ECCV 2026 Papers

We introduce the challenging Sterkfontein Caves dataset comprising ten underground scenes from a UNESCO World Heritage Site, and use it to find a new simple baseline method that beats existing low-light reconstruction methods upon it. Each scene is of complex surface geometry and high-frequency texture from cave rock structures, including humanmade markings on the rock face. The captured images exhibit varied or uncontrolled lighting over a large dynamic range, with glare artefacts, and low signal-to-noise ratios from the challenging dark real-world capture scenario.Weproposeaviewsynthesisbenchmarkforlow-lightRAWandsRGB reconstruction. All tested NeRF and Gaussian splatting baseline methods struggleonthisdata,withthebestperformingmethodintermsofreliability andaveragePSNRbeingourRaw-Nerfactomethod.Wediscusstheseerrors in detail to find directions of future work for the community in overcoming the significant challenges that remain in low-light high-detail scenes. Our dataset is publicly available at https://visual.cs.brown.edu/sterkfontein.

Parametric 3D human models such as SMPL have driven significant advances in human pose and shape estimation, yet their simplified kinematics limit biomechanical realism. The recently proposed SKEL model addresses this limitation by re-rigging SMPL with an anatomically accurate skeleton. However, estimating SKEL parameters directly remains challenging due to limited training data, perspective ambiguities, and the inherent complexity of human articulation. In this work, we propose SKEL-CF, a new framework for estimating SKEL parameters. SKEL-CF adopts a standard transformer-based encoder–decoder architecture. The encoder first produces coarse predictions of the camera extrinsics and SKEL parameters. The decoder then iteratively refines these predictions across multiple layers, with explicit and auxiliary supervision applied at each layer. To provide anatomically consistent training data, we convert the existing SMPL-based dataset into a SKEL-aligned version, called HMR-SKEL. This new dataset offers high-quality supervision for SKEL estimation. In addition, to reduce depth and scale ambiguity, we explicitly incorporate camera intrinsic estimation into the SKEL-CF pipeline and show that it is important for accurate reconstruction. Extensive experiments validate the effectiveness of the proposed design. On the challenging MOYO dataset, SKEL-CF achieves 85.0 MPJPE / 51.4 PA-MPJPE, significantly outperforming the previous SKEL-based state-of-the-art HSMR (104.5 / 79.6). These results establish SKEL-CF as a promising framework for human motion analysis, facilitating the use of computer vision techniques in biomechanicsrelated analysis. Our implementation is available on the project page: https://pokerman8.github.io/SKEL-CF/.

Accurate document layout analysis remains a critical bottleneck for document parsing systems, due to the intricate coupling among heterogeneous document layout elements, geometric distortions (e.g., paper warping and bending, perspective variations), and reading order within diverse layout structures. Existing approaches typically rely on fragmented multi-stage pipelines or computationally heavy generative Transformer architectures, leading to error propagation and limited efficiency. In this paper, we present RT-DocLayout, a highly efficient end-to-end framework for document layout analysis, designed as a front-end for document parsing tasks. The proposed model unifies classification, detection, pixel-level segmentation, and reading order prediction for layout elements within a single 33M-parameter architecture. Built upon the RT-DETR, our key contribution is a unified multi-task formulation within a single query-based decoder that simultaneously classifies, regresses bounding box, generates masks, and constructs relationship to reason reading order. By jointly learning geometric and structural representations, RT-DocLayout introduces multi-task optimization that substantially improves robustness under real-world document distortions. Extensive experiments on OmniDocBench v1.5 and Real5-OmniDocBench demonstrate state-ofthe-art performance in document layout analysis while maintaining realtime inference speed(132.1 FPS). When coupled with downstream OCR engines, RT-DocLayout significantly improves full-document reconstruction quality, providing a scalable and practical foundation for real-world document intelligence systems. The code and models are available at https://github.com/PaddlePaddle/ PaddleOCR.

Diffusion models achieve strong image and video generation quality but remain slow at inference due to multi-step denoising. Caching accelerates inference by reusing intermediate features across timesteps, yet its effectiveness depends critically on refresh scheduling (when to perform full recomputation). Existing schedules fall into three paradigms: fixed-interval, online heuristic, and global optimization. The first two may miss temporal sensitivity structure or accumulated segment-level degradation, while the last incurs substantial offline cost. We propose Equal-Risk Caching (ERC), which builds a one-dimensional temporal risk profile from a single fixed-length probing cache curve and uses cumulative risk mass as a global surrogate for segment sensitivity. ERC places refresh boundaries by equal-mass partitioning on the cumulative risk axis, automatically allocating denser recomputation to high-risk regions and longer reuse to low-risk regions. This reduces the offline profiling cost of schedule construction from O(T ·L) to O(T). Experiments on multiple diffusion models demonstrate an overall competitive quality– speed trade-off.

Recent advancements have successfully adapted autoregressive language models to process multimodal signals, such as images and actions. Since raw action signals are continuous, effective tokenization is essential to map high-dimensional inputs into compact discrete tokens for autoregressive processing. However, existing discrete action tokenizers often suffer from high reconstruction loss, failing to preserve the finegrained dynamics required for precise control. This “discretization bottleneck” significantly limits the performance ceiling of downstream VisionLanguage-Action (VLA) models. To address this, we propose M2Tok, a Multi-head Multi-codebook Action Tokenizer designed to minimize reconstruction error and enhance policy performance. Our approach introduces two key structural innovations: (1) we decompose the latent action features into multiple heads, enabling the model to implicitly align specific heads with distinct action dimensions; (2) we assign independent codebooks to each head for quantization. By leveraging the combinatorial nature of multiple codebooks, we significantly expand the representational expressivity of the tokenizer, leading to substantially lower reconstruction loss compared to previous methods. We evaluate the M2Tok-based VLA on the RoboTwin, Simpler-Env, and 3 zero-shot real-world tasks. Experimental results demonstrate our method not only achieves superior reconstruction fidelity but also significantly boosts the success rate of VLA models. Comprehensive ablation studies further confirm the effectiveness of the multi-head and multi-codebook mechanisms. Code is available at https://github.com/cpaaax/M2Tok.

In intelligent transportation systems, roadside 3D object detection provides wide-area perception crucial for traffic understanding, cooperative early warning, and safe autonomous driving. However, existing methods suffer from high sensitivity to camera extrinsics; even slight deviations (whether manifesting as transient jitter or persistent drift) can be significantly amplified by projective geometry. This cascade results in severe feature misalignment and degraded localization. To mitigate this limitation, we propose RECO, a region-aware extrinsic compensation framework that corrects extrinsics using piecewise 6-DoF pose offsets. RECO predicts a learnable range boundary to partition the scene into near and far regions, estimating region-specific pose corrections. A differentiable sigmoid gate then smoothly blends the two compensated geometries to preserve continuous BEV sampling and facilitate stable optimization. To supervise the refinement of extrinsics, we introduce an auxiliary reprojection loss that compares 2D bounding boxes projected from 3D ground truth against 2D annotations, optimizing it jointly with the standard detection objective. Extensive experiments on the DAIR-V2X-I and Rope3D benchmarks under extrinsic perturbations demonstrate consistent improvements over state-of-the-art baselines across both yaw and z-axis deviations. RECO also generalizes from transient perturbations to persistent shifts, maintaining highly competitive performance under strict calibration uncertainty.

SAM 3D Body (3DB) achieves state-of-the-art accuracy in monocular 3D human mesh recovery, yet its inference latency of several seconds per image precludes real-time application. We present Fast SAM 3D Body, a training-free acceleration framework that reformulates the 3DB inference pathway to achieve interactive rates. By decoupling serial spatial dependencies and applying architecture-aware pruning, we enable parallelized multi-crop feature extraction and streamlined transformer decoding. Moreover, to extract the joint-level kinematics (SMPL) compatible with existing humanoid control and policy learning frameworks, we replace the iterative mesh fitting with a direct feedforward mapping, accelerating this specific conversion by over 10,000×. Overall, our framework delivers up to a 10.25× end-to-end speedup while maintaining largely on-par reconstruction fidelity, even surpassing 3DB on benchmarks such as LSPET. We demonstrate its utility by deploying Fast SAM 3D Body in a vision-only teleoperation system that—unlike methods reliant on wearable IMUs—enables real-time humanoid control and the direct collection of manipulation policies from a single RGB stream.

We present DynHair, a novel method for tracking and mod-eling dynamic hair for human head avatars. From video input, we recon-struct a dynamic head avatar with an explicit strand-based hair represen-tation using structured 3D Gaussian Splatting. In contrast to the faceregion of human head avatars, which can be modeled with 3D Gaus-sians that are attached or generated with respect to some expressive3D head model, hair is particularly challenging as it exhibits dynamicmotion effects. Therefore, we present a novel method that models thedynamic deformations of the hair strands using a temporal network thatis conditioned on angular velocity and acceleration of the head, as wellas relative gravity. Specifically, an LSTM encodes the motion historyand modulates per-point strand features via FiLM conditioning whichis then used by an MLP to produce physically plausible displacementsto the canonical hairstyle. We jointly optimize this motion and appear-ance representation of the hair, with a 3DGS-based representation ofthe face-region, via differentiable Gaussian splatting with photometric,geometric, and physics-based supervision. As a result, we obtain hair⋆ Equal contribution. † Equal supervision.tracking of the training video data and an animatable head avatar withcontrollable hair dynamics. In our experiments, we demonstrate state-of-the-art performance in terms of hair dynamics, temporal consistency,and generalization across subjects. Our data and code are available athttps://dynhair.is.tue.mpg.de/.

Autonomous robotic systems require spatio-temporal understanding of dynamic environments to ensure reliable navigation and interaction. While Vision-Language Models (VLMs) provide open-world semantic priors, they lack grounding in 3D geometry and temporal dynamics. Conversely, geometric perception captures structure and motion but remains semantically sparse. We propose SNOW (Scene Understanding with Open-World Knowledge), a training-free and backboneagnostic framework for unified 4D scene understanding that integrates VLM-derived semantics with point cloud geometry and temporal consistency. SNOW processes synchronized RGB images and 3D point clouds, using HDBSCAN clustering to generate object-level proposals that guide SAM2-based segmentation. Each segmented region is encoded through our proposed Spatio-Temporal Tokenized Patch Encoding (STEP), producing multimodal tokens that capture localized semantic, geometric, and temporal attributes. These tokens are incrementally integrated into a 4D Scene Graph (4DSG), which serves as 4D prior for downstream reasoning. A lightweight SLAM backend anchors all STEP tokens spatially, providing the global reference alignment, and ensuring unambiguous spatial grounding across time. The resulting 4DSG forms a queryable, unified world model through which VLMs can directly interpret spatial scene structure and temporal dynamics. Experiments on a diverse set of benchmarks demonstrate that SNOW enables precise 4D scene understanding and spatially grounded inference, setting new stateof-the-art performance in several settings, highlighting the importance of structured 4D priors for embodied reasoning and autonomous robotics.

Visual grounding is an essential capability of Visual Language Models (VLMs) to understand the real physical world. Previous state-of-the-art grounding visual language models usually have large model sizes, making them heavy for deployment and slow for inference. However, we notice that the sizes of visual encoders are nearly the same for small and large VLMs and the major difference is the sizes of the language models. Small VLMs fall behind larger VLMs in grounding because of the difference in language understanding capability rather than visual information handling. To mitigate the gap, we introduce ‘Efficient visual Grounding language Models’ (EGM): generate many mid-quality tokens (from small models) to match the performance of large VLMs with few high-quality but expensive tokens. This method is deploymentfriendly, and yields better end-to-end latency: On the RefCOCO benchmark, our EGM-Qwen3-VL-8B demonstrates 91.4 IoU with an average of 737ms (5.9× faster) latency while Qwen3-VL-235B demands 4,320ms to reach 90.5 IoU. To validate our approach’s generality, we further set up a new amodal grounding setting that requires the model to predict both the visible and occluded parts of the objects. Experiments show our method consistently improves both vanilla and amodal grounding capabilities of small models to match or outperform larger models, thereby improving efficiency for visual grounding.

Text-to-image diffusion models have achieved unprecedented success but still struggle to produce high-quality results under limited sampling budgets. Existing training-free sampling acceleration methods are typically developed independently, leaving the overall performance and compatibility among these methods unexplored. In this paper, we bridge this gap by systematically elucidating the design space, and our comprehensive experiments identify the sampling time schedule as the most pivotal factor. Inspired by the geometric properties of diffusion models revealed through the Frenet-Serret formulas, we propose constant total rotation schedule (TORS), a scheduling strategy that ensures uniform geometric variation along the sampling trajectory. TORS outperforms previous training-free acceleration methods and produces highquality images with 10 sampling steps on Flux.1-Dev and Stable Diffusion 3.5. Extensive experiments underscore the adaptability of our method to unseen models, hyperparameters, and downstream applications. Code is available at https://github.com/zju-pi/TORS.

Galaxy clusters are powerful probes of astrophysics and cosmology through gravitational lensing: their mass, dominated by 85% dark matter, distorts background light. Yet, mass reconstruction lacks the scalability and large-scale benchmarks to process the hundreds of thousands of clusters expected from forthcoming wide-field surveys. We introduce a fully automated method to reconstruct clustersurfacemassdensityfromphotometryandgravitationallensingobservables. Central to our approach is DARKCLUSTERS-15K, our new dataset of 15,000 mock cluster observations with paired mass and photometry maps, the largest to date, spanning multiple redshifts and simulation frameworks. We train a plug-and-play diffusion prior on DARKCLUSTERS-15K that learns the statistical relationship between mass and light, and draw approximate posterior samples constrained by weakand strong-lensing observables, yielding principled reconstructions with well-calibrated empirical uncertainties. Our approach requires no expert tuning, runs in minutes rather than hours, achieves higher accuracy, and matches experttuned reconstructions of the MACS 1206 cluster. We release our method and DARKCLUSTERS-15K to support upcoming wide-field cosmological surveys.

Emerging multi-modal world models attempt to jointly generate videos across diverse modalities (e.g., RGB, depth, and mask), yet they fail to fully exploit the priors of existing foundation models. We propose M2-REPA, the first representation alignment method tailored for multi-modal video generation. Our key insight is that foundation models trained on different modality spaces naturally capture distinct domain-specific priors, acting as complementary “experts.” Specifically, we first decouple modality-specific features from the diffusion model’s intermediate representations, then align each with its corresponding expert foundation model. To this end, we design two synergistic objectives: a multi-modal representation alignment loss that enforces feature-toexpert matching, and a modality-specific decoupling regularization that encourages complementarity across different modalities. This design enables joint optimization, exploiting priors from multiple foundation models. Extensive experiments demonstrate that our method outperforms baselines in visual quality and long-term consistency.

Cinematic camera control relies on a tight feedback loop between director and cinematographer, where camera motion and framing are continuously reviewed and refined. Recent generative camera systems can produce diverse, text-conditioned trajectories, but they lack this “director in the loop” and have no explicit supervision of whether a shot is visually desirable. This results in in-distribution camera motion but poor framing, off-screen characters, and undesirable visual aesthetics. In this paper, we introduce VERTIGO, the first framework for visual preference optimization of camera trajectory generators. Our framework leverages a real-time graphics engine (Unity) to render 2D visual previews from generated camera motion. A cinematically fine-tuned vision–language model then scores these previews using our proposed caption-based consistency similarity mechanism, which aligns renders with text prompts. This process provides the visual preference signals for Direct Preference Optimization (DPO) post-training. Both quantitative evaluations and user studies on Unity renders and diffusion-based Camera-to-Video pipelines show consistent gains in condition adherence, framing quality, and perceptual realism. Notably, VERTIGO reduces the character off-screen rate from 38% to nearly 0% while preserving the geometric fidelity of camera motion. User study participants further prefer VERTIGO over baselines across composition, consistency, prompt adherence, and aesthetic quality, confirming the perceptual benefits of our visual preference post-training.

Diffuser-based imaging systems replace conventional lenses with optical diffusers to enable compact lensless cameras, but their performance critically depends on computational reconstruction from compressed measurements. Deep unrolled networks have emerged as a leading framework by incorporating physical forward models into iterative optimization. However, existing methods typically rely on simplified assumptions, most notably shift-invariant point spread functions (PSFs), which fail to capture spatially-varying aberrations and geometric distortions in real hardware. In this paper, we propose learning physics-based forward model corrections for diffuser-based imaging. We construct a physically interpretable forward model with learnable correction terms based on Eigen-Kernel decomposition to handle spatially-varying aberrations. This design preserves the structure of the physical model while improving its fidelity to real measurements. The corrected forward model is integrated into an unrolled reconstruction framework (EK-DUN), leading to consistent improvements in reconstruction accuracy. Experimental results on real-world data demonstrate that our method achieves stateof-the-art performance compared to existing unrolled baselines.

Self-supervised stereo matching holds great promise by eliminating the reliance on expensive ground-truth data. Its dominant paradigm, based on photometric consistency, is however fundamentally hindered by the occlusion challenge—an issue that persists regardless of network architecture. The essential insight is that for any occluders, valid feedback signals can only be derived from the unoccluded areas on one side of the occluder. Existing methods attempt to address this by focusing on the erroneous feedback from the other side, either by identifying and removing it, or by introducing additional regularities for correction on that basis. Nevertheless, these approaches have failed to provide a complete solution. This work proposes a more fundamental solution. The core idea is to transform the fixed state of one-sided valid and one-sided erroneous signals into a probabilistic acquisition of valid feedback from both sides of an occluder. This is achieved through a complete framework, centered on a pseudo-stereo inputs strategy that decouples the input and feedback, without introducing any additional constraints. Qualitative results visually demonstrate that the occlusion problem is resolved, manifested by fully symmetrical and identical performance on both flanks of occluding objects. Quantitative experiments thoroughly validate the significant performance improvements resulting from solving the occlusion challenge. The code is available at https://github.com/qrzyang/PseudoStereo.

Modern vision-language-model (VLM) based graphical user interface (GUI) agents are expected not only to execute actions accurately but also to respond to user instructions with low latency. While existing research on GUI-agent security mainly focuses on manipulating action correctness, the security risks related to response efficiency remain largely unexplored. In this paper, we introduce SlowBA, a novel backdoor attack that targets the responsiveness of VLM-based GUI agents. The key idea is to manipulate response latency by inducing excessively long reasoning chains under specific trigger patterns. To achieve this, we propose a two-stage reward-level backdoor injection (RBI) strategy that first aligns the long-response format and then learns trigger-aware activation through reinforcement learning. In addition, we design realistic pop-up windows as triggers that naturally appear in GUI environments, improving the stealthiness of the attack. Extensive experiments across multiple datasets and baselines demonstrate that SlowBA can significantly increase response length and latency while largely preserving task accuracy. The attack remains effective even with a small poisoning ratio and under several defense settings. These findings reveal a previously overlooked security vulnerability in GUI agents and highlight the need for defenses that consider both action correctness and response efficiency. Code can be found in https://github.com/tu-tuing/SlowBA.

3D human avatars have shown impressive visual fidelitydriven by pose-conditioned models, yet they still lack the physical abil-ity required for interactions with each other and environments. Althoughrecent studies have made various attempts to incorporate physical char-acteristics into 3D avatars, they only exhibit limited physical deforma-tions, often leading to constrained interaction behaviors. To resolve thisissue, we present PIAvatar, a framework to simultaneously enable physi-cally aware interactions between avatar-avatar and avatar-environment,and a non-rigid deformable human body simulation. In this work, ourkey insight is to decouple kinematic velocity from deformation gradi-ent. When external forces act on avatars, the kinematic velocity inducesstress which hinders the avatar’s ability to achieve a desired pose. Inaddition, we integrate a skeletal framework within the avatar. It allowsestimating its poses and real-time tracking in a closed form, even duringnon-rigid physical interactions. Our approach is implemented within aconventional Material Point Method framework to ensure physically con-sistent dynamics. We lastly evaluate the method on both human-objectand human-human interaction scenarios to assess its behavior under di-verse interaction settings.

Recent advances in Streaming Video Understanding has enabled a new interaction paradigm where models respond proactively to user queries. Current proactive VideoLLMs rely on per-frame triggering decision making, which suffers from an efficiency-accuracy dilemma. We propose Em-Garde, a novel framework that decouples semantic understanding from streaming perception. At query time, the InstructionGuided Proposal Parser transforms user queries into structured, perceptually grounded visual proposals; during streaming, a Lightweight Proposal Matching Module performs efficient embedding-based matching to trigger responses. Experiments on StreamingBench and OVO-Bench demonstrate consistent improvements over prior models in proactive response accuracy and efficiency, validating an effective solution for proactive video understanding under strict computational constraints. Code and model are available at https://air-embodied-brain.github.io/ Em-Garde/.