ECCV 2026 论文列表

Virtual immunohistochemistry (IHC) multiplex staining hasemerged as a highly promising non-destructive solution in digital pathol-ogy. However, existing universal IHC staining models typically guide thegeneration of specific biomarkers by injecting task-specific text prompts,which often struggle to bridge the semantic gap with complex patholog-ical structures, leading to the "manifold drift" problem. Moreover, thesemodels lack specificity in modeling heterogeneous tissue components.In this paper, we propose HUSE, a Histocomponent-driven Universalmodel for virtual IHC multiplex Staining via joint manifold Evolution.HUSE introduces Joint Manifold Anchoring (JMA) as a global solution totreat H&E images as persistent structural scaffolds, effectively anchoringthe generation trajectory to the intrinsic image manifold. Complementingthis, a localized refinement scheme named Histocomponent-driven Mix-ture of Experts (Hi-MoE) and Representation Conflict Gating (RCG)is introduced to precisely capture local topological features of complexmanifolds. Furthermore, we pioneer an mIF-driven multi-marker IHCgeneration paradigm to construct large-scale 1(H&E) : N (IHC) train-ing matrices; evaluations by senior pathologists verify the feasibility ofthis paradigm as well as its exceptional clinical realism and structuralfidelity. Extensive experiments on two distinct types of datasets demon-strate that HUSE achieves state-of-the-art performance. Code is releasedat https://github.com/DeepMed-Lab-ECNU/HUSE.

Real-world collaborative perception agents typically deploydiverse sensor configurations and network architectures. While feature-level fusion methods address this heterogeneity by aligning feature spaces,they require collaborative training and large communication overhead. Incontrast, object-level fusion sidesteps these constraints, as transmittedbounding boxes are directly usable across varying architectures. How-ever, standard late fusion relies only on box coordinates and confidencescores, which provide insufficient information for robust association andleave the system vulnerable to localization noise and asynchronous de-lays. To address this, we propose HERO, an uncertainty-guided object-level collaborative perception method that equips proposals with evi-dential statistics and decouples the two fusion decisions, aggregating se-mantic evidence across matched proposals while selecting geometry fromthe most reliable source, to perform robust association and fusion underpose noise and asynchrony. Experiments on OPV2V-H and DAIR-V2Xshow that HERO matches the SOTA feature-level fusion methods whiletransmitting only about 2 KB per frame of sparse object proposals. Un-der localization error and latency scenarios, HERO further outperformsfeature-level SOTA models.

Hairstyle transfer has practical applications such as virtual try-on, yet remains challenging when the source and reference exhibit large head-pose discrepancies. We propose H-Adapter, which improves pose robustness by training with a region-specific loss that disentangles hair and non-hair objectives and thereby induces spatially disentangled cross-attention, from which a source-aligned hair edit mask is derived to guide diffusion-based inpainting. Experiments on pose-agnostic and posedifferent subsets demonstrate strong quantitative results, including the best FID, FID , and CLIP-I under pose differences, while maintaining CLIP competitive non-hair preservation and improving qualitative fidelity to fine-grained reference hairstyle details. Beyond source-conditioned transfer, H-Adapter supports practical extensions including reference-guided text-to-image generation, auxiliary prompt-based hair color control, and compatibility with an identity-preserving IP-Adapter variant. We also introduce a VLM-as-a-judge protocol and observe consistent gains in hairstyle faithfulness, non-hair preservation, and artifact quality.

Diffusion models have achieved remarkable success in image generation, yet their deployment remains constrained by the heavy computational cost and the need for numerous inference steps. Previous efforts on fewer-step distillation attempt to skip redundant steps by training compact student models, yet they often suffer from heavy retraining costs and degraded generalization. In this work, we take a different perspective: we accelerate smartly, not evenly, applying smaller speedups to early semantic stages and larger ones to later redundant phases. We instantiate this phase-aware strategy with two experts that specialize in slow and fast denoising phases. Surprisingly, instead of investing massive effort in retraining student models, we find that simply equipping the base model with lightweight LoRA adapters achieves both efficient acceleration and strong generalization. We refer to these two adapters as Slow-LoRA and Fast-LoRA. Through extensive experiments, our method achieves up to 5× acceleration over the base model while maintaining comparable visual quality across diverse benchmarks. Remarkably, the LoRA experts are trained with only 1 samples on a single V100 within one hour, yet the resulting models generalize strongly on unseen prompts. Code is available at https://zhuobaidong.github.io/Glance/.

Estimating category-level 9-DoF object poses (rotation, trans-lation, scale) is fundamental for robotic manipulation and autonomoussystems. Generative approaches, particularly diffusion-based models, haveachieved strong accuracy by effectively bridging the sim-to-real domaingap through large-scale synthetic training. However, it is difficult fordiffusion-based approaches to achieve both high accuracy and real-timeperformance due to their reliance on a multi-step iterative inference pro-cess. To address this limitation, we propose Geometric Prior GeodesicFlow Matching (G2FM), an efficient, simulation-free training objectivethat learns a deterministic vector field to directly map noise to pose dis-tributions. G2FM learns a vector field mapping noise to pose distribu-tions with only three ODE steps, achieving high accuracy and real-timeperformance (> 30 FPS). To avoid the structural distortion and singular-ities caused by Euclidean approximations, we construct a geometrically-consistent trajectory that operates directly on the SO(3) manifold bytransporting rotations along geodesic paths. Furthermore, to dynami-cally refine the generative flow during inference and enhance accuracywithout retraining, we design a training-free geometric guidance mech-anism by leveraging Chamfer distance gradients to project pose correc-tions directly onto the SO(3) tangent space. Extensive experiments showthat G2FM delivers a favorable accuracy-efficiency-generalization trade-off for category-level RGB-D 9-DoF pose estimation. Code will be re-leased upon publication.

Multi-object tracking (MOT) via UAV fleets is crucial for low-altitudeapplications. However, existing multi-view multi-object tracking (MVMOT) meth-ods and datasets are constrained by fixed camera configurations and limited cov-erage of complex scenarios, thereby failing to address real-world applications. Tofill this gap, we first conduct a systematic study of MOT under arbitrary camerasetups and introduce MDMOT, a novel multi-UAV benchmark covering diversereal-world scenarios, exposing practical challenges like random object entry/exitand cross-view appearance inconsistencies. We further propose FusionTrack, anend-to-end MVMOT framework that abandons the traditional decoupled track-then-associate paradigm. It jointly optimizes tracking and association via bidirec-tional object fusion between a Tracklet Memory Pool and a Trajectory IdentityPool, leveraging spatio-temporal context to enhance representation discriminabil-ity. For inference, we design View-aware Hierarchical Clustering with neighborfiltering to ensure intra-view exclusivity and inter-view consistency in cross-viewassociation. Extensive experiments verify that FusionTrack achieves state-of-the-art performance in both single- and multi-view tracking tasks, outperforming ex-isting methods by at least 2.3% in relative improvement on our MDMOT bench-mark. Project page: https://github.com/aircas501/FusionTrack

Self-supervised learning for imaging inverse problems is in-creasingly important in photon-limited settings, where acquiring cleanground truth is impractical and reconstruction must remain stable underdataset and acquisition shifts. This challenge is amplified under Poissonnoise, whose signal-dependent statistics interact with sampling operators(e.g., CFA mosaicing). Meanwhile, foundation vision encoders trained atweb scale offer distortion-invariant, content-related representations thatgeneralize well across domains, suggesting a promising route to buildpriors that transfer beyond the training distribution without expensivefine-tuning. This paper proposes an ADMM-inspired unrolled plug-and-play solver for Poisson inverse problems that decouples a closed-formdata-consistency update from a parameter-efficient prior. The prior isimplemented as a lightweight decoder operating on frozen CLIP RN50dense multi-scale features, adapting foundation representations with lesstrainable parameters. For self-supervision, the method integrates GR2Rmeasurement-domain re-corruption with an Equivariant Imaging regu-larizer via virtual acquisitions. Experiments on Poisson CFA demosaic-ing and deblurring show competitive quality, improved robustness undershifts, and self-supervised performance approaching supervised training.

Accurate process supervision remains a critical challenge forlong-horizon robotic manipulation. A primary bottleneck is that currentvideo MLLMs, trained primarily under a Supervised Fine-Tuning (SFT)paradigm, function as passive “Observers” that recognize ongoing eventsrather than evaluating the current state relative to the final task goal.In this paper, we introduce PRIMO R1 (Process Reasoning InducedMOnitoring), a 7B framework that transforms video MLLMs into ac-tive “Critics”. We leverage outcome-based Reinforcement Learning to in-centivize explicit Chain-of-Thought generation for progress estimation.Furthermore, our architecture constructs a structured temporal inputby explicitly anchoring the video sequence between initial and currentstate images. Supported by the proposed PRIMO Dataset and Bench-mark, extensive experiments across diverse in-domain environments andout-of-domain real-world humanoid scenarios demonstrate that PRIMOR1 achieves state-of-the-art performance. Quantitatively, our 7B modelachieves a 50% reduction in the mean absolute error of specialized rea-soning baselines, demonstrating significant relative accuracy improve-ments over 72B-scale general MLLMs. Furthermore, PRIMO R1 exhibitsstrong zero-shot generalization on difficult failure detection tasks. Weestablish state-of-the-art performance on the RoboFail benchmark with67.0% accuracy, surpassing closed-source models like OpenAI o1 6.0%.The project website is: 10-oasis-01.github.io/primo-r1-website.

Large Vision-Language Models (VLMs) typically process hun-dreds or even thousands of visual tokens per image or video frame, in-curring quadratic attention costs and significant information redundancy.Existing token-reduction methods either ignore the textual query or relyon deep attention maps, which become unstable under aggressive prun-ing and often lead to degraded semantic alignment. To address this, weintroduce FlashVLM, a text-guided visual token selection frameworkthat dynamically adapts visual inputs to the given query. Rather thanrelying on noisy attention weights, FlashVLM computes an explicit cross-modal similarity between projected image tokens and normalized textembeddings within the LLM space. It then fuses this extrinsic text-imagerelevance with intrinsic visual saliency using log-domain weighting andtemperature-controlled sharpening. Furthermore, a diversity-preservingpartitioning mechanism retains a minimal yet representative set of back-ground tokens to preserve global context. Evaluated under identical to-ken budgets and protocols, FlashVLM achieves beyond-lossless compres-sion—slightly surpassing the unpruned baseline while discarding up to77.8% of visual tokens on LLaVA-1.5. Remarkably, it maintains 92.8%performance even under an aggressive 94.4% compression rate. Extensiveexperiments across 14 image and video benchmarks demonstrate thatFlashVLM establishes a new state-of-the-art in efficiency-performancetrade-offs, exhibiting strong robustness and broad generalization acrossmainstream VLMs.

Learning-based image-to-point-cloud (I2P) registration hasgarnered increasing attention in recent years. Nevertheless, existing meth-ods still struggle with severe outliers under challenging scenarios withunseen, low-inlier, or distorted cases. A fast and robust 2D-3D correspon-dence pruning method is therefore highly desirable. Recently, a promis-ing scheme lifts 2D-3D correspondences to 3D-3D correspondences usingdepth priors, casting correspondence pruning as a Sim(3) registrationproblem. However, depth priors estimated from monocular images areinherently noisy, which undermines the reliability of this scheme. In thispaper, to explicitly model non-negligible depth noise, we reformulatecorrespondence pruning as an extended Sim(3) registration problem andpropose a simple yet effective pruning algorithm termed Ex-Sim(3)-Reg.We further provide a theoretical analysis to justify the effectiveness ofour method. Extensive experiments on the 7-Scenes, RGBD-V2, Scan-Net, and TUM datasets demonstrate that Ex-Sim(3)-Reg achieves up to24.7% improvement in registration recall over state-of-the-art baselinemethods. Code is released at github.com/anpei96/ex-sim3-demo.

We introduce EgoSim, a closed-loop egocentric world sim-ulator that generates spatially consistent interaction videos and per-sistently updates the underlying 3D scene state for continuous simu-lation. Existing egocentric simulators either lack explicit 3D ground-ing, causing structural drift under viewpoint changes, or treat the sceneas static, failing to update world states across multi-stage interactions.EgoSim addresses both limitations by modeling 3D scenes as updatableworld states. We generate embodiment interactions via a Geometry-action-aware Observation Simulation model, with spatial consistencyfrom an Interaction-aware State Updating module. To overcome the crit-ical data bottleneck posed by the difficulty in acquiring densely alignedscene–interaction training pairs, we design a scalable pipeline that ex-tracts static point clouds, camera trajectories, and embodiment actionsfrom in-the-wild large-scale monocular egocentric videos. Extensive ex-periments demonstrate that EgoSim significantly outperforms existingmethods in terms of visual quality, spatial consistency, and generalizationto complex scenes and in-the-wild dexterous interactions, while support-ing cross-embodiment transfer to robotic manipulation. Project page isat: egosimulator.github.io.

RGB-T detectors leverage the complementary strengths ofvisible and thermal infrared modalities, achieving robust performanceunder challenging conditions. Many of them resort to heavy dual back-bones and exhaustive cross-modality fusion across the entire image, lead-ing to impractically high computational costs. We observe that most im-age regions are smooth backgrounds (e.g., sky, ground) that can be easilyhandled by lightweight single-modality models. In light of this observa-tion, we propose a sparse fusion mechanism for efficient RGB-T detec-tion: first rapidly scanning the image to identify the proposals and thencarefully examining the remaining sparse proposals via feature fusion.We propose a two-stage framework to instantiate this mechanism, whichperforms detection in two stages: 1) a lightweight and modality-specificdetection stage that produces high-recall RoIs, and 2) a fusion-driven ex-amination and refinement stage that filters out the false positives and re-fines the bounding boxes. This design enables the detector to adaptivelyallocate more computational resources to the potential foregrounds, im-proving the efficiency while ensuring detection accuracy. Extensive ex-periments show that our method achieves competitive performance withsubstantially fewer parameters and lower cost, while maintaining strongscalability to high-resolution images.

Arbitrary slice super-resolution reconstructs isotropic vol-umes from anisotropic clinical acquisitions by synthesizing intermedi-ate slices at arbitrary scales. However, treating this ill-posed inverseproblem as unconstrained residual-based regression risks hallucinatinganatomically implausible structures or altering the originally observeddata. To address both concerns, this paper presents the Dual-PriorNull-Space Learning (DP-NSL) framework, which reformulates the taskas a constrained recovery process guided by two complementary priors.A Measurement-Consistent Projection (MCP) enforces a DeterministicObservation Prior : the reconstruction undergoes an exact orthogonalprojection that reproduces every acquired slice with zero error, confin-ing all learned details to the unobservable null space. Within this nullspace, a Mixture-of-Splines (MoS) module imposes a Geometric Conti-nuity Prior by dynamically mixing B-spline experts of different analyticorders, allowing each anatomical region to be modeled with a content-aware level of continuity. To promote spatial coherence, a Local SpatialConsistency Decoder (LSCD) further injects local inductive bias. Ex-periments on three CT and one MRI benchmark show that DP-NSLoutperforms existing approaches while strictly preserving measurementconsistency. Code is available at https://github.com/DeepMed-Lab-ECNU/Medical-Image-Reconstruction.

Reconstructing building wireframe from airborne LiDAR pointclouds yields a compact, topology-centric representation that enablesstructural understanding beyond dense meshes. Yet a key limitation per-sists: conventional methods have failed to achieve accurate wireframe re-construction in regions afflicted by significant noise, sparsity, or internalcorners. This failure stems from the inability to establish an adaptivesearch space to effectively leverage the rich 3D geometry of large, sparsebuilding point clouds. In this work, we address this challenge with De-launay Canopy, which utilizes the Delaunay graph as a geometric priorto define a geometrically adaptive search space. Central to our approachis Delaunay Graph Scoring, which not only reconstructs the underlyinggeometric manifold but also yields region-wise curvature signatures torobustly guide the reconstruction. Built on this foundation, our cornerand wire selection modules leverage the Delaunay-induced prior to focuson highly probable elements, thereby shaping the search space and en-abling accurate prediction even in previously intractable regions. Exten-sive experiments on the Building3D Tallinn city and entry-level datasetsdemonstrate state-of-the-art wireframe reconstruction, delivering accu-rate predictions across diverse and complex building geometries.

Multimodal deception detection aims to identify deceptivebehavior by analyzing audiovisual cues for forensics and security. Inthese high-stakes settings, investigators need verifiable evidence con-necting audiovisual cues to final decisions, along with reliable general-ization across domains and cultural contexts. However, existing bench-marks provide only binary labels without intermediate reasoning cues.Datasets are also small with limited scenario coverage, leading to short-cut learning. We address these issues through three contributions. First,we construct reasoning datasets by augmenting existing benchmarks withstructured cue-level descriptions and reasoning chains, enabling modelsto output auditable reports. Second, we release T4-Deception, a mul-ticultural dataset based on the unified “To Tell the Truth” televisionformat implemented across four countries. With 1695 samples, it is thelargest non-laboratory deception detection dataset. Third, we proposetwo modules for robust learning under small-data conditions. StabilizedIndividuality-Commonality Synergy (SICS) refines multimodal represen-tations by combining learnable global priors with sample-adaptive resid-uals and applying polarity-aware recalibration. Distilled Modality Con-sistency (DMC) aligns modality-specific predictions with the fused mul-timodal predictions via knowledge distillation to prevent unimodal short-cut learning. Experiments on three established benchmarks and our noveldataset demonstrate that our method achieves state-of-the-art perfor-mance in both in-domain and cross-domain scenarios, while exhibitingsuperior transferability across diverse cultural contexts. The datasets andcode are available at this link.

Credentials, such as corporate qualifications and personalidentification, serve as common forms of evidentiary materials. The forgeryof such credentials poses significant risks to security, making its detectiona long-standing area of widespread concern. With the advancement of AI-Generated Content technologies, forgery techniques can now largely pre-serve visual characteristics, making purely vision-based detection meth-ods insufficient to meet practical demands. Instead, a growing number offorgeries necessitate cognitive reasoning for detection. While the devel-opment of Multimodal Large Language Models (MLLMs) presents newopportunities for complex credential forgery detection, the scarcity ofdatasets and benchmarks severely impedes progress in this critical do-main. To bridge this gap, we introduce CogniCred — the first large-scale,multimodal dataset built upon complex, real-world credentials with care-fully crafted cognitive-level forgeries. Along with this dataset, we proposea dedicated benchmark, namely CogniCredBench, that evaluates modelperformance across five distinct types of cognitive forgeries. We conductcomprehensive evaluations across 21 mainstream MLLMs, revealing sub-stantial performance disparities and a strong bias toward text-based rea-soning, highlighting that current models still fall short in balanced multi-modal reasoning for cognitive forgery detection. Our datasets are publiclyavailable at https://huggingface.co/datasets/sjs0606/CogniCred.

Existing stylized motion generation models have shown theirremarkable ability to understand specific style information from the stylemotion, and insert it into the content motion. However, capturing intra-style diversity, where a single style should correspond to diverse motionvariations, remains a significant challenge. In this paper, we propose aclustering-based framework, ClusterStyle, to address this limitation.Instead of learning an unstructured embedding from each style motion,we leverage a set of prototypes to effectively model diverse style pat-terns across motions belonging to the same style category. We considertwo types of style diversity: global-level diversity among style motionsof the same category, and local-level diversity within the temporal dy-namics of motion sequences. These components jointly shape two struc-tured style embedding spaces, i.e., global and local, optimized via align-ment with non-learnable prototype anchors. Furthermore, we augmentthe pretrained text-to-motion generation model with the Stylistic Mod-ulation Adapter (SMA) to integrate the style features. Extensive ex-periments demonstrate that our approach outperforms existing state-of-the-art models in stylized motion generation and motion style transfer.Project page: https://1233chen.github.io/ClusterStyle/.

Transformer architectures have achieved strong performancein image denoising, but their computational and memory costs remainhigh. We revisit self-attention (SA) from a Bayesian perspective and showthat standard SA mainly exploits first-order inter-patch statistics, resem-bling a learnable non-local averaging scheme. However, this formulationneglects second-order intra-patch statistics, which are important for cap-turing local pixel dependencies. To address this, we introduce a BayesianSA formulation that jointly models first-order inter-patch and second-order intra-patch statistics. This leads to a lightweight denoising Trans-former, termed NLformer, featuring a dual-branch attention design andan efficient feed-forward module. Experiments on several benchmarksshow that NLformer outperforms existing lightweight denoising networksand substantially narrows the gap to full-size Transformer-based denois-ers while maintaining low model complexity.

Large vision-language models incur substantial inference costsbecause high-resolution inputs introduce thousands of visual tokens, manyof which are redundant for a given query. Existing pruning methods of-ten combine query relevance and token diversity, yet these objectivescan conflict under aggressive compression: relevance-driven selection mayoverconcentrate the budget on correlated local evidence, while diversity-driven selection may suppress indispensable tokens or retain distinctbut uninformative regions. We introduce AnchorPrune, a training-freeframework that first constructs a protected relevance anchor and thenexpands it with complementary visual context. AnchorPrune adaptivelydetermines the anchor size from the novelty profile of relevance-rankedtokens, preserving a compact set of query-critical evidence, and allocatesthe remaining budget through importance-weighted novelty to recover in-formative, non-redundant context relative to the anchor. This ordered de-sign prevents contextual expansion from displacing indispensable querycues while improving overall visual coverage. AnchorPrune is lightweight,architecture-aware, and requires neither retraining nor model modifica-tion. Across image and video vision-language models and benchmarks, itconsistently improves the accuracy–efficiency trade-off over training-freebaselines, particularly under severe compression. On LLaVA-NeXT-7B,AnchorPrune preserves 97.6% of full-token performance using only 160of 2,880 visual tokens. These results establish relevance-anchored contex-tual expansion as an effective principle for efficient multimodal inference.Code is available at https://github.com/MULTI-cau/AnchorPrune.