ECCV 2026 论文列表

Cross-view object geo-localization (CVOGL) aims to locatea target object from a query view (e.g., ground or drone) within ageo-tagged reference image (e.g., satellite). Existing approaches heav-ily rely on 2D appearance matching and are constrained by limiteddatasets lacking geometric metadata, diverse prompts, and standardfield-of-view imagery. To address these intertwined challenges, we firstintroduce CMA-Loc, a large-scale, high-fidelity building dataset compris-ing over 220,000 ground-satellite and drone-satellite pairs. It providesmulti-modal prompts (points, boxes, masks) and camera poses to en-able flexible target referring and explicit spatial modeling. Furthermore,we propose a novel single-stage Geometry-Aware Geo-localization frame-work (GAGeo), built upon the permutation-equivariant 3D foundationmodel π 3 . By seamlessly integrating visual features, referring prompts,and learnable task tokens, our model adapts the inherited 3D prior tojointly predict bounding boxes, segmentation masks, and camera posesin a single forward pass. Additionally, we introduce a contrastive lossthat utilizes the satellite view as a universal anchor, implicitly aligningground and drone representations to enable zero-shot ground-to-drone lo-calization without requiring triplet training data. Extensive experimentsdemonstrate that our approach significantly outperforms state-of-the-artmethods, exhibiting exceptional generalization ability in unseen scenesand novel cross-view setups.

Audio-visual continual test-time adaptation involves contin-ually adapting a source audio-visual model at test-time, to unlabelednon-stationary domains, where either or both modalities can be distribu-tionally shifted, which hampers online cross-modal learning and eventuallyleads to poor accuracy. While previous works have tackled this problem,we find that SOTA methods suffer from catastrophic forgetting wherethe model’s performance drops well below even the source model dueto continual parameter updates at test-time. In this work, we first showthat adapting only the modality fusion layer to a target domain not onlyimproves performance on that domain but can also enhance performanceon subsequent domains. Based on this strong cross-task transferability ofthe fusion layer’s parameters, we propose a method, AVReCAP, that im-proves test-time performance of the models without access to any sourcedata. Our approach works by using a selective parameter retrieval mech-anism that dynamically retrieves the best fusion layer parameters from abuffer using only a small batch of test data. These parameters are thenintegrated into the model, adapted to the current test distribution, andsaved back for future use. Extensive experiments on benchmark datasetsinvolving unimodal and bimodal corruptions show our proposed AVReCAPsignificantly outperforms existing methods while minimizing catastrophicforgetting.

Affordance grounding aims to localize where to interact withan object, a fundamental capability for embodied agents. Yet progressis bottlenecked by data: manual annotation is prohibitively expensiveand confines existing datasets to a narrow set of predefined objectand affordance categories. We introduce Affogato, a framework foropen-vocabulary affordance grounding centered on Affogato-750K, alarge-scale dataset of 750K 3D affordance heatmaps paired with naturallanguage queries. We build it with a fully automated pipeline that or-chestrates foundation models to generate them at scale without humanlabeling. It covers significantly more diverse categories than any existingdataset. For reliable evaluation, we further provide 5K human-verified testpairs. We also present Espresso-3D and Espresso-2D, simple yet effec-tive models with a unified architecture across both modalities. Pretrainingon Affogato-750K improves both Espresso and prior methods and yieldsthe largest gains on unseen object and affordance categories, showingthat it provides broadly transferable supervision across architectures.

Multimodal large language models have demonstrated strongdocument reasoning capabilities by incorporating explicit thinking pro-cesses. While this capability significantly improves performance on chal-lenging tasks, current models apply such deep reasoning uniformly to allquestions, resulting in unnecessary computational overhead for simpletask. This not only degrades user experience but also negatively impactaccuracy on benchmark datasets. We identify the critical need for adap-tive thinking mechanisms that can intelligently determine when to en-gage reasoning based on question complexity. To address this, we proposeAdaThinking-E, a novel reinforcement learning framework that learnsadaptive thinking through one-token entropy regulation. Our key insightis that model confidence in the decision to engage thinking (or not)can be quantified through entropy analysis of the predicted probabilitydistribution at critical decision tokens. This observation motivates ourentropy-governed reward mechanism: the training process naturally tran-sitions from high-entropy exploration, where the model experiments withdifferent thinking strategies, to low-entropy convergence with confident,generalizable decision-making policies. Crucially, this approach enablesmodels to intrinsically discover when to think without requiring manualintervention or external difficulty labels. Extensive experiments demon-strate that our approach enables models to be both accurate on complexproblems and efficient on simple ones across diverse document tasks.Code is available at https://github.com/PriNing/AdaThinking-E.

3D semantic occupancy prediction is crucial for fine-grainedscene understanding, yet its advancement in privacy-sensitive indoor en-vironments is fundamentally hindered by the scarcity of large-scale an-notated 3D data. To overcome this limitation, we explore learning indoor3D semantic occupancy prediction from abundant, uncalibrated in-the-wild internet videos while simultaneously bypassing the extensive manualannotation. Specifically, we introduce YouTube-Occ, including an auto-mated data pipeline that leverages 2D and 3D foundation models to pro-cess raw web videos, estimating camera geometry, reconstructing scenepoint clouds, and enriching them with dense semantic pseudo-labels.However, these plausible pseudo-labels fail to yield performance gains un-der naive supervision. To address this impasse, we further propose a pre-training framework driven by feature distillation with a dual-alignmentstrategy. Within it, an intra-frame alignment utilizes a voxel-anchoredGaussianization module to align 3D features with corresponding 2D pri-ors, whereas a cross-scene alignment achieves global semantic consistencyvia class-prototype distillation. Empirically, YouTube-Occ delivers con-sistent gains across three mainstream architectures on the NYUv2 andOcc-ScanNet benchmarks, especially under limited-data conditions. Wewill publicly release our code and data, hoping to inspire future research.

Intelligent cabin is gradually evolving from passive voicebased interfaces into multimodal collaborative systems supported by multi-camera and multi-sensor setups, where spatial reasoning becomes a core capability for enabling proactive perceptual interaction. Unlike traditional single-scene spatial tasks, cabin environments involve both interior and exterior vehicle regions, multiple occupants and traffic participants, and inherently require cross-view observations. To fill this research gap, we introduce CabinSI, a benchmark for evaluating spatial intelligence in cross-cabin environments, consisting of two components: RelCabin for relational reasoning tasks and RefCabin for spatial referring localization tasks. The benchmark is built upon real-world captured multi-view cabin data and systematically covers in-cabin, out-of-cabin, and cross-cabin scenarios under single-view, multi-view, and cross-cabinview settings. Furthermore, we propose a cognitive-map-based framework that projects multi-view observations onto a normalized top-down plane and constructs an explicit spatial graph as model input, allowing multimodal large language models to focus on structured reasoning. Experimental results demonstrate that such explicit spatial representation significantly improves the stability of cross-view reasoning in MLLMs. Data and code is available at CabinSI.

Cortical surface reconstruction (CSR) is widely used in neuroimag-ing and is crucial for quantitative analyses of cerebral cortical thickness andsulcal morphology. While CSR is a mature technology when applied routinelyto adult T1 -weighted brain MRI data of the cerebrum, it remains underexploredfor a broad range of MRI contrasts, resolutions, ages, species, and brain struc-tures such as the cerebellum. To address this challenge, we propose XSurfer,a contrast- and resolution-agnostic CSR framework that performs optimizationon single images using an untrained neural network such that training data arenot needed. Specifically, XSurfer starts from segmentations of 3D MRI data,applies sulcal cerebrospinal fluid (CSF) recovery in sulci to address limitationscaused by partial-volume effects, identifies CSF voxels in deep sulci, and extractstopologically correct initial white matter (WM) surfaces and pseudo-target pialsurfaces. It then employs an untrained neural network to predict velocity fields,which are driven by mesh-based and geometrically-constrained loss functionterms to achieve diffeomorphic surface deformation, thereby reconstructing accu-rate pial surfaces. We evaluated XSurfer across different contrasts, resolutionsand seven datasets spanning adult, fetal, infant, ex vivo human, and non-humanprimate brain; XSurfer provides accurate CSR for all cases, surpassing state-of-the-art learning-based methods and recon-all-clinical recently providedby FreeSurfer as a more general-purpose CSR method. By enabling precise,contrast- and resolution-agnostic CSR without pretraining, XSurfer facilitatescross-domain multi-modal neuroimaging and standardizes morphometric analysisfor neuroscience research and clinical applications. The source code is publiclyavailable at: https://github.com/birthlab/XSurfer.

The rapid deployment of visual localization (VL) in aug-mented/virtual reality (AR/VR) and autonomous systems makes privacy-preserving localization a critical societal necessity. Existing VL systemsrely on cloud-based 3D scene representation storage and transmit clientside features to a remote server, exposing users to potential reconstruc-tion attacks from intercepted privacy preserving representations. The aimof this paper is therefore to investigate what sensitive information canan adversary actually recover, by comparing multiple privacy-preservingsolutions within the literature. We define privacy as the inability to re-cover personally identifiable information from these representations, ac-knowledging that general scene details do not inherently represent a pri-vacy breach. To model adversarial behavior and to thoroughly measure amethod’s degree of privacy preservation, we introduce a new privacy at-tack that trains a conditional diffusion model to reconstruct images fromprivacy-preserving representations. This reconstruction quality serves asa direct proxy for the amount of sensitive information each represen-tation contains. Leveraging this unified attack protocol, we present thefirst comprehensive comparison across major families of privacy preserv-ing VL methods. Our analysis reveals substantial differences in privacyleakage between representations and highlights limitations in current de-sign assumptions. Code and trained checkpoints may be found here:https://github.com/MaximePi/benchmark-privacy-inversion.

Vision-language object detectors (VLODs) such as YOLO-World and Grounding DINO exhibit strong zero-shot generalization, buttheir performance degrades under distribution shift. Test-time adapta-tion (TTA) offers a practical way to adapt models online using onlyunlabeled target data. However, despite substantial progress in TTAfor vision-language classification, TTA for VLODs remains largely unex-plored. The only prior method relies on a mean-teacher framework thatintroduces significant latency and memory overhead. To this end, weintroduce VLOD-TTA, a TTA method that leverages dense proposaloverlap and image-conditioned prompts to adapt VLODs with low addi-tional overhead. VLOD-TTA combines (i) an IoU-weighted entropy ob-jective that emphasizes spatially coherent proposal clusters and mitigatesconfirmation bias from isolated boxes, and (ii) image-conditioned promptselection that ranks prompts by image-level compatibility and aggregatesthe most informative prompt scores for detection. Our experiments acrossdiverse distribution shifts, including artistic domains, adverse drivingconditions, low-light imagery, and common corruptions, indicate thatVLOD-TTA consistently outperforms standard TTA baselines and theprior state-of-the-art method using YOLO-World and Grounding DINO.Code: https://github.com/imatif17/VLOD-TTA

As Vision Language Models (VLM) are deployed across safety-critical applications, understanding and controlling their behavioral pat-terns has become increasingly important. Existing behavioral controlmethods face significant limitations: system prompting is a popular ap-proach but could easily be overridden by user instructions, while ap-plying activation-based steering vectors requires invasive runtime accessto model internals, precluding deployment with API-based services andclosed-source models. Finding steering methods that transfer across mul-tiple VLMs is still an open area of research. To this end, we introduce vi-sual input based steering for output redirection (VISOR++), a novel ap-proach that achieves behavioral control through optimized visual inputsalone. We demonstrate that a single VISOR++ image can be generatedfor two architecturally diverse VLMs that by itself can emulate each oftheir steering vectors. By crafting universal visual inputs that induce tar-get activation patterns for an ensemble of models, VISOR++ eliminatesthe need for runtime model access while remaining deployment-agnostic.This means that when an underlying model supports multimodal capa-bility, model behaviors can be steered by inserting an image input com-pletely replacing runtime steering vector based interventions. We firstdemonstrate the effectiveness of the VISOR++ images on open-accessmodels such as LLaVA-1.5-7B and IDEFICS2-8B along three alignmentdirections: refusal, sycophancy and survival instinct. Both the model-specific steering images and the jointly optimized images achieve perfor-mance parity closely following that of steering vectors for both positiveand negative steering tasks. We also show early promise of VISOR++images in achieving directional behavioral shifts for unseen models thatinclude both open-access and closed-access models. At the same time,VISOR++ images are able to preserve 99.9% performance on 14,000unrelated MMLU evaluation samples highlighting their specificity to in-ducing only behavioral shifts.

Scienti_x001C_c research based on multimodal clinical data (includ-ing medical imaging) requires coordinating clinical, radiological, pro-gramming, and biostatistical expertise, a fragmented process that bottle-necks discovery. We present Veritas (Veri_x001C_able Epistemic Reasoningfor Image-Derived Hypothesis Testing via Agentic Systems), a clinicalco-scientist: a multi-agent system that autonomously tests natural-language hypotheses and produces a fully auditable evidence trail, trac-ing every conclusion through executable outputs from analysis plan tosegmentation masks to statistical code to _x001C_nal verdict. Unlike prior AI-scientist systems, which mainly operate on tabular or text data, Veritasgrounds autonomous discovery directly in medical images. It decomposesthe work_x001D_ow into four phases handled by role-specialized agents, and in-troduces an epistemic evidence label framework that mechanically classi-_x001C_es outcomes as Supported, Refuted, Underpowered, or Invalid byjointly evaluating signi_x001C_cance, e_x001B_ect direction, and study power. This dis-tinction is critical in medical imaging, where non-signi_x001C_cant results oftenre_x001D_ect insu_x001E_cient sample size rather than absent e_x001B_ects. We construct atiered benchmark of 64 hypotheses spanning six complexity levels acrosscardiac and brain glioma MRI datasets. Veritas reaches 81.4% verdictaccuracy with frontier models and 71.2% with locally-hosted open-weightmodels (8_x0015_30B), outperforming all single-model baselines in both classes.It also produces the highest rate of independently veri_x001C_able statisticaloutputs (86.6%), so even its failures remain diagnosable through arti-fact inspection. Structured multi-agent decomposition thus substitutesfor model scale while preserving the veri_x001C_ability that scienti_x001C_c discoverydemands. We release code, hypothesis bank, and evaluation pipeline athttps://github.com/LucZot/veritas.

Diffusion model advances have enabled powerful text-guidedimage editing, but also raise ethical and legal risks such as deepfakesand unauthorized use. To prevent these risks, adversarial attack-basedimage immunization has emerged as a promising defense against AI-driven semantic manipulation. Yet, most existing approaches requireimage-specific optimization or additional neural networks at inferencetime, hindering scalability and practicality. In this paper, we proposethe first universal adversarial perturbation-based image immunizationframework that generates a single, image-agnostic adversarial perturbationspecifically designed for diffusion-based editing pipelines. Inspired by UAPused in targeted attacks, our method aims to generate a UAP that inducesdiffusion models to misinterpret the input image as a specific semantictarget. Simultaneously, it suppresses original content to misdirect themodel’s attention during editing, thereby effectively blocking unauthorizededits by overwriting the image’s original semantics via the UAP. Extensiveexperiments show that our method, as the first universal immunizationapproach, significantly outperforms several baselines in the UAP setting.Notably, despite the inherent difficulty of universal perturbations, ourmethod achieves competitive or superior performance compared to image-specific methods under a more restricted perturbation budget, while alsoexhibiting strong black-box transferability across diverse diffusion models.

Photorealistic volumetric rendering of CT scans greatly ben-e_x001C_ts clinical work_x001D_ows, yet neural approaches such as Neural RadianceFields (NeRF) and 3D Gaussian Splatting (3DGS) require prohibitiveper-scan optimization (hours for NeRF, about 30 minutes for 3DGS),making them impractical in clinical settings. We propose Render-FM,a feedforward model that eliminates this bottleneck by directly regress-ing 6D Gaussian Splatting (6DGS) parameters from a CT volume ina single 2.8-second forward pass, a 500× speedup over per-scan opti-mization. To bridge the domain gap between natural scene reconstruc-tion and medical volumetric rendering, we introduce Anatomy-GuidedPriming (AGP), which incorporates segmentation masks and transferfunctions as structural and appearance priors, information that exist-ing Gaussian splatting methods overlook. Built on an nnU-Net-inspired3D U-Net trained on diverse CT scans, Render-FM predicts per-voxel6DGS parameters and supports immediate real-time rendering. Unlikeper-scan methods, it generalizes to unseen anatomies, novel transferfunctions, and enables compositional organ visualization with zero ad-ditional preparation time. Optional 89-second _x001C_ne-tuning further im-proves quality, surpassing per-scan optimized baselines. Project page:https://gaozhongpai.github.io/renderfm/.

Diffractive Snapshot Spectral Imaging (DSSI) encodes spec-tral and spatial information into a single 2D measurement via a diffrac-tive optical element (DOE), offering a compact and light-efficient al-ternative to conventional Coded-Aperture Snapshot Spectral Imaging(CASSI). However, the spatially uneven coding of DSSI, which is mainlyconcentrated in texture-rich regions, has led existing reconstruction net-works to rely heavily on global attention modules to propagate spectralfeatures into texture-less regions, making them short-sighted in locallycoded areas and computationally redundant. To address these challenges,we propose Local-to-Global Spectral Reconstruction Network (LGSR-Net), a two-stage progressive framework comprising complementary lo-cal and global subnetworks. The local part extracts information fromtexture-rich regions under direct supervision from effectively coded ar-eas, while the global part models long-range dependencies to facilitateglobal information flow. We introduce a spatially guided channel atten-tion module to emphasize informative regions and an edge-aware gatedfusion mechanism to integrate local features into the global subnetwork’sdecoders. Moreover, a coding-aware sample selection strategy and a tai-lored loss function are designed to improve training efficiency and re-construction fidelity. Extensive experiments on both simulated and realDSSI datasets show that LGSRNet achieves state-of-the-art reconstruc-tion performance with significantly lower computational cost and latency.

LiDAR scene flow is the task of estimating per-point 3D mo-tion between consecutive point clouds. Recent methods achieve centimeter-level accuracy on popular autonomous vehicle (AV) datasets, but aretypically only trained and evaluated on a single sensor. In this paper, weaim to learn general motion priors that transfer to diverse and unseenLiDAR sensors. However, prior work in LiDAR semantic segmentationand 3D object detection demonstrate that naively training on multipledatasets yields worse performance than single dataset models. Interest-ingly, we find that this conventional wisdom does not hold for motionestimation, and that state-of-the-art scene flow methods greatly benefitfrom cross-dataset training without architectural modification. We positthat low-level tasks such as motion estimation may be less sensitive tosensor configuration; indeed, our analysis shows that models trained onfast-moving objects (e.g., from highway datasets) perform well on fast-moving objects, even across different datasets. Informed by our analy-sis, we propose UniFlow, a feedforward model that unifies and trainson multiple large-scale LiDAR scene flow datasets with diverse sensorplacements and point cloud densities. Our frustratingly simple solutionestablishes a new state-of-the-art on Waymo and nuScenes, improvingover prior work by 5.1% and 35.2% respectively. Moreover, UniFlowachieves state-of-the-art accuracy on unseen datasets like TruckScenesand AEVAScenes, outperforming prior dataset-specific models by 30.1%and 22.5% respectively. See our project page for additional visuals.

The practical deployment of autoregressive (AR) Vision-Language-Action (VLA) models is severely constrained by inference speed bottlenecks, particularly in high-frequency and dexterous manipulation tasks. While recent studies have explored Jacobi decoding as a more efficient alternative to traditional autoregressive decoding, its practical benefits are marginal due to the lengthy iterations. To address this problem, we introduce consistency distillation to teach the model to predict multiple correct action tokens in each iteration, thereby reducing the total iterations. While the distillation brings moderate speedup, we identify that certain redundancy iterations remain a critical limitation. To tackle this, we propose an adaptive early-exit decoding strategy that moderately relaxes convergence conditions, which further improves average inference efficiency. Experimental results show that the proposed method achieves more than 4× inference acceleration across different base models while maintaining high task success rates in both simulated and real-world robot tasks. These experiments validate that our approach provides an efficient and general paradigm for accelerating multimodal decision-making in robotics. Our code and videos could be found in the supplementary files.

Novel-view synthesis of dynamic scenes, crucial for AR/VRapplications, remains a challenging problem. Recent methods adapt rep-resentations like 3D Gaussian Splatting (3DGS) and Neural RadianceFields (NeRF) for dynamic scenes by incorporating time as the fourth di-mension (4D representations). These 4D representations still suffer fromaliasing artifacts, especially when generating novel views from divergentviewpoints (zoom-in/zoom-out operations). While using 3D smoothingfilters like those proposed in Mip-Splatting might seem like a possible so-lution, they fail to account for local motion and also exhibit aliasing. Toaddress this, we propose a motion-aware 3D smoothing filter specificallydesigned for 4D representations. Our approach adapts the filter strengthbased on local motion information, effectively mitigating aliasing withoutcompromising rendering quality. This is achieved by estimating the jointdensity function of time and focal-to-depth ratio using a non-parametricestimation method. During inference, we sample from this joint distribu-tion to determine the appropriate smoothing filter. This flexible strategycan be integrated with various 4D representations. Our evaluations onstandard datasets demonstrate superior performance compared to state-of-the-art methods.

The frontier of visual reasoning is shifting toward models likeOpenAI o3, which can intelligently create and operate tools to trans-form images for problem-solving, also known as thinking-with-images inchain-of-thought. Yet existing benchmarks fail to fully capture this ad-vanced capability. Even Visual Search, the most common benchmark forcurrent thinking-with-images methods, tests only basic operations suchas localization and cropping, offering little insight into more complex,dynamic, and tool-dependent reasoning. We introduce TIR-Bench, acomprehensive benchmark for evaluating agentic thinking-with-imagesacross 13 diverse tasks, each requiring novel tool use for image process-ing and manipulation in chain-of-thought. We evaluate 22 multimodallarge language models (MLLMs), from leading open-sourced and propri-etary models to those with explicit tool-use augmentation. Results showthat TIR-Bench is universally challenging, and strong performance re-quires genuine thinking-with-images capabilities. Finally, we present apilot study comparing direct versus agentic fine-tuning.

Vision-language models (VLMs) are increasingly deployedin high-stakes settings where reliable uncertainty quantification (UQ)is as important as predictive accuracy. Extended reasoning via chain-of-thought (CoT) prompting or reasoning-trained models has becomeubiquitous in modern VLM pipelines, yet its effect on UQ reliabilityremains poorly understood. Our results show that reasoning tends todegrade the quality of many uncertainty estimates, even when it im-proves task accuracy. We identify implicit answer conditioning as theprimary mechanism: as reasoning traces converge on a conclusion beforethe final answer is generated, token probabilities increasingly reflect con-sistency with the model’s own reasoning trace rather than uncertaintyabout correctness. In effect, the model becomes overconfident in its an-swer. In contrast, agreement-based consistency remains robust and oftenimproves under reasoning, making it a practical choice for uncertaintyestimation in reasoning-enabled VLMs.