td><td>2.054</td><td>1.733</td><td>1.054</td><td>10.532</td><td>7.692</td><td>0.753</td><td>1.228</td><td>1.209</td><td>3.037</td></tr><tr><td>Jiang et al3</td><td>1.019</td><td>1.126</td><td>1.911</td><td>2.341</td><td>3.757</td><td>11.821</td><td>2.154</td><td>0.604</td><td>4.010</td><td>3.193</td></tr><tr><td>Li et al4</td><td>5.768</td><td>7.834</td><td>1.757</td><td>1.295</td><td>5.370</td><td>10.766</td><td>-</td><td>30.07</td><td>-</td><td>8.98*</td></tr><tr><td>Average</td><td>2.232</td><td>3.163</td><td>1.476</td><td>1.282</td><td>4.953</td><td>7.776</td><td>0.982</td><td>8.024</td><td>1.763</td><td></td></tr></table>
+
+![](images/7b8e8b8ab874deecdf5e7d6a75971d02e8e9b765ff81b297d02a906c91599ca0.jpg)  
+Figure 5. From left to right, it shows robustness metric  $R_{p}$  and  $R_{r}$  for LiDAR and visual sequences respectively. Note: This is a summary of results for all sequences, with weights based on the trajectory length. The area under the curve (AUC) represents the robustness  $(R_{p}, R_{r})$ . The x-axis shows velocity thresholds for classifying estimated velocities as inliers and the y-axis is F-1 score.
+
+![](images/55cfc6031a312bd26283cd3f30215cbe3a40100a153c8fd52f3f5b533cf7c593.jpg)
+
+![](images/1ddcbb645616be2d30870b196461b2b337a9d4108afc101e9f4d046bbd9ca7d8.jpg)
+
+![](images/ddea1daf4391967666e91c1708573d8df339b08b3c6cbb9141b9eeb3029f5647.jpg)
+
+the algorithms required more processing times to achieve reasonably high accuracy. Liu et al. at the University of Hong Kong, who achieved the highest accuracy, recorded an Absolute Trajectory Error (ATE) of  $0.588\mathrm{m}$ , but each iteration took 51.3 seconds. In the visual track, Peng et al. from the Samsung Research Center attained a leading ATE of  $0.547\mathrm{m}$ , but each iteration required 183 seconds.
+
+LiDAR Track Discussion To enhance robustness in all-weather environments, a pivotal question arises: what are the limits of current LiDAR SLAM algorithms? Table 3 presents a summary of the ATE/RPE errors observed in real-world geometrically degraded, simulated, and mixed degradation scenarios across the top five teams. In real-world geometrically degraded environments, we observed the first limitation: existing LiDAR solutions struggle in confined spaces such as caves. Almost all algorithms perform worse in Cave (Figure 3S) and Laureal_Cavern environments (Figure 3N) with average ATE 0.696 and 2.432 meters respectively. One reason is that cave environments is the most confined space which may lack sufficient geometric features, compared with other environments like tunnel (Figure 3J) and Urban (Figure 3B) and Nuclear scenes (Figure 3U).
+
+In simulated environments, we encountered the second limitation: existing methods are tailored for fixed motion patterns like vehicles and struggle with unpredictable motion patterns. Since our simulation sequence is derived from TartanAir [47], featuring aggressive and random motion patterns, it might not follow the usual velocity distribution domain and has not been thoroughly tested. As shown in Table 3 (simulation), it resulted in significantly higher ATE errors  $(5.239\mathrm{m})$  compared to real-world's  $(0.836\mathrm{m})$ .
+
+In mixed degradation environments, the third limitation we found is that existing methods cannot actively select the most informative measurements to adapt to new environments. For example, in the Long Corridor sequence (Figure 3 A) with low lighting and a featureless environment, the scenario is intricately designed for collaborative usage of LiDAR and Visual sensors. Algorithms are expected to utilize visual information in geometrically degraded environments while disregarding it in darkness. Similarly, the Block LiDAR sequences (Figure 3 V) aim to simulate sensor drop scenarios frequently encountered in robotic applications. These sequences involve alternating periods of LiDAR and Visual data loss, challenging algorithms to promptly detect sensor failures and switch to the other modality for SLAM. However, the average ATE error  $(1.61\mathrm{m})$  in Table 3 (mixed degradation) is much higher than in geometric degradation scenarios  $(0.836\mathrm{m})$ , suggesting the above limitations. Details are in the supplementary.
+
+Visual Track Discussion To improve the robustness, a pivotal question arises: what are the limits of current visual SLAM? To what extent does image quality impact a visual SLAM system? Table 4 shows a summary of ATE errors from real-world and simulation scenarios from the awarded 4 teams. In real-world scenarios, the first limitation is the lack of anti-noise capability in current methods, especially in low-quality image settings. We assessed visual odometry accuracy in visually degraded environments, including low lighting, sunlight overexposure, flashing lights, smoke-filled rooms, and nighttime outdoor settings (Figure 3). Trajectories in conditions like the Smoke Room and Outdoor Night exhibit significantly higher average ATE er
+
+Table 5. Robustness Performance on Geometric Degradation. Red numbers are robustness ranking. Larger values indicate better robustness.  
+
+<table><tr><td rowspan="2" colspan="2">Team</td><td colspan="6">Geometric Degradation (Real World)</td><td colspan="3">Simulation</td><td colspan="3">Mix Degradation</td><td rowspan="2">Average</td></tr><tr><td>Urban</td><td>Tunnel</td><td>Cave</td><td>Nuclear_1</td><td>Nuclear_2</td><td>Laurel_Caverns</td><td>Factory</td><td>Ocean</td><td>Sewerage</td><td>Long Corridor</td><td>Multi Floor</td><td>Block Lidar</td></tr><tr><td>Liu et al1</td><td rowspan="5">Rp↑</td><td>0.811</td><td>0.865</td><td>0.736</td><td>0.747</td><td>0.504</td><td>0.816</td><td>0.157</td><td>0.135</td><td>0.135</td><td>0.396</td><td>0.529</td><td>0.371</td><td>0.516</td></tr><tr><td>Weitong et al2</td><td>0.773</td><td>0.838</td><td>0.690</td><td>0.871</td><td>0.739</td><td>0.152</td><td>0.153</td><td>0.149</td><td>0.081</td><td>0.407</td><td>0.456</td><td>0.345</td><td>0.471</td></tr><tr><td>Kim et al4</td><td>0.747</td><td>0.865</td><td>0.737</td><td>0.870</td><td>0.689</td><td>0.777</td><td>0.014</td><td>0.001</td><td>0.010</td><td>0.410</td><td>0.285</td><td>0.345</td><td>0.479</td></tr><tr><td>Yibin et al3</td><td>0.650</td><td>0.783</td><td>0.557</td><td>0.721</td><td>0.490</td><td>0.481</td><td>0.219</td><td>0.560</td><td>0.019</td><td>0.175</td><td>0.260</td><td>0.370</td><td>0.440</td></tr><tr><td>Zhong et al5</td><td>0.567</td><td>0.683</td><td>0.204</td><td>0.680</td><td>0.426</td><td>0.447</td><td>0.103</td><td>0.084</td><td>0.122</td><td>0</td><td>0</td><td>0</td><td>0.276</td></tr><tr><td>Liu et al1</td><td rowspan="5">Rr↑</td><td>0.888</td><td>0.893</td><td>0.816</td><td>0.857</td><td>0.778</td><td>0.861</td><td>0.692</td><td>0.719</td><td>0.652</td><td>0.753</td><td>0.689</td><td>0.643</td><td>0.770</td></tr><tr><td>Weitong et al2</td><td>0.886</td><td>0.892</td><td>0.811</td><td>0.893</td><td>0.840</td><td>0.598</td><td>0.688</td><td>0.720</td><td>0.514</td><td>0.766</td><td>0.706</td><td>0.644</td><td>0.746</td></tr><tr><td>Kim et al4</td><td>0.680</td><td>0.893</td><td>0.816</td><td>0.893</td><td>0.803</td><td>0.808</td><td>0.198</td><td>0.125</td><td>0.300</td><td>0.802</td><td>0.422</td><td>0.642</td><td>0.615</td></tr><tr><td>Yibin et al3</td><td>0.624</td><td>0.707</td><td>0.462</td><td>0.733</td><td>0.540</td><td>0.467</td><td>0.862</td><td>0.962</td><td>0.496</td><td>0.477</td><td>0.342</td><td>0.213</td><td>0.573</td></tr><tr><td>Zhong et al5</td><td>0.731</td><td>0.745</td><td>0.437</td><td>0.757</td><td>0.603</td><td>0.523</td><td>0.688</td><td>0.704</td><td>0.643</td><td>0</td><td>0</td><td>0</td><td>0.485</td></tr><tr><td colspan="2">AverageRp</td><td>0.710</td><td>0.807</td><td>0.585</td><td>0.778</td><td>0.570</td><td>0.535</td><td>0.130</td><td>0.186</td><td>0.073</td><td>0.278</td><td>0.306</td><td>0.286</td><td></td></tr><tr><td colspan="2">Average Rr</td><td>0.762</td><td>0.826</td><td>0.668</td><td>0.827</td><td>0.713</td><td>0.651</td><td>0.626</td><td>0.646</td><td>0.521</td><td>0.560</td><td>0.432</td><td>0.428</td><td></td></tr></table>
+
+Table 6. Robustness Performance on Visual Degradation. Red numbers are robustness ranking. Larger values indicate better robustness.  
+
+<table><tr><td rowspan="2" colspan="2">Team</td><td colspan="6">Visual Degradation (Real World)</td><td colspan="3">Simulation</td><td rowspan="2">Average</td></tr><tr><td>Lowlight 1</td><td>Lowlight 2</td><td>Over Exposure</td><td>Flash Light</td><td>Smoke Room</td><td>Outdoor Night</td><td>End of World</td><td>Moon</td><td>Western Desert</td></tr><tr><td>Peng et al1</td><td></td><td>0.357</td><td>0.227</td><td>0.264</td><td>0.203</td><td>0.536</td><td>0.270</td><td>0.699</td><td>0.893</td><td>0.806</td><td>0.472</td></tr><tr><td>Thien et al3</td><td></td><td>0.045</td><td>0.070</td><td>0.240</td><td>0.156</td><td>0.131</td><td>0.075</td><td>0</td><td>0.031</td><td>0</td><td>0.083</td></tr><tr><td>Jiang et al4</td><td>Rp↑</td><td>0.046</td><td>0.039</td><td>0.194</td><td>0.088</td><td>0.242</td><td>0.095</td><td>0</td><td>0</td><td>0</td><td>0.078</td></tr><tr><td>Li et al2</td><td></td><td>0.342</td><td>0.187</td><td>0.257</td><td>0.208</td><td>0.322</td><td>0.142</td><td>0</td><td>0.006</td><td>0</td><td>0.162</td></tr><tr><td>Peng et al1</td><td></td><td>0.641</td><td>0.581</td><td>0.744</td><td>0.650</td><td>0.878</td><td>0.670</td><td>0.975</td><td>0.975</td><td>0.974</td><td>0.787</td></tr><tr><td>Thien et al3</td><td></td><td>0.413</td><td>0.445</td><td>0.610</td><td>0.269</td><td>0.474</td><td>0.487</td><td>0.177</td><td>0.315</td><td>0</td><td>0.354</td></tr><tr><td>Jiang et al4</td><td>Rr↑</td><td>0.453</td><td>0.452</td><td>0.619</td><td>0.252</td><td>0.542</td><td>0.577</td><td>0.002</td><td>0</td><td>0.157</td><td>0.339</td></tr><tr><td>Li et al2</td><td></td><td>0.642</td><td>0.574</td><td>0.657</td><td>0.651</td><td>0.773</td><td>0.660</td><td>0.0</td><td>0.305</td><td>0</td><td>0.473</td></tr><tr><td colspan="2">AverageRp</td><td>0.198</td><td>0.131</td><td>0.239</td><td>0.164</td><td>0.308</td><td>0.146</td><td>0.175</td><td>0.232</td><td>0.202</td><td></td></tr><tr><td colspan="2">Average Rr</td><td>0.537</td><td>0.513</td><td>0.658</td><td>0.456</td><td>0.667</td><td>0.598</td><td>0.288</td><td>0.399</td><td>0.283</td><td></td></tr></table>
+
+rors (4.95m, 7.76m respectively). Smoke and night scenes challenge feature extraction and tracking and introduce sensor noise, emphasizing the need for robust anti-noise algorithms. The second limitation is that existing methods struggle to overcome aggressive motion. Even in simulations with relatively good image quality, most methods still show significant average ATE errors (8.024m) on the Moon sequence (Figure 3 Y). See more details in supplementary.
+
+# 4.2. Robustness Evaluation
+
+The F-1 robustness curve provides a detailed evaluation under different error tolerances, while  $R_{p}$  and  $R_{r}$  provide an overall measure of robustness that is not dependent on a specific decision threshold. This suggests the flexibility of our new robustness metric. Figure 5 clearly shows the robustness performance of all teams regarding position and rotation, highlighting the differences in robustness between the methods. The values of  $R_{p}$  and  $R_{r}$ , which are the AUC under F-1 curves, are displayed in the brackets of the legends. It reveals that Liu et al. and Peng et al. are the most robust solutions for LiDAR and visual tracks, respectively.
+
+Is the robustness metric robust? We did extensive experiments on our robustness metric against diverse sensor degradation settings. Table 5 and 6 present a summary of robustness performance in real-world, simulated, and mixed degradation scenarios for both LiDAR and visual sequences. In Table 5, we observe that the average values of  $R_{p}$  and  $R_{r}$  for the mixed degradation environment (0.290,
+
+0.473) are significantly lower than geometric degradation environments (0.664, 0.754). This observation suggests that the majority of SLAM algorithms exhibit reduced robustness in mixed degradation environments as compared to geometrically degraded ones. This aligns with our expectations and verifies the effectiveness of the robustness metric.
+
+Our robustness ranking, indicated by small red numbers, differs from those on ATE ranking shown in Table 3 and 4. This is because our F-1 score-based metrics  $R_{p}$  and  $R_{r}$  jointly consider precision and recall rate to provide a balanced evaluation of SLAM performance, considering both metrics across the full spectrum of thresholds. In contrast, the ATE focuses solely on precision, neglecting the trajectory's completeness (recall). We also evaluate our metric on various synthetic trajectories in the supplementary.
+
+# 5. Conclusion
+
+We introduce SubT-MRS, a comprehensive SLAM dataset with various sensor data, locomotion patterns, and over 30 degradation types in simulation and real-world settings to push SLAM towards all-weather environments. Additionally, we introduce a new robustness metric to evaluate the reliability of SLAM systems, enhancing robot safety control. 29 teams have tested SubT-MRS in the organized SLAM challenge and we expect that it will serve as a critical benchmark for future SLAM development.
+
+# References
+
+[1] https://www.darpa.mil/program/darpa-subterranean-challenge. 1  
+[2] Chunge Bai, Tao Xiao, Yajie Chen, Haoqian Wang, Fang Zhang, and Xiang Gao. Faster-lio: Lightweight tightly coupled lidar-inertial odometry using parallel sparse incremental voxels. IEEE Robotics and Automation Letters, 7(2):4861-4868, 2022. 6  
+[3] Michael Burri, Janosch Nikolic, Pascal Gohl, Thomas Schneider, Joern Rehder, Sammy Omari, Markus W Achtelik, and Roland Siegwart. The EuRoC micro aerial vehicle datasets. The International Journal of Robotics Research, 2016. 1, 2, 3, 6  
+[4] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 6  
+[5] Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel, and Juan D. Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6): 1874-1890, 2021. 1  
+[6] Nicholas Carlevaris-Bianco, Arash K Ushani, and Ryan M Eustice. University of michigan north campus long-term vision and lidar dataset. The International Journal of Robotics Research, 35(9):1023-1035, 2016. 3  
+[7] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J Taylor, and Kostas Daniilidis. M3ed: Multi-robot, multi-sensor, multi-environment event dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015–4022, 2023. 3  
+[8] Kenny Chen, Brett T. Lopez, Ali-akbar Agha-mohammadi, and Ankur Mehta. Direct lidar odometry: Fast localization with dense point clouds. IEEE Robotics and Automation Letters, 7(2):2000–2007, 2022. 6  
+[9] Long Chen, Yuchen Li, Chao Huang, Bai Li, Yang Xing, Daxin Tian, Li Li, Zhongxu Hu, Xiaoxiang Na, Zixuan Li, et al. Milestones in autonomous driving and intelligent vehicles: Survey of surveys. IEEE Transactions on Intelligent Vehicles, 8(2):1046-1056, 2022. 2  
+[10] Weichen Dai, Yu Zhang, Shenzhou Chen, Donglei Sun, and Da Kong. A multi-spectral dataset for evaluating motion estimation systems. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5560-5566. IEEE, 2021. 2  
+[11] Frank Dellaert, Michael Kaess, et al. Factor graphs for robot perception. Foundations and Trends® in Robotics, 6(1-2): 1-139, 2017. 6  
+[12] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the UZH-FPV drone racing dataset. In IEEE Int. Conf. Robot. Autom. (ICRA), 2019. 1, 3  
+[13] Rodolphe Dubois, Alexandre Eudes, and Vincent Frémont. Airmuseum: a heterogeneous multi-robot dataset for stereovisual and inertial simultaneous localization and mapping. In
+
+2020 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI), pages 166-172, 2020. 3  
+[14] Naser El-Sheimi, Haiying Hou, and Xiaoji Niu. Analysis and modeling of inertial sensors using allan variance. IEEE Transactions on Instrumentation and Measurement, 57(1): 140-149, 2008. 4  
+[15] Dapeng Feng, Yuhua Qi, Shipeng Zhong, Zhiqiang Chen, Yudu Jiao, Qiming Chen, Tao Jiang, and Hongbo Chen. S3e: A large-scale multimodal dataset for collaborative slam. arXiv preprint arXiv:2210.13723, 2022. 2  
+[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4340-4349, 2016. 3  
+[17] Wenliang Gao. A ros package tool to analyze the imu performance, 2018. 4  
+[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI Vision Benchmark Suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 1  
+[19] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 3  
+[20] Dongjiao He, Wei Xu, Nan Chen, Fanze Kong, Chongjian Yuan, and Fu Zhang. Point-lio: Robust high-bandwidth light detection and ranging inertial odometry. Advanced Intelligent Systems, page 2200459, 2023. 6  
+[21] Michael Helmberger, Kristian Morin, Beda Berner, Nitish Kumar, Giovanni Cioffi, and Davide Scaramuzza. The hilti slam challenge dataset. IEEE Robotics and Automation Letters, 7(3):7518-7525, 2022. 1, 3  
+[22] Giseop Kim and Ayoung Kim. Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4802-4809. IEEE, 2018. 6  
+[23] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. Vivid++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, 2022. 2  
+[24] Keith YK Leung, Yoni Halpern, Timothy D Barfoot, and Hugh HT Liu. The utias multi-robot cooperative localization and mapping dataset. The International Journal of Robotics Research, 30(8):969–974, 2011. 2  
+[25] Hyungtae Lim, Suyong Yeon, Suyong Ryu, Yonghan Lee, Youngji Kim, Jaeseong Yun, Euigon Jung, Donghwan Lee, and Hyun Myung. A single correspondence is enough: Robust global registration to avoid degeneracy in urban environments. page Accepted. To appear, 2022. 6  
+[26] Yicheng Lin, Shuo Wang, Yunlong Jiang, and Bin Han. Breaking of brightness consistency in optical flow with a lightweight cnn network, 2023. 6  
+[27] Zheng Liu, Xiyuan Liu, and Fu Zhang. Efficient and consistent bundle adjustment on lidar point clouds. IEEE Transactions on Robotics, 2023. 6
+
+[28] Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year,  $1000\mathrm{km}$ : The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3-15, 2017. 3  
+[29] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research, 36(2):142-149, 2017. 2  
+[30] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Transactions on Robotics, 34(6): 1425-1440, 2018. 6  
+[31] Thien Hoang Nguyen, Shenghai Yuan, and Lihua Xie. Vrslam: A visual-range simultaneous localization and mapping system using monocular camera and ultra-wideband sensors, 2023. 6  
+[32] Xiongfeng Peng, Zhihua Liu, Weiming Li, Ping Tan, SoonYong Cho, and Qiang Wang. Dvi-slam: A dual visual inertial slam network, 2023. 6  
+[33] Bernd Pfrommer, Nitin Sanket, Kostas Daniilidis, and Jonas Cleveland. Penncosyvio: A challenging visual inertial odometry benchmark. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3847-3854. IEEE, 2017. 1, 3  
+[34] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics, 34(4):1004-1020, 2018. 1, 6  
+[35] Milad Ramezani, Yiduo Wang, Marco Camurri, David Wisth, Matias Mattamala, and Maurice Fallon. The newer college dataset: Handheld lidar, inertial and vision with ground truth. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4353-4360. IEEE, 2020. 1  
+[36] Joern Rehder, Janosch Nikolic, Thomas Schneider, Timo Hinzmann, and Roland Siegwart. Extending kalibr: Calibrating the extrinsic of multiple imus and of individual axes. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 4304-4311. IEEE, 2016. 3, 4  
+[37] John G Rogers, Arthur Schang, Carlos Nieto-Granda, John Ware, John Carter, Jonathan Fink, and Ethan Stump. The darpa subt urban circuit mapping dataset and evaluation metric. In Experimental Robotics: The 17th International Symposium, pages 391–401. Springer, 2021. 3  
+[38] Sebastian Scherer, Vasu Agrawal, Graeme Best, Chao Cao, Katarina Cujic, R Darnley, R DeBortoli, E Dexheimer, B Drozd, R Garg, et al. Resilient and modular subterranean exploration with a team of roving and flying robots. Field Robotics, 2022. 4  
+[39] David Schubert, Thore Goll, Nikolaus Demmel, Vladyslav Usenko, Jörg Stuckler, and Daniel Cremers. The tum vi benchmark for evaluating visual-inertial odometry. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1680-1687. IEEE, 2018. 1, 3  
+[40] Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-icp. In Robotics: science and systems, page 435. Seattle, WA, 2009. 4, 5
+
+[41] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Rus Daniela. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5135-5142. IEEE, 2020. 1  
+[42] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In Proc. of the International Conference on Intelligent Robot Systems (IROS), 2012. 2  
+[43] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1  
+[44] Yulun Tian, Yun Chang, Long Quang, Arthur Schang, Carlos Nieto-Granda, Jonathan P How, and Luca Carlone. Resilient and distributed multi-robot visual slam: Datasets, experiments, and lessons learned. arXiv preprint arXiv:2304.04362, 2023. 2  
+[45] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters, 8(2):1029–1036, 2023. 6  
+[46] Chen Wang, Dasong Gao, Kuan Xu, Junyi Geng, Yaoyu Hu, Yuheng Qiu, Bowen Li, Fan Yang, Brady Moon, Abhinav Pandey, Aryan, Jiahe Xu, Tianhao Wu, Haonan He, Daning Huang, Zhongqiang Ren, Shibo Zhao, Taimeng Fu, Pranay Reddy, Xiao Lin, Wenshan Wang, Jingnan Shi, Rajat Talak, Kun Cao, Yi Du, Han Wang, Huai Yu, Shanzhao Wang, Siyu Chen, Ananth Kashyap, Rohan Bandaru, Karthik Dantu, Jiajun Wu, Lihua Xie, Luca Carlone, Marco Hutter, and Sebastian Scherer. PyPose: A library for robot learning with physics-based optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6  
+[47] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual SLAM. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 3, 7  
+[48] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. Tartanvo: A generalizable learning-based vo. In Conference on Robot Learning, pages 1761-1772. PMLR, 2021. 1  
+[49] P. Wenzel, R. Wang, N. Yang, Q. Cheng, Q. Khan, L. von Stumberg, N. Zeller, and D. Cremers. 4Seasons: A cross-season dataset for multi-weather SLAM in autonomous driving. In Proceedings of the German Conference on Pattern Recognition (GCPR), 2020. 2  
+[50] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. FAST-LIO2: fast direct lidar-inertial odometry. IEEE Transactions on Robotics, abs/2107.06829, 2022. 6  
+[51] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. Fast-lio2: Fast direct lidar-inertial odometry. IEEE Transactions on Robotics, 38(4):2053–2073, 2022. 6  
+[52] Lintong Zhang, Marco Camurri, and Maurice Fallon. Multicamera lidar inertial extension to the newer college dataset. arXiv preprint arXiv:2112.08854, 2021. 3  
+[53] Shibo Zhao, Hengrui Zhang, Peng Wang, Lucas Nogueira, and Sebastian Scherer. Super odometry: Imu-centric
+
+lidar-visual-inertial estimator for challenging environments. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8729-8736. IEEE, 2021. 4  
+[54] David Zuniga-Noël, Alberto Jaenal, Ruben Gomez-Ojeda, and Javier Gonzalez-Jimenez. The uma-vi dataset: Visual-inertial odometry in low-textured and dynamic illumination environments. The International Journal of Robotics Research, 39(9):1052-1060, 2020. 1, 3

2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dca19bf33ee81173ed05a5c0bc919e488852a9738c80476b1513dca7daa91725
+size 961835

2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/5f051557-0067-4cd2-89f8-ff27dddafa6f_content_list.json

@@ -0,0 +1,1430 @@
+[
+    {
+        "type": "text",
+        "text": "Summarize the Past to Predict the Future: Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation",
+        "text_level": 1,
+        "bbox": [
+            96,
+            130,
+            875,
+            176
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Razvan-George Pasca $^{1*}$  Alexey Gavyushin $^{1*}$  Muhammad Hamza $^{2}$  Yen-Ling Kuo $^{3}$  Kaichun Mo $^{4}$  Luc Van Gool $^{1,5,6}$  Otmar Hilliges $^{1}$  Xi Wang $^{1}$  ETH Zurich Univ. of Zurich Univ. of Virginia NVIDIA KU Leuven INSAIT, Sofia",
+        "bbox": [
+            145,
+            202,
+            831,
+            256
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/0f23b4421a71e33743c550be111aa1ad4300ae983910743e70882576c09612d4.jpg",
+        "image_caption": [
+            "Figure 1. TransFusion: Multimodal fusion transformer for short-term object interaction anticipation. Given a video sequence of past observations, the task aims to predict a set of objects visible in the current frame that will be interacted with in the future, i.e. in the activation frame that is  $\\Delta$  frames away from the current prediction frame. Additionally, the task requires estimating the bounding box, the associated action described by a verb-noun pair, and the time to contact for each predicted object. We propose TransFusion, a multimodal fusion architecture that uses language summaries of past actions and salient objects to effectively predict future object interactions."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            147,
+            267,
+            831,
+            500
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            233,
+            604,
+            313,
+            619
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We study object interaction anticipation in egocentric videos. This task requires an understanding of the spatiotemporal context formed by past actions on objects, coined action context. We propose TransFusion, a multimodal transformer-based architecture for short-term object interaction anticipation. Our method exploits the representational power of language by summarizing the action context textually, after leveraging pre-trained vision-language foundation models to extract the action context from past video frames. The summarized action context and the last observed video frame are processed by the multimodal fusion module to forecast the next object interaction. Experiments on the Ego4D next active object interaction dataset show the effectiveness of our multimodal fusion model and highlight the benefits of using the power of foundation models and language-based context summaries in a task where",
+        "bbox": [
+            75,
+            635,
+            472,
+            878
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "vision may appear to suffice. Our novel approach outperforms all state-of-the-art methods on both versions of the Ego4D dataset. A project video and code are available at https://eth-ait.github.io/transfusion-proj/.",
+        "bbox": [
+            496,
+            604,
+            893,
+            667
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1. Introduction",
+        "text_level": 1,
+        "bbox": [
+            500,
+            691,
+            633,
+            709
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The ability to predict plausible future human-object interactions is important for many assistive technologies. The task of short-term object interaction anticipation [18, 22] is defined as predicting which object a user will interact with next, what action will be performed and when in the future, given an egocentric video input. Providing an effective solution would help artificial agents, such as care robots, assist humans in their daily activities, as well as enable safety applications such as child and elderly monitoring.",
+        "bbox": [
+            496,
+            718,
+            893,
+            854
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "This task is challenging because it requires reasoning about user intent [6, 35, 43], which is not directly observable. For example, if the video shows the user holding",
+        "bbox": [
+            496,
+            854,
+            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_footnote",
+        "text": "*Authors contributed equally.",
+        "bbox": [
+            94,
+            886,
+            254,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "18286",
+        "bbox": [
+            480,
+            944,
+            519,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "a vegetable, the most likely next action is to slice it, but only with a knife and cutting board present. Without those, there is significantly more uncertainty. Therefore, understanding what has happened in the past and what relevant objects are present is critical for predicting the next object interaction. We collectively refer to these concepts as action context. Existing interaction forecasting methods do not explicitly model action context and rely on neural networks to extract that information from fixed-sized video chunks [5, 18, 22, 32].",
+        "bbox": [
+            75,
+            90,
+            472,
+            241
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose to use language summarization of the past as an explicit representation of action context. This filters out ambiguities caused by visual clutter. It also puts focus on objects and actions directly relevant to the ongoing and likely future activities. For the object interaction anticipation, we next introduce a multimodal fusion approach. Specifically, we propose a transformer-based multimodal fusion model, TransFusion. As shown in Figure 1, our model takes language summaries of the action context and the last observed single frame of a video as input and predicts future object interactions. We employ an existing image captioning system [53] to generate a concise description of past actions given by pairs of verbs and nouns. We use CLIP [38], a vision-language model trained on pairs of images and captions, to detect salient objects in the prediction frame of interest, and deem these objects task-relevant. The verb-noun pairs together with the list of salient objects form the summary of the past action context. Leveraging the vast knowledge contained in pre-trained vision-language foundation models yields more generic representations of the past action context and enables us to better generalize towards various scenarios, in particular rare ones not well-represented in the dataset.",
+        "bbox": [
+            75,
+            243,
+            472,
+            589
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our experiments show that our language-based summaries can effectively represent the past action context. We evaluate on both versions of the challenging egocentric dataset Ego4D [22]. Our approach improves interaction prediction accuracy and outperforms state-of-the-art models pre-trained on large video datasets [7, 40]. Such improvements show that large vision-language models can be useful even for a task that seems purely visual.",
+        "bbox": [
+            75,
+            592,
+            470,
+            712
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In summary, our contributions are:",
+        "bbox": [
+            76,
+            714,
+            307,
+            729
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- A strategy for generating concise language-based summaries of action context over longer timespans from frame-wise language-based summaries, for instance obtained using vision-language foundation models.",
+            "- TransFusion, a multimodal fusion model for short-term object interaction anticipation combining action context descriptions and visual features.",
+            "- Experiments on both versions of the Ego4D dataset showing improvements over state-of-the-art methods on the short-term object interaction anticipation task, in particular for long-tailed classes."
+        ],
+        "bbox": [
+            94,
+            732,
+            467,
+            896
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2. Related work",
+        "text_level": 1,
+        "bbox": [
+            500,
+            89,
+            637,
+            104
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Object interaction anticipation. Several works have been proposed for human-object interaction prediction. Furnari et al. [18] first introduced next-active-object (NAO) prediction, which predicts the active/not active label for each object using its motion trajectory features, but doesn't consider action anticipation. Closely related to NAO, Bertasius et al. [5] introduced action-objects, which considers objects that capture the human actor's visual attention or tactile interaction. Bertasius et al. use the data from a stereo camera system in a two-stream network, integrating RGB and depth information. They show that certain aspects of human actions can be predicted by exploiting the spatial configuration of the objects and the actor's head. Their approach is not easily scalable due to the need for stereo data.",
+        "bbox": [
+            496,
+            118,
+            890,
+            329
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In addition to different tasks, Liu et al. [32] propose a model that improves NAO prediction by incorporating future hand trajectory modeling. They obtain impressive results using 3D CNNs, confirming that hand motion cues and scene dynamics have important explanatory power in predicting the short-term future. Nonetheless, their approach also encounters scaling limitations because additional ground-truth labels for hand trajectories are needed.",
+        "bbox": [
+            496,
+            333,
+            890,
+            455
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The Ego4D dataset [22] proposed object interaction anticipation. Apart from locating the NAO, the task in Ego4D also requires the prediction of associated nouns and verbs, as well as the time to contact (TTC) for the future interacted objects. The official Ego4D baseline uses the current frame and past video frames as input and employs a two-stage approach to first learn noun and box prediction, and afterwards learn verb and TTC prediction. InternVideo [7] similarly leverages a DINO [62] model to detect objects in the first stage, and a VideoMAE [49]-pretrained ViT [14] spacetime attention model to predict verbs and TTC conditioned on the object predictions in the second stage. StillFast [40] presents an end-to-end method for NAO prediction, fusing multi-scale high-resolution image and low-resolution video features and combining local RoI features with global scene features to boost the interaction anticipation performance. GANOv2 [46, 47] and its predecessor [48] extract features from objects detected in past observed frames and fuse them with global frame features, forwarding them to a StillFast predictor and a motion-object-separating encoder-decoder architecture, respectively.",
+        "bbox": [
+            496,
+            457,
+            890,
+            773
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Action anticipation. Besides localizing future object interactions, predicting the next action steps is an equally important task that intelligent systems need to perform to interact with humans. A longer prediction timespan requires a better grasp of the structure behind the succession of actions. A pioneering advancement in this field is the introduction of the EPIC-Kitchens 100 (EK100) [9] dataset, which facilitates research in action anticipation within egocentric",
+        "bbox": [
+            496,
+            780,
+            890,
+            900
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "18287",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "videos. Significant research efforts have been dedicated to exploring the action anticipation task utilizing EK100. Anticipative Video Transformer (AVT) [20] is one of the best-performing architectures on the EK100 [9] action anticipation benchmark. AVT uses a pre-trained Vision Transformer [14] as the image feature backbone and attends to the previously observed video frames to anticipate future actions. MeMViT [54] improves AVT by exploiting a longer temporal context, highlighting the importance of modeling a long sequence of previous actions. Other model architectures such as recurrent neural networks [17, 55], multimodal temporal CNNs [25] and transformers [19, 51] are also used to model the temporal contexts. Note that these methods are designed for action classification using video data and adapting them for object detection is not trivial. It is worth noting that the action anticipation task from the EK100 benchmark is a subtask of our short-term object interaction anticipation task. Our task not only predicts the action label of verb-noun pairs, but we also predict the location of the object and the time to contact, which is more challenging than the action anticipation task itself. TransFusion focuses on detecting the next active objects, as all evaluation metrics are conditioned on the box IoU. Therefore, comparing our work with the state-of-the-art methods for the action anticipation task on EK100 is out of the scope of this work. The Ego4D dataset itself further introduces a long-term action anticipation benchmark, which [61] tackles using pre-trained visual grounding models.",
+        "bbox": [
+            76,
+            90,
+            472,
+            513
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Vision-language models. Our approach combines the visual and language modalities to improve object interaction anticipation. Most recent multimodal vision-language architectures are based on Transformer [50] fusion schemes applied on features extracted by various encoders [4, 12, 26]. We follow their spirit by employing self-attention-based modality fusion. Our work differs in one significant aspect: the language features represent the past action context. This serves as an additional signal to disambiguate the actor's intent given similar visual scenes. The model must learn to predict from the multimodal fusion of language context summaries and visual cues in a considerably smaller data regime.",
+        "bbox": [
+            76,
+            521,
+            468,
+            718
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "More broadly, there are numerous transformer-based works aimed at learning cross-modal representations between vision and language [24, 28, 36, 38, 44, 45, 58, 59, 64] using terabyte-scale datasets crawled from multiple internet sources. When combined with large language models such as GPT-2 [37], BERT [13], Sentence-BERT [41], and the T5 variants [8, 39], these vision-language models can be applied to several downstream tasks such as VQA [3] and image captioning [56, 57]. Similar to other vision-language models that are trained with language input, our approach leverages state-of-the-art image captioning models [53] to provide a consistent description of past actions.",
+        "bbox": [
+            76,
+            719,
+            472,
+            901
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3. Language-guided prediction",
+        "text_level": 1,
+        "bbox": [
+            498,
+            89,
+            759,
+            107
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We use language-based context summaries to represent the spatial-temporal context formed by past actions and object relationships. The context summary offers an explicit yet compact representation, which allows us to focus on task-relevant content and filter out unimportant information from the likely cluttered environment. To this end, we propose TransFusion, a transformer-based model that leverages a joint-attention fusion mechanism to effectively learn to predict object interactions in the future from the past action context. In the following, we describe our framework, which takes video clips as input and anticipates the next object interaction. We begin by defining the object interaction anticipation task (Sec. 3.1). We then show how to generate a summary for a prediction frame from an egocentric video to represent the action context provided by the past frames (Sec. 3.2). Finally, we present our multimodal fusion model's architecture (Sec. 3.3).",
+        "bbox": [
+            496,
+            114,
+            893,
+            375
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1. Task definition",
+        "text_level": 1,
+        "bbox": [
+            498,
+            385,
+            650,
+            400
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given a video  $\\mathcal{V} = \\{f_1,\\dots ,f_T\\}$  of length  $T$  as input, our goal is to predict the next object interaction  $\\mathcal{O}$  in the last observed video frame  $f_{T}$ , possibly using older frames to extract context information. We call  $f_{T}$  the prediction frame.",
+        "bbox": [
+            496,
+            409,
+            890,
+            470
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As illustrated in Figure 1, the task is to anticipate the object interaction in  $f_{T}$ .  $\\mathcal{O}_T = \\{o_1^T,\\dots ,o_N^T\\}$  denotes the set of future object interactions, with  $N > 1$  whenever multiple objects are interacted with simultaneously. In other words, multiple  $o_i^T$  can be detected in the prediction frame  $f_{T}$ . For example,  $N = 2$  when one hand is holding two mugs in a kitchen scene, or the left hand is touching a plant while the right hand is moving a pot in a garden scene. Each object interaction  $o_i^T = (b_i^T,n_i^T,v_i^T,t_i^T)$  consists of the object bounding box  $b_{i}^{T}$ , the semantic object label (noun)  $n_i^T$ , the action label (verb)  $v_{i}^{T}$ , and the time to contact (TTC)  $t_i^T$ . The task is to predict the next active object location  $b_{i}^{T}$  in the prediction frame  $f_{T}$ , the associated action described by the verb-noun pair  $(n_i^T,v_i^T)$ , and the time  $t_i^T$  to the point of contact when the interaction is going to start. The task is set up such that the actual object interaction takes place after a buffer time  $\\Delta$  on frame  $f_{T + \\Delta}$  and the predicted time to contact  $t_i^T = \\Delta$ . Ego4D uses  $\\Delta \\geq 0.033s$ . This requires anticipating future actions and increases the task difficulty.",
+        "bbox": [
+            496,
+            472,
+            893,
+            758
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2. Summarizing the past",
+        "text_level": 1,
+        "bbox": [
+            500,
+            770,
+            705,
+            787
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Anticipating object interactions in the future requires an understanding of what happened in the past. We first infer the per-frame action context from each egocentric video. This process consists of extracting action descriptions and held objects. These information sources are generated and aggregated independently across frames and finally forwarded as the action context to our anticipation model.",
+        "bbox": [
+            496,
+            795,
+            893,
+            902
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "18288",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/dd1f4c7a5e8616dfd8819f244e47559a7fb781f4fd0cbae7c8aa31790cffbb6f.jpg",
+        "image_caption": [
+            "Figure 2. Overview of the TransFusion model. TransFusion takes the prediction frame and the action context summary as input and predicts the bounding box of the next-active object, the noun-verb pair describing the associated action, and the time to contact (TTC). Feature maps of different scales are extracted from the visual encoder and then fused via a multimodal fusion module with the encoded language features. Their output is then processed by a regular feature pyramid network (FPN), denoted as multi-scale region proposal networks, before being fed into the Faster R-CNN detector."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            76,
+            87,
+            893,
+            255
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Extracting frame-wise action descriptions. Inspired by the label format used in egocentric video datasets [10, 22], we aim to describe the past actions by sequences of verb-noun pairs (e.g. \"wash tomato\", \"cut tomato\"), so-called action descriptions. For each video frame  $f$ , we extract the action currently performed by the agent, represented by a single pair  $(v, n)$  of a verb  $v$  and a noun  $n$ .",
+        "bbox": [
+            75,
+            343,
+            472,
+            450
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We use an off-the-shelf image captioning model [53] to generate multiple full-sentence captions for each frame of interest, so as to find repeatedly occurring outputs and reduce noise: we perform part-of-speech tagging [2] on these captions and collect candidate pairs consisting of verbs followed by nouns within some cutoff distance  $d \\in \\mathbb{N}_0^+$ . Setting  $d > 0$  allows us to account for additional words appearing between a verb-noun pair (e.g. \"eat apple\" from \"A person eating a red apple\") and thereby obtain more valid candidate pairs from the captions. However, larger cutoff distances may introduce more spurious detections (e.g. \"eat gathering\", from \"A person eating while at a gathering\"). We use  $d = 4$  in our experiments. The most frequent candidate pair across the frame's captions is selected as the verb-noun action description of this frame. If no valid verb-noun pair is found, the frame will have no action description.",
+        "bbox": [
+            75,
+            459,
+            472,
+            700
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Extracting frame-wise salient objects. Another characteristic shared by many daily action sequences is that the next active object is likely to have already appeared in the actor's view before the upcoming interaction. The current environment in the recording may also provide model conditioning. Therefore, we select a set of salient objects in each frame as (1) the potential candidates for future active objects and (2) a proxy of the surroundings.",
+        "bbox": [
+            75,
+            703,
+            468,
+            824
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Encouraged by its previous successful use in action-related tasks [30, 52], we use CLIP [38], a pre-trained open-vocabulary zero-shot classifier, to compute the cosine similarities between the CLIP image embedding of a frame and the embeddings of the object categories from the Ego4D do",
+        "bbox": [
+            75,
+            825,
+            470,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "main. We choose to keep the highest-scoring  $k$  objects as the set of salient objects for a given frame. A small  $k$  may fail to provide sufficient information, while a large  $k$  makes capturing the true salient objects more likely, yet introduces more noise due to false detections. We set  $k = 3$ .",
+        "bbox": [
+            496,
+            343,
+            893,
+            419
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Aggregating summaries over multiple frames. After the frame-wise summarization step, we have, for each frame  $t$ , at most one verb-noun pair describing the frame-wise action  $a_{t} = (n,v) \\in \\mathcal{A}_{t}^{F}$ . We then apply a cross-frame aggregation scheme to all  $\\mathcal{A}_{t}^{F}$  so as to identify contiguous segments of frames exhibiting the same activity, while accounting for individual noisy frames possibly not matching the framework summaries of their neighbors. The aggregation processes the frames in temporal order. We consider frames belonging to the same atomic action as an action segment, which is initiated by observing a number of reoccurrences of some verb-noun pair and terminated by a lack of its occurrence within a number of contiguous frames. After the aggregation, we obtain  $\\mathcal{A}$ , the sequence of all action segments in the video. For each prediction frame, we construct its action context from a number of action segments preceding the frame, and the objects from  $\\mathcal{N}_s$  in that frame.",
+        "bbox": [
+            496,
+            431,
+            895,
+            688
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "By aggregating across frames, we thus capture dominant actions in the past, reduce noise in the generated action context and represent long timespans of similar activity by short textual descriptions. The complete action context is provided to our model by concatenating the textual representations of the aggregation results of any desired subset of the two context representations  $\\{\\mathcal{A},\\mathcal{N}_s\\}$ , where we drop the index  $t$  for convenience. For more details, see the Supp. Mat.",
+        "bbox": [
+            496,
+            688,
+            893,
+            809
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3. TransFusion: Multimodal fusion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            816,
+            790,
+            832
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Figure 2 visualizes the architecture of the proposed multimodal TransFusion model. TransFusion employs two different base encoders, for language and image input respectively. The input prediction frame  $f_{T}$  is processed by a regu",
+        "bbox": [
+            496,
+            839,
+            893,
+            902
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "18289",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/ec1da6b14eb441a83d93e834c2f4a065392b12cc24c9c1f485faa788592b3b40.jpg",
+        "image_caption": [
+            "Figure 3. Multimodal fusion module. We first project the CNN feature map and the language tokens to a common dimensionality before adding the specific embeddings. We concatenate the visual and language tokens and feed them to  $M$  self-attention layers. At the output, the fused visual tokens are projected back into the initial feature map shape."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            87,
+            433,
+            348
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "lar CNN-based visual encoder [23], producing a set of multiscale feature maps  $\\mathcal{F}_{\\mathcal{V}}^{s}(f_{T})$ , with  $s$  indicating the scale. The frame's action context is encoded by a language encoder [41] and yields  $\\mathcal{F}_{\\mathcal{L}}(\\mathcal{C}_L)$ . The feature maps belonging to different scale levels from the visual encoder are fused with the encoded language features via a joint-attention scheme. The result is then processed by a regular feature pyramid network (FPN) [29] before being forwarded to the Faster R-CNN [42] detector.",
+        "bbox": [
+            75,
+            455,
+            468,
+            590
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Multimodal fusion. For simplicity, we describe the multimodal fusion module for a single input scale  $s$ , which we hence omit from notation. See Figure 3 for a detailed view of a single Transformer encoder layer together with the input projection stages. For the visual modality, we first split the image into a sequence of tokens. Specifically, the visual features  $\\mathcal{F}_{\\mathcal{V}}(f_T) \\in \\mathbb{R}^{C \\times H \\times W}$  are reshaped to  $\\mathbb{R}^{N \\times (P^2 \\cdot C)}$  where  $P$  is the token size and  $N = \\frac{H \\times W}{P^2}$  the number of tokens. We add unique type embeddings as well as a positional encoding to each modality. Both visual and language tokens are first mapped to  $\\mathbb{R}^D$  via a linear projection before being fed to the fusion module in a concatenated form. The fusion block consists of a regular Transformer encoder layer replicated  $M$  times. Finally, the visual tokens are regrouped into their initial shape to compose the feature map for the feature pyramid network (FPN).",
+        "bbox": [
+            75,
+            594,
+            472,
+            838
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Multiscale fusion. The multimodal fusion block outputs multiple fused feature maps of varying scales:  $\\mathcal{F}_{fi} \\in \\mathbb{R}^{C_i \\times H_i \\times W_i}$ , where  $i$  is the scale index. They are fed into the FPN, which has different parameters for each feature",
+        "bbox": [
+            76,
+            839,
+            472,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "map, allowing the network to learn specialized features for each downsampling ratio. The output of the FPN is processed by a Faster R-CNN detector.",
+        "bbox": [
+            496,
+            90,
+            890,
+            136
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Prediction head networks. The Faster R-CNN detector outputs the bounding box  $b$  through a regression layer, an associated object  $n$  and verb  $v$ , and time to contact  $t$ .",
+        "bbox": [
+            496,
+            138,
+            890,
+            184
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training objective. We train TransFusion using the standard Faster R-CNN [42] objective  $\\mathcal{L}_{box}$  for bounding box prediction, which consists of localization and objectness terms; further, cross-entropy losses  $\\mathcal{L}_{noun}$  and  $\\mathcal{L}_{verb}$  for noun and verb prediction respectively, and an L1 loss  $\\mathcal{L}_{ttc}$  for time to contact prediction. The overall objective is",
+        "bbox": [
+            496,
+            186,
+            890,
+            277
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {b o x}} + \\mathcal {L} _ {\\text {n o u n}} + \\mathcal {L} _ {\\text {v e r b}} + \\mathcal {L} _ {\\text {t t c}}. \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            573,
+            289,
+            890,
+            306
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As defined in Faster R-CNN,  $\\mathcal{L}_{box}$  is",
+        "bbox": [
+            519,
+            316,
+            764,
+            330
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {b o x}} = \\frac {1}{N _ {\\mathrm {c l s}}} \\sum_ {i} \\mathcal {L} _ {\\mathrm {c l s}} \\left(p _ {i}, p _ {i} ^ {*}\\right) + \\frac {\\lambda}{N _ {\\mathrm {r e g}}} \\sum_ {i} p _ {i} ^ {*} \\cdot L _ {\\mathrm {r e g}} \\left(b _ {i} - b _ {i} ^ {*}\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            500,
+            340,
+            890,
+            388
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $p_i$  is the predicted probability of a box being classified as a foreground object (true positive),  $p_i^*$  is the ground-truth label;  $b_i$  and  $b_i^*$  are the predicted and ground-truth box coordinates.  $N_{\\mathrm{reg}}$  is the number of bounding boxes used in training,  $N_{\\mathrm{cls}}$  the detection network's image batch size. We leave  $\\lambda = 11$  as in the reference implementation.",
+        "bbox": [
+            496,
+            390,
+            890,
+            481
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Experiments",
+        "text_level": 1,
+        "bbox": [
+            500,
+            494,
+            632,
+            511
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Dataset. We use both versions (v1 and v2) of the dataset provided for the short-term object interaction anticipation task in Ego4D [22]. It contains various activities ranging from gardening work to cooking or car parts changing. There are in total 64,798 (v2: 165,451) annotated video clips of  $\\sim 5$  minutes each with frame rate  $fps = 30$ . The dataset has a highly imbalanced long tail distribution, including 87 nouns and 74 verbs (v2: 128 and 81 resp.). For instance, the \"take\" class encompasses almost  $40\\%$  of all the verb labels in version 1, while \"mold\" appears only in one video. The train set consists of about 28k (v2: 98k) samples, while the rest are roughly equally split between validation and test.",
+        "bbox": [
+            496,
+            518,
+            890,
+            715
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Evaluation. As in the official Ego4D benchmark [22], we use the Top-5 mean Average Precision (mAP) to evaluate the performance of individual predictions. It considers two constraints: The IoU constraint requires that the predicted boxes are counted as hits only if there is an intersection over union IoU  $\\geq 0.5$  with the ground-truth box. The TTC constraint requires that the predicted time-to-contact  $t_i$  is close enough to the ground-truth  $\\hat{t}_i$ , i.e.  $|t_i - \\hat{t}_i| < T_\\delta, T_\\delta = 0.25$ . The following metrics are considered:",
+        "bbox": [
+            496,
+            718,
+            890,
+            853
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "- Noun refers to the Noun-Box Top-5 mAP. In addition to the IoU constraint, a correct match with the ground-truth noun is required.",
+        "bbox": [
+            517,
+            854,
+            890,
+            900
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "18290",
+        "bbox": [
+            480,
+            944,
+            519,
+            955
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/3d5187d91dda355a3a175239bb85fe7d7ec5cf23bc0fd5522819e758861d34ff.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Ver.</td><td>Model</td><td>N↑</td><td>N-V ↑</td><td>N-T ↑</td><td>A↑</td></tr><tr><td rowspan=\"5\">v1</td><td>FRCNN+Rnd. [22]</td><td>20.45</td><td>2.22</td><td>3.86</td><td>0.44</td></tr><tr><td>FRCNN+SF [22]</td><td>20.45</td><td>6.78</td><td>6.17</td><td>2.45</td></tr><tr><td>InternVideo [7]</td><td>24.60</td><td>9.18</td><td>7.64</td><td>3.40</td></tr><tr><td>StillFast [40]</td><td>19.51</td><td>9.95</td><td>6.45</td><td>3.49</td></tr><tr><td>TF (ours)</td><td>24.90</td><td>10.06</td><td>7.57</td><td>3.54</td></tr><tr><td rowspan=\"4\">v2</td><td>FRCNN+SF [22]</td><td>26.15</td><td>9.45</td><td>8.69</td><td>3.61</td></tr><tr><td>StillFast [40]</td><td>25.06</td><td>13.29</td><td>9.14</td><td>5.12</td></tr><tr><td>GANov2 [46]</td><td>25.67</td><td>13.60</td><td>9.02</td><td>5.16</td></tr><tr><td>TF (ours)</td><td>30.43</td><td>13.45</td><td>10.38</td><td>5.18</td></tr></table>",
+        "bbox": [
+            78,
+            88,
+            470,
+            232
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1. Comparison with the state-of-the-art. Our method TransFusion (TF) outperforms all state-of-the-art approaches on 3 out of 4 metrics, on both versions of the Ego4D test set. We report the top-5 mean average precision for the Noun (N), NounVerb (N-V), Noun-TTC (N-T), and Noun-Verb-TTC (A) metrics.",
+        "bbox": [
+            75,
+            239,
+            468,
+            309
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Noun-Verb refers to the Noun-Verb-Box Top-5 mAP. Correct noun and verb match with the IoU constraint.",
+            "- Noun-TTC refers to the Noun-Box Top-5 mAP. Correct noun matches with the TTC constraint.",
+            "- Noun-Verb-TTC (\"Overall\") refers to the overall Top-5 mAP, intersecting correct noun-verb matches with IoU and TTC constraints."
+        ],
+        "bbox": [
+            94,
+            327,
+            467,
+            435
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Note that since detecting the next-active objects is the main focus of the task, all evaluation metrics are conditioned on the IoU constraint.",
+        "bbox": [
+            75,
+            439,
+            467,
+            483
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation details. We make use of the pre-trained Torchvision [34] Faster R-CNN. ResNet-50 [23] is used as the visual encoder and Sentence-BERT (SBERT) [41] is used as the language encoder. We freeze ResNet-50 and apply the multimodal fusion scheme on top. We use the RAdam [31] optimizer with a weight decay of  $2e^{-4}$ , and an effective batch size of 32. Besides the default Faster R-CNN background class in nouns, we also add it in the verb class, which improves the Noun-Verb mAP by almost 2 points. We reduce the region proposal network's sampling batch size and the detection network's image batch size and use data augmentation techniques to reduce overfitting. The number of verb-noun pairs used in the input is defined as the context length  $L_{c}$ . On average, an action segment described by one verb-noun phrase corresponds to one second with 30 frames, as computed from the GT annotations. Unless specified otherwise, we use  $L_{c} = 3$ . We perform postprocessing involving non-maximum suppression on the predicted bounding boxes, as elaborated in the Supp. Mat.",
+        "bbox": [
+            75,
+            486,
+            468,
+            773
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1. Comparison with state-of-the-art",
+        "text_level": 1,
+        "bbox": [
+            76,
+            785,
+            366,
+            801
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We first compare our method against the baselines proposed in the Ego4D work [22]. The FRCNN + SF method employs a two-stage approach: a ResNet-based Faster R-CNN detector first produces bounding box regressions and noun predictions from the prediction frame, without using any video features. In the second stage, a SlowFast [16] 3D",
+        "bbox": [
+            75,
+            809,
+            467,
+            900
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/074104493356d0743a5070d5e10fd5fb5890bc65383919de0f728cf803c48a64.jpg",
+        "image_caption": [
+            "Figure 4. Prediction of Noun-Verb in relation to time-to-contact. Histogram of the time-to-contact labels in Ego4Dv1 is shown in blue bars. Performance measured in Noun-Verb is plotted as a function of time-to-contact. We compare our method to Ego4D's official FRCNN + SF [22] baseline."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            89,
+            885,
+            251
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "CNN video processing module performs region-of-interest (RoI) pooling [21] on the corresponding SlowFast 3D CNN video features. The pooled video features are then used by RoI heads to predict the verb and TTC values. The FR-CNN + Rnd. baseline, with results published on v1 only, drops the second stage of FRCNN + SF and uses randomly drawn verb and TTC predictions, following their respective distributions in the training set.",
+        "bbox": [
+            496,
+            332,
+            890,
+            453
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Another baseline on v1 is given by InternVideo [7], a method achieving strong results on the short-term object interaction anticipation task with a two-stage strategy similar to FRCNN+SF: bounding boxes and noun labels are first obtained using a DINO-based object detector [62], after which a VideoMAE-pretrained [49] ViT [14] model leverages space-time attention to predict the corresponding verb classes and TTC values.",
+        "bbox": [
+            496,
+            455,
+            890,
+            575
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We further compare TransFusion to the recent StillFast [40] approach on both dataset versions. StillFast uses a Faster-RCNN-based spatial stream operating on a high-resolution version of the input image, and an X3D-based [15] temporal stream operating on a low-resolution version of the input video. The two streams each produce a multiscale feature map. Both feature maps are fused into a combined feature pyramid, from which local RoI features and global scene features are extracted, merged by a fusion network, and finally used by linear layers to predict each boxes' noun, verb and TTC.",
+        "bbox": [
+            496,
+            578,
+            890,
+            743
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "GANOV2 [46] enhances StillFast by adding a multi-head guided attention mechanism to its temporal stream, achieving favorable performance on version 2 of the dataset. The mechanism attends to object detections produced from the network's temporal branch, together with patches from the 3-D temporal feature stack, in a cross-fusion manner, and uses the fused tokens as the input to StillFast's feature pyramid and subsequent architecture.",
+        "bbox": [
+            496,
+            747,
+            890,
+            867
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1 demonstrates that TransFusion outperforms all state-of-the-art methods in terms of the N, N-V and A met-",
+        "bbox": [
+            500,
+            869,
+            890,
+            898
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "18291",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/edbb1320c1404778505b6b5f81388d0d766cb2538b2dca253dc38b7479457ed9.jpg",
+        "image_caption": [
+            "Figure 5. Classification performance on top/tail categories. We show the relative and absolute gains of TransFusion over Ego4D's FRCNN + SF baseline [22] on the validation set, for Noun-Only and Verb-Only mAP (without any box IoU constraint). Relative improvements are written on top of the red bars."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            112,
+            80,
+            439,
+            212
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "rics on version 1 of the dataset, and approaches the state-of-the art performance on the N-T metric while requiring a fraction of the competitor's resources during training [7] (see the Supp. Mat. for details). On the dataset's second version, TransFusion performs best in terms of N, N-T and A, while approaching the runner-up method on N-V. These favorable results validate the advantages of modeling action context explicitly for the interaction anticipation task.",
+        "bbox": [
+            75,
+            287,
+            468,
+            409
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We henceforth analyze our performance gains over the official Ego4D FRCNN + SF baseline on the first version of the Ego4D dataset. Figure 4 shows the long tail distribution of the time-to-contact labels in the validation set, split into buckets. We see that TransFusion consistently outperforms the baseline on the N-V metric for all TTC ranges. The results in Figure 5 indicate that our method not only outperforms the Ego4D method on top noun and verb classes, but also achieves bigger gains in the tail classes, demonstrating the generalization ability of using our language-based context summaries. Top and tail buckets represent roughly  $50\\%$  of the samples in the validation set.",
+        "bbox": [
+            75,
+            410,
+            468,
+            589
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Figure 6 shows that the Ego4D approach manages to detect the most salient objects in the environment, but fails to reliably associate the bounding boxes to the correct action phrase. In contrast, TransFusion is able to correctly associate the drawer bounding box to \"open drawer\" and to predict the correct action associated with the phone. More examples are provided in the Supp. Mat.",
+        "bbox": [
+            75,
+            592,
+            468,
+            698
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2. Ablation studies",
+        "text_level": 1,
+        "bbox": [
+            76,
+            709,
+            236,
+            724
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Evaluation of action context representation. In Table 2, we ablate different variants of context representation and compare against the representation that uses the groundtruth (GT) action labels provided by Ego4D in the action context, to assess the quality of our generated context summaries. The observed TTC Average Precision difference measures around  $2\\%$  and was thus omitted from the comparison. We report on the Noun and Noun-Verb metrics as we expect major differences in semantic-related tasks.",
+        "bbox": [
+            75,
+            733,
+            468,
+            868
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Experiments show that we achieve results approaching GT, demonstrating the effectiveness of our context repre",
+        "bbox": [
+            76,
+            869,
+            468,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2760f6325fb033de5848b12ab2a0db695cc01daa0951789615e11bb639c27c08.jpg",
+        "image_caption": [
+            "GT: open drawer 0.1s",
+            "(a) Ego4D: take container 0.78s, take container 0.46s, take pot 0.42s, take lid 0.46s"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            102,
+            687,
+            207
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f347836900e00440eff16f08c3b1a20da37d3039bebd62ce7fbfd2bcac128d1c.jpg",
+        "image_caption": [
+            "GT: take phone 1.0s",
+            "(b) Ego4D: take table 0.78s, take container 0.85s, take computer 0.86s, take phone 1.11s"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            700,
+            102,
+            880,
+            205
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f275f6477bf296d642ce09639b28fbbdb8f8c32b7a37ecd75be9cb706813e18f.jpg",
+        "image_caption": [
+            "(c) TF (ours): open drawer 0.56s, open drawer 0.36s, open drawer 0.44s, take lid 0.6s"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            506,
+            250,
+            684,
+            352
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/5e1888e52b63b20e82773356a1746d7e8c773d7b1012884f11f87efe7f330a24.jpg",
+        "image_caption": [
+            "(d) TF (ours): take phone 0.87s, open computer 0.63s, operate tablet 0.73s, take phone 0.98s",
+            "Figure 6. Qualitative results of the proposed approach. Results of Ego4D's FRCNN + SF baseline [22] are shown in the top row and TransFusion in the bottom row. Green denotes the ground truth bounding box, with the labels shown on top of each column."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            699,
+            250,
+            879,
+            353
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/a61616aac5c4d615da69ed141d1e3f176671e6c2f7d79f6a060e06ca8f2cfc20.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Rep.</td><td>∅</td><td>A</td><td>Nh</td><td>Ns</td><td>A+Nh</td><td>A+Ns</td><td>GT</td></tr><tr><td>N↑</td><td>17.71</td><td>20.09</td><td>19.33</td><td>19.75</td><td>19.03</td><td>20.19</td><td>21.71</td></tr><tr><td>N-V↑</td><td>6.14</td><td>7.16</td><td>6.86</td><td>7.05</td><td>7.12</td><td>7.55</td><td>7.86</td></tr></table>",
+        "bbox": [
+            504,
+            464,
+            883,
+            510
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2. Comparison of different action context representations. We evaluate different variations of context representations on the predicted noun (N) and noun-verb (N-V), and compare them against ground-truth annotations (GT) provided by v1 of the Ego4D validation set. GT only contains action descriptions of verb-noun pairs. Action context represented by  $\\mathcal{A} + \\mathcal{N}_s$  achieves the best N-V performance.",
+        "bbox": [
+            498,
+            513,
+            890,
+            611
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "sentations. Using a vision-only version of our model  $(\\emptyset)$  yields markedly worse results than when adding any type of language conditioning. Action descriptions  $\\mathcal{A}$  consistently boost the verb classification performance. Adding salient objects  $\\mathcal{N}_s$  to  $\\mathcal{A}$  further improves task performance. These together offer a reasonable picture of using language summaries to represent the action context. In addition, we notice a domain gap in the choice of words between the generated labels and the ground-truth ones. In some cases, even if the description is satisfactory for a human evaluator, the information may be too vague to be directly useful to the model. Phrasing differences between the generated captions and the training corpus of the language encoders may further help to improve the performance.",
+        "bbox": [
+            496,
+            613,
+            890,
+            824
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As an additional type of action context, we hypothesize that the sequence of objects held by the actor may be indicative of the overall task and can help the model better infer the immediate next steps. Motivated by this, we use an existing hand-object interaction predictor [11] jointly with an",
+        "bbox": [
+            496,
+            825,
+            890,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "18292",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/940a5f7209592e370d384e48631731a1fa40247041117243d0139f6e40d41cee.jpg",
+        "image_caption": [
+            "Figure 7. Comparison to video features. TransFusion outperforms TransFusion-Video over different context lengths measured by Noun-Verb mAP. TransFusion using GT-based context summaries as language input represents an upper performance bound."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            84,
+            411,
+            219
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "object detector [65] to look for held objects and their semantic labels in each frame. Usually, such objects occur only in a subset of frames (e.g. due to limited visibility of the actor's hands), which results in a lower performance of the resulting  $\\mathcal{N}_h$  and  $\\mathcal{A} + \\mathcal{N}_h$  representations. We hence use  $\\mathcal{A} + \\mathcal{N}_s$  to represent the action context, considering the good performance achieved. Ablations on the hyperparameters  $d$ , the cutoff distance, and  $k$ , the number of salient objects per frame, show that the performance is less sensitive to their settings. More details can be found in the Supp. Mat.",
+        "bbox": [
+            75,
+            287,
+            468,
+            439
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Comparison to video features. We evaluate whether language summaries can better represent action context than video features for the object-anticipation task. To do so, we implement TransFusion-Video, which takes video features extracted from the Ego4D SlowFast [16] model as input and uses an identical fusion module as in TransFusion.",
+        "bbox": [
+            75,
+            441,
+            468,
+            531
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 7 shows that TransFusion consistently outperforms TransFusion-Video.  $L_{c} \\times 30$  frames are taken in TransFusion-Video. For video features, the domain gap between pre-training and target datasets seems to be too large to allow effective generalization to a diverse dataset such as Ego4D. We suspect that the feature representations are more entangled due to the need to represent temporal, visual, and semantic aspects simultaneously. With language, temporal information is directly encoded in the syntax, i.e. the succession of words. Given a video clip of one second, language can summarize it in two words whereas corresponding video features use more than 12 times the space (e.g. smaller SBERT [41] features of  $2 \\times 384$  compared to SlowFast [16] features of  $4 \\times 2304$ ). Therefore, our language-based context representation has an advantage when it comes to longer video sequences. Overall, TransFusion obtains a  $21\\%$  boost in N-V performance. See the Supp. Mat. for details.",
+        "bbox": [
+            75,
+            535,
+            468,
+            806
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Length of action context. A significant parameter in obtaining good performance is the length of action context, i.e., how many past actions to include in the input. An adequate context length is important to allow the model to pick up more specific activity patterns. Some activities that have a more structured nature or certain repeating patterns (e.g.",
+        "bbox": [
+            75,
+            810,
+            468,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/a85093ff60befe94cb1db30be749e073ca104ce1fc6217a43d1fa2f1130037a9.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Language Encoder</td><td>Token size</td><td>N↑</td><td>N-V↑</td></tr><tr><td>SBERT/BERT [13, 41]</td><td>384</td><td>20.19</td><td>7.55</td></tr><tr><td>SBERT/RoBERTa [33, 41]</td><td>768</td><td>19.78</td><td>7.78</td></tr><tr><td>GPT-2 [37]</td><td>768</td><td>18.98</td><td>7.08</td></tr><tr><td>Flan-T5 [8]</td><td>1024</td><td>18.26</td><td>6.61</td></tr></table>",
+        "bbox": [
+            509,
+            88,
+            883,
+            161
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3. Language encoder ablations. We compare SBERT using BERT and RoBERTa backbones, with GPT-2 and Flan-T5.",
+        "bbox": [
+            498,
+            172,
+            890,
+            200
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "cooking or repairing a bicycle) are easier to model with an increased context length compared to more random activities such as playing basketball. Figure 7 shows that the overall performance plateaus at  $L_{c} = 2$  and higher.",
+        "bbox": [
+            496,
+            210,
+            890,
+            270
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Language encoder. A language encoder is used to process the context summaries in TransFusion. To understand the impact of different language encoders, we experiment with multiple LLMs including SBERT with a RoBERTa backbone [33, 41], GPT-2 [37], and FLAN-T5 [8], using version 1 of the validation set. Table 3 summarizes the performance of different language encoders, showing that the SBERT models perform best. We consider TTC less relevant here. We hypothesize that the difference comes from the training objectives of language models. SBERT is trained to map semantically similar sentences closer to each other in the embedding space. This is particularly useful when the model needs to understand descriptions with semantically close words, e.g. \"cut carrots\" and \"slice carrots\". GPT-2 and FLAN-T5 are optimized using generative objectives, which makes them overly sensitive to the input word choice.",
+        "bbox": [
+            496,
+            273,
+            892,
+            515
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Fusion module. We provide ablations regarding the fusion module in the Supp. Mat.",
+        "bbox": [
+            498,
+            517,
+            890,
+            547
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5. Conclusion and discussion",
+        "text_level": 1,
+        "bbox": [
+            500,
+            568,
+            743,
+            584
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We propose TransFusion, a multimodal fusion model that anticipates object interactions by considering the past action context represented by language summaries. Our successful experiments on a challenging egocentric video dataset demonstrate how language summaries and the vast knowledge contained in vision-language foundation models can improve object interaction anticipation, highlighting the representational power of language. It appears promising to investigate fusing even more modalities, such as motion cues from optical flow. We emphasize that the proposed summarization approach is not limited to this task. Language provides a universal interface to complement the visual input with information encoded in foundation models or task-specific pipelines. Future work can leverage language summarization for other video reasoning tasks [60, 63], or extend it temporally to describe possible future activities [1, 27] for long-term interaction anticipation.",
+        "bbox": [
+            496,
+            595,
+            890,
+            852
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements. This work was supported by an ETH Zurich Postdoctoral Fellowship. We thank Adrian Spurr for the helpful comments and Velko Vechev for the voiceover.",
+        "bbox": [
+            498,
+            854,
+            890,
+            900
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "18293",
+        "bbox": [
+            480,
+            944,
+            517,
+            955
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            78,
+            90,
+            173,
+            104
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[1] Chaitanya Ahuja and Louis-Philippe Morency. Language2Pose: Natural Language Grounded Pose Forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719–728, 2019. 8",
+            "[2] Alan Akbik, Duncan Blythe, and Roland Vollgraf. Contextual String Embeddings for Sequence Labeling. In Proceedings of the 27th international conference on computational linguistics, pages 1638-1649, 2018. 4",
+            "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual Question Answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425-2433, 2015. 3",
+            "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. CoRR, abs/2104.00650, 2021. 3",
+            "[5] Gedas Bertasius, Hyun Soo Park, Stella X. Yu, and Jianbo Shi. First-Person Action-Object Detection with EgoNet. In Proceedings of Robotics: Science and Systems, 2017. 2",
+            "[6] Jerome S Bruner. Intention in the Structure of Action and Interaction. Advances in Infancy Research, 1981. 1",
+            "[7] Guo Chen et al. InternVideo-Ego4D: A Pack of Champion Solutions to Ego4D Challenges. arXiv preprint arXiv:2211.09529, 2022.2,6,7",
+            "[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling Instruction-Finetuned Language Models, 2022. 3, 8",
+            "[9] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling Egocentric Vision. CoRR, abs/2006.13256, 2020. 2, 3",
+            "[10] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-Kitchens-100. International Journal of Computer Vision, 130(1):33-55, 2022. 4",
+            "[11] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. EPIC-KITCHENS VISOR Benchmark: VVideo Segmentations and Object Relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2022. 7",
+            "[12] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. TransVG: End-to-End Visual Grounding with Transformers. CoRR, abs/2104.08541, 2021.3",
+            "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional"
+        ],
+        "bbox": [
+            78,
+            114,
+            470,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Transformers for Language Understanding. CoRR, abs/1810.04805, 2018.3,8",
+            "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. CoRR, abs/2010.11929, 2020. 2, 3, 6",
+            "[15] Christoph Feichtenhofer. X3D: Expanding Architectures for Efficient Video Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 203-213, 2020. 6",
+            "[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition. CoRR, abs/1812.03982, 2018. 6, 8",
+            "[17] Antonino Furnari and Giovanni Maria Farinella. Rolling-Unrolling LSTMs for Action Anticipation from First-Person Video. CoRR, abs/2005.02190, 2020. 3",
+            "[18] Antonino Furnari, Sebastiano Battiato, Kristen Grauman, and Giovanni Maria Farinella. Next-Active-Object Prediction from Egocentric Videos. CoRR, abs/1904.05250, 2019. 1, 2",
+            "[19] H. Girase, N. Agarwal, C. Choi, and K. Mangalam. Latency Matters: Real-Time Action Forecasting Transformer. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18759-18769, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3",
+            "[20] Rohit Girdhar and Kristen Grauman. Anticipative Video Transformer. In ICCV, 2021. 3",
+            "[21] Ross B. Girshick. Fast R-CNN. CoRR, abs/1504.08083, 2015.6",
+            "[22] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Wesley Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4D: Around the World in 3,000 Hours of Egocentric Video. CoRR, abs/2110.07058, 2021. 1, 2, 4, 5, 6, 7"
+        ],
+        "bbox": [
+            501,
+            92,
+            893,
+            900
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "18294",
+        "bbox": [
+            480,
+            945,
+            519,
+            955
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. CoRR, abs/1512.03385, 2015. 5, 6",
+            "[24] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-BERT: Aligning Image Pixels with Text by Deep Multi-Modal Transformers. CoRR, abs/2004.00849, 2020.3",
+            "[25] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. EPIC-Fusion: Audio-Visual Temporal Binding for Egocentric Action Recognition. CoRR, abs/1908.08498, 2019. 3",
+            "[26] Weicheng Kuo, Fred Bertsch, Wei Li, AJ Piergiovanni, Mohammad Saffar, and Anelia Angelova. FindIt: Generalized Localization with Natural Language Queries. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 502-520. Springer, 2022. 3",
+            "[27] Yen-Ling Kuo, Xin Huang, Andrei Barbu, Stephen G McGill, Boris Katz, John J Leonard, and Guy Rosman. Trajectory Prediction with Linguistic Representations. In 2022 International Conference on Robotics and Automation (ICRA), pages 2868-2875. IEEE, 2022. 8",
+            "[28] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is More: ClipBERT for Video-and-Language Learning via Sparse Sampling. CoRR, abs/2102.06183, 2021. 3",
+            "[29] Tsung-Yi Lin, Piotr Dólar, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature Pyramid Networks for Object Detection. CoRR, abs/1612.03144, 2016. 5",
+            "[30] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen Clip Models Are Efficient Video Learners. In European Conference on Computer Vision, pages 388-404. Springer, 2022. 4",
+            "[31] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the Variance of the Adaptive Learning Rate and Beyond. CoRR, abs/1908.03265, 2019. 6",
+            "[32] Miao Liu, Siyu Tang, Yin Li, and James M. Rehg. Forecasting Human Object Interaction: Joint Prediction of Motor Attention and Egocentric Activity. CoRR, abs/1911.10967, 2019. 2",
+            "[33] 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. CoRR, abs/1907.11692, 2019.8",
+            "[34] TorchVision maintainers and contributors. TorchVision: PyTorch's Computer Vision Library. https://github.com/pytorch/vision, 2016.6",
+            "[35] David W Orr. The Nature of Design: Ecology, Culture, and Human Intention. Oxford University Press, 2002. 1",
+            "[36] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. EGOVLPV2: Egocentric Video-Language Pre-training With Fusion in the Backbone, 2023. 3"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[37] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language Models are Unsupervised Multitask Learners. OpenAI Blog, 2019. 3, 8",
+            "[38] 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. CoRR, abs/2103.00020, 2021. 2, 3, 4",
+            "[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. CoRR, abs/1910.10683, 2019. 3",
+            "[40] Francesco Ragusa et al. StillFast: An End-to-End Approach for Short-Term Object Interaction Anticipation. arXiv preprint arXiv:2304.03959, 2023. 2, 6",
+            "[41] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. CoRR, abs/1908.10084, 2019. 3, 5, 6, 8",
+            "[42] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. CoRR, abs/1506.01497, 2015. 5",
+            "[43] Paschal Sheeran. Intention-Behavior Relations: A Conceptual and Empirical Review. European Review of Social Psychology, 12(1):1-36, 2002. 1",
+            "[44] Chen Sun, Fabien Baradel, Kevin Murphy, and Cordelia Schmid. Contrastive Bidirectional Transformer for Temporal Representation Learning. CoRR, abs/1906.05743, 2019. 3",
+            "[45] Hao Tan and Mohit Bansal. LXMERT: Learning Cross-Modality Encoder Representations from Transformers. CoRR, abs/1908.07490, 2019. 3",
+            "[46] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Enhancing Next Active Object-Based Egocentric Action Anticipation With Guided Attention. In 2023 IEEE International Conference on Image Processing (ICIP), pages 1450-1454. IEEE, 2023. 2, 6",
+            "[47] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Guided Attention for Next Active Object @ EGO4D STA Challenge. arXiv preprint arXiv:2305.16066, 2023. 2",
+            "[48] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Leveraging Next-Active Objects for Context-Aware Anticipation in Egocentric Videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 8657–8666, 2024. 2",
+            "[49] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders Are Data-Efficient Learners for Self-Supervised Video Pre-Training. Advances in Neural Information Processing Systems, 35:10078–10093, 2022. 2, 6",
+            "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. CoRR, abs/1706.03762, 2017. 3"
+        ],
+        "bbox": [
+            501,
+            92,
+            890,
+            898
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "18295",
+        "bbox": [
+            480,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "[51] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-Anticipation Transformer for Online Action Understanding, 2023. 3",
+            "[52] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Action-CLIP: A New Paradigm for Video Action Recognition. arXiv preprint arXiv:2109.08472, 2021. 4",
+            "[53] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework. CoRR, abs/2202.03052, 2022. 2, 3, 4",
+            "[54] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MeMViT: Memory-Augmented Multiscale Vision Transformer for Efficient Long-Term Video Recognition. CoRR, abs/2201.08383, 2022. 3",
+            "[55] Yu Wu, Linchao Zhu, Xiaohan Wang, Yi Yang, and Fei Wu. Learning to Anticipate Egocentric Actions by Imagination. CoRR, abs/2101.04924, 2021. 3",
+            "[56] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, Attend and Tell: Neural Image Caption Generation With Visual Attention. In International Conference on Machine Learning, pages 2048-2057. PMLR, 2015. 3",
+            "[57] Ting Yao, Yingwei Pan, Yehao Li, Zhaofan Qiu, and Tao Mei. Boosting Image Captioning With Attributes. In Proceedings of the IEEE International Conference on Computer Vision, pages 4894-4902, 2017. 3",
+            "[58] Moon Ye-Bin, Jisoo Kim, Hongyeob Kim, Kilho Son, and Tae-Hyun Oh. TextMania: Enriching Visual Feature by Text-driven Manifold Augmentation, 2023. 3",
+            "[59] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. MERLOT: Multimodal Neural Script Knowledge Models. CoRR, abs/2106.02636, 2021. 3",
+            "[60] Kuo-Hao Zeng, Tseng-Hung Chen, Ching-Yao Chuang, Yuan-Hong Liao, Juan Carlos Niebles, and Min Sun. Leveraging Video Descriptions to Learn Video Question Answering. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 8",
+            "[61] Ce Zhang, Changcheng Fu, Shijie Wang, Nakul Agarwal, Kwonjoon Lee, Chiho Choi, and Chen Sun. Object-Centric Video Representation for Long-term Action Anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6751-6761, 2024. 3",
+            "[62] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection. In The Eleventh International Conference on Learning Representations, 2022. 2, 6",
+            "[63] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video Summarization With Long Short-Term Memory. In European Conference on Computer Vision, pages 766-782. Springer, 2016. 8",
+            "[64] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning Video Representations From Large Language Models, 2022. 3"
+        ],
+        "bbox": [
+            78,
+            90,
+            468,
+            898
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "ref_text",
+        "text": "[65] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple Multi-Dataset Detection. CoRR, abs/2102.13086, 2021. 8",
+        "bbox": [
+            501,
+            90,
+            890,
+            132
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "18296",
+        "bbox": [
+            480,
+            945,
+            517,
+            955
+        ],
+        "page_idx": 10
+    }
+]