IEEE/CVF Conference on Computer Vision and Pattern Recognition Β· 1047 papers
OmniDocLayout: Towards Diverse Document Layout Generation via Coarse-to-Fine LLM Learning
Hengrui Kang (Shanghai Jiao Tong University), Conghui He (Shanghai Artificial Intelligence Laboratory)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMultimodality
π― What it does: Researched and implemented the LLM model OmniDocLayout-LLM for diverse document layout generation, and constructed the OmniDocLayout-1M dataset with a scale of millions.
OmniVTG: A Large-Scale Dataset and Training Paradigm for Open-World Video Temporal Grounding
Minghang Zheng (Peking University), Yang Liu (Peking University)
CodeRetrievalTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodalityBenchmarkChain-of-Thought
π― What it does: Proposed a large-scale, semantically rich open-world video temporal localization dataset called OmniVTG, and designed a three-stage training paradigm based on Self-Correction Chain-of-Thought (Self-Correction CoT) to enhance the localization capability of multi-modal large language models (MLLMs) on rare concepts.
OmniZip: Audio-Guided Dynamic Token Compression for Fast Omnimodal Large Language Models
Keda Tao (Zhejiang University), Huan Wang (Westlake University)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelVideoMultimodalityAudio
π― What it does: Propose OmniZip, a no-training, audio-guided dynamic audio-video token compression framework for accelerating inference in Omnimodal large language models;
OmniZip: Learning a Unified and Lightweight Lossless Compressor for Multi-Modal Data
Yan Zhao (Shanghai Jiao Tong University), Li Song (Ant Group)
CodeCompressionTransformerMixture of ExpertsImageTextMultimodalityTabularBiomedical DataAudio
π― What it does: Developed a unified, lightweight lossless compression model called OmniZip, supporting multimodal data including images, text, voice, tactile signals, medical images, gene sequences, and databases;
On the Role of Temporal Granularity in the Robustness of Spiking Neural Networks
Mengting Xu (Zhejiang University), Gang Pan (Zhejiang University)
CodeOptimizationAdversarial AttackSpiking Neural NetworkImageTime Series
π― What it does: This paper studies the robustness of spiking neural networks (SNN) from a time granular perspective, proposing a time granular attack (TG-Attack) based on gradients at each time step, as well as a time sensitivity value (TSV) calculated using Hessian information, and incorporating time granular regularization (TG-Reg) during training to enhance robustness.
π― What it does: Proposes the One-Shot Flow, Any-Time Frame framework for video frame interpolation in event cameras, capable of instantly retrieving bidirectional optical flow at any time point and generating high-quality interpolated frames after a single forward pass.
OneSparse: A Unified Framework for Sparse Activation Layers in Vision Models
Xingkui Zhu (Huazhong University Of Science And Technology), Xiang Bai (Xiaohongshu Inc)
CodeClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerMixture of ExpertsImage
π― What it does: Proposed the OneSparse framework, unifying MoE and memory modules into a three-stage process of dispatchβprocessβcombine, and designed the Nexus Layer as a sparse activation layer in visual models.
Open-Vocabulary Domain Generalization in Urban-Scene Segmentation
Dong Zhao, Zhun Zhong (Hefei University of Technology)
CodeSegmentationDomain AdaptationAutonomous DrivingComputational EfficiencyTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
π― What it does: Proposes a new task of open-vocabulary domain generalization semantic segmentation (OVDG-SS), constructs a comprehensive benchmark for urban driving scenarios, and introduces a text-image correlation refinement module based on state space, named S2-Corr, to achieve robust segmentation under unseen domains and unseen categories.
CodeSegmentationTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: Proposed OpenDPR, a training-free vision-centric diffusion-guided prototype retrieval framework for open-vocabulary change detection, further enhancing change localization under weakly supervised conditions through the S2C module.
OpenFS: Multi-Hand-Capable Fingerspelling Recognition with Implicit Signing-Hand Detection and Frame-Wise Letter-Conditioned Synthesis
Junuk Cha (Korea Advanced Institute of Science and Technology), Han-Mu Park (Korea Electronics Technology Institute)
CodeRecognitionGenerationPose EstimationTransformerDiffusion model
π― What it does: Propose a multi-hand recognizable finger spelling recognition and synthesis framework OpenFS, which includes an implicit hand-finger detector and a generator based on letter frame-level conditions;
CodeOptimizationKnowledge DistillationData-Centric LearningTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
π― What it does: Propose a unified two-stage training paradigm (SFT+RL) to construct the multimodal reasoning model OpenMMReasoner, and open-source the complete data, code, and model.
OpenVision 2: A Family of Generative Pretrained Visual Encoders for Multimodal Learning
Yanqing Liu (University of California Santa Cruz), Cihang Xie (University of California Santa Cruz)
CodeComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Simplify the original OpenVision model into a generative pre-training framework that uses only the image encoder and text decoder, removing the text encoder and contrastive loss, and training solely with the title generation loss.
π― What it does: Redefine optical flow estimation as continuous transport dynamics, learning time-related velocity fields and solving for pixel trajectories via Euler ODE to obtain optical flow.
π― What it does: This paper proposes an offline vectorized map construction framework called OptiMVMap based on the principle of 'select first, then fuse,' which utilizes multi-vehicle perspectives to compensate for the limitations of single-vehicle viewpoints, thereby generating more complete and topologically consistent high-definition maps;
ORCA: Orchestrated Reasoning with Collaborative Agents for Document Visual Question Answering
Aymen Lassoued (Digital Research Center of Sfax), Yousri Kessentini (Computer Vision Center)
CodeTransformerLarge Language ModelAgentic AIMixture of ExpertsVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
π― What it does: Propose the ORCA multi-agent framework, achieving single-page document visual question answering through five stages: reasoning decomposition, specialized agent collaboration, debate self-inspection, format validation, and other processes;
ORIC: Benchmarking Object Recognition under Contextual Incongruity in Large Vision-Language Models
Zhaoyang Li (University of California San Diego), Hao Su (University of California San Diego)
CodeRecognitionLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
π― What it does: Constructed an object recognition benchmark (ORIC-Bench) tailored for context-inconsistent scenarios, and automatically generated positive and negative samples through a dual sampling strategy combining LLM and CLIP; evaluated 20 large-scale vision-language models and two open-source detectors on this benchmark, further enhancing model robustness by applying Visual-RFT (Visual Reinforcement Fine-Tuning) on 600 ORIC-style training samples.
π― What it does: Proposes a CAD-free object pose estimation framework, OrienPose, under single-image scenarios, achieving pose prediction via direction-guided novel view synthesis.
π― What it does: Propose a unified cross-image editing framework called OrionEdit, which can achieve attribute transfer, style alignment, and multi-subject fusion under multiple reference images.
π― What it does: This paper studies how to migrate DINOv3 to remote sensing tasks, proposing the ORSATR-X framework, and incorporating the Weber Local Adapter (WLA) and Multi-scale Aggregation Module (MSAM) into the framework to enhance robustness to low-contrast targets and scale variations.
Orthogonal Spatial-Aware Multi-View Anchor Graph Clustering for Incomplete Remote Sensing Data
Yongshan Zhang (Sun Yat-sen University), Zhihua Cai (China University of Geosciences)
CodeGraph Neural NetworkImage
π― What it does: Propose a clustering framework OSMAGC for incomplete remote sensing multi-view data, achieving clustering through spatially aware anchor graph learning and orthogonal spatial regularization.
π― What it does: Propose a 4D occupancy-based robot video generation framework, ORV, capable of generating high-fidelity, controllable single-view, multi-view, and simulation-to-real robot operation videos.
OS-Oracle: A Comprehensive Framework for Cross-Platform GUI Critic Models
Zhenyu Wu (Shanghai Jiaotong University), Zichen Ding (Shanghai AI Laboratory)
CodeData SynthesisSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmark
π― What it does: Proposed a cross-platform GUI evaluation framework, OS-Oracle, for judging each operation step of computer use agents in mobile, Web, and desktop environments;
π― What it does: Propose a multi-GPU parallel inference framework named Otil, which accelerates diffusion model inference by synchronizing only the most informative latent variable blocks.
π― What it does: Proposes a training-free framework called P-Flow, which customizes dynamic visual effects in videos by optimizing text prompts during inference.
PA-Attack: Guiding Gray-Box Attacks on LVLM Vision Encoders with Prototypes and Attention
Hefei Mei (City University of Hong Kong), Minjing Dong (City University of Hong Kong)
CodeAdversarial AttackTransformerVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Targeting Large Vision-Language Models with shared visual encoders, the authors propose a gray-box attack method called PA-Attack, which can weaken the model's performance across multiple tasks by generating adversarial perturbations within the visual encoder without accessing the large language model's parameters.
PACT: Phase-Like Transition Constraints in Adapter-Based Continual Learning of Vision-Language Models
Xuan Wang (Tsinghua University), Jungong Han (Tsinghua University)
CodeTransformerVision Language ModelImage
π― What it does: Propose the PACT framework, which in visual-language model continual learning applies phase-class transfer constraints to task-specific adapters after convergence, creating two smooth states of 'freezing' and 'melting' to enhance cross-task transferability and stability.
π― What it does: Propose a Transformer-based Pano360 framework that performs global alignment of multi-image stitching using 3D perspective geometric consistency and automatically generates seamless panoramas.
PanoEnv: Exploring 3D Spatial Intelligence in Panoramic Environments with Reinforcement Learning
Zekai Lin (University of Glasgow), Xu Zheng (Hong Kong University of Science and Technology)
CodeData SynthesisReinforcement LearningVision Language ModelImageMultimodalityBenchmark
π― What it does: Propose PanoEnv, a large-scale panoramic visual question answering (VQA) benchmark for evaluating and enhancing the 3D spatial reasoning capabilities of Vision-Language Models (VLMs).
Quan Kong (Zhejiang University), Cong Wang (Zhejiang University)
CodeComputational EfficiencyVision Language ModelVideo
π― What it does: This paper proposes a parallel prediction-validation inference framework called ParallelVLM to accelerate the autoregressive decoding of video large language models.
CodeComputational EfficiencyLarge Language ModelSupervised Fine-TuningMultimodality
π― What it does: Propose an implicit modal decomposition (IMoD) method for parameter-efficient fine-tuning of multi-modal large language models, enabling the model to balance learning between text and non-text modalities while maintaining low parameter count and mergeability.
ParaUni: Enhance Generation in Unified Multimodal Model with Reinforcement-driven Hierarchical Parallel Information Interaction
Jiangtong Tan (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
CodeGenerationReinforcement LearningVision Language ModelDiffusion modelMultimodality
π― What it does: Propose the ParaUni framework, which parallelly extracts visual information from all layers of a Vision-Language Model (VLM), aggregates them through a Layer Integration Module (LIM) as conditional inputs to a diffusion model, and introduces a Layer-wise Dynamic Adjustment Mechanism (LDAM) during the reinforcement learning phase to apply noise perturbations to specific layers based on different rewards, thereby improving image generation quality and multi-reward optimization.
π― What it does: Proposes a partially weakly supervised object detection framework called PWOOD, which achieves efficient detection by utilizing only horizontal boxes or single-point weak annotations along with massive unlabeled data.
PCA-Seg: Revisiting Cost Aggregation for Open-Vocabulary Semantic and Part Segmentation
Jianjian Yin (Nanjing University of Science and Technology), Fumin Shen (University of Electronic Science and Technology of China)
CodeSegmentationTransformerMixture of ExpertsVision Language ModelImage
π― What it does: Propose the PCA-Seg model, which addresses the knowledge interference between class-level semantics and spatial context in open-vocabulary semantics and part segmentation through mechanisms such as parallel cost aggregation, expert-driven perceptual learning (EPL), and feature orthogonal decoupling (FOD).
PDCR: Perception-Decomposed Confidence Reward for Vision-Language Reasoning
Hee Suk Yoon (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
CodeLarge Language ModelReinforcement LearningMultimodality
π― What it does: Proposed the Perception-Decomposed Confidence Reward (PDCR), which classifies multimodal reasoning steps into visual perception and text reasoning categories using unsupervised visual dependency scores. It normalizes confidence improvements within each category to provide reinforcement learning with more stable, non-diluted step-level rewards, significantly enhancing multimodal reasoning performance.
π― What it does: This paper proposes a dual-teacher dual-student manifold distillation framework called PDD for medical image anomaly detection, which integrates the global context from VMamba-Tiny with the local structural priors from Wide-ResNet50 to construct a unified high-dimensional manifold, achieving more robust normal pattern learning through multi-scale fusion and diversity distillation.
CodeSegmentationTransformerVision Language ModelImageTextBenchmark
π― What it does: Proposes the PEARL framework, achieving training-free open-vocabulary semantic segmentation by inserting Procrustes alignment into the final self-attention of ViT and performing text-aware Laplacian propagation on a small grid without additional training.
π― What it does: Propose a new perception evaluation metric called Perception Characteristics Distance (PCD) to quantify the reliable detection range of perception systems under different distances and environmental conditions, and calculate the average PCD (aPCD) based on this metric; meanwhile, construct the SensorRainFall dataset, which provides real distance, bounding box, and segmentation annotations under controllable rainy weather and lighting conditions.
π― What it does: This paper proposes a variable-rate neural video compression framework PNVC-C/PNVC-CR based on color separation and grade chain GAN to enhance perceptual and objective compression quality.
Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning
Chi Zhang (Wuhan University), Jing Zhang (Wuhan University)
CodeReinforcement LearningVision Language ModelMultimodality
π― What it does: Introduce a perceptual inspection mechanism in reinforcement learning for vision-language models (VLMs), constructing a perceptual checklist and using its results as rewards and gating to enhance the model'sεεθ½ε (synergy) between perception and reasoning.
PETAR: Localized Findings Generation with Mask-Aware Vision-Language Modeling for PET Automated Reporting
Danyal Maqbool (University of Wisconsin-Madison), Tyler J. Bradshaw (Microsoft)
CodeSegmentationGenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBiomedical DataComputed TomographyPositron Emission Tomography
π― What it does: Created the first large-scale PET/CT dataset, PETARSeg-11K, and developed a 3D mask-aware vision-language model, PETAR-4B, based on this dataset to generate accurate tumor localization reports.
PFGNet: A Fully Convolutional Frequency-Guided Peripheral Gating Network for Efficient Spatiotemporal Predictive Learning
Xinyong Cai (Sichuan University), Yuankai Wu (Sichuan University)
CodeComputational EfficiencyConvolutional Neural NetworkVideoTime Series
π― What it does: Propose a fully convolutional PFGNet that achieves adaptive receptive fields via pixel-level frequency-guided peripheral frequency gating, addressing spatial and temporal modeling in spatiotemporal prediction learning.
π― What it does: Proposes a single-channel, training-agnostic virtual try-on framework called PG-VTON, which can naturally fit reference clothing onto a target person in a single diffusion inference.
PHASE-Net: Physics-Grounded Harmonic Attention System for Efficient Remote Photoplethysmography Measurement
Bo Zhao, Zitong Yu (Harbin Institute of Technology)
CodeComputational EfficiencyRepresentation LearningConvolutional Neural NetworkVideoBiomedical DataPhysics Related
π― What it does: PHASE-Net achieves temporal modeling of remote photoplethysmography (rPPG) through a physics-driven damped resonator model, and proposes a zero-FLOPs ZAS (Zero-FLOPs Axial Swapper) and Adaptive Spatial Filter (ASF) to enhance spatial feature interaction and attention selection.
PhaSR: Generalized Image Shadow Removal with Physically Aligned Priors
Chia-Ming Lee (National Yang Ming Chiao Tung University), Chih-Chung Hsu (National Yang Ming Chiao Tung University)
CodeRestorationDepth EstimationTransformerImage
π― What it does: Proposes a dual-layer physical alignment shadow removal framework called PhaSR, which achieves efficient, mask-free shadow removal by leveraging global normalization and cross-modal attention.
PhenoYieldNet: Learning Crop-Aware Phenological Responses for Multi-Crop Yield Prediction
Yu Luo (University of Sydney), Kun Hu (Edith Cowan University)
CodeTransformerContrastive LearningImageMultimodalityTime SeriesAgriculture Related
π― What it does: Propose a unified multi-crop yield prediction framework, PhenoYieldNet, capable of learning crop-specific growth cycle responses and associating them with meteorological changes;
Phrase-Grounding-Aware Supervised Fine-Tuning for Chart Recognition via Side-Masked Attention
Koichiro Ito (Japan Aerospace Exploration Agency)
CodeRecognitionTransformerSupervised Fine-TuningVision Language ModelMultimodality
π― What it does: Proposes a chart recognition method that introduces phrase alignment supervision during the fine-tuning stage of visual language models.
PhyGaP: Physically-Grounded Gaussians with Polarization Cues
Jiale Wu (Zhejiang University), Yifan Peng (The University Of Hong Kong)
CodeGenerationGaussian SplattingImagePhysics Related
π― What it does: Propose a physics-based 3D Gaussian splatting method called PhyGaP, which utilizes polarization information to achieve reflection decomposition and relightable rendering of glossy objects.
PhysGen: Physically Grounded 3D Shape Generation for Industrial Design
Yingxuan You (EPFL), Pascal Fua (EPFL)
CodeGenerationOptimizationTransformerFlow-based ModelRectified FlowAuto EncoderMeshPhysics Related
π― What it does: Propose PhysGen, a 3D shape generation framework that combines flow matching with physics-guided methods, capable of meeting specified aerodynamic performance while maintaining geometric feasibility.
Physics-Consistent Diffusion for Efficient Fluid Super-Resolution via Multiscale Residual Correction
Zhihao Li (Hong Kong University of Science and Technology (Guangzhou)), Guangtao Zhang (SandGold AI Research)
CodeSuper ResolutionDiffusion modelPhysics Related
π― What it does: Propose a fluid super-resolution method ReMD based on physical consistency diffusion, enhancing high-frequency details through multi-scale residual correction.
CodeGenerationData SynthesisTransformerNeural Radiance FieldGaussian SplattingImageMultimodalityPhysics Related
π― What it does: This paper proposes the PhysIR Splat framework, combining 3D Gaussian Splatting with thermal infrared radiative transfer models, achieving physically consistent thermal infrared 3D reconstruction and novel view synthesis.
PixDLM: A Dual-Path Multimodal Language Model for UAV Reasoning Segmentation
Shuyan Ke (Xiamen University), Rongrong Ji (Xiamen University)
CodeSegmentationLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityChain-of-Thought
π― What it does: Proposed the UAV Reasoning Segmentation task, constructed a large-scale high-resolution dataset named DRSeg, and designed the dual-path pixel-level multimodal language model PixDLM as the baseline.
π― What it does: This paper proposes a camera localization method called PlanaReLoc based on planar primitives, which achieves 6-DoF camera pose estimation by cross-modal matching between planes recovered from a single image and texture-free 3D plane maps.
π― What it does: A pluggable structured pruning framework named PPCL is proposed for Diffusion Transformer (DiT), which can flexibly prune depth and width without requiring retraining.
π― What it does: Propose a framework (PnP-CM) that uses consistency models as a Plug-and-Play prior to solve both linear and nonlinear inverse problems.
PoInit-of-View: Poisoning Initialization of Views Transfers Across Multiple 3D Reconstruction Systems
Weijie Wang (University of Trento), Bruno Lepri (Fondazione Bruno Kessler)
CodeAdversarial AttackGaussian SplattingImage
π― What it does: Propose the PoInit-of-View attack, which adversarially corrupts input views during the initialization phase of structured light reconstruction in 3D reconstruction systems.
Point Cloud as a Foreign Language for Multi-modal Large Language Model
Sneha Paul (Concordia University), Nizar Bouguila (Concordia University)
CodeClassificationGenerationTransformerLarge Language ModelReinforcement LearningMultimodalityPoint Cloud
π― What it does: Propose a multi-modal large language model called SAGE, which is end-to-end and does not rely on pre-trained 3D encoders. It can directly convert raw point clouds into discrete tokens and input them into LLMs for natural language generation and reasoning.
PointAlign: Feature-Level Alignment Regularization for 3D Vision-Language Models
Yuanhao Su (University Of Science And Technology Of China), Qi Fan (University Of Science And Technology Of China)
CodeClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningTextPoint Cloud
π― What it does: This paper proposes PointAlign, a regularization method for feature-level alignment of intermediate point cloud tokens in 3D vision-language models.
CodeGenerationGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextMeshGraph
π― What it does: Utilizing large language models (LLM) to generate CAD models through pointer-based command sequences enables explicit selection of B-rep edges/faces during generation, achieving advanced editing operations such as chamfer and fillet, while ensuring semantic consistency and topological integrity through multi-step generation.
π― What it does: This paper proposes a pose-free global spherical 3D Gaussian Splatting framework for reconstructing high-quality scenes and camera poses from unlabelled 360-degree videos.
PositionIC: Unified Position and Identity Consistency for Image Customization
Junjie Hu (Meituan), Wenqiang Zhang (Fudan University)
CodeGenerationData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelNeural Radiance FieldAuto EncoderImageTextMultimodalityBenchmark
π― What it does: Propose the PositionIC framework to achieve high-fidelity identity consistency and fine-grained spatial controllability during multi-subject image customization.
PPISP: Physically-Plausible Compensation and Control of Photometric Variations in Radiance Field Reconstruction
Isaac Deutsch (NVIDIA), Zan Gojcic (NVIDIA)
CodeRestorationNeural Radiance FieldGaussian SplattingImagePhysics Related
π― What it does: Propose a differentiable, physically plausible ISP processing pipeline and controller to address photometric inconsistencies in multi-view reconstruction caused by camera optical characteristics and image signal processing (ISP).
PQDT: Pseudo-Query Dual Transformer for Robust Point Cloud Restoration
Haoqing Wu (Mercedes-Benz AG), Jochen Garcke (University of Bonn)
CodeRestorationTransformerPoint Cloud
π― What it does: Proposes a unified Pseudo-Query Dual Transformer (PQDT) for point cloud restoration, addressing multiple degradations such as missing data, noise, and deformation.
π― What it does: This paper proposes the PR-IQA framework, which evaluates the quality of sparsely viewed new views generated by diffusion models using cross-perspective reference images, and introduces quality maps into 3D Gaussian Splatting (3DGS) training, significantly improving the quality of sparse-view 3D reconstruction and novel view synthesis.
Predict Before You Explore: Predictive Planning with Specialized Memory for Embodied Question Answering
Bowen Yuan (Nanjing University of Posts and Telecommunications), Bing-Kun Bao (Hefei University of Technology)
CodeRobotic IntelligenceTransformerReinforcement LearningVision Language ModelVision-Language-Action ModelImageTextMultimodality
π― What it does: Propose a prediction-driven EQA framework called Pred-EQA, combining hierarchical prediction planning with structured memory to achieve coherent long-term exploration and efficient evidence collection.
CodeClassificationComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelImageText
π― What it does: Propose a new service model reprogramming method called AReS, which primes the local encoder through a one-time API call, followed by local visual reprogramming to achieve efficient migration of closed-box service models
π― What it does: Propose the ProcessMaker framework, which utilizes Diffusion Transformers to generate cross-domain process sequences with self-adaptive steps.
Prompt-Anchored Vision-Text Distillation for Lifelong Person Re-identification
Wen Wen, Shiliang Zhang (University Of Electronic Science And Technology Of China)
CodeRecognitionDomain AdaptationKnowledge DistillationMeta LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageText
π― What it does: Propose a lifelong person re-identification framework without examples, PAD (Prompt-Anchored Vision-Text Distillation), which uses a frozen text encoder as a cross-domain semantic anchor. It achieves cross-domain semantic alignment and adaptive learning through learnable text prompts (TA-Prompt) and visual prompts (VA-Prompt), and employs knowledge distillation using a teacher model generated by EMA in the visual branch, maintaining overall model stability and plasticity.
Qihong Tang (Nanjing University), Yang Gao (Nanjing University)
CodeObject DetectionTransformerMixture of ExpertsContrastive LearningImage
π― What it does: Proposes a Prompt-Free Universal Region Proposal Network (PF-RPN) that can adaptively generate potential target boxes using visual features without any textual or visual prompts;
π― What it does: This paper proposes PromptStereo, a zero-shot stereo matching framework that achieves stronger iterative refinement by replacing GRU with Prompt Recurrent Unit (PRU) and combining structural prompts (SP) and motion prompts (MP).
Proof-of-Perception: Certified Tool-Using Multimodal Reasoning with Compositional Conformal Guarantees
Arya Fayyazi (University of Southern California), Haleh Akrami (Nuro)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
π― What it does: Propose the ProofβofβPerception (PoP) framework, modeling multi-modal reasoning as executing a directed acyclic graph with conformal certificates, and dynamically allocating computational resources based on node-level uncertainty using a controller.
Prototype-based Causal Intervention for Multi-Label Image Classification
Yanmin Li (National University of Defense Technology), Weidong Bao (National University of Defense Technology)
CodeClassificationImageBenchmark
π― What it does: Propose the ProCI framework, which reduces confusion associations in multi-label image classification under image-level labels through dynamic confusion memory and adaptive causal intervention.
π― What it does: Proposes a training-free concept erasure method based on prototype guidance, which can effectively eliminate general concepts during diffusion model inference.
π― What it does: In the federated semi-supervised learning framework, we propose using classifier learnable weights as a unified 'proxy' to simultaneously alleviate inter-client (external) and label-unlabeled data mismatch (internal) heterogeneity, and enhance model performance through two modules: Global Proxy Tuning (GPT) and Uncertain Class Proxy Learning (ICPL).
PRUE: A Practical Recipe for Field Boundary Segmentation at Scale
Gedeon Muhawenayo (Arizona State University), Hannah Kerner (Arizona State University)
CodeSegmentationConvolutional Neural NetworkImageAgriculture Related
π― What it does: This paper proposes a PRUE method for large-scale, cross-regional agricultural field boundary segmentation. By combining the U-Net architecture, EfficientNet-B7 encoder, log-cosh Dice loss, boundary weighting, brightness/scale augmentation, and channel shuffling techniques, it achieves robust segmentation of Sentinel-2 L2A images and generates high-precision field boundaries on the FTW benchmark.
π― What it does: This paper proposes an adaptive reconstruction-aware pruning scheduler (RPS) and a 3D differential Gaussian (3D-DoG) primitive capable of simultaneously representing positive and negative densities, achieving significant model compression and visual quality improvement under the 3D Gaussian splatting framework.
CodeAutonomous DrivingOptimizationComputational EfficiencyHyperparameter SearchVision Language ModelMultimodality
π― What it does: Propose Prune2Drive, a plug-and-play visual token pruning framework designed to accelerate the inference of multi-view vision-language models in autonomous driving.
π― What it does: Proposed a scalable multi-agent personalized image generation framework, leveraging a single-agent generative model to construct 350K high-quality multi-agent training data, and subsequently employing frame-level position encoding and Pairwise Subject-Consistency Rewards for post-training reinforcement learning.
π― What it does: Proposes Perception Encoder Audiovisual (PEAV), a unified multi-modal encoder capable of simultaneously aligning and embedding audio, video, and text, achieving state-of-the-art performance across multiple tasks such as audio, video, music, and speech under zero-shot settings.
PV-Ground: Text-Guided Point-Voxel Interaction for 3D Visual Grounding
Junpeng Shang (Zhejiang University), Dongfang Ma (Zhejiang University)
CodeObject DetectionSegmentationConvolutional Neural NetworkTransformerVision Language ModelTextPoint Cloud
π― What it does: Propose a text-guided point-voxel interaction framework, PV-Ground, for 3D visual localization tasks, which achieves efficient fusion of text and point cloud features while preserving high-resolution geometric details.
PvP: Data-Efficient Humanoid Robot Learning with Proprioceptive-Privileged Contrastive Representations
Mingqi Yuan (HK PolyU), Wenjun Zeng (EIT)
CodeRobotic IntelligenceReinforcement LearningAuto EncoderContrastive LearningTabularTime Series
π― What it does: In this study, the authors propose the PvP framework, which enhances representation learning and policy training in whole-body control (WBC) tasks through contrastive learning between self-sensory states and privileged states.
Quantum-Gated Task-interaction Knowledge Distillation for Pre-trained Model-based Class-Incremental Learning
Linjie Li (Beijing University of Posts and Telecommunications), Yang Ji (Beijing University of Posts and Telecommunications)
CodeClassificationKnowledge DistillationImage
π― What it does: Studying class-incremental learning based on pre-trained models, proposing a task-interaction knowledge distillation (QKD) framework based on quantum gates;
QuietPrune: Query-Guided Early Token Pruning for Vision-Language Models
Tianxiao Gao (Ant Group), Chenguang Ma (Ant Group)
CodeComputational EfficiencyTransformerVision Language ModelMultimodalityBenchmark
π― What it does: In visual-language models, this paper proposes the QuietPrune method, which performs query-guided early visual token pruning within ViT to reduce computational costs.
π― What it does: Propose Qwen-Image-Layered, an end-to-end diffusion model that can decompose a single RGB image into multiple layers of semantically disentangled RGBA layers, achieving naturally editable image representations.
R4-CGQA: Retrieval-based Vision Language Models for Computer Graphics Image Quality Assessment
Zhuangzi Li (Nanyang Technological University), Weisi Lin (Nanyang Technological University)
CodeRetrievalVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
π― What it does: Proposed a retrieval-enhanced two-stream framework named R4-CGQA to improve the ability of large models in computer graphics image quality assessment (CGQA), and constructed a dataset containing 3.5K CG images and six-dimensional quality descriptions.
RAAS: LLM Agentic System Architecture Search with GRPO
Jiayi Yang (Wuhan University), Mang Ye (Wuhan University)
CodeNeural Architecture SearchLarge Language ModelAgentic AIText
π― What it does: Proposed a robust architecture-adaptive search framework RAAS for automatically discovering high-quality multi-agent workflows in Agentic Supernet;
π― What it does: Proposes the RaGS framework, which utilizes 3D Gaussian splats to fuse 4D mmWave radar with monocular images, achieving sparse and efficient 3D object detection.
RAGTrack: Language-aware RGBT Tracking with Retrieval-Augmented Generation
Hao Li (Army Engineering University of PLA), Huchuan Lu (Dalian University of Technology)
CodeObject TrackingTransformerLarge Language ModelVision Language ModelMultimodalityRetrieval-Augmented Generation
π― What it does: Proposes the RAGTrack framework, which utilizes a multi-modal Transformer Encoder to fuse RGB, TIR, and text information, and achieves language-driven RGBT tracking through Adaptive Token Fusion (ATF) and Context-aware Reasoning Module (CRM);
Mingxiu Cai (Northeastern University), Xiatian Zhu (University of Surrey)
CodeAnomaly DetectionTransformerMixture of ExpertsImageRetrieval-Augmented Generation
π― What it does: Propose an unsupervised anomaly detection method RAID based on the retrieval-augmented generation (RAG) framework, which uses retrieved normal samples to guide the generation of fine-grained anomaly maps.
CodeSegmentationComputational EfficiencyRepresentation LearningData-Centric LearningTransformerMixture of ExpertsAuto EncoderMultimodalityTime SeriesBenchmark
π― What it does: Proposes RAMENβa tunable resolution multimodal encoder for unified representation learning of Earth observation data, achieving cross-modal transfer through self-supervised mask reconstruction.
Random Wins All: Rethinking Grouping Strategies for Vision Tokens
Qihang Fan (MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences), Ran He (MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences)
π― What it does: Proposed an extremely simple random grouping strategy as an alternative to the complex grouping methods in Vision Transformers, and validated its effectiveness across various visual tasks (image classification, object detection, instance segmentation, semantic segmentation, point cloud segmentation, and vision-language tasks).