ECCV 2026 Papers

Remote sensing underpins crucial applications such as disas-ter relief and ecological field surveys, where systems must understandcomplex scenes and constraints and make reliable decisions. Currentremote-sensing benchmarks mainly focus on evaluating perception andreasoning capabilities of multimodal large language models (MLLMs).They fail to assess planning capability, stemming either from the diffi-culty of curating and validating planning tasks at scale or from evaluationprotocols that are inaccurate and inadequate. To address these limita-tions, we introduce NeSy-Route, a large-scale neuro-symbolic benchmarkfor constrained route planning in remote sensing. Within this benchmark,we introduce an automated data-generation framework that integrateshigh-fidelity semantic masks with heuristic search to produce diverseroute-planning tasks with provably optimal solutions. This allows NeSy-Route to comprehensively evaluate planning across 10,821 route-planningsamples, nearly 10 times larger than the largest prior benchmark. Fur-thermore, a three-level hierarchical neuro-symbolic evaluation protocol isdeveloped to enable accurate assessment and support fine-grained analy-sis on perception, reasoning, and planning simultaneously. Our compre-hensive evaluation of various state-of-the-art MLLMs demonstrates thatexisting MLLMs show significant deficiencies in perception and planningcapabilities. We hope NeSy-Route can support further research and de-velopment of more powerful MLLMs for remote sensing.The dataset andcode are available at https://mingyang1010.github.io/NeSy-Route/.

As AI systems are being integrated more rapidly into di-verse and complex real-world environments, the ability to perform holis-tic reasoning over an implicit query and an image to localize a target isbecoming increasingly important. However, recent reasoning segmenta-tion methods fail to sufficiently elicit the visual reasoning capabilities ofthe base model. In this work, we present Segment Anything Reasoner(StAR), a comprehensive framework that refines the design space frommultiple perspectives—including parameter-tuning scheme, reward func-tions, learning strategies and answer format—and achieves substantialimprovements over recent baselines. In addition, for the first time, wesuccessfully introduce parallel test-time scaling to the segmentation task,pushing the performance boundary even further. To eXtend the scopeand depth of reasoning covered by existing benchmark, we also con-struct the ReasonSeg-X, which compactly defines reasoning types and in-cludes samples that require deeper reasoning. Leveraging this dataset, wetrain StAR with a rollout-expanded selective-tuning approach to activatethe base model’s latent reasoning capabilities, and establish a rigorousbenchmark for systematic, fine-grained evaluation of advanced methods.With only 5k training samples, StAR achieves significant gains over itsbase counterparts across extensive benchmarks, demonstrating that ourmethod effectively brings dormant reasoning competence to the surface.

Model merging has emerged as a promising paradigm for enabling multi-task capabilities without additional training. However, traditional basic merging methods often experience performance degradation due to parameter conflicts, even when applied to similar tasks. While recent personalized merging frameworks successfully preserve taskspecific information to maintain performance, they typically incur storage overhead. In this paper, we propose Decomposition, Thresholding, and Scaling (DTS), an approximation-based personalized merging framework that pushes task-specific storage efficiency. DTS first applies singular value decomposition to the task-specific information and retains only a small subset of singular values and vectors. It then introduces a novel thresholding strategy that partitions singular vector elements into groups and assigns a scaling factor to each group. To enable generalization to unseen tasks, we further extend DTS with a variant that fuses task-specific information in a data-free manner based on the semantic similarity of task characteristics. Extensive experiments demonstrate that DTS consistently outperforms state-of-the-art baselines while requiring only 1% extra storage per task. Furthermore, experiments on unseen tasks show that the DTS variant achieves significantly better generalization performance. Our code is available at https://github.com/krumpguo/DTS.

Recent advances in single image-to-3D generation have en-abled high-quality asset synthesis, yet extending these capabilities toindoor scene generation remains challenging. Existing methods focus onasset-level generation while neglecting the structural layout, which isessential for downstream applications and serves as the spatial anchorfor grounding assets. However, a single image with a limited field ofview lacks the spatial coverage to recover a coherent global layout. Tothis end, we use a 360° image represented in equirectangular projec-tion (ERP) and propose InSpace, a structure-aware framework for 3Dindoor scene generation. InSpace comprises three stages: (1) estimat-ing partial scene geometry as spatial priors, (2) generating coarse scene* †Work done during an internship at NAVER LABS. Co-corresponding authorsArtifactsMisplacementMisplacementERP Image ERP ImageInput Result Floating Input Result ResultResultSceneGen SAM3D InSpace (Ours)(a) Current Single Image to 3D Scene Generation (b) ERP Image to 3D Scene GenerationFig. 2: (a) Existing single-image methods generate individual assets without structurallayout, causing floating, misplacement, and artifacts. (b) InSpace uses an ERP imageto generate complete indoor scenes with structural layout and well-grounded assets.structure with view-selective cross-attention, and (3) producing detailedlayout and asset geometry with textures through a global-local hybridattention, using flow matching. We also propose ERP-FRONT, a pairedERP-Image-to-3D indoor scene dataset based on 3D-FRONT. Exper-iments show that InSpace generates complete 3D indoor scenes withstructural layout, along with separate textured assets from a single ERPimage, achieving strong performance across 3D and 2D metrics.

While recent flow-matching 3D generative models (e.g., Vec-Set) adopt structured representations, their tokens share global context,causing conventional training-free editing to suffer from semantic arti-facts such as collapsed preserved regions or incomplete transformations.To address this, we propose TanGO, a training-free framework thatenables adaptive per-token steering in the tangent space of generativedynamics. To realize this selective control, we formulate a one-step opti-mal control rule and determine the strength of each token’s control signalusing a von Mises-Fisher inspired directional discrepancy derived fromthe source and target velocity fields. Experiments show that TanGOsubstantially reduces structural artifacts and achieves state-of-the-artperformance, outperforming existing 3D editing baselines. The code ispublicly available at https://github.com/siw00-lim/TanGO.

Accurately reconstructing complex full multi-object scenesfrom sparse observations remains a core challenge in computer vision anda key step toward scalable and reliable simulation for robotics. In thiswork, we introduce RecGen, a generative framework for probabilisticjoint estimation of object and part shapes, as well as their pose underocclusions and partial visibility from one or multiple RGB-D images. Byleveraging compositional synthetic scene generation and strong 3D shapepriors, RecGen generalizes across diverse object types and real-worldenvironments. RecGen achieves state-of-the-art performance on complex,heavily occluded datasets, robustly handling severe occlusions, symmetricobjects, objects parts, and intricate geometry and texture. Despite usingnearly 80% fewer training meshes than the previous state of the artSAM3D, RecGen outperforms it by 30.1% in geometric shape quality,9.1% in texture reconstruction, and 33.9% in pose estimation. We willrelease our code, training data and evaluation benchmark at our website.

Thermal infrared (TIR) imaging facilitates perception, se-curity surveillance, and industrial inspection. Compared with visiblespectrum, TIR observations are typically resolution-limited and texture-scarce. More critically, thermal diffusion smooths true geometric bound-aries into low-frequency blur, whereas non-uniform sensor response andfixed-pattern noise introduce structured high-frequency artifacts, yield-ing a pronounced signal-to-noise inversion. Consequently, reconstructinga high-resolution (HR) 3D TIR scene and synthesizing high-quality novelviews from only low-resolution (LR) infrared inputs remains challeng-ing. To address this, we propose the first imaging-calibrated TIR super-resolution 3D Gaussian Splatting (3DGS) framework that reconstructshigh-resolution 3D TIR scenes using only LR inputs. Specifically, we ex-plicitly model infrared imaging degradation on the clean radiation signalsrendered by 3DGS using physics-informed priors, enabling decoupling ofthe underlying thermal radiation from sensor-induced artifacts. To allevi-ate texture scarcity, an intensity-conditioned tone adapter is designed toapply local affine residual modulation, thereby surfacing latent thermalgradients. Finally, we implement a frequency-aware curriculum learn-ing strategy that leverages edge priors from a super-resolution teachernetwork, and employs a cosine ramp-up schedule to smoothly shift theoptimization focus from coarse low-frequency structures to fine high-frequency details. Extensive evaluations across TIR benchmarks showthat our method significantly outperforms existing works in both high-resolution synthesis quality and quantitative temperature fidelity.

Multi-view computer vision pipelines typically rely on accu-rate sparse keypoints and robust descriptors. While incorporating linefeatures has shown clear benefits for matching and pose estimation, exist-ing point-line approaches remain inefficient: they detect points and linesseparately, use increasingly heavy networks, and depend on CPU-boundheuristics that hinder real-time performance. We introduce a Unified Effi-cient Points and Lines (UPAL) feature extractor that jointly extracts key-points, line segments, and feature descriptors within a single lightweightarchitecture. A shared backbone provides common representations thatfeed different branches for point and line features. Line segments are recov-ered through an accelerated post-processing stage, an enhanced and highlyefficient variant of the LSD algorithm. UPAL matches or exceeds state-of-the-art performance in both point and line applications while significantlyreducing computational cost, achieving, for instance, a 4× speedup and10× smaller memory footprint over the ALIKED + DeepLSD pipeline.Code is publicly available at https://github.com/francois141/upal.

Detecting anomalies in images and video is an essential task for multi-ple real-world problems, including industrial inspection, computer-assisted diag-nosis, and environmental monitoring. Anomaly detection is typically formulatedas a one-class classification problem, where the training data consists solely ofnominal values, leaving methods built on this assumption susceptible to train-ing label noise. We present Bootstrap Aggregation Anomaly Filtering (BAAF),a method that transforms an arbitrary one-class classifier-based anomaly detec-tor into a fully unsupervised method. This is achieved by leveraging the uniqueintrinsic properties of anomaly detection: anomalies are uncommon in the sam-pled data and generally heterogeneous. These properties enable us to design amodified Bootstrap Aggregation method that uses multiple independently trainedinstances of supervised one-class classifiers to filter the training dataset for anoma-lies. This transformation requires no modifications to the underlying anomaly de-tector; only the algorithmically selected data bags used for training change. Wedemonstrate empirically that our method can transform a wide variety of one-classclassifier-based image anomaly detectors into unsupervised ones. Consequently,we present the first fully unsupervised logical anomaly detection method for im-ages. We also demonstrate that our method achieves state-of-the-art performancein fully unsupervised anomaly detection on the MVTec AD, ViSA, Real-IAD andMvTec LOCO AD datasets. As improvements to one-class classifiers are made,our method directly transfers those improvements to the unsupervised domain,linking the domains.

Federated learning (FL) enables collaborative learning ofcomputer vision models, where privacy and regulatory constraints pre-vent centralizing data across devices or organizations. However, prac-tical FL deployments often exhibit severe class imbalance and labelskew, causing standard aggregation protocols to overfit dominant clientsand degrade minority-class performance. We propose a data-free, class-wise contribution estimation and aggregation framework based on logitmaximization (CELM) that does not require raw data, client metadata,or auxiliary public datasets at the server. The FL server probes clientupdates to obtain class-wise evidence scores and assembles a cross-clientevidence matrix, which quantifies both per-class competence and classcoverage. Using this matrix, we compute contribution weights that up-weight clients providing discriminative evidence for underrepresentedclasses. The resulting aggregation is stable due to simplex constraints andmomentum smoothing, and remains compatible with standard FL train-ing pipelines. We evaluate the approach on representative vision bench-marks under controlled non-IID and pathological label splits, demon-strating that CELM-based aggregation improves robustness to imbalanceand statistical heterogeneity, while yielding better performance with-out requiring any additional data exchange. The code is available at:https://github.com/asimukaye/celm.

Computer-aided design (CAD) plays a fundamental role inmodern manufacturing by providing the high precision required for in-dustrial production. Recent large language model based approaches for-mulate CAD generation as a sequence prediction problem and haveachieved promising results. However, existing methods and evaluationprotocols primarily emphasize visual similarity, while overlooking precisegeometric parameters and correct metric scale. Small numerical devia-tions that are negligible at the shape-level may still violate industrial tol-erance requirements, a problem further compounded by current autore-gressive paradigms that utilize command sequence representations, ag-gressively quantize numerical parameters to ease LLM prediction. In thiswork, we present Pointer-CAD v2. Compared with v1 [36], this versiondirectly predicts continuous values, bypassing the need for quantized nu-merical parameters and thereby eliminating quantization errors. Specif-ically, we propose a unified framework that decouples parameter rea-soning from geometric construction through a Plan-Then-Constructparadigm. Our method first produces a structured design plan withexplicit metric scale parameters. These parameters are organized intoa dictionary and directly referenced during sequence generation via apointer mechanism, eliminating discretization errors and ensuring dimen-sionally consistent execution. In addition, we construct a new large-scaledataset with plan-level annotation and introduce three hierarchical geom-etry accuracy metrics to evaluate parametric fidelity at the vertex, edge,and face levels. Extensive experiments demonstrate that Pointer-CADv2 consistently outperforms existing baselines and achieves substantialimprovements in geometric accuracy, enabling reliable CAD generationfor precision-critical engineering applications. Our code is available athttps://github.com/Snitro/Pointer-CAD-v2.

Normalizing Flows (NFs) are powerful generative models ca-pable of exact density estimation and sampling. However, their strictinvertibility often forces the model to exhaust its capacity on low-levelpixel details, hindering the capture of high-level semantic structures.While Masked Image Modeling (MIM) has excelled in representationlearning, its integration into generative pipelines has remained largelymodular and disjointed. In this paper, we propose MIMFlow, a unifiedend-to-end framework that jointly optimizes latent semantics, pixel re-construction, and generative flow. By employing a VAE encoder to infersemantic latent from masked images, MIMFlow achieves a principleddecoupling of the generative task: the Normalizing Flow focuses on mod-eling a simplified, low-frequency semantic manifold, while a specializeddecoder handles high-frequency synthesis. This design effectively resolvesthe inherent capacity bottleneck of NFs, allowing the model to prioritizeglobal structural coherence over redundant noise. Empirical results onImageNet 256×256 show that MIMFlow-L reaches 71.3% linear prob-ing accuracy and an FID of 2.50. Despite using only 128 tokens (50%fewer than standard models), it yields a 32.8% performance gain oversimilar-scale NF baselines.

The evaluation of long-term video quality understanding re-mains an open challenge for large vision-language models (LVLMs). Ex-isting video quality benchmarks predominantly focus on short clips andisolated distortions, overlooking the temporal continuity, cumulative degra-dation, and reasoning complexity inherent in long-duration content. Toaddress these limitations, we present LongVQUBench, a comprehen-sive benchmark for long-term video quality understanding. LongVQUBenchcontains over 1,200 diverse videos spanning movies, documentaries, surveil-lance footage, egocentric recordings, and animated content, accompa-nied by 1,500 multiple-choice and open-ended questions for validationand testing. To assess perceptual reasoning across different temporalscopes, we introduce three progressively complex evaluation levels: (i)local event quality understanding (LQU) for analyzing localized distor-tions; (ii) cross-event quality reasoning (CQR) for integrating multipledegraded events; and (iii) global quality understanding (GQU) for holis-tic perceptual evaluation over extended durations. Furthermore, a needledistortion question-answering (NDQA) paradigm is embedded across allthree levels, where spatial or temporal artifacts are sparsely inserted toprobe fine-grained detection and reasoning capabilities. Extensive ex-periments on 14 state-of-the-art LVLMs reveal significant performancedegradation with increasing video length and reasoning depth, highlight-ing their limited capacity for long-range temporal integration and percep-tual attribution. We envision LongVQUBench as a foundational step to-ward the systematic, hierarchical, and explainable evaluation of LVLMs’long-term video quality understanding.

Despite advances in indoor scene generation, synthesizingcoherent building exteriors consistent with generated interiors remainslargely unexplored. Existing methods can generate floor plans and walllayouts but typically stop at a structural shell, lacking stylistically con-sistent facades and roofs. Completing these exteriors is challenging be-cause the footprint, wall geometry, and opening semantics must remainfixed—constraints that unconstrained generative models often violate.We introduce ShellMaker, a language-guided exterior completion frame-work that operates under these structural constraints. Given a build-ing scaffold and a text style prompt, ShellMaker generates a completeexterior mesh with PBR materials by combining parametric roof gen-eration, LLM-based part-aware prompt refinement, joint wall–roof ma-terial retrieval, and geometry-aware assembly. Operating on a format-agnostic scaffold representation, ShellMaker generalizes to indoor gener-ators, CityGML, and CAD inputs, while maintaining structural consis-tency and improving architectural coherence over retrieval and uncon-strained generative baselines. The project page is available at https://ruiqixu37.github.io/ShellMaker_web/.

End-to-end multi-modal planning has been widely adopted to model the uncertainty of driving behavior, typically by scoring candidate trajectories and selecting the optimal one. Existing approaches generally fall into two categories: scoring a large static trajectory vocabulary, or scoring a small set of dynamically generated proposals. While static vocabularies often suffer from coarse discretization of the action space, dynamic proposals provide finer-grained precision and have shown stronger empirical performance on existing benchmarks. However, it remains unclear whether dynamic generation is fundamentally necessary, or whether static vocabularies can already achieve comparable performance when they are sufficiently dense to cover the action space. In this work, we start with a systematic scaling study of Hydra-MDP, a representative scoring-based method, revealing that performance consistently improves as trajectory anchors become denser, without exhibiting saturation before computational constraints are reached. Motivated by this observation, we propose SparseDriveV2 to push the performance boundary of scoring-based planning through two complementary innovations: (1) a scalable vocabulary representation with a factorized structure that decomposes trajectories into geometric paths and velocity profiles, enabling combinatorial coverage of the action space, and (2) a scalable scoring strategy with coarse factorized scoring over paths and velocity profiles followed by fine-grained scoring on a small set of composed trajectories. By combining these two techniques, SparseDriveV2 scales the trajectory vocabulary to be 32× denser than prior methods, while still enabling efficient scoring over such super-dense candidate set. With a lightweight ResNet-34 as backbone, SparseDriveV2 achieves 92.0 PDMS and 90.1 EPDMS on NAVSIM, with 89.15 Driving Score and 70.00 Success Rate on Bench2Drive. Code and model are released at https://github.com/swc-17/SparseDriveV2.

Visual Autoregressive Models (VAR) generate images throughnext-scale prediction, producing all tokens within each scale in paral-lel. We show that this parallel decoding constitutes a mean-field-styleapproximation that discards spatial dependencies among same-scale to-kens, causing locally incoherent samples regardless of backbone capacity– a limitation of the decoding rule. Addressing this limitation, we intro-duce the Logit Refiner, a lightweight autoregressive module that restoresintra-scale dependencies by sequentially sampling tokens conditioned onfrozen backbone features. Adding only ∼10% parameters and less than5% of the base model’s training compute, it plugs into any pretrainedVAR checkpoint without retraining. Controlled ablations isolate jointintra-scale sampling – rather than additional capacity or training – asthe critical ingredient. Across backbones from 310M to 2B parameterson class-conditional ImageNet 256×256, the refiner consistently improvesgeneration quality, enabling a 1.1B-parameter model to surpass one twiceits size. The approach further generalizes to text-to-image generation,confirming that the mean-field bottleneck persists across VAR variantsand is effectively alleviated by our method.Project page: https://compvis.github.io/logit-refiner/.

This paper introduces an inverse-rendering framework fornatural-light, uncalibrated photometric stereo (PS) that leverages back-ground cues captured alongside the target object. Natural-light PS (NaPS)acquires shading variations by moving or rotating the camera and objectunder fixed, uncontrolled illumination, such as indoor lighting, whilemaintaining their relative geometry. However, uncalibrated NaPS, inwhich the lighting conditions are unknown, remains inherently ill-posed.To tackle this challenge, we propose explicitly reconstructing the lightingenvironment from the image background, which is typically masked outin prior work, thereby converting uncalibrated NaPS into a tractableinverse-rendering problem. Specifically, we move and rotate the camera-object pair while keeping their relative pose fixed, and reconstruct thesurrounding illumination directly from the observed background usinga 3D Gaussian Splatting (3DGS) representation. We then optimize thetarget object’s shape and reflectance via inverse rendering under thereconstructed illumination. Experimental results demonstrate that ourinverse-rendering-based approach yields more accurate estimates of bothgeometry and reflectance than existing learning-based PS methods, espe-cially under realistic near-field indoor conditions.

Text Image Machine Translation (TIMT) aims to translatetext embedded in images in the source-language into target-language,requiring synergistic integration of visual perception and linguistic un-derstanding. Existing TIMT methods, whether cascaded pipelines orend-to-end multimodal large language models (MLLMs), struggle withhigh-resolution text-rich images due to cluttered layouts, diverse fonts,and non-textual distractions, resulting in text omission, semantic drift,and contextual inconsistency. To address these challenges, we proposeGLoTran, a global-local dual visual perception framework for MLLM-based TIMT. GLoTran integrates a low-resolution global image withmulti-scale region-level text image slices through pre-local region trans-lation replay and an instruction-guided alignment strategy, conditioningMLLMs to maintain scene-level contextual consistency while faithfullycapturing fine-grained textual details. Moreover, to realize this dual-perception paradigm, we construct GLoD, a large-scale text-rich TIMTdataset comprising 510K high-resolution global-local image-text pairscovering diverse real-world scenarios. Extensive experiments demonstratethat GLoTran substantially improves translation completeness and ac-curacy over state-of-the-art MLLMs.

Event cameras enable high-frequency visual perception withmicrosecond latency, offering advantages for dynamic scenes. However,event-based small object detection remains challenging due to sparseasynchronous measurements and weak object responses that are easilydisrupted by noise. Limited spatial support causes small-object signalsto lose temporal continuity, resulting in fragmented and unstable pre-dictions. To address this issue, we propose a physics-guided advection-consistent modeling framework, termed PACT, which formulates eventevolution as a motion-driven feature transport process. Instead of rely-ing solely on local spatio-temporal aggregation, PACT propagates fea-tures along estimated velocity fields and enforces trajectory-level consis-tency through advection constraints. This design preserves weak eventresponses over time and prevents their degradation under complex back-ground interference. Technically, PACT integrates motion-aware featureextraction with a differentiable advection-based transport operator, en-abling coherent motion representation and effective noise suppressionduring temporal evolution. Extensive experiments on benchmark event-based datasets demonstrate that PACT consistently outperforms state-of-the-art methods, achieving improvements of 20.72% in IoU and 15.03%in accuracy while maintaining comparable computational efficiency. Thecode will be made publicly available.