ECCV 2026 Papers

Video understanding benchmarks have long centered on sin-gle-camera settings, where modern multi-modal language models achievestrong performance across image and video tasks. Yet, the real world runson multi-camera networks: autonomous vehicles, security systems, androbots all gather data across many simultaneous views. We argue thatthis is not simply “more” of the single-camera problem; it is fundamen-tally different. Multi-camera reasoning requires handling context thatscales with the number of views, resolving occlusions visible from only asubset of cameras, judging which views matter, and integrating evidenceacross perspectives that may overlap or diverge. Current models strugglewith exactly these challenges, yet no benchmark systematically targetsthem. We introduce CrossView, a multi-camera video question-answeringbenchmark spanning autonomous driving, security surveillance, egocen-tric/exocentric video, and robotics. Evaluation of proprietary models,such as GPT-5.2, and open-source models, like Qwen3-VL, reveals con-sistently low accuracy, with open-source models trailing by a wide mar-gin. Performance scales strongly with a model’s ability to jointly pro-cess multiple viewpoints, positioning CrossView as a rigorous bench-mark for multi-camera video. We open-source our code and dataset athttps://utaustin-swarmlab.github.io/CrossView.

Modern one-step diffusion models achieve impressive qual-ity through distribution-based timestep distillation. Yet, they rely on acritical assumption: Teacher and Student must inhabit the same latentspace. This Shared-Space constraint prevents knowledge transfer frommodern high-capacity Teachers (e.g., SD 3.5 and Flux) into compact,deployment-friendly Students such as SD 1.5, whose latent resolutionand VAE parameterization differ from the Teacher. We formalize thisoverlooked regime as Cross-Space Distillation, where Teacher andStudent differ in both latent resolution and VAE space. To enable distil-lation under this mismatch, we introduce the Bridge (Bϕ ), a lightweightlatent interface that maps Student latents into the Teacher space with-out modifying the Student backbone. Bridge combines a frozen StudentVAE decoder as a spatial prior with a compact learnable projector, andis trained with latent reconstruction and attention fidelity objectives forstable Teacher-space alignment. Across diverse modern Teachers, Bridgeenables substantial gains for compact one-step Students; for example, itimproves SD 1.5 from 5.4 to 9.4 HPSv3 while preserving one-step in-ference, low latency, and broad ecosystem compatibility. These resultsshow that heterogeneous large Teachers can be distilled into efficient,deployable backbones through a lightweight latent-space interface.

With the recent rapid progress in generative models, im-age editing has made remarkable advances, yet achieving faithful editsthat precisely modify only the target regions while strictly preservingall other regions remains challenging. Since externally provided regionannotations are often difficult to obtain in practice, a growing body ofwork seeks to improve preservation by automatically inferring edit andnon-edit regions, and then enforcing consistency on the latter. However,these approaches still suffer from inaccurate region estimation and heuris-tic correction strategies that distort the native inference process, makingmethods designed for fidelity themselves a new source of artifacts. Wepropose SR-Edit, an image editing framework that overcomes these is-sues via iterative self-refinement. Specifically, at each iteration, SR-Editfirst (i) extracts progressively precise and self-consistent region separa-tion from the model’s own predictions by lightweight post-processing,and then (ii) enforces preservation in non-edit areas through correctionupdates that remain aligned with the original sampling dynamics. Ex-tensive experiments demonstrate that SR-Edit achieves superior preser-vation and overall image quality compared to existing editing techniques.

3D Gaussian Splatting (3DGS) enables photorealistic real-time novel view synthesis, yet placing a virtual camera to capture a de-sired frame remains largely manual. Existing language-guided approachesin 3D scenes mainly focus on object-centric grounding, determining whatto observe but rarely controlling how it should appear in a single frame,such as subject orientation or frame layout. To address this limitation,we introduce a new task, Text-Instructed Viewpoint Grounding (TIVG),which aims to identify a 6-DoF camera pose in a 3D Gaussian scenewhose rendered frame aligns with a text instruction. To solve this task,we propose CapFrame, a partially differentiable framework that con-verts language into geometric pseudo labels for camera pose optimization.CapFrame follows a Retrieve–Translate–Refine pipeline: it retrieves rele-vant views and ranks them through a Question-Evaluation process withMLLMs, translates the instruction into orientation and layout pseudo la-bels, and refines the camera pose via differentiable optimization with lay-out and orientation losses in 3DGS. Experiments on 38 real-world sceneswith 135 instructions indicate that CapFrame produces viewpoints betteraligned with texts than heuristic viewpoint search and adapted trajectorygeneration baselines, validated by VLM metrics, MLLM judges, and userstudies. Code is available at: https://github.com/jirongli/CapFrame

Autonomous driving operates in partially observable envi-ronments where actors may become fully occluded by other vehicles orinfrastructure. Most end-to-end driving systems implicitly couple actorexistence to instantaneous observations, causing actor hypotheses to de-grade or disappear during prolonged occlusion and removing potentiallycritical agents from downstream prediction and planning. We introduceBeyondSight, a permanence-aware end-to-end driving framework thatdecouples actor existence from observability by maintaining persistentactor hypotheses over time. BeyondSight propagates actor queries tem-porally and updates them with observation-conditioned evidence, en-abling joint perception, prediction, and planning to reason about ac-tors even when they are temporarily unobservable. To enable principledtraining and evaluation of persistence-aware models, we further introducenuScenes-Permanence, an extension of nuScenes that provides super-vision and observability-conditioned evaluation for unobservable actors.Experiments show that BeyondSight substantially improves reasoningunder occlusion, increasing detection performance for unobservable ac-tors from 0 to 0.249 mAP while reducing planning error from 0.61 to0.54 L2avg . These results highlight object permanence as an importantmodeling principle for robust end-to-end autonomous driving.

Video emotion analysis is typically framed as a static clas-sification problem, treating each clip as an independent labeled unit.However, such a formulation overlooks a key psychological fact: emotionschange as a result of cumulative reactions to consecutive causal events. Tobridge this gap, we introduce DAR (Dynamic Affective Reasoning), thefirst large-scale benchmark for viewer-centric affect transitions and causalreasoning over consecutive video events. DAR contains 15,087 videos and36,908 event-aligned affective segments annotated with 27 emotion cat-egories. Unlike existing video-based emotion datasets, DAR presents anew viewer-centric perspective on fine-grained emotional expressions andtransitions, and provides dense, temporally grounded, and causally ex-plicit reasoning chains. Based on DAR, we formally define three challeng-ing tasks: affective segmentation, fine-grained emotion classification, andaffective reasoning. Complementing this benchmark, we propose DAR-R1, a two-stage framework that combines supervised fine-tuning withGroup Relative Policy Optimization. Experiments across 10+ MLLMsshow that DAR-R1 sets a new state-of-the-art for dynamic affective rea-soning, in terms of both emotional localization and affective reasoning.Project page: https://github.com/Zhang-Zhiyan/DAR.

Understanding a painting is never a single act. Art histori-ans may analyze the same work through concepts of style, iconography,or historical context, dimensions that are not interchangeable, and eachcarries distinct semantic relationships between the visual and the tex-tual. Vision-Language Models (VLMs) like CLIP, which learn a singleshared embedding space, collapse this richness into a single homogeneousalignment, thereby losing the multi-relational structure that defines art-historical reasoning. We introduce CANVAS (Contrastive Art-awareNetwork for Vision-Language Alignment with Sheaves), a frameworkfor learning relation-aware multimodal representations inspired by sheaftheory. Each artwork is projected into multiple embeddings conditionedon the type of relation (i.e., the context), and a novel contrastive lossencodes contextual information during training, with no dependency onexternal data at inference. We evaluate on three newly introduced bench-marks of artworks for multi-relational art understanding: WikiArt+,derived from WikiArt and Wikipedia, HertzianaDP, from the Biblio-theca Hertziana collection, and SemArt+, refined from the SemArtdataset. In multimodal retrieval and art understanding, CANVAS out-performs the baselines, supporting the view that multi-relational align-ment is not just theoretically motivated but also practically essential.

Zero-shot anomaly detection aims to identify defects in ar-bitrary novel domains; however, existing models assume that the auxil-iary data contains a rich diversity of anomalies, neglecting the far morecomplex and unpredictable variations in real-world target domains. Thisstudy introduces DIVE, the first approach to investigate the scenario oflimited auxiliary anomaly priors and resolve the resulting substantial per-formance degradation. Through a shallow-and-deep text embedding in-jection strategy during visual encoding, DIVE learns to abstract genericanomaly concepts shared across the auxiliary training domain and di-verse target domains. Moreover, we propose a disentanglement mech-anism to tackle the suboptimal alignment between visual embeddingsentangled with object semantics and object-agnostic textual prompts.Experiments demonstrate that, under the setting of limited anomalypatterns in auxiliary data, DIVE outperforms SOTA baselines by up to16.2% and 28.5% on two classification metrics, and 23.4%, 24.1%, and47.0% on three segmentation metrics, in terms of average performanceacross twelve datasets. Furthermore, it maintains highly competitive per-formance when auxiliary data exhibits sufficient anomaly diversity.

Existing video object removal methods excel at inpaintingcontent “behind” the object and correcting appearance-level artifactssuch as shadows and reflections. However, when the removed object hasmore significant interactions, such as collisions with other objects, cur-rent models fail to correct them and produce implausible results. Wepresent VOID, a video object removal framework designed to performphysically-plausible inpainting in these complex scenarios. To train themodel, we generate a new paired dataset of counterfactual object re-movals using Kubric and HUMOTO, where removing an object requiresaltering downstream physical interactions. During inference, a vision-language model identifies regions of the scene affected by the removedobject. These regions are then used to guide a video diffusion model thatgenerates physically consistent counterfactual outcomes. Experiments onboth synthetic and real data show that our approach better preservesconsistent scene dynamics after object removal compared to prior videoobject removal methods. We hope this framework sheds light on howto make video editing models better simulators of the world throughhigh-level causal reasoning.

Video diffusion models lack explicit geometric supervisionduring training, leading to inconsistency artifacts such as object defor-mation, spatial drift, and depth violations in generated videos. To addressthis limitation, we propose a geometry-based reward model that leveragespretrained geometric foundation models to evaluate multi-view consis-tency through cross-frame reprojection error. Unlike previous geometricmetrics that measure inconsistency in pixel space, where pixel intensitymay introduce additional noise, our approach conducts error computa-tion in a pointwise fashion, yielding a more physically grounded and ro-bust error metric. Furthermore, we introduce a geometry-aware samplingstrategy that filters out low-texture and non-semantic regions, focusingevaluation on geometrically meaningful areas with reliable correspon-dences to improve robustness. We apply this reward model to align videodiffusion models through two complementary pathways: post-training ofa bidirectional model via SFT or Reinforcement Learning and inference-time optimization of a Causal Video Model (e.g., Streaming video genera-tor) via test-time scaling with our reward as a path verifier. Experimentalresults validate the effectiveness of our design, demonstrating that ourgeometry-based reward provides superior robustness compared to othervariants. By enabling efficient inference-time scaling, our method offersa practical solution for enhancing open-source video models without re-quiring extensive computational resources for retraining.

Existing Multimodal Large Language Models (MLLMs) of-ten suffer from hallucinations in long video understanding (LVU), pri-marily due to the imbalance between textual and visual tokens. Ob-serving that MLLMs handle shorter but more accurate visual inputswell, recent LVU works alleviate hallucinations by automatically parsingthe vast visual data into manageable segments that can be effectivelyprocessed by MLLMs. SFT-based tool-calling methods can serve thispurpose, but they typically require vast amounts of fine-grained, high-quality data and suffer from constrained tool-calling trajectories. Wepropose a novel VideoTIR that leverages Reinforcement Learning (RL)to encourage proper usage of comprehensive multi-level toolkits for ef-ficient long video understanding. VideoTIR explores both Zero-RL andSFT cold-starting to enable MLLMs to retrieve and focus on meaning-ful video segments/images/regions, enhancing long video understandingboth accurately and efficiently. To reduce redundant tool-calling in theearly RL-stage and accelerate convergence, we propose Toolkit ActionGrouped Policy Optimization (TAGPO), which enhances the efficiencyof the calling process through the finer stepwise reward assignment. Ad-ditionally, we develop a sandbox-based trajectory synthesis frameworkto generate high-quality trajectory data. Extensive experiments on threelong-video QA benchmarks demonstrate the effectiveness and efficiencyof our method.

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.