CVPR 2026 Papers — Page 24
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
Neighbor GRPO: Contrastive ODE Policy Optimization Aligns Flow Models
Dailan He (The Chinese University of Hong Kong Mmlab), Hongsheng Li (The Chinese University of Hong Kong Mmlab)
Reinforcement Learning from Human FeedbackFlow-based ModelContrastive LearningImageTextOrdinary Differential Equation
🎯 What it does: Propose Neighbor GRPO, achieving RLHF alignment in flow matching models from a contrastive learning perspective, without requiring SDE transformation, while retaining the advantages of deterministic ODE sampling.
Neighbor-Aware Localized Concept Erasure in Text-to-Image Diffusion Models
Zhuan Shi (McGill University), Golnoosh Farnadi (McGill University)
GenerationSafty and PrivacyConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Propose a training-free, three-stage neighborhood-aware local concept erasure framework called NLCE, which can precisely remove specified concepts in text-to-image diffusion models while preserving the visual expression of adjacent concepts.
NeighborMAE: Exploiting Spatial Dependencies between Neighboring Earth Observation Images in Masked Autoencoders Pretraining
Liang Zeng (KU Leuven), Maarten Vergauwen (KU Leuven)
ClassificationSegmentationTransformerAuto EncoderImage
🎯 What it does: Proposes the NeighborMAE framework, which leverages adjacent Earth observation images to jointly complete the reconstruction task of the Masked AutoEncoder, thereby learning spatial dependencies.
NeoVerse: Enhancing 4D World Model with in-the-wild Monocular Videos
Yuxue Yang (School of Artificial Intelligence University of Chinese Academy of Sciences), Zhaoxiang Zhang (School of Artificial Intelligence University of Chinese Academy of Sciences)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelRectified FlowGaussian SplattingWorld ModelVideoMultimodality
🎯 What it does: Propose NeoVerse, an expandable 4D world model capable of pose-free 4D Gaussian Splatting reconstruction from monocular video, online generation of new trajectory videos, and support for various downstream applications.
NERFIFY: A Multi-Agent Framework for Turning NeRF Papers into Code
Seemandhar Jain (University of California San Diego), Manmohan Chandraker (University of California San Diego)
AI Code AssistantLarge Language ModelVision Language ModelNeural Radiance FieldImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Automatically convert NeRF papers into trainable Nerfstudio plugins, achieving fully executable code.
NESTOR: A Nested MOE-based Neural Operator for Large-Scale PDE Pre-Training
Dengdi Sun (Anhui University), Bin Luo (Anhui University)
Representation LearningMixture of ExpertsPhysics Related
🎯 What it does: This paper proposes NESTOR, a PDE pre-trained neural operator based on a nested Mixture-of-Experts (MoE) architecture.
Nestwork: Conditional 3D Furnished House Layout Generation through Latent Heterogeneous Graph Diffusion
Shuhan Miao (Massachusetts Institute of Technology), Junling Zhuang (Georgia Institute of Technology)
GenerationGraph Neural NetworkLarge Language ModelDiffusion modelNeural Radiance FieldTextMultimodalityGraph
🎯 What it does: Propose the Nestwork framework, which utilizes heterogeneous graphs (rooms and furniture nodes with multiple spatial relationships) to perform conditional 3D layout generation in a latent diffusion space, and supports control from semantic graphs or natural language;
Neu-PiG: Neural Preconditioned Grids for Fast Dynamic Surface Reconstruction on Long Sequences
Julian Kaltheuner, Reinhard Klein
RestorationRepresentation LearningPoint Cloud
🎯 What it does: This paper proposes Neu-PiG, a fast dynamic surface reconstruction method based on preconditioned multi-scale latent grids, which achieves spatiotemporally consistent high-precision 3D reconstruction from long sequences of unstructured point clouds.
Neural Collapse in Test-Time Adaptation
Xiao Chen (Tsinghua University), Zhi Wang (Sichuan University)
ClassificationDomain AdaptationContrastive LearningImage
🎯 What it does: Proposed a test-time adaptation method called NCTTA based on Neural Collapse, extending NC for the first time at the sample level, revealing the alignment collapse between sample features and classifier weights (NC3+) and correcting pseudo-label errors using a hybrid objective.
Neural Differentiation in Deep Networks: A Theoretical Framework for Expressivity and Representational Diversity
Boyuan Wang (Lancaster University), Richard Jiang (Shanghai Jiaotong University)
ClassificationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a theoretical framework called Neural Differentiation and introduces the Neural Differentiation Index (NDI) to measure the functional uniqueness of individual neurons within a network. Based on NDI, we design an NDI-guided sparsification algorithm (NDP) that achieves structured sparsification without significantly reducing accuracy or even improving it.
Neural Distribution Prior for LiDAR Out-of-Distribution Detection
Zizhao Li (University of Melbourne), Kourosh Khoshelham (University of Melbourne)
Data SynthesisAnomaly DetectionTransformerPoint Cloud
🎯 What it does: Propose the Neural Distribution Prior (NDP) framework, which models the network output logit distribution using an attention mechanism and adaptively reweights OOD scores, while generating diverse auxiliary OOD samples through Perlin noise synthesis and introducing the Soft Outlier Exposure (SOE) strategy.
Neural Dynamic GI: Random-Access Neural Compression for Temporal Lightmaps in Dynamic Lighting Environments
Jianhui Wu (University of Science and Technology of China), Chao Li (Zhejiang University)
CompressionAuto EncoderVideo
🎯 What it does: Propose Neural Dynamic GI (NDGI), which uses neural networks to compress temporal lightmaps to support real-time dynamic global illumination.
Neural Field-Based 3D Surface Reconstruction of Microstructures from Multi-Detector Signals in Scanning Electron Microscopy
Shuo Chen (Zhejiang University), Guofeng Zhang (Zhejiang University)
GenerationNeural Radiance FieldImagePoint CloudMesh
🎯 What it does: A hybrid framework based on neural fields, NFH-SEM, achieves high-fidelity 3D surface reconstruction of microstructures using multi-view and quadrant backscattered electron (BSE) detector images.
Neural Gabor Splatting: Enhanced Gaussian Splatting with Neural Gabor for High-frequency Surface Reconstruction
Haato Watanabe (University of Tokyo), Nobuyuki Umetani (University of Tokyo)
RestorationGaussian SplattingImagePoint Cloud
🎯 What it does: Proposes a neural Gabor splatting method, which encodes spatial and view-dependent color variations in each Gaussian primitive using a lightweight MLP, enabling one-time representation of high-frequency textures; introduces a frequency-aware densification strategy that controls the number of primitives by adding/removing primitives based on frequency-domain error.
Neural Mixture Density Processes
Yi Ding (National University of Defense Technology), Guangquan Cheng (National University of Defense Technology)
OptimizationRepresentation LearningMeta LearningConvolutional Neural NetworkTransformerMixture of ExpertsImageTabular
🎯 What it does: Proposed a new hybrid density neural process (Neural Mixture Density Process, NMDP) that represents tasks through learnable mixture weights on the Dirichlet simplex and achieves fast adaptation to function distributions under a meta-learning framework.
Neural-Centric Video Processing Pipeline for Unified Multi-Task Inference
Seyeon Lee (Korea Advanced Institute of Science and Technology), Dongsu Han (Korea Advanced Institute of Science and Technology)
ClassificationRecognitionObject DetectionTransformerNeural Radiance FieldVideo
🎯 What it does: Directly encode videos into implicit neural representations (INR), extracting task-specific features from intermediate layers via micro-adapters, skipping pixel decoding and preprocessing to support multi-task inference.
Neuro-Cognitive Reward Modeling for Human-Centered Autonomous Vehicle Control
Zhuoli Zhuang (University of Technology Sydney), Chin-Teng Lin (University of Technology Sydney)
Autonomous DrivingReinforcement Learning from Human FeedbackConvolutional Neural NetworkTransformerReinforcement LearningMultimodalityBiomedical Data
🎯 What it does: Collected EEG, eye movement, and control data from 20 drivers in a simulated driving environment, and constructed a reward model based on EEG event-related potentials (ERP) prediction. Subsequently, the model's predictions were embedded into the reinforcement learning (RL) reward to enhance autonomous vehicles' collision avoidance capabilities.
Neurodynamics-Driven Coupled Neural P Systems for Multi-Focus Image Fusion
Bo Li (Xi'an Jiaotong University), Jun Liu (Xi'an Jiaotong University)
Spiking Neural NetworkImageBenchmark
🎯 What it does: This paper proposes a neural dynamics-driven coupled neural P system (ND-CNPFuse) for multi-focus image fusion.
NeuroFlow: Toward Unified Visual Encoding and Decoding from Neural Activity
Weijian Mai (Shanghai Artificial Intelligence Laboratory), Jiamin Wu (Shanghai Artificial Intelligence Laboratory)
GenerationRetrievalComputational EfficiencyRepresentation LearningTransformerDiffusion modelFlow-based ModelAuto EncoderContrastive LearningImageBiomedical DataMagnetic Resonance ImagingOrdinary Differential Equation
🎯 What it does: Developed a unified visual encoding and decoding framework called NeuroFlow, capable of generating images from brain activity and predicting brain activity from images within the same model.
NeuROK: Generative 4D Neural Object Kinematics
Chen Geng, Jiajun Wu
GenerationData SynthesisTransformerMeshTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: Proposes a framework named NEUROK, capable of generating four-dimensional (spatial+temporal) simulated dynamic sequences of objects based solely on a single static 3D shape and physical conditions (e.g., force, velocity). The framework achieves category-agnostic global dynamic simulation by learning a low-dimensional latent space to represent all possible object states and solving dynamics using Lagrangian mechanics within this space.
NeuroRule: Bridging Vision and Logic with Differentiable Rule Induction
Muhammad Zarar, Zhiyong Feng (Tianjin University)
Object DetectionSegmentationExplainability and InterpretabilityTransformerImageBenchmarkChain-of-Thought
🎯 What it does: Proposed an end-to-end differentiable rule induction framework called NeuroRule, integrating pixel-level visual perception with symbolic logic reasoning to generate interpretable scene graphs;
NeuroSeg Meets DINOv3: Transferring 2D Self-Supervised Visual Priors to 3D Neuron Segmentation via DINOv3 Initialization
Yik San Cheng (University of Sydney), Weidong Cai (University of Sydney)
SegmentationRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningBiomedical Data
🎯 What it does: Propose the NeurINO model, by migrating the 2D self-supervised visual prior DINOv3 to a 3D convolutional network, achieving efficient neuronal segmentation and reconstruction.
Next-Scale Autoregressive Models for Text-to-Motion Generation
Zhiwei Zheng (University of Pennsylvania), Mingmin Zhao (University of Pennsylvania)
GenerationTransformerVision-Language-Action ModelVideoTextMultimodality
🎯 What it does: Designed and implemented MoScale, a text-to-action generation framework based on next-scale autoregressive methods.
Next-Scale Prediction: A Self-Supervised Approach for Real-World Image Denoising
Yiwen Shan (Sichuan University), Yuanbiao Gou (Sichuan University)
RestorationSuper ResolutionConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposes the Next-Scale Prediction (NSP) framework, achieving self-supervised image denoising by first denoising at low resolution and then predicting high resolution, resolving the conflict between noise decorrelation and detail preservation.
NexusFlow: Unifying Disparate Tasks under Partial Supervision via Invertible Flow Networks
Fangzhou Lin (Worcester Polytechnic Institute), Ziming Zhang (Worcester Polytechnic Institute)
Object DetectionObject TrackingSegmentationDepth EstimationAutonomous DrivingFlow-based ModelImageVideo
🎯 What it does: Propose the NexusFlow framework to address multi-task learning problems under partially supervised settings, with different tasks and domain differences.
NG-GS: NeRF-guided 3D Gaussian Splatting Segmentation
Yi He (Beijing Jiaotong University), Haibin Ling (Westlake University)
SegmentationNeural Radiance FieldGaussian SplattingPoint Cloud
🎯 What it does: Proposed a NeRF-guided 3D Gaussian scattering segmentation framework, NG-GS, capable of refining object boundaries within a 3D Gaussian model
NI-Tex: Non-isometric Image-based Garment Texture Generation
Hui Shan (Zhejiang University), Xiangru Huang (Westlake University)
GenerationData SynthesisPose EstimationDiffusion modelImageVideoMesh
🎯 What it does: Propose a model named NI-Tex that can generate high-quality PBR textures even when there are non-isometric and topological differences between the image and target geometry.
NIL: No-data Imitation Learning
Mert Albaba (ETH Zurich), Michael J. Black (Max Planck Institute for Intelligent Systems)
SegmentationData SynthesisRobotic IntelligenceTransformerReinforcement LearningDiffusion modelImageVideo
🎯 What it does: Using a single 2D video generated by a pre-trained video diffusion model as an expert demonstration, combined with physical simulation and a discriminator-free reward signal, to achieve 3D motion skill learning without any 3D demonstration data.
NimbusGS: Unified 3D Scene Reconstruction under Hybrid Weather
Yanying Li (Ocean University of China), Yong Du (Ocean University of China)
RestorationGaussian SplattingImage
🎯 What it does: Propose NimbusGS, a unified framework that recovers high-quality 3D scenes when multi-view inputs are affected by various or mixed adverse weather conditions (fog, rain, snow).
NitroGen: An Open Foundation Model for Generalist Gaming Agents
Loïc Magne (Nvidia), Linxi Fan (Nvidia)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningVision-Language-Action ModelDiffusion modelFlow-based ModelVideoBenchmark
🎯 What it does: Trained a general-purpose game agent named NITROGEN using 40,000 hours of publicly available game videos with player action annotations for behavior cloning, and provided a multi-game, multi-task evaluation suite.
No Calibration, No Depth, No Problem: Cross-Sensor View Synthesis with 3D Consistency
Cho-Ying Wu (Bosch Research North America and Bosch Center for Artificial Intelligence), Liu Ren (Bosch Research North America and Bosch Center for Artificial Intelligence)
GenerationNeural Radiance FieldGaussian SplattingImageMultimodality
🎯 What it does: Proposes a cross-sensor view synthesis method that aligns RGB with other perception modalities (e.g., infrared, near-infrared, SAR) and generates pixel-level consistent X images without calibration or depth information, addressing the bottleneck of traditional 3D prior dependencies.
No Hard Negatives Required: Concept Centric Learning Leads to Compositionality without Degrading Zero-shot Capabilities of Contrastive Models
Hai X. Pham (Samsung AI Center), Brais Martinez (Samsung AI Center)
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper proposes the C2LIP method, which significantly enhances the compositional reasoning ability of CLIP-style vision-language models through concept-centric contrastive learning and parameter-free cross-modal attention pooling.
No Labels, No Look-Ahead: Unsupervised Online Video Stabilization with Classical Priors
Tao Liu (Nanjing University of Science and Technology), Shibo Wen (Jilin University)
RestorationConvolutional Neural NetworkOptical FlowVideoMultimodalityBenchmark
🎯 What it does: Propose an unsupervised, online video stabilization framework that integrates a classic three-stage pipeline (motion estimation, motion propagation, motion compensation) with multithreaded buffering, achieving real-time stabilization without requiring paired stabilized/unstabilized video data.
No Need For Real Anomaly: MLLM Empowered Zero-Shot Video Anomaly Detection
Zunkai Dai (Beijing University of Posts and Telecommunications), Yuanyuan Qiao (Beijing University of Posts and Telecommunications)
Anomaly DetectionTransformerLarge Language ModelVision Language ModelVideoMultimodality
🎯 What it does: Designed an end-to-end zero-shot video anomaly detection framework, LAVIDA, which leverages multi-modal large language models (MLLM) and pseudo anomaly samplers to achieve cross-scenario anomaly detection without requiring training on real abnormal videos.
No Way To Steal My Face: Proactive Defense Against Identity-Preserving Personalized Generation
Lizhi Xiong (Nanjing University of Information Science and Technology), Zhangjie Fu (Nanjing University of Information Science and Technology)
Safty and PrivacyDiffusion modelScore-based ModelImage
🎯 What it does: Proposed a generic active identity protection framework called IDGuardian, which suppresses identity information leakage in both training-based and training-free personalized diffusion models through two-phase perturbations.
Node-RF: Learning Generalized Continuous Space-Time Scene Dynamics with Neural ODE-based NeRFs
Hiran Sarkar (Technical University of Munich), Benjamin Busam (Technical University of Munich)
GenerationData SynthesisRepresentation LearningNeural Radiance FieldVideo
🎯 What it does: By combining neural ODE with NeRF, a continuous-time, extrapolatable spatiotemporal implicit field is learned to achieve long-term video prediction and novel view rendering.
Noise-Aware Few-Shot Learning through Bi-directional Multi-View Prompt Alignment
Lu Niu (Southeast University), Cheng Xue (Southeast University)
ClassificationMeta LearningPrompt EngineeringVision Language ModelImage
🎯 What it does: Propose NA-MVP, a noise-aware few-shot learning framework realized through bidirectional multi-view prompt alignment and reversible optimal transport.
Nonlinear Color Transfer via Learnable Bezier Flows
Junhyoung Lee (Easywith), Jangho Kim (Kookmin University)
Image TranslationImage HarmonizationMixture of ExpertsFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: This paper proposes a nonlinear color transfer framework (NCT) based on learnable Bezier curves, achieving smooth and realistic migration of color distribution from content images to target style images.
Nonparametric Deep Fine-grained Clustering with Low-Rank Guided Vision-Language Model
Xulun Ye (Ningbo University), Kun Zhou (Shenzhen University)
OptimizationRepresentation LearningVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose a non-parametric deep fine-grained clustering framework based on a low-rank guided vision-language model (VLM), which can simultaneously learn discriminative features and dynamically infer the number of clusters under unlabeled and unknown class number scenarios.
NoOVD: Novel Category Discovery and Embedding for Open-Vocabulary Object Detection
Yupeng Zhang (Tianjin University), Liang Wan (Tianjin University)
Object DetectionKnowledge DistillationPrompt EngineeringVision Language ModelImageText
🎯 What it does: Proposes a novel open-vocabulary object detection framework called NoOVD, aiming to address the category mismatch problem between training and testing phases.
NoRD: A Data-Efficient Vision-Language-Action Model that Drives without Reasoning
Ishaan Rawal (University Of California Berkeley), Wei Zhan (University Of California Berkeley)
Autonomous DrivingTransformerSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelImageMultimodality
🎯 What it does: Proposed a visual-language-action model called NORD that achieves autonomous driving without a reasoning step.
Not All Birds Look The Same: Identity-Preserving Generation For Birds
Aaron Sun (University of Massachusetts), Subhransu Maji (University of Massachusetts)
GenerationTransformerSupervised Fine-TuningDiffusion modelImageBenchmark
🎯 What it does: Propose the NABLA benchmark and train an identity-preserving bird image generation model
NOVA: Sparse Control, Dense Synthesis for Pair-Free Video Editing
Tianlin Pan (Nanjing University), Chenyang Si (Nanjing University)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: Propose the NOVA framework, which combines semantic control via sparse keyframes with dense original video information synthesis to achieve unpaired video editing.
NOWA: Null-space Optical Watermark for Invisible Capture Fingerprinting and Tamper Localization
Edwin Vargas (Rice University), Ashok Veeraraghavan (Rice University)
Safty and PrivacyConvolutional Neural NetworkImage
🎯 What it does: Propose a hybrid optical-digital framework that embeds a phase mask in the camera aperture to generate an invisible Null-space Optical Watermark (NOWA), and achieves measurement-consistent reconstruction through the Null-Space Network (NSN), while utilizing null-space projection for pixel-level tamper localization.
NS-Diff: Fluid Navier-Stokes Guided Video Diffusion via Reinforcement Learning
Zijun Deng (Wangxuan Institute of Computer Technology, Peking University), Yuxin Peng (Wangxuan Institute of Computer Technology, Peking University)
GenerationTransformerReinforcement LearningDiffusion modelOptical FlowVideoBenchmarkPhysics Related
🎯 What it does: Propose NS-Diff, a reinforcement learning framework integrating physical constraints for generating physically plausible videos within video diffusion models.
NTK-Guided Implicit Neural Teaching
Chen Zhang (University of Hong Kong), Ngai Wong (University of Hong Kong)
Computational EfficiencyRepresentation LearningImageAudio
🎯 What it does: This paper proposes a sampling strategy called NINT based on the neural tangent kernel (NTK) to accelerate the training of implicit neural representations (INR).
NuWa: Deriving Lightweight Class-Specific Vision Transformers for Edge Devices
Ziteng Wei (Huazhong University of Science and Technology), Yun Yang (Swinburne University of Technology)
ClassificationComputational EfficiencyTransformerImage
🎯 What it does: To address the resource constraints of edge devices, the NuWa method is proposed, which can quickly derive lightweight models tailored to specific categories from pre-trained Vision Transformers (ViT) without requiring post-training fine-tuning.
NVGS: Neural Visibility for Occlusion Culling in 3D Gaussian Splatting
Brent Zoomers (Hasselt University), Nick Michiels (Hasselt University)
Computational EfficiencyGaussian SplattingMesh
🎯 What it does: Implement visibility learning and instance culling in the 3D Gaussian Splatting scenario by using a small MLP to predict view-dependent visibility and cull occluded Gaussians before rendering, significantly reducing memory usage and improving frame rate.
OASIS: On-Demand Hierarchical Event Memory for Streaming Video Reasoning
Zhijia Liang (Sun Yat Sen University), Guanbin Li (Sun Yat Sen University)
Computational EfficiencyLarge Language ModelVideoRetrieval-Augmented Generation
🎯 What it does: Proposes the OASIS framework, using hierarchical event memory and two-stage reasoning to achieve streaming video inference.
Object-Generalized Re-Identification: A Step Towards Universal Instance Perception
Shuoyi Chen (Wuhan University), Mang Ye (Wuhan University)
RecognitionDomain AdaptationRepresentation LearningMeta LearningTransformerImageBenchmark
🎯 What it does: Proposes a cross-category generalizable object re-identification paradigm OG-ReID and constructs the MGOR framework to achieve target-domain-free adaptation for universal identity representation learning.
Object-WIPER: Training-Free Object and Associated Effect Removal in Videos
Saksham Singh Kushwaha (University of Texas at Dallas), Kuldeep Kulkarni (Adobe Research)
RestorationGenerationTransformerDiffusion modelVideo
🎯 What it does: This work proposes a training-free framework called Object-WIPER based on a pre-trained text-to-video diffusion transformer (DiT), which enables complete removal of target objects along with their shadows, reflections, and related visual effects in videos, achieving semantically consistent and temporally coherent inpainting.
ObjectMorpher: 3D-Aware Image Editing via Deformable 3DGS
Yuhuan Xie (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
Image HarmonizationGenerationDiffusion modelGaussian SplattingImagePoint Cloud
🎯 What it does: Propose ObjectMorpher, a method that elevates target instances in 2D images to high-quality 3D Gaussian Splatting (3DGS) via an image-to-3D generative model, allowing users to perform real-time non-rigid deformation in 3D space by dragging control points, and finally using a diffusion model to integrate lighting, color, and boundaries of the edited results, achieving fine-grained, controllable, and physically plausible image editing.
Obstruction Reasoning for Robotic Grasping
Runyu Jiao (Fondazione Bruno Kessler), Fabio Poiesi (Fondazione Bruno Kessler)
Robotic IntelligenceSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose UNOGrasp, a Vision-Language Model (VLM) capable of performing obstacle reasoning based on vision and language and generating grasping plans, and create a large-scale UNOBench dataset;
OccAny: Generalized Unconstrained Urban 3D Occupancy
Anh-Quan Cao (Valeo.ai), Tuan-Hung Vu (Valeo.ai)
Autonomous DrivingKnowledge DistillationTransformerNeural Radiance FieldImagePoint Cloud
🎯 What it does: Proposes OccAny, an uncalibrated, domain-agnostic urban 3D occupancy prediction model capable of performing zero-shot inference on sequences, monocular, or panoramic images while generating high-quality occupancy and segmentation features.
Occluded Human Body Capture with Frequency Domain Denoising Prior
Buzhen Huang (Southeast University), Yangang Wang (Southeast University)
Pose EstimationTransformerDiffusion modelVideoBenchmark
🎯 What it does: Propose a monocular video occlusion human motion capture method based on frequency domain denoising prior, modeling occluded motion as a wavelet coefficient selection problem.
Occlusion-Aware SORT: Observing Occlusion for Robust Multi-Object Tracking
Chunjiang Li (Sichuan University), Liangyin Chen (Sichuan University)
Object TrackingVideo
🎯 What it does: Proposed Occlusion-Aware SORT (OA-SORT), which enhances position association and Kalman filter stability in multi-object tracking by observing occlusions and integrating three training-free plugins: Occlusion-Aware Module, Occlusion-Aware Offset, and Bias-Aware Momentum.
OccuFly: A 3D Vision Benchmark for Semantic Scene Completion from the Aerial Perspective
Markus Gross (Fraunhofer Institute IVI), Henri Meeß (Fraunhofer Institute IVI)
SegmentationDepth EstimationImagePoint CloudBenchmark
🎯 What it does: This paper introduces the OccuFly dataset, the first to establish a 3D semantic scene benchmark from an aerial perspective using a camera.
OctoMed: Data Recipes for State-of-the-Art Multimodal Medical Reasoning
Timothy Ossowski (University of Wisconsin Madison), Hoifung Poon (Microsoft Research)
Knowledge DistillationData-Centric LearningTransformerSupervised Fine-TuningMultimodalityBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: Constructed the largest-scale and most structured medical multimodal reasoning dataset, and transferred the teacher model's reasoning capabilities to the student model through supervised fine-tuning (SFT), proposing a data recipe centered on diversified reasoning trajectory lengths and cross-modal coverage.
OctoNav: Towards Generalist Embodied Navigation
Chen Gao (Beihang University), Si Liu (Beihang University)
Robotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the OctoNav-Bench benchmark and the OctoNav-R1 model, supporting generalized navigation with multi-modal, multi-capability free instructions;
Octopus: History-Free Gradient Orthogonalization for Continual Learning in Multimodal Large Language Models
Yuehao Liu (Shanghai Jiao Tong University), Chao Ma (Shanghai Jiao Tong University)
TransformerLarge Language ModelSupervised Fine-TuningMultimodality
🎯 What it does: Proposes the Octopus framework to achieve continuous learning without historical data in multimodal large language models; mitigates catastrophic forgetting through two-stage fine-tuning and historical-free gradient orthogonalization (HiFGO).
OctoT2I: A Self-Evolving Agentic Text-to-Image Router
Xu Jiang (Peking University), Jian Zhang (Peking University)
GenerationComputational EfficiencyTransformerLarge Language ModelAgentic AIImageText
🎯 What it does: Proposed and implemented OctoT2I, a multi-round adaptive router that leverages LLM to control multiple text-to-image (T2I) models, achieving efficient and quality-controllable text-to-image generation.
OddGridBench: Exposing the Lack of Fine-Grained Visual Discrepancy Sensitivity in Multimodal Large Language Models
Tengjin Weng (Shenzhen University), Zhong Ming (Tsinghua Shenzhen International Graduate School)
Large Language ModelReinforcement LearningVision Language ModelImageBenchmark
🎯 What it does: Proposed the OddGridBench benchmark for systematically evaluating the perceptual sensitivity of multi-modal large language models (MLLMs) in fine-grained visual difference detection; simultaneously designed the OddGrid-GRPO reinforcement learning framework, enhancing the model's visual difference discrimination ability through curriculum training and distance-aware rewards.
ODGS-SLAM: Omnidirectional Gaussian Splatting SLAM
Stefan Spiss (University of Innsbruck), Matthias Harders (University of Innsbruck)
OptimizationGaussian SplattingSimultaneous Localization and MappingMultimodality
🎯 What it does: Developed a panoramic visual SLAM system, ODGS-SLAM, based on 3D Gaussian distribution rendering, which can simultaneously achieve localization and dense mapping.
Off The Grid: Detection of Primitives for Feed-Forward 3D Gaussian Splatting
Arthur Moreau (Huawei Noah's Ark Lab), Eduardo Pérez-Pellitero (Huawei Noah's Ark Lab)
GenerationConvolutional Neural NetworkGaussian SplattingImagePoint Cloud
🎯 What it does: Propose Off-The-Grid 3D Gaussian Splatting, leveraging sub-pixel keypoint detection to locate high-quality primitives at sub-pixel precision and achieving fast, pose-agnostic 3D reconstruction through self-supervised training.
OLATverse: A Large-scale Real-world Object Dataset with Precise Lighting Control
Xilong Zhou (Max Planck Institute for Informatics), Christian Theobalt (Max Planck Institute for Informatics)
Data SynthesisMultimodalityBenchmark
🎯 What it does: Created the OLATverse large-scale real-object dataset, containing approximately 9 million images, 765 objects, and providing multimodal data including precise lighting, camera parameters, object masks, normals, and diffuse reflectance.
Olbedo: An Albedo and Shading Aerial Dataset for Large-Scale Outdoor Environments
Shuang Song (Ohio State University), Rongjun Qin (Ohio State University)
RestorationData SynthesisConvolutional Neural NetworkSupervised Fine-TuningDiffusion modelImageBenchmark
🎯 What it does: Propose and release the Olbedo dataset, achieving albedo and shadow separation in large-scale real outdoor environments from a drone perspective
OlmoEarth: Stable Latent Image Modeling for Multimodal Earth Observation
Henry Herzog (Allen Institute for AI), Patrick Beukema (Allen Institute for AI)
TransformerSupervised Fine-TuningAuto EncoderContrastive LearningImageMultimodalityTime SeriesBenchmark
🎯 What it does: Built a multi-modal, multi-temporal Earth observation foundation model called OlmoEarth, adopting a self-supervised learning framework that supports multi-source data such as remote sensing images and maps.
OMG-Avatar: One-shot Multi-LOD Gaussian Head Avatar
Jianqiang Ren, Steven Hoi (Alibaba Group)
GenerationTransformerGaussian SplattingVideo
🎯 What it does: Propose OMG-Avatar, a method for instant generation of multi-resolution, high-quality, animatable Gaussian head models from a single image.
OMG-Bench: A New Challenging Benchmark for Skeleton-based Online Micro Hand Gesture Recognition
Haochen Chang (Sun Yat-sen University), Erwei Yin (Academy of Military Sciences)
RecognitionPose EstimationGraph Neural NetworkTransformerContrastive LearningGraphBenchmark
🎯 What it does: Constructed the OMG-Bench large-scale hand skeleton online micro-gesture recognition dataset, and proposed the Hierarchical Memory-Augmented Transformer (HMATr), an end-to-end framework that unifies detection and classification.
OMGTex: One-stage Multi-style Facial Texture Reconstruction without Geometry Guidance
Zitong Xiao (Chinese University of Hong Kong), Xiaoguang Han (Chinese University of Hong Kong)
GenerationTransformerDiffusion modelFlow-based ModelImage
🎯 What it does: Proposed a geometry-free end-to-end diffusion framework called OMGTex for reconstructing high-quality editable UV textures from single multi-style human face images.
Omni IIE Bench: Benchmarking the Practical Capabilities of Image Editing Models
Yujia Yang (University of Chinese Academy of Sciences), Hongzhu Yi (Tencent)
MultimodalityBenchmark
🎯 What it does: Propose the Omni IIE Bench benchmark to evaluate the single-round consistency and multi-round coordination capabilities of instruction-driven image editing models, ensuring data quality through rigorous manual screening.
Omni-3DEdit: Generalized Versatile 3D Editing in One-Pass
Liyi Chen (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
RestorationData SynthesisPose EstimationOptimizationGaussian SplattingImagePoint Cloud
🎯 What it does: This paper proposes a unified one-time 3D editing framework called Omni-3DEdit, capable of performing multiple editing tasks such as object deletion, addition, and appearance modification in a multi-view latent space.
Omni-AD: A Large-scale and Versatile Benchmark for Industrial Anomaly Detection
Dahu Shi (Zhejiang University), Xing Wei (Xi'an Jiaotong University)
Anomaly DetectionConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelImageTextBenchmark
🎯 What it does: Proposes Omni-AD—a large-scale industrial anomaly detection dataset containing approximately 35,000 images, 150 product categories, and 16 industrial scenarios. Three MLLM evaluation subtasks (anomaly discrimination, classification, and localization) are defined on this dataset, along with detailed annotation and QA generation processes.
Omni-Attack: Adversarial Attacks on Open-Ended VQA in Black-Box Multimodal LLMs
Kai Hu (Carnegie Mellon University), Matt Fredrikson (Carnegie Mellon University)
Adversarial AttackLarge Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper constructs an adversarial robustness benchmark named AdvRobustBench for multi-modal large language models and proposes Omni-Attack, a transfer-based black-box attack method.
Omni-Attribute: Open-vocabulary Attribute Encoder for Visual Concept Personalization
Tsai-Shien Chen (Snap Inc.), Sergey Tulyakov (Snap Inc.)
GenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelFlow-based ModelContrastive LearningMultimodality
🎯 What it does: Propose an open-vocabulary attribute encoder called Omni-Attribute, which extracts attribute-specific representations from images and enables attribute personalization and combination generation in new visual contexts.
Omni-Fake: Benchmarking Unified Multimodal Social Media Deepfake Detection
Tianxiao Li (University of Liverpool), Guangliang Cheng (University of Liverpool)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageVideoMultimodalityBenchmarkAudio
🎯 What it does: This paper constructs a unified four-modal deepfake detection benchmark called Omni-Fake, and proposes the Omni-Fake-R1 model based on Qwen2.5-Omni-7B, which can simultaneously accomplish detection, localization, and interpretation.
Omni-MMSI: Toward Identity-attributed Social Interaction Understanding
Xinpeng Li (University of Texas at Dallas), Yapeng Tian (University of Texas at Dallas)
RecognitionObject DetectionTransformerLarge Language ModelSupervised Fine-TuningVideoTextMultimodalityRetrieval-Augmented GenerationChain-of-ThoughtAudio
🎯 What it does: Propose the Omni-MMSI task and the Omni-MMSI-R pipeline to address the problem of identifying identity attribution from raw audio and video and reasoning about multi-party social interactions.
Omni-Supervised Motion Editing: Balancing Change and Invariance through Positive-Negative Learning
Zhenwu Shi (East China Normal University), Shaohui Lin (East China Normal University)
GenerationRetrievalTransformerDiffusion modelContrastive LearningTextSequential
🎯 What it does: Propose an Omni-Supervised Positive-Negative Learning framework called OmniME for natural language-based 3D human motion editing, which can precisely modify target regions while preserving the original motion unchanged.
Omni2Sound: Towards Unified Video-Text-to-Audio Generation
Yusheng Dai (Monash University), Jun Zhu (Monash University)
GenerationData SynthesisTransformerAgentic AIVision Language ModelDiffusion modelMultimodalityBenchmark
🎯 What it does: Proposed a unified audio generation framework called Omni2Sound, which can simultaneously support three generation tasks: video-text-audio (VT2A), video-audio (V2A), and text-audio (T2A).
OmniBrainBench: A Comprehensive Multimodal Benchmark for Brain Imaging Analysis Across Multi-stage Clinical Tasks
Zhihao Peng (Chinese University of Hong Kong), Yixuan Yuan (Chinese University of Hong Kong)
TransformerLarge Language ModelVision Language ModelContrastive LearningMultimodalityBiomedical DataPositron Emission TomographyBenchmark
🎯 What it does: Created an end-to-end multimodal evaluation benchmark called OmniBrainBench, covering 15 brain imaging modalities, 5 clinical stages, and 9,527 VQA pairs.
OmniDocLayout: Towards Diverse Document Layout Generation via Coarse-to-Fine LLM Learning
Hengrui Kang (Shanghai Jiao Tong University), Conghui He (Shanghai Artificial Intelligence Laboratory)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMultimodality
🎯 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.
OmniFM: Toward Modality-Robust and Task-Agnostic Federated Learning for Heterogeneous Medical Imaging
Meilin Liu (Shenyang University of Technology), Jing Shan (Shenyang University of Technology)
ClassificationSegmentationSuper ResolutionRetrievalFederated LearningTransformerImageTextMultimodalityBiomedical Data
🎯 What it does: Propose OmniFM, a general and frequency-aware federated learning framework that works across various medical imaging tasks (classification, segmentation, super-resolution, VQA, multi-modal fusion) and heterogeneous modalities without requiring rewriting of the optimization process.
OmniFood8K: Single-Image Nutrition Estimation via Hierarchical Frequency-Aligned Fusion
Dongjian Yu (Yunnan University), Shuqiang Jiang (Chinese Academy of Sciences)
ClassificationData SynthesisImageMultimodalityBenchmark
🎯 What it does: Constructed the OmniFood8K dataset containing 8,036 Chinese dishes, proposed the NutritionSynth115K synthetic dataset, and designed an end-to-end nutritional estimation framework based on single RGB images.
OmniGen2: Towards Instruction-Aligned Multimodal Generation
Chenyuan Wu (Beijing Academy of Artificial Intelligence), Zheng Liu (University of Science and Technology of China)
GenerationTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelDiffusion modelRectified FlowAuto EncoderImageVideoTextMultimodalityBenchmark
🎯 What it does: Designed and trained a unified multimodal generation model called OmniGen2, which supports text→image generation, image editing, and context generation, achieving systematic instruction alignment.
OmniGround: A Comprehensive Spatio-Temporal Grounding Benchmark for Real-World Complex Scenarios
Hong Gao (Southeast University), Min-Ling Zhang (Southeast University)
Object DetectionObject TrackingLarge Language ModelVision Language ModelVideoBenchmark
🎯 What it does: Constructed the OmniGround large-scale video localization benchmark, proposed a multi-directional forward-backward tracking + voting (FBR) efficient annotation process, the DeepSTG multi-dimensional quality evaluation framework, and a training-agnostic two-stage PG-TAF localization method.
OmniLottie: Generating Vector Animations via Parameterized Lottie Tokens
Yiying Yang, Xingjun Ma
GenerationTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodalityBenchmark
🎯 What it does: Propose OmniLottie, an autoregressive Lottie generation framework based on a pre-trained Vision-Language Model (VLM), capable of generating high-quality editable vector animations from text, image, or video multimodal instructions.
OmniSonic: Towards Universal and Holistic Audio Generation from Video and Text
Weiguo Pian (University of Texas at Dallas), Yapeng Tian (University of Texas at Dallas)
GenerationTransformerMixture of ExpertsVision Language ModelDiffusion modelFlow-based ModelVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Proposes the UniHAGen task, designing the OmniSonic model to achieve panoramic audio synthesis under joint conditions of video and text, covering environmental sounds and human voices both inside and outside the screen.
OmniVGGT: Omni-Modality Driven Visual Geometry Grounded Transformer
Haosong Peng (Hong Kong University of Science and Technology), Ziwei Liu (Nanyang Technological University)
Pose EstimationDepth EstimationRobotic IntelligenceTransformerContrastive LearningImageMultimodality
🎯 What it does: Propose OmniVGGT, a spatial foundation model that can leverage an arbitrary number of auxiliary geometric modalities (e.g., depth maps, camera intrinsics, extrinsics) during both training and inference phases, to improve multi-task 3D perception;
OmniVTG: A Large-Scale Dataset and Training Paradigm for Open-World Video Temporal Grounding
Minghang Zheng (Peking University), Yang Liu (Peking University)
RetrievalTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 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)
CompressionComputational EfficiencyTransformerLarge Language ModelVideoMultimodalityAudio
🎯 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)
CompressionTransformerMixture of ExpertsImageTextMultimodalityTabularBiomedical DataAudio
🎯 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;
OMoBlur: An Object Motion Blur Dataset and Benchmark for Real-World Local Motion Deblurring
Dingchuan Yu (Zhejiang University), Qi Li (Zhejiang University)
RestorationConvolutional Neural NetworkOptical FlowImageBenchmark
🎯 What it does: Proposed a large-scale physically realistic local motion blur dataset called OMoBlur, and designed a dedicated deblurring network OMDNet based on this dataset;
On the Role of Temporal Granularity in the Robustness of Spiking Neural Networks
Mengting Xu (Zhejiang University), Gang Pan (Zhejiang University)
OptimizationAdversarial AttackSpiking Neural NetworkImageTime Series
🎯 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.
On Token's Dilemma: Dynamic MoE with Drift-Aware Token Assignment for Continual Learning of Large Vision Language Models
Chongyang Zhao (University of New South Wales), Dong Gong (University of New South Wales)
TransformerLarge Language ModelMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: Proposed a new dynamic Mixture of Experts (MoE) framework, LLaVA-DyMoE, to address catastrophic forgetting in large vision-language models during continuous learning caused by routing drift.
One Algorithm to Align Them All
Boyi Pang (Harbin Institute of Technology), Evgeny Burnaev (Harbin Institute of Technology)
GenerationData SynthesisFlow-based ModelRectified FlowImageVideoMultimodalityMesh
🎯 What it does: This work proposes a general multimodal algorithm that utilizes the Rectified Flow model to achieve joint reasoning, thereby generating structure-aligned image, video, and 3D model pairs; by performing velocity-guided joint transport on a line segment in the latent space and incorporating smooth regularization and anchor point velocity correction, it can achieve structure-consistent dual-sample generation across multiple domains without retraining the model.
One Layer's Trash is Another Layer's Treasure: Adaptive Layer-wise Visual Token Selection in LVLMs
Yongru Chen (Hikvision Research Institute), Jilin Hu (East China Normal University)
Computational EfficiencyVision Language ModelMultimodality
🎯 What it does: Proposed a hierarchical adaptive visual token selection framework ALVTS for dynamically compressing visual tokens during LVLM inference.
One Model, Many Budgets: Elastic Latent Interfaces for Diffusion Transformers
Moayed Haji-Ali (Rice University), Aliaksandr Siarohin (Snap Inc.)
GenerationComputational EfficiencyTransformerDiffusion modelRectified FlowImageVideoMultimodality
🎯 What it does: Propose a scalable latent interface that decouples the computation of Diffusion Transformers from input resolution through a lightweight Read/Write cross-attention module, enabling dynamic adjustment of inference budget.
One Patch to Caption Them All: A Unified Zero-Shot Captioning Framework
Lorenzo Bianchi (ISTI-CNR), Fabrizio Falchi (ISTI-CNR)
GenerationTransformerVision Language ModelDiffusion modelImageText
🎯 What it does: A unified zero-shot image/region captioning framework is proposed, adopting a Patch-Centric approach based on image patches, which can generate natural language descriptions for arbitrary sub-regions (single patches, non-continuous regions, mouse trajectories, etc.) without requiring region-level supervision.
One Token, Two Fates: A Unified Framework via Vision Token Manipulation Against MLLMs Hallucination
Zhan Fa (Nanjing University), Yinghuan Shi (Nanjing University)
Explainability and InterpretabilityComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: Propose a unified training-free framework that alleviates hallucinations in multimodal large language models by enhancing and trimming visual tokens
One-Shot Flow, Any-Time Frame: A Bidirectional Warping Framework for Event-Based Video Frame Interpolation
Linghui Fu (Beijing University of Technology), Yongjian Deng (Nankai University)
RestorationGenerationRecurrent Neural NetworkOptical FlowTime SeriesSequential
🎯 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.