ECCV 2026 论文列表

Recent advances in sparse voxel representations have significantly improved the quality of 3D content generation, enabling highresolution modeling with fine-grained geometry. However, existing frameworks suffer from severe computational inefficiencies due to the quadratic complexity of attention mechanisms in their two-stage diffusion pipelines. In this work, we propose Ultra3D, an efficient 3D generation framework that significantly accelerates sparse voxel modeling without compromising quality. Our method leverages the compact VecSet representation to efficiently generate a coarse object layout in the first stage, reducing token count and accelerating voxel coordinate prediction. To refine pervoxel latent features in the second stage, we introduce Part Attention, a geometry-aware localized attention mechanism that restricts attention computation within semantically consistent part regions. This design preserves structural continuity while avoiding unnecessary global attention, achieving up to a 6.7× speed-up in latent generation. To support this mechanism, we construct a scalable part annotation pipeline that converts raw meshes into part-labeled sparse voxels. Extensive experiments demonstrate that Ultra3D supports high-resolution 3D generation at 1024 resolution and achieves state-of-the-art performance in both visual fidelity and user preference.

Trained policies for real-world robotics rely on computer vi-sion components, typically in the form of pre-trained visual encoders.These encoders are an essential component and it has been shown thattheir power does not emerge from training on robotics downstream lossesalone. Pre-training with auxiliary losses in the form of computer-visionpre-text tasks is a defining factor and heavily conditions agent perfor-mance in robotics tasks. In this unprecedented large-scale study, we ran966 navigation episodes of static point goal navigation in a real-worldbuilding for 24km and asked which components really matter for thecomputer vision aspects of robotics: we evaluate state-of-the art visualencoders in realistic conditions. We explore the usefulness of heteroge-neous multi-teacher distillation leading to encoders with multiple dif-ferent and complementary skills. We investigate how much informationfrom these encoders is necessary for robotics by bottlenecking them in aprincipled and “spatially useful” way and we show that this leads to theemergence of interpretable features linked to affordances. We also arguethat training policies on RGB data alone does not lead to an optimalusage of visual features and show this by finetuning policies pre-trainedon privileged information. All in all, we paint a more complete picture ofwhat aspects of computer vision are relevant for real-world navigation.

Vision-Language Models (VLMs) exhibit strong visual rea-soning capabilities, yet they still struggle with 3D understanding. Inparticular, VLMs often fail to infer a text-consistent goal 6D pose of atarget object in a 3D scene. However, we find that with some inference-time techniques and iterative reasoning, VLMs can achieve dramatic per-formance gains. Concretely, given a 3D scene represented by an RGB-Dimage (or a compositional scene of 3D meshes) and a text instructionspecifying a desired state change, we repeat the following loop: observethe current scene; evaluate whether it is faithful to the instruction; pro-pose a pose update for the target object; apply the update; and ren-der the updated scene. Through this closed-loop interaction, the VLMeffectively acts as an agent. We further introduce three inference-timetechniques that are essential to this closed-loop process: (i) multi-viewreasoning with supporting view selection, (ii) object-centered coordinatesystem visualization, and (iii) single-axis rotation prediction. Withoutany additional fine-tuning or new modules, our approach surpasses priormethods at predicting the text-guided goal 6D pose of the target object.It works consistently across both closed-source and open-source VLMs.Moreover, when combining our 6D pose prediction with simple robot mo-tion planning, it enables more successful robot manipulation than recentVision-Language-Action models (VLAs). Finally, we conduct an ablationstudy to demonstrate the necessity of each proposed technique.

LiDAR-based Human Mesh Reconstruction (HMR) is crucialfor understanding human activities in real-world environments. How-ever, the inherent sparsity of LiDAR point clouds often causes severeloss of anatomical details, limiting reconstruction accuracy. Althoughrecent methods have made promising progress, most still follow a decou-pled pipeline that first estimates skeletal joints and then reconstructsthe mesh, leading to error propagation under sparse observations. Toaddress this issue and fully exploit both the topological guidance of theskeleton on the mesh and the spatial constraints of the surface mesh onthe skeleton, we propose SynHMR, a Synergistic Joint-Mesh Modelingframework for robust LiDAR-based HMR. Specifically, we first initializeboth the skeletal joints and the mesh vertices as queries equipped withpositional information for co-optimization. We then design a Geometry-Aware Transformer that leverages point cloud features to apply the at-tention mechanism between joint and vertex queries, enabling their it-erative refinement through mutual information exchange. Furthermore,we introduce a Noise-Augmented Learning strategy that injects per-turbed joint queries during training. This reduces the impact of noisewhile encouraging bidirectional association between joints and mesh. Ex-tensive experiments on LiDARHuman26M, SLOPER4D, and Human-M3 demonstrate that SynHMR achieves state-of-the-art performanceamong LiDAR-based HMR approaches. The project page is availableat https://github.com/ShiRui1208/SynHMR.

Contrastive vision–language models (VLMs) such as CLIP,EVACLIP, ImageBind, and SigLIP exhibit impressive zero-shot general-ization, yet their performance on specific data distributions can furtherbe enhanced by improving the synchronization of the vision and languagerepresentations. For this purpose, we present SyncVL, a bi-directionalVL synchronization framework that enables fully unsupervised adapta-tion of VLMs. SyncVL jointly refines visual and textual representationsthrough two lightweight transformer-based encoder-decoder synchroniz-ers (Vision-to-Text and Text-to-Vision) that progressively project VLrepresentation into a shared latent subspace without requiring labeleddata. To ensure robust synchronization, we form groups using multipleimage-views and text templates. Both synchronizers are trained using areward function motivated by group relative policy optimization whilekeeping the original VLMs frozen. Additionally, a mutual-distillationmechanism is proposed to iteratively reinforce each synchronizer. Weextensively evaluate SyncVL on 24 benchmark datasets spanning unsu-pervised VL adaptation, out-of-distribution matching, unsupervised clus-tering, cross-modal retrieval, object detection, and segmentation. Acrossall tasks, SyncVL consistently improves over state-of-the-art baselines,establishing a new standard for unsupervised adaptation of VLMs.

Autonomous agents require representations that are gen-eral, causal, and physically structured to operate in real-time stream-ing environments. However, current vision foundation models remainfragmented, specializing narrowly in image semantic perception, offlinetemporal modeling, or spatial geometry. This paper introduces Om-niStream, a unified streaming visual backbone that effectively per-ceives, reconstructs, and acts from diverse visual inputs. By incorporat-ing causal spatiotemporal attention and 3D rotary positional embeddings(3D-RoPE), our model supports efficient, frame-by-frame online process-ing of video streams via a persistent KV-cache. We pre-train OmniStreamusing a synergistic multi-task framework coupling static and temporalrepresentation learning, streaming geometric reconstruction, and vision-language alignment on 29 datasets. Extensive evaluations show that, evenwith a strictly frozen backbone, OmniStream achieves consistently com-petitive performance with specialized experts across image and videoprobing, streaming geometric reconstruction, complex video and spatialreasoning, and robotic manipulation. Rather than pursuing benchmark-specific dominance, our work demonstrates the viability of training a sin-gle, versatile vision backbone that generalizes across semantic, spatial,and temporal reasoning, i.e., a meaningful step towards general-purposevisual understanding for interactive and embodied agents.

Vision-Language-Action (VLA) models excel in static manipulation but struggle in dynamic environments with moving targets. This performance gap primarily stems from a scarcity of dynamic manipulation datasets and the reliance of mainstream VLAs on single-frame observations, restricting their spatiotemporal reasoning capabilities. To address this, we introduce DOMINO, a large-scale dataset and benchmark for generalizable dynamic manipulation, featuring 35 tasks with hierarchical complexities, over 110K expert trajectories, and a multidimensional evaluation suite. Through comprehensive experiments, we systematically evaluate existing VLAs on dynamic tasks, explore effective training strategies for dynamic awareness, and validate the generalizability of dynamic data. Furthermore, we propose PUMA, a dynamicsaware VLA architecture. By integrating scene-centric historical optical flow and specialized world queries to implicitly forecast object-centric future states, PUMA couples history-aware perception with short-horizon prediction. Results demonstrate that PUMA achieves state-of-the-art performance, yielding a 6.3% absolute improvement in success rate over baselines. Moreover, we show that training on dynamic data fosters robust spatiotemporal representations that transfer to static tasks.

High Dynamic Range (HDR) imaging is indispensable forautonomous driving to ensure robust perception under changing light-ing conditions, such as tunnel exits or night driving. However, exist-ing learning-based HDR methods suffer from excessive computationalcosts and latency, making them impractical for real-time edge deploy-ment. In this paper, we propose SplitHDR, an extremely lightweightHDR framework tailored for machine vision. Leveraging advanced sen-sor technologies designed to minimize temporal displacement, our ap-proach avoids relying on heavy alignment modules and introduces aluminance-aware dynamic inference mechanism. Specifically, the modelconstructs a saturation mask based on input intensity, efficiently routingsaturated regions to a multi-frame fusion mode and unsaturated regionsto a lightweight denoising mode. This task decomposition ensures thatcomputational resources are allocated only where necessary. Experimen-tal results demonstrate that our method achieves state-of-the-art effi-ciency with only 0.002M parameters and 3.63 GMACs, reducing the pa-rameter count by over 99% compared to existing multi-exposure models.Despite the extreme compression, the HDR images reconstructed by ourapproach effectively preserve critical structural details, yielding highlycompetitive object detection performance crucial for safety-critical au-tonomous systems.

Autonomous driving world models are expected to work effectively across three core dimensions: state, action, and reward. However, existing methods are typically restricted to fragmented modality modeling, short-horizon drift, and imprecise action control, while lacking intrinsic mechanisms for policy evaluation. In this paper, we introduce OmniNWM, an Omniscient panoramic Navigation World Model that addresses all three dimensions within a consistent probabilistic framework. For State, OmniNWM generates panoramic videos of RGB, semantics, metric depth, and 3D occupancy, ensuring pixel-level alignment across modalities with joint distribution modeling. To mitigate autoregressive exposure bias, we propose a structured panoramic forcing strategy to stabilize long-horizon generation via stochastic manifold thickening. For Action, we introduce canonical geometric action encoding with normalized panoramic Plücker ray-maps. This representation decouples motion dynamics from sensor intrinsics, enabling precise, zero-shot trajectory control across heterogeneous datasets and camera configurations. For Reward, we derive intrinsic occupancy-grounded dense rewards directly from generated 3D volumes, establishing a reliable closed-loop simulation cycle for evaluating diverse planning agents. Extensive experiments demonstrate that OmniNWM achieves SOTA performance in generation fidelity and control precision, with remarkable zero-shot robustness to novel scenes on nuPlan and in-house datasets with distinct camera rigs.

Open-Vocabulary Multi-Object Tracking (OVMOT) repre-sents a critical challenge requiring the detection and tracking of diverseobject categories beyond predefined training classes, including both seen(base) and unseen (novel) categories in real-world videos. Existing ap-proaches are fundamentally limited by their reliance on traditional closed-set tracking paradigms trained only on base classes and synthetic imagepairs, failing to exploit the rich temporal dynamics inherent in videosequences for universal tracking capabilities. We present SOVTrack, anovel self-supervised framework that harnesses SAM2’s universal represen-tations to address general-purpose tracking from raw video data withoutmanual supervision. Our approach introduces Dual-Direction PseudoLabeling (DDPL) to automatically generate high-quality pseudo-labelsthrough bidirectional temporal validation while dynamically assessingthe difficulty of each training example, and employs a dual-branch archi-tecture with Multi-Cue Adversarial Distillation (MCAD) to effectivelytransfer SAM2’s multi-cue pixel-level universal tracking capabilities toobject-level tracking scenarios while preserving its generalization poweracross diverse object categories. Experimental results demonstrate thatSOVTrack achieves state-of-the-art performance using only 10K videoframes (2% of existing methods’ data requirements), delivering substantialimprovements in both base and novel category tracking performance.

We propose SMP-UWGS, an underwater reconstruction frame-work based on 3D Gaussian Splatting (3DGS) that couples scalablemulti-partition geometry with physical light transport modeling. Ourmethod introduces two key components: SMP-GAUSSIAN, a multi-partitionarchitecture enabling distributed optimization for large-scale scenes, andDPR-Net (Differentiable Physical Rendering Network), which combinesWaterParamPredict module for estimating water optical parameters witha Dual-Branch Differential Refinement (DBDR) module to model atten-uation and backscattering. With tailored loss functions, the frameworkjointly estimates water optical parameters and improves color fidelityand geometric accuracy. Experiments on public datasets show that SMP-UWGS achieves state-of-the-art efficiency while enabling scalable, high-fidelity underwater reconstruction, benefiting applications such as ecolog-ical monitoring, habitat mapping, and autonomous underwater robotics.

Existing vision-and-language navigation (VLN) models pri-marily reason over past and current visual observations, while largelyignoring the future visual dynamics induced by actions. As a result,they often lack an effective understanding of the causal relationship be-tween actions and how the visual world changes, limiting robust decision-making. Humans, in contrast, can “imagine” the near future by leveragingaction–dynamics causality, which improves both environmental under-standing and navigation choices. Inspired by this capability, we proposeLatentPilot, a new paradigm that exploits future observations duringtraining as a valuable data source to learn action-conditioned visual dy-namics, while requiring no access to future frames at inference. Con-cretely, we propose a flywheel-style training mechanism that iterativelycollects on-policy trajectories and retrains the model to better match theagent’s behavior distribution, with an expert takeover triggered whenthe agent deviates excessively. LatentPilot further learns visual latenttokens without explicit supervision; these latent tokens attend globallyin a continuous latent space and are carried across steps, serving asboth the current output and the next input, which enabling the agentto “dream ahead” and reason about how actions will affect subsequentobservations. Experiments on R2R-CE, RxR-CE, and R2R-PE bench-marks achieve new SOTA results, and real-robot tests across diverseenvironments demonstrate LatentPilot’s superior understanding of envi-ronment–action dynamics in scene.

Recent advances in Neural Radiance Fields and 3D Gaus-sian Splatting have demonstrated immense potential for large-scale UAV-based 3D reconstruction. However, real-world data collection often spansmultiple times of the day, violating the fundamental photometric consis-tency assumption. Such multi-temporal illumination inconsistencies in-evitably lead to severe color artifacts, coupled shadows, and degradedgeometric accuracy. Due to the lack of UAV datasets that systematicallycapture the same areas under varying lighting conditions, this crucialchallenge remains largely underexplored. To bridge this gap, we introduceSkyLume, a large-scale, real-world UAV dataset specifically designed forstudying illumination-robust 3D reconstruction and urban scene model-ing. Our dataset features three primary contributions: (1) We collect over100k high-resolution UAV images (nadir and four oblique views) span-ning 10 diverse urban regions, with each region meticulously capturedacross three distinct time periods to systematically isolate illuminationvariations. (2) To enable rigorous evaluation, we provide comprehensiveper-scene LiDAR scans, offering highly accurate ground truth for as-sessing depth, surface normals, and geometric fidelity. (3) For inverserendering and appearance decoupling tasks, we introduce the Tempo-ral Consistency Coefficient (TCC), a novel metric designed to quantifycross-time rendering stability. We envision SkyLume as a foundationalbenchmark that will advance research and real-world evaluation in large-scale inverse rendering, robust geometry reconstruction, and novel viewsynthesis. Project Page: skylume.

Monocular 3D lane detection offers a cost-effective solutionfor autonomous driving perception but is inherently challenged by theill-posed problem of inferring 3D geometry from a 2D image. Standard-Definition (SD) maps offer a lightweight prior for road topology, promis-ing enhanced accuracy for distant or occluded lanes. However, their in-herent inaccuracies and representation as road-level centerlines, whichmisalign with the actual lanes to be perceived, pose significant challengesfor direct integration. To tackle this, we propose SGC-Lane, a novelframework that advances monocular 3D lane detection through SD mapguidance and lane completion. Specifically, we propose an FV-GuidedMap Encoder that generates a lane probability map and structured mapfeatures, where the feature channels are aligned with potential lane in-stances. This probability map directs attention toward lane-specific re-gions in the front view (FV), refining image features and initializing lanequeries by the Local Spatial Cross-Attention module for precise geome-try decoding. In addition, to mitigate short-range truncation, we designa Step Consistency Completion (SCC) Head that is used to predict andcluster lane endpoints, effectively recovering missing segments to producecomplete 3D lane representations. On the large-scale OpenLane dataset,our method achieves a 65.6% F1-score, showing improved performanceagainst existing approaches with the assistance of only coarse road-levelSD map priors. The improvements are especially evident in challengingconditions, such as roads with sharp curves and complex intersections.Code is released at https://github.com/FuqingJIang/SGC-Lane.

Clinical cardiac MRI spans diverse contrasts, sampling tra-jectories, accelerations, scanners, and patient populations, yet many de-terministic deep-learning reconstruction methods remain protocol spe-cific. We present the Scalable Deep Unrolled Model (SDUM), which in-tegrates a Restormer-based unrolled reconstructor, per-cascade coil sen-sitivity estimation, sampling aware weighted data consistency, and uni-versal conditioning on cascade index and acquisition metadata. A sin-gle SDUM model achieves state-of-the-art performance across all CM-RxRecon2025 tracks without task-specific fine-tuning and outperformsPromptMR+ on CMRxRecon2024 by +0.55 dB. Scaling experimentsshow near-logarithmic gains with depth up to 18 cascades (r=0.986,R2 =0.973) and continued but diminishing gains from data scaling (32.72 dBat 40% to 33.18 dB at 100%). SDUM also generalizes in a zero-shot set-ting to unseen in-house chemical exchange saturation transfer (CEST)MRI (43.57 dB PSNR, 0.9769 SSIM). When trained separately on fastMRIbrain, SDUM surpasses PC-RNN by +1.8 dB. These results supportSDUM as a scalable framework for robust MRI reconstruction acrossheterogeneous acquisition settings beyond cardiac MRI.Code: https://github.com/NVIDIA-Medtech/NV-Raw2insights-MRIModel: https://huggingface.co/nvidia/NV-Raw2insights-MRI

While Text-to-Image (T2I) models have shown remarkablesuccess in generating photorealistic visual content, they still strugglewith the rigorous semantic alignment and logical reasoning requiredfor scientific imagery. Inspired by Peirce’s Semiotic Triad, we introduceScientific Image Reasoning (SciIR), a comprehensive resource for trainingand evaluation of scientific image generation. We formalize scientificreasoning into three core dimensions: Entity Structure (Icon), ScientificProcess (Index ), and Scientific Law (Symbol ). Specifically, to overcomethe scarcity of training data in scientific image generation, we elaboratelycreate SciIR-82k, a large-scale dataset containing over 80,000 high-qualityscientific image-text pairs from cutting-edge publications. The datasetis hierarchically organized according to the semiotic dimensions andincorporates a Scientific Reasoning Chain-of-Thought (Sci-RCoT) toexplicitly model underlying visual logic. For evaluation, we propose SciIR-Bench, which aligns with these three semiotic levels and employs anAtomic Checklist to convert the outcome-oriented scientific accuracyinto process-oriented, verifiable, fine-grained questions. Our extensiveexperiments reveal significant deficiencies in current models’ scientificreasoning capabilities. Furthermore, by fine-tuning on the SciIR-82kdataset, we developed the Qwen-Image-SciIR model, which achieves asubstantial improvement on the SciIR-Bench, increasing the final scorefrom 35% to 43%, laying a solid foundation for future advances in scientificimage generation.

Backdoor attacks, which induce Convolutional Neural Net-works (CNNs) to behave maliciously when a predefined trigger is present,pose serious security risks. While such threats also extend to VisionTransformers (ViTs), previous studies have suggested that existing back-door attacks remain highly effective on ViTs and can evade commondefense mechanisms—often without significantly compromising accuracy.In this paper, we revisit this claim and demonstrate that such con-clusions are overly optimistic, largely due to inadequate adaptation ofCNN-based defenses to ViTs. We show that, with proper adjustments,existing backdoor attacks can in fact be effectively mitigated. More-over, we propose a more robust attack strategy: by introducing slightperturbations to the trigger, existing attacks can be made significantlymore resistant to various defenses. We hope that our findings—bothon the defensibility of current attacks with correct adaptations and theproposed enhanced attack—will inspire deeper investigation into thebackdoor robustness of Vision Transformers. Our code is available athttps://github.com/PKU-ML/ViT_backdoor.

Point clouds are one of the most fundamental and widelyused 3D representations, serving as the most basic geometric represen-tation of 3D shapes. Nevertheless, most existing 3D printing pipelinesrequire a watertight mesh as input, preventing the direct use of pointclouds for fabrication. A common workaround is to reconstruct meshesfrom point clouds; however, the resulting meshes often contain geomet-ric artifacts, such as incorrect faces or topological inconsistencies, thatare difficult to repair and may lead to printing failures. To overcomethese limitations, we propose PrintAnything, a novel framework thatlearns to produce executable 3D printing G-code directly from 3D pointclouds without requiring mesh reconstruction. To enable point cloudsto serve as direct input for slice-wise toolpath generation, we introducea slice-wise point projection strategy that transforms unstructured 3Dpoint clouds into slice-aligned 2D representations consistent with layer-by-layer nature of fused deposition modeling in 3D printing. To eliminatemesh dependency and provide a unified representation that bridges pointclouds and G-code, we propose Geometric plan (G-plan) map, a com-pact 2D representation composed of occupancy, region, and flow mapsthat encode the geometric and extrusion properties required for toolpathsynthesis in 3D printing. As a result, our proposed method accuratelygenerates printable G-code directly from point clouds, enabling a prac-tical and fully mesh-free pipeline for 3D printing. The code is publiclyavailable at https://github.com/Sangminhong/PrintAnything.

Real-world image dehazing remains a challenging problem due to the significant domain gap between synthetic training data and natural haze, often leading to poor generalization. While unsupervised Generative Adversarial Networks (GANs) attempt to bridge this gap, they frequently suffer from training instability and hallucinatory artifacts. Recently, Reinforcement Learning (RL) has emerged as a promising alternative; however, current RL-based restoration paradigms predominantly rely on diffusion models, leading to prohibitive computational costs and ineffective exploration. To address these bottlenecks, we propose Dehaze-RL, an efficient framework tailored for unpaired real-world dehazing. Bypassing the expensive iterative sampling of diffusion models, we design an efficient policy network to predict hybrid actions in a Physics Hybrid Action Space. By explicitly estimating physical parameters, enhancement factors, and a gating policy map, we anchor the agent’s exploration in reliable physical priors, effectively resolving the exploration dilemma. These components are dynamically integrated via a gating fusion mechanism. Furthermore, we introduce a Multi-Granularity Reward Mechanism to provide comprehensive feedback, seamlessly aligning the outputs with human perception without structural distortions. Extensive experiments demonstrate that Dehaze-RL outperforms stateof-the-art methods in both visual fidelity and quantitative metrics.