ECCV 2026 Papers

Point-supervised crowd counting and localization enable end-to-end training by casting prediction as set matching with one-to-oneHungarian assignment. However, point annotations provide no explicitscale information, making distance-based assignment prone to neighbor-induced ambiguity in congested regions, where nearby ground-truth pointscompete for similar proposals and corrupt supervision. We propose LocalSpacing-Aware Hungarian Matching (SAH-matcher), a drop-in replace-ment that derives a local spacing prior from k-nearest-neighbor distancesand performs per-target rescaling of the geometric cost, inducing a moreselective effective matching region in dense areas while remaining toler-ant in sparse ones. To quantify assignment behavior, we introduce Com-petitive Ambiguity Score (CAS) and Hijacking Rate (HR) for within-epoch ambiguity and severe hijacking failures, and combine them withInstability Rate (IR) to measure cross-epoch consistency. Experiments onmultiple benchmarks show improved assignment reliability and trainingstability, with particularly clear counting and localization gains in denseor locally ambiguous settings, modest overhead, and favorable transfer-ability across the evaluated point-based frameworks. Code is available athttps://github.com/kaijiang77/SAHCC.

Real-world applications require 6D pose estimation to beaccurate, fast, and scalable to unseen objects. This paper introducesWAPR, a zero-shot wide-angle pose re_x001C_nement model that re_x001C_nes candi-date poses with rotational deviations up to 90^0013._Weather-Conditioned_Depth_Anything_1490_Weather_Conditioned_Depth_Anything_ECCV_2026___C....abstract.txt

Accurate object geometry estimation is essential for manydownstream tasks, including robotic manipulation and physical interac-tion. Although vision is the dominant modality for shape perception, itbecomes unreliable under occlusions or challenging lighting conditions.In such scenarios, tactile sensing provides direct geometric informationthrough physical contact. However, reconstructing global 3D geometryfrom sparse local touches alone is fundamentally underconstrained. Wepresent TouchAnything, a framework that leverages a pretrained large-scale 2D vision diffusion model as a semantic and geometric prior for3D reconstruction from sparse tactile measurements. Unlike prior workthat trains category-specific reconstruction networks or learns diffusionmodels directly from tactile data, we transfer the geometric knowledgeencoded in pretrained visual diffusion models to the tactile domain.Given sparse contact constraints and a coarse class-level description ofthe object, we formulate reconstruction as an optimization problem thatenforces tactile consistency while guiding solutions toward shapes con-sistent with the diffusion prior. Our method reconstructs accurate ge-ometries from only a few touches, outperforms existing baselines, andenables 3D reconstruction of previously unseen object instances. Ourproject page is https://grange007.github.io/touchanything/.

Cycle self-training (CST) breaks the shared classifier assumption of the standard self-training framework, which is effective for unsupervised domain adaptation and exploits unlabeled target data by training with target pseudo-labels. CST introduces a target classifier and employs an inner-outer loop updating strategy, addressing the issue of unreliable pseudo-labels and enabling pseudo-labels to generalize across domains. Despite its success in image classification, extending CST to object detection faces three main challenges. First, the upper bound of CST in object detection is constrained by three types of unreliable pseudo-labels, such as classification error alone, localization error alone, and their combination. Second, since object detection involves detecting multiple target objects, directly applying CST leads to training instability. Third, a wider numerical range of regression coordinates leads to exploding losses. To this end, we apply CST to both classification and regression and propose the Dual-Stream Bilevel-Cycle Optimization framework. Specifically, we construct CST upon Mean Teacher to prevent training instability and use extra normalization to map the regression bounding box into a standardized space, effectively addressing exploding losses. Also, we provide a theoretical derivation of the regression bound. Extensive experiments across four cross domain standard scenarios demonstrate that our framework achieves considerable results.

Memorization in machine learning models enables high per-formance on rare in-distribution samples by capturing their atypicalpatterns. However, it also causes harmful retention of noise and out-liers, degrading generalization. While memorization has been extensivelystudied in both supervised and self-supervised learning in the visiondomain, it remains unexplored in multi-modal contrastive learning. Weaddress this gap by introducing MultiMem, the first metric designed toquantify memorization in multi-modal contrastive learning. Through oursystematic analysis, we demonstrate that cross-modal semantic misalign-ment has the strongest influence on memorization, with text being thedominant modality driving memorization, followed by video, image, andaudio. We show that targeted augmentations applied across all modalitieseffectively reduce memorization as measured by our MultiMem metricand improve model performance. Overall, this work establishes the firstframework for measuring and mitigating memorization in multi-modalcontrastive learning, preventing harmful data retention and contributingto higher-performing models. Full version with appendix available athttp://arxiv.org/abs/2606.22220

Ensuring the reliability of deep learning models in real-timeindustrial defect detection is critical for high-stakes quality inspection.To mine uncertain samples within continuous industrial media streams,thereby enhancing the reliability of the detection system, this paperproposes a streaming active learning method based on the Fuzzy Dual-dimensional Uncertainty (FuDU) framework. Specifically, we first designa Prototype-based Global Uncertainty Quantification (PGUQ) moduleon the backbone to evaluate image-level uncertainty via normal/defectivefeature prototypes. A Dual-entropy defect Uncertainty Evaluator (DeUE)is then integrated into the detection head to quantify box-level uncer-tainty. Finally, by modeling uncertainty as systematic error, we pro-pose a fuzzy dual-dimensional uncertainty-aware strategy that lever-ages fuzzy inference to fuse dual-dimensional uncertainties, enabling ex-pert knowledge-driven adaptive sampling decisions. Comprehensive ex-periments demonstrate that FuDU is efficient and flexible, making itwell-suited for challenging industrial inspection tasks such as the de-tection of nuclear fuel rod defects. Our code is publicly available at:https://github.com/wangzhaoyang-508/FuDU.

Understanding human visual attention on a scene over timehas applications in domains such as autonomous driving, interface de-sign, inferring cognitive states. Modeling human visual scanpaths has his-torically relied on specialized architectures with hand-crafted geometricpriors. While these architectures successfully model _x001C_xation sequences,their rigid structural biases restrict easy extendability and _x001D_exible con-ditioning. For instance, integrating task-speci_x001C_c instructions or adaptingto distinct viewer identities traditionally requires custom, disjoint ar-chitectural additions. We address this limitation by framing scanpathprediction purely as a discrete sequence modeling task. By mapping spa-tial coordinates into an autoregressive vocabulary, we leverage the pre-trained representations of Large Vision-Language Models (LVLMs). Thisformulation naturally absorbs diverse factors of variation: simple prompt-ing modi_x001C_cations allow for global conditioning, such as providing vieweridentities to capture personalized biases, or task-speci_x001C_c objectives (likefree-viewing versus visual search). The framework can also easily inte-grate the modeling of per-_x001C_xation attributes, such as individual _x001C_xa-tion durations alongside spatial locations. Crucially, this autoregressivealignment enables the scalable, exact computation of per-_x001C_xation log-likelihoods, directly equivalent to the commonly used Information Gain(IG) metric. Our model, DeepGaze3.5-VL, establishes a new state-of-the-art across multiple datasets, achieving 2.18 bits of IG on MIT1003,a 46% improvement over matched-backbone DeepGaze III. This advan-tage persists even when baselines use identical high-capacity vision en-coders. Beyond predictive performance, our generative framework servesas a powerful computational tool for direct behavioral interventions, al-lowing for controlled in-silico simulations that would be experimentallydi_x001E_cult or impossible to conduct in vivo. We demonstrate this abilityby performing controlled interventions on the durations of pre-saccadic_x001C_xations, recovering known oculomotor phenomena purely from data.

Visuo-Tactile policies leveraging optical tactile sensors haveshown great promise in contact-rich manipulation. These sensors achievehigh spatial resolution and multi-dimensional force sensing by utilizingan internal camera to monitor the deformation of their elastic gel sur-face, thereby indirectly inferring tactile cues. Despite their advantages,extracting fine-grained contact states necessary for contact-rich manipu-lation remains an open challenge. Existing methods typically use eitherraw images or cumulative motion fields to represent tactile cues. However,both are prone to perception ambiguity. Raw tactile images mainly cap-ture appearance changes, while cumulative motion fields only reflect theaggregate gel deformation. Consequently, distinct fine-grained contactstates can exhibit highly similar patterns, making it difficult to explicitlydistinguish subtle contact variations. To address this issue, we explorethe dynamic priors of tactile motion and discover that the correlationbetween transient and cumulative motion can explicitly distinguish fine-grained contact states. Based on this insight, we propose a motion-awaretactile representation to facilitate contact-rich manipulation. Beyond tac-tile representation, effective fusion of tactile and visual modalities is alsocritical. Most existing fusion methods either directly concatenate featuresfrom each modality or train modality-specific networks separately andfuse their outputs. However, these strategies struggle to simultaneouslymodel cross-modal interactions and preserve modality-specific character-istics. In this work, we take advantage of the Mixture-of-Transformersarchitecture and propose a unified modality-aware visuo-tactile policythat captures cross-modal complementarity while maintaining modality-specific properties. Extensive experiments on four challenging contact-rich manipulation tasks indicate the superior performance of our method.

Understanding whether modern vision-language-action poli-cies internally represent task progress is central to interpretable robotcontrol. This paper presents a progress-based SafeSAE-VLA analysis ofOpenVLA on 750 LIBERO episodes, replacing a degenerate binary targetwith a suite-normalized high-vs.-low split over relative geometric progressand testing layer-20 sparse-autoencoder (SAE) features (dsae = 16,384)with non-parametric differential analysis and false-discovery-rate (FDR)control. The recovered signal is strong and sparse. Of 1,881 active fea-tures, 1,117 are significant (FDR < 0.05), a linear probe reaches 0.918AUROC (area under the ROC curve), and top-20 features alone retain0.894 AUROC. Performance is stable across suites (up to 0.985 on goaland 0.947 on object), while motion-only controls are markedly weaker(0.572–0.711), indicating separability is not explained by gross move-ment magnitude. Dense raw-activation probes are equal or stronger thanSAE readouts (e.g., raw LR 0.975 vs. full-SAE 0.947 under a pooledanalysis with bootstrap confidence intervals). We therefore position thecontribution not as predictive dominance but as showing that nearly allof the progress signal survives in a compact set of named, inspectable,intervention-compatible directions. Layer sweeps, split-robustness checks,and a success-labeled audit (in which the progress-tuned top-20 correctlyscores an at-chance 0.471 while the broader SAE basis reaches 0.968)support this scoping, and direct on-OpenVLA feature-setting interven-tions produce specific, task-local behavioral shifts rather than generalclosed-loop repair.

Remote Sensing Image Change Captioning (RSICC) aimsto describe changes between bi-temporal remote sensing images andholds significant research and application value. However, most exist-ing methods rely on conventional deep learning architectures, and thelimited model capacity constrains performance. Although large-modelpost-training techniques have achieved great success in general domains,their direct transfer to RSICC remains challenging due to data scarcityand the need for fine-grained change understanding. To address this, wepropose RSICCLLM, the first post-training framework for large vision-language models in RSICC. Specifically, we design a data genera-tion paradigm, release the instruction dataset RSICI, and establish atask-specific RSICC benchmark. We further introduce Difference-aware Supervised Fine-tuning to explicitly extract change repre-sentations and guide the model in perceiving and understanding tem-poral differences. In addition, we propose Dual-Negative PreferenceOptimization (DNPO), which employs two complementary negative-sample construction strategies to construct the preference dataset RSICPand further refine model performance. Extensive experiments validatethe superior capability of RSICCLLM, which achieves outstanding re-sults with only 7B parameters, surpassing models of substantially largerscales. The code and dataset will be made publicly available at https://github.com/keaill/RSICCLLM.

Deepfake generators evolve rapidly, making exhaustive datacollection and repeated retraining impractical. Unlike generic multi-tasksettings, deepfake specialists share a common binary objective (Real vs.Fake) and mainly differ in generator-specific artifacts. However, naiveparameter arithmetic can induce unintended decision-boundary shifts,causing unstable ranking behavior and degraded AUC under domainshift. We propose R2 M, a training-free merging framework that decom-poses task vectors into a shared component and generator-specific resid-uals, linking parameter-space updates to logit-space behavior. Offlinespectral construction identifies shared and residual subspaces, and on-line routing selects residuals via first-order gradient-residual alignment,predicting per-sample logit updates under linearization. R2 M is bothefficient and interpretable: expensive computations are performed of-fline, while online inference requires only a single forward–backward passwith lightweight inner-product routing. The same formulation enablesdiagnostic analysis through margin and routing statistics. Experimentsdemonstrate consistently strong performance across in-domain, cross-domain, and unseen settings, highlighting R2 M as an interpretable andscalable approach to training-free deepfake model merging.

The proliferation of multimedia content on social platformshas fueled multimodal misinformation, where images are used to rein-force false claims. Consequently, Multimodal Fact-Checking (MFC) hasemerged as an increasingly important research area. However, currentprogress is hindered by a reliance on synthetic training data and cu-rated benchmarks that fail to capture the complexity of in-the-wild data.Furthermore, existing detection models rely on restricted intra-modalityconsistency or unconstrained all-to-all fusion, failing to capture nuancedrelations between posts and external evidence. To address these limi-tations, we introduce X-POSE, a benchmark of real-world, community-annotated multimodal posts from X (formerly Twitter), augmented withfull-length news articles retrieved via VLM-optimized search. Addition-ally, we propose TRENT, a novel MFC model that performs evidence tri-angulation using three parallel cross-attention streams alongside a rela-tional fusion mechanism that explicitly models entailment and contradic-tion. Extensive evaluations demonstrate that TRENT consistently out-performs state-of-the-art specialized models and commercial VLMs. Thecode, prompt templates, and dataset are available at https://github.com/stevejpapad/evidence-triangulation.

The key challenge in articulated 3D object generation from a single image is accurately predicting the underlying kinematic structure. Existing methods either infer kinematic parameters directly from a static image that lacks dynamic part-level kinematic relationships, or estimate parameters from visual dynamics generated from a single image, which is prone to accumulated errors of two steps. Moreover, the limited scale and diversity of existing annotated datasets further hinder generalization to complex, real-world objects. To overcome these limitations, we propose to learn the joint distribution of visual dynamics and kinematic parameters. Recognizing that articulated objects can be formulated as dynamic systems, we propose a unified Part World Model called PWM-ArtGen. To leverage unannotated data, this model couples action diffusion and image diffusion with independent diffusion timesteps, which enables visual branch co-training. We further curate a photorealistic dataset of 19.7k part-level image pairs without kinematic annotations, to support co-training. Experiments demonstrate that PWM-ArtGen substantially outperforms existing baselines in the resting state and exhibits strong zero-shot generalization to out-of-distribution objects. We will release our code at https://github.com/Wentap123/PWM-ArtGen.

Graphic design forms the cornerstone of modern visual com-munication, serving as a vital medium for promoting cultural and com-mercial events. Recent advances have explored automating this processusing Large Multimodal Models (LMMs), yet existing methods often pro-duce geometrically inaccurate layouts and lack the iterative, layer-specificediting required in professional workflows. To address these limitations,we present PosterCopilot, a framework that advances layout reasoningand controllable editing for professional graphic design. Specifically, weintroduce a progressive three-stage training strategy that equips LMMswith geometric understanding and aesthetic reasoning for layout design,consisting of Perturbed Supervised Fine-Tuning, Reinforcement Learn-ing for Visual-Reality Alignment, and Reinforcement Learning from Aes-thetic Feedback. Furthermore, we develop a complete workflow thatcouples the trained LMM-based design model with generative models,enabling layer-controllable, iterative editing for precise element refine-ment while maintaining global visual consistency. Extensive experimentsdemonstrate that PosterCopilot achieves geometrically accurate and aes-thetically superior layouts, offering fine-grained, layer-wise controllabilityfor professional iterative design.

Text-guided video editing with diffusion models is imprac-tically slow, hindered by costly multi-step sampling and inversion. Wepresent OSVE, the first framework to successfully adapt one-step Text-to-Image (T2I) models for high-quality video editing, addressing the corechallenges of inversion, editability, and temporal consistency. To by-pass slow iterative inversion, we train a learnable encoder that pre-dicts the initial noise for each frame in a single forward pass. This en-coder is trained with a novel Structure-Aware Editing (SAE) loss on acurated dataset of structurally-aligned image pairs, teaching it to pre-serve the source video’s geometry during edits. For temporal coherence,we introduce Unified-Frame Editing (UFE), a technique that concate-nates frame latents to facilitate cross-frame attention in a single gen-eration step. Furthermore, for long videos, a sliding-window strategywith an anchor frame maintains global consistency. Our extensive ex-periments demonstrate that OSVE achieves editing quality compara-ble or superior to state-of-the-art multi-step methods, while operatingapproximately 155–171 times faster. This breakthrough paves the wayfor practical, real-time video editing applications. Code is available athttps://github.com/KU-VGI/OSVE.

Recent advances in video generation have made AI-synthesizedcontent increasingly difficult to distinguish from real footage. We pro-pose a physics-based authentication signature that real cameras producenaturally, but that generative models cannot faithfully reproduce. Ourapproach exploits the Moiré effect: the interference fringes formed whena camera views a compact two-layer grating structure. We utilize theMoiré motion invariant as a verification criterion: under real image for-mation, fringe phase and grating image displacement are linearly cou-pled by optical geometry, independent of viewing distance and gratingstructure. A verifier extracts both signals from video and tests theircorrelation. We validate the invariant on both real-captured and AI-generated videos from multiple state-of-the-art generators, and find thatreal and AI-generated videos produce significantly different correlationsignatures, suggesting a robust means of differentiating them. Our workdemonstrates that deterministic optical phenomena can serve as physi-cally grounded, verifiable signatures against AI-generated video.

Offset learning has recently demonstrated remarkable prowessin capturing the structural diversity of objects. Consequently, it ap-pears to be a natural fit for modeling the discriminative features ofvisual targets in tracking. Yet existing leading offline offset-based meth-ods struggle to cope with the ever-changing appearance and motionof the target during tracking. To address this limitation, we presentODONet, a Transformer-based tracker that performs online dynamicoffset network. Its Dynamic Visual–Motion Feature Interaction (DV-MFI) method fuses target appearance with historical box information,yielding appearance features and motion-aware embeddings. These arethen fed to an Online Dynamic Offset Propagation (ODOP) methodthat explicitly captures inter-frame motion trends, augments them withoffline offsets, and steers deformable attention toward an adaptive re-ceptive field. To our knowledge, ODONet is the first tracker to prop-agate dynamic offsets online, with excellent performance on publiclyavailable datasets, demonstrating its robustness to appearance and mo-tion drift. Code, models, and additional information are available athttps://github.com/WhiteButterflies/ODONet.

Generating lifelike facial animation for dyadic conversationsrequires reconciling high-level cognitive intent with precise low-level mo-tor reflexes, yet existing methods fall short in the semantic understandingof dialogue context and in precise dynamic control. In this paper, we pro-pose MindFlow, a dual-pathway generative framework inspired by theVentral-Dorsal pathway model in neuroscience, which decouples genera-tion into two collaborative streams, thereby harmonizing deep semanticreasoning with fine-grained control. In the Ventral module, we trans-form the conventional Sentence-Action approach into a novel Chunk-State approach that models raw acoustic streams as a context-aware,evolving emotional state chain, capturing subtle paralinguistic nuancesand mid-utterance emotional shifts missed by sentence-level modeling.The Dorsal module features a conditional autoregressive flow matchingnetwork for high-fidelity facial motion, driven by high-frequency acousticcues and modulated by emotion states, plus a Selective Acoustic Injectorfor adaptive audio gating to ensure robustness in talking-and-listeningdynamics without interference. Extensive experiments demonstrate thatMindFlow achieves superior semantic appropriateness and motion natu-ralness compared to state-of-the-art baselines.

Deploying video anomaly detection (VAD) in the real worldis often constrained by the scarcity, privacy, and cost of collecting realabnormal footage. We propose PA-VAD, a novel pseudo-only frame-work that trains an anomaly detector without using any real abnormalvideos, by pairing real normal videos with diffusion-synthesized pseudo-abnormal videos generated from a small set of real normal images. Be-yond proposing a generation-driven training pipeline, we make a keyempirical discovery: pseudo anomalies exhibit a characteristic spatiotem-poral magnitude bias in feature space, which can dominate MultipleInstance Learning and degrade generalization if left unaddressed. Tocounter this pseudo-induced bias, we introduce the Domain-Aligned Reg-ularized Module (DARM), which combines domain alignment with usage-aware memory updates to balance prototype coverage and stabilize opti-mization under biased pseudo supervision. Extensive experiments demon-strate that PA-VAD achieves 98.2% AUC on ShanghaiTech, 82.5% onUCF-Crime, and 95.1% on XD-Violence, and further improves general-ization to unseen anomaly classes in open-set evaluations. Notably, PA-VAD surpasses the best real-abnormal WVAD baselines on ShanghaiTechand XD-Violence by +0.6% and +0.9% , respectively, and improves overthe UVAD state of the art on UCF-Crime by +1.9% —showing that high-accuracy VAD is attainable without collecting real abnormal videos.