ECCV 2026 论文列表

Knowledge-based Visual Question Answering (KB-VQA) re-quires models to retrieve visual entities matching the query image fromlarge-scale encyclopedic knowledge bases and answer related questions.Existing multimodal Retrieval Augmented Generation (mRAG) meth-ods rely on global visual features to match candidate entities, yet whenthe knowledge base contains numerous visually similar entities, the re-triever struggles to distinguish them, populating the candidate set withvisually similar but factually mismatched distractors. Since subsequentprocessing steps such as noise filtering are also confined to this fixedcandidate set, errors from failed retrieval inevitably propagate to the fi-nal answer. To address these challenges, we propose MMAgent-R2 , anagentic mRAG framework that integrates visual reranking and active re-jection as its internal verification mechanism. Visual reranking directlycompares query and candidate images, capturing discriminative detailsbeyond textual descriptions to precisely identify the target entity amongsimilar candidates; active rejection discards unreliable results and re-trieves additional candidates when no confident match is found, movingbeyond the fixed candidate pool. We design a composite reward functionwith step-level verification rewards and achieve joint optimization of ex-ternal retrieval, internal verification, and answer generation via GRPOtraining. Experiments on InfoSeek, E-VQA, and MMhops demonstratethat MMAgent-R2 achieves state-of-the-art performance, with particu-larly notable advantages in challenging retrieval scenarios and complexmulti-image multi-hop reasoning tasks.

Concealed object segmentation aims to segment objects thatblend into their surroundings, presenting significant challenges due tothe high similarity between objects and backgrounds. Existing methodsrely on task-specific designs and independent training, leading to limitedgeneralization and structural redundancy. Although a unified parameter-shared model for diverse concealed scenarios is highly desirable andpromising, it is hindered by two challenges: large representation gapsacross scenarios and the intrinsic difficulty of concealed targets. From afrequency-domain perspective, we observe that different concealed sce-narios exhibit more discriminative and interpretable spectral character-istics compared to those in the RGB domain. These findings motivate usto approach unified concealed scene segmentation from the perspectiveof frequency modulation. To this end, we propose Frequency Director(FreqDirect), a dynamic frequency adaptation framework that directsspectral representations across diverse concealed scenes. It features amixture of frequency experts to perform adaptive routing over frequencycomponents, capturing heterogeneous task-specific patterns, along witha spatial commonality anchor to complement the spatial and structurecues. These components collectively enable effective and efficient adap-tation to diverse concealed segmentation tasks within a single model.Extensive experiments on eight concealed scenarios demonstrate thatFreqDirect surpasses existing unified models and achieves performancecomparable to, or better than, specialist models.

Change captioning is the task of generating natural languagedescriptions that explain the changes between a pair of images. Althoughdifferent change types (e.g., color shifts, object additions) exhibit distinctvisual cues and require specialized reasoning processes, existing methodsoften overlook these distinctions. To address this limitation, we proposeMulti-Expert Diagnosis for Image Change (MEDIC), a novel frameworkthat introduces change-type awareness by explicitly modeling changecategories. We build our MEDIC as a memory network to dynamicallyretrieve type-relevant visual patterns conditioned on the input. This de-sign allows each expert to flexibly capture diverse variations within eachchange type and focus on the most informative cues for its designatedchange type. By routing inputs through type-specialized experts andlearning dedicated representations for each change category, MEDICgenerates more precise and type-aware change descriptions. Extensiveexperiments demonstrate that proposed MEDIC consistently outper-forms across diverse and challenging datasets. The code is available athttps://github.com/VisualAIKHU/MEDIC.

Multi-view 3D Visual Question Answering (MV3D-VQA) re-quires integrating partial observations into a coherent 3D scene repre-sentation and selecting informative viewpoints for multi-step spatial rea-soning. However, current multimodal LLMs are typically trained withsparse, answer-level supervision, which often yields inconsistent cross-view reasoning and brittle view selection. We present DR-MV3D (DenseReward for MV3D-VQA), a map-grounded learning framework that pro-vides dense, verifiable rewards to supervise the reasoning process. Our ap-proach decomposes MV3D-VQA into (i) allocentric global map construc-tion, (ii) question-conditioned view-trajectory planning, and (iii) egocen-tric grounding for answer prediction. To make intermediate steps learn-able without manual annotations, we introduce two rewards: a global con-sistency reward that aligns the predicted map with geometry-consistentpseudo targets from frozen 3D vision foundation models (e.g., VGGT+ SAM3), and a local trajectory reward that supervises ordered view-point selection. We optimize the full pipeline with trajectory-level pol-icy optimization (GRPO). Experiments on MindCube, VSI-Bench, andBLINK (MV) show that DR-MV3D consistently improves over strongmulti-image baselines, supporting the effectiveness of process-level densesupervision for multi-view 3D reasoning.Code is available at: https://github.com/kaist-cvml/dr-mv3d

Linear probing on frozen features is the standard evalua-tion protocol for vision foundation models, yet classical theory for d-dimensional classifiers demands Ω(d) labeled samples, far more thanpractitioners actually need. We resolve this discrepancy with the first non-vacuous PAC-Bayes generalization bounds for linear probing on frozenvision foundation-model features. The key insight is that foundation-modelfeatures do not fill Rd : covariance spectra decay steeply, concentratingvariance in a subspace of effective dimension deff ≪ d. A data-dependentGaussian prior aligned to this spectral structure yields a certificate whosecomplexity is governed by deff and the classification margin γ rather thanthe ambient dimension d. We validate the theory on ImageNet across 12vision encoders spanning discriminative, language-supervised, generative,and predictive pretraining: the bound is non-vacuous for all discrimi-natively pretrained models and ranks among the tightest PAC-Bayescertificates reported for any vision model. Three empirical predictionsfollow: generative pretraining yields markedly less certifiable features thandiscriminative pretraining, deff alone ranks foundation models withoutany labels, and larger models tend to be more certifiable because scalesteepens spectral decay.

Understanding and forecasting audience reactions to videocontent are crucial for improving content creation, recommendation sys-tems, and media analysis. To enable audience reaction prediction andother content engagement applications, we introduce Video2Reaction,a multimodal dataset that maps short movie segments to a distributionof induced emotions of viewers in the wild, as expressed through socialmedia. Video2Reaction spans more than 10,000 videos and serves asa reliable benchmark as well as a training resource for audience reactionprediction. To enable cost-e!ective continuous annotations as reactionsmay change over time, we develop a two-stage multi-agent pipeline usingonly open-source LLMs, achieving 86% correctness under blind humanverification despite the inherently noisy and subjective nature of the task.We establish the first benchmark for video-to-reaction-distribution pre-diction in the wild and show that pretrained foundation video modelsfail in zero-shot settings, while finetuning transforms them into state-of-the-art predictors capable of modeling both full reaction distribu-tions and dominant responses from video alone. However, the task re-mains challenging: even the strongest methods achieve only 77% Top-3F1 in dominant reaction prediction (LLaVA-Next), highlighting a sub-stantial gap in modeling collective audience reaction. Dataset and codeare available at our project page: https://information-fusion-lab-umass.github.io/video2reaction-bench.github.io/.

Although diffusion-based real-world image restoration (Real-IR) has achieved remarkable progress, efficiently leveraging ultra-large-scale pre-trained text-to-image (T2I) models and fully exploiting theirpotential remain significant challenges. To address this issue, we proposeResFlow-Tuner, an image restoration framework based on the state-of-the-art flow matching model, FLUX.1-dev, which integrates unifiedmulti-modal fusion (UMMF) with test-time scaling (TTS) to achieveunprecedented restoration performance. Our approach fully leveragesthe advantages of the Multi-Modal Diffusion Transformer (MM-DiT) ar-chitecture by encoding multi-modal conditions into a unified sequencethat guides the synthesis of high-quality images. Furthermore, we in-troduce a training-free test-time scaling paradigm tailored for imagerestoration. During inference, this technique dynamically steers the de-noising direction through feedback from a reward model (RM), therebyachieving significant performance gains with controllable computationaloverhead. Extensive experiments demonstrate that our method achievesstate-of-the-art performance across multiple standard benchmarks. Thiswork not only validates the powerful capabilities of the flow match-ing model in low-level vision tasks but, more importantly, proposes anovel and efficient inference-time scaling paradigm suitable for largepre-trained models. The source code is publicly available at https://github.com/Rorschach-1010/ResFlow-Tuner.

In visual object tracking, SAM2 stands out in deformation adaptability and distinguishing similar objects due to its pixel-level masks, whereas traditional discriminative trackers have advantages under occlusion and motion blur thanks to global semantic feature matching. To leverage this complementarity, we propose TR-MoE, a Temporal Reliability-aware Mixture-of-Experts framework. Unlike conventional fusion strategies that rely on static or per-frame confidence, we reframe tracker fusion as a time-series modeling problem. Specifically, we introduce a Mamba-based Trajectory Refiner (MTR) that processes the historical sequence of tracker states—including relative deviations and consistency scores—to capture the temporal evolution of error regimes. This enables the model to adaptively allocate fusion weights among experts based on historical temporal context. Furthermore, to achieve precise geometric correction, we design a KAN-enhanced Disentangled Head. Leveraging Kolmogorov–Arnold Networks, this module factorizes the refinement into independent directional offsets, dynamic scale magnitudes, and fusion weights. Extensive experiments show that TR-MoE effectively adapts to varying failure patterns, significantly boosting robustness and establishing new state-of-the-art performance. The source code is available at https://github.com/Avant1v/TR-MoE.

We introduce PPTArena, a benchmark for PowerPoint edit-ing that evaluates how agents modify real slides from natural-languageinstructions. Unlike benchmarks that rely on image-PDF renderings ortext-to-slide generation, PPTArena features 100 decks with over 1,300human-curated edits across 2,125 slides, spanning text, charts, anima-tions, and professional master styles. Each edit pairs a ground-truthdeck with a target rubric and is scored by two Vision-Language Model(VLM) judges: one rates instruction following from structural diffs, theother visual quality from slide images. On top of this benchmark, wepresent PPTPilot, a structure-aware agent that plans semantic edit se-quences, routes between programmatic tools and deterministic XMLoperations, and verifies each result in an iterative plan-edit-check loop.PPTPilot outperforms strong VLM-based agents by more than 10 per-centage points on compound, layout-sensitive, and cross-slide edits, withlarge gains in visual fidelity and deck-wide consistency. Despite this, allagents still struggle on long-horizon, document-scale tasks, underscoringhow hard reliable PowerPoint editing remains. We publicly release ourcode: https://github.com/michaelofengend/PPTArena.

Parametric body models provide the structural basis formany human-centric tasks, yet existing models often rely on costly 3Dscans and learned shape spaces that are proprietary and demographi-cally narrow. We introduce Anny, a simple, fully differentiable, and scan-free human body model grounded in anthropometric knowledge from theMakeHuman community. Anny defines a continuous, interpretable shapespace, where phenotype parameters (e.g. gender, age, height, weight)control blendshapes spanning a wide range of human forms—across ages(from infants to elders), body types, and proportions. Calibrated usingWHO population statistics, Anny provides realistic and demographicallygrounded human shape variation within a single unified model. We re-lease the Anny body model and its code under the Apache 2.0 license.Thanks to its openness and semantic control, Anny serves as a versa-tile foundation for 3D human modeling—supporting millimeter-accuratescan fitting, controlled synthetic data generation, and Human Mesh Re-covery (HMR). We further introduce Anny-One, a collection of 780kphotorealistic images generated with Anny, showing that despite its sim-plicity, HMR models trained with Anny can match the performance ofthose trained with scan-based body models.

Few-step video generation has been significantly advancedby consistency distillation. However, the performance of consistency-distilled models often degrades as more sampling steps are allocatedat test time, limiting their effectiveness for any-step video diffusion.We argue that this limitation arises because consistency distillation re-places the original probability-flow ODE trajectory with a consistency-sampling trajectory, weakening the desirable test-time scaling behaviorof ODE sampling. To address this limitation, we introduce AnyFlow,the first any-step video diffusion distillation framework based on flowmaps. Instead of distilling a model for only a few fixed sampling steps,AnyFlow optimizes the full ODE sampling trajectory. To this end, weshift the distillation target from endpoint consistency mapping (zt → z0 )to flow-map transition learning (zt → zr ) over arbitrary time inter-vals. We further propose Flow Map Backward Simulation, whichdecomposes a full Euler rollout into shortcut flow-map transitions, en-abling efficient on-policy distillation that reduces test-time errors (i.e.,discretization error in few-step sampling and exposure bias in causalgeneration). Extensive experiments across both bidirectional and causalarchitectures, at scales ranging from 1.3B to 14B parameters, demon-strate that AnyFlow achieves performance comparable to or better thanconsistency-based counterparts in the few-step regime, while support-ing flexible and scalable sampling under varying step budgets. Code isreleased at https://github.com/NVLabs/AnyFlow.

Vision–Language Models (VLMs) are increasingly deployedin safety-critical applications, yet remain vulnerable to backdoor attacks.Existing methods primarily manipulate final outputs, often producingreasoning traces that are inconsistent or easily detectable. In this paper,we propose ReShift, the novel aha-moment-driven reasoning-level back-door framework that explicitly redirects the internal chain-of-thought(CoT) trajectory while preserving surface-level coherence. ReShift intro-duces a Poisoned Reasoning-Aware Data Construction (PRDC) pipelineand a Supervised–Reinforcement Joint Optimization (SRJO) strategyto induce stable trigger-conditioned reasoning shifts. We further for-malize Entropy Rebound as a principled signal for characterizing rea-soning redirection and provide theoretical guarantees linking entropygaps to trajectory-level divergence. Extensive experiments demonstratethat ReShift achieves high attack success rates while maintaining clean-task performance and realistic reasoning traces, substantially improv-ing stealthiness against existing defenses. Code can be found at https://github.com/AlbertZhaoCA/ReShift.

In precipitation forecasting, not only accuracy but also tem-poral resolution is critical. However, increasing temporal resolution isconstrained by observational limitations and the computational cost ofdense discrete modeling. To overcome this limitation, we reformulate pre-cipitation forecasting as a continuous-time dynamical system and pro-pose RainODE, a framework that models precipitation evolution in la-tent space using a Neural ODE. This formulation enables derivative-consistent temporal dynamics and captures the dominant large-scaleadvective motion of precipitation systems. Nevertheless, a purely de-terministic ODE struggles to represent non-advective intensity changessuch as localized growth, decay, and sub-grid variability, often lead-ing to over-smoothed predictions. To address this issue, we introducea stochastic source modeling module based on a Brownian Bridge for-mulation, which refines residual intensity variations and restores fine-grained structures while preserving advective consistency. By combiningdeterministic continuous dynamics with stochastic refinement, RainODEenables arbitrary-time inference while maintaining sharp predictions. Ex-periments on SEVIR and the newly introduced Radar-based Precipita-tion Integrated Dataset (RAPID) demonstrate consistent improvementsacross multiple temporal intervals and precipitation regimes. The codeis available at https://github.com/SeongYE/RainODE.

We present GuidedSceneGen, a text-to-3D generation frame-work that produces metrically accurate, globally consistent, and semanti-cally interpretable indoor scenes. Unlike prior text-driven methods thatoften suffer from geometric drift or scale ambiguity, our approach main-tains an absolute world coordinate frame throughout the entire generationprocess. Starting from a textual scene description, we predict a global 3Dlayout encoding both semantic and geometric structure, which serves as aguiding proxy for downstream stages. A semantics- and depth-conditionedpanoramic diffusion model then synthesizes 360° imagery aligned with theglobal layout, substantially improving spatial coherence. To explore un-observed regions, we employ a video diffusion model guided by optimizedcamera trajectories that balances coverage and collision avoidance, achiev-ing up to 10× faster sampling compared to exhaustive path exploration.The generated views are fused using 3D Gaussian Splatting, yielding aconsistent and fully navigable 3D scene in absolute scale. GuidedSceneGenenables accurate transfer of object poses and semantic labels from lay-out to reconstruction, and supports progressive scene expansion withoutre-alignment. Quantitative results and a user study demonstrate greater3D consistency and layout plausibility compared to recent text-to-3Dbaselines. Project page: https://d3ixi.github.io/GuidedSceneGen/

Despite remarkable progress toward general-purpose videomodels, a critical question remains unanswered: how far are these modelsfrom achieving true multimodal reasoning? Existing benchmarks fail toaddress this question rigorously, as they remain constrained by straight-forward task designs and fragmented evaluation metrics that neglectcomplex multimodal reasoning. To bridge this gap, we introduce CLVG-Bench, an evaluation framework designed to probe video models’ zero-shot reasoning capabilities via Context Learning in Video Generation.CLVG-Bench comprises more than 1,000 high-quality, manually anno-tated metadata across 6 categories and 47 subcategories, covering com-plex scenarios including physical simulation, logical reasoning, and inter-active contexts. To enable rigorous and scalable assessment, we furtherpropose an Adaptive Video Evaluator (AVE) that aligns with human ex-pert perception using minimal annotations, delivering interpretable tex-tual feedback across diverse video context tasks. Extensive experimentsreveal a striking answer to our central question: while state-of-the-art(SOTA) video models, such as Seedance 2.0, demonstrate competenceon certain understanding and reasoning subtasks, they fall substantiallyshort with logically grounded and interactive generation tasks (achiev-ing success rates < 25% and ∼ 0%, respectively), exposing multimodalreasoning and physical grounding as critical bottlenecks. By systemati-cally quantifying these limitations, the proposed method provides action-able feedbacks and a clear roadmap toward truly robust, general-purposevideo models. CLVG-Bench and code are released at here.

Pixel-level annotation remains a major bottleneck in medicalimage segmentation, making weak supervision an attractive yet under-constrained alternative. We propose OBBSeg, an intermediate super-vision paradigm guided by Oriented Bounding Boxes (OBBs) thatbridges the gap between full and weak supervision. By jointly encodingspatial extent and orientation, OBBs provide compact geometric super-vision that better aligns with elongated or anisotropic lesions, reducingthe ambiguity of coarse box annotations. To mitigate the inherent rect-angular bias of OBBs, we introduce a Mask-to-OBB loss, a differen-tiable formulation that enforces geometric consistency between predictedmasks and OBB regions. Furthermore, we incorporate prompt-drivensemantic guidance through two complementary modules—PAFE andDBFE—which enhance foreground representation and suppress back-ground interference. Extensive experiments on 13 datasets across 5imaging modalities show that OBBSeg not only outperforms existingweakly supervised methods but also achieves performance comparableto fully supervised approaches, demonstrating its potential for efficientand scalable medical image segmentation. The code is available athttps://github.com/StarLxc3/OBBSeg.

On-device AI for resource-constrained systems, such as wear-ables and IoT sensors, requires models that are both highly efficientand reliable for mission-critical applications. Out-of-Distribution (OOD)detection is vital for ensuring this reliability, yet mainstream OOD re-search has focused on achieving state-of-the-art (SOTA) accuracy withoutregard for inference-time computational cost. This has led to powerful butcomplex methods (e.g., CLIP-based or supervised learning-based models)that are orders of magnitude too slow for real-world on-device deployment.We argue that the more meaningful benchmark for this domain is maxi-mizing accuracy within a practical, constrained inference computationalbudget. However, reducing inference complexity is typically achieved atthe expense of lower performance, exhibiting a trade-off between inferencespeed and accuracy. This paper breaks that trade-off. We introduce anovel framework that establishes an accuracy-SOTA under constrainedinference complexity. This is achieved through two key innovations: (i)a latent-space de-biasing technique that learns to remove pixel-intensitybias during training, drastically reducing inference complexity, and (ii) anintelligent negative sampling strategy that generates high-quality negativesamples to significantly boost performance. On both a real edge deviceand standard GPU infrastructure, extensive experiments reveal thatour method surpasses existing accuracy-SOTA OOD detection methodswithin a TinyML-appropriate inference budget.

An efficient and accurate system for detecting errors in pro-cedural tasks is crucial for supporting human needs in daily life, asit can provide instant notifications and guide people to correct mis-takes. In this work, we study real-time online error detection in pro-cedural videos from a simple but overlooked perspective: the predictionbehavior of action detectors themselves. Instead of designing complexarchitectures or specialized supervision, we observe that action detec-tors naturally exhibit different prediction characteristics depending ontheir sensitivity to input dynamics and temporal context. We thereforepropose ESTANet (Error-Sensitive and Temporally-vArying Network),a lightweight framework that detects errors by exploiting inconsistenciesamong action predictions produced by a small set of action detectors.We construct standard and error-sensitive action detectors that behavesimilarly on correct executions but respond differently when errors occur.Meanwhile, detectors operating with different temporal contexts furtheramplify prediction inconsistencies when the procedure deviates from theintended sequence. During inference, we detect errors by aggregating mis-matches between standard and error-sensitive predictions through ma-jority voting to flag frames that contain errors. Extensive experimentson EgoPER, Assembly-101-O, and EPIC-Tent-O demonstrate that ES-TANet achieves state-of-the-art performance in online error detectionwhile maintaining real-time efficiency with a lightweight architecture.Our results highlight that leveraging the intrinsic properties of actiondetectors can yield a powerful and practical solution for online error de-tection without increasing architectural design complexity. Our code isavailable at: https://github.com/robert80203/ESTANet

Self-supervised pretraining has transformed computer visionby enabling data-efficient fine-tuning, yet high-resolution pretraining typ-ically requires server-scale infrastructure, limiting custom in-domain train-ing for many research laboratories. Masked Autoencoders (MAE) reducecomputation by encoding only visible tokens, but combining MAE withhierarchical downsampling architectures has remained structurally chal-lenging due to dense grid priors and mask-aware design compromises. Weintroduce AFFMAE, a masking-friendly hierarchical pretraining frame-work built on adaptive, off-grid token merging. AFFMAE removes dense-grid assumptions while preserving hierarchical scalability during pre-training and fine-tuning. To support this architecture, we developed nu-merically stable mixed-precision Triton kernels and a lightweight, point-based decoder that can be directly repurposed as a segmentation head.On high-resolution microscopy segmentation, AFFMAE matches MAEfinetuning performance on foot process width estimation with ViT back-bone at equal parameter counts while being 2x faster during pre-trainingand halving peak memory usage. Furthermore, AFFMAE achieves up to5x throughput speedups fine-tuning at the 1024px resolution, providinghigh-resolution model training on desktop hardware. Code available athttps://github.com/najafian-lab/affmae.