ECCV 2026 论文列表

Open-Vocabulary Camouflaged Object Segmentation aimsto segment camouflaged objects from unseen categories. Existing two-stage methods typically treat category semantics as a one-time priorfor segmentation, thereby lacking deep and iterative interaction betweentextual semantics and visual spatial structures. To address this limita-tion, we propose an iterative two-stage refinement framework based onsemantic context and spatial structure. Specifically, we first design aspatial-structure-aware category re-ranker that leverages segmentationmasks generated by SAM2 to reorder CLIP’s candidate categories ac-cording to region-level visual consistency. Furthermore, we introduce asemantic-context-based segmentation modulator that injects the refinedcategory information into SAM2, guiding SAM2 to produce more preciseand discriminative segmentation results. Notably, the proposed re-rankerand modulator are jointly optimized in an iterative manner, forming aclosed-loop refinement process that enables mutual guidance between se-mantic representations and spatial features. Through iterative updates,semantic context and spatial structure are progressively enhanced. Ex-tensive experiments demonstrate that the proposed method significantlyoutperforms existing open-vocabulary methods and camouflaged objectsegmentation approaches. In particular, the proposed method achieves a14.4% improvement in the cSm metric compared with OVCoser.

We propose Self-Augmented Residual 3D Gaussian Splat-ting, a novel framework for stabilizing uncertainty quantification andenhancing uncertainty-aware supervision in Next-Best-View selection foractive scene reconstruction. To efficiently estimate scene coverage, SA-ResGS generates geometry-consistent Self-Augmented point clouds (SA-Points) via triangulation between observed training views and rasterizedextrapolated views. To address the lack of learning signals in underrep-resented regions within sparse, wide-baseline settings, we introduce thefirst skip-connection-inspired residual learning strategy tailored for 3DGS.This mechanism amplifies gradient flow to weakly contributing, high-uncertainty Gaussians. Our contributions are threefold: (1) a physicallygrounded, diversified view selection strategy; (2) an uncertainty-awareresidual supervision scheme that improves gradient flow and learningstability; and (3) implicitly debiased uncertainty quantification resultingfrom constrained view selection and residual supervision. Experiments onNeRF Synthetic, Mip-NeRF 360, and challenging extended benchmarkfrom Deep Blending and Tanks and Temples demonstrate that SA-ResGSconsistently outperforms state-of-the-art competing methods in bothreconstruction quality and view selection robustness.

Cross-spectral image synthesis from visible to infrared facessigni_x001C_cant challenges due to the severe scarcity of infrared data in criti-cal scenarios. Existing methods primarily achieve style transfer throughpixel-level mappings, which often su_x001B_er from semantic misalignment andunrealistic spectral characteristics. To address this problem, we rethinkcross-spectral generation as a decoupling and reconstruction process ofshared-speci_x001C_c features between modalities and propose SHASP, a novelgenerative network. Speci_x001C_cally, an encoding-decoupling module is de-signed at the front of the generator, which consists of a structure contentencoder and a spectral feature encoder to decouple the shared-speci_x001C_cfeatures. A decoding-reconstruction module is designed at the backendof the generator to perform feature combination and reconstruction. Wefurther impose semantic consistency constraints on shared features toensure precise cross-modal alignment. Extensive experiments on aerial,driving, and monitoring scenarios demonstrate that our method outper-forms state-of-the-art approaches. This work reveals that there are com-mon representations that can be mined at the semantic structure levelin cross-spectral data, and modality gaps can be accurately modeled viaa decoupling-reconstruction mechanism.

The adoption of powerful diffusion models is hindered bytheir significant inference latency. Recent “cache-then-forecast” schemesalleviate this issue by accelerating DiTs using derivative-based polynomi-als, but they suffer from severe quality degradation at high accelerationratios. Our analysis reveals its root cause: the discrete extrapolation per-formed on representations that are misaligned with the continuous diffu-sion trajectory and are numerically unstable. Thus, accelerated DiTs suf-fer from accumulated spatial errors, noisy derivative amplification, andhigh-order instability. We therefore reformulate accelerated inference asstable macro-trajectory extrapolation in ordinary differential equation(ODE) space. Instead of predicting intermediate features, we align fore-casting with the model’s Global Drift (GD), i.e., the end-to-end stateevolution, thereby eliminating feature inconsistency and memory over-head. However, even this smooth macro-trajectory remains vulnerableto the derivative fallacy: its higher-order temporal derivatives are intrin-sically noisy. Thus, we introduce a derivative-free barycentric Lagrangeextrapolator to effectively bypass derivative instability and approxima-tion error. We further propose a bounded Phase Mapping that regu-larizes the extrapolation domain, suppressing oscillatory error growth.These elements collectively constitute ResilPhase, a noise-resilient ac-celeration framework. Experiments on FLUX.1-dev and HunyuanVideodemonstrate state-of-the-art fidelity under aggressive acceleration ratios.Code is publicly available at https://github.com/zqc214/ResilPhase.

We introduce ReconSplat, a feed-forward model for 3D scenereconstruction that aims to address the longstanding trade-off betweenplausible view generation for unobserved regions and geometric consis-tency, providing both geometrically aligned novel views and sharp depthestimates. Our approach builds on 3D Gaussian splatting (3DGS) as anintermediate differentiable scene representation and integrates it with amulti-view latent diffusion model (MV-LDM) trained to act simultane-ously as a refiner and an inpainter for appearance and scene geometry.We enforce geometric consistency by guiding the diffusion process withvariational 3D latent features for appearance and geometry, encoded bythe feed-forward 3DGS representation and rasterized to 2D latent space.ReconSplat produces both photorealistic novel views and accurate depthmaps on real-world benchmarks, RealEstate10K and DL3DV-10K, out-performing existing methods in challenging extrapolation setups. No-tably, ReconSplat allows the extrapolation of unseen and challengingviewpoints jointly with coherent and precise scene geometry. Code andadditional visual results are available on our project page.

Synthesizing multi-pass videos is important for autonomousdriving. While current repair-based methods often struggle with out-of-distribution artifacts, camera-controlled methods often produce 3D-inconsistent results due to sparse LiDAR cues. We propose ReCamDriv-ing, a purely vision-based framework that achieves camera-controlledgeneration by leveraging dense, structurally complete 3DGS renderingsas geometric guidance. Specifically, to prevent the model from overfit-ting to a trivial repair solution when conditioning on 3DGS render-ings, we adopt a two-stage progressive training paradigm: the first stageuses camera poses for coarse control, while the second stage incorpo-rates 3DGS renderings for fine-grained viewpoint and geometric guid-ance. Furthermore, to align training and inference camera transforma-tion patterns, we propose a 3DGS-based cross-trajectory data curationstrategy, enabling consistent lateral-trajectory supervision from single-pass videos. Based on this strategy, we construct the ParaDrive dataset,containing approximately 110K parallel-trajectory video pairs. Exten-sive experiments demonstrate that ReCamDriving achieves state-of-the-art camera controllability and structural consistency. Project website:https://recamdriving.github.io/.

Interpretable machine learning aims to develop models whoseinternal reasoning is transparent and aligned with human perception.While prototype-based networks provide intuitive part-based interpreta-tions, most existing methods rely on flat or independently learned proto-type sets, failing to capture the hierarchical and relational organization ofvisual concepts. Inspired by theories of structured part-whole perception,we introduce ProtoMappingNet, a prototype-based architecture thatlearns hierarchically structured visual concepts through learnable cross-layer prototype mappings. These mappings align representations acrosstexture, part, and object levels, enabling coherent parent-child relationsbetween prototypes. To quantitatively evaluate the learned hierarchy,we propose a relational evaluation framework consisting of HierarchicalRelational Consistency (HRC) and Hierarchical Spatial Stability (HSS),which measure cross-level relational consistency and spatial containmentwithout requiring part annotations. Experiments show that ProtoMap-pingNet produces coherent multi-level explanations and substantially im-proves structural consistency through bidirectional relational modelingwhile maintaining competitive classification accuracy. Our code is avail-able at https://github.com/pk9403/ProtoMappingNet

LLM agents often fail in long-horizon tasks due to com-pounding simulation errors. We propose ProAct, a two-stage frameworkfor internalizing foresight. First, Grounded LookAhead Distillation (GLAD) performs supervised fine-tuning (SFT) on search-derived trajectories,compressing complex search trees into concise causal reasoning chains.This allows agents to learn lookahead logic without inference-time com-putational overhead. Second, we introduce Monte-Carlo Critic (MC-Critic), a plug-and-play auxiliary value estimator for policy-gradient al-gorithms (PPO/GRPO). By using lightweight rollouts to provide low-variance value signals, MC-Critic enables stable optimization without ex-pensive model-based approximations. Experiments on stochastic (2048)and deterministic (Sokoban) tasks show ProAct significantly boosts plan-ning accuracy. Notably, a 4B model trained with ProAct outperformsopen-source baselines and rivals state-of-the-art closed-source models,demonstrating strong generalization to unseen environments. The codesand models are available at https://github.com/GreatX3/ProAct.

It is intuitive for human vision system to perceive the objectposition, area and structure, which corresponds to the tasks of detection,segmentation, and pose estimation. Compared to the modern detectorsor segmentors, existing pose estimators only cover specific classes (e.g.,human body) and restricted structures. Thus, recent works gather previ-ous single-class pose datasets to construct multi-class pose datasets, butthe richness of classes and structures are still limited. To this end, we la-bel a new pose dataset upon ImageNet dataset, which contains extensiveclasses and structures. Due to the structure diversity in single class, wedecompose into two sub-tasks based on structure prototypes: prototypelabeling and pose annotating. In this way, we obtain 3500+ structureprototypes of 720 semantic classes. Although different prototypes mayhave diverse structures, there are some shared keypoints. Therefore, wepropose an appropriate baseline with query-based model and matching-based supervision, where a query will learn the shared information acrosskeypoints and focus on estimating the keypoints with similar character-istics. Extensive analyses and experiments indicate the quality of ourdataset and the effectiveness of our method.

Multimodal Large Language Models (MLLMs) have recentlyachieved remarkable success in visual-language understanding, demon-strating superior high-level semantic alignment within their vision en-coders. An important question thus arises: Can these encoders serve asversatile vision backbones, capable of reliably performing classic vision-centric tasks as well? To address the question, we make the following con-tributions: (i) We identify that the vision encoders within MLLMs exhibitdeficiencies in their dense feature representations, as evidenced by theirsuboptimal performance on dense prediction tasks (e.g., semantic seg-mentation, depth estimation); (ii) We propose VersaViT, a well-roundedvision transformer that instantiates a novel multi-task framework for col-laborative post-training. This framework facilitates the optimization ofthe vision backbone via lightweight task heads with multi-granularity su-pervision; (iii) Extensive experiments across various downstream tasksdemonstrate the effectiveness of our method, yielding a versatile visionbackbone suited for both language-mediated reasoning and pixel-levelunderstanding. The project page is available here.

Due to the extensive application of UAVs in modern war-fare, anti-UAV single object tracking has garnered increasing researchinterest. However, prevailing multi-modal methods heavily rely on theassumption of perfectly calibrated sensors, leaving the critical challengeof tracking under realistic spatio-temporal misalignment largely unad-dressed. To tackle this issue, we formally introduce the novel task ofuncalibrated RGBT anti-UAV tracking and mathematically formulateits observation process as a unified state-space model. Based on thistheoretical foundation, we propose OCTA-SOT, an Online Cross-modalTrajectory Adjustment Single Object Tracking framework. As a com-pletely training-free and plug-and-play module, it can be directly appliedto off-the-shelf base trackers. During inference, it adaptively selects reli-able trajectory information to update inter-modality mappings in realtime, while simultaneously calibrating unconfident trackers and theirresults. At its core, a customized Kalman-driven mechanism dynami-cally adjusts these mappings, thereby achieving robust target tracking.Extensive experiments on the challenging Anti-UAV300 dataset demon-strate the exceptional effectiveness of our method. Without requiring anytraining, OCTA-SOT effectively enhances the DIMP tracker, achievingabsolute improvements of 9.3% in AUC, 12.9% in Precision, and 11.9%in Normalized Precision, while simultaneously outperforming a series ofcurrent state-of-the-art methods. The source code is publicly availableat https://github.com/xkliu-eps/AntiUAV_RGBT_Tracking/.

Recent advances in 3D Gaussian Splatting have demonstratedunprecedented success in novel view synthesis. However, the substantialinference and storage overhead driven by high-order Spherical Harmon-ics (SH) are primary bottlenecks for mobile platforms. In this paper,we present Flux-GS, a real-time Gaussian Splatting method designedto achieve high-fidelity rendering with significantly reduced overhead forresource-constrained mobile platforms. We first propose a Monte CarloSpecular Energy Aggregator, sampling third-order radiance residualsand aggregating specular energy into a compact latent space. In thisway, our method effectively preserves visually salient lighting features inlower-order bands without expensive distillation or pre-training. To miti-gate the high-frequency details lost during compression, we introduce anAttribute-Conditioned SH Enhancement module. This module predictsGaussian-aware offsets based on intrinsic Gaussian attributes, which en-hance the first-order SH representation prior to inference, without extrainference costs. Furthermore, the original single-view gradient-based den-sification is prone to producing excessive Gaussians and overfitting to aa. Rotation Opacity Opacity Opacity b.Scale 7% 2% 4% 6%5%PositionRotation SH5%Rotation 35%18%23% 3DGS Third-orderScaleSH SH Rest52%13%SH ScalePosition Position81% 18%13% 18%61%↓ 26%↓ First-orderThird-order SH First-order SH Flux-GS Ground Truth Flux-GS SH RestFig. 2: Gaussian parameter distribution and Spherical Harmonic fidelityanalysis. a. Per Gaussian memory footprint across 3DGS variants. Flux-GS achievessignificant compression (61% and 26% reductions) by optimizing Spherical Harmonics(SH) coefficients and decoupling SH into the base and view-independent components.b. Qualitative comparison demonstrates that Flux-GS with only first-order SH canrender high-fidelity high-frequency details comparable to 3DGS.certain view. We address these limitations by proposing a Multi-viewAlpha-based Densification and Pruning strategy. By leveraging multi-view guidance, we ensure multi-view structure consistency and the pre-cise removal of redundant primitives. Extensive experiments demonstratethat Flux-GS achieves substantial parameter reduction while maintain-ing competitive visual quality, offering a robust and scalable solution forreal-time mobile rendering. Code: https://xiaobiaodu.github.io/flux-gs-project/.

Domain Adaptive Object Detection (DAOD) adapts detec-tors from labeled source domains to unlabeled target domains. Currentteacher-student methods rely solely on proposal-level pseudo-boxes assupervision, which suffer from substantial noise under domain shift. Wemake an intriguing empirical observation: image-level predictions, ob-tained by aggregating proposal scores, achieve significantly higher ac-curacy than proposal-level pseudo-boxes. This superiority stems fromaggregation’s inherent noise reduction and robustness to localization er-rors. Despite this finding, image-level predictions have been completelyoverlooked as a supervision source in DAOD, though extensively usedin weakly supervised detection. We propose Multi-Granularity Pseudo-Labeling (MGPL), which for the first time incorporates image-level su-pervision alongside proposal-level pseudo-boxes in teacher-student DAOD.Our framework supervises the student detector at both granularities:proposal-level boxes provide spatial precision while image-level predic-tions offer reliable category information. We further introduce coarse-to-fine collaborative filtering, where image-level predictions guide class-adaptive threshold selection for proposals, reducing false negatives whilemaintaining precision. Extensive experiments demonstrate state-of-the-art performance across multiple benchmarks. For example, on PascalVOC → Clipart1k, we achieve 51.4% mAP, surpassing previous best andfully-supervised Oracle by 2.3% and 6.4% respectively. Our work revealsthe untapped potential of multi-granularity learning in domain adapta-tion. Code will be made publicly available.

Web agents—autonomous systems that navigate and executetasks on the web on behalf of users—have the potential to transformhow people interact with the digital world. However, the most capableweb agents today rely on proprietary models with undisclosed trainingdata and recipes, limiting scientific understanding, reproducibility, andcommunity-driven progress. We believe agents for the open web shouldbe built in the open. To this end, we introduce (1) MolmoWebMix, alarge and diverse mixture of browser task demonstrations and web-GUIperception data and (2) MolmoWeb a family of fully open multimodalweb agents. Specifically, MolmoWebMix combines over 100K synthetictask trajectories from multiple complementary generation pipelines with30K+ human demonstrations, atomic web-skill trajectories, and GUIperception data, including referring expression grounding and screen-shot question answering. MolmoWeb agents operate as instruction-conditioned visual-language action policies: given a task instruction anda webpage screenshot, they predict the next browser action, requiring noaccess to HTML, accessibility trees, or specialized APIs. Available in 4Band 8B size, on browser-use benchmarks like WebVoyager, Online-Mind2Web, and DeepShop, MolmoWeb agents achieve state-of-the-art results outperforming similar scale open-weight-only models suchas Fara-7B, and Holo1-7B. MolmoWeb-8B also surpasses set-of-marks(SoM) agents built on much larger closed frontier models like GPT-4o.We further demonstrate consistent gains through test-time scaling viaparallel rollouts with best-of-N selection, achieving 94.7% and 60.5%pass@4 (compared to 78.2% and 35.3% pass@1) on WebVoyager andOnline-Mind2Web respectively. We release model checkpoints, train-ing data, code, and a unified evaluation harness to enable reproducibilityand accelerate open research on web agents (GitHub).

Synthetic data mitigates the data scarcity problem in au-tonomous driving perception. However, the synthetic-to-real gap leads toperformance degradation, hindering real-world model generalization. Al-though current methods leverage diffusion models for photorealistic styletransfer to bridge this gap, they critically ignore a practical asymmetry:while synthetic data possesses perfect pixel-level annotations, real-worldstyle reference images generally lack corresponding labels. Consequently,existing methods relying on symmetric semantic guidance suffer fromeither prohibitive annotation costs or severe semantic misalignment. Toaddress this dilemma, we formally propose a novel task: AsymmetricStyle Transfer for Autonomous Driving (ASTAD), which requires se-mantically consistent transfer using only labeled synthetic content andunlabeled real-world references. We further introduce the ASTModel, atraining-free two-stage framework designed to bridge this domain gapunder asymmetric constraints. ASTModel first extracts a coarse seman-tic prior from the unlabeled target, followed by dynamic prior refinementand class-consistent style injection during the denoising process. Exten-sive experiments demonstrate that ASTModel significantly outperformsexisting methods in downstream perception utility and structural fidelity,while offering a 3.2× inference speedup. This work aligns synthetic-to-real adaptation with practical constraints, holding the potential to ac-celerate the scalable deployment of robust autonomous driving systems.Code: https://github.com/Dingyi-Yao/ASTAD.

Learned image compression (LIC) is bottlenecked by theneed to store independent models for each rate-distortion operating point.Existing variable bit-rate (VBR) methods aim to reduce this overhead viadense parameter modulation, but forcing a shared backbone to approx-imate divergent mappings causes severe feature entanglement. Specifi-cally, low-rate smoothing gradients inherently conflict with the preser-vation of high-frequency textural details, leading to sub-optimal perfor-mance. To resolve this, we propose MixCompress, a unified VBR frame-work based on sparse structural specialization. While sparsely gatedMixture-of-Experts (MoE) routing successfully mitigates gradient con-flict, it operates on a fixed computational budget. To address the in-creased representational demands of higher bit-rates we introduce a Mixture-of-Depths (MoD) extension to dynamically scale model capacity. Com-bined with Conditional Auxiliary Transforms (CAT) for dynamic sub-band energy modulation, our hierarchical framework effectively dynam-ically scales capacity. Extensive evaluations demonstrate that MixCom-press not only matches individually optimized single-rate baselines butcan even surpass them, establishing a new Pareto frontier for computa-tionally efficient image coding.

Metric reasoning is a critical and challenging task for VisionLanguage Models (VLMs), playing a pivotal role in embodied AI taskssuch as robotic manipulation and autonomous navigation. However, cur-rent spatial reasoning remains bottlenecked by rigid pixel-level supervi-sion; such localized optimization often compromises general multimodalintelligence, triggering performance degradation or catastrophic forget-ting of broad reasoning capabilities. To address these limitations, weintroduce Metric-Bench, a focused benchmark designed to guide metric-spatial reasoning using contextual information. By incorporating in-imagereference objects with known physical dimensions, Metric-Bench guidesmodels to implicitly learn the 2D-to-3D mapping without camera in-trinsics. We further present MetricReasoner, a task-adapted reinforce-ment fine-tuning recipe for reference-grounded metric reasoning, usingstructured prompts and verifiable numerical rewards. Extensive experi-ments on Metric-Bench demonstrate that our approach significantly en-hances spatial metric understanding, outperforming existing and evenlarger proprietary models by 43.1%, while improving downstream em-bodied performance over a spatial-specialized counterpart by 30.4% onRoboSpatial overall accuracy and 9.3% on ERQA, and additionally deliv-ering consistent gains on general benchmarks (15.9% on V⋆Bench, 88.9%on BLINK), indicating that the proposed adaptation does not necessarilycompromise general VLM capabilities.

Medical report understanding from real-world document im-ages is essential for generating patient-facing explanations and enablingstructured information exchange in clinical systems.Existing VLMs and LLMs have shown strong performance on documentunderstanding, but structured understanding of medical reports remainsinsufficiently benchmarked. Therefore, we introduce MedRepBench, abenchmark with 1,925 de-identified Chinese medical report images span-ning diverse departments, patient demographics, and acquisition formats.In MedRepBench, we mainly focus on report-grounded interpretationrather than evaluating diagnostic reasoning, treatment recommendation,or the integration of patient history. The interpretation is defined asstructured extraction of report fields (e.g., item, value, unit, referencerange, abnormal flag) plus a patient-facing explanation grounded strictlyin the report content.The benchmark primarily evaluates end-to-end VLMs, and also includesa controlled text-only setting (high-quality OCR + LLM) to approximatean upper bound when character recognition errors are minimized. Ourevaluation framework provides two complementary protocols: (1) an ob-jective protocol measuring field-level recall of structured items, and (2)an automated subjective protocol that uses an LLM-based judge to scorefactuality, interpretability, and reasoning quality under a fixed prompt.Using the objective metric as a reward signal, we also provide a lightweightGRPO-based alignment baseline for a mid-sized VLM, which improvesfield-level recall by up to 6%.Finally, we analyze practical limitations of OCR+LLM pipelines, includ-ing layout-related errors and additional system latency, showing the needfor robust end-to-end vision-based medical report understanding. Thedataset and evaluation resources are publicly available on HuggingFace.

Human-object interaction (HOI) video generation has gar-nered increasing attention due to its promising applications in digitalhumans, e-commerce, advertising, and robotics imitation learning. How-ever, existing methods face two critical limitations: (1) a lack of effec-tive mechanisms to inject multi-view information of the object into themodel, leading to poor cross-view consistency, and (2) heavy relianceon fine-grained annotations including hand mesh and body templatesfor modeling interaction occlusions, due to limited training data. To ad-dress these challenges, we introduce ByteLOOM, a Diffusion Transformer(DiT)-based framework that generates realistic human-object interactionvideo with geometrically consistent object illustration, using simplifiedhuman conditioning and 3D object inputs. We first propose an RCM-cache mechanism that leverages Relative Coordinate Maps (RCM) as auniversal representation to maintain object’s geometry consistency andprecisely control 6-DoF object transformations in the meantime. To com-pensate HOI dataset scarcity and leverage existing datasets, we furtherdesign a training curriculum that enhances model capabilities in a pro-gressive style and relaxes the demand of hand mesh. Extensive experi-ments demonstrate that our method faithfully preserves human identityand the object’s multi-view geometry, while maintaining smooth motionand object manipulation. Project page: neutrinoliu.github.io/byteloom/