ECCV 2026 论文列表

Video understanding requires models to continuously track and update world state during playback. Although existing benchmarks have advanced video understanding evaluation across multiple dimensions, they provide limited visibility into how models maintain world state over time. We propose SVCBench, a Streaming Video Counting Benchmark that repositions counting as a minimal, controlled probe for diagnosing models’ world-state maintenance capability. We decompose this capability into object counting and event counting, forming 8 fine-grained subcategories. Object counting covers tracking currently visible objects and cumulative unique identities, while event counting covers detecting instantaneous actions and tracking complete activity cycles. SVCBench contains 406 videos with frame-by-frame annotations of 10,071 event occurrences and object state changes, yielding 1,000 streaming QA pairs with 4,576 query points distributed along video timelines. By observing state maintenance trajectories through streaming multipoint queries, we design three complementary metrics to diagnose numerical precision, trajectory consistency, and temporal awareness. Evaluations of mainstream video-language models show that current models still exhibit significant deficiencies in spatial-temporal state maintenance, with especially poor performance on periodic event counting. SVCBench provides a diagnostic framework for measuring and improving state maintenance in video understanding systems. Our code and data are available at https://buaa-colalab.github.io/SVCBench.

State space models, especially Visual State Space Duality (VSSD), have emerged as efficient linear-time alternatives to Transformers for dense visual tasks. However, we observe that VSSD compresses spatial context into a global aggregation that suppresses high-frequency responses, causing excessive boundary smoothing in remote sensing semantic segmentation. To address this, we propose CRISP, a calibration framework with two components. Its core, the Duality Calibration Operator (DCO), restores local contrast and boundary responses through residual injection and frequency calibration within the VSSD backbone, without altering its linear complexity. To retain the recovered detail, an Orthogonal Multi-Prototype (OMP) head assigns multiple orthogonally constrained prototypes per class to model large intra-class variance. Extensive experiments on Potsdam, Vaihingen, and LoveDA show that, with only ∼30M parameters, CRISP achieves consistent gains in mean F1 (mF) and mIoU while remaining competitive with state-of-the-art methods. Code is available at https://github.com/crazylifeha/crisp.

Natural images exhibit strong geometric regularities: local structures, such as edges, corners, and textures, appear in many orientations and mirror configurations. Since Vision Transformers (ViTs) operate on square image patches, these transformations naturally correspond to the dihedral symmetry group D 8, also known as the octic group. Recent work has shown that ViTs can be made reflection equivariant and more e!cient than standard ViTs simultaneously by implementing the linear layers in the Fourier domain of the reflection group. In this work, we extend the equivariance to reflections and rotations and analyze the scalability of the resulting networks. Our Quick ViTs, based on octic equivariant linear layers, achieve 5.33x reductions in FLOPs and up to 8x reductions in memory compared to ordinary linear layers. By analyzing the arithmetic intensity of these layers, we identify theoretical limits on how much the FLOP savings translate into throughput improvements on modern GPUs. However, these limitations disappear as the embedding dimensions increase. Enabled by their computational e!ciency, we conduct a broader empirical evaluation of equivariant ViTs than in previous work. Upon training supervised (DeiT-III) and self-supervised (DINOv2) on ImageNet-1K, we find that our Quick ViTs match or exceed baseline accuracy while at the same time providing substantial e!ciency gains.

4D scene synthesis from monocular videos has made significant progress in recent years. However, existing methods are typically constrained by view interpolation. As a result, they struggle to infer unseen regions beyond the observed views. In this paper, we reformulate the task as 4D scene synthesis with unseen regions, which extends beyond traditional interpolation settings. Camera-conditioned video generation enables unseen region synthesis by guiding generation along specified cameras. However, these methods lack explicit 3D priors and are optimized with random camera trajectories. This design leads to severe inconsistencies under large trajectory deviations. To address this limitation, we build a unified training and inference framework with panoramic trajectory guidance. While this design improves cross-view consistency, the panoramic representation alone fails to model dynamic content effectively. Object motion in panoramic space introduces scale and shape distortions. To address this, we propose PanoGaussian, a unified PanoramicGaussian representation that distills the panoramic representation into an explicit dynamic Gaussian representation to capture dynamic physical priors of the 4D scene. Experiments demonstrate that PanoGaussian achieves consistent 4D scene synthesis even under large viewpoint variations.

While prior research on text-to-image generation has predominantly focused on biases in human depictions, demographic bias in generated objects remains relatively underexplored. We introduce SODA (Stereotyped Object Diagnostic Audit)1, a novel framework for systematically measuring these biases through automated attribute discovery and three standardized metrics: Base vs. Demographic Divergence (BDS), Cross-Demographic Disparity (CDS), and Visual Attribute Concentration (VAC). Applying SODA to 8,000 images across five state-of-the-art models and eight object categories (e.g., cars), we find that “neutral” prompts produce outputs most visually similar to middle-aged and White people, suggesting these groups are implicitly over-represented in model defaults. Furthermore, demographic cues trigger highly skewed stereotypical outputs: 26.6% of object-model-demographic combinations produce results where all 20 generated images share the exact same attribute value (e.g., rose gold laptops for women). Finally, prompt-level debiasing reduces inter-group disparity but paradoxically collapses within-group diversity, replacing one stereotype with another. SODA offers a practical pipeline for making these implicit associations measurable, serving as a step toward more responsible AI development.

This paper addresses the problem of monocular metric depth estimation in aerial UAV imagery. Although recent data-driven methods have achieved remarkable progress in ground-level scenarios, models trained primarily on street-view and indoor datasets exhibit significant domain gaps when applied to aerial viewpoints. To tackle these challenges, we introduce AerialMetric, a benchmark dataset designed to evaluate and facilitate the adaptation of monocular metric depth estimation under UAV aerial viewpoints. The dataset consists of four complementary subsets collected from different sources, jointly covering real-world photogrammetry data, controlled aerial acquisition settings, photorealistic synthetic scenes, and in-the-wild Internet imagery. In total, AerialMetric provides 52K real-world and 16K synthetic image–depth pairs with reliable metric ground truth. Based on this dataset, we conduct systematic evaluations of existing state-of-the-art models under aerial settings and investigate the impact of viewpoint, altitude, and camera parameters on metric depth prediction. In addition, by fine-tuning representative metric depth model on our dataset, we establish a comprehensive aerial benchmark and achieve state-of-the-art performance across diverse aerial imagery. Our dataset, code, and model weight are publicly available at https://kuieless.github.io/AerialMetric-ECCV2026-page/.

Recent large multimodal models (LMMs) have made rapid progress on visual grounding, document understanding, and diagram reasoning tasks. However, their ability to convert Printed Circuit Board (PCB) schematic diagrams into machine-readable, spatially weighted netlist graphs that capture component attributes, connectivity, and geometry remains largely underexplored, even though such graph representations are the backbone of practical electronic design automation (EDA) workflows. To bridge this gap, we introduce OmniSch, the first comprehensive benchmark designed to assess LMMs on schematic understanding and spatial netlist graph construction. OmniSch contains 1,854 real-world schematic diagrams and supports four tasks: (1) visual grounding for schematic entities, with 109.9K grounded instances aligning 423.4K diagram semantic labels to their visual regions; (2) diagram-to-graph reasoning, which requires understanding topological relationships among diagram elements; (3) geometric reasoning, constructing layout-dependent edge weights for each connection; and (4) tool-augmented agentic reasoning for visual search, which invokes external tools to accomplish (1) - (3). Our results reveal substantial gaps in current LMMs’ ability to interpret schematic engineering artifacts, including unreliable fine-grained grounding, brittle layout-to-graph parsing, inconsistent global connectivity reasoning, and inefficient visual exploration.

Novel view synthesis under trajectory changes is essential for autonomous driving simulation, yet existing methods struggle to generate consistent videos when extrapolating to unseen viewpoints such as lane shifts. Reconstruction-based approaches maintain geometric consistency but degrade in visual quality under large viewpoint changes, while reconstruct-then-restore methods often produce visually plausible frames yet suffer from geometric and temporal inconsistencies. We present GeoV2V, a geometry-grounded video-to-video diffusion framework that synthesizes temporally coherent driving videos along novel trajectories by integrating LiDAR and depth-derived geometric priors with full-video conditioning in the Wan 2.1 backbone. A key challenge in learning cross-lane transformations is the lack of synchronized multitrajectory supervision in real-world datasets. To address this, we also introduce Para4D, a synthetic dataset capturing synchronized multi-lane driving videos within dynamic scenes, which provides explicit supervision for lane-shift generation and serves as a strong training prior for geometry-consistent view synthesis. Experiments on Waymo, nuScenes, and Para4D demonstrate state-of-the-art performance in visual quality, geometric consistency, and robustness to complex lighting dynamics under large trajectory shifts. Project page: https://xiyuche.github.io/ GeoV2V/

Most recent advances in 3D generative modeling rely on diffusion or flow-matching formulations. We instead explore a fully autoregressive alternative and introduce GaussianGPT, a transformer-based model that directly generates 3D Gaussians via next-token prediction, thus facilitating full 3D scene generation. We first compress Gaussian primitives into a discrete latent grid using a sparse 3D convolutional autoencoder with vector quantization. The resulting tokens are serialized and modeled using a causal transformer with 3D rotary positional embedding, enabling sequential generation of spatial structure and appearance. Unlike diffusion-based methods that refine scenes holistically, our formulation constructs scenes step-by-step, naturally supporting completion, outpainting, controllable sampling via temperature, and flexible generation horizons. This formulation leverages the compositional inductive biases and scalability of autoregressive modeling while operating on explicit representations compatible with modern neural rendering pipelines, positioning autoregressive transformers as a complementary paradigm for controllable and context-aware 3D generation.

Self-supervised novel view synthesis (NVS) remains challenging to scale, despite the abundance of video data, largely due to the brittleness of training on realistic videos and the hard-to-predict scaling behavior of multi-network system designs. We introduce RayDer, a unified, feed-forward transformer that consolidates camera estimation, scene reconstruction, and rendering into a single backbone, turning self-supervised NVS into a well-posed single-model scaling problem. A minimal dynamic state, treated as a nuisance factor, absorbs time-varying content and enables stable training on unconstrained real-world video. Importantly, RayDer keeps static-scene NVS as its target task: dynamic content is leveraged purely as scalable supervision, not reconstructed as in dynamicscene (4D) NVS. Across multiple model sizes and orders of magnitude in data, RayDer exhibits clean power-law scaling with data and compute, and outperforms static-scene data mixtures. On a large number of benchmarks, RayDer achieves strong zero-shot open-set performance competitive with state-of-the-art supervised approaches. Project Page: https://compvis.github.io/rayder.

Video World Models are interactive video generation modelsthat predict future world states based on user actions and history videoframes. A critical challenge in video world models is the lack of memory,causing inconsistent generated scenes over extended durations. Previousmethods explored rule-based context frame retrieval as memory, but theyfail to generalize in scenarios with scene occlusions and dynamic objects.We propose MemLearner, a learning-based adaptive context querymethod using query tokens to bridge context and predicted tokens. Byleveraging the video generation model itself for context querying, Mem-Learner exploits pre-trained visual priors without training additionalmodules from scratch, and incorporates e_x001E_cient strategies for trainingand inference. We collect a dataset of long videos with scene occlusionsand dynamic objects, paired with camera pose annotations, and proposea multi-dataset training strategy leveraging both annotated renderedand unannotated real-world videos. Extensive experiments demonstratethat MemLearner signi_x001C_cantly outperforms prior video world modelsin terms of scene consistency and memory, particularly under challengingocclusion and dynamic scenarios.

Existing referring image segmentation (RIS) models passively process static images captured from fixed perspectives, limiting their applicability in Embodied AI, where agents must perform active perception in the continuous 360◦ environments. To bridge this gap, we introduce a novel task: Active Panoramic Referring Segmentation (APRS). In this setting, an agent is required to adjust its viewing direction (∆θ,∆ϕ) to explore the 360◦ environment, seeking the object specified by a user instruction for segmentation. To tackle this challenging task, we propose PanoSeeker, a memory-augmented agent for efficient APRS. Rather than relying on heuristic scanning, PanoSeeker integrates a Vision-Language Model (VLM) with EgoSphere—an explicit spatial visual memory. By progressively integrating sequential local observations into a unified 360◦ representation, EgoSphere enables the agent to plan efficient and non-redundant search trajectories. Once the target is found, the agent performs active viewpoint alignment and outputs the segmentation mask. Furthermore, we curate an expert-annotated search trajectory dataset with memory timelines for Supervised FineTuning, followed by Reinforcement Learning post-training to explicitly optimize PanoSeeker’s exploration efficiency. Extensive experiments on our newly established APRS benchmark demonstrate that PanoSeeker achieves superior search efficiency and segmentation accuracy, significantly outperforming adapted state-of-the-art baselines.

Multi-object tracking degrades in crowded scenes. The reliability of geometry and appearance shifts under occlusion, uniform appearance, and fast motion. Consequently, association scores suffer from mode-dependent miscalibration. We observe that these failures concentrate into a small set of recurring error modes. This pattern renders single global thresholds ineffective. To address this issue, we propose a diagnose-then-correct strategy via Mode-Conditioned Residual Calibration (MCRC). It is a lightweight auxiliary module that identifies the dominant failure mode from pairwise and set-level geometric cues. The module then applies a mode-conditioned residual correction to the association cost matrix. For highly ambiguous tracking pairs, MCRC selectively activates an uncertainty-gated semantic refiner. This refiner incorporates simple acceptance checks to ensure representation stability, thereby avoiding global computational overhead. Extensive evaluations on DanceTrack, SportsMOT, MOT17, and MOT20 confirm its effectiveness. MCRC consistently improves identity preservation when inserted into diverse trackers, maintaining minimal added overhead.

Structured Light (SL) is a popular technique that delivers precise 3D shape acquisition across diverse materials and object sizes. SL requires, however, careful modeling of the set-up, including geometric and radiometric calibration of the camera-projector pair; without it, reconstruction quality degrades. We propose NeLU3D, a neural inverse SL method without explicitly modeling the projector. We calibrate the camera–projector system using a neural network that maps 3D positions to a set of projected patterns. Then, with as few as four monochromatic images (or two RGB images), our approach uses differentiable volume rendering to fit a surface to match SL captures. We scan over twenty-five objects of different shapes and reflectances to demonstrate the feasibility and quality of our method in a handful of projector-camera set-ups, including an extremely low-cost projector and an analog projector with fixed RGB pattern. We also showcase sub-millimeter accuracy with suboptimal patterns, where previous methods recover noisy 3D surfaces. We release an open-source implementation at https://github.com/ geometryprocessing/neural-lookup.

Recent advances in video diffusion models have remarkably improved camera-controlled video generation, but most methods rely solely on supervised fine-tuning (SFT), leaving online reinforcement learning (RL) post-training largely underexplored. In this work, we introduce an online RL post-training framework that optimizes a pretrained video generator for precise camera control. To make RL effective in this setting, we design a verifiable geometry reward that delivers dense segment-level feedback to guide model optimization. Specifically, we estimate the 3D camera trajectories for both generated and reference videos, divide each trajectory into short segments, and compute segment-wise relative poses. The reward function then compares each generated-reference segment pair and assigns an alignment score as the reward signal, which helps alleviate reward sparsity and improve optimization efficiency. Moreover, we construct a comprehensive dataset featuring diverse large-amplitude camera motions and scenes with varied subject dynamics. Extensive experiments show that our online RL post-training clearly outperforms SFT baselines across multiple aspects, including camera-control accuracy, geometric consistency, and visual quality, demonstrating its superiority in advancing camera-controlled video generation.

Human motion forecasters are increasingly accurate and fast, but reliable deployment requires uncertainty estimates that are structured, calibrated, and efficient. Bayesian and ensemble-based uncertainty estimates often require repeated stochastic inference [15,26], while conformal calibration alone does not provide an epistemic signal or preserve trajectory covariance structure [14,50]. We introduce SPARC (SinglePass Adaptive Risk Calibration), a Bayesian–conformal uncertainty layer for motion forecasting. A deterministic MLP backbone predicts the future mean, and a conjugate Bayesian last layer converts time-domain feature leverage into an analytic horizon-wise epistemic scale κ (x). This t scale inflates a graph-temporal Gaussian covariance without changing its correlation structure, and split conformal calibration produces 95% marginal prediction tubes with finite-sample validity under exchangeability. The key interface is the structured factorization κ (x)Σ (x), t str,t which injects feature-space epistemic uncertainty into trajectory densities without Monte Carlo sampling. Across nine dataset/protocol blocks and deterministic, multimodal, and calibration baselines, SPARC ranks first on NLL and on the combined MPJPE+NLL criterion while retaining competitive point accuracy and efficient calibrated tubes. Ranking windows by κ separates high-error cases, making the scale usable as a lightweight risk monitor.

Cross-view geo-localization is challenging due to drastic viewpoint changes and large appearance discrepancies between street-level and satellite imagery. Although existing methods often use geometric warping to expose co-visible cues, such transformations rely on restrictive spatial assumptions and inevitably introduce severe visual distortions under view-dependent visibility, yielding noisy supervision and fragile correspondences. To overcome this, we propose a novel joint-view consensus-guided learning framework that entirely bypasses explicit geometric warping. Instead of forcing rigid spatial alignment, we dynamically mine and adaptively strengthen a semantic consensus directly within the feature space. Specifically, an auxiliary joint-view pathway during training enables direct cross-view interaction, allowing each view to selectively aggregate corroborative evidence into a unified consensus representation. To resolve feature heterogeneity among the singleand joint-view streams, we introduce global pattern probes acting as a semantic dictionary to project divergent modalities into a strictly aligned metric space. Guided by a consensus-mediated contrastive objective, single-view embeddings are explicitly pulled toward the joint-view anchor during training, distilling this consensus-mining capability into the single-view encoders for robust retrieval at inference. Extensive experiments demonstrate that our method achieves state-of-the-art performance across four standard benchmarks, underscoring the importance of discovering cross-view semantic consensus for reliable geo-localization.

Photorealistic 3D face avatars are increasingly deployed asreusable digital assets across applications such as telepresence, anima-tion, and personalized media. At the same time, vision–language models(VLMs) can infer sensitive attributes from rendered images with open-ended semantic reasoning without any fine-tuning. This creates a newprivacy challenge: once a 3D face avatar is shared, any of its renderingscan be analyzed to extract high-level facial attributes. Existing defenseslargely operate in 2D image space and do not address identity-preservingsemantic manipulation of 3D facial representations.We propose 3D FaceShell, a framework for steering VLM interpreta-tions of faces rendered from 3D models while preserving geometric fidelityand facial identity. 3D FaceShell augments the original 3D representa-tion with a learnable Gaussian shell that produces subtle, spatially dis-tributed perturbations optimized through multi-view embedding align-ment. The perturbations are designed to be visually inconspicuous yetsufficient to redirect VLM-based attribute inference in a view-consistentExtensive experiments on reconstructed celebrity face avatars and mul-tiple black-box VLMs demonstrate that 3D FaceShell significantly in-creases attribute injection and mismatch rates while maintaining highperceptual similarity and identity consistency. Our results show that itis possible to manipulate VLM-level semantic interpretation of 3D faceswithout compromising their human-recognizable appearance.

A fundamental challenge in 3D reconstruction and robotic localization is co-visibility: determining which image pairs share overlapping visible surfaces, particularly in scenarios with minimal overlap. We demonstrate that VGGT implicitly encodes co-visibility as an emergent behavior: without any supervision for this task, its internal representations exhibit a clear hierarchical structure mirroring that of large language models — early layers build a 3D-aware scene representation, while late layers act as dedicated co-visibility reasoners. In particular, we identify layer L17 as a negative anchor that consistently routes nonco-visible pairs for this backbone, regardless of the evaluation setting, providing task-grounded evidence of layer specialization in a geometrygrounded foundation model. Building on this, we introduce Co-VGGT, which freezes VGGT and trains only a lightweight layer-wise mixtureof-experts head (∼7.5M parameters) to classify co-visibility from RGB alone, treating each layer as a specialized expert whose geometric abstraction is adaptively weighted per input pair. On the Co-VisiON benchmark, Co-VGGT surpasses the human annotation baseline and improves over prior work by more than 25% pairwise and 10% multiview. Pairwise predictions are well-calibrated (ECE=0.030), enabling direct use as edge weights in visibility graphs for downstream SfM and SLAM pipelines without post-hoc correction. Code and data available1.