IEEE/CVF Conference on Computer Vision and Pattern Recognition Β· 1047 papers
View-Aware Semantic Alignment for Aerial-Ground Person Re-Identification
Quan Zhang (Sun Yat-sen University), Jianhuang Lai (Sun Yat-sen University)
CodeRecognitionRetrievalGraph Neural NetworkTransformerMixture of ExpertsImage
π― What it does: Proposes a perspective-aware semantic alignment framework named ViSA for aerial-ground person re-identification. The framework achieves separation and fusion of viewpoint-invariant and viewpoint-specific features through a perspective-decoupled Transformer, an Expert-Driven Token Generation Module (ETGM), and a Dual-Branch Local Fusion Module (DLFM).
ViHOI: Human-Object Interaction Synthesis with Visual Priors
Songjin Cai (South China University of Technology), Changxing Ding (South China University of Technology)
CodeGenerationData SynthesisVision Language ModelDiffusion modelImageTextMultimodality
π― What it does: Design the ViHOI framework, which leverages a Vision-Language Model (VLM) to extract visual and textual priors, compresses them into compact tokens via a Q-Former adapter, and integrates with a diffusion generation model. During training, real-action rendered images serve as visual priors, while during inference, reference images are synthesized using text-to-image models, enabling the generation of more realistic and physically feasible 3D human-object interaction actions.
ViKey: Enhancing Temporal Understanding in Videos via Visual Prompting
Yeonkyung Lee (Yonsei University), Seong Jae Hwang (Yonsei University)
CodeTransformerPrompt EngineeringVision Language ModelVideo
π― What it does: Propose the VIKEY framework, which enhances the temporal reasoning capabilities of existing VideoLLMs without additional training by overlaying explicit frame numbering (e.g., 'frame #01') as visual prompts on each video frame, combined with keyword-frame mapping (KFM) technology.
ViRC: Enhancing Visual Interleaved Mathematical CoT with Reason Chunking
Lihong Wang (Jilin University), Zhe Li (Ant Digital Technologies, Ant Group)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelMultimodalityBenchmarkChain-of-Thought
π― What it does: Proposed the VIRC framework, which decomposes multimodal mathematical reasoning into key reasoning units (CRUs) using the 'Reason Chunking' mechanism, and constructed the CRUX dataset along with a three-stage training strategy;
VIRO: Robust and Efficient Neuro-Symbolic Reasoning with Verification for Referring Expression Comprehension
Hyejin Park (Pohang University of Science and Technology), Jungseul Ok (Pohang University of Science and Technology)
CodeObject DetectionDepth EstimationLarge Language ModelVision Language ModelImageVideoTextMultimodality
π― What it does: Propose the VIRO (Verification-Integrated Reasoning Operators) framework, which embeds a lightweight verifier in each step of neuro-symbolic reasoning to explicitly handle no-target scenarios.
π― What it does: Propose the CodeBrain framework, which performs a unified 'virtual full-stack scanning' to complete missing modalities in brain MRI;
Virtual Immunohistochemistry Staining with Dual-Aligned Multi-Task Feature Guidance
Shigeng Xie (Dalian University Of Technology), Fengyu Cong (Dalian University Of Technology)
CodeImage TranslationGenerationGenerative Adversarial NetworkContrastive LearningBiomedical Data
π― What it does: This paper proposes a virtual immunohistochemistry (VIS) model based on bidirectional alignment multi-task features, which significantly improves the quality of virtual staining by utilizing auxiliary task features to guide the generator network at the feature level.
Cheng Shi (University of Hong Kong), Sibei Yang (Sun Yat-sen University)
CodeObject DetectionSegmentationTransformerImage
π― What it does: Proposes LaSt-ViT, a frequency-aware selective aggregation mechanism that utilizes CLS tokens to aggregate stable foreground patches, thereby eliminating the 'lazy aggregation' and sparse feature anomalies that appear in ViT under different supervision modes;
CodeRetrievalTransformerVision Language ModelContrastive LearningMultimodality
π― What it does: Achieve lifetime person re-identification (LReID) using pre-trained vision-language models (VLM), enhancing cross-domain knowledge transfer and memory retention by explicitly separating global and local attributes.
Vision-Language Model Guided Source-Free Domain Adaptation via Optimal Transport
Shuo Han (Xidian University), Xiangrong Zhang (Xidian University)
CodeClassificationDomain AdaptationKnowledge DistillationVision Language ModelContrastive LearningImageMultimodality
π― What it does: Leverage pre-trained vision-language models (VLM) as external semantic priors, combining optimal transport (OT) alignment with bidirectional distillation to achieve source-agnostic domain adaptation.
π― What it does: Proposed a visual memory framework called VisMem based on cognitive theory, which injects short-term perceptual memory and long-term semantic memory into visual language models, and dynamically inserts latent visual memories into the autoregressive inference process through special memory invocation tokens.
VISTA: A Test-Time Self-Improving Video Generation Agent
Do Xuan Long (Google), Sercan Γ. Arik (Google)
CodeGenerationTransformerLarge Language ModelAgentic AIPrompt EngineeringVideoTextMultimodalityAudio
π― What it does: Proposed an adaptive multi-agent framework called VISTA, which achieves self-improvement in text-to-video generation by decomposing user prompts, generating candidate videos, performing multi-dimensional critique on the best video, and automatically rewriting prompts during inference.
π― What it does: Proposed a zero-shot anomaly detection framework called VisualAD that uses only visual Transformers. It freezes the ViT and inserts learnable abnormal and normal global tokens into its input sequence. These tokens interact across multiple layers of features through spatially aware cross-attention (SCA) and self-alignment function (SAF), ultimately generating pixel-level anomaly maps and image-level anomaly scores.
VITAL: Vision-Encoder-centered Pre-training for LMMs in Visual Quality Assessment
Ziheng Jia (Shanghai Jiaotong University), Xiongkuo Min (Shanghai Jiaotong University)
CodeTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageVideoMultimodality
π― What it does: Constructed a large-scale machine-annotated dataset containing 458 million visual-language pairs, and developed the VITAL series of multimodal large models centered on visual encoders for visual quality assessment.
ViTPrompt: Training-Free Prompt Refinement with Visual Tokens for Open-Vocabulary Detection
Yitong Qin (University of Electronic Science and Technology of China), Heng Tao Shen (University of Electronic Science and Technology of China)
CodeObject DetectionTransformerPrompt EngineeringVision Language ModelImage
π― What it does: Propose ViTPrompt, a training-free, two-phase inference open-domain object detection adaptation framework that leverages RoI visual tokens from initial high-confidence detections to enhance text prompts, jointly improving localization and classification.
VKG-QA: Visual Knowledge Graph-based Question Answer for Large Multimodal Models
Yuntao Du (Shandong University), Lizhen Cui (Shandong University)
CodeTransformerLarge Language ModelImageGraphBenchmarkChain-of-Thought
π― What it does: Proposes a visual knowledge graph-based question answering benchmark, VKG-QA, systematically evaluating the ability of large-scale multimodal models in visual structured knowledge reasoning.
VLM4RSDet: Collaborative Optimization with Vision-Language Model for Enhancing Remote Sensing Object Detection
Shuohao Shi (National University of Defense Technology), Xin Xu (National University of Defense Technology)
CodeObject DetectionConvolutional Neural NetworkTransformerVision Language ModelImage
π― What it does: Propose a collaborative optimization framework called VLM4RSDet, which jointly trains a traditional closed-set remote sensing object detector with a vision-language model. During inference, only the detector is retained to avoid additional computational costs.
VQRAE: Representation Quantization Autoencoders for Multimodal Understanding, Generation and Reconstruction
Sinan Du (Tsinghua University), Chun Yuan (Tsinghua University)
CodeRestorationGenerationTransformerVision Language ModelAuto EncoderMultimodality
π― What it does: Propose VQRAE, a vector-quantized representation autoencoder that achieves unified visual representation generation, understanding, and reconstruction;
π― What it does: Propose an end-to-end joint video super-resolution and low-light enhancement framework called VSRELL, which can directly restore low-light, low-resolution videos into high-resolution normally illuminated sequences.
π― What it does: Proposed a new visual autoregressive generation acceleration framework called VVS, which significantly reduces the number of forward passes of the target model by partially skipping verification steps during inference.
π― What it does: Proposed the WaDi framework, which significantly improves single-step text generation speed and quality through weight-direction-aware distillation of first-order diffusion models using the low-rank rotation adapter LoRaD.
CodeAutonomous DrivingSupervised Fine-TuningReinforcement LearningVision Language ModelFlow-based ModelImageTextMultimodality
π― What it does: WAM-Flow treats vehicle trajectory planning as a discrete flow matching problem, achieving a tunable coarse-to-fine reasoning process through parallel denoising.
Wavelet-based Frame Selection by Detecting Semantic Boundary for Long Video Understanding
Wang Chen (Xiamen University), Xiawu Zheng (Xiamen University)
CodeVision Language ModelVideo
π― What it does: For long video understanding, an untrained wavelet transform framework named WFS-SB is proposed, achieving frame selection by detecting semantic boundaries
WebGym: Scaling Training Environments for Long-Horizon Visual Web Agents with Realistic Tasks
Hao Bai (Microsoft), Spencer Whitehead (Microsoft)
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
π― What it does: Built WebGym, an open-source training environment containing nearly 300,000 real web tasks, supporting evaluation, and asynchronous rollout.
π― What it does: Created and released the controllable human video generation benchmark WYD, providing fine-grained multi-class annotations and corresponding evaluation frameworks
What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models
Yingqi Fan (Eastern Institute of Technology), Xiaoyu Shen (Eastern Institute of Technology)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerContrastive LearningMultimodality
π― What it does: This study constructs the EmbedLens probe framework to perform fine-grained analysis of visual tokens in multi-modal large language models (MLLMs), revealing that visual tokens can be categorized into three types: sink, dead, and alive, and demonstrating that alive tokens carry most of the image semantics. Further quantitative experiments show that visual self-attention and feed-forward networks inside MLLMs contribute little to most tasks, and shallow processing has minimal impact on visual tokens, suggesting that visual inputs should be directly injected into intermediate layers.
π― What it does: This paper studies how to find the optimal ranking metric between precision and recall in binary classification problems. The authors prove that all FΞ² scores (i.e., weighted harmonic means) can produce meaningful rankings and construct a set of rankings along a 'shortest path.' Subsequently, they define and compute the optimal Ξ² value using Karcher mean and Kendall distance, and provide a closed-form solution.
π― What it does: This paper addresses the cross-sparse correspondence problem between sparse lines and texture-rich images, proposing the SFA-DIFT framework that achieves high-precision correspondence by fusing spatial and frequency domain alignment.
π― What it does: Proposes a generic transferable attack framework named UPA-RFAS for vision-language-action (VLA) models, and verifies its efficiency under black-box, cross-model, cross-task, and cross-domain settings.
π― What it does: This paper proposes a novel Any-Scenario ReID (AS-ReID) task, constructs a large-scale multispectral aerial-ground human image dataset WHU-MARS, and designs a Unified Alignment and Discrimination (UAD) framework based on a single network to achieve unified retrieval across modalities and viewpoints.
Why Does RL Generalize Better Than SFT? A Data-Centric Perspective on VLM Post-Training
Aojun Lu (Sichuan University), Yanan Sun (Sichuan University)
CodeComputational EfficiencyData-Centric LearningSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodality
π― What it does: Investigated the fundamental reasons for the generalization gap between reinforcement learning (RL) and supervised fine-tuning (SFT) in vision-language models (VLMs), and proposed a Difficulty-Curated SFT (DC-SFT) method based on difficulty screening;
Why Not Hyperparameter-Friendly Optimisation? A Monotonic Adaptive Norm Rescaling Approach For Long-Tailed Recognition
Shuo Zhang (University of Oxford), Tingting Zhu (University of Oxford)
CodeClassificationRecognitionImage
π― What it does: This paper proposes an adaptive monotonic normalization (SAMN) method, which addresses the weight scale imbalance problem in long-tailed recognition by directly applying monotonic constraints on the weight norms of each class during two-stage retraining.
WISER: Wider Search, Deeper Thinking, and Adaptive Fusion for Training-Free Zero-Shot Composed Image Retrieval
Tianyue Wang (Chinese Academy of Sciences), Jinqiao Wang (Chinese Academy of Sciences)
CodeRetrievalLarge Language ModelVision Language ModelMultimodalityBenchmark
π― What it does: Proposes a training-agnostic zero-shot compositional image retrieval framework called WISER, which enhances retrieval quality by fusing T2I and I2I retrieval paths through a retrieval-validate-improve loop.
π― What it does: Constructed WiTTA-Bench, a systematic benchmark for evaluating test-time adaptation (TTA) methods in WiFi human activity recognition (HAR), providing two adaptation protocols (OTTA and TTDA), 20 representative methods, three categories of physics-induced domain shifts (cross-environment, cross-subject, cross-device), and releasing a new dataset, WiHAR-Dual, to cover cross-device scenarios.
π― What it does: Propose the WorldStereo framework, leveraging two geometric memories (Global Geometric Memory GGM and Spatial Stereo Memory SSM) to achieve high-quality video generation from single views and consistent 3D reconstruction.
CodePose EstimationDepth EstimationConvolutional Neural NetworkGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
π― What it does: A zero-shot training framework for aligning ground images to satellite images called Wrivinder was constructed. It utilizes multi-view ground photos to reconstruct sparse geometry via SfM, then densifies it using 3D Gaussian Splatting, and generates a consistent zenith view through semantic ground planes and monocular depth scale information. Finally, it achieves geometrically accurate GPS positioning by performing self-supervised depth template matching with satellite images.
X-PCR: A Benchmark for Cross-modality Progressive Clinical Reasoning in Ophthalmic Diagnosis
Gui Wang (Shenzhen University), Linlin Shen (Shenzhen University)
CodeTransformerLarge Language ModelVision Language ModelImageMultimodalityBiomedical DataBenchmarkChain-of-Thought
π― What it does: Proposed the X-PCR benchmark, using a six-stage progressive reasoning chain and cross-modal reasoning tasks to evaluate the capabilities of multimodal large language models in ophthalmic diagnosis.
π― What it does: Propose X-WIN, a chest X-ray world model that achieves 3D CT knowledge distillation into 2D X-ray representations by predicting CT projections in the latent space.
π― What it does: Proposes the YOSE framework, achieving efficient video object removal by processing only masked regions based on the Diffusion Transformer;
π― What it does: Proposed the Energy-based Joint Distribution Adversarial Training (EB-JDAT) method, which unifies the tasks of classification, robustness, and generation.
π― What it does: Propose the Pixel-Equivalent Latent Compositing (PELC) principle and implement a lightweight Transformer-based DecFormer to perform pixel-equivalent mixing in the latent space within Diffusion models, enhancing the boundary quality of masked editing and inpainting.
π― What it does: This paper proposes Yume1.5, a long-sequence interactive virtual world video model that can be controlled via keyboard, generate text instructions, and perform real-time inference.
Zoo3D: Zero-Shot 3D Object Detection at Scene Level
Andrey Lemeshko (Higher School Of Economics), Maksim Kolodiazhnyi (Lomonosov Moscow State University)
CodeObject DetectionComputational EfficiencyTransformerVision Language ModelImagePoint Cloud
π― What it does: Proposed two zero-training/self-supervised open-vocabulary 3D object detection frameworks, Zoo3D0 and Zoo3D1, capable of directly performing object detection on point clouds, images with known or unknown camera poses.