ECCV 2026 Papers

Pathology whole slide image (WSI) analysis is crucial fordisease diagnosis and understanding. While pathology foundation modelshave advanced WSI analysis, their scalability remains limited by thelarge resolution of individual WSIs and the small number of availableslides. Existing methods often use the [CLS] token from tile-level ViTsfor efficiency, but this design overlooks spatially informative tokens thatencode fine-grained pathological patterns. Incorporating all spatial tokenscould enrich representation but is computationally prohibitive, revealinga fundamental trade-off between efficiency and representational richness.We observe high redundancy among spatial tokens, suggesting that se-lectively retaining the most informative ones can better balance thesegoals. To address this, we propose TokRet, a token retention and com-pression module that preserves the most informative spatial tokens whileeliminating redundancy. Built upon TokRet, we develop PathTokScale, aslide-level self-supervised pretraining framework that leverages efficientlong-contextual transformer modeling to scale training to the billion-tokenlevel—over 20× larger than previous WSI studies—without requiring mas-sive WSI collections. Experiments show that PathTokScale consistentlyoutperforms existing pathology foundation models across diagnostic, prog-nostic, and biomarker prediction tasks. Remarkably, using only 10% ofthe number of WSIs as used by TITAN, it achieves comparable per-formance through efficient token scaling, paving the way for scalable,high-performing pathology foundation models.

Dense 3D semantic occupancy perception is critical for mo-bile robots operating in pedestrian-rich environments, yet it remains un-derexplored compared to its application in autonomous driving. To ad-dress this gap, we present MobileOcc, a semantic occupancy dataset formobile robots operating in crowded human environments. Our dataset isbuilt using an annotation pipeline that incorporates static object occu-pancy annotations and a novel mesh optimization framework explicitlydesigned for human occupancy modeling. It reconstructs deformable hu-man geometry from 2D images, then refines and optimizes it using as-sociated LiDAR point data. Using MobileOcc, we establish benchmarksfor two tasks: i) Occupancy prediction and ii) Pedestrian velocity predic-tion, using different methods, including monocular, stereo, and panopticoccupancy, with metrics and baseline implementations for reproduciblecomparison. Beyond occupancy prediction, we further assess our annota-tion method on 3D human pose estimation datasets. Results demonstratethat our method exhibits robust performance across different datasets.Our code and dataset are released at https://autonomousrobots.nl/paper_websites/mobileocc

Reconstructing high-fidelity 3D scenes from sparse views re-mains a central problem in generalizable neural rendering. Existing gen-eralizable 3D Gaussian Splatting (3DGS) methods often exhibit geomet-ric artifacts in sparse-view settings, since supervision based solely on 2Dphotometric losses cannot resolve depth and correspondence ambigui-ties. To address this issue, we propose MAC-Splat, a training frameworkbuilt around direct 3D consistency supervision. MAC-Splat builds onthe MASt3R geometric backbone and a frozen DINOv3 encoder to ob-tain semantically informed 2D correspondences, which serve as geometricanchors for 3D supervision. Using these anchors, we define the Multi-Attribute Consistency (MAC) loss. This objective jointly regularizes the3D attributes of matched Gaussians, including their position, shape,and appearance, by enforcing agreement in a common world coordinateframe. The formulation is robust to outliers and respects the geometry ofcovariance matrices, which leads to stable training under sparse-view con-ditions. Experiments on ScanNet++ show that MAC-Splat outperformsstrong baselines, with particularly large gains under different overlapregimes. In particular, it improves average PSNR over Splatt3R by morethan 4.5 dB, reduces LPIPS, and maintains performance as the camerapose gap increases. These results indicate that a direct, multi-attribute3D consistency objective, when combined with high-quality correspon-dences, is effective for addressing the ill-posed sparse-view reconstructionproblem.

Multi-modal Large Language Models have demonstrated im-pressive capabilities in vision-language tasks, yet they remain highly sus-ceptible to hallucinations, in which generated text diverges from the ac-tual visual evidence. Current mitigation strategies, such as self-refinementand specialized decoding, often incur significant latency due to multi-round iterations or reliance on high-performance auxiliary models, limit-ing their real-time practical utility. In this paper, we systematically ana-lyze the hidden states of MLLMs across the entire inference pipeline andreveal that hallucination-inducing redundant features persist not onlyin image tokens during the prefill stage but also propagate into Imageand Answer Value Caches during decoding. Motivated by the observationthat Gaussian noise injection can effectively suppress these redundancies,we propose HalMask (Hallucination Feature Masking), a lightweightand real-time self-correction framework. HalMask employs specializedImage and Text Noise Modulation Modules, optimized via a dynamic reg-ularization term, to accurately discern and mask hallucination-triggeringfeatures. During inference, it applies a dynamic Value Cache maskingmechanism that continuously filters redundant information across bothgeneration phases without disrupting critical semantic cues. Extensiveexperiments on multiple benchmarks demonstrate that HalMask signifi-cantly reduces hallucinations while introducing only marginal computa-tional overhead.

At the heart of composed visual data retrieval is the fusionof a reference visual input and a textual modification into a single query.While current state-of-the-art methods utilize multimodal large languagemodels for this fusion, their complexity introduces prohibitive query-time latency, limiting their scalability. We instead revisit the efficacy ofsimple linear interpolation within an embedding space, and introduceSRAIN, the first framework that dynamically predicts query-specific in-terpolation weights. The key challenge lies in the fact that the qualityof an interpolation weight should be measured by the interpolated em-bedding’s discriminability from negatives as well as its proximity to truetargets; this makes collecting and predicting optimal weights intractable.We overcome this bottleneck through two key innovations: batch-wiserank-aware weight estimation during training, and a compact memorybank that synthesizes hard negatives during inference. SRAIN achievesthe best in composed video retrieval and matches the current state of theart in composed image retrieval, all while substantially reducing query-time latency compared to MLLM-based alternatives.

Multiple small object tracking (MSOT) is critical for ap-plications such as anti-UAV systems and security surveillance, yet tra-ditional frame-based cameras struggle to track fast-moving small ob-jects in complex environments due to their low frame rates and lim-ited dynamic range. Event cameras, with their ability to continuouslyrecord subtle brightness changes, can naturally overcome these limita-tions. However, most existing event-based tracking methods follow a“convert-then-detect-and-track” pipeline. This pipeline sacrifices the in-herent high temporal resolution of event data, leading to fragmentedtrajectories of fast-moving objects. Moreover, it introduces significantbackground redundancy during framing, which reduces computational ef-ficiency. To handle these issues, we introduce “instance segmentationas tracking”, a novel paradigm that formulates event-based MSOT asan instance segmentation task in the spatio-temporal dimension. Follow-ing this paradigm, we first design Ev-ISNet, which leverages 3D sparseconvolutions to extract per-voxel features while simultaneously predict-ing object confidence and motion direction. Then we construct an eventgraph and progressively cluster trajectory instances by predicting edge-wise instance affinities. By leveraging intra-graph and inter-graph associ-ation modules, our method achieves highly efficient streaming inference.To address the lack of large-scale benchmarks for event-based MSOT,we build EV-UAV-Track, a comprehensive dataset featuring per-eventinstance-level annotations. Extensive experiments demonstrate that Ev-ISNet consistently outperforms state-of-the-art MSOT methods, achiev-ing over 30% improvement in MOTA score compared to frame-basedtracking methods.

Di!usion and flow-matching models dominate conditionalimage generation, yet inference-time scaling for these models is far lessdeveloped than for autoregressive language models. Because final qual-ity is highly sensitive to the initial noise seed, many approaches spendextra compute on seed search or resampling under a black-box reward,but typically maintaining a constant memory footprint throughout in-ference. We show that relaxing this constraint enables an underexploredinference-time scaling axis: by front-loading exploration, evaluating manyseeds early, and pruning aggressively, we can use a fixed compute budgetmore e!ectively. Progressive Seed Pruning (PSP) scores intermediate de-noised estimates and progressively narrows the candidate set so that onlypromising trajectories are fully denoised, while keeping the total num-ber of model evaluations fixed. Across di!usion and flow-matching back-bones, PSP consistently improves reward-guided selection and achieveshigher GenEval scores (automated) and better human evaluation onprompt-alignment than best-of-N , importance-sampling, and tree-searchbaselines at matched compute. Project page: vision.caltech.edu/psp.Code: github.com/rogerioagjr/psp

Human body parsing datasets and benchmarks define partlabels for intact anatomies and focus on body parts and clothing. Widelyused datasets implicitly assume intact anatomies and therefore do not in-clude residual-limb or prosthetic classes. This leaves people with limb de-ficiencies excluded from body parsing model training and standard evalu-ation, creating a fairness gap. To the best of our knowledge, InclusiveHuman-10K (IH-10K) is the first human parsing benchmark centered on in-dividuals with limb deficiencies, with explicit labels for residual limbs,prostheses, and mobility assistive devices. Specifically, IH-10K containsmore than 10k images and around 15k person instances from everyday,sports, and clinical scenes, each with pixel-wise body parsing masks. Thisbenchmark can support e-commerce, human-machine interaction, reha-bilitation and assistive technology, human-centric image understanding,and editing. Moreover, we extend conventional body parsing categorieswith explicit classes for residual limbs, prostheses, and mobility assis-tive devices. On our IH-10K, body parsing models can be trained andevaluated on disability-related regions for this population with limb de-ficiencies, instead of treating them as background or noise. To assesswhether disability-related regions are systematically underserved, we re-port group-wise evaluation metrics across disability-related and commonclasses. Our experiments show that supervised models struggle on theseregions due to large appearance variation and broken limb structure.Meanwhile, zero-shot methods perform poorly, suggesting that these con-cepts are absent from previous benchmarks and under-covered in open-vocabulary pretraining. This further highlights the representational andmeasurement gaps and signifies the necessity of our IH-10K benchmark.We hope this work helps build vision systems that serve people with limbdeficiencies more reliably and fairly, increasing the likelihood of mean-ingful social benefit.

This work introduces and demonstrates the first system capable of imaging fast-moving extended non-rigid objects through strong atmospheric turbulence at high frame rate. Event cameras are a novel sensing architecture capable of estimating high-speed imagery at thousands of frames per second. However, on their own, event cameras are unable to disambiguate scene motion from turbulence. In this work, we overcome this limitation using event-based light field cameras: By simultaneously capturing multiple views of a scene, event-based light field cameras and machine learning-based reconstruction algorithms are able to disambiguate motion-induced dynamics, which produce events that are strongly correlated across views, from turbulence-induced dynamics, which produce events that are weakly correlated across view. Tabletop experiments demonstrate event-based light field can overcome strong turbulence while imaging high-speed objects traveling at up to 16,000 pixels per second.

Aerial-Ground Person Re-IDentification (AG-ReID) aims toretrieve the same person across heterogeneous aerial and ground cameraplatforms. Although great progress, existing methods remain subopti-mal due to the direct feature alignment across views, overlooking view-specific cues. To address this issue, we propose a novel Hierarchical Hy-perbolic Representation (HiHR) framework for AG-ReID. More specif-ically, we first extract multi-granularity features based on pre-trainedvisual-text encoders. Then, we propose a Text-guided Multi-granularityFusion (TMF) to fuse multi-granularity features and enhance the repre-sentation ability of identity features. Furthermore, we introduce the Hi-erarchical Hyperbolic Learning (HHL) to construct a hierarchical featurestructure in a hyperbolic space. This hierarchy includes a coarse level thatensures identity separability and cross-view consistency, and a fine levelthat preserves view-specific discriminative cues. A a result, our proposedframework can effectively aggregate view-invariant and view-specific dis-criminative features for AG-ReID. Extensive experiments on four AG-ReID benchmarks demonstrate the effectiveness of our framework. Thesource code is available at https://github.com/YangQiWei3/HiHR.

Pre-trained image restoration models often fail on out-of-distribution (OOD) real-world degradations. Adapting to these domainsis challenging as real-world data lacks paired ground truth, and unsuper-vised methods often require unstable architectural changes. We proposeGenerative Manifold Distillation (GMD), which reframes domain adapta-tion as geometric manifold alignment. GMD operates in a strictly unpairedsetting, requiring only low-quality (LQ) target observations. By leveragingthe flow-matching dynamics of a frozen text-to-image foundation model,GMD projects off-manifold restorations onto the natural image manifoldto generate high-quality pseudo-targets. To ensure stability, a quality-gated manifold filter rejects off-manifold samples, while source-anchoredtrajectory regularization prevents error accumulation. Ultimately, GMDdistills a powerful generative prior into an efficient restoration network.Experiments demonstrate that GMD seamlessly adapts to new distribu-tions using only LQ inputs, drastically improving perceptual quality withzero architectural modifications or added inference latency.

Reasoning segmentation requires localizing targets based oncomplex, implicit queries. Current end-to-end models typically entangleperception and deduction into an opaque black box, severely limitinginterpretability and scalability. To address this, we propose GEAR-Seg(Grounded Explainable Agent for Reasoning Segmentation), an explic-itly decoupled agent that shifts the paradigm by translating visual pixelsinto dense, attribute-rich text. By decoupling class-agnostic segmenta-tion, semantic description, and Large Language Model (LLM) deduc-tion, GEAR-Seg transforms implicit reasoning into an explicit, track-able logic chain. As a zero-shot inference framework, it achieves highlycompetitive performance across diverse reasoning and fine-grained re-ferring segmentation benchmarks. Furthermore, GEAR-Seg inherentlyfunctions as a highly scalable data engine. Utilizing this engine, we con-struct GEAR-131K, a massive benchmark (over 38k images, 656k QA-mask pairs) introducing a multifaceted taxonomy tailored for complexreal-world manipulation-oriented reasoning. Finally, distillation exper-iments demonstrate that lightweight models supervised exclusively byour automated pipeline closely match the upper-bound performance ofcostly human-annotated baselines.

Effective and generalizable control in video generation re-mains a significant challenge. While many methods rely on ambiguousor task-specific signals, we argue that a fundamental disentanglement of"appearance" and "motion" provides a more robust and scalable path-way. We propose FlexAM, a unified framework built upon a novel 3Dcontrol signal. This signal represents video dynamics as a point cloud,introducing three key enhancements: multi-frequency positional encodingto distinguish fine-grained motion, depth-aware encoding, and a flexiblecontrol signal for balancing precision and generalization. This represen-tation allows FlexAM to effectively disentangle appearance and motion,enabling a wide range of tasks including I2V/V2V editing, camera con-trol, and spatial object editing. Extensive experiments demonstrate thatFlexAM achieves superior performance across all evaluated tasks. Codesare available at https://github.com/IGL-HKUST/FlexAM.

Multimodal large language models (MLLMs) are increas-ingly considered as a foundation for embodied agents, yet it remainsunclear whether they can reliably reason about the long-term phys-ical consequences of actions from an egocentric viewpoint. We studythis gap through a new task, Egocentric Scene Prediction with LOng-horizon REasoning: given an initial-scene image and a sequence of atomicaction descriptions, a model is asked to predict the final scene afterall actions are executed. To enable systematic evaluation, we introduceEXPLORE-Bench, a benchmark curated from real first-person videosspanning diverse scenarios. Each instance pairs long action sequenceswith structured final-scene annotations, including object categories, vi-sual attributes, and inter-object relations, which supports fine-grained,quantitative assessment. Experiments on a range of proprietary andopen-source MLLMs reveal a significant performance gap to humans,indicating that long-horizon egocentric reasoning remains a major chal-lenge. We further analyze test-time scaling via stepwise reasoning andshow that decomposing long action sequences can improve performanceto some extent, while incurring non-trivial computational overhead. Over-all, EXPLORE-Bench provides a principled testbed for measuring andadvancing long-horizon reasoning for egocentric embodied perception.

Real-world deployment of embodied agents requires activeexploration, visual grounding, and interactive intent disambiguation. How-ever, existing frameworks often rely on privileged simulator states or as-sume complete instructions, bypassing realistic deployment challenges.To bridge this gap, we present REAL, an agentic framework for open-world mobile manipulation. REAL establishes sim-to-real-consistent en-vironment APIs without oracle perception and integrates a simulateduser to enable human-in-the-loop interaction. Within this environment,we design diverse task compositions to drive data collection, supervisedfine-tuning, and online reinforcement learning, systematically optimizingagent performance. To comprehensively evaluate this approach, we in-troduce REAL-Bench, a benchmark spanning 241 tasks across activeexploration, visual distraction, articulated manipulation, and interactivedisambiguation. Experimental results demonstrate that our trained agentoutperforms leading commercial closed-source VLMs on interactive taskswith a 56.9% success rate. Further empirical analysis reveals that ourhierarchical training pipeline successfully aligns the model’s tool-use ca-pabilities while maintaining robust open-vocabulary reasoning under ex-tended exploration horizons. Finally, we deploy and evaluate our frame-work on a physical dual-arm mobile robot, where it achieves a 78.3%end-to-end success rate over 60 real-world episodes. These physical trialsdemonstrate robust zero-shot transferability to unseen household scenar-ios, validating that our sim-to-real-consistent design successfully bridgesthe reality gap for long-horizon mobile manipulation. Code is availableat https://github.com/InternRobotics/REAL.

While surface normal estimation from point clouds is a fun-damental problem in computer vision, there has been limited explorationon LiDAR point clouds. Recent approaches adopt Transformer-basedmodels designed for large-scale point clouds to achieve single-step esti-mation, still struggling to estimate surface normals with high computa-tional efficiency and accuracy. In this paper, we propose an efficient sur-face normal estimation method for LiDAR point clouds, termed ESNE.It converts point clouds into a sequence data format and leverages astate space model to capture long-range dependencies. To capture lo-cal structure, the proposed method introduces a local attention moduleincorporating an efficient and explicit search for neighboring points onthe sequence. Furthermore, the proposed method introduces a featurerefinement module using structural variability as guidance to enhancethe shape representations of local regions. Experimental results on large-scale LiDAR point cloud datasets demonstrate that the proposed methodimproves the mean accuracy by over 33.4% and achieves more than twicethe inference speed compared to state-of-the-art methods.

Deep learning-based gaze estimation methods suffer severeperformance degradation in cross-domain scenarios. Our frequency anal-ysis reveals that overfitting to source-domain low-frequency componentsassociated with domain-specific factors, such as illumination, background,and skin tone, is a critical cause of cross-domain degradation. Basedon this finding, we propose the Adversarial Low-Frequency Perturba-tion (ALFP) framework, which employs a Perturbation Generation Net-work (PGNet) to adversarially generate instance-specific perturbationimages and mix their low-frequency amplitude statistics with those ofsource images through an AdaIN-based mechanism. The resulting inputs,with perturbed low-frequency distributions, are used to train the GazeEstimation Network (GazeNet). Adversarial training between PGNet andGazeNet progressively mitigates GazeNet’s overfitting to source-domainlow-frequency components, thereby improving cross-domain generaliza-tion. Experiments show that ALFP outperforms state-of-the-art meth-ods across various cross-domain settings, improving over the baseline byup to 34.9% and the best existing method by 15.0%, without requiringtarget-domain data.

DiT video generation is latency-intensive due to iterativefull-frame denoising, while prior cloud-edge methods largely rely on staticinter-step decoupling and cannot leverage inter-frame similarity or adaptto system dynamics. We propose EcoVideo, an entropy-orchestratedframework for dynamic inter-frame decoupling: early-stage self-attentionentropy provides a training-free estimate of frame-wise information den-sity for frame selection; a cloud large model denoises sparse high-entropykeyframes; and an edge lightweight model reconstructs the remainingframes via motion-aware interpolation with refinement for temporal sta-bility. EcoVideo further adapts the keyframe budget and edge refine-ment depth to real-time bandwidth and compute availability, optimizingend-to-end latency under constraints. Experiments on representative DiTvideo generators show improved quality–efficiency trade-offs and up to2.9× end-to-end speedup in low-bandwidth, compute-limited edge set-tings. Code is available at https://github.com/IF-LAB-PKU/EcoVideo.

Conditional diffusion models frequently suffer from language-image misalignments. Due to the ambiguity of intermediate noise cor-rupted latents, assessing prompt adherence currently requires complet-ing the entire sampling trajectory. This late-stage evaluation incurs evenhigher computational costs during test-time scaling strategies, such asBest-of-N (BoN) sampling, as all misaligned trajectories must finish gen-eration before being discarded. To tackle this, we propose NoisyCLIP, anoise-aware twin-tower model that enables early language-to-latent align-ment estimation. By learning a vision encoder on noise-corrupted latents,we allow the model to "see" through the ambiguity of intermediate diffu-sion steps. To facilitate this training, we investigate noise-data augmen-tation sampling strategies and introduce two new benchmark datasets:Noisy-Conceptual-Captions and Noisy-GenAI-Bench. When applied asan early-stopping criterion for BoN, NoisyCLIP at half cost matches orbeats frozen CLIP at full cost. Ultimately, this transforms alignment as-sessment from an expensive final check into a continuous monitoring tool,drastically reducing compute costs without sacrificing semantic fidelity.