CVPR 2025 Papers — Page 19
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2871 papers
OmniMMI: A Comprehensive Multi-modal Interaction Benchmark in Streaming Video Contexts
Yuxuan Wang (Beijing Institute for General Artificial Intelligence), Zilong Zheng (Beijing Institute for General Artificial Intelligence)
RecognitionGenerationData SynthesisOptimizationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoMultimodalityBenchmarkAudio
🎯 What it does: This paper presents OmniMMI—a comprehensive benchmark for evaluating multimodal interaction (visual and audio) capabilities in streaming video environments, and designs the M4 framework based on this benchmark to achieve efficient real-time multimodal inference.
OmniSplat: Taming Feed-Forward 3D Gaussian Splatting for Omnidirectional Images with Editable Capabilities
Suyoung Lee (Seoul National University), Kyoung Mu Lee (Seoul National University)
GenerationData SynthesisOptimizationTransformerGaussian SplattingImage
🎯 What it does: Using the Yin-Yang grid to decompose the panoramic image into two quasi-perspective views, directly estimating the 3D Gaussian parameters on a pre-trained feedforward 3D Gaussian splatting network, and achieving rapid panoramic image reconstruction and new viewpoint synthesis through cross-view attention and Yin-Yang rasterization.
OmniStereo: Real-time Omnidireactional Depth Estimation with Multiview Fisheye Cameras
Jiaxi Deng (Sun Yat-sen University), Gang Chen (Sun Yat-sen University)
Depth EstimationConvolutional Neural NetworkImage
🎯 What it does: We propose OmniStereo, a panoramic depth estimation framework based on Cassini projection and a lightweight stereo matching network, capable of generating high-quality panoramic depth maps in real-time.
OmniStyle: Filtering High Quality Style Transfer Data at Scale
Ye Wang (Jilin University), Rui Ma (Jilin University)
Image TranslationData SynthesisTransformerDiffusion modelImage
🎯 What it does: A large-scale high-quality style transfer dataset, OmniStyle-1M, has been constructed, and an end-to-end style transfer framework, OmniStyle, based on the DiT architecture has been proposed.
On Denoising Walking Videos for Gait Recognition
Dongyang Jin (Southern University of Science and Technology), Shiqi Yu (Southern University of Science and Technology)
RecognitionDiffusion modelVideo
🎯 What it does: The DenoisingGait method is proposed, which uses knowledge-driven and geometry-driven feature matching through a diffusion model to denoise walking videos, enhancing pedestrian gait recognition accuracy.
On the Consistency of Video Large Language Models in Temporal Comprehension
Minjoon Jung (Seoul National University), Angela Yao (National University of Singapore)
RecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVideoTextMultimodalityChain-of-Thought
🎯 What it does: This study investigates the consistency issues of large video language models in temporal understanding, proposing specialized evaluation datasets (Charades-CON, ActivityNet-CON) and consistency detection methods (rephrasing, displacement, holistic/compositional validation), and conducts a comprehensive evaluation across multiple models.
On the Generalization of Handwritten Text Recognition Models
Carlos Garrido-Munoz (University of Alicante), Jorge Calvo-Zaragoza (University of Alicante)
RecognitionDomain AdaptationRecurrent Neural NetworkAuto EncoderText
🎯 What it does: A large-scale evaluation of the domain generalization ability of handwritten text recognition models in the absence of prior data was conducted.
On the Out-Of-Distribution Generalization of Large Multimodal Models
Xingxuan Zhang (Tsinghua University), Peng Cui (Tsinghua University)
Domain AdaptationDrug DiscoveryTransformerLarge Language ModelVision Language ModelImageMultimodalityBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: The system evaluates and analyzes the zero-shot generalization ability of large multimodal models (LMM) in fields such as synthetic images, natural distribution drift, and medical and molecular imaging, finding that their performance significantly declines outside the training domain.
On the Zero-shot Adversarial Robustness of Vision-Language Models: A Truly Zero-shot and Training-free Approach
Baoshun Tong (Sun Yat-sen University), Jian Yin (Sun Yat-sen University)
ClassificationSegmentationAdversarial AttackTransformerVision Language ModelContrastive LearningImage
🎯 What it does: A completely untrained zero-shot adversarial robustness enhancement method AOM is proposed, which constructs anchors using a small amount of Gaussian noise and performs a linear interpolation in the embedding space to repair adversarial samples.
On-Device Self-Supervised Learning of Low-Latency Monocular Depth from Only Events
Jesse J. Hagenaars (MAVLab TU Delft), Guido C.H.E. de Croon (MAVLab TU Delft)
Depth EstimationAutonomous DrivingRecurrent Neural NetworkContrastive LearningImageVideo
🎯 What it does: This paper implements online, device-side self-supervised learning for monocular depth estimation based on event cameras on small unmanned aerial vehicles (UAVs), achieving real-time training and inference through an improved Contrast Maximization process.
Once-Tuning-Multiple-Variants: Tuning Once and Expanded as Multiple Vision-Language Model Variants
Chong Yu (Fudan University), Zhongxue Gan (Fudan University)
TransformerVision Language ModelMultimodality
🎯 What it does: A Once-Tuning-Multiple-Variants (OTMV) framework is proposed, which can dynamically expand to generate multiple variants of visual-language models after a single tuning.
ONDA-Pose: Occlusion-Aware Neural Domain Adaptation for Self-Supervised 6D Object Pose Estimation
Tao Tan (University of Chinese Academy of Sciences), Qiulei Dong (University of Chinese Academy of Sciences)
Object DetectionPose EstimationDomain AdaptationNeural Radiance FieldImage
🎯 What it does: This paper proposes a self-supervised 6D pose estimation method called ONDA-Pose, which utilizes a CAD-like lighting field to transform unlabeled real images into a synthetic domain, and generates pseudo-pose labels through a global pose refinement module, further self-supervising the pose estimator.
One Diffusion to Generate Them All
Duong H. Le (Allen Institute for AI), Jiasen Lu (Allen Institute for AI)
Image TranslationGenerationData SynthesisTransformerDiffusion modelFlow-based ModelImageTextMultimodality
🎯 What it does: We propose OneDiffusion, a unified diffusion model capable of simultaneously performing multiple tasks such as text-to-image, image-to-image, multi-view generation, and identity customization.
One is Plenty: A Polymorphic Feature Interpreter for Immutable Heterogeneous Collaborative Perception
Yuchen Xia (Beijing University of Posts and Telecommunications), Jinglin Li (Beijing University of Posts and Telecommunications)
Autonomous DrivingExplainability and InterpretabilityPrompt EngineeringPoint Cloud
🎯 What it does: This paper proposes PolyInter, a polymorphic feature interpreter designed to align intermediate features from different perception networks to the semantic space of the target vehicle in immutable heterogeneous cooperative perception scenarios.
One Model for ALL: Low-Level Task Interaction Is a Key to Task-Agnostic Image Fusion
Chunyang Cheng (Jiangnan University), Josef Kittler (University of Surrey)
TransformerAuto EncoderImage
🎯 What it does: This paper proposes GIFNet, a model that achieves unified multi-task image fusion and single-modal enhancement through the interaction of low-level tasks.
One-for-More: Continual Diffusion Model for Anomaly Detection
Xiaofan Li (East China Normal University), Yuan Xie (East China Normal University)
Anomaly DetectionDiffusion modelImage
🎯 What it does: A sustainable learning diffusion model (CDAD) is proposed for continuous anomaly detection.
One-Minute Video Generation with Test-Time Training
Karan Dalal (NVIDIA), Xiaolong Wang (NVIDIA)
GenerationData SynthesisTransformerDiffusion modelVideoText
🎯 What it does: A Test-Time Training (TTT) layer is added to a pre-trained diffusion Transformer, utilizing its trainable neural network hidden states to generate a coherent one-minute video from a text storyboard.
One-shot 3D Object Canonicalization based on Geometric and Semantic Consistency
Li Jin (Shandong University), Baoquan Chen (Peking University)
Object DetectionPose EstimationOptimizationTransformerLarge Language ModelVision Language ModelPoint CloudMesh
🎯 What it does: This paper proposes a one-shot 3D object normalization framework that requires only a single prior model for each category to align the target object to a canonical coordinate system under any pose.
One-Step Event-Driven High-Speed Autofocus
Yuhan Bao (Zhejiang University), Kaiwei Wang (Zhejiang University)
Image TranslationOptimizationImage
🎯 What it does: This study investigates a method for single-step autofocus using event cameras and grayscale Laplacian information.
One-Way Ticket: Time-Independent Unified Encoder for Distilling Text-to-Image Diffusion Models
Senmao Li (Nankai University), Jian Yang (Nankai University)
GenerationComputational EfficiencyKnowledge DistillationDiffusion modelImageText
🎯 What it does: This paper studies a time-invariant unified encoder (TiUE) that implements a non-cyclic text-to-image diffusion model distillation, significantly improving inference speed while maintaining high-quality diversity.
One2Any: One-Reference 6D Pose Estimation for Any Object
Mengya Liu (ETH Zurich), Federico Tombari (Google)
Pose EstimationDiffusion modelImage
🎯 What it does: A novel 6D pose estimation method for new objects can be achieved using a single-view RGB-D reference image (One2Any).
Online Task-Free Continual Learning via Dynamic Expansionable Memory Distribution
Fei Ye (University of Electronic Science and Technology of China), Adrian G. Bors (University of York)
Data SynthesisAuto EncoderImage
🎯 What it does: A dynamic expandable memory distribution (DEMD) framework is proposed to address the problem of catastrophic forgetting in task-free continual learning without task or category information, combining dynamic memory systems with long-term memory systems for online learning.
Online Video Understanding: OVBench and VideoChat-Online
Zhenpeng Huang (Nanjing University), Limin Wang (Nanjing University)
RecognitionSegmentationRetrievalRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: A benchmark for real-time streaming video, OVBench, has been constructed, and a multimodal large language model called VideoChat-Online has been proposed, which can efficiently handle infinitely long video streams.
OnlineAnySeg: Online Zero-Shot 3D Segmentation by Visual Foundation Model Guided 2D Mask Merging
Yijie Tang (National University of Defense Technology), Kai Xu (National University of Defense Technology)
Object DetectionSegmentationVision Language ModelPoint Cloud
🎯 What it does: In real-time incremental modeling scenarios, 2D masks generated by visual foundation models are used to quickly retrieve spatial overlap relationships through voxel hashing, and 3D instance segmentation results are obtained by merging in a zero-shot manner.
OODD: Test-time Out-of-Distribution Detection with Dynamic Dictionary
Yifeng Yang (Shanghai Jiao Tong University), Nanyang Ye (Tianjin University)
Anomaly DetectionComputational EfficiencyImageBenchmark
🎯 What it does: The research proposes a testing-time OOD detection method named OODD, which achieves detection by dynamically maintaining an OOD dictionary during the inference phase, without the need for model fine-tuning or additional training.
Open Ad-hoc Categorization with Contextualized Feature Learning
Zilin Wang (University of Michigan), Liu Ren (Bosch Center for AI)
ClassificationExplainability and InterpretabilityRepresentation LearningTransformerVision Language ModelContrastive LearningImage
🎯 What it does: A framework for open ad-hoc categorization (OAK) is proposed, which can automatically discover and predict known and novel categories with only a small number of labeled samples and unlabeled images.
Open Set Label Shift with Test Time Out-of-Distribution Reference
Changkun Ye (Australian National University), Nick Barnes (Australian National University)
ClassificationDomain AdaptationImage
🎯 What it does: A three-stage EM algorithm based on ID classifiers and OOB classifiers is proposed, which can estimate and correct the open set label distribution shift (OSLS) and improve target domain classification performance without retraining the model.
Open-Canopy: Towards Very High Resolution Forest Monitoring
Fajwel Fogel, Philippe Ciais (Laboratory of Climate and Environmental Sciences)
Object DetectionSegmentationConvolutional Neural NetworkTransformerSupervised Fine-TuningImagePoint CloudBenchmark
🎯 What it does: A publicly available VHR forest canopy height dataset covering 87,000 square kilometers in France was constructed, and the Open-Canopy and Open-Canopy-∆ benchmarks were proposed.
Open-Vocabulary Functional 3D Scene Graphs for Real-World Indoor Spaces
Chenyangguang Zhang (Tsinghua University), Francis Engelmann (Stanford University)
Object DetectionRobotic IntelligenceTransformerLarge Language ModelVision Language ModelImagePoint CloudGraph
🎯 What it does: Proposes a task to generate open vocabulary functional 3D scene graphs (including objects, interactive components, and their functional relationships) from RGB-D images;
Open-World Amodal Appearance Completion
Jiayang Ao (University of Melbourne), Krista A. Ehinger (University of Melbourne)
RestorationSegmentationGenerationDiffusion modelImageBenchmark
🎯 What it does: This study proposes a training-free open-world blind spot completion framework that utilizes text queries to achieve complete reconstruction of any occluded object, outputting RGBA layers that can be directly used for image editing and 3D reconstruction.
Open-World Objectness Modeling Unifies Novel Object Detection
Shan Zhang (Australian Institute for Machine Learning), Anton van den Hengel (Australian Institute for Machine Learning)
Object DetectionTransformerImage
🎯 What it does: The OWOBJ method is proposed, which addresses the issue of unknown objects being misclassified as known categories or background in open-world, few-shot, and zero-shot object detection by modeling the joint probability of objectness and category information.
OpenHumanVid: A Large-Scale High-Quality Dataset for Enhancing Human-Centric Video Generation
Hui Li (Fudan University), Siyu Zhu (Baidu Inc)
GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelVideoText
🎯 What it does: This paper presents the OpenHumanVid, a large-scale high-quality human video dataset, and utilizes LoRA fine-tuning on an extended Diffusion Transformer to enhance the quality of human video generation.
OpenING: A Comprehensive Benchmark for Judging Open-ended Interleaved Image-Text Generation
Pengfei Zhou (Shanghai AI Laboratory), Kaipeng Zhang (Shanghai AI Laboratory)
GenerationTransformerLarge Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the OpenING intertwined image-text generation benchmark (5,400 examples, 56 tasks) and designs the IntJudge evaluation model to support large-scale assessment of intertwined generation.
OpenMIBOOD: Open Medical Imaging Benchmarks for Out-Of-Distribution Detection
Max Gutbrod (Regensburg Medical Image Computing), Christoph Palm (Regensburg Medical Image Computing)
ClassificationAnomaly DetectionImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A benchmark for OOD detection in medical imaging, OpenMIBOOD, has been established, and 24 post-hoc methods have been evaluated on it.
OpenSDI: Spotting Diffusion-Generated Images in the Open World
Yabin Wang (Xi'an Jiaotong University), Xiaopeng Hong (Harbin Institute of Technology)
RecognitionObject DetectionSegmentationTransformerPrompt EngineeringVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: This paper presents the challenge of identifying images generated by diffusion models under open-world conditions (OpenSDI) and constructs a corresponding large-scale dataset, OpenSDID. Subsequently, a collaborative scheme, SPM, which integrates multiple pre-trained models, is proposed, and based on this, the MaskCLIP model is implemented to simultaneously perform image-level detection and pixel-level localization.
Opportunistic Single-Photon Time of Flight
Sotiris Nousias (University of Toronto), Kyros N. Kutulakos
Object DetectionDepth EstimationOptical FlowPoint Cloud
🎯 What it does: Using a single-photon SPAD camera to passively detect unknown, asynchronous, offline laser pulse signals, and achieving three-dimensional reconstruction and laser positioning based on their frequency structure.
Optical-Flow Guided Prompt Optimization for Coherent Video Generation
Hyelin Nam (KAIST), Jong Chul Ye (KAIST)
GenerationData SynthesisOptimizationPrompt EngineeringDiffusion modelOptical FlowVideoBenchmark
🎯 What it does: The MotionPrompt framework is proposed to enhance the temporal consistency and motion smoothness of text-to-video diffusion models through flow-guided prompt optimization.
OPTICAL: Leveraging Optimal Transport for Contribution Allocation in Dataset Distillation
Xiao Cui (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
OptimizationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the OPTICAL framework, which transforms the minimization of uniform distances in dataset distillation into a dynamic allocation of contributions from real sample pairs to synthetic samples through Optimal Transport, thereby capturing instance-level geometric structures more accurately.
OpticalNet: An Optical Imaging Dataset and Benchmark Beyond the Diffraction Limit
Benquan Wang (Nanyang Technological University), Bo An (Nanyang Technological University)
Image TranslationRestorationSuper ResolutionConvolutional Neural NetworkTransformerImageBenchmark
🎯 What it does: An OpticalNet dataset was constructed, and an optical imaging benchmark based on the concept of building blocks was designed to achieve deep learning mapping from diffraction images to sub-wavelength object images.
Optimal Transport-Guided Source-Free Adaptation for Face Anti-Spoofing
Zhuowei Li (Rutgers University), Yifan Xing (Amazon Web Services AI Labs)
Domain AdaptationContrastive LearningImage
🎯 What it does: This study investigates a source-end data-free few-shot adaptation framework based on prototypes and optimal transport, which can adapt a face anti-spoofing model using only a minimal amount of data during client testing.
Optimizing for the Shortest Path in Denoising Diffusion Model
Ping Chen (China Unicom), Shiguo Lian (China Unicom)
RestorationGenerationOptimizationDiffusion modelImage
🎯 What it does: A denoising diffusion model based on the shortest path, ShortDF, is proposed to optimize residual propagation for faster sampling.
Optimus-2: Multimodal Minecraft Agent with Goal-Observation-Action Conditioned Policy
Zaijing Li (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision Language ModelVideoMultimodality
🎯 What it does: This paper proposes the Optimus-2 agent, which combines a multimodal large language model (MLLM) for high-level planning and designs a Goal-Observation-Action Conditioned Policy (GOAP) for low-level control, addressing the modeling problem of the relationship among observation, action, and sub-goals; simultaneously, it constructs the MGOA dataset with 25,000 videos.
OralXrays-9: Towards Hospital-Scale Panoramic X-ray Anomaly Detection via Personalized Multi-Object Query-Aware Mining
Bingzhi Chen (Beijing Institute of Technology), Yishu Liu (Harbin Institute of Technology)
Object DetectionAnomaly DetectionTransformerContrastive LearningImageBenchmark
🎯 What it does: This paper proposes a panoramic dental X-ray anomaly detection benchmark for hospital scale, OralXrays-9, and implements automatic detection of nine types of oral anomalies based on a Transformer-based Multi-Objective Query-Aware Mining (MOQAM) framework.
Order-One Rolling Shutter Cameras
Marvin Anas Hahn (Trinity College Dublin), Tomas Pajdla (Czech Technical University in Prague)
🎯 What it does: A unified theory of the 'Order-one Rolling Shutter (RS1) camera' is proposed, providing its parameterization, geometric properties, and the corresponding minimum camera pose estimation problem.
Order-Robust Class Incremental Learning: Graph-Driven Dynamic Similarity Grouping
Guannan Lai (Southwestern University of Finance and Economics), Xin Yang (Southwestern University of Finance and Economics)
ClassificationGraph Neural NetworkImage
🎯 What it does: A dynamic similarity grouping method for class-incremental learning (GDDSG) is proposed, which dynamically partitions classes with low similarity into different groups using a graph coloring algorithm, and trains independent sub-models for each group.
ORIDa: Object-centric Real-world Image Composition Dataset
Jinwoo Kim (Yonsei University), Seon Joo Kim (Yonsei University)
Image TranslationRestorationGenerationData SynthesisSupervised Fine-TuningDiffusion modelImage
🎯 What it does: This paper presents a large-scale, real-world object composition dataset ORIDa, which includes 200 unique objects, approximately 30,000 images, and provides fact-counterfactual (F-CF) sets and fact-only image (F-Only) sets; it also trains and evaluates object removal and insertion methods based on diffusion models on this dataset.
OSDFace: One-Step Diffusion Model for Face Restoration
Jingkai Wang (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
RestorationDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a single-step diffusion model OSDFace, which extracts prior information from low-quality faces using a Visual Representation Embedder (VRE) and achieves high-quality restoration with good facial identity preservation by combining identity loss and GAN guidance.
OSLoPrompt: Bridging Low-Supervision Challenges and Open-Set Domain Generalization in CLIP
Mohamad Hassan N C (Indian Institute of Technology Bombay), Biplab Banerjee (Indian Institute of Technology Bombay)
GenerationDomain AdaptationTransformerLarge Language ModelPrompt EngineeringDiffusion modelImage
🎯 What it does: A low-sample open-domain generalization task, LSOSDG, is proposed, and the OSLOPROMPT framework is designed to achieve domain-independent generalization and open-set detection with very few samples.
OSMamba: Omnidirectional Spectral Mamba with Dual-Domain Prior Generator for Exposure Correction
Gehui Li (Peking University), Jian Zhang (Peking University)
RestorationDiffusion modelImage
🎯 What it does: This paper proposes a global frequency domain exposure correction network, OSMamba, based on a state space model, aimed at addressing the issues of illumination correction and structure recovery under extreme exposure.
OSV: One Step is Enough for High-Quality Image to Video Generation
Xiaofeng Mao (Fudan University), Wenhan Luo (Hong Kong University of Science and Technology)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImageVideoOrdinary Differential Equation
🎯 What it does: A two-stage training framework (OSV) is proposed to achieve high-quality image-to-video generation in one step.
Ouroboros3D: Image-to-3D Generation via 3D-aware Recursive Diffusion
Hao Wen (Beihang University), Lu Sheng (Beihang University)
GenerationData SynthesisDiffusion modelGaussian SplattingImagePoint Cloud
🎯 What it does: Using a recursive diffusion process, a single image is transformed into a multi-view consistent image and a corresponding high-quality 3D model.
Overcoming Shortcut Problem in VLM for Robust Out-of-Distribution Detection
Zhuo Xu (Huazhong University of Science and Technology), Yifan Liang (Huazhong University of Science and Technology)
SegmentationAnomaly DetectionPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This paper addresses the shortcut problem of visual language models (such as CLIP) in OOD detection and proposes the OSPCoOp method, which enhances robustness through background separation and mask-guided regional regularization.
OverLoCK: An Overview-first-Look-Closely-next ConvNet with Context-Mixing Dynamic Kernels
Meng Lou (Hong Kong University), Yizhou Yu (Hong Kong University)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A fully convolutional network backbone called OverLoCK is proposed, which uses a Deeply Divided Strategy (DDS) to divide the network into three sub-networks: Base-Net, Overview-Net, and Focus-Net. Within Focus-Net, context-mixed dynamic convolution (ContMix) is employed to achieve adaptive long-range modeling.
OVO-Bench: How Far is Your Video-LLMs from Real-World Online Video Understanding?
Junbo Niu (Shanghai Artificial Intelligence Laboratory), Jiaqi Wang (Shanghai Innovation Institute)
Large Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper presents OVO-Bench, a benchmark specifically designed to evaluate the online understanding capabilities of video large language models, covering three modes: backward reasoning, real-time visual perception, and forward proactive response.
OW-OVD: Unified Open World and Open Vocabulary Object Detection
Xing Xi (South China University of Technology), Yu Qiu (South China University of Technology)
Object DetectionLarge Language ModelVision Language ModelImage
🎯 What it does: A detector called OW-OVD is proposed, which possesses both open vocabulary and open world object detection capabilities;
PACT: Pruning and Clustering-Based Token Reduction for Faster Visual Language Models
Mohamed Dhouib (Polytechnic Institute of Paris), Aymen Shabou (Credit Agricole)
CompressionComputational EfficiencyTransformerVision Language ModelImageVideoMultimodality
🎯 What it does: To address the computational and memory consumption issues caused by a large number of visual tokens during inference in visual language models, this paper proposes pruning visual tokens and merging clusters in the early layers of the model, significantly reducing inference costs.
Paint by Inpaint: Learning to Add Image Objects by Removing Them First
Navve Wasserman (Weizmann Institute of Science), Ron Kimmel (Technion - Israel Institute of Technology)
Image TranslationGenerationTransformerLarge Language ModelVision Language ModelDiffusion modelImageText
🎯 What it does: By first removing the target objects from the images and utilizing the source-target image pairs generated from this process, a diffusion model is trained to achieve object addition in images based on unmasked text instructions.
PanDA: Towards Panoramic Depth Anything with Unlabeled Panoramas and Mobius Spatial Augmentation
Zidong Cao (Hong Kong University of Science and Technology), Lin Wang (Nanyang Technological University)
Depth EstimationImage
🎯 What it does: This paper conducts a systematic analysis of the performance of the traditional Depth Anything model on panoramic images and proposes the PanDA framework based on semi-supervised learning, utilizing large-scale unlabeled panoramic data and Mobius space augmentation to enhance the zero-shot and generalization capabilities of panoramic depth estimation.
PanoGS: Gaussian-based Panoptic Segmentation for 3D Open Vocabulary Scene Understanding
Hongjia Zhai (Zhejiang University), Guofeng Zhang (Zhejiang University)
Object DetectionSegmentationGaussian SplattingPoint Cloud
🎯 What it does: We propose PanoGS, a vocabulary-open scene understanding framework based on 3D Gaussian Splatting, capable of achieving joint segmentation of 3D semantics and instances.
Panorama Generation From NFoV Image Done Right
Dian Zheng (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
GenerationDiffusion modelContrastive LearningImage
🎯 What it does: A separable diffusion model PanoDecouple is designed, using DistortNet to guide panoramic distortion and ContentNet to complete the content. Additionally, the Distort-CLIP evaluation metric and Distort-FID metric are proposed to address the 'visual cheating' phenomenon where existing methods sacrifice distortion accuracy to enhance visual quality.
PanSplat: 4K Panorama Synthesis with Feed-Forward Gaussian Splatting
Cheng Zhang (Monash University), Jianfei Cai (Monash University)
GenerationData SynthesisGaussian SplattingImage
🎯 What it does: A full convolutional 3D Gaussian splatting method (PanSplat) aimed at 4K resolution is proposed, which can quickly synthesize new views from two wide baseline panoramas.
ParaHome: Parameterizing Everyday Home Activities Towards 3D Generative Modeling of Human-Object Interactions
Jeonghwan Kim (Seoul National University), Hanbyul Joo (Seoul National University)
Object TrackingGenerationData SynthesisDiffusion modelVideoTextMultimodality
🎯 What it does: This paper presents the ParaHome system, which utilizes 70 synchronized RGB cameras, an IMU motion capture suit, and gloves, combined with 3D ArUco cube markers, to capture the three-dimensional movements of humans, hands, and multiple objects (including disassemblable mechanical parts) in everyday home scenes, and generates corresponding textual descriptions; a HOI dataset was collected from 38 participants over 486 minutes and 207 recordings.
Parallel Sequence Modeling via Generalized Spatial Propagation Network
Hongjun Wang (NVIDIA), Sifei Liu (University of Hong Kong)
ClassificationGenerationTransformerImage
🎯 What it does: This paper proposes the Generalized Spatial Propagation Network (GSPN), a two-dimensional linear propagation attention mechanism specifically designed for visual tasks, which achieves efficient sequence modeling while maintaining spatial coherence.
Parallelized Autoregressive Visual Generation
Yuqing Wang (University of Hong Kong), Xihui Liu (University of Hong Kong)
GenerationData SynthesisTransformerImageVideo
🎯 What it does: This paper proposes a non-local parallel autoregressive visual generation framework called PAR, which achieves efficient parallel sampling by utilizing the dependencies between tokens.
Parameter Efficient Mamba Tuning via Projector-targeted Diagonal-centric Linear Transformation
Seokil Ham (Korea Advanced Institute of Science and Technology), Changick Kim (Korea Advanced Institute of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: This study investigates parameter-efficient fine-tuning methods for the Mamba architecture, finding that the Projector is crucial for downstream tasks, and proposes the ProDiaL method.
Parameter-efficient Fine-tuning in Hyperspherical Space for Open-vocabulary Semantic Segmentation
Zelin Peng (Shanghai Jiao Tong University), Wei Shen (Shanghai Jiao Tong University)
SegmentationTransformerContrastive LearningImage
🎯 What it does: This paper proposes H-CLIP, which performs open vocabulary semantic segmentation for CLIP through symmetric parameter-efficient fine-tuning in hyperspherical space.
Parameterized Blur Kernel Prior Learning for Local Motion Deblurring
Zhenxuan Fang (Xidian University), Guangming Shi (Xidian University)
RestorationConvolutional Neural NetworkImage
🎯 What it does: To address the problem of Local Motion Deblurring, this paper proposes a novel framework based on the learning of parameterized blur kernel priors (PGDN), which can estimate the blur kernel (length, angle, curvature) for each pixel without the need for manual annotations, and utilizes this prior information to guide a dual-branch deblurring network for refined recovery.
Parametric Point Cloud Completion for Polygonal Surface Reconstruction
Zhaiyu Chen (Technical University of Munich), Xiao Xiang Zhu (Technical University of Munich)
RestorationGenerationTransformerPoint Cloud
🎯 What it does: A framework called PaCo is proposed, which is based on parameterized point cloud completion to directly recover planar primitives from incomplete point clouds, achieving polygon surface reconstruction.
PARC: A Quantitative Framework Uncovering the Symmetries within Vision Language Models
Jenny Schmalfuss (University of Stuttgart), Jose M. Alvarez (NVIDIA)
RecognitionData-Centric LearningTransformerPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: This study investigates the prompt sensitivity of visual language models (VLM) and proposes the PARC framework to systematically evaluate the model's performance under different prompt variations.
PartGen: Part-level 3D Generation and Reconstruction with Multi-view Diffusion Models
Minghao Chen (Visual Geometry Group, University of Oxford), Andrea Vedaldi (Visual Geometry Group, University of Oxford)
SegmentationGenerationDiffusion modelMesh
🎯 What it does: Proposes the PartGen pipeline, which generates editable multi-part 3D models from text, images, or unstructured 3D objects.
PartRM: Modeling Part-Level Dynamics with Large Cross-State Reconstruction Model
Mingju Gao (Tsinghua University), Hao Zhao (Peking University)
GenerationData SynthesisContrastive LearningPoint CloudMesh
🎯 What it does: A 4D model called PartRM is proposed, based on large-scale 3D Gaussian reconstruction, which can simultaneously learn appearance, geometry, and part-level motion.
PassionSR: Post-Training Quantization with Adaptive Scale in One-Step Diffusion based Image Super-Resolution
Libo Zhu (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
RestorationSuper ResolutionDiffusion modelImage
🎯 What it does: Low-bit post-training quantization is performed on a first-order diffusion super-resolution model, proposing PassionSR, which achieves 8/6 bit quantization while maintaining visual quality close to full precision.
Patch Matters: Training-free Fine-grained Image Caption Enhancement via Local Perception
Ruotian Peng (South China University of Technology), Di Hu (Renmin University of China)
Object DetectionTransformerLarge Language ModelVision Language ModelImageTextBenchmark
🎯 What it does: A training-free divide-then-aggregate method is proposed, which generates finer-grained and less hallucinatory image descriptions by semantically and spatially partitioning images and performing hierarchical aggregation and semantic filtering at both local and global levels.
PatchDEMUX: A Certifiably Robust Framework for Multi-label Classifiers Against Adversarial Patches
Dennis Jacob (University of California Berkeley), Prateek Mittal (Princeton University)
ClassificationObject DetectionAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposes PatchDEMUX, a multi-label adversarial patch defense framework that can be scaled to any single-label Certifiable Defense against Patch Attacks (CDPA);
PatchDPO: Patch-level DPO for Finetuning-free Personalized Image Generation
Qihan Huang (Zhejiang University), Jie Song (Zhejiang University)
GenerationData SynthesisDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes PatchDPO, which adds an additional training phase to a personalized image generation model without the need for fine-tuning, enhancing generation quality through patch-level direct preference optimization.
PatchGuard: Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies
Mojtaba Nafez (Sharif University of Technology), Mohammad Hossein Rohban (Sharif University of Technology)
Anomaly DetectionTransformerImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: This paper proposes PatchGuard, a visual Transformer architecture that combines foreground-aware pseudo-anomaly generation and attention strength regularization to enhance adversarial robustness in anomaly detection and localization.
PatchVSR: Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution
Shian Du (Tsinghua University), Xiangyang Ji (Tsinghua University)
RestorationGenerationSuper ResolutionDiffusion modelVideoText
🎯 What it does: Utilize a pre-trained video diffusion model for patch video super-resolution, breaking the model's resolution limitations and generating high-quality details;
Pathways on the Image Manifold: Image Editing via Video Generation
Noam Rotstein (Technion - Israel Institute of Technology), Ron Kimmel (Technion - Israel Institute of Technology)
Image TranslationRestorationGenerationTransformerVision Language ModelImageVideoBenchmark
🎯 What it does: Redefines the image editing task as a video generation process, utilizing a video model to generate continuous frames from the original image to the target edit, achieving high-fidelity editing through automatically generated temporal editing subtitles and frame selection.
Patient-Level Anatomy Meets Scanning-Level Physics: Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model
Ziyuan Yang (Sichuan University), Yi Zhang (Sichuan University)
RestorationFederated LearningTransformerLarge Language ModelPrompt EngineeringImageBiomedical DataComputed Tomography
🎯 What it does: A dual-layer physically driven federated learning framework that combines scanning protocols and anatomical hints for low-dose CT image denoising is proposed.
Pattern Analogies: Learning to Perform Programmatic Image Edits by Analogy
Aditya Ganeshan (Brown University), Daniel Ritchie (Brown University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: By providing a pair of example patterns (A, A') to express the desired structural editing, a conditional diffusion model is utilized to automatically perform programmatic, structure-aware editing of the target pattern B; there is no need to first infer the pattern generation program.
PAVE: Patching and Adapting Video Large Language Models
Zhuoming Liu, Yin Li
Object DetectionTransformerImage
🎯 What it does: This paper proposes a multi-scale feature fusion framework based on Transformer to enhance object detection performance.
Pay Attention to the Foreground in Object-Centric Learning
Pinzhuo Tian (Shanghai University), Alex Kot (Nanyang Technological University)
Object DetectionSegmentationTransformerContrastive LearningImage
🎯 What it does: An unsupervised foreground-background distinction indicator is proposed, which is integrated with a slot attention-based object-centric learning model to enhance foreground object segmentation in complex background scenes.
PBR-NeRF: Inverse Rendering with Physics-Based Neural Fields
Sean Wu (ETH Zurich), Christos Sakaridis (ETH Zurich)
RestorationGenerationNeural Radiance FieldImagePhysics Related
🎯 What it does: Perform inverse rendering on multi-view images, jointly estimating geometry, material (Disney BRDF parameters), and lighting, and achieving relightable new view synthesis within the NeRF framework.
PCDreamer: Point Cloud Completion Through Multi-view Diffusion Priors
Guangshun Wei (Shandong University), Changjian Li (University of Edinburgh)
RestorationGenerationData SynthesisTransformerDiffusion modelPoint Cloud
🎯 What it does: This paper proposes a point cloud completion framework based on multi-view diffusion models for generating images as a precursor, combined with multi-modal fusion and confidence constraints.
PCM : Picard Consistency Model for Fast Parallel Sampling of Diffusion Models
Junhyuk So (POSTECH), Eunhyeok Park (POSTECH)
GenerationData SynthesisComputational EfficiencyRobotic IntelligenceDiffusion modelImage
🎯 What it does: Proposes the Picard Consistency Model (PCM), which directly predicts fixed points during the Picard iteration process to enhance parallel sampling speed.
PDFactor: Learning Tri-Perspective View Policy Diffusion Field for Multi-Task Robotic Manipulation
Jingyi Tian (Xi'an Jiaotong University), Wei Tang (University of Illinois at Chicago)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelMultimodality
🎯 What it does: The PDFactor framework is proposed, which utilizes a strategy of combining three-view projection and three-plane feature learning to directly generate 6-DoF robotic arm actions from visual and language instructions.
PEACE: Empowering Geologic Map Holistic Understanding with MLLMs
Yangyu Huang (Microsoft Research), Furu Wei (Microsoft Research)
Large Language ModelPrompt EngineeringImageTextMultimodalityBenchmark
🎯 What it does: Proposed the GeoMap-Bench benchmark and the GeoMap-Agent framework to achieve multimodal understanding and question answering of geological images.
PEER Pressure: Model-to-Model Regularization for Single Source Domain Generalization
Dong Kyu Cho (New York University), Sanghack Lee (Seoul National University)
Domain AdaptationContrastive LearningImageBenchmark
🎯 What it does: This paper proposes a single-source domain generalization method called PEER (Parameter-Space Ensemble with Entropy Regularization), which alleviates mid-term OOD performance fluctuations caused by data augmentation and improves generalization to unknown target domains by averaging the parameters of the proxy model and the main model and maximizing the feature information of both through entropy regularization.
Percept, Memory, and Imagine: World Feature Simulating for Open-Domain Unknown Object Detection
Aming Wu (Xidian University), Cheng Deng (Xidian University)
Object DetectionDomain AdaptationConvolutional Neural NetworkGraph Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes a 'World Feature Simulation' (WFS) framework that utilizes multi-layer perception, fragmented memory, and imagination mechanisms to synthesize virtual unknown features under the condition of no auxiliary data, thereby enhancing the robustness of object detectors against unknown objects in open domains.
Perception Tokens Enhance Visual Reasoning in Multimodal Language Models
Mahtab Bigverdi (University of Washington), Ranjay Krishna (University of Washington)
Object DetectionDepth EstimationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: The introduction of perceptual tokens in multimodal language models, along with the AURORA training framework, enables the model to generate visual intermediate representations such as depth maps and bounding boxes during inference, thereby enhancing its reasoning capabilities for 3D depth and 2D counting tasks.
Perceptual Inductive Bias Is What You Need Before Contrastive Learning
Junru Zhao (Carnegie Mellon University), Tai Sing Lee (Carnegie Mellon University)
SegmentationDepth EstimationRepresentation LearningContrastive LearningImage
🎯 What it does: Introduce intermediate visual constructs such as shape prototypes, reflectance, and shadows before contrastive learning as a pre-training phase to accelerate convergence and improve downstream task performance.
Perceptual Video Compression with Neural Wrapping
Muhammad Umar Karim Khan (Sony Interactive Entertainment), Yiannis Andreopoulos (Sony Interactive Entertainment)
CompressionConvolutional Neural NetworkFlow-based ModelVideo
🎯 What it does: A neural packaging scheme is proposed, which inserts trainable preprocessors and postprocessors before and after traditional video codecs (AV1, VVC) to enhance perceptual quality and rate-quality performance.
Perceptually Accurate 3D Talking Head Generation: New Definitions, Speech-Mesh Representation, and Evaluation Metrics
Lee Chae-Yeon (POSTECH), Tae-Hyun Oh (POSTECH)
GenerationTransformerAuto EncoderContrastive LearningVideoMesh
🎯 What it does: Three major perceptual criteria for 3D talking head generation are proposed (temporal synchronization, lip readability, expressiveness), and a speech-grid synchronized representation space, corresponding perceptual loss, and evaluation metrics are designed to enhance the perceptual accuracy of 3D talking heads.
Period-LLM: Extending the Periodic Capability of Multimodal Large Language Model
Yuting Zhang (Hong Kong University of Science and Technology), Kaishun Wu (Hong Kong University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: A multimodal large language model named Period-LLM is proposed, specifically designed to handle periodic tasks, and a cross-modal periodic evaluation benchmark is constructed.
PerLA: Perceptive 3D Language Assistant
Guofeng Mei (Fondazione Bruno Kessler), Yiming Wang (CSIRO)
RecognitionObject DetectionGraph Neural NetworkLarge Language ModelPoint Cloud
🎯 What it does: We propose PerLA, a 3D language assistant that can simultaneously capture local details and global context, achieving more accurate LLM interactions through point cloud processing.
PERSE: Personalized 3D Generative Avatars from A Single Portrait
Hyunsoo Cha (Seoul National University), Hanbyul Joo (Seoul National University)
GenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: Generate animatable and editable personalized 3D generative avatars from a single portrait.
Person De-reidentification: A Variation-guided Identity Shift Modeling
Yi-Xing Peng (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
RecognitionSafty and PrivacyImage
🎯 What it does: This study explores the problem of De-ReID (De-Identification), proposing a new framework that achieves de-identification through a mutation-guided identity transfer mechanism while maintaining the re-identification capability for other accessible individuals.
PersonaBooth: Personalized Text-to-Motion Generation
Boeun Kim (University of Birmingham), Hyung Jin Chang
GenerationTransformerDiffusion modelContrastive LearningVideoTextMultimodality
🎯 What it does: A personalized momentum task is proposed, the PerMo dataset is constructed, and the PersonaBooth scheme is developed to achieve text-based personalized action generation.
PersonaHOI: Effortlessly Improving Face Personalization in Human-Object Interaction Generation
Xinting Hu (Max Planck Institute for Informatics), Bernt Schiele (Max Planck Institute for Informatics)
GenerationDiffusion modelImageText
🎯 What it does: The PersonaHOI framework is proposed, which combines general StableDiffusion with a personalized face diffusion model to achieve identity-consistent human-computer interaction image generation without training or fine-tuning.
Personalized Preference Fine-tuning of Diffusion Models
Meihua Dang (Stanford University), Jiaming Song (Luma AI)
GenerationRecommendation SystemOptimizationReinforcement LearningVision Language ModelDiffusion modelImageText
🎯 What it does: A multi-reward optimization objective PPD is proposed, utilizing VLM to extract user preference embeddings and personalizing Stable Cascade through cross-attention, thereby achieving alignment and generation of multiple user preferences.