ECCV 2026 Papers

In real-world deployment under strict real-time constraints,weather and imaging variations induce significant distribution shifts,severely degrading detectors. Single-Domain Generalized Object Detec-tion aims to mitigate this issue, yet existing methods rarely investi-gate—at the level of problem formulation—the generalization capabil-ity of real-time detectors under such constrained inference budgets. Tothis end, we introduce Real-Time Single-Domain Generalized ObjectDetection (RT-SDGOD), which focuses on how real-time detectors canachieve cross-domain generalization under zero extra inference overheadby relying solely on training-time representation learning. We observethat, under domain shift, DETR-based real-time detectors mainly de-grade through increased missed detections, rooted in limited and un-stable object-level discriminative evidence. Based on this, we proposeRT-SDGDet, a multi-evidence collaborative modeling framework for RT-SDGOD. The core idea is to enable multiple queries of the same ob-ject to collaboratively cover more sufficient discriminative evidence whilemaintaining the stability of such evidence modeling across views. Specifi-cally, we use one-to-many (O2M) supervision to construct stable object-specific query groups, and further design Discriminative Evidence Di-versity Learning (DEDL) and Dual-view Evidence Consistency Learning(DvECL) to expand object-level evidence coverage and improve evidencestability under appearance perturbations, respectively. Since all compo-nents are introduced only during training, our method incurs no ex-tra inference overhead. Extensive experiments show that the proposedmethod achieves better generalization performance than existing ap-proaches across multiple unseen target domains.

Recovering 3D human motion from sparse wearable sensorsis challenging due to the inherent noise, occlusion, and instability ofreal-world tracking signals. Existing approaches are restricted to fixedsensor configurations and, when built on diffusion models, are computa-tionally prohibitive for real-time deployment. We propose OmniPoser, auniversal framework that recovers full-body SMPL motion from a flex-ible combination of heterogeneous inputs, such as 3D keypoints, sparseIMU rotations, head-mounted 6-DoF poses, or any subset thereof, withina single model. OmniPoser introduces a cross-modal masking mechanismpaired with a Masked Conditioning Encoder, enabling one architectureto incorporate diverse sensor configurations through a masked design. Adecoupled dual-stream generative backbone generates complete motionvia Conditional Flow Matching, requiring only a single ODE integrationstep at inference. Across three settings and seven benchmarks, includingin-the-wild Nymeria and Ego-Exo4D dataset, OmniPoser performs onpar with state-of-the-art methods while running at over 700 FPS. Thecode will be open-sourced.

Out-of-Distribution (OOD) detection is essential for ensur-ing the robustness and reliability of object detection systems deployed insafety-critical applications. While prior research has mainly focused onuni-modal detectors or vision-language model (VLM) based classifiers,the potential of VLM-based object detectors in OOD scenarios remainsunderexplored. In this work, we take the first step toward building OODobject detection methods upon VLMs. We identify two key challenges:(i) learning discriminative features to separate in-distribution (ID) fromOOD instances, and (ii) designing scoring functions consistent with VLMprobabilistic outputs. Hence, we introduce Negative Label Guided Atten-tion and Scoring (NegAS). To address (i), we propose a negative labelguided attention module (NegA), where LLM-generated, visually-similarbut semantically-different negative labels are used to guide attention to-ward potential OOD background regions. To address (ii), we introducea novel sigmoid-based OOD scoring function (NegS) that leverages bothID and negative labels, producing strong responses for ID instances andsuppressed responses for OOD ones. Extensive experiments demonstratethat our approach improves OOD detection performance by a large mar-gin while maintaining competitive ID accuracy, e.g., reducing the FPR95by 11.4% on the COCO dataset and 25.5% on the OpenImages datasetcompared to the baseline model. While initially designed for dense VLMdetectors like YOLO-World, we successfully adapt NegAS to GroundingDINO, a query-based VLM transformer, achieving significant improve-ments, demonstrating the generalizability of our framework. The code isavailable at https://github.com/yjzzyj/NegAS.

Tracking multi-person 3D human meshes from in-the-wildvideos is a highly challenging problem due to complex interactions, fre-quent occlusions, and severe truncation inherent in unconstrained envi-ronments. While recent approaches have improved robustness againstthese issues, they largely overlook the critical challenge prevalent inreal-world footage: frequent shot changes. These abrupt transitions incamera viewpoints often cause existing methods to lose track of humanidentities and fail in reconstructing temporally coherent trajectories. Al-though several recent works have explored 3D human mesh tracking un-der shot changes, they are still limited to single-person scenarios, mak-ing them inadequate for real-world videos where multiple people interactand appear simultaneously. To address this limitation, we propose Multi-THuMBS (Multi-person Tracking of 3D Human Meshes Beyond VideoShots) that leverages a state-of-the-art 3D scene prior to reconstructthe two boundary frames in a single shared 3D space. Human meshesare then registered within the shared 3D space, maintaining per-personidentity and motion consistency across shot changes. Extensive experi-ments demonstrate that our approach yields significant improvements in3D human mesh recovery, camera pose estimation, and identity track-ing, thereby ensuring high-fidelity motion reconstruction with consistentidentity preservation across shots compared to previous state-of-the-artmethods.

Many real-world applications of spatial intelligence, such asrobotic control, autonomous driving, and automated assembly, requirespatial reasoning across multiple sequential steps. However, the extent towhich current Multimodal Large Language Models (MLLMs) possess thiscapability remains largely unexplored. Inspired by LEGO construction, arecreational activity that critically relies on multi-step spatial reasoning,we introduce LEGO-Puzzles: a benchmark designed to systematicallyevaluate the spatial reasoning capabilities of MLLMs from basic spatialunderstanding to multi-step planning. LEGO-Puzzles contains two tasksets. The Elementary set covers 11 visual question-answering (VQA)tasks with 1, 100 carefully curated samples to test elementary spatialreasoning skills that are cruical for LEGO assembly. The Planning setdirectly requires the model to generate a step-by-step plan for assemblinga target LEGO structure, where the tasks are organized into subsets withdifferent planning horizons ranging up to 8. Our evaluation of 29 state-of-the-art MLLMs shows that even the strongest models struggle withelementary reasoning tasks in LEGO construction, falling at least 20%behind human performance. The planning accuracy also quickly dropsto 0% as the number of planning steps increases, whereas our humanparticipants solve all the tasks perfectly. Switching the output formatfrom multiple choice to image generation degrades model performanceeven further, leading to zero accuracy even for planning 3 steps. Over-all, LEGO-Puzzles reveals critical limitations in current MLLMs’ spatialreasoning capabilities and highlights the need for substantial advances.

Analyzing fine-grained skill activities (e.g., sports, surgery) requires not only recognizing visual patterns but also performing stepby-step visual reasoning that leads to the final judgment. While recent advances in action quality assessment have achieved remarkable progress in evaluating performance, existing models remain black boxes, where they lack the ability to explicitly reveal the reasoning processes underlying their judgments. To address this limitation, we propose Latent Visual Diffusion Reasoning (LVDR), a novel framework that integrates keypoint-guided Monte Carlo Tree Search (MCTS) to model and visualize the latent visual reasoning process. LVDR not only produces more accurate skill assessments but also uncovers the critical visual reasoning sequences that contribute to the final evaluation. Extensive experiments across four datasets spanning diverse sports and surgical domains demonstrate that LVDR achieves competitive quantitative performance while providing interpretable visual reasoning trajectories leading to the final predictions. Source codes and models can be found through the following link: https://github.com/XiruiTeng/LVDR_Official.git.

Existing Stochastic 3D Human Motion Prediction models are fundamentally constrained by hard-coding the skeleton kinematics, severely limiting generalization, preventing cross-dataset training, and requiring complex data retargeting. We introduce EquiFusion, the first kinematics-agnostic model to solve this bottleneck, implementing a latent diffusion model with a permutation equivariant architecture. EquiFusion treats the kinematics’ connectivity as an explicit input parameter, ensuring its internal computations are inherently agnostic to joint ordering and graph structure. This novel design enables truly cross-dataset generalization to unseen kinematics and unlocks novel zero-shot directions, such as motion prediction from partial or occluded observations and targeted limb generation. EquiFusion achieves state-of-the-art results on major benchmarks, being up to 75% more compact than previous kinematics-specific methods, while achieving faster training and inference. EquiFusion thus establishes a new, flexible standard for robust human motion prediction.

Knowledge-Based Visual Question Answering (KB-VQA) aimsto evaluate whether Visual Language Models (VLMs) can retrieve, ground,and reason over external structured knowledge beyond visual evidence.In practice, answer accuracy is widely adopted as the primary evalu-ation metric, implicitly treating correctness as a proxy for knowledge-grounded reasoning. However, for existing KB-VQA benchmarks, thisproxy relies on critical assumptions that are often overlooked and ren-dered unreliable by benchmark issues: annotated answer must be deriv-able from the associated knowledge base, question must be well-posedwith su!cient constraints, and visual setting must meaningfully requiregrounded disambiguation. In this work, we show that these assumptionsare systematically violated in existing KB-VQA benchmarks. Our au-dit reveals substantial instances with missing or contradicted answersand underspecified questions that render accuracy a misleading metric.Furthermore, we find that existing datasets rely on visually trivial, single-entity scenes that bypass the need for sophisticated visual-to-knowledgemapping. We demonstrate that even with controlled architectures, theseflaws lead to distorted model rankings and overestimations of reasoningcapabilities. To address this, we introduce (1) a principled audit-and-repair protocol that restores answer derivability and question clarity,and (2) a controlled multi-entity augmentation protocol that introducesvisual ambiguity to challenge initial retrieval and grounded reasoning.Re-evaluation under corrected and augmented settings yields markedlydi"erent performance trends. Our findings call for rethinking evaluationprotocols and designing more interaction-aware KB-VQA benchmarksthat prioritize verifiable reasoning over simple matching. 1

The physical world is not merely visual; it is governed byrigorous structural and procedural constraints. Yet, the evaluation ofvision-language models (VLMs) remains heavily skewed toward percep-tual realism, prioritizing the generation of visually plausible 3D lay-outs, shapes, and appearances. Current benchmarks rarely test whethermodels grasp the step-by-step processes and physical dependencies re-quired to actually build these artifacts—a capability essential for au-tomating design-to-construction pipelines. To address this, we introduceDreamHouse, a novel benchmark for physical generative reasoning:the capacity to synthesize artifacts that concurrently satisfy geomet-ric, structural, constructability, and code-compliance constraints. Weground this benchmark in residential timber-frame construction, a do-main with fully codified engineering standards and objectively verifiablecorrectness. We curate over 26,000 structures spanning 13 architecturalstyles—each verified to construction-document standards (LOD 350)—and develop a deterministic 10-test structural validation framework. Un-like static benchmarks that assess only final outputs, DreamHouse sup-ports iterative agentic interaction. Models observe intermediate buildstates, generate construction actions, and receive structured environ-mental feedback, enabling a fine-grained evaluation of planning, struc-tural reasoning, and self-correction. Extensive experiments with state-of-the-art VLMs reveal substantial capability gaps that are largely invis-ible on existing leaderboards. These findings establish physical validityas a critical evaluation axis orthogonal to visual realism, highlightingphysical generative reasoning as a distinct and underdeveloped fron-tier in multimodal intelligence. The benchmark is available at https://luluyuyuyang.github.io/dreamhouse/.

We present HairWeaver, a diffusion-based pipeline that ani-mates a single human image with realistic and expressive hair dynamics.While existing methods successfully control body pose, they lack specificcontrol over hair, and as a result, fail to capture the intricate hair mo-tions, resulting in stiff and unrealistic animations. HairWeaver overcomesthis limitation using two specialized modules: a Motion-Context-LoRAto integrate motion conditions and a Style-Alignment-LoRA to pre-serve the subject’s photoreal appearance across different data domains.These lightweight components are designed to guide a video diffusionbackbone while maintaining its core generative capabilities. By trainingon a specialized dataset of dynamic human motion generated from aCG simulator, HairWeaver affords fine control over hair motion and ulti-mately learns to produce highly realistic hair that responds naturally tomovement. Comprehensive evaluations demonstrate that our approachsets a new state of the art, producing lifelike human hair animationswith dynamic details.

End-to-end planning has emerged as a dominant paradigmfor autonomous driving, where recent models often adopt a scoring-selection framework to choose trajectories from a large set of candidates,with diffusion-based decoding showing strong promise. However, directlyselecting from the entire candidate space remains difficult to optimize,and Gaussian perturbations used in diffusion often introduce unrealis-tic trajectories that complicate the denoising process. In addition, fortraining these models, reinforcement learning (RL) has shown promise,but existing end-to-end RL approaches typically rely on a single coupledreward without structured signals, limiting optimization effectiveness.To address these challenges, we propose HAD, an end-to-end planningframework with a Hierarchical Diffusion Policy that decomposes planninginto a coarse-to-fine process. To improve trajectory generation, we intro-duce Structure-Preserved Trajectory Expansion, which produces realisticcandidates while maintaining kinematic structure. For policy learning, wedevelop Metric-Decoupled Policy Optimization (MDPO) to enable struc-tured RL optimization across multiple driving objectives. Extensive ex-periments show that HAD achieves new state-of-the-art performance onboth NAVSIM and HUGSIM, outperforming prior arts by a huge margin:+2.3 EPDMS on NAVSIM and +4.9 Route Completion on HUGSIM.

Feed-forward foundation models for multi-view 3-dimensional(3D) reconstruction have been trained on large-scale datasets of perspec-tive images; when tested on wide field-of-view images, e.g., from a fish-eye camera, their performance degrades. This degradation arises fromchanges in spatial arrangements of pixels induced by the non-linear pro-jection model that maps 3D points onto the 2D image plane. While onemay surmise that training on fisheye images would resolve this problem,there are far fewer fisheye images with ground truth than perspectiveimages, which limits generalization. To enable inference on imagery ex-hibiting high radial distortion, we propose Fisheye3R, a novel adaptationframework that extends these multi-view 3D reconstruction foundationmodels to natively accommodate fisheye inputs without performance re-gression on perspective images. To address the scarcity of fisheye imagesand ground truth, we introduce flexible learning schemes that supportself-supervised adaptation using only unlabeled perspective images andsupervised adaptation without any fisheye training data. Extensive ex-periments across three foundation models, including VGGT, π 3 , and Ma-pAnything, demonstrate that our approach consistently improves camerapose, depth, point map, and field-of-view estimation on fisheye images.Code is available at https://github.com/android-xr/fisheye3r.

This work presents Lapis, a linear-attention-based pixel-space generative framework that achieves efficient and high-fidelity depthestimation with one-step diffusion. While generative frameworks havesignificantly advanced monocular depth estimation with superior detailfidelity, the O(N 2 ) complexity of standard attention and the multi-stepdenoising process introduce prohibitive computational costs when scal-ing them to high-resolution image applications. Although linear atten-tion and one-step prediction are intuitively viable, directly applying themleads to poor structural consistency, detail loss, and noise. Lapis recti-fies these limitations through a coarse-to-fine hierarchy. Specifically, aPatch-level Consistency Module restores structural coherence by inte-grating semantic and spatial priors. Subsequently, a Pixel-level Refine-ment Module recovers sharp geometric boundaries via skip-connection-based pixel correspondence. Furthermore, to mitigate sampling noise in-herent in one-step diffusion, we leverage the manifold assumption andadopt a direct x-prediction strategy to target the clean data manifold.Extensive evaluations on multiple benchmarks demonstrate that Lapisconsistently achieves state-of-the-art (SOTA) accuracy and boundarysharpness across various resolutions, reducing inference latency by upto 7.6× at 1080P and 10.9× at 1440P resolution compared to previousSOTA generative models.

Language-guided prompt-based continual learning emergesas a promising approach to tackle catastrophic forgetting in a dynamicenvironment under rehearsal-free constraint. However, existing state-of-the-art (SOTA) requires additional resources, such as LLM-generateddescriptors and an LLM decoder that may not be available in a real ap-plication. Meanwhile, the more straightforward method experiences sub-optimality due to the high similarity between language text embeddings.To address this problem, we propose a resource-efficient novel language-guided approach and algorithm that incorporates two novel ideas: (1)mixed vision and language modality for prompt generation, and (2) Dif-ferential prefix tuning for the model training process. Our experimentalanalysis shows that our method outperforms existing language-guidedprompt-based methods, i.e., up to 30%, 26%, and 10% for final aver-age accuracy, cumulative average accuracy, and final forgetting mea-sure, respectively. The historical analysis confirms our method’s stability-plasticity balance in every task. Our extended analysis shows that ourmethod consistently achieves better performance in various prompt lengthsand ViT layers. For further study and reproducibility, we also providerigorous analysis, details, and source code of our method in the sup-plementary document. The implementation of the proposed method isavailable at https://github.com/anwarmaxsum/D-VLAM.

LiDAR-based 3D object detection is essential for autonomousdriving systems. However, traditional Ego-only Perception (Eo-Perception)suffers from limited perspective and occlusions in a complex outdoor en-vironment, leading to performance bottlenecks. Recently, research onmulti-agent Collaborative Perception (Co-Perception) has demonstratedexcellent performance, but high communication costs and accumulatedpose error hinder its application. To address this, we explore a novelC2E (Co-Perception to Eo-Perception) paradigm through the Multi-to-Single (M2S) agent contrastive knowledge distillation framework. OurM2S framework first designs Multi-Level Feature Enhancement moduleto provide more stable features, and introduces Auxiliary Point CloudReconstruction and Multi-Teacher Contrastive Distillation mechanismsto mitigate domain gaps in point cloud and feature distributions withinthe C2E paradigm. Benefiting from this, our M2S can retain the ex-cellent performance of collaborative perception while effectively avoidingthe drawbacks, such as communication delays and positioning errors. Ex-tensive experiments on the V2XSet, V2V4Real and DAIR-V2X datasetsshow the effectiveness and generalizability of our M2S framework whencombined with the state-of-the-art CoSDH model and other excellent 3Ddetectors. Our M2S framework can deliver up to a 8.64% improvementin 3D mAP performance without introducing any communication costs.

Zero-shot skeleton-based action recognition aims to identifyunseen actions via semantic transfer. The existing methods often strug-gle with the topological mismatch between skeleton and text manifolds,leading to loose decision boundaries due to weak constraints and fail-ing to adapt to unseen distributions during static inference. To miti-gate this challenge, we propose ProtoFM, a novel generative frameworkbased on conditional flow matching and prototype guidance. First, in-stead of heterogeneous skeleton-text matching, ProtoFM introduces agenerative matching paradigm that performs homogeneous skeleton andpseudo-skeleton matching by synthesizing features via ODE-based flowintegration. Second, we propose a prototype-guided center loss that an-chors flow-trajectory states to learnable class prototypes, encouragingcompact semantic clusters and improved class separation. Last, at theinference phase, we develop a training-free prototype-guided rectificationstrategy that exploits a global memory bank for progressive test-timerefinement. Extensive experimental results obtained on multiple bench-marking datasets demonstrate the superiority of the proposed ProtoFMmethod over the state-of-the-art approaches.

Recent 3D generative models produce high-quality geometryfrom a single image using large-scale priors and diffusion architectures.However, they assume complete visibility and single-view inputs, limitingapplicability in multi-view, occluded, or editing scenarios. Although priorworks address these challenges individually, they lack a unified frameworkfor controllable 3D completion under diverse conditioning signals.We present Axolotl3D, a multi-modal and occlusion-aware 3D generationmodel that jointly conditions on images, visibility masks, camera param-eters, and a partial point cloud. The point cloud serves as a geometricanchor promoting faithful shape completion, while camera parametersensure consistent multi-view alignment in a shared 3D coordinate sys-tem. A unified training strategy synthesizes diverse conditioning regimesfrom large-scale 3D data, enabling robust cross-modal reasoning.Experiments on Toys4K and OmniObject3D demonstrate state-of-the-art performance under both clean and occluded settings, as well as strongresults in real-world reconstruction and geometry-consistent editing.

Traditional high-quality 3D scanning and reconstruction typ-ically relies on human labor to plan the scanning procedure. With therapid development of embodied systems such as drones and robots, thereis a growing demand of performing accurate 3D scanning and recon-struction in an fully automated manner. We introduce Auto3R, a data-driven uncertainty quantification model that is designed to automatethe 3D scanning and reconstruction of scenes and objects, includingobjects with non-lambertian and specular materials. Specifically, in aprocess of iterative 3D reconstruction and scanning, Auto3R can makeefficient and accurate prediction of uncertainty distribution over poten-tial scanning viewpoints, without knowing the ground truth geometryand appearance. Through extensive experiments, Auto3R achieves supe-rior performance that outperforms the state-of-the-art methods, partic-ularly on challenging viewpoints. We also deploy Auto3R on a robotarm equipped with a camera and demonstrate that Auto3R can beused to effectively digitize real-world 3D objects and delivers ready-to-use and photorealistic digital assets. Our code is available at https://tomatoma00.github.io/auto3r.github.io/.

Deep neural networks often show poor generalization whentrained on biased datasets presenting spurious relations between sam-ples and target labels. Unsupervised debiasing methods aim at mitigat-ing models’ dependency on bias, without relying on bias information.Bias shortcuts are typically learned early in training, with the modelrapidly overfitting the few samples not sharing the same related at-tributes. Avoiding bias conflicting memorization for mining bias infor-mation useful for mitigation (e.g., in the form of pseudo-labels) is still afundamental open issue in this context, with existing approaches propos-ing to learn for a few epochs or to rely on unbiased validation sets for earlystopping. In this work, we propose to exploit intermediate layers’ featuresfor achieving bias mitigation robust to the memorization problem, replac-ing the when to stop with a where to look paradigm. Specifically, we pro-pose AracNet, an unsupervised debiasing framework for self-mitigation.Attaching one linear layer to each network’s block, which we refer to asShallow Monitor, we obtain a useful debiasing signal capable of miti-gating bias dependency in the same model when randomly reinitialized.Our results on typical benchmark datasets with single and multiple biasesconfirm the effectiveness of the proposed approach, paving the way forfurther research in understanding how bias propagates in trained models.Code is available at https://github.com/Malga-Vision/AracNet