ECCV 2026 论文列表

Radar sensors provide reliable perception under adverse weatherand lighting conditions, but their sparse, noisy, and weakly semanticmeasurements make dense semantic segmentation challenging. Most ex-isting radar segmentation methods rely on grid-based encodings andpairwise interactions, which struggle to capture the higher-order rela-tional structure formed by multiple radar returns from the same phys-ical object. We introduce a unified higher-order structural alignmentframework for multi-view radar segmentation. The proposed method re-fines radar feature representations using learnable hypergraphs to cap-ture higher-order dependencies among spatially related responses. To en-sure consistency across heterogeneous radar projections, we further alignview-specific features using Unbalanced Optimal Transport (UOT), en-abling correspondence-free alignment under varying measurement den-sities and partial observations. An adaptive attention mechanism thenfuses complementary radar views while emphasising structurally informa-tive responses under sparsity and noise. The resulting architecture learnsstructurally consistent representations across Range Angle (RA), RangeDoppler (RD), and Angle Doppler (AD) views and is trained using super-vised segmentation together with cross-view consistency regularisation.Experiments on the CARRADA and RADIal benchmarks demonstrateconsistent improvements over strong radar-specific baselines, achieving63.8% mIoU on CARRADA and 83.4% mIoU on RADIal, im-proving the previous best methods by +1.7 and +2.3 mIoU, respec-tively. These results highlight the importance of higher-order relationalmodelling for robust radar perception.

Current expressive avatar systems rely heavily on visual cuesand often fail when faces are occluded or emotions remain internal. Wepresent Mind-to-Face, the first framework to decode non-invasive elec-troencephalogram (EEG) signals directly into high-fidelity facial expres-sions. We build a dual-modality recording setup that captures synchro-nized EEG and multi-view facial video during emotion-eliciting stimuli,providing precise supervision for neural-to-visual learning. Our modeluses a CNN-Transformer encoder to map EEG signals into dense 3Dposition maps that sample over 65k vertices, capturing fine-scale geom-etry and subtle emotional dynamics, and renders them through a mod-ified 3D Gaussian Splatting pipeline for photorealistic, view-consistentresults. Extensive evaluations show that EEG alone can reliably predictdynamic, subject-specific facial expressions, including subtle emotionalresponses, demonstrating that neural signals contain far richer affectiveand geometric information than previously assumed. Mind-to-Face es-tablishes a new paradigm for neural-driven avatars, enabling personal-ized, emotion-aware telepresence and cognitive interaction in immersive

Visual Prompting (VP) has emerged as a parameter-efficient paradigm for adapting pre-trained models to downstream tasks. While existing VP methods predominantly operate in the input space, extending VP to deep activations confronts a fundamental Capacity-Integrity Paradox: increasing prompt capacity for better task adaptation inevitably triggers a parameter explosion and severe semantic disruption, where deep prompts override pre-trained representations. To strike a delicate tradeoff within this Capacity-Integrity Paradox, we conducted a series of preliminary experiments. Interestingly, these experiments reveal a strong correlation among prompt channels: applying shared prompts across grouped channels not only drastically reduces parameters but also improves performance. Motivated by this structural correlation and the need to systematically master the tradeoff, we propose Low-Rank Tensor Visual Prompting (LoTeP). By modeling deep prompts as low-rank tensors, LoTeP elegantly achieves extreme parameter compression. Concurrently, it enables precise control over the prompt capacity via layerwise rank decaying, effectively preserving the semantic integrity of deep activations. Extensive experiments demonstrate the superior generality, effectiveness, and efficiency of our approach. Overall, LoTeP consistently outperforms the state-of-the-art LoR-VP by an average margin of over 2.5% across all evaluated scenarios, while adding less than 0.1% of the backbone parameters over the LoR-VP baseline.

A major challenge in autonomous driving is the “long tail” ofsafety-critical edge cases, which often emerge from unusual combinationsof common traffic elements. Synthesizing these scenarios is crucial, yetcurrent controllable generative models provide incomplete or entangledguidance, preventing the independent manipulation of scene structure,object identity, and ego actions. We introduce CompoSIA, a composi-tional driving video simulator that disentangles these traffic factors, en-abling fine-grained control over diverse adversarial driving scenarios. Tosupport controllable identity replacement of scene elements, we proposea noise-level identity injection, allowing pose-agnostic identity generationacross diverse element poses, all from a single reference image. Further-more, a hierarchical dual-branch action control mechanism is introducedto improve action controllability. Such disentangled control enables ad-versarial scenario synthesis—systematically combining safe elements intodangerous configurations that entangled generators cannot produce. Ex-tensive comparisons demonstrate superior controllable generation qualityover state-of-the-art baselines, with a 17% improvement in FVD for iden-tity editing and reductions of 30% and 47% in rotation and translationerrors for action control. Furthermore, downstream stress-testing revealssubstantial planner failures: across editing modalities, the average colli-sion rate of 3s increases by 173%.

World models that forecast scene evolution by generatingfuture video frames devote the bulk of their capacity to photometricdetails, yet the resulting predictions often remain geometrically inconsis-tent. We present VGGT-World, a geometry world model that side-stepsvideo generation entirely and instead forecasts the temporal evolution offrozen geometry-foundation-model (GFM) features. Concretely, we re-purpose the latent tokens of a frozen VGGT as the world state and traina lightweight temporal flow transformer to autoregressively predict theirfuture trajectory. Two technical challenges arise in this high-dimensional(d=1024) feature space: (i) standard velocity-prediction flow matchingcollapses, and (ii) autoregressive rollout suffers from compounding ex-posure bias. We address the first with a clean-target (z-prediction) pa-rameterization that yields a substantially higher signal-to-noise ratio,and the second with a two-stage latent flow-forcing curriculum that pro-gressively conditions the model on its own partially denoised rollouts.Experiments on KITTI, Cityscapes, and TartanAir demonstrate thatVGGT-World significantly outperforms the strongest baselines in depthforecasting while running 3.6-5× faster with only 0.43B trainable pa-rameters, establishing frozen GFM features as an effective and efficientpredictive state for 3D world modeling. The code will be open-source athttps://github.com/SimonSun0810/VGGT-World.

Low-light image enhancement is an essential computer visiontask to improve image contrast and to decrease the effects of colorbias and noise. Many existing interpretable deep-learning algorithmsexploit the Retinex theory as the basis of model design. However,previous Retinex-based algorithms, that consider reflected objects asideal Lambertian ignore specular reflection in the modeling process andconstruct the physical constraints in image space, limiting generalizationof the model. To address this issue, we preserve the specular reflectioncoefficient and reformulate the original physical constraints in theimaging process based on the Kubelka-Munk theory, thereby constructingconstraint relationship between illumination, reflection, and detection, theso-called triple physical constraints (TPCs) theory. Based on this theory,the physical constraints are constructed in the feature space of the modelto obtain the TPC network (TPCNet). Comprehensive quantitativeand qualitative benchmark and ablation experiments confirm that theseconstraints effectively improve the performance metrics and visual qualitywithout introducing new parameters, and demonstrate that our TPCNetoutperforms other state-of-the-art methods on 10 datasets. The code isavailable at https://github.com/2020shijingyi/TPCNet

Rapid progress in embodied AI is bringing robots closer to performing complex tasks in real-world environments. In these settings, robots must operate over long time horizons, making decisions based on experiences collected over hours or even days. Vision-language models (VLMs) have recently shown strong capabilities in planning and control, making them promising high-level controllers for embodied agents. However, current VLMs can process only a limited number of images at once, highlighting the need for more efficient mechanisms to manage long-term memory in embodied contexts. To meaningfully evaluate these models for long-horizon control, a benchmark must target scenarios where memory is essential. Existing long-video QA benchmarks neglect embodied challenges like object manipulation and navigation, which require lowlevel skills and fine-grained reasoning over past interactions. Moreover, effective memory integration in embodied agents involves both recalling relevant historical information and executing actions based on that information, making it essential to study these aspects together. In this work, we introduce FindingDory, a new benchmark for long-range embodied tasks in the Habitat simulator. FindingDory evaluates memorycentric capabilities across 60 tasks requiring sustained engagement and contextual awareness in an environment. The tasks can also be procedurally extended to longer and more challenging versions, enabling scalable evaluation of memory and reasoning. We further present baselines that integrate state-of-the-art closed-source and fine-tuned open-source VLMs with low-level navigation policies, assessing their performance on these memory-intensive tasks and highlighting key areas for improvement.

Non-line-of-sight (NLOS) imaging aims to reconstruct hid-den scenes from time-resolved light transport, enabling vision around cor-ners. While classical physics-based methods provide principled inversion,learning-based approaches often rely on dense 3D backbones with cubiccomputational complexity. We propose a hybrid deep learning frame-work that combines a physics-guided initialization with a triplane-basedbackbone for high-fidelity volumetric reconstruction. The initialization isprovided by a learnable Enhanced Light-Cone Transform (ELCT), whichproduces a stable physics-consistent coarse volume, while the learnedbackbone replaces expensive O(N 3 ) 3D processing with scalable O(N 2 )triplane feature extraction. ConvNeXt-style residual convolutions, Restormerattention, and axis-aware cross-attention jointly refine structure and re-cover missing geometry. Experiments on synthetic and real NLOS datademonstrate improved reconstruction fidelity compared to representativephysics-based and learning-based baselines.

Visual Place Recognition (VPR) has been traditionally for-mulated as a single-image retrieval task. Using multiple views offers clearadvantages, yet this setting remains relatively underexplored and existingmethods often struggle to generalize across diverse environments. In thiswork, we introduce UniPR-3D, the first VPR architecture that effectivelyintegrates geometry-aware information from multiple views. UniPR-3Dbuilds on a VGGT backbone capable of encoding multi-view 3D represen-tations, which we adapt by designing feature aggregators and fine-tuningthem for the place recognition task. To construct our descriptor, wejointly leverage VGGT’s 3D tokens, but also intermediate 2D ones. Basedon their distinct characteristics, we design dedicated aggregation modulesfor 2D and 3D features, allowing our descriptor to capture fine-grainedtexture patterns while also reasoning across viewpoints. To further en-hance generalization, we incorporate both single- and multi-frame aggre-gation schemes, along with a variable-length sequence retrieval strategy.Our experiments show that UniPR-3D sets a new state of the art, out-performing both single- and multi-view baselines and highlighting theeffectiveness of geometry-grounded tokens for VPR. Our code and mod-els will be made publicly available on Github. We will release the codeand datasets publicly on https://github.com/dtc111111/UniPR-3D.

Recent 4D Gaussian Splatting (4DGS) methods achieve im-pressive dynamic scene reconstruction but often rely on piecewise linearvelocity approximations and short temporal windows. This disjointedmodeling leads to severe temporal fragmentation, forcing primitives tobe repeatedly eliminated and regenerated to track complex nonlinear dy-namics. This makeshift approximation eliminates the long-term temporalidentity of objects and causes an inevitable proliferation of Gaussians,hindering scalability to extended video sequences. To address this, wepropose TRiGS, a novel 4D representation that utilizes unified, contin-uous geometric transformations. By integrating SE(3) transformations,hierarchical Bézier residuals, and learnable local anchors, TRiGS mod-els geometrically consistent rigid motions for individual primitives. Thiscontinuous formulation preserves temporal identity and effectively mit-igates unbounded memory growth. Extensive experiments demonstratethat TRiGS achieves high fidelity rendering on standard benchmarkswhile uniquely scaling to extended video sequences (e.g., 600 to 1200frames) without severe memory bottlenecks, significantly outperformingprior works in temporal stability.

We introduce Spatial Amsan, a benchmark for evaluatingstate-based spatial reasoning and action evaluation in egocentric manip-ulation videos. The term amsan (Korean for mental arithmetic) reflectsthe core challenge: constructing and updating an internal spatial state toevaluate whether each manipulation advances toward a goal. The bench-mark comprises a tangram task (7 colored blocks) and a wooden puz-zle task (16 blocks), totaling 282 scenarios with 1,801 evaluation turns.Models are evaluated through a multi-turn protocol that diagnosticallyisolates four capabilities: visual perception, temporal tracking, spatialcontact reasoning, and goal-directed action evaluation. Our zero-shotevaluation of nine VLMs spanning frontier API models and open-weightlocal models reveals that contact reasoning is a universal bottleneck (12–16% on the wooden puzzle despite 53–96% on block identification) andthat action evaluation degrades from above 79% to below 16% over 8turns while perception remains stable. These results highlight a fun-damental gap between perceiving scene changes and reasoning abouttheir correctness over time. We release all data, annotations, and codeat https://github.com/Blanchard-lab/SpatialAmsan.

Different stages of manipulation tasks exhibit varying lev-els of difficulty, suggesting stage-dependent motion speeds and temporalprediction horizons. However, existing IL-based visuomotor policies typ-ically imitate the execution speed of expert demonstrations and operatewith a fixed temporal prediction horizon, limiting flexibility and over-all task throughput. In this paper, we introduce AutoSpeed, a model-agnostic learning framework that enables existing visuomotor policies topredict trajectories with stage-adaptive motion speeds, without requir-ing speed or stage annotations. We treat future trajectories at differentspeeds as candidate optimization targets, evaluate each candidate us-ing a composite cost that trades off prediction error against predictionhorizon, and optimize the policy toward the minimum-cost candidate.With a fixed-length action sequence, speed modulation adjusts the ef-fective temporal prediction horizon: simple stages are executed fasterwith a longer prediction horizon, whereas complex stages are executedmore slowly with a shorter prediction horizon. Specifically, we implementspeed modulation in the frequency domain via the discrete cosine trans-form (DCT), which enables smooth, non-integer speed scaling and thuspreserves motion continuity. Extensive evaluations show that AutoSpeedsubstantially reduces task execution time while also improving successrates. Under the AutoSpeed framework, the inferred motion speeds ex-hibit a strong correspondence with task stages.

Adversarial patch attacks that aim to fool deep neural network models in the real world prove effective in ‘every’ possible network type, including vision-language models (VLMs). Surprisingly, most of the existing adversarial patch generation algorithms develop patches of fixed shape (presumably rectangular or circular in a few instances). Due to such biases toward the shape of an adversarial patch, existing adversarial patch detectors, including VLMs, which yield high accuracy, provide a false sense of security. Our extensive experiments equivalently showcase the bias towards the fixed shape of adversarial patches and their high detection, which might leave a hole in adversarial patch defense against adaptive patches. In response, we have first proposed a novel deformable adversarial patch generation algorithm with multifold aims: (i) to highlight the vulnerabilities of state-of-the-art (SOTA) image classification algorithms and (ii) to help provide a strong foundation for developing a universal adversarial patch detector. Introducing smooth and deformable patches that significantly increase practical threats in real-world scenarios. Using the proposed algorithm, we developed large-scale adversarial patch detection datasets comprising 140,000 adversarial patch images for rigorous robustness evaluation and generalizable adversarial detection. The proposed deformable patches effectively fool several SOTA image classification models. While the defense is not the primary focus, our preliminary study highlights that detecting the proposed deformable patches is hard. Still, they can enhance security against various patch shapes, including fixed and deformable patches.

Contrastive pre-training has propelled video-text alignment, yet models often inherit the critical limitations of their image-text predecessors like CLIP, resulting in entangled representations. These challenges are severely exacerbated by two fundamental properties in the video domain: Temporal Misalignment, where textual descriptions often correlate only to specific, constrained temporal windows, leaving other frames text-irrelevant; and Semantic Asymmetry, which dictates a sparse, bidirectional, and non-equivalent relevance between frame-level visual details and caption-level concepts. This failure persists whether captions are short and temporally disjoint, creating ambiguity, or long and detailed, fostering entanglement between static objects and their temporal evolution. In this paper, we establish theoretical conditions that enable flexible alignment between video and text representations across the temporal dimension and at varying levels of granularity. Building on these theoretical insights, we introduce MoVA—Modular Long Video–Text Alignment—which learns dual asymmetric projections: a text-side projection that adaptively selects frame-aware subspaces of the caption, and a video-side projection that disentangles text-relevant visual concepts. Our framework ensures that the model can preserve global cross-modal semantics while disentangling evolving, frame-specific concepts and scale naturally to long captions and videos. Empirical evaluations show that MoVA outperforms existing methods in multiple video-text alignment tasks, demonstrating the effectiveness of our method.

What if a world simulation model could render not an imag-ined environment but a city that actually exists? Prior generative worldmodels synthesize visually plausible yet artificial environments by imag-ining all content. We present Seoul World Model (SWM), a city-scale world model grounded in the real city of Seoul. SWM anchors au-toregressive video generation through retrieval-augmented conditioningon nearby street-view images. However, this design introduces severalchallenges, including temporal misalignment between retrieved referencesand the dynamic target scene, limited trajectory diversity and data spar-sity from vehicle-mounted captures at sparse intervals. We address thesechallenges through cross-temporal pairing, a large-scale synthetic datasetenabling diverse camera trajectories, and a view interpolation pipelinethat synthesizes coherent training videos from sparse street-view images.We further introduce a Virtual Lookahead Sink to stabilize long-horizongeneration by continuously re-grounding each chunk to a retrieved im-age at a future location. We evaluate SWM against recent video worldmodels across three cities: Seoul, Busan, and Ann Arbor. SWM out-performs existing methods in generating spatially faithful, temporallyconsistent, long-horizon videos grounded in actual urban environmentsover trajectories reaching hundreds of meters, while supporting diversecamera movements and text-prompted scenario variations.

End-to-end (E2E) autonomous driving algorithms requirerigorous closed-loop validation in simulation environments offering highvisual fidelity, strong interactivity, and real-time performance. Exist-ing approaches, from game engines to static neural rendering, inher-ently trade off these requirements and struggle with the dynamic scenecomposition essential for E2E testing. To bridge this gap, we proposea novel decoupled 3D Gaussian Splatting (3DGS) framework tailoredfor large-scale E2E evaluation. We fundamentally decompose scenes intoa high-fidelity static background and manipulable dynamic agents us-ing an object-centric canonical representation. To resolve resulting rep-resentational conflicts, we introduce three targeted modules: (1) assetcompression via perceptual pruning and vector quantization for real-time traffic rendering; (2) map-guided geometric registration leveragingsemantic topology to strictly align trajectories; and (3) proxy-based re-lighting transferring ambient illumination for seamless photometric inte-gration. Extensive experiments demonstrate that DecoupleGS achievesa balanced fidelity-efficiency trade-off, improves metric and photomet-ric consistency, and provides a practical closed-loop sensor simulationplatform for E2E autonomous driving evaluation.

Portrait animation has witnessed tremendous quality im-provements thanks to recent advances in video diffusion models. How-ever, these 2D methods often compromise 3D consistency and speed, lim-iting their applicability in real-world scenarios, such as digital twins ortelepresence. In contrast, 3D-aware feedforward facial animation methods– built upon 3D representations, such as neural radiance fields or Gaus-sian splatting – ensure 3D consistency and achieve faster inference speed,but come with inferior expression details. In this paper, we address thisportrait animation trilemma (speed, 3D consistency, and expressiveness)and propose a pipeline that instantly converts an in-the-wild single im-age into a 3D-consistent, fast yet expressive animatable representationvia a feed-forward encoder. Unlike previous computationally intensiveglobal fusion mechanisms (e.g., multiple attention layers) for fusing 3Dstructural and animation information, our design employs an efficientlightweight local fusion strategy to achieve high animation expressivity.Furthermore, our animation representation is decoupled from the face’s3D representation and learns motion implicitly from data, eliminatingthe dependency on pre-defined parametric models that often constrainanimation capabilities. Our method runs at 107.31 FPS for animationand pose control, representing a 3-4 order of magnitude speedup versusthe state of the art while achieving comparable animation quality, thussurpassing alternative designs that trade speed for quality or vice versa.

Test-Time Adaptation (TTA) methods commonly updatethe affine parameters of normalization layers to adapt deployed modelsunder distribution shifts. However, per-channel affine parameters per-form axis-aligned scaling and shifting, making them geometrically inca-pable of correcting cross-channel structural changes induced by distribu-tion shift. To address this limitation, we propose MixTTA, a lightweightplug-in module that equips normalization layers with a low-rank cross-channel transformation, enabling inter-channel mixing at each layer. Toensure that the low-rank branch captures only cross-channel interac-tions, we also propose Decoupling Projection that enforces strict separa-tion from the diagonal affine path, along with Spectral Projection thatprevents rank-1 collapse under non-stationary test streams. MixTTAcan be seamlessly integrated into any existing normalization-based TTAmethod. Experiments in both standard and wild TTA settings show con-sistent improvements over strong baselines while mitigating adaptationfailure under challenging conditions. The source code is publicly availableat https://github.com/delta6189/MixTTA.

Amodal segmentation aims to recover the full shape of objects that are partially occluded; however, most existing approaches are trained and assessed on in-domain benchmarks and exhibit limited generalization in zero-shot scenarios. Human-intuition-guided depth-aware (HIDA) is proposed as a plug-and-play framework that combines two frozen vision foundation models with a lightweight trainable segmentation network to estimate amodal masks using only a visible bounding box. HIDA initially derives a visible prior mask and a pseudo-depth map (PDM) to enable target-centered cropping, which reduces irrelevant context while highlighting occlusion boundaries and hidden regions. An occlusion-aware token-routing soft MoE module is also introduced within a lightweight ViT encoder to adaptively handle varied occlusion patterns and strengthen representations of invisible regions. Evaluations on COCOA-cls, KINS, and D2SA demonstrate that HIDA attains state-of-the-art performance in fully supervised settings and state-of-the-art zero-shot results on COCOA-cls and D2SA while training only a lightweight segmentation network. Code is available at: https://github.com/lipenghuai/HIDA.