IEEE/CVF Conference on Computer Vision and Pattern Recognition Β· 1047 papers
See What I Mean: Aligning Vision and Language Representations for Video Fine-grained Object Understanding
Boyuan Sun (Nankai University), Qibin Hou (Nankai University)
CodeObject DetectionSegmentationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityBenchmark
π― What it does: Constructed the NL-Refer dataset and performed cross-attention alignment training on multimodal large language models, enabling the model to precisely locate target objects in videos based solely on text prompts.
See What We Cannot See: A Geo-guided Reasoning Benchmark for Object Counting under Adverse Earth Observation Conditions
Jiayi Wang (Wuhan University), Zhenzhong Chen (Wuhan University)
CodeObject DetectionTransformerLarge Language ModelImageMultimodalityPoint CloudBenchmark
π― What it does: Proposed a large-scale remote sensing object counting dataset called GROC, constructed a controllable degradation synthesis and interactive annotation engine, and further developed a Geo-guided Reasoning Agent based on a multi-modal large model;
CodeRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringVision-Language-Action ModelMultimodalityBenchmark
π― What it does: Built a binary switch instruction state control benchmark and proposed the StaR multimodal reasoning method, enabling agents to first perceive the current state, infer the target state, and then determine whether a switch is needed when executing switch operations;
Seeing Both Sides: Towards Bidirectional Semantic Alignment for Open-Vocabulary Camouflaged Object Segmentation
Guohui Zhang (Dalian Minzu University), Fasheng Wang (Dalian Minzu University)
CodeSegmentationConvolutional Neural NetworkTransformerPrompt EngineeringVision Language ModelContrastive LearningImage
π― What it does: Proposed the BaCLIP framework, which achieves open-vocabulary camouflage object segmentation through bidirectional semantic alignment;
π― What it does: This paper proposes a frequency-guided self-supervised monocular depth estimation framework that considers the complementarity between depth and pose networks;
Seeing is Improving: Visual Feedback for Iterative Text Layout Refinement
Junrong Guo (University of Science and Technology of China), Hongtao Xie (University of Science and Technology of China)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodality
π― What it does: Propose a self-improving text layout generation framework VFLM, combining a generate-render-reflect-modify cycle, leveraging visual feedback to achieve iterative optimization.
Seeing through Light and Darkness: Sensor-Physics Grounded Deblurring HDR NeRF from Single-Exposure Images and Events
Yunshan Qi (Beihang University), Jia Li (Beihang University)
CodeRestorationGenerationNeural Radiance FieldImageMultimodalityPhysics Related
π― What it does: A unified perceptual physics-driven NeRF framework called See-NeRF is constructed, achieving HDR scene deblurring and novel view synthesis by leveraging single-exposure blurred LDR images and corresponding event data.
SegEarth-R2: Towards Comprehensive Language-guided Segmentation for Remote Sensing Images
Zepeng Xin (Xi'an Jiaotong University), Xiangyong Cao (Xi'an Jiaotong University)
CodeSegmentationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Proposed a new large language-driven remote sensing image segmentation framework called SegEarth-R2, and constructed the LaSeRS dataset covering four dimensions: hierarchical fine-grainedness, target diversity, inference requirements, and linguistic variation
SegMo: Co-Designing Content-Aware Sparsity and Locally-Cohesive Segment Parallelism for Efficient VLM Inference
Haojuan Li (Shanghai Jiao Tong University), Jiaqi Wang (Shanghai Innovation Institute)
CodeComputational EfficiencyVision Language ModelVideo
π― What it does: Proposes the SegMo framework, which jointly optimizes long video VLM inference through content-aware sparsification and localized coherent segment parallelization.
Selection-as-Nonlinearity: Bridging Attention and Activation via a Joint Game-Decision Lens for Interpretable, Discriminative Visual Representations
Sudong Cai (Hong Kong Polytechnic University), Bing Wang (Hong Kong Polytechnic University)
CodeClassificationObject DetectionExplainability and InterpretabilityTransformerImage
π― What it does: Proposes the 'Select-as-Nonlinearity' (SaN) explanation framework to reveal why self-attention performs poorly without FFN, and based on this designs the CSaN mechanism, which enhances attention expressiveness through hierarchical budget calibration and public-private readout while maintaining standard attention regularization.
π― What it does: Proposes a training-free encoder editing method based on Contrastive Subnet Erasure (CSE), which removes specified class concepts in visual models without affecting other model capabilities.
π― What it does: Proposed a self-critical distillation network (SCD-Net) for generating commonsense descriptions of videos, addressing the weaknesses of traditional models in visual-semantic alignment and the lack of mutual reasoning among different commonsense categories.
π― What it does: Proposed the SAVN-CE task, i.e., semantic audio-visual navigation in continuous 3D environments, and implemented continuous reasoning and navigation toward targets through MAGNet;
π― What it does: Proposes Semantic-First Diffusion (SFD), which in a three-stage asynchronous denoising process first generates semantic latent representations and then guides texture generation, significantly accelerating training and improving image quality.
π― What it does: Propose a semi-supervised medical image segmentation framework called SemiGDA based on generative adversarial distribution alignment. It uses dual encoders to align the latent distributions of images and segmentation masks, and enhances multi-scale semantic consistency through a jump adapter, significantly improving segmentation performance in low-annotation scenarios.
π― What it does: Propose a dynamic caching framework called SenCache, which determines whether to cache based on the local sensitivity of the network to input perturbations (noise latent variables and time steps), to accelerate diffusion model inference.
SenseSearch: Empowering Vision-Language Models with High-Resolution Agentic Search-Reasoning via Reinforcement Learning
Yong Xien Chng (SenseTime Research), Lewei Lu (SenseTime Research)
CodeRetrievalSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Built SenseSearch, an agentic visual-language model capable of adaptively invoking image search, text search, and image cropping tools during multi-round reasoning.
Yuqiao He (Hunan University), Wenbin He (Hunan University)
CodeRestorationOptimizationComputational EfficiencyConvolutional Neural NetworkImagePhysics Related
π― What it does: A self-supervised geometric degradation estimation framework named SGDE is studied for correcting reconstruction distortions caused by mask mismatch in CASSI systems.
π― What it does: Propose Structured 2D Gaussians (SGI), generating local 2D Gaussian primitives through multi-scale seed regions and lightweight MLP to achieve efficient and compact representation of high-resolution images.
Shedding Light on VLN Robustness: A Black-box Framework for Indoor Lighting-based Adversarial Attack
Chenyang Li (Nanyang Technological University), Yang Liu (Nanyang Technological University)
CodeAdversarial AttackMultimodality
π― What it does: Propose a black-box adversarial attack framework based on indoor lighting (ILA), which adjusts global illumination intensity or switches lights on/off at critical moments to evaluate and undermine the robustness of vision-and-language navigation (VLN) models.
π― What it does: Propose the SIGMA unified framework, which supports multi-attribute cross-insertion in text and image conditions, enabling controllable multi-source image generation.
π― What it does: Propose the SimLBR scheme, which learns a compact decision boundary for real images by mixing a small amount of pseudo-fake information with real images in the DINOv3 latent space;
Simple-ViLMedSAM: Simple Text Prompts Meet Vision-Language Models for Medical Image Segmentation
Chengcan Qian (Nanjing University of Aeronautics and Astronautics), Xuyun Wen (Nanjing University of Aeronautics and Astronautics)
CodeSegmentationTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningImageBiomedical DataComputed Tomography
π― What it does: Propose a medical image segmentation framework called Simple-ViLMedSAM based on CLIP and SAM, which can perform zero-shot and few-shot segmentation using only simple text labels.
π― What it does: Proposes the SinGeo framework, achieving robust cross-view geolocation with a single model across different perspectives and FoV through Dual Discriminative Learning and Curriculum Learning.
SIR: Structured Image Representations for Explainable Robot Learning
Paul Mattes (Karlsruhe Institute of Technology), Rudolf Lioutikov (Karlsruhe Institute of Technology)
CodeExplainability and InterpretabilityRepresentation LearningRobotic IntelligenceGraph Neural NetworkTransformerVision-Language-Action ModelImagePoint Cloud
π― What it does: Proposes the SIR method, which converts images into scene graphs (Scene Graph), generates task-related subgraphs through learnable sparsification, and inputs the subgraph as an intermediate state into the GCIL policy, thereby achieving interpretable robot learning.
π― What it does: This paper investigates the memorization problem in diffusion models, proposing a theoretical framework that explains memorization as arising from the sharpness of the score function of the empirical distribution, and enhancing generalization through smoothing techniques.
SO-Bench: A Structural Output Evaluation of Multimodal LLM
Di Feng (Apple), Afshin Dehghan (Apple)
CodeTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: Constructed SO-Bench to evaluate the ability of multimodal LLMs in visual structured output tasks and enhanced model performance through training.
CodeExplainability and InterpretabilityRobotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningMixture of ExpertsFlow-based ModelImageVideoTextMultimodalityChain-of-Thought
π― What it does: Proposed a hierarchical foundation model called SocialNav, which integrates high-level semantic reasoning with low-level trajectory generation to achieve social, interpretable robot navigation.
π― What it does: Developed a dynamic acceleration framework called SODA based on fine-grained sensitivity-aware mechanisms, combining caching and pruning techniques to significantly improve inference efficiency while maintaining or enhancing the generation quality of Diffusion Transformers.
Solvability of the Viewing Graph Under the Affine Camera Model
Gabriele Pedroni (Politecnico di Milano), Federica Arrigoni (Politecnico di Milano)
CodeGraph
π― What it does: This paper studies the solvability of the viewing graph under the affine camera model, and provides its linear system formulation and decision algorithm;
SOTA: Self-adaptive Optimal Transport for Zero-Shot Classification with Multiple Foundation Models
Zhanxuan Hu (Yunnan Normal University), Huafeng Li (Kunming University of Science and Technology)
CodeClassificationVision Language ModelImageMultimodalityBiomedical DataBenchmark
π― What it does: This paper proposes a training-agnostic adaptive optimal transport framework called SOTA, which can fuse outputs from various base models (VLMs and VFM) to enhance zero-shot classification performance.
SpaceDrive: Infusing Spatial Awareness into VLM-based Autonomous Driving
Peizheng Li (Mercedes-Benz AG), Andreas Zell (University of Tbingen)
CodeDepth EstimationAutonomous DrivingTransformerLarge Language ModelVision Language ModelImagePoint Cloud
π― What it does: Propose SpaceDrive, an end-to-end autonomous driving framework that integrates 3D spatial encoding into VLM, directly replacing digital tokens with a unified 3D position encoding for regression prediction;
π― What it does: Leverage teacher-student distillation to transfer high-resolution ViT inference capabilities from a sliding-window-based approach to a single-channel ViT without architectural modifications, enabling efficient dense feature extraction at arbitrary resolutions for open-vocabulary segmentation.
CodePose EstimationComputational EfficiencySupervised Fine-TuningSimultaneous Localization and MappingImage
π― What it does: Propose ON3R, an online-trained neural 3D regressor that utilizes sparse matching to predict 3D coordinates and solves camera poses in sparse views via P3P-RANSAC
Spatial Matters: Position-Guided 3D Referring Expression Segmentation
Yabing Wang (Xi'an Jiaotong University), Sanping Zhou (Xi'an Jiaotong University)
CodeSegmentationConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningTextPoint Cloud
π― What it does: Propose Position3D, an end-to-end method capable of precisely segmenting target objects in point cloud scenes based on natural language descriptions.
π― What it does: Introduce a spatial retrieval paradigm, incorporating offline geographic images (e.g., Google Maps street view/satellite images) as additional input to enhance multi-task autonomous driving performance.
Spatiotemporal Pyramid Flow Matching for Climate Emulation
Jeremy A. Irvin (Stanford University), Duncan Watson-Parris (University of California, San Diego)
CodeData SynthesisTransformerFlow-based ModelTime SeriesPhysics Related
π― What it does: This paper proposes a novel spatiotemporal pyramid flow matching model (SPF) that can efficiently generate climate simulation results in parallel across multiple timescales, supporting direct sampling at different spatial and temporal resolutions;
Specificity-aware reinforcement learning for fine-grained open-world classification
Samuele Angheben (University of Trento), Yiming Wang (Fondazione Bruno Kessler)
CodeClassificationLarge Language ModelReinforcement LearningVision Language ModelImageMultimodality
π― What it does: Propose SpeciaRL, a specificity-aware reinforcement learning framework designed to enhance model specificity in open-world fine-grained image classification tasks while maintaining accuracy.
Spectral Super-Resolution via Adversarial Unfolding and Data-Driven Spectrum Regularization: From Multispectral Satellite Data to NASA Hyperspectral Image
Si-Sheng Young (National Cheng Kung University), Chia-Hsiang Lin (National Cheng Kung University)
π― What it does: Proposed an algorithm to convert Sentinel-2 multispectral images into NASA AVIRIS-level hyperspectral images, achieving joint reconstruction of spectral super-resolution from 12 bands to 186 bands and 5ΞΌm spatial resolution.
π― What it does: By periodically shrinking Gaussian scales (scale reset) during the 3D Gaussian Splatting (3DGS) training process and applying entropy constraints to Ξ± blending weights, the Gaussian list per pixel is significantly shortened, accelerating learning; combined with a resolution scheduler further enhances efficiency.
π― What it does: Proposed a RGB-to-RAW conversion method called SpiralDiff based on diffusion models, achieving a cross-camera unified model through signal-dependent noise weighting and CamLoRA adapter.
SplitFlux: Learning to Decouple Content and Style from a Single Image
Yitong Yang (Shanghai University of Finance and Economics), Shuting He (Shanghai University of Finance and Economics)
CodeGenerationSupervised Fine-TuningVision Language ModelDiffusion modelImageText
π― What it does: For a single image, we propose SplitFlux, which achieves decoupling of image content and style by fine-tuning the single-stream blocks in the Flux model, and allows the decoupled content to be re-embedded into any new context.
SSM-Aware Token-Efficient VMamba via Adaptive Patch Pruning and Merging for Person Re-Identification
Huiyuan Huang (Kookmin University), Sang Min Yoon (Kookmin University)
CodeRecognitionComputational EfficiencyImage
π― What it does: Proposes an efficient person re-identification framework called TE-VMamba based on a variable state space model, which significantly reduces the number of tokens while maintaining recognition accuracy.
π― What it does: Train a multi-task model to learn seven different pixel-level tasks on partially labeled synthetic data and achieve cross-domain generalization on real datasets.
π― What it does: Propose a lightweight, plug-and-play identity-preserving video generation framework called Stand-In, which achieves identity control by adding a conditional image branch to a pre-trained video diffusion model, generating high-quality videos consistent with the text while keeping the reference image unchanged.
π― What it does: Proposes STARFlow-V, an end-to-end video generation framework based on autoregressive regularized flows, which adopts a global-local architecture, flow score matching denoising, and efficient sampling via Jacobi iteration, enabling multi-task generation (T2V, I2V, V2V) and streaming generation under text, image, and video conditions.
π― What it does: This paper proposes a monocular video 3D head avatar reconstruction method called STAvatar based on 3D Gaussian Splatting, which can generate high-fidelity, animatable head avatars from monocular video.
STCast: Adaptive Boundary Alignment for Global and Regional Weather Forecasting
Hao Chen (Hong Kong University of Science and Technology), Lei Bai (Shanghai AI Laboratory)
CodeTransformerMixture of ExpertsTime SeriesPhysics Related
π― What it does: Propose the STCast framework, combining adaptive regional boundaries and dynamic monthly expert allocation to unify global and regional weather prediction.
π― What it does: This paper proposes an infrared-visible video fusion method based on a streaming diffusion model, which can achieve real-time inference while maintaining high quality.
π― What it does: Propose a single-model SL-HOI to achieve one-stage open-vocabulary human-object interaction detection, utilizing the DINOv3 backbone for localization and the vision head for interaction classification.
Structural Graph Probing of Vision-Language Models
Haoyu He (Northeastern University), Qi R. Wang (Northeastern University)
CodeExplainability and InterpretabilityGraph Neural NetworkVision Language ModelContrastive LearningMultimodality
π― What it does: This paper constructs a neural correlation graph (neural topology) for each layer to study group-level interactions and structures within VLMs, and employs graph models for behavior prediction, structural alignment, and causal intervention analysis.
π― What it does: Designed and implemented a camera-based 3D object detection framework called STUR3D, which projects and propagates 2D and 3D detection results from time series, injects geometric depth information into the 2D detection head, and finally maps the improved 2D results to precise 3D queries for detection.
Subspace Alignment for CLIP-based Continual Learning via Canonical Correlation Analysis
Huan Zhang (Wuhan University), Fan Lyu (Computer Vision Center Universitat Autonoma De Barcelona)
CodeRepresentation LearningSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Propose a subspace alignment framework called CCA-CL based on Canonical Correlation Analysis to address the visual-textual asymmetric drift problem that occurs in CLIP during continuous learning.
π― What it does: Propose a sub-cloud based point cloud registration framework SuP, transforming low-overlap registration into a problem of mining high-overlap sub-cloud pairs;
Superman: Unifying Skeleton and Vision for Human Motion Perception and Generation
Xinshun Wang (Peking University), Mengyuan Liu (Peking University)
CodeGenerationPose EstimationTransformerLarge Language ModelMultimodalitySequential
π― What it does: Propose a unified generative framework called Superman, which can simultaneously accomplish 3D pose estimation, motion prediction, and motion interpolation;
π― What it does: This paper analyzes the non-semantic noise that occurs in masked image modeling (MIM) self-supervised learning and proposes a post-processing method called SOAP (Semantically Orthogonal Artifact Projection) to suppress this noise, thereby improving the performance of downstream tasks under zero-shot scenarios.
SurgCoT: Advancing Spatiotemporal Reasoning in Surgical Videos through a Chain-of-Thought Benchmark
Gui Wang (Shenzhen University), Linlin Shen (Shenzhen University)
CodeObject DetectionObject TrackingTransformerLarge Language ModelVideoMultimodalityBenchmarkChain-of-Thought
π― What it does: This paper proposes SurgCoT, a fine-grained spatiotemporal reasoning benchmark spanning seven surgical categories and 35 surgical procedures, constructed through three-stage chain reasoning and quintuple annotations (Question, Options, Knowledge, Clues, Answer) to build structured reasoning chains.
SVHalluc: Benchmarking Speech-Vision Hallucination in Audio-Visual Large Language Models
Chenshuang Zhang (KAIST), Tae-Hyun Oh (KAIST)
CodeAnomaly DetectionLarge Language ModelVideoMultimodalityBenchmarkAudio
π― What it does: Proposed the SVHalluc benchmark to assess the hallucination behavior of audio-visual large language models during speech and vision alignment, and evaluate existing models.
π― What it does: This paper proposes SWIFT, a few-shot untrained video attribution method based on sliding window reconstruction, for identifying whether a video originates from a specified generative model;
Symphony: A Cognitively-Inspired Multi-Agent System for Long-Video Understanding
Haiyang Yan (Institute of Automation, Chinese Academy of Sciences), Mengyi Liu (Kuaishou Technology)
CodeRecognitionRetrievalTransformerLarge Language ModelAgentic AIVision Language ModelVideoRetrieval-Augmented Generation
π― What it does: Designed and implemented a multi-agent system called Symphony for long video understanding, simulating human cognition for task decomposition and collaboration
Synergistic Bleeding Region and Point Detection in Laparoscopic Surgical Videos
Jialun Pei (Chinese University of Hong Kong), Pheng-Ann Heng (Chinese University of Hong Kong)
CodeObject DetectionSegmentationTransformerOptical FlowVideoBiomedical Data
π― What it does: Proposes a dual-task online detection framework for simultaneous detection of blood cavity regions and blood spots in laparoscopic surgery videos, and constructs the first public blood cavity detection dataset SurgBlood.
Tackling Alignment Ambiguity in Person Retrieval through Conversational Attribute Mining
Hao Zou (University Of Electronic Science And Technology Of China), Mingzhu Cai (University Of Electronic Science And Technology Of China)
CodeRetrievalTransformerLarge Language ModelVision Language ModelMultimodality
π― What it does: This paper proposes the CECA framework, which utilizes multimodal large language models for conversational attribute mining to enhance the alignment between text and images with fine-grained attribute information, thereby improving the performance of text-to-image person retrieval.
TALO: Pushing 3D Vision Foundation Models Towards Globally Consistent Online Reconstruction
Fengyi Zhang (University of Queensland), Yadan Luo (University of Queensland)
CodeAutonomous DrivingOptimizationSimultaneous Localization and MappingVideoPoint Cloud
π― What it does: To maintain global geometric consistency across submaps in online 3D vision foundation models (e.g., VGGT, Ο3, MapAnything), this paper proposes TALO, an alignment framework based on Thin Plate Spline (TPS) and global control points, and introduces a point-agnostic submap registration method based on camera poses.
π― What it does: To address the challenge of identifying known classes and discovering unknown classes in online unlabeled streams, this paper proposes the TALON framework for test-time adaptive learning, which continuously updates the feature extractor and class prototypes during inference to enable immediate learning of newly emerging classes.
π― What it does: Proposed a task-oriented infrared imaging framework that models infrared image recovery as a single-step diffusion process, combining dynamic clock estimation with spectral regularization to achieve high-quality visual recovery and semantic consistency;
Taming Sampling Perturbations with Variance Expansion Loss for Latent Diffusion Models
Qifan Li (University of Electronic Science and Technology of China), Shuhang Gu (University of Electronic Science and Technology of China)
CodeGenerationDiffusion modelAuto EncoderImage
π― What it does: Propose a 'Variance Expansion Loss' (VE Loss) that constructs a more robust latent space against sampling perturbations without relying on KL regularization, thereby improving the generation quality of latent diffusion models.
TAPE: Task-Adaptive Prototype Evolution in Audio-Language Models for Fully Few-shot Class-incremental Audio Classification
Yunlong Gao (Dalian University of Technology), Xinyue Liu (Dalian University of Technology)
CodeClassificationRecognitionTransformerVision Language ModelAudio
π― What it does: The study proposes a task-adaptive prototype evolution framework based on an audio-language model to address catastrophic forgetting and overfitting in fully few-shot incremental audio classification tasks.
π― What it does: Designed a lightweight task-aware RAW-to-RGB processing framework called TA-ISP, which adaptively reconstructs RAW images at global, regional, and pixel levels using a multi-scale pixel modulation module to provide optimized representations for pre-trained visual models.
TaskIT: Memory-Efficient Fine-Tuning of Multi-LoRA LLMs via Cross-Task Importance Transfer
Cheng Fang (City University of Hong Kong), Bin Guo (Northwestern Polytechnical University)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningImageTextMultimodalityAudio
π― What it does: This paper proposes the TaskIT framework, which addresses sparse fine-tuning of multiple LoRA LLMs on memory-constrained mobile/edge devices. It leverages cross-task importance transfer to predict the significance of uninserted modules and constructs an accurate memory prediction model through block-level activation dependency analysis. This enables optimal selection of LoRA module positions, quantities, and ranks under a given memory budget, achieving efficient model fine-tuning.
Taxonomy-Aware Representation Alignment for Hierarchical Visual Recognition with Large Multimodal Models
Hulingxiao He (Peking University), Yuxin Peng (Peking University)
CodeRecognitionLarge Language ModelReinforcement LearningVision Language ModelContrastive LearningMultimodality
π― What it does: Proposes a framework named TARA that aligns visual and label representations within large multimodal models (LMM) by leveraging biological foundation models (BFM) to achieve hierarchical visual recognition.
π― What it does: This work injects 3D geometry awareness into the DINOv3 visual foundation model through a self-supervised LoRA method, achieving more robust feature extraction for partial 3D shapes.
π― What it does: This paper proposes an algorithm called TESO for online tracking of the extrinsic and intrinsic parameters of stereo cameras, which uses random optimization on the essential matrix manifold and a robust essential matrix error based on kernel correlation to real-time correct camera rotation drift.
TESSERA: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis
Zhengpeng Feng (University of Cambridge), Srinivasan Keshav (University of Cambridge)
CodeClassificationSegmentationRepresentation LearningRecurrent Neural NetworkTransformerContrastive LearningMultimodalityTime SeriesAgriculture Related
π― What it does: Proposes TESSERAβa pixel-level multimodal (Sentinel-1/2) time series foundation model that generates efficient embeddings for sparse satellite observations using self-supervised learning;
Fengyi Zhang (University of Queensland), Yadan Luo (University of Queensland)
CodeAutonomous DrivingGaussian SplattingSimultaneous Localization and MappingOptical FlowImagePoint Cloud
π― What it does: Proposes a 3D occupancy prediction framework TT-Occ that generates time-aware 3D Gaussians using visual foundation models during testing and supports voxelization at arbitrary resolutions.
Test-Time Alignment of Text-to-Image Diffusion Models via Null-Text Embedding Optimisation
Taehoon Kim (University of Edinburgh), Timothy Hospedales (University of Edinburgh)
CodeGenerationDiffusion modelImageText
π― What it does: Propose a framework called Null-TTA that achieves alignment in text-to-image diffusion models during inference by optimizing unconditional text embeddings (null-text).
π― What it does: Proposes the task TE2A, which performs instant adaptation for action prediction between egocentric and exocentric views during testing, and introduces the Dual-Clue enhanced Prototype Growing Network (DCPGN) scheme
Test-Time Multi-Prompt Adaptation for Open-Vocabulary Remote Sensing Image Segmentation
Ting Yang (Tianjin University), Qinghua Hu (Tianjin University)
CodeSegmentationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImage
π― What it does: Propose a test-time multi-prompt adaptation method (TMPA), which generates diverse context-aware text descriptions and adaptively calibrates prompt embeddings during inference to alleviate text ambiguity in open-vocabulary remote sensing image semantic segmentation, thereby improving segmentation accuracy.
π― What it does: Propose a validator-free test-time adaptation framework PDF, which leverages uncertainty-driven data augmentation voting and delay feedback-guided lightweight perturbation learning to enhance the robustness of vision-language-action models under minor environmental changes.
TextPecker: Rewarding Structural Anomaly Quantification for Enhancing Visual Text Rendering
Hanshen Zhu (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
CodeGenerationAnomaly DetectionReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodality
π― What it does: Proposes TextPecker, which employs a structure-aware anomaly detection-based reward strategy to perform reinforcement learning on text generation models, thereby enhancing the structural integrity and semantic consistency of visual text rendering.
Texvent: Asynchronous Event Data Simulation via Text Prompt
Ruofei Wang (Hong Kong Baptist University), Renjie Wan (Hong Kong Baptist University)
CodeGenerationData SynthesisDepth EstimationTransformerLarge Language ModelPrompt EngineeringTextMultimodality
π― What it does: Texvent proposes a training-free text-to-event (T2E) simulation framework that directly generates asynchronous, high-resolution event data from text prompts using multimodal large language models.
TF-SSD: A Strong Pipeline via Synergic Mask Filter for Training-free Co-salient Object Detection
Zhijin He (Xi'an Jiaotong-Liverpool University), Jimin Xiao (Xi'an Jiaotong-Liverpool University)
CodeObject DetectionTransformerImageBenchmark
π― What it does: Proposed a new no-training method called TF-SSD for cooperative significant object detection (CoSOD), generating and filtering masks through the synergistic effects of SAM and DINO.
π― What it does: This paper proposes TGSFormer, an expandable time Gaussian splatting framework for completing the construction of 3D semantic scenes from continuous first-person perspective observations.
π― What it does: Propose TGTrack, a unified multi-modal single-target tracking framework that explicitly guides the model to capture the evolution of targets and scenes over time through temporal generation learning tasks, achieving efficient tracking across RGB, depth, thermal imaging, event streams, and natural language descriptions.
The Coherence Trap: When MLLM-Crafted Narratives Exploit Manipulated Visual Contexts
Yuchen Zhang (School of Software Engineering, Xi'an Jiaotong University), Zhedong Zheng (University of Macau)
CodeClassificationData SynthesisAnomaly DetectionLarge Language ModelVision Language ModelMultimodality
π― What it does: Constructed a large-scale semantically aligned multimodal fake news dataset (MDSM) and proposed a framework (AMD) that utilizes multimodal large language models (MLLM) for generating and detecting forged text.
π― What it does: Proposed a source-free continuous test-time adaptation framework called GOLD, which utilizes a low-rank golden subspace to perform lightweight feature updates, addressing the trade-off between efficiency and generalization.
The Image as Its Own Reward: Reinforcement Learning with Adversarial Reward for Image Generation
Weijia Mao (Show Lab, National University of Singapore), Mike Zheng Shou (Show Lab, National University of Singapore)
CodeGenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelContrastive LearningImageText
π― What it does: This study proposes Adv-GRPO, a reinforcement learning framework based on adversarial rewards, aimed at improving the quality of text-to-image generation.