ECCV 2026 论文列表

The nonparametric identification of causal mediation effectsfrom heterogeneous multimodal observational data remains an opentheoretical problem. Existing single-modal methods cannot handlecross-modal causal relationships, while multimodal learning methodslack causal inference capabilities, preventing reliable mediation pathwayidentification and effect quantification under structural uncertainty.We propose the Multimodal Structure-Informed Guided MediationAnalysis (MM-SIGMA) framework, which achieves automated cross-modal mediation pathway identification and end-to-end uncertaintypropagation through probabilistic causal structure discovery. Weestablish nonparametric identification conditions for the Cross-ModalNatural Direct Effect and Cross-Modal Natural Indirect Effect fromheterogeneous multimodal observational data, and build MM-SIGMAupon this theoretical foundation. Specifically, it employs multimodalvariational autoencoders for causality-preserving latent representation,differentiable Flow-Structural Equation Models for asymptoticallyconsistent latent structure learning, Cross-Modal Path Stability Scoringfor high-confidence pathway identification, and Efficient InfluenceFunctions with Bayesian Model Averaging for end-to-end uncertaintypropagation. Experiments on synthetic data demonstrate state-of-the-art performance under structural uncertainty, nonlinearity, andcross-modal heterogeneity. On the HPP dataset, MM-SIGMA identifiescross-modal mediation pathways connecting sleep, fundus imaging, andcardiovascular health, revealing undiscovered cross-system mechanisms.

Reconstructing 3D Human-Object Interaction (HOI) from asingle image is challenging due to the diversity of the reconstructed tar-gets. Ordinarily, the human body follows a fixed parametric template,whereas objects exhibit diverse and irregular shapes. Such differenceslead to inconsistent geometric priors and scale variations, making jointmodeling within a unified framework highly difficult. Existing vertex-level methods struggle to effectively handle this structural discrepancy.We propose to bridge this gap by introducing a unified triplane featuremap that models both human and object meshes within a shared 3Dlatent space. Our framework comprises three components: (1) a HOI Tri-plane Variational AutoEncoder (HOIT-VAE) that learns a compact andstable triplane latent space from large-scale human-object mesh data; (2)a HOI Triplane Transformer (HOIT-T) that predicts triplane featuresfrom a single image, regularized by the learned priors of HOIT-VAE;(3) a pose-fitting method for estimating the human pose from the re-constructed human mesh. Experiments on the BEHAVE and InterCapdatasets show that our approach achieves competitive reconstruction per-formance, validating the effectiveness of the unified triplane representa-tion in capturing the mesh structures and interactions between humansand objects.

Automatically detecting abnormal events in videos is cru-cial for modern autonomous systems, yet existing Video Anomaly De-tection (VAD) benchmarks lack the scene diversity, balanced anomalycoverage, and temporal complexity needed to reliably assess real-worldperformance. Meanwhile, the community is increasingly moving towardVideo Anomaly Understanding (VAU), which requires deeper seman-tic and causal reasoning but remains difficult to benchmark due to theheavy manual annotation effort it demands. In this paper, we introducePistachio, a new VAD/VAU benchmark constructed entirely through acontrolled, generation-based pipeline. By leveraging recent advances invideo generation models, Pistachio provides precise control over scenes,anomaly types, and temporal narratives, effectively eliminating the bi-ases and limitations of Internet-collected datasets. Our pipeline inte-grates scene-conditioned anomaly assignment, multi-step storyline gen-eration, and a temporally consistent long-form synthesis strategy thatproduces coherent 41-second videos with minimal human intervention.Extensive experiments demonstrate the scale, diversity, and complexityof Pistachio, revealing new challenges for existing methods and motivat-ing future research on dynamic and multi-event anomaly understanding.

Inferring physical properties from a single image is funda-mentally under-constrained. Attributes such as density, elasticity, andyield stress govern how objects move, yet they are largely invisible ina static frame. Existing physics-aware methods attempt to resolve thisambiguity through task-specific fine-tuning or supervised property es-timation, but both strategies struggle to generalize across diverse ma-terials and scenes. We observe that different motions expose comple-mentary physical cues. Building on this observation, we propose Phy-MAGIC, a training-free framework that actively probes physical prop-erties by synthesizing targeted motions from a single image. Specifically,PhyMAGIC uses a pretrained image-to-video model to construct mo-tion probes that generate diverse dynamic sequences from the inputimage. A vision-language model then analyzes these sequences to esti-mate physical parameters, each accompanied by a confidence score. Pa-rameters with low confidence trigger targeted prompt refinement, whichgenerates additional probe motions to gather complementary evidence.Once all parameters reach sufficient confidence, PhyMAGIC compilesthem into a complete physical specification and executes it in a differ-entiable Material Point Method simulator initialized from 3D Gaussianreconstructions. Experiments on diverse real-world scenes demonstratethat PhyMAGIC achieves stronger text-motion alignment and higherhuman-rated physical plausibility than state-of-the-art open-source videogenerators and physics-aware baselines. Code is available at: https://mengsiwei.github.io/MAGIC/.

Continuous 6-DoF pose estimation is essential for au-tonomous UAV operations. Yet, existing visual odometry and SLAMmethods accumulate drift and yield only relative, up-to-scale trajecto-ries. Single-frame geo-localization, in turn, discards temporal continu-ity and remains too slow for real-time use. We present OrthoTrack, atraining-free system that estimates continuous 6-DoF UAV trajectoriesusing only publicly available orthophotos and surface models as a mapprior. OrthoTrack matches keyframes against the orthophoto and liftscorrespondences to metric 3D via the surface model. It then propagatesthese map-anchored correspondences to intermediate frames with opticalflow, producing absolute, metrically scaled poses at every frame with-out GPS or post-hoc alignment. We also introduce the MovingDroneDataset, a large-scale benchmark pairing photorealistic UAV sequenceswith dense 6-DoF ground truth and co-registered multi-modal geodataincluding multi-temporal orthophotos. On MovingDrone and real-worldbenchmarks, OrthoTrack runs in real time on a single GPU. It outper-forms all baselines by a large margin, even those receiving oracle scaleand alignment. By relying on publicly available geodata, OrthoTrackenables deployment to new regions without site-specific adaptation.

Fine-grained Cross-View Localization (CVL) estimates theprecise position and orientation of a ground-level image by aligningit with geo-referenced aerial imagery, offering a scalable alternative toGlobal Navigation Satellite Systems (GNSS) in challenging urban envi-ronments. Existing datasets rely on data collected with high-end sensorsuites, which inherently limit image diversity and scalability. While in-the-wild images are abundant, their noisy geo-tags make them unsuit-able for reliable evaluation. To bridge this gap, we introduce OpenCVL,a large-scale, diverse, and open dataset containing 617,388 ground-aerialimage pairs spanning 41 cities across four European countries. All imagesare sourced from permissive platforms, ensuring long-term accessibilityand supporting open and reproducible research. The training set com-bines images captured with high-end sensors with diverse in-the-wild im-agery. We further develop a data curation framework that filters and cor-rects pose annotations to construct reliable in-the-wild evaluation data.In addition, OpenCVL includes dedicated cross-area and snowy test setsto assess generalization and robustness. Experiments with a state-of-the-art CVL model on OpenCVL show that incorporating noisy in-the-wilddata consistently improves performance on clean test sets, suggesting apromising direction for scaling CVL with diverse real-world imagery.

We present a deployed system for on-orbit wildfire detec-tion aboard OroraTech’s thermal infrared Constellation Plane 1 (OTC-P1), operating under a demanding set of joint constraints: sub-megabytemodel footprint, sub-150 ms per-batch TensorRT FP16 inference on anNVIDIA Jetson Xavier NX, and an end-to-end alert pipeline targetingunder 10 minutes from satellite overpass to fire event communication.The system operates on uncalibrated mid-wave infrared (MWIR) single-band imagery at 200 m ground sampling distance, where fires frequentlyappear as sub-pixel or single-pixel thermal anomalies under extremeclass imbalance—challenges not addressed by the contextual thermal-thresholding pipelines (MODIS, VIIRS) that currently dominate opera-tional fire monitoring.We present an empirical study of lightweight dense representation learn-ing for this regime using a proprietary nine-satellite MWIR dataset. Wecompare dense masked autoencoding (DenseMAE) and a hybrid Dense-MAE+EMA (exponential moving average) distillation variant, and eval-uate representations via linear probing and full-distribution pixel-levelaverage precision (AP) under extreme class imbalance. DenseMAE pre-training enables compact downstream models that sit on the latency–accuracy Pareto frontier: our fastest SSL-pretrained model achieves 0.640test AP and 0.69 event-level Fire-F1 with 65.34 ms TensorRT FP16 la-tency per batch (224 × 224, B=8) and a 0.52 MB engine, achieved with-out pruning or compression. The best-performing configuration reaches0.699 test AP and 0.744 Fire-F1 while remaining below 1 MB. Bothoutperform a supervised baseline (up to 0.650 AP) under comparableconstraints.

Online reinforcement learning (RL) has become an effectivetool for aligning text-to-image models with human preferences, but itsefficiency is often limited by group-relative variance reduction. Exist-ing methods generate multiple trajectories per prompt to estimate localbaselines, which introduces synchronization overhead and can weaken thelearning signal when group rewards become homogeneous. We proposeMonolithic Policy Optimization (MPO), a group-free online RL frame-work based on a single-stream training principle: one prompt, one trajec-tory, and one policy update. MPO retains stochastic differential equation(SDE) exploration within each trajectory, while replacing group-wisebaselines with a persistent Bayesian value tracker. The tracker main-tains a history-aware reward estimate for each prompt and adapts itsuncertainty using a Girsanov-inspired policy-drift proxy, providing stableglobal advantage estimation under non-stationary policy updates. Acrosscompositional generation, visual text rendering, and human-preferencealignment benchmarks, MPO consistently improves alignment qualityover group-based baselines. It also delivers a 26× wall-clock speedupand a 5× sample-efficiency gain under matched training settings.

While most image deblurring techniques directly restore thespatial image variable, we propose an amplitude and phase decomposi-tion recognizing the importance of accurate phase estimation in recover-ing sharp image details. To that end, we first develop novel linear min-imum mean squared (LMMSE) estimators of the amplitude and phaseof the blurred, noisy image observation. An iterative optimization algo-rithm follows that recovers the sharp image using the aforementionedLMMSE estimators. Finally, matrix parameters that are statistically de-termined and fixed in the iterative algorithm are now learned using atraining dataset of clean and degraded observations. Our deblurring en-gine is dubbed UPADNet – Unrolled Phase and Amplitude Decompo-sition Network, such that each iteration of the underlying phase andamplitude recovery algorithm is parameterized and trained end-to-end.Experiments over benchmark evaluation datasets such as GoPro, Real-Blur and COCO datasets confirm that UPADNet outperforms state ofthe art deep networks including those based on algorithm unrolling inthe image domain. The benefits of UPADNet are even more pronouncedin high noise and limited training data regimes.

Out-of-distribution (OOD) detection is crucial for the safedeployment of deep neural models in applications such as autonomousdriving. With the emerging multimodal nature of modern applications,recent attention has shifted toward OOD detection in multimodal set-tings. However, current multimodal OOD detection methods fail to fullyexploit the synergy among modalities: they treat all modalities equally,disregarding their varying detection performance, and they are unableto capture the diverse uncertainty information encoded at the logit level.In this paper, we propose to exploit the dark knowledge within unimodalexperts as the key to revealing their synergy. To this end, we introducea self multimodal OOD distillation framework, which leverages logits asuncertainty-aware soft targets to train a holistic model that operates inthe joint embedding space of all modalities. Specifically, the proposedframework accounts for the negative effects of underperforming modal-ities and effectively fuses both the rich feature-level knowledge and thelogit-level knowledge of modalities. As a result, our method improvesthe performance of current state-of-the-art multimodal OOD detectionmethods, achieving gains of up to 30% across diverse OOD detectionbenchmarks, spanning two tasks and five multimodal OOD datasets.

Recent advances in diffusion models have enabled impressivevideo editing capabilities, yet production-grade Video Object Insertion(VOI) remains challenging due to inadequate 4D scene understandingand a lack of proper optical interactions, such as shadows and reflec-tions. To address these limitations, we present InsertAnywhere, a com-prehensive VOI framework that achieves geometrically grounded objectplacement and optics-aware video synthesis. Our approach first lever-ages a 4D-aware mask generation module that allows users to anchor anobject’s 3D pose in a single frame. The framework automatically prop-agates this placement across the video, accurately handling local scenedynamics and occlusions. To synthesize realistic physical lighting inter-actions, we introduce Optics-Aware Representation Alignment, a novelstrategy that utilizes an extended mask to guide feature extraction, en-abling optical effects to seamlessly extend beyond the inserted object’sboundary. Finally, to overcome the lack of training data for such phenom-ena, we construct and open-source ROSE++, a specialized quadrupletdataset tailored for the supervised learning of optical effects. Extensiveexperiments demonstrate that InsertAnywhere produces geometricallyplausible and photometrically realistic insertions in complex real-worldscenarios, significantly outperforming existing research and commercialgenerative tools.

While multimodal Large Language Models (MLLMs) excel at offline video understanding, an interesting question of how far they are from serving as a real-time procedural coach remains unknown. Such a role typically requires an MLLM to continuously monitor the execution, detect mistakes, and provide corrective guidance in a closed-loop interaction. In this paper, we construct GuideMe, the first multi-domain benchmark for streaming video that supports training and evaluation of MLLMs for closed-loop interactive task guidance. It comprises 2,458 videos spanning 223.7 hours across diverse domains (e.g., cooking, object manipulation, daily-life guidance, and fitness), with 47,775 interaction samples covering next-step instructions, completion feedback, error detection, and corrective guidance. To evaluate existing models on GuideMe, we design a three-component assessment framework to measure the capabilities of representative MLLMs, which consists of temporal-semantic bipartite matching for sequence-level alignment, behavioral classification for intervention timing, and LLM-as-a-Judge for content quality. Extensive experiments highlight a critical performance asymmetry: despite excelling at providing instructions, existing MLLMs consistently fail to identify execution errors and respond with corrective feedback. Code and data are released at (cid:140)homepage.

Music-to-dance generation aims to translate auditory sig-nals into expressive human motion. Despite promising progress, existingmethods remain underexplored in achieving high-quality generation un-der a strict efficiency budget, due to (i) generative strategies that sup-port high-fidelity few-step sampling, and (ii) per-step model architecturesthat are optimized for efficient yet refined long-horizon motion synthesis,thereby degrading downstream user experience in terms of responsive-ness and visual fidelity. Thus, we propose FlowerDance, which not onlygenerates refined motion with physical plausibility and artistic expres-siveness, but also achieves significant generation efficiency on inferencespeed and memory utilization. Specifically, FlowerDance combines Mean-Flow with Physical Consistency Constraints, which enables high-qualitymotion generation with only a few sampling steps. Moreover, Flower-Dance leverages a simple but efficient model architecture with BiMamba-based backbone and Channel-Level Cross-Modal Fusion, which generateslong-horizon dance with efficient non-autoregressive manner. Meanwhile,FlowerDance supports motion editing, enabling users to interactively re-fine dance sequences. Extensive experiments on AIST++ and FineDanceshow that FlowerDance achieves state-of-the-art results in both motionquality and generation efficiency. Code is available at https://sun-happy-ykx.github.io/FlowerDance/.

Rendering views using 3D scene representations such as Gaus-sian Splatting (3DGS), Neural Radiance Fields (NeRF), meshes, or evenpoint clouds produces artifacts when input views are sparse or targetviews lie far from the input. Recent work mitigates these artifacts usingdiffusion-based generative priors, but is specialized to individual repre-sentations and require custom architectures or extensive retraining. Wepresent FixAnything, a single model for fixing a wide range of render-ing artifacts. It does so by repurposing a pretrained video generativemodel, leveraging its implicit multi-view priors with only minimal mod-ification and lightweight finetuning. Our key insight is that even noisily-rendered sequences preserve camera motion and coarse scene structure,allowing cleanup to be formulated as video-to-video translation. To con-trol what scene structure should be preserved, we introduce a binarymask denoting the clean pixels, enabling the model to anchor its outputto high-quality inputs (e.g. training views) while refining the rest. To en-courage FixAnything to produce 3D-consistent renderings that supportdownstream reconstruction, we use camera pose accuracy (recovered viastructure-from-motion) as a reward signal for direct preference optimiza-tion (DPO). Across four distinct 3D representations, FixAnything con-sistently improves rendering quality with lightweight finetuning, demon-strating that a single generalist video prior can replace multiple specialistrefinement pipelines. The simplicity of the framework enables immediateadoption of stronger future video models without architectural redesign.

Accurate estimation of forest biomass, a major carbon sink, relies heavily on tree-level traits such as height and species. Unoccupied Aerial Vehicles (UAVs) capturing high-resolution imagery from a single RGB camera offer a cost-effective and scalable approach for mapping and measuring individual trees. We introduce BIRCH-Trees, the first benchmark for individual tree height and species estimation from tree-centered UAV images, spanning three datasets: temperate forests, tropical forests, and boreal plantations. We also present DINOvTree, a unified approach using a Vision Foundation Model (VFM) backbone with task-specific heads for simultaneous height and species prediction. Through extensive evaluations on BIRCH-Trees, we compare DINOvTree against commonly used vision methods, including VFMs, as well as biological allometric equations. We find that DINOvTree achieves top overall results with accurate height predictions and competitive classification accuracy while using only 54% to 58% of the parameters of the second-best approach.

Visual search in 3D environments requires embodied agentsto actively explore their surroundings and acquire task-relevant evidence.However, existing visual search and embodied AI benchmarks, includ-ing EQA, typically rely on static observations or constrained egocen-tric motion, and thus do not explicitly evaluate fine-grained viewpoint-dependent phenomena that arise under unrestricted 5-DoF viewpointcontrol, such as disambiguating object attributes observable only fromspecific angles. To address this limitation, we introduce E3VS-Bench,a benchmark for embodied 3D visual search where agents must controltheir viewpoints in 5-DoF to gather viewpoint-dependent evidence forquestion answering. E3VS-Bench consists of 99 high-fidelity 3D scenesreconstructed using 3D Gaussian Splatting and 2,014 question-drivenepisodes. 3D Gaussian Splatting enables photorealistic free-viewpointrendering that preserves fine-grained visual details (e.g., small text andsubtle attributes) often degraded in mesh-based simulators, thereby al-lowing the construction of questions that cannot be answered from asingle view and instead require active inspection across viewpoints in5-DoF. We evaluate multiple state-of-the-art VLMs and compare theirperformance with humans. Despite strong 2D reasoning ability, all mod-els exhibit a substantial gap from humans, highlighting limitations inactive perception and coherent viewpoint planning specifically under full5-DoF viewpoint changes. The benchmark, code, and dataset are publiclyavailable at https://k0uya.github.io/e3vs-proj/.

Adapting pre-trained text-to-image diffusion models, whether to learn new visual concepts or erase unwanted ones, is routinely evaluated on its intended effects alone. We argue this framing is incomplete. Through sparse autoencoder analysis and zero-shot classification, we demonstrate that adaptation systematically damages semantically unrelated concepts in ways that aggregate metrics structurally cannot surface: when damage is severe enough for FID and KID to respond, the model is already nearly unusable; when the model remains functional, FID and KID stay flat while specific classes silently suffer worst-case zero-shot accuracy drops of up to 18.9 points and concept-level distributions shift dramatically. This pattern appears at both ends of the adaptation spectrum (concept customization and concept unlearning), suggesting it is a systematic consequence of weight-level modification rather than an artifact of any particular method. To surface this hidden drift before deployment, we introduce DriftScope, a prompt-level diagnostic tool that takes any two model checkpoints and returns a ranked list of tokens whose visual concepts have shifted most between them. DriftScope optimizes a soft prompt to attribute drift at the token level without requiring access to real data or model internals. The result is an interpretable, concept-level audit that aggregate evaluation cannot provide.

The long-standing vision of general-purpose robots hinges on theirability to understand and act upon natural language instructions. Vision-Language-Action (VLA) models have made remarkable progress toward thisgoal, yet their generated actions can still misalign with the given instructions.In this paper, we investigate test-time verification as a means to shrinkthe “intention-action gap.” We first characterize the test-time scaling lawsfor embodied instruction following and demonstrate that jointly scaling thenumber of rephrased instructions and generated actions greatly increasestest-time sample diversity, often recovering correct actions more efficientlythan scaling each dimension independently. To capitalize on these scalinglaws, we present CoVer, a contrastive verifier for vision–language–actionalignment, and show that our architecture scales gracefully with additionalcomputational resources and data. We then introduce CoVer-VLA, a hierar-chical test-time verification pipeline using the trained verifier. At deployment,our framework precomputes a diverse set of rephrased instructions from aVision-Language-Model (VLM), repeatedly generates action candidates foreach instruction, and then uses the verifier to select the optimal high-levelprompt and low-level action chunks. Compared to scaling policy pre-training⋆ †Equal contribution. § Equal advising.on the same data, our verification approach yields 22% gains in-distributionand 13% out-of-distribution on the SIMPLER benchmark, with a further45% improvement in real-world experiments. On the PolaRiS benchmark,CoVer-VLA achieves 14% gains in task progress and 9% in success rate.

While Audio-Visual Language Models (AVLMs) have achievedremarkable progress over recent years, their reliability is bottleneckedby cross-modal hallucination. A particularly pervasive manifestation isvideo-driven audio hallucination: models routinely exploit visual short-cuts to hallucinate expected sounds, discarding true auditory evidence.To counteract this deeply ingrained visual dominance, we propose Audio-Contrastive Preference Optimization (ACPO). This dual-axis preferencelearning framework introduces an output-contrastive objective to penal-ize visual descriptions masquerading as audio facts, alongside an input-contrastive objective that swaps audio tracks to explicitly penalize gener-ation invariant to the true auditory signal. Extensive experiments demon-strate that ACPO establishes highly faithful audio grounding and mit-igates audio hallucination. Project page: https://vision.cs.utexas.edu/projects/acpo/