ECCV 2024 Papers — Page 11
European Conference on Computer Vision · 2387 papers
How Many Unicorns Are in This Image? A Safety Evaluation Benchmark for Vision LLMs
Haoqin Tu (University Of Santa Cruz), Cihang Xie (National University Of Singapore)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextBenchmark
🎯 What it does: Proposed a security evaluation benchmark named Unicorn to test the OOD generalization and adversarial robustness of Vision LLMs.
How to Train the Teacher Model for Effective Knowledge Distillation
Shayan Mohajer Hamidi (University of Waterloo), Ahmed Hussein Salamah (University of Waterloo)
ClassificationKnowledge DistillationImage
🎯 What it does: Proposed and verified that using mean squared error (MSE) loss to train the teacher model can enhance knowledge distillation (KD) effectiveness, replacing the traditional cross-entropy (CE) teacher with an MSE teacher in various KD methods, and observed performance improvements in the student model.
How Video Meetings Change Your Expression
Sumit Sarin (Columbia University), Carl Vondrick (Columbia University)
Image TranslationExplainability and InterpretabilityAuto EncoderGenerative Adversarial NetworkVideo
🎯 What it does: Discover and explain spatiotemporal expression differences between video conferencing (VC) and face-to-face (F2F) conversations using the generative domain translation framework FacET, and convert VC videos into F2F visual effects.
HowToCaption: Prompting LLMs to Transform Video Annotations at Scale
Nina Shvetsova (Goethe University Frankfurt), Hilde Kuehne (University of Oxford)
Data SynthesisRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextAudio
🎯 What it does: We leverage large language models (e.g., Vicuna-13B) to prompt ASR subtitles from a large number of instructional videos, generating video subtitles in a human-written style and predicting timestamps for each subtitle, thereby constructing a dataset of 25M video-text pairs named HowToCaption without requiring manual annotation.
HPE-Li: WiFi-enabled Lightweight Dual Selective Kernel Convolution for Human Pose Estimation
Toan D. Gian (VinUniversity), Van-Dinh Nguyen (VinUniversity)
Pose EstimationKnowledge DistillationConvolutional Neural NetworkMultimodality
🎯 What it does: Designed and implemented a lightweight dual selective kernel convolutional network (HPE-Li) based on WiFi CSI, generating precise 3D skeleton poses through multimodal sensors (camera + WiFi), and using a teacher network for supervised learning in the student network.
HPFF: Hierarchical Locally Supervised Learning with Patch Feature Fusion
Junhao Su (Southeast University), Chenyang Si (Southeast University)
ClassificationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: Proposed the HPFF method, combining hierarchical local supervision learning with patch feature fusion to enhance local learning performance and reduce GPU memory consumption.
HSR: Holistic 3D Human-Scene Reconstruction from Monocular Videos
Lixin Xue (ETH Zürich), Otmar Hilliges (ETH Zürich)
GenerationNeural Radiance FieldOptical FlowVideo
🎯 What it does: Propose a globally consistent framework for simultaneously reconstructing dynamic humans and static scenes from monocular RGB videos.
Human Hair Reconstruction with Strand-Aligned 3D Gaussians
Egor Zakharov (ETH Zürich), Otmar Hilliges (ETH Zürich)
GenerationDiffusion modelImageVideo
🎯 What it does: Proposes the Gaussian Haircut method, which achieves high-quality, editable hairstyle reconstruction using multi-view images through dual representations (hair strand polylines and 3D Gaussians aligned with hair strands).
Human Motion Forecasting in Dynamic Domain Shifts: A Homeostatic Continual Test-time Adaptation Framework
Qiongjie Cui (Nanjing University Of Science And Technology), Weiqing Li
Pose EstimationDomain AdaptationKnowledge DistillationRepresentation LearningTime SeriesSequential
🎯 What it does: This paper proposes a Homeostatic Continuous Test-Time Adaptation (HoCoTTA) method based on a teacher-student framework for human motion prediction under continuously changing target domains.
Human Pose Recognition via Occlusion-Preserving Abstract Images
Saad Manzur (University of California Irvine), Wayne B Hayes (University of California Irvine)
Data SynthesisPose EstimationConvolutional Neural NetworkImage
🎯 What it does: A two-stage network for 3D human pose estimation based on abstract images that retain occlusion information was studied. The real image is first converted into an occlusion-preserving abstract image, and then the viewpoint and pose are predicted from this image.
Human-in-the-Loop Visual Re-ID for Population Size Estimation
Gustavo Perez (University of California Berkeley), Subhransu Maji (University of Massachusetts Amherst)
RecognitionRetrievalReinforcement Learning from Human FeedbackGraph Neural NetworkImage
🎯 What it does: Investigated the use of human-in-the-loop visual Re-ID systems to estimate animal population sizes.
HumanRefiner: Benchmarking Abnormal Human Generation and Refining with Coarse-to-fine Pose-Reversible Guidance
Guian Fang (Sun Yat-sen University), Xiaodan Liang (Sun Yat-sen University)
Object DetectionGenerationData SynthesisAnomaly DetectionVision Language ModelDiffusion modelImageMultimodalityBenchmark
🎯 What it does: Constructed a large-scale human abnormal benchmark AbHuman, and proposed HumanRefiner, an end-to-end coarse-to-fine, reversible pose-guided generation process, significantly improving the accuracy of text-to-image models in human generation.
HUMOS: Human Motion Model Conditioned on Body Shape
Shashank Tripathi (Max Planck Institute for Intelligent Systems), Carsten Stoll (Epic Games)
GenerationData SynthesisTransformerAuto EncoderMeshSequential
🎯 What it does: Propose a generative motion model called HUMOS based on body shape conditions that can generate natural, physically feasible, and dynamically stable human motions for different body shapes in one go.
HVCLIP: High-dimensional Vector in CLIP for Unsupervised Domain Adaptation
Noranart Vesdapunt (Amazon), Pradeep Natarajan (Amazon)
Domain AdaptationVision Language ModelGenerative Adversarial NetworkImageBenchmark
🎯 What it does: In unsupervised domain adaptation tasks, robust fine-tuning of CLIP's pre-trained knowledge is achieved by mapping CLIP model features into a high-dimensional super vector space, combined with three strategies: forgetting inhibition, domain difference reduction, and feature enhancement.
Hybrid Video Diffusion Models with 2D Triplane and 3D Wavelet Representation
Kihong Kim (VIVE STUDIOS), Jaejun Yoo (UNIST)
GenerationData SynthesisConvolutional Neural NetworkTransformerDiffusion modelVideo
🎯 What it does: Designed and trained a hybrid video diffusion model (HVDM) that utilizes 2D tri-plane and 3D wavelet encoding to achieve high-quality video generation.
HybridBooth: Hybrid Prompt Inversion for Efficient Subject-Driven Generation
Shanyan Guan (vivo Mobile Communication Co., Ltd), Mingyu You (vivo Mobile Communication Co., Ltd)
GenerationPrompt EngineeringVision Language ModelDiffusion modelImage
🎯 What it does: Proposes the HybridBooth framework to rapidly generate high-quality, theme-driven images from a single image.
HYDRA: A Hyper Agent for Dynamic Compositional Visual Reasoning
Fucai Ke (Monash University), Hamid Rezatofighi (Monash University)
Large Language ModelReinforcement LearningAgentic AIVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose a multi-stage dynamic compositional visual reasoning framework named HYDRA, which utilizes a planner, reinforcement learning controller, and reasoner to achieve recursive and incremental visual reasoning.
HYPE: Hyperbolic Entailment Filtering for Underspecified Images and Texts
Wonjae Kim (NAVER AI Lab), Sangdoo Yun (NAVER AI Lab)
RetrievalData-Centric LearningVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposed a filtering method called HYPE based on hypercurve embedding to extract high specificity and well-aligned samples from large-scale noisy image-text pair datasets.
Hypernetworks for Generalizable BRDF Representation
Fazilet Gokbudak (University of Cambridge), A. Cengiz Oztireli
CompressionComputational EfficiencyRepresentation LearningPhysics Related
🎯 What it does: Proposed a hypernetwork-based universal BRDF representation method that can estimate the BRDF of unknown materials from sparse measurement samples and compress it into an extremely low-dimensional embedding.
HyperSpaceX: Radial and Angular Exploration of HyperSpherical Dimensions
Chiranjeev Chiranjeev (Indian Institute of Technology Jodhpur), Richa Singh (Indian Institute of Technology Jodhpur)
ClassificationRecognitionConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes the HyperSpaceX framework, which simultaneously explores radial and angular features in multiple hyperspheres, and introduces the DistArc loss to enhance performance in image classification and face recognition.
HyTAS: A Hyperspectral Image Transformer Architecture Search Benchmark and Analysis
Fangqin Zhou (Eindhoven University of Technology), Ran Piao (Eindhoven University of Technology)
Neural Architecture SearchTransformerImageBenchmark
🎯 What it does: Propose the HyTAS benchmark, systematically evaluate the effectiveness of 12 zero-cost proxies in searching for Transformers across five hyperspectral datasets, and analyze the relationship between proxies and model features.
I Can't Believe It's Not Scene Flow!
Ishan Khatri (Stack AV), James Hays (CMU)
Object DetectionObject TrackingAutonomous DrivingOptical FlowPoint Cloud
🎯 What it does: This paper points out that existing scene flow methods perform poorly on small targets (such as pedestrians and cyclists), and proposes a novel evaluation metric called Bucket Normalized EPE, as well as a simple baseline called TrackFlow based on 3D detection + Kalman filtering to address this issue.
I-MedSAM: Implicit Medical Image Segmentation with Segment Anything
Xiaobao Wei (Peking University), Shanghang Zhang (Peking University)
SegmentationTransformerSupervised Fine-TuningImageBiomedical Data
🎯 What it does: Propose I-MedSAM, a medical image segmentation framework that combines the Segment Anything Model (SAM) with implicit neural representations (INR)
I2-SLAM: Inverting Imaging Process for Robust Photorealistic Dense SLAM
Gwangtak Bae (Seoul National University), Young Min Kim (Seoul National University)
Neural Radiance FieldGaussian SplattingSimultaneous Localization and MappingImageVideo
🎯 What it does: Propose the I²-SLAM module, which inverts the image formation process and integrates it into a dense SLAM pipeline, enabling the reconstruction of sharp HDR 3D maps from arbitrarily captured videos containing motion blur and automatic exposure changes.
IAM-VFI : Interpolate Any Motion for Video Frame Interpolation with motion complexity map
Kihwan Yoon (University of Seoul), Jinwoo Jeong (Korea Electronics Technology Institute)
RestorationGenerationConvolutional Neural NetworkFlow-based ModelOptical FlowVideo
🎯 What it does: Propose the IAM-VFI method, which classifies regions with different motion complexity in videos and generates motion complexity maps using MCENet to achieve high-quality video frame interpolation for arbitrary motion.
Idea2Img: Iterative Self-Refinement with GPT-4V for Automatic Image Design and Generation
Zhengyuan Yang (Microsoft), Lijuan Wang (Microsoft)
GenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelMultimodality
🎯 What it does: Propose the Idea2Img framework, which utilizes GPT-4V for multi-modal iterative self-improvement to generate images from high-level design concepts.
Idempotent Unsupervised Representation Learning for Skeleton-Based Action Recognition
Lilang Lin (Peking University), Jiaying Liu (Peking University)
RecognitionRepresentation LearningTransformerDiffusion modelContrastive LearningVideo
🎯 What it does: Proposes an unsupervised skeleton action recognition method based on the Isometric Generative Model (IGM), combining generative pre-training with contrastive learning to achieve isometric constraints in the feature space.
Identity-Consistent Diffusion Network for Grading Knee Osteoarthritis Progression in Radiographic Imaging
Wenhua Wu (University of Sydney), Zhiyong Wang (University of Sydney)
Image TranslationGenerationDiffusion modelContrastive LearningImageBiomedical Data
🎯 What it does: This paper proposes an Identity Consistent Diffusion Network (IC-RDN) that can generate knee X-ray images 12 months after baseline X-ray scans and utilize both baseline images and generated images to predict the progression of knee osteoarthritis (KOA) in terms of Kellgren-Lawrence (KL) grades.
IDOL: Unified Dual-Modal Latent Diffusion for Human-Centric Joint Video-Depth Generation
Yuanhao Zhai (State University of New York at Buffalo), Lijuan Wang (Microsoft)
GenerationData SynthesisDepth EstimationConvolutional Neural NetworkDiffusion modelImageVideoMultimodality
🎯 What it does: Propose a unified bimodal diffusion model called IDOL that can simultaneously generate human videos and their corresponding depth maps.
IFTR: An Instance-Level Fusion Transformer for Visual Collaborative Perception
Shaohong Wang (Zhejiang University), Eryun Liu (Zhejiang University)
Object DetectionAutonomous DrivingTransformerImage
🎯 What it does: Proposes a visual collaborative perception framework based on Instance-Level Fusion Transformer (IFTR), which enhances the 3D detection performance of multi-vehicle cameras by leveraging shared instance features.
IG Captioner: Information Gain Captioners are Strong Zero-shot Classifiers
Chenglin Yang (Johns Hopkins University), Jiahui Yu (Google DeepMind)
ClassificationRetrievalVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes training a generative captioning model as a zero-shot classifier, reducing reliance on language priors through information gain evaluation.
IGNORE: Information Gap-based False Negative Loss Rejection for Single Positive Multi-Label Learning
Gyeong Ryeol Song, Jee-Hyong Lee (Sungkyunkwan University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: Proposes an information gap-based method for rejecting misjudged negative samples (IGNORE), which utilizes CAM generated from a single positive label to create a masked image, and identifies and rejects misjudged negative samples during training by comparing the logits differences between the original image and the masked image.
iHuman: Instant Animatable Digital Humans From Monocular Videos
Pramish Paudel (Tribhuvan University), Ajad Chhatkuli (Tribhuvan University)
GenerationPose EstimationGaussian SplattingVideoPoint CloudMesh
🎯 What it does: Instantly generate animatable 3D digital humans from monocular video, providing high-quality meshes and renderable Gaussian splats.
Image Compression for Machine and Human Vision With Spatial-Frequency Adaptation
Han Li (Shanghai Jiao Tong University), Hongkai Xiong (Shanghai Jiao Tong University)
CompressionConvolutional Neural NetworkImage
🎯 What it does: A lightweight adapter framework called Adapt-ICMH is studied, which transfers a pre-trained human visual image compression model to machine vision tasks while balancing bitrate and task accuracy.
Image Demoireing in RAW and sRGB Domains
Shuning Xu, Jiantao Zhou (University of Macau)
RestorationConvolutional Neural NetworkTransformerImage
🎯 What it does: Propose an image demoiréing network (RRID) that leverages dual-domain information from RAW and sRGB. It achieves multi-scale stripe removal by inserting a Gated Feedback Module (GFM) and a Frequency Selection Module (FSM) into the U-Net structure, and restores color through RGB Guided ISP (RGISP) that learns device-specific ISP.
Image Manipulation Detection With Implicit Neural Representation and Limited Supervision
Zhenfei Zhang (University at Albany State University of New York), Jun-Wei Hsieh (National Yang Ming Chiao Tung University)
Anomaly DetectionConvolutional Neural NetworkNeural Radiance FieldContrastive LearningImage
🎯 What it does: Proposed a unified image tampering detection framework compatible with weakly supervised and unsupervised training modes.
Image-adaptive 3D Lookup Tables for Real-time Image Enhancement with Bilateral Grids
Wontae Kim (Seoul National University), Nam Ik Cho (Seoul National University)
RestorationConvolutional Neural NetworkImage
🎯 What it does: Designed an image enhancement model combining bilateral grids and 3D LUT to achieve real-time spatial-aware enhancement.
Image-Feature Weak-to-Strong Consistency: An Enhanced Paradigm for Semi-Supervised Learning
Zhiyu Wu (Peking University), Jinshi Cui (Peking University)
ClassificationRepresentation LearningData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: Proposes a new paradigm called Image-Feature Weak-Strong Consistency (IFMatch), improving semi-supervised learning by combining feature-level and image-level perturbations in a three-branch structure;
Image-to-Lidar Relational Distillation for Autonomous Driving Data
Anas Mahmoud (University of Toronto), Steven Waslander (University of Toronto)
SegmentationAutonomous DrivingKnowledge DistillationRepresentation LearningConvolutional Neural NetworkPrompt EngineeringVision Language ModelContrastive LearningImageTextPoint Cloud
🎯 What it does: This paper proposes a new framework for knowledge distillation of LiDAR point cloud encoders based on 2D vision-language foundation models (CLIP, DINOv2), with the core idea of approximating 2D representation space through cross-modal and intra-modal relational loss, thereby achieving more structurally consistent 3D representations.
Images are Achilles' Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Models
Yifan Li (Renmin University of China), Ji-Rong Wen (Renmin University of China)
Safty and PrivacyAdversarial AttackPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Conduct systematic experiments on the harmless alignment of multimodal large language models (MLLMs), demonstrating that image inputs act as backdoors for alignment, and proposing a three-phase automatic attack method HADES to hide and amplify harmful information in images for exploitation.
Imaging Interiors: An Implicit Solution to Electromagnetic Inverse Scattering Problems
Ziyuan Luo (Hong Kong Baptist University), Renjie Wan (Hong Kong Baptist University)
OptimizationPhysics Related
🎯 What it does: Propose a forward optimization method based on implicit neural representations, using a continuous multi-layer perceptron to reconstruct the relative permittivity and induced currents in electromagnetic inverse scattering problems, avoiding the difficulties of traditional inverse estimation and matrix inversion.
Imaging with Confidence: Uncertainty Quantification for High-dimensional Undersampled MR Images
Frederik Hoppe (RWTH Aachen University), Holger Rauhut (LMU Munich)
RestorationOptimizationBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose a debiased estimator based on Total Variation (TV) minimization, and use it to construct pixel-level confidence intervals to achieve confidence quantification in undersampled magnetic resonance imaging (MRI).
iMatching: Imperative Correspondence Learning
Zitong Zhan (University at Buffalo), Chen Wang (InnoPeak Technology)
Pose EstimationOptimizationSimultaneous Localization and MappingVideo
🎯 What it does: Propose a self-supervised feature correspondence learning framework called iMatching, which uses bundle adjustment as the supervisory signal for lower-level optimization, and achieves end-to-end training of the feature matching network through bilevel optimization.
IMMA: Immunizing text-to-image Models against Malicious Adaptation
Amber Yijia Zheng (Purdue University), Raymond A. Yeh (Purdue University)
GenerationSafty and PrivacyMeta LearningDiffusion modelImageText
🎯 What it does: Before releasing pre-trained text-to-image diffusion models, immunization training is conducted to make the models difficult to adapt to malicious fine-tuning (e.g., replicating artistic styles, recovering deleted concepts, and personalizing content).
Implicit Concept Removal of Diffusion Models
Zhili Liu (Hong Kong University of Science and Technology), James Kwok
RestorationGenerationSupervised Fine-TuningPrompt EngineeringDiffusion modelImageBenchmark
🎯 What it does: Proposed a geometry-based implicit concept removal method called Geom-Erasing to eliminate implicit concepts such as watermarks, QR codes, and text in diffusion models.
Implicit Filtering for Learning Neural Signed Distance Functions from 3D Point Clouds
Shengtao Li (Tsinghua University), Yu-Shen Liu (Tsinghua University)
RestorationPoint Cloud
🎯 What it does: Propose a bilateral filtering-based implicit filtering method to improve the signed distance function (SDF) learned by neural networks, thereby enhancing the quality of surface reconstruction from 3D point clouds.
Implicit Neural Models to Extract Heart Rate from Video
Pradyumna Chari (University of California Los Angeles), Achuta Kadambi (University of California Los Angeles)
Computational EfficiencyRepresentation LearningConvolutional Neural NetworkNeural Radiance FieldVideoBiomedical Data
🎯 What it does: Propose a heart rate extraction framework based on implicit neural representations (INR), which improves the signal-to-noise ratio (SNR) of remote photoplethysmography (rPPG) signals by decomposing facial videos into Appearance-Blood (A-B) components to separate blood flow information;
Implicit Steganography Beyond the Constraints of Modality
Sojeong Song (Korea Advanced Institute of Science and Technology), Junmo Kim (Korea Advanced Institute of Science and Technology)
Safty and PrivacyComputational EfficiencyNeural Radiance FieldAuto EncoderImageVideoMultimodalityAudio
🎯 What it does: Proposes a cross-modal steganography framework INRSteg based on implicit neural representations (INR), which can embed secret data of multiple modalities (image, audio, video, 3D shapes) into a carrier of another modality, and supports embedding multiple secret messages in one operation.
Implicit Style-Content Separation using B-LoRA
Yarden Frenkel (Tel Aviv University), Danny Cohen-Or
GenerationTransformerPrompt EngineeringDiffusion modelImage
🎯 What it does: Propose the B-LoRA method, which utilizes LoRA to fine-tune two specific Transformer blocks in SDXL, achieving implicit style-content separation for a single image.
Improving 2D Feature Representations by 3D-Aware Fine-Tuning
Yuanwen Yue, Jan Eric Lenssen (ETH Zurich)
SegmentationDepth EstimationConvolutional Neural NetworkTransformerSupervised Fine-TuningGaussian SplattingImagePoint Cloud
🎯 What it does: This paper proposes a two-stage method: first, the semantic features of a 2D foundation model (e.g., DINOv2) are upscaled to a 3D efficient Gaussian representation, achieving high-resolution feature reconstruction through multi-view consistency and RGB guidance; subsequently, the 3D-aware features rendered are used to fine-tune the original 2D foundation model, enhancing its performance on downstream tasks such as semantic segmentation and depth estimation.
Improving 3D Semi-supervised Learning by Effectively Utilizing All Unlabelled Data
Sneha Paul (Concordia University), Nizar Bouguila (Concordia University)
ClassificationRepresentation LearningTransformerContrastive LearningPoint Cloud
🎯 What it does: Proposes the AllMatch framework, which fully utilizes all unlabeled point cloud samples to enhance semi-supervised learning performance in 3D classification tasks through three modules: adaptive hard augmentation, reverse learning, and contrastive learning.
Improving Adversarial Transferability via Model Alignment
Avery Ma (University of Toronto), Jindong Gu (University of Oxford)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: Fine-tune the source model to make its predictions consistent with an independently trained witness model, generating adversarial perturbations with higher transferability.
Improving Agent Behaviors with RL Fine-tuning for Autonomous Driving
Zhenghao Peng (Ucla), Justin Fu (Waymo)
Autonomous DrivingTransformerSupervised Fine-TuningReinforcement LearningTime SeriesSequentialBenchmark
🎯 What it does: This paper builds upon pre-trained large-scale traffic behavior prediction models and uses reinforcement learning based on policy gradients to perform closed-loop fine-tuning of behavior models in simulated driving scenarios, aiming to enhance the realism and safety of vehicle interaction behaviors.
Improving Diffusion Models for Authentic Virtual Try-on in the Wild
Yisol Choi (Korea Advanced Institute of Science and Technology), Jinwoo Shin (OMNIOUS.AI)
Image TranslationGenerationTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: Designed and trained IDM-VTON, a diffusion model-based virtual try-on system capable of generating realistic and detail-preserving try-on images under any pose.
Improving Domain Generalization in Self-Supervised Monocular Depth Estimation via Stabilized Adversarial Training
Yuanqi Yao (Harbin Institute of Technology), Junjun Jiang (Harbin Institute of Technology)
Depth EstimationDomain AdaptationConvolutional Neural NetworkGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Proposes the SCAT framework, introducing stable adversarial training in self-supervised monocular depth estimation to enhance domain generalization capability.
Improving Feature Stability during Upsampling -- Spectral Artifacts and the Importance of Spatial Context
Shashank Agnihotri (University of Mannheim), Margret Keuper (University of Mannheim)
RestorationSegmentationDepth EstimationAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: Investigate spectral artifacts that occur during the upsampling process, proposing the use of transposed convolution with large kernels (LCTC) to enhance feature stability in pixel-level predictions, and verifying its effectiveness in image restoration, semantic segmentation, and disparity estimation.
Improving Geo-diversity of Generated Images with Contextualized Vendi Score Guidance
Reyhane Askari Hemmat (FAIR at Meta), Adriana Romero-Soriano (FAIR at Meta)
GenerationData SynthesisDiffusion modelScore-based ModelImage
🎯 What it does: Interpolate potential diffusion models (LDM) during the inference phase, proposing the Contextualized Vendi Score Guidance (c-VSG) mechanism, which utilizes a memory pool and a small number of real samples to guide the generation process, making generated images in different regions more geographically diverse.
Improving Hyperbolic Representations via Gromov-Wasserstein Regularization
Yifei Yang (Wuhan University), Gilad Lerman (University of Minnesota)
ClassificationRepresentation LearningImageGraph
🎯 What it does: Introduce Gromov-Wasserstein regularization into existing hyperbolic neural networks (HNNs), enabling the model to better preserve the original geometric structure when mapping Euclidean features to hyperbolic space.
Improving image synthesis with diffusion-negative sampling
Alakh Desai (University of California San Diego), Nuno Vasconcelos (University of California San Diego)
GenerationData SynthesisPrompt EngineeringDiffusion modelImageText
🎯 What it does: Proposes a method that utilizes Diffusion Negative Sampling (DNS) to generate images that are most inconsistent with text prompts, and then generates negative prompts (DNP) from these images, bridging the semantic gap between diffusion models and humans in concept negation.
Improving Intervention Efficacy via Concept Realignment in Concept Bottleneck Models
Nishad Singhi (University of Tübingen), Zeynep Akata (Technische Universität München)
ClassificationRecurrent Neural NetworkImageBenchmark
🎯 What it does: Proposed a trainable concept intervention relocalization module to enhance the intervention effectiveness of concept bottleneck models (CBM/CEM/IntCEM) during testing.
Improving Knowledge Distillation via Regularizing Feature Direction and Norm
Yuzhu Wang, Shu Kong (Zhejiang University)
Knowledge DistillationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper enhances knowledge distillation by regularizing the feature direction and norm in the embedding layer of the student network, and proposes the Dino-loss loss function;
Improving Medical Multi-modal Contrastive Learning with Expert Annotations
Yogesh Kumar (Aalto University), Pekka Marttinen (Aalto University)
ClassificationRetrievalRepresentation LearningData-Centric LearningTransformerVision Language ModelContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Developed the eCLIP model, integrating radiologists' eye-tracking heatmaps and mixup augmentation into the CLIP framework to enhance the effectiveness of medical multimodal contrastive learning.
Improving Neural Surface Reconstruction with Feature Priors from Multi-View Images
Xinlin Ren (Fudan University), Xiangyang Xue (Fudan University)
GenerationConvolutional Neural NetworkTransformerNeural Radiance FieldImage
🎯 What it does: This paper proposes a pixel-level and block-level multi-view consistency loss in the feature space by combining the pre-trained feature priors extracted from multi-view images with Neural Surface Reconstruction (NSR). It systematically evaluates 13 models across seven types of pre-training tasks (MIM, IC, SS, MDE, SM, IM, MVS) and constructs two improved NSR variants, MVS-NeuS and Match-NeuS.
Improving Point-based Crowd Counting and Localization Based on Auxiliary Point Guidance
I-HSIANG CHEN, Sy-Yen Kuo (National Taiwan University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: Proposed a point-level crowd counting and localization method based on auxiliary point guidance (APG), combined with implicit feature interpolation (IFI) to stabilize the proposal-target matching process, significantly improving counting and localization accuracy.
Improving Robustness to Model Inversion Attacks via Sparse Coding Architectures
Sayanton V. Dibbo (Dartmouth College), Michael Teti (Los Alamos National Laboratory)
Safty and PrivacyConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: This paper proposes a network architecture (SCA) that alternately embeds sparse coding layers, defending against model inversion attacks by actively suppressing irrelevant private information in intermediate representations during training.
Improving Text-guided Object Inpainting with Semantic Pre-inpainting
Yifu Chen (Fudan University), Tao Mei (HiDream.ai Inc)
GenerationKnowledge DistillationTransformerDiffusion modelImageTextMultimodality
🎯 What it does: Proposes a two-stage text-guided object inpainting method called CAT-Diffusion, which first pre-fills the semantic features of the target object in a multi-modal feature space, and then guides the diffusion model to generate high-fidelity objects through a reference adapter layer.
Improving Unsupervised Domain Adaptation: A Pseudo-Candidate Set Approach
Aveen Dayal (Indian Institute of Technology Hyderabad), Vineeth N Balasubramanian (Indian Institute of Technology Hyderabad)
Domain AdaptationImageBenchmark
🎯 What it does: Proposed an unsupervised domain adaptation refinement method (UDPCS) based on pseudo-candidate sets, which can be post-trained on any existing UDA method to improve target domain accuracy.
Improving Video Segmentation via Dynamic Anchor Queries
Yikang Zhou (Wuhan University), Shuicheng Yan (Skywork AI)
Object TrackingSegmentationTransformerVideo
🎯 What it does: Proposed dynamic anchor query (DAQ) and emergence/disappearance simulation (EDS) strategies to enhance video segmentation models' ability to detect and track newly emerging and disappearing targets.
Improving Virtual Try-On with Garment-focused Diffusion Models
Siqi Wan (University of Science and Technology of China), Tao Mei (HiDream.ai Inc)
Image TranslationImage HarmonizationGenerationVision Language ModelDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: Proposes GarDiff, a garment-focused diffusion model for virtual try-on, capable of generating high-quality, detail-rich wearable effect images.
Improving Vision and Language Concepts Understanding with Multimodal Counterfactual Samples
Chengen Lai (Xidian University), Guangneng Hu (Xidian University)
Data SynthesisTransformerVision Language ModelDiffusion modelGenerative Adversarial NetworkContrastive LearningMultimodality
🎯 What it does: By automatically generating multimodal adversarial examples (text and image) and incorporating them into contrastive learning training, the performance of vision-language models in concept understanding and compositional reasoning is improved.
Improving Zero-Shot Generalization for CLIP with Variational Adapter
Ziqian Lu (Zhejiang University), Xi Li (Zhejiang University)
ClassificationRecognitionDomain AdaptationRepresentation LearningTransformerPrompt EngineeringVision Language ModelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: Propose a Prompt-based Variational Adapter (PVA), which aligns visual and semantic features in a shared latent space by combining learnable text prompts and a spherical variational autoencoder (SVAE) within the CLIP model. This significantly reduces bias toward base classes in GZSL scenarios, classifies test samples into base and novel classes via cosine similarity in the latent space, and refines adapter outputs using residual connections to enhance generalization for novel classes.
Improving Zero-shot Generalization of Learned Prompts via Unsupervised Knowledge Distillation
Marco Mistretta (University of Florence), Andrew D. Bagdanov (University of Florence)
Knowledge DistillationRepresentation LearningPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: Propose an unsupervised knowledge distillation prompt learning framework KDPL to improve the zero-shot generalization capability of lightweight vision-language models (VLMs).
In Defense of Lazy Visual Grounding for Open-Vocabulary Semantic Segmentation
Dahyun Kang (Pohang University of Science and Technology), Minsu Cho (Pohang University of Science and Technology)
SegmentationTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: Propose a two-step lazy visual grounding framework, first discovering object masks through unsupervised normalized cut on DINO features, then aligning text descriptions with masks via CLIP for open-vocabulary semantic segmentation.
iNeMo: Incremental Neural Mesh Models for Robust Class-Incremental Learning
Tom Fischer (Saarland University), Eddy Ilg (Saarland University)
ClassificationPose EstimationKnowledge DistillationRepresentation LearningContrastive LearningMesh
🎯 What it does: Propose an incrementally learnable 3D neural grid model (iNeMo) that can perform classification and pose estimation while continuously adding new categories, and maintain robustness in out-of-distribution (OOD) scenarios.
Inf-DiT: Upsampling any-resolution image with memory-efficient diffusion transformer.
Zhuoyi Yang (Tsinghua University), Jie Tang (Tsinghua University)
GenerationSuper ResolutionTransformerDiffusion modelImage
🎯 What it does: This paper proposes Inf-DiT, a diffusion transformer-based infinite-resolution image upscaling model that achieves efficient upscaling for images of arbitrary sizes using unidirectional block attention (UniBA).
Infinite-ID: Identity-preserved Personalization via ID-semantics Decoupling Paradigm
Yi Wu (University of Science and Technology of China), Bin Li (University of Science and Technology of China)
GenerationTransformerVision Language ModelDiffusion modelMultimodality
🎯 What it does: Propose the Infinite-ID framework, which achieves separation and fusion of identity information and semantic information to enhance personalized text-to-image generation quality by using identity-enhanced training (disabling text cross-attention) during training and adopting hybrid attention and AdaIN-mean mechanisms during inference.
InfMAE: A Foundation Model in The Infrared Modality
Fangcen Liu, Deyu Meng (Sun Yat-sen University)
Object DetectionSegmentationConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Propose InfMAE, a foundational model for infrared images, achieving self-supervised pre-training through information-aware masking, multi-scale encoder, and infrared decoder, demonstrating excellent performance in infrared semantic segmentation, object detection, and small target detection tasks.
InfoNorm: Mutual Information Shaping of Normals for Sparse-View Reconstruction
Xulong Wang (Zhejiang University), Yanchao Yang (University of Hong Kong)
GenerationNeural Radiance FieldContrastive LearningImage
🎯 What it does: Propose an SDF-NeRF framework that utilizes mutual information to constrain surface normals under sparse viewpoints, significantly improving the quality of 3D surface reconstruction.
Information Bottleneck Based Data Correction in Continual Learning
Shuai Chen (CRISE Institute of Automation Chinese Academy of Sciences), Kaiqi Huang (CRISE Institute of Automation Chinese Academy of Sciences)
Representation LearningData-Centric LearningImage
🎯 What it does: Propose an information bottleneck-based continual learning data correction framework (IBCL), which combines two modules: information bottleneck task-agnostic constraint (IBTA) and information bottleneck sample-free data surrogate (IBDS), along with experience replay methods, to alleviate catastrophic forgetting and data bias caused by interactions between new and old tasks.
Insect Identification in the Wild: The AMI Dataset
Aditya Jain (Mila - Quebec AI Institute), David Rolnick (Mila - Quebec AI Institute)
ClassificationRecognitionDomain AdaptationConvolutional Neural NetworkTransformerImageBenchmark
🎯 What it does: This paper constructs the first large-scale insect identification benchmark dataset, AMI, containing 2.5 million images collected by humans and 2,893 images captured by automated insect traps, with a total of 5,364 moth species and 52,948 individuals annotated. The dataset aims to evaluate the performance of fine-grained insect recognition and domain transfer in field deployment.
InsMapper: Exploring Inner-instance Information for Vectorized HD Mapping
zhenhua xu, Hengshuang Zhao (University of Hong Kong)
Object DetectionAutonomous DrivingTransformer
🎯 What it does: In the vehicle perception scenario, the InsMapper framework based on Transformer is proposed for online detection of road elements in vectorized high-definition maps;
Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation
Arpit Garg (University of Adelaide), Gustavo Carneiro (University of Surrey)
ClassificationImage
🎯 What it does: Propose a noise rate estimation method based on a graphical model to improve the sample selection curriculum in instance-dependent noise label learning, and seamlessly integrate with existing state-of-the-art (SOTA) noise label learning (LNL) methods.
Instant 3D Human Avatar Generation using Image Diffusion Models
Nikos Kolotouros (Google Research), Cristian Sminchisescu (Google Research)
GenerationData SynthesisVision Language ModelDiffusion modelNeural Radiance FieldImageTextMesh
🎯 What it does: Propose the AvatarPopUp method, which enables controlling human pose and body shape from text, image, or a combination of both inputs, rapidly (2–10 seconds) generating bindable-skeleton, textured full-body 3D human avatars.
Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator
Niki Amini-Naieni, Ronald Clark
GenerationNeural Radiance FieldImage
🎯 What it does: This paper constructs a parametric model for uncalibrated uncertainty calibration curves based on DINOv2 features, and uses MLP to learn mapping from raw probabilities to calibrated probabilities;
InstaStyle: Inversion Noise of a Stylized Image is Secretly a Style Adviser
Xing Cui (Beijing University of Posts and Telecommunications), Zhaofeng He (Beijing University of Posts and Telecommunications)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: Utilize the DDIM reverse process to convert a single style reference image into noise, and sample from this noise to generate new stylized images, while learning generalizable style tokens through prompt refinement.
InstructGIE: Towards Generalizable Image Editing
Zichong Meng (Northeastern University), Yanzhi Wang (Northeastern University)
Image TranslationTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Propose an image editing framework called InstructGIE aimed at enhancing generalization for unseen editing tasks.
Instruction Tuning-free Visual Token Complement for Multimodal LLMs
Dongsheng Wang (Shenzhen University), Hanwang Zhang (Nanyang Technological University)
RestorationGenerationTransformerVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: Propose a visual token completion framework (VTC) that does not require instruction tuning, enabling multimodal LLMs to recover missing visual tokens by reconstructing priors, thereby improving answer accuracy.
InstructIR: High-Quality Image Restoration Following Human Instructions
Marcos V. Conde (University of Würzburg), Radu Timofte (University of Würzburg)
RestorationConvolutional Neural NetworkLarge Language ModelPrompt EngineeringVision Language ModelImageText
🎯 What it does: Proposed a universal image restoration model called InstructIR based on human natural language instructions, capable of restoring images in various tasks such as noise reduction, de-raining, de-fogging, deblurring, and low-light enhancement.
Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection
Xinhao Luo (Institute of Automation, Chinese Academy of Sciences), Guoqi Li (Institute of Automation, Chinese Academy of Sciences)
Object DetectionConvolutional Neural NetworkSpiking Neural NetworkImageTime Series
🎯 What it does: Propose a spiking neural network framework named SpikeYOLO for high-performance, low-energy consumption object detection, and design an I-LIF neuron with integer-value training and spike-driven inference; implemented and verified on the COCO 2017 static dataset and Gen1 neuromorphic event stream dataset.
Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization
Naiyu Yin (Rensselaer Polytechnic Institute), Qiang Ji (Rensselaer Polytechnic Institute)
Domain AdaptationRepresentation LearningAuto EncoderImage
🎯 What it does: This paper proposes a three-stage Causal Markov Blanket Representation Learning (CMBRL) framework, first learning latent variables using an identifiable variational autoencoder, then rapidly identifying the Markov Blanket (CMB) set of the target variable through mutual information testing, and finally constructing a domain-invariant predictor based on the CMB.
Integration of Global and Local Representations for Fine-grained Cross-modal Alignment
Seungwan Jin (Hanyang University), Kyungsik Han (Hanyang University)
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper proposes the Fashion-FINE model, which achieves fine-grained visual-text alignment in the fashion domain through the fusion of global and local features.
Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks
Lingzhuang Meng (China University of Petroleum (East China)), Wenjie Liu (China University of Petroleum (East China))
Knowledge DistillationAdversarial AttackDiffusion modelImage
🎯 What it does: This paper proposes an Inter-Class Topology Alignment (ICTA) framework, which uses Position Exploration Sample (PES) to more comprehensively simulate black-box target models in the decision space, thereby training substitute models with higher similarity and improving the success rate of target attacks.
Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition
Yisong Wang (Peking University), Junsong Yuan (University at Buffalo)
RecognitionConvolutional Neural NetworkRecurrent Neural NetworkGraph Neural NetworkTransformerVision Language ModelVideo
🎯 What it does: This paper proposes an interactive center spatiotemporal context reasoning framework for human-object interaction (HOI) recognition in multi-person videos.
Interactive 3D Object Detection with Prompts
Ruifei Zhang (Chinese University of Hong Kong), Guanbin Li (Baidu Inc)
Object DetectionTransformerPrompt EngineeringImageMultimodalityPoint Cloud
🎯 What it does: Designed and implemented a multimodal interactive 3D object detection framework based on 2D interactive prompts, 3D detection, and 3D refinement for efficient annotation and improvement of 3D detection results.
InterFusion: Text-Driven Generation of 3D Human-Object Interaction
Sisi Dai (National University of Defense Technology), Ruizhen Hu (Shenzhen University)
GenerationData SynthesisPose EstimationTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelNeural Radiance FieldImageTextMultimodalityMeshRetrieval-Augmented Generation
🎯 What it does: Designed and implemented a two-stage text-driven 3D human-computer interaction generation framework called InterFusion, which first generates text-corresponding anchor poses and then achieves high-quality HOI scenes through a local-global optimization process that combines geometric priors and text prompts.
Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection
Yongwei Nie (South China University Of Technology), Hongmin Cai (South China University Of Technology)
Anomaly DetectionAuto EncoderVideo
🎯 What it does: An unsupervised video anomaly detection framework was studied, achieving label-free learning by alternately training a one-class classification (OCC) model and a weakly supervised (WS) model to generate pseudo-labels.
InternVideo2: Scaling Foundation Models for Multimodal Video Understanding
Yi Wang (OpenGVLab, Shanghai AI Laboratory), Limin Wang (OpenGVLab, Shanghai AI Laboratory)
RecognitionRepresentation LearningTransformerLarge Language ModelContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: This paper proposes InternVideo2, a large-scale video foundation model constructed through three-stage progressive training (occlusion-free reconstruction, cross-modal contrastive learning, next token prediction), achieving state-of-the-art (SOTA) performance on multiple video tasks.
Interpretability-Guided Test-Time Adversarial Defense
Akshay Kulkarni (UC San Diego), Tsui-Wei Weng (UC San Diego)
Explainability and InterpretabilityAdversarial AttackImageBenchmark
🎯 What it does: Proposes a training-agnostic test-time adversarial defense method that leverages neuron interpretability information to rank the importance of internal neurons, retaining only those related to the predicted class during forward propagation;
INTRA: Interaction Relationship-aware Weakly Supervised Affordance Grounding
Ji Ha Jang (Seoul National University), Se Young Chun (Seoul National University)
SegmentationRepresentation LearningTransformerLarge Language ModelContrastive LearningImageText
🎯 What it does: Propose a weakly supervised interaction-aware object region localization method, INTRA, which generates action region maps through text-conditioned generation and contrastive learning using only appearance images.
Intrinsic Single-Image HDR Reconstruction
Sebastian Dille (Simon Fraser University), Yagiz Aksoy
RestorationConvolutional Neural NetworkSupervised Fine-TuningImagePhysics Related
🎯 What it does: This paper proposes a physics-inspired single-image HDR reconstruction pipeline, decomposing the problem into dynamic range expansion for shadows and reflectance, and ultimately fusing them to obtain high-quality HDR images.