ECCV 2024 Papers — Page 7
European Conference on Computer Vision · 2387 papers
DreamSampler: Unifying Diffusion Sampling and Score Distillation for Image Manipulation
Jeongsol Kim (KAIST), Jong Chul Ye (KAIST)
RestorationGenerationOptimizationDiffusion modelScore-based ModelAuto EncoderImageStochastic Differential Equation
🎯 What it does: Proposes DreamSampler, a unified framework that combines reverse diffusion sampling and score distillation for image editing, restoration, vectorization, and inverse problem solving.
DreamScene: 3D Gaussian-based Text-to-3D Scene Generation via Formation Pattern Sampling
Haoran Li (University of Science and Technology of China), Peng Yuan Zhou (Aarhus University)
GenerationLarge Language ModelDiffusion modelScore-based ModelGaussian SplattingTextPoint Cloud
🎯 What it does: Propose DreamScene, a text-driven 3D scene generation framework based on 3D Gaussians.
DreamScene360: Unconstrained Text-to-3D Scene Generation with Panoramic Gaussian Splatting
Shijie Zhou (University of California, Los Angeles), Achuta Kadambi (University of California, Los Angeles)
GenerationDepth EstimationTransformerLarge Language ModelVision Language ModelDiffusion modelGaussian SplattingImageTextMultimodalityPoint Cloud
🎯 What it does: Proposes an end-to-end pipeline called DreamScene360 for generating 360° panoramas from text and converting them into 3D scenes.
DreamStruct: Understanding Slides and User Interfaces via Synthetic Data Generation
Yi-Hao Peng (Carnegie Mellon University), Amy Pavel (University of Texas Austin)
ClassificationRecognitionGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a synthetic structured visual data generation method based on large language models (LLMs) to generate code and render visual data for creating annotated data to build slides and user interfaces.
DreamView: Injecting View-specific Text Guidance into Text-to-3D Generation
Junkai Yan (Sun Yat-sen University), WEI-SHI ZHENG
GenerationVision Language ModelDiffusion modelScore-based ModelImageText
🎯 What it does: DreamView achieves customizable and consistent text-to-3D generation by injecting view-specific text.
DriveDreamer: Towards Real-world-driven World Models for Autonomous Driving
Xiaofeng Wang (GigaAI), Jiwen Lu (GigaAI)
Autonomous DrivingRecurrent Neural NetworkTransformerVision-Language-Action ModelDiffusion modelWorld ModelVideoTextMultimodality
🎯 What it does: Built a world model called DriveDreamer based on real driving videos, which can generate controllable driving videos under structured traffic conditions (e.g., HDMap, 3D bounding boxes) and text prompts, and utilize action sequences to predict future videos and driving strategies.
DriveLM: Driving with Graph Visual Question Answering
Chonghao Sima (Shanghai AI Lab), Hongyang Li (Shanghai AI Lab)
Autonomous DrivingTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodalityGraphChain-of-Thought
🎯 What it does: Proposed the DriveLM task, i.e., end-to-end autonomous driving based on graph-structured visual question answering (GVQA), and constructed two large-scale graph question-answering datasets with logical dependencies: DriveLM-nuScenes and DriveLM-CARLA. Corresponding evaluation metrics were provided, and the DriveLM-Agent VLM baseline model was proposed based on this.
DrivingDiffusion: Layout-Guided Multi-View Driving Scenarios Video Generation with Latent Diffusion Model
Li Xiaofan (Baidu Inc), Ye Xiaoqing (Baidu Inc)
GenerationData SynthesisAutonomous DrivingDiffusion modelOptical FlowImageVideo
🎯 What it does: This study proposes DrivingDiffusion, a multi-perspective, temporally consistent video generation framework based on 3D layouts, capable of automatically generating high-quality, cross-perspective and cross-frame consistent driving scene videos from synthetic or real 3D traffic scene layouts.
Dropout Mixture Low-Rank Adaptation for Visual Parameters-Efficient Fine-Tuning
Zhengyi Fang (Shanghai Jiao Tong University), Lizhuang Ma (Shanghai Jiao Tong University)
ClassificationComputational EfficiencySupervised Fine-TuningImage
🎯 What it does: This paper proposes a vision parameter-efficient fine-tuning framework called DMLoRA based on dynamic training structures, which enhances model robustness and performance by leveraging multi-branch low-rank adaptation and phased scale learning.
DSA: Discriminative Scatter Analysis for Early Smoke Segmentation
Lujian Yao (East China University of Science and Technology), Kaijie Zhao (East China University of Science and Technology)
SegmentationContrastive LearningImage
🎯 What it does: Propose discriminative divergence analysis (DSA) as the loss function during training for early smoke segmentation, enhancing the resolution of the feature embedding space.
DSMix: Distortion-Induced Saliency Map Based Pre-training for No-Reference Image Quality Assessment
Jinsong Shi (Nanjing University of Aeronautics and Astronautics), Jie Qin (Nanjing University of Aeronautics and Astronautics)
Knowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposed a self-supervised pre-training framework combined with Cut-Mix data augmentation based on Distortion Sensitivity Map (DSM), dynamically assigning mixed labels through DSM and introducing semantic features via knowledge distillation to achieve no-reference image quality assessment;
Dual-Camera Smooth Zoom on Mobile Phones
Renlong Wu (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
GenerationSupervised Fine-TuningGaussian SplattingImage
🎯 What it does: Proposed the dual-camera smooth zoom (DCSZ) task, and constructed a virtual camera 'data factory' to generate synthetic training data, thereby fine-tuning existing frame interpolation models to achieve smooth preview.
Dual-Decoupling Learning and Metric-Adaptive Thresholding for Semi-Supervised Multi-Label Learning
Jia-Hao Xiao (Nanjing University of Aeronautics and Astronautics), Sheng-Jun Huang (Nanjing University of Aeronautics and Astronautics)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a dual decoupled learning framework (D2L) and a metric adaptive threshold (MAT) strategy for semi-supervised multi-label learning to generate high-quality pseudo labels and enhance model performance.
Dual-level Adaptive Self-Labeling for Novel Class Discovery in Point Cloud Segmentation
Ruijie Xu (ShanghaiTech University), Xuming He (ShanghaiTech University)
SegmentationConvolutional Neural NetworkPoint Cloud
🎯 What it does: Propose a dual-layer adaptive self-annotation framework for new class discovery in point cloud semantic segmentation, addressing issues of class imbalance and missing spatial contextual information.
Dual-Path Adversarial Lifting for Domain Shift Correction in Online Test-time Adaptation
Yushun Tang (Southern University of Science and Technology), Zhihai He (Southern University of Science and Technology)
ClassificationDomain AdaptationTransformerGenerative Adversarial NetworkImage
🎯 What it does: Proposed a Transformer-based online test-time adaptation method called Dual-Path Adversarial Lifting (DPAL), which inserts a domain offset token before each Transformer layer and alternates between noise prediction (Prediction) and feature update (Update) to progressively remove domain shift, ultimately achieving fine-grained classification of target domain samples;
Dual-Rain: Video Rain Removal using Assertive and Gentle Teachers
Tingting Chen (National University Of Singapore), Robby T. Tan (National University Of Singapore)
RestorationDepth EstimationTransformerVideo
🎯 What it does: Propose the Dual-Rain video rain rendering removal method, which utilizes assertive and gentle teachers for self-supervised learning on unlabeled real rainy videos, and enhances de-raining performance by generating hard rain and easy rain samples through RainMix augmentation.
Dual-stage Hyperspectral Image Classification Model with Spectral Supertoken
Peifu Liu (Beijing Institute of Technology), Jianan Li (Beijing Institute of Technology)
ClassificationTransformerImage
🎯 What it does: Proposed the Dual-stage Spectral Supertoken Classifier (DSTC), which forms spectral supertokens through clustering of spectral derivative features and utilizes Transformers for classification to achieve pixel-level high-precision remote sensing image classification.
DualBEV: Unifying Dual View Transformation with Probabilistic Correspondences
Peidong Li (Zhijia Technology), Dixiao Cui (Zhijia Technology)
Object DetectionDepth EstimationAutonomous DrivingConvolutional Neural NetworkImagePoint Cloud
🎯 What it does: Propose the DualBEV framework, unifying 3D-2D and 2D-3D view transformations, achieving BEV feature extraction through dual-perspective probabilistic correspondence.
DualDn: Dual-domain Denoising via Differentiable ISP
Ruikang Li (Shanghai Artificial Intelligence Laboratory), Tianfan Xue (Chinese University of Hong Kong)
RestorationConvolutional Neural NetworkTransformerImage
🎯 What it does: Designed and trained the DualDn dual-domain denoising network, connecting raw domain denoising with sRGB domain denoising into an end-to-end trainable framework using differentiable ISP.
DVLO: Deep Visual-LiDAR Odometry with Local-to-Global Feature Fusion and Bi-Directional Structure Alignment
Jiuming Liu (Shanghai Jiao Tong University), Hesheng Wang (Shanghai Jiao Tong University)
Pose EstimationAutonomous DrivingSimultaneous Localization and MappingImageMultimodalityPoint CloudBenchmark
🎯 What it does: Propose a visual-LiDAR odometry network DVLO based on local-to-global fusion, achieving efficient and fine-grained multimodal feature fusion through bidirectional structural alignment.
DyFADet: Dynamic Feature Aggregation for Temporal Action Detection
Le Yang (Xi'an Jiaotong University), Fan Li (Xi'an Jiaotong University)
Object DetectionConvolutional Neural NetworkTransformerVideo
🎯 What it does: Proposed a temporal action detection framework DyFADet based on dynamic feature aggregation (DFA), addressing the issues of insufficient feature discriminability and poor compatibility of the detection head in traditional models.
Dyn-Adapter: Towards Disentangled Representation for Efficient Visual Recognition
Yurong Zhang (Shanghai Jiao Tong University), Li Song (Shanghai Jiao Tong University)
RecognitionComputational EfficiencyTransformerImageVideo
🎯 What it does: Proposed the Dyn-Adapter framework, leveraging dynamic early exit and bidirectional sparsification techniques, achieving improved inference efficiency and maintaining or enhancing accuracy while keeping pre-trained model parameters frozen in PETL.
Dynamic Data Selection for Efficient SSL via Coarse-to-Fine Refinement
Aditay Tripathi (Indian Institute of Science), Anirban Chakraborty (Indian Institute of Science)
Computational EfficiencyRepresentation LearningData-Centric LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposes DySCoF, a dynamic data subset selection framework that dynamically adjusts subset size and instance importance based on the model's learning progress during training to improve self-supervised learning (SSL) efficiency.
Dynamic Guidance Adversarial Distillation with Enhanced Teacher Knowledge
Hyejin Park (Ewha Womans University), Dongbo Min (Ewha Womans University)
ClassificationKnowledge DistillationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Propose a dynamic guided adversarial distillation (DGAD) framework that leverages three mechanisms: misclassification-aware partitioning (MAP), error-corrected label exchange (ELS), and prediction consistency regularization (PCR) to enhance the accuracy and robustness of student models on natural images and adversarial examples.
Dynamic Neural Radiance Field From Defocused Monocular Video
Xianrui Luo (Huazhong University of Science and Technology), Zhiguo Cao (Huazhong University of Science and Technology)
RestorationGenerationDepth EstimationNeural Radiance FieldOptical FlowVideo
🎯 What it does: Proposes DRF2, a dynamic neural radiance field model capable of recovering clear dynamic scenes from defocused monocular videos.
Dynamic Retraining-Updating Mean Teacher for Source-Free Object Detection
Trinh Le Ba Khanh (Sungkyunkwan University), Jae Wook Jeon (Sungkyunkwan University)
Object DetectionDomain AdaptationTransformerImage
🎯 What it does: This paper proposes a dynamic retraining-updating Mean Teacher framework for source-free unsupervised object detection (SFOD) tasks.
DynamiCrafter: Animating Open-domain Images with Video Diffusion Priors
Jinbo Xing (Chinese University of Hong Kong), Tien-Tsin Wong (Chinese University of Hong Kong)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelImageVideoText
🎯 What it does: Developed an open-domain image animation method called DynamiCrafter, achieving image-to-video animation in video diffusion models through a dual-stream image injection scheme.
DynMF: Neural Motion Factorization for Real-time Dynamic View Synthesis with 3D Gaussian Splatting
Angelos Kratimenos, Kostas Daniilidis (University of Pennsylvania)
GenerationData SynthesisGaussian SplattingImageVideo
🎯 What it does: This paper proposes DynMF, which achieves efficient and real-time dynamic view synthesis by decomposing the motion of dynamic scenes into a small number of learnable basis trajectories, and representing the motion of each point as a sparse linear combination of these trajectories.
DynoSurf: Neural Deformation-based Temporally Consistent Dynamic Surface Reconstruction
Yuxin Yao (City University of Hong Kong), Wenping Wang (Texas A&M University)
RestorationGraph Neural NetworkPoint CloudMesh
🎯 What it does: Proposes an unsupervised learning framework named DynoSurf, capable of reconstructing temporally consistent, topologically identical dynamic meshes from continuous dynamic point cloud sequences without relying on shape priors, ground truth surfaces, or temporal correspondences.
DySeT: a Dynamic Masked Self-distillation Approach for Robust Trajectory Prediction
Mozghan Pourkeshavarz (Huawei), Amir Rasouli (Huawei)
Autonomous DrivingKnowledge DistillationTransformerReinforcement LearningPoint CloudTime SeriesSequential
🎯 What it does: In the self-supervised pre-training framework DySeT, the paper proposes a dynamic masking strategy and self-distillation mechanism. It first selects trajectory and lane tokens with high information content that reflect complex driving behaviors through a sampling network, then distills the complete scene representation to the masked sub-model, thereby enhancing the quality and generalization ability of behavior representation learning.
DεpS: Delayed ε-Shrinking for Faster Once-For-All Training
Aditya Annavajjala (Georgia Institute of Technology), Alexey Tumanov (Cisco Research)
Computational EfficiencyKnowledge DistillationNeural Architecture SearchConvolutional Neural NetworkImage
🎯 What it does: The delayed ε-shrink method is proposed within the once-for-all framework, which first partially preheats the full model, then gradually introduces subnetworks, and efficiently trains with shared weights through ε-Shrinking learning rate scheduling and IKD-Warmup.
E.T. the Exceptional Trajectory: Text-to-camera-trajectory generation with character awareness
Robin Courant (Ecole Polytechnique), Vicky Kalogeiton (Ecole Polytechnique)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelVideoText
🎯 What it does: This paper proposes a new dataset called Exceptional Trajectories (E.T.) and trains a diffusion model named Director based on this dataset. The model generates movie camera trajectories in a global coordinate system from text descriptions and main character trajectories. Additionally, a contrastive language-trajectory embedding model named CLaTr is constructed to evaluate the generated trajectories.
E3M: Zero-Shot Spatio-Temporal Video Grounding with Expectation-Maximization Multimodal Modulation
Peijun Bao (Nanyang Technological University), Alex Kot
RetrievalVision Language ModelContrastive LearningVideoTextMultimodalityBenchmark
🎯 What it does: Propose a zero-shot spatiotemporal video localization method called E3M, which directly locates target objects in videos during testing using a pre-trained CLIP model.
E3V-K5: An Authentic Benchmark for Redefining Video-Based Energy Expenditure Estimation
Shengxuming Zhang (Zhejiang University), Mingli Song (Zhejiang University)
RecognitionPose EstimationTransformerVideoTabularTime SeriesBenchmark
🎯 What it does: This paper first constructs the E3V-K5 dataset of motion video energy expenditure based on COSMED K5 ground-truth measurements, and proposes the E3SFormer model, which simultaneously performs action recognition and energy regression using human skeletal videos to estimate energy expenditure at the video level.
EA-VTR: Event-Aware Video-Text Retrieval
Zongyang Ma (MAIS Institute of Automation Chinese Academy of Sciences), Weiming Hu (MAIS Institute of Automation Chinese Academy of Sciences)
Data SynthesisRetrievalTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Improving video-text retrieval by leveraging event content augmentation and event time augmentation to construct a richer dataset, and proposing a multi-granularity video encoding and event learning mechanism to enhance retrieval accuracy and event understanding capabilities.
EAFormer: Scene Text Segmentation with Edge-Aware Transformers
Haiyang Yu (Fudan University), Xiangyang Xue (Fudan University)
SegmentationTransformerImage
🎯 What it does: Propose the Edge-Aware Transformers (EAFormer) model, combining explicit text edge extraction with encoder cross-attention to improve edge precision in scene text segmentation.
EAGLES: Efficient Accelerated 3D Gaussians with Lightweight EncodingS
Sharath Girish (University of Maryland), Abhinav Shrivastava (University of Maryland)
Computational EfficiencyGaussian SplattingImageBenchmark
🎯 What it does: This paper proposes the EAGLES method, which significantly reduces storage and runtime memory, achieving faster training and rendering while maintaining high-quality reconstruction through attribute quantization of 3D Gaussian splines, progressive training, and impact pruning.
Early Anticipation of Driving Maneuvers
Abdul Wasi Lone (Indian Institute of Information Technology Hyderabad), C. V. Jawahar (Indian Institute of Information Technology Hyderabad)
Autonomous DrivingTransformerMultimodality
🎯 What it does: A new task, ADM, was created, and a multi-perspective, multi-modal DAAD dataset was collected for early prediction before driving actions begin.
Early Preparation Pays Off: New Classifier Pre-tuning for Class Incremental Semantic Segmentation
Zhengyuan Xie (Nankai University), Xialei Liu (UESTC)
ClassificationSegmentationRepresentation LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: Propose a new classifier pre-tuning method called NeST, which helps better initialize classifiers in new tasks for class-incremental semantic segmentation
EAS-SNN: End-to-End Adaptive Sampling and Representation for Event-based Detection with Recurrent Spiking Neural Networks
Ziming Wang (Zhejiang University), Huajin Tang (Zhejiang University)
Object DetectionRecurrent Neural NetworkSpiking Neural NetworkSequential
🎯 What it does: Propose a learnable adaptive event sampling module based on recursive convolutional SNN, integrating it with subsequent detection networks into an end-to-end trainable event detection framework; simultaneously design two techniques, Residual Potential Dropout (RPD) and SpikeAware Training (SAT), to enhance the synergy between sampling and the network.
Easing 3D Pattern Reasoning with Side-view Features for Semantic Scene Completion
Linxi Huan (Wuhan University), Xianwei Zheng (Wuhan University)
SegmentationConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: This paper proposes the SidePaint strategy, which addresses the challenge of 3D pattern reasoning in semantic scene completion by performing context filling on three 2D feature slices from side views.
EBDM: Exemplar-guided Image Translation with Brownian-bridge Diffusion Models
Eungbean Lee (Yonsei University), Kwanghoon Sohn (Yonsei University)
Image TranslationConvolutional Neural NetworkTransformerDiffusion modelImageOrdinary Differential Equation
🎯 What it does: Propose an example-guided image translation framework EBDM based on the Brownian bridge diffusion model, which can generate high-quality images with the style of the example under the condition of using only structural control (such as masks, edges, poses) and example images.
Echoes of the Past: Boosting Long-tail Recognition via Reflective Learning
Qihao Zhao (Beijing University of Chemical Technology), Jun Liu (Singapore University of Technology and Design)
ClassificationRecognitionImageBenchmark
🎯 What it does: Propose a reflection learning framework that improves long-tailed image recognition performance through three stages: review, summarize, and correct.
EchoScene: Indoor Scene Generation via Information Echo over Scene Graph Diffusion
Guangyao Zhai (Technical University of Munich), Benjamin Busam (Technical University of Munich)
GenerationGraph Neural NetworkDiffusion modelGraph
🎯 What it does: Propose a dual-branch diffusion model called EchoScene based on information echo, enabling the generation of controllable 3D indoor scenes from semantic scene graphs;
EcoMatcher: Efficient Clustering Oriented Matcher for Detector-free Image Matching
Peiqi Chen (Wuhan University), Ming Yang (Ant Group)
Pose EstimationConvolutional Neural NetworkTransformerImagePoint CloudBenchmark
🎯 What it does: This paper proposes a detector-free matcher called EcoMatcher based on Context Cluster, which achieves dense matching between image pairs without using feature point detectors.
EDformer: Transformer-Based Event Denoising Across Varied Noise Levels
Bin Jiang (Nanjing University), Zhan Ma (Nanjing University)
RestorationTransformerVideoSequential
🎯 What it does: This paper constructs the first real-world event denoising dataset, ED24, by collecting background activity noise from the DAVIS346 camera under different lighting conditions, and proposes an event-level denoising model called EDformer based on Transformer.
Edge-Guided Fusion and Motion Augmentation for Event-Image Stereo
Fengan Zhao (University of Science and Technology of China), Junlin Xiong (University of Science and Technology of China)
Depth EstimationRecurrent Neural NetworkImageMultimodality
🎯 What it does: Integrate stereo data from event cameras and optical cameras to achieve more accurate disparity estimation through edge-guided feature fusion and motion enhancement.
Editable Image Elements for Controllable Synthesis
Jiteng Mu (University of California San Diego), Taesung Park (Adobe Research)
GenerationData SynthesisConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderImage
🎯 What it does: Proposes an editable image elements representation, partitioning input images into manipulable regions, achieving high-fidelity reconstruction and spatial editing through convolutional encoders, Transformer lightweight decoders, and diffusion models.
EditShield: Protecting Unauthorized Image Editing by Instruction-guided Diffusion Models
Ruoxi Chen (Zhejiang University of Technology), Lichao Sun (Lehigh University)
GenerationSafty and PrivacyVision Language ModelDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: This paper proposes a protection method called EditShield, which uses non-targeted perturbations to disrupt the latent representations of instruction-driven diffusion models, making unauthorized image edits unable to produce results that comply with the instructions.
EDTalk: Efficient Disentanglement for Emotional Talking Head Synthesis
Shuai Tan (Shanghai Jiao Tong University), ye pan
GenerationFlow-based ModelAuto EncoderVideoMultimodalityAudio
🎯 What it does: Propose an efficient decoupled speaker avatar generation framework called EDTalk that can simultaneously control lip shape, head pose, and emotional expression, driven by video or audio.
Effective Lymph Nodes Detection in CT Scans Using Location Debiased Query Selection and Contrastive Query Representation in Transformer
Qinji Yu (Shanghai Jiao Tong University), Dakai Jin (DAMO Academy, Alibaba Group)
Object DetectionTransformerContrastive LearningBiomedical DataComputed Tomography
🎯 What it does: Proposed a Transformer-based lymph node detection framework, LN-DETR, combining positional bias query selection, contrastive query learning, and multi-scale 2.5D semantic fusion to achieve end-to-end 3D CT lymph node detection.
Efficient 3D-Aware Facial Image Editing via Attribute-Specific Prompt Learning
Amandeep Kumar (Mohamed bin Zayed University of Artificial Intelligence), Rao Muhammad Anwer (Mohamed bin Zayed University of Artificial Intelligence)
Image TranslationGenerationTransformerPrompt EngineeringVision Language ModelNeural Radiance FieldGenerative Adversarial NetworkContrastive LearningImageTextMesh
🎯 What it does: This paper proposes a 3D-Aware facial image editing framework based on attribute-specific prompt learning, which can control facial attributes (such as hair color, expression, etc.) from different perspectives through natural language prompts while maintaining perspective consistency and identity fidelity;
Efficient Active Domain Adaptation for Semantic Segmentation by Selecting Information-rich Superpixels
Yuan Gao (University of Science and Technology of China), Bohai Tu (University of Science and Technology of China)
SegmentationDomain AdaptationImageBenchmark
🎯 What it does: In the active domain adaptation task of semantic segmentation, an active learning framework is proposed that combines low uncertainty superpixel fusion with two-stage superpixel acquisition, achieving annotation with only 640 clicks per image.
Efficient and Versatile Robust Fine-Tuning of Zero-shot Models
Sungyeon Kim (POSTECH), Suha Kwak (POSTECH)
ClassificationSegmentationRetrievalRepresentation LearningSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Efficient and robust fine-tuning of large zero-shot image-text pre-trained models (e.g., CLIP) for multiple tasks including classification, cross-modal retrieval, and open-vocabulary segmentation.
Efficient Bias Mitigation Without Privileged Information
Mateo Espinosa Zarlenga (University of Cambridge), Alice Xiang (Sony AI)
ClassificationDomain AdaptationImage
🎯 What it does: Proposes a bias mitigation framework named TAB without group information, which partitions samples and generates group-balanced datasets by leveraging the complete training history of an auxiliary model, followed by retraining a robust model from this dataset.
Efficient Cascaded Multiscale Adaptive Network for Image Restoration
Yichen Zhou (National University of Singapore), Teck Khim Ng (National University of Singapore)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: Proposed an efficient cascaded multi-scale adaptive network (ECMA) for image restoration tasks such as deblurring, denoising, and super-resolution.
Efficient Depth-Guided Urban View Synthesis
sheng miao, Yiyi Liao (Zhejiang University)
Data SynthesisDepth EstimationAutonomous DrivingConvolutional Neural NetworkNeural Radiance FieldImagePoint Cloud
🎯 What it does: Propose EDUS, a generalizable depth-guided view synthesis framework for urban street scenes, which can achieve fast forward inference under sparse image inputs and support efficient per-scenario fine-tuning.
Efficient Diffusion Transformer with Step-wise Dynamic Attention Mediators
Yifan Pu (Tsinghua University), Xiu Li (Tsinghua University)
GenerationComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: Discover redundancy in query-key interactions within diffusion Transformers, proposing the use of mediator tokens to compress attention and dynamically adjust the number of mediator tokens based on denoising steps, achieving an efficient diffusion Transformer with linear complexity;
Efficient Diffusion-Driven Corruption Editor for Test-Time Adaptation
Yeongtak Oh (Seoul National University), Sungroh Yoon (Seoul National University)
RestorationDomain AdaptationComputational EfficiencyKnowledge DistillationDiffusion modelContrastive LearningImageVideo
🎯 What it does: This paper proposes Decorruptor, which fine-tunes the Latent Diffusion Model using an improved corruption modeling scheme, enabling the generation of clear images by editing corrupted images during testing and accelerating inference.
Efficient Few-Shot Action Recognition via Multi-Level Post-Reasoning
Cong Wu (Jiangnan University), Josef Kittler (University of Surrey)
RecognitionMeta LearningTransformerVideoMultimodality
🎯 What it does: By freezing the CLIP vision and text encoders, we achieve efficient few-shot action recognition through multi-level post-reasoning and interactive spatiotemporal reasoning.
Efficient Frequency-Domain Image Deraining with Contrastive Regularization
Ning Gao (Beihang University), Yue Deng (Beihang University)
RestorationTransformerContrastive LearningImage
🎯 What it does: This paper proposes FADformer, a frequency-domain based Transformer framework, which achieves efficient single-image deraining using a frequency-domain convolution mixer and prior-gated feed-forward networks.
Efficient Image Pre-Training with Siamese Cropped Masked Autoencoders
Alexandre Eymaël (University of Liège), Marc Van Droogenbroeck (University of Liège)
Representation LearningTransformerAuto EncoderImageVideo
🎯 What it does: Proposed the CropMAE method for self-supervised pre-training using only a single image.
Efficient Inference of Vision Instruction-Following Models with Elastic Cache
Zuyan Liu (Tsinghua University), Jiwen Lu (Tsinghua University)
Computational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: Propose the Elastic Cache method, which compresses and accelerates the KV cache during inference for large-scale vision instruction following models (e.g., LLaVA, Qwen-VL). By adopting importance-driven cache merging (anchor-point + bucket merging) during the instruction encoding phase and using fixed-point removal (retaining the initial and latest KV pairs) during the output generation phase, a training-agnostic multi-stage acceleration is achieved.
Efficient Learning of Event-based Dense Representation using Hierarchical Memories with Adaptive Update
Uday Kamal (Georgia Institute of Technology), Saibal Mukhopadhyay (Georgia Institute of Technology)
Object DetectionSegmentationDepth EstimationTransformerVideo
🎯 What it does: Propose a hierarchical memory-based and adaptively updatable dense representation learning framework for event cameras, which maps sparse events to multi-level latent memory spaces via cross-attention, and updates higher-level memory only when necessary to achieve event-level efficient inference;
Efficient NeRF Optimization - Not All Samples Remain Equally Hard
Juuso Korhonen (Nokia Technologies), Juho Kannala (Aalto University)
OptimizationComputational EfficiencyNeural Radiance FieldImage
🎯 What it does: This paper proposes an online hard sample mining method to accelerate NeRF (Neural Radiance Field) training and significantly reduce memory consumption.
Efficient Neural Video Representation with Temporally Coherent Modulation
Seungjun Shin (Samsung Advanced Institute of Technology), Dokwan Oh (Samsung Advanced Institute of Technology)
RestorationSuper ResolutionCompressionFlow-based ModelOptical FlowVideo
🎯 What it does: Propose a neural video representation framework called NVTM that leverages time-coherent modulation to efficiently learn video content and support tasks such as compression, super-resolution, and frame interpolation.
Efficient Pre-training for Localized Instruction Generation of Procedural Videos
Anil Batra (University of Edinburgh), Frank Keller (University of Edinburgh)
GenerationRetrievalTransformerContrastive LearningVideoTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes an automatic method for filtering and replacing ASR transcripts with human-written instructions (Sieve & Swap), and implements step localization and text generation for cooking videos using a pre-trained Procedure Transformer (ProcX).
Efficient Snapshot Spectral Imaging: Calibration-Free Parallel Structure with Aperture Diffraction Fusion
Tao Lv (Nanjing University), Xun Cao (Nanjing University)
RestorationTransformerImagePhysics Related
🎯 What it does: Designed and implemented an uncalibrated parallel snapshot spectral imaging system called PCCADIS, combining aperture diffraction with RGB prior fusion, and proposed an adaptive feature encoding cascaded Transformer (SSFCT) for fusion and reconstruction.
Efficient Training of Spiking Neural Networks with Multi-Parallel Implicit Stream Architecture
Zhigao Cao (Xi'an Jiaotong University), Zigang Huang
Computational EfficiencySpiking Neural NetworkImage
🎯 What it does: Proposes an implicit training method for Spiking Neural Networks (SNN) based on a multi-parallel implicit flow (MPIS) architecture, achieving fast convergence while maintaining low latency, low memory consumption, and low sparsity.
Efficient Training with Denoised Neural Weights
Yifan Gong (Snap Inc.), Jian Ren (Snap Inc.)
Image TranslationTransformerVision Language ModelDiffusion modelGenerative Adversarial NetworkImageText
🎯 What it does: Built a weight generator based on a diffusion model to predict initial weights for image-to-image translation GAN models, enabling one-time fast inference followed by fine-tuning.
Efficient Unsupervised Visual Representation Learning with Explicit Cluster Balancing
Ioannis Maniadis Metaxas (Queen Mary University of London), Ioannis Patras (Queen Mary University of London)
Object DetectionSegmentationRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Proposes ExCB, a framework that employs online cluster balancing in self-supervised visual representation learning.
Efficient Vision Transformers with Partial Attention
Xuan-Thuy Vo (University of Ulsan), Kang-Hyun Jo (University of Ulsan)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: Proposed a Partial Attention mechanism, which divides image tokens into two groups: foreground (important) and background (unimportant). The model efficiently captures global and local information by using Hybrid Multi-Head Self-Attention (MMSA) and Single-Query Attention (SQA) respectively, and builds an efficient and scalable visual Transformer model called PartialFormer.
EGIC: Enhanced Low-Bit-Rate Generative Image Compression Guided by Semantic Segmentation
Nikolai Körber (Technical University of Munich), Björn Schuller (Technical University of Munich)
SegmentationCompressionConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a low-bitrate generative image compression method called EGIC, which can achieve smooth transitions on the distortion-perception curve using a single model.
EgoBody3M: Egocentric Body Tracking on a VR Headset using a Diverse Dataset
Amy Zhao (Meta Platforms Inc), Robert Y. Wang (Meta Platforms Inc)
Pose EstimationConvolutional Neural NetworkRecurrent Neural NetworkVideoBenchmark
🎯 What it does: Proposed a controller-free full-body pose tracking scheme utilizing four built-in cameras in a VR headset and released the first large-scale real-world dataset EgoBody3M.
EgoCVR: An Egocentric Benchmark for Fine-Grained Composed Video Retrieval
Thomas Hummel (University of Tübingen), Zeynep Akata (TU Munich)
RetrievalLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes a fine-grained first-person perspective video retrieval benchmark called EgoCVR, evaluates multiple vision-language models on this benchmark, and designs a training-agnostic re-ranking framework named TFR-CVR.
EgoExo-Fitness: Towards Egocentric and Exocentric Full-Body Action Understanding
Yuan-Ming Li (Sun Yat-sen University), Wei-Shi Zheng
ClassificationRecognitionConvolutional Neural NetworkTransformerVision-Language-Action ModelVideoTextBenchmark
🎯 What it does: This paper introduces the EgoExo-Fitness dataset, which collects synchronized egocentric (front and downward views) and exocentric (front, left-front, right-front) full-body fitness action videos. It provides rich annotations, including two-level time boundaries, technical keypoint validation, natural language comments, and action quality scores. Based on this dataset, five benchmark tasks are constructed: action classification, action localization, cross-perspective sequence verification, cross-perspective skill judgment, and a novel instruction execution verification.
EgoLifter: Open-world 3D Segmentation for Egocentric Perception
Qiao Gu (University of Toronto), Chris Sweeney (Meta Reality Labs)
SegmentationTransformerContrastive LearningGaussian SplattingVideoBenchmark
🎯 What it does: Designed and implemented the EgoLifter system, which can automatically perform 3D scene reconstruction and open-world (without fixed semantic categories) 3D instance segmentation from egocentric videos.
EgoPet: Egomotion and Interaction Data from an Animal's Perspective
Amir Bar (UC Berkeley), Trevor Darrell
Robotic IntelligenceSupervised Fine-TuningOptical FlowVideoBenchmark
🎯 What it does: This paper constructs an 84-hour animal first-person perspective dataset named EgoPet, containing 6,646 video segments, and proposes three benchmark tasks: Visual Interaction Prediction (VIP), Motion Trajectory Prediction (LP), and Visual-to-Self-Perception Prediction (VPP).
EgoPoseFormer: A Simple Baseline for Stereo Egocentric 3D Human Pose Estimation
Chenhongyi Yang (University of Edinburgh), Cem Keskin (Meta Reality Labs)
Pose EstimationTransformerAuto EncoderImage
🎯 What it does: Proposed a two-stage Transformer architecture, EgoPoseFormer, for egocentric 3D human pose estimation under stereo perspectives.
EgoPoser: Robust Real-Time Egocentric Pose Estimation from Sparse and Intermittent Observations Everywhere
Jiaxi Jiang (ETH Zürich), Christian Holz (ETH Zürich)
Pose EstimationComputational EfficiencyTransformerPoint CloudMeshTime Series
🎯 What it does: To address the sparse and discontinuous head and hand position information provided by head-mounted devices, this paper proposes a real-time full-body pose estimation method called EgoPoser, which can maintain high-accuracy and coherent pose outputs even when hands are out of the field of view.
EINet: Point Cloud Completion via Extrapolation and Interpolation
Pingping Cai (University of South Carolina), Song Wang (University of South Carolina)
RestorationTransformerPoint Cloud
🎯 What it does: Propose a new point cloud completion framework called EINet, which uses extrapolation in the feature space to complete missing shapes and interpolation in the feature space to upsample point clouds.
Elegantly Written: Disentangling Writer and Character Styles for Enhancing Online Chinese Handwriting
Yu Liu (University Putra Malaysia), Cunrui Wang (Dalian Minzu University)
GenerationTransformerSequential
🎯 What it does: This paper proposes an online Chinese handwriting trajectory beautification method based on a sequence Transformer, which can learn writing styles from a small number of user samples and optimize and beautify the handwriting trajectories while preserving the original text content.
Elevating All Zero-Shot Sketch-Based Image Retrieval Through Multimodal Prompt Learning
Mainak Singha (Indian Institute of Technology Bombay), Biplab Banerjee (Indian Institute of Technology Bombay)
RetrievalTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper proposes a bidirectional multimodal Prompt learning framework called SpLIP for zero-shot and fine-grained sketch image retrieval (ZS-SBIR, GZS-SBIR, FG-ZS-SBIR).
Eliminating Feature Ambiguity for Few-Shot Segmentation
Qianxiong Xu (Nanyang Technological University), Rui Zhao (SenseTime Research)
SegmentationMeta LearningTransformerContrastive LearningImage
🎯 What it does: Designed and proposed a pluggable disambiguation network, AENet, to eliminate feature ambiguity in few-shot segmentation, enhancing the matching quality between query and support foreground features, thereby significantly improving the segmentation performance of multiple baseline models.
Eliminating Warping Shakes for Unsupervised Online Video Stitching
Lang Nie (Beijing Jiaotong University), Yao Zhao (Beijing Jiaotong University)
OptimizationComputational EfficiencyConvolutional Neural NetworkVideoBenchmark
🎯 What it does: Proposes an online unsupervised video stitching and stabilization framework named StabStitch, specifically addressing the 'warping shake' problem in video stitching;
ELSE: Efficient Deep Neural Network Inference through Line-based Sparsity Exploration
Zeqi Zhu (Snap lnc.), Orlando Moreira (Snap lnc.)
Object DetectionPose EstimationComputational EfficiencyVideo
🎯 What it does: Proposes a row-based sparsity exploration method called ELSE to suppress events during deep neural network inference, thereby reducing computational load.
Elucidating the Hierarchical Nature of Behavior with Masked Autoencoders
Lucas Stoffl (Ecole Polytechnique Federale de Lausanne), Alexander Mathis (Ecole Polytechnique Federale de Lausanne)
Explainability and InterpretabilityRepresentation LearningTransformerAuto EncoderVideoTime SeriesBenchmark
🎯 What it does: This paper proposes a hierarchical masked autoencoder called hBehaveMAE, which is validated on a newly constructed synthetic basketball dataset Shot7M2 and the extended human action benchmark hBABEL.
Elysium: Exploring Object-level Perception in Videos through Semantic Integration Using MLLMs
Han Wang (Bytedance Inc), Can Huang (Bytedance Inc)
Object TrackingTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodality
🎯 What it does: Constructed and released a large-scale video object perception dataset ElysiumTrack-1M (1.27 million trajectories + descriptions), and proposed an end-to-end multimodal large language model Elysium that can directly complete video-level and object-level tasks; visual token compression is achieved through T-Selector, balancing frame rate and performance.
Embedding-Free Transformer with Inference Spatial Reduction for Efficient Semantic Segmentation
Hyunwoo Yu (Sogang University), Suk-Ju Kang (Sogang University)
SegmentationComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposed a Transformer-based Encoder-Decoder structure called EDAFormer, combining Embedding-Free Attention and Inference Spatial Reduction to achieve efficient semantic segmentation.
Embodied Understanding of Driving Scenarios
Yunsong Zhou (OpenDriveLab at Shanghai AI Lab), Hongyang Li (Shanghai Jiao Tong University)
Autonomous DrivingTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVision-Language-Action ModelImageTextMultimodality
🎯 What it does: This paper proposes the Embodied Language Model (ELM), aiming to achieve a four-dimensional, full-space, long-term temporal embodied understanding of driving scenarios.
Embracing Events and Frames with Hierarchical Feature Refinement Network for Object Detection
Hu Cao (Technical University of Munich), Alois C. Knoll
Object DetectionConvolutional Neural NetworkMultimodality
🎯 What it does: Propose a Hierarchical Feature Refinement Network (HFRN) and a Cross-Modal Adaptive Feature Refinement (CAFR) module to integrate features from event cameras and traditional frame cameras, improving target detection performance in complex environments.
Emergent Visual-Semantic Hierarchies in Image-Text Representations
Morris Alper (Tel Aviv University), Hadar Averbuch-Elor (Tel Aviv University)
Representation LearningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This study explores the understanding of visual semantic hierarchies in image-text representations by existing vision-language models (VLMs), proposes a radial embedding (RE) framework to optimize such hierarchical understanding, and contributes the HierarCaps dataset as a benchmark.
Emerging Property of Masked Token for Effective Pre-training
Hyesong Choi (Ewha Womans University Hyundai Motor Company), Dongbo Min (Ewha Womans University Hyundai Motor Company)
ClassificationRepresentation LearningTransformerImage
🎯 What it does: This paper addresses the inefficiency of Mask Token in MIM pre-training by proposing the Masked Token Optimization (MTO) method, which enhances pre-training efficiency by analyzing and optimizing the learning process of mask tokens in Transformers.
EMIE-MAP: Large-Scale Road Surface Reconstruction Based on Explicit Mesh and Implicit Encoding
Wenhua Wu (Shanghai Jiao Tong University), Hesheng Wang (Shanghai Jiao Tong University)
Autonomous DrivingNeural Radiance FieldImageMultimodalityPoint Cloud
🎯 What it does: Propose an EMIE-MAP method combining explicit mesh and implicit encoding for large-scale road surface reconstruction, capable of simultaneously outputting RGB maps, semantic maps, and elevation maps.
EMO: Emote Portrait Alive - Generating Expressive Portrait Videos with Audio2Video Diffusion Model under Weak Conditions
Linrui Tian (Alibaba Group), Liefeng Bo (Alibaba Group)
GenerationTransformerDiffusion modelImageVideoMultimodalityAudio
🎯 What it does: Generate synchronized and expressive talking and singing videos of arbitrary duration directly from a single portrait photo and audio input using diffusion models.
EmoTalk3D: High-Fidelity Free-View Synthesis of Emotional 3D Talking Head
Qianyun He (Nanjing University), Hao Zhu (Huawei)
GenerationData SynthesisRecurrent Neural NetworkTransformerNeural Radiance FieldGaussian SplattingPoint CloudMeshAudio
🎯 What it does: Built a three-stage mapping framework from audio to 3D geometry to appearance, achieving high-fidelity 3D talking head synthesis with controllable emotions and free viewpoints.
Empowering Embodied Visual Tracking with Visual Foundation Models and Offline RL
Fangwei Zhong (Peking University), Hao Chen (City University of Macau)
Object TrackingRecurrent Neural NetworkReinforcement LearningVision Language ModelImageVideo
🎯 What it does: Designed a framework that integrates visual foundation models with offline reinforcement learning for efficiently training visual tracking agents with robustness and generalizability in physical phenomena.
Encapsulating Knowledge in One Prompt
Qi Li (National University of Singapore), Xinchao Wang (National University of Singapore)
ClassificationKnowledge DistillationRepresentation LearningConvolutional Neural NetworkPrompt EngineeringGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Proposes a new knowledge transfer paradigm called Knowledge in One Prompt (KiOP), which encapsulates knowledge from multiple models into a single visual prompt without modifying the source models or accessing the original training data.
End-to-End Rate-Distortion Optimized 3D Gaussian Representation
Henan Wang (University of Science and Technology of China), Zhibo Chen (Microsoft Research Asia)
CompressionGaussian SplattingImage
🎯 What it does: Under the 3D Gaussian Splatting framework, an end-to-end rate-distortion optimization (RDO) model is proposed, achieving flexible bitrate control and compressing the model volume to less than 1/40 of the original through dynamic Gaussian pruning, adaptive spherical harmonic pruning, and entropy-constrained vector quantization.
Energy-Clibrated VAE with Test Time Free Lunch
Yihong Luo (Hong Kong University of Science and Technology), Jing Tang (Hong Kong University of Science and Technology)
RestorationGenerationFlow-based ModelAuto EncoderImageStochastic Differential Equation
🎯 What it does: Propose Energy-Calibrated VAE (EC-VAE), which achieves high-quality sample generation during the inference phase without requiring MCMC by using conditional EBM to perform short MCMC calibration on VAE-generated samples during training.