ECCV 2026 论文列表

Vision-Language Models (VLMs) have demonstrated remark-able success across diverse visual tasks, yet their performance degradesin complex visual environments. Existing enhancement approaches re-quire additional training, rely on external segmentation tools, or operateat coarse-grained levels, overlooking VLMs’ innate attention capabilities.To bridge this gap, we investigate VLMs’ attention patterns and discoverthat: (1) visual complexity strongly correlates with attention entropy,negatively impacting reasoning performance; (2) attention progressivelyrefines from global scanning in shallow layers to focused convergencein deeper layers, with the degree of convergence determined by visualcomplexity; (3) theoretically, under a multiplicative decomposition as-sumption, we show that contrasting attention maps between general andtask-specific queries approximately separates visual signal into seman-tic and visual noise components. Building on these insights, we proposeContrastive Attention Refinement for Visual Enhancement (CARVE),a training-free method that extracts task-relevant visual signals throughattention contrasting at the pixel level. Experiments on seven bench-marks show that CARVE improves performance on visual perceptiontasks, with notable gains on both recent and earlier-generation models.Our analysis reveals how visual complexity affects attention mechanismsand demonstrates an effective strategy for improving visual reasoningthrough attention contrasting.

Diffusion prior-based methods have shown impressive resultsin real-world image super-resolution (ISR), yet two key challenges per-sist: balancing pixel-level fidelity with semantic quality, and adapting todiverse degradations. Existing dual-branch approaches freeze the pixelmodule during semantic training, but the semantic branch can still ex-pand capacity within the pixel subspace, precluding genuine perceptualimprovement. Moreover, using a single static adapter cannot generalizeacross heterogeneous real-world corruptions. To address both issues, wepropose FreqOrtho-SR, which comprises: Frequency-guided Mixture ofLoRA Experts (FreqMoE), it routes inputs to specialized experts via anon-parametric FFT-based degradation-feature extractor that encodesfrequency-domain signatures, enabling stable and interpretable special-ization across corruption types; and Orthogonal Gradient Projection(OGP), which reframes the dual-objective optimization as a subspace-constrained problem: by extracting the pixel-fidelity subspace via SVDon combined expert weight deltas and projecting semantic gradients ontoits null space, OGP guarantees orthogonality between the two objectives,enabling genuinely complementary learning without mutual interference.Experiments show that FreqOrtho-SR achieves competitive overall per-formance and a strong fidelity-perception trade-off across multiple bench-marks with efficient single-step inference. The source code of our methodcan be found at sonhm3029/FreqOrtho-SR.

Few-shot Test-Time Domain Adaptation (FSTT-DA) seeksto adapt models to novel domains using only a handful of unlabeledtarget samples. This setting is more realistic than typical domain adap-tation setups, which assume access to target data during source training.However, prior FSTT-DA approaches fail to effectively leverage sourcedomain-specific knowledge, relying on shallow batch normalization up-dates, prompt-based methods that treat the model as a black box, orensembling strategies that do not capture cross-domain relationships. Toaddress these limitations, we introduce a new FSTT-DA framework thatintegrates LoRA fine-tuning with model merging. In our approach, sep-arate LoRA modules are fine-tuned on CLIP’s vision encoder for eachsource domain. Since LoRA modifies only a small fraction of the model’sparameters, it retains the base model’s generalized knowledge while in-ternally learning domain-specific features. To adapt the learned knowl-edge to a specific target domain, we propose a hypernetwork trainedvia meta-learning that generates per-column merging factors to combineLoRA modules. Given a small batch of target images, the hypernetworkproduces merging weights that fuse source LoRA modules into a singleadapted representation. Our results demonstrate state-of-the-art perfor-mance across various domain adaptation datasets. Our code is publiclyavailable at https://github.com/nahbois4321/DA-MergeLoRA.

Automated chest X-ray report generation has recently bene-fited from reinforcement learning (RL) and large language models. How-ever, RL training often suffers from instability or limited explorationdue to fixed Kullback-Leibler (KL) regularization and a static referencepolicy that accumulates KL pressure over time. We propose Response-Weighted and Validation-Anchored Policy Optimization (REVA-PO), aRL framework that stabilizes long-term training via Response-WeightedRegularization (RER) and Validation-Anchored Policy Reset (VAPR).RER dynamically adjusts per-response KL weights based on advantageand reference-policy entropy, relaxing constraints for high-quality re-sponses while tightening them for low-quality ones. Complementarily,VAPR periodically synchronizes the reference and current policies tothe best validation checkpoint, resetting accumulated regularization pres-sure to expand the viable exploration space. To ensure a robust startingpoint, we employ a three-stage pipeline consisting of warm-up training,classifier-guided supervised fine-tuning, and RL. Extensive evaluationson MIMIC-CXR and IU-Xray demonstrate that REVA-PO sets newstate-of-the-art benchmarks in both linguistic quality and clinical accu-racy. Notably, BLEU-4 improves by 5.1% on MIMIC-CXR and 3.6% onIU-Xray, while CheXpert F1 and RadGraph F1 scores increase by 4.5%and 12.8%, respectively, over prior leading methods. The code is publiclyavailable at https://github.com/LiGuo12/REVA_PO/.

End-to-end transformer architectures have driven significant progress in multi-object tracking by unifying detection and association into a single, heuristic-free framework. Despite these benefits, poor detection performance and the inherent conflict between detection and association in a joint architecture remain critical concerns. Recent approaches aim to mitigate these issues by employing advanced denoising or label assignment strategies, or by incorporating detection priors from external object detectors. In this paper, we propose SelfMOTR, a simple yet highly effective detector-free alternative that decouples proposal discovery from association using self-generated internal detection priors. Through extensive analysis and ablation studies, we show that end-to-end transformer trackers with joint detection–association decoding retain substantial hidden detection capacity, and we provide a practical detector-free mechanism for leveraging it. To shed light on these joint decoding dynamics, we draw inspiration from attention sink analyses in large language models, leveraging Track Attention Mass to show that standard generic queries exhibit unbalanced attention, frequently struggling to weigh track context against novel object discovery. SelfMOTR achieves highly competitive performance in complex, dynamic environments, yielding 69.2 HOTA on DanceTrack and leading with 71.1 HOTA on the Bird Flock Tracking (BFT) dataset. Project page: https://medem23.github.io/SM.

Image-based high-throughput plant phenotyping leveragesinstance-level understanding of individual leaves growth over experimen-tal time, yet existing datasets lack temporal depth and annotation con-sistency to jointly benchmark segmentation, tracking, and growth stageclassification. We introduce PhenoLeaf-TS, a time-series dataset of 17,082top-down RGB images spanning 21 Arabidopsis thaliana genotypes, to-talling 318 plant replicates, each annotated with colour-coded leaf in-stance masks that maintain consistent identity throughout the growthsequence. We define three benchmark computer vision tasks with stan-dardised protocols and evaluate 21 distinct models: 9 instance segmenta-tion architectures, 6 multi-object trackers, and 6 classification architec-tures. Mask R-CNN with ResNet-50 achieves 73.2 mAP for leaf segmen-tation, ByteTrack reaches 84.1% MOTA for tracking individual leaves insequences, and Swin-T achieves 91.7% accuracy for growth stage classifi-cation. Cross-dataset experiments on CVPPP and Komatsuna show thatfine-tuning from PhenoLeaf-TS weights yields up to +51 mAP over zero-shot transfer. The dataset, the evaluation code, and the trained modelsare publicly accessible at https://pisyntor.github.io/PhenoLeaf-TS.

We present UniQueR, a unified query-based feedforward framework for efficient and accurate 3D reconstruction from unposed images. Existing feedforward models such as DUSt3R, VGGT, and AnySplat typically predict per-pixel point maps or pixel-aligned Gaussians, which remain fundamentally 2.5D and limited to visible surfaces. In contrast, UniQueR formulates reconstruction as a sparse 3D query inference problem. Our model learns a compact set of 3D anchor points that act as explicit geometric queries, enabling the network to infer scene structure— including geometry in occluded regions—in a single forward pass. Each query encodes spatial and appearance priors directly in global 3D space (instead of per-frame camera space) and spawns a set of 3D Gaussians for differentiable rendering. By leveraging unified query interactions across multi-view features and a decoupled cross-attention design, UniQueR achieves strong geometric expressiveness while substantially reducing memory and computational cost. Experiments on Mip-NeRF 360 and VR-NeRF demonstrate that UniQueR surpasses recent feedforward baselines in both rendering quality and geometric accuracy, using an order of magnitude fewer primitives than dense alternatives.

We present Spectral Consistent Flow (SC-Flow), a 3D med-ical image translation framework with a single function evaluation (1-NFE) in the latent space. This approach reformulates medical imagetranslation as a stochastic Brownian bridge process that directly con-structs a mapping between source and target modalities by predictingthe support regularized mean velocity field. To mitigate modality entan-glement, over-smoothing, and artifacts induced by the implicit low-passmodulation of the latent average velocity, we introduce a Spectral Consis-tency Corrector that dynamically regularizes the evolution of the powerspectral density via learnable frequency-domain gain modulation. Thismechanism establishes an explicit bridge between spatial textures andspectral energy flow, enabling the model to recover fine-grained anatom-ical fidelity while maintaining global structural coherence. Extensive ex-periments on four datasets demonstrate that SC-Flow delivers signifi-cantly more accurate, consistent, and robust performance across varioustranslation scenarios.

Personalized image retouching aims to adapt retouching stylesof individual users from reference examples, but existing methods oftenrequire user-specific fine-tuning or fail to generalize effectively. To addressthese challenges, we introduce RefRetouch, a general framework forpersonalized image retouching that instantly adapts to user retouchingstyles without any test-time fine-tuning. It employs an asymmetric auto-encoder to encode the retouching style from paired examples into a con-tent disentangled latent representation that enables faithful transfer ofthe retouching style to new images. To adaptively apply the encoded re-touching style to new images, we further propose retrieval-augmented re-touching (RAR), which retrieves and aggregates style latents from refer-ence pairs most similar in content to the query image. With these compo-nents, RefRetouch enables superior and generic content-aware retouch-ing personalization across diverse scenarios, including single-reference,multi-reference, and mixed-style settings, while also generalizing out ofthe box to photorealistic style transfer.

Spin plays a crucial role in many ball sports due to its effecton the trajectory of the ball. Vision-based estimation of the ball’s spinduring a game with conventional cameras is challenging due to the ball’ssmall size, high speed, and fast rotation. To address these challenges, wepropose an event-based active vision system that can track unmodifiedballs and measure their spin in real time. The system consists of an eventcamera for its high temporal resolution and minimal motion blur, high-speed pan/tilt galvanometer mirrors to keep the ball in the field of view,and a low-latency focus-tunable telephoto lens to increase the spatial res-olution on the ball and keep it in focus. To track the ball, we use a hybridapproach that combines 2D event-based detection for centering and 3Dpositions from a ball localization system for re-initialization. For high-accuracy spin estimation, we propose an offline method that performscontrast maximization on the sphere (s-CMax). This method achievesstate-of-the-art accuracy on static balls across multiple sports (table ten-nis, baseball, tennis, and golf), with mean magnitude and axis errors of1.2% and 1.5 degrees, respectively. We then develop a low-latency onlinemethod for table tennis as a case study in real-time applications. Thismethod uses an uncertainty-aware convolutional neural network trainedon pseudo-ground-truth spin labels from the offline approach, combinedwith a GPU-accelerated batch implementation of contrast maximizationfor refinement. We demonstrate reliable tracking and spin estimationwith a three-view setup during professional table tennis matches, withhigh accuracy (8.8% magnitude and 6.4 degrees axis mismatch w.r.t. theoffline method), 3 ms latency, and 750 Hz throughput.

Long-term aerial localization requires matching live flightimagery against archival reference maps, demanding feature represen-tations that are invariant to severe appearance changes and perceptualaliasing. However, current benchmarks report only aggregate accuracyover predominantly man-made environments, masking severe terrain-dependent performance gaps. As a result, localization performance overunstructured natural landscapes—where self-similar textures and drasticseasonal changes dominate—remains effectively unmeasured. To addressthis, we introduce Leaving the City, the first large-scale aerial benchmarkdesigned to isolate and quantify terrain-dependent localization gaps. Cap-tured via a microlight aircraft, our dataset comprises 1,379 km of flighttrajectories flown repeatedly to capture distinct seasonal variations. Wepair high-frame-rate imagery and inertial measurements with semanticterrain masks, multi-year-old orthophotos, and precise 6-DoF groundtruth. Evaluating state-of-the-art matchers through our terrain-stratifiedprotocol reveals a systematic bias: methods that succeed on man-madesurfaces degrade sharply over natural terrain undergoing strong appear-ance change. By exposing where current methods fail, our benchmarkprovides a rigorous foundation for developing robust, all-terrain aeriallocalization. The dataset and code are publicly available.4

Accurate anatomical structure segmentation in medical im-age is essential for diagnosis and treatment planning. While recent inter-active segmentation foundation models enhance generalization throughlarge-scale multimodal pretraining, they still depend on precise promptsand can fail in underrepresented clinical contexts (e.g., small organs-at-risk). We present AtlasSegFM, an atlas-guided framework that cus-tomizes off-the-shelf foundation models to new clinical contexts with asingle annotated example. AtlasSegFM 1) performs atlas-query regis-tration to generate context-aware prompts, 2) refines the segmentationwith a frozen foundation model, and 3) applies a lightweight adaptivefusion module to combine atlas priors with foundation-model inputs andpredictions. Extensive experiments on six public and in-house datasetsacross radiotherapy and vascular scenarios show consistent gains, withthe largest improvements on small and delicate structures. AtlasSegFMprovides a lightweight, deployable solution for one-shot customization ofsegmentation foundation models in real-world clinical workflows.

Masked autoencoding has emerged as a prominent paradigmfor self-supervised learning on 3D point clouds, achieving competitiveperformance across downstream tasks. Unlike its 2D counterpart, 3Dmasked autoencoding directly reconstructs spatial coordinates, makingit inherently susceptible to positional leakage. In this work, we identifythat the decoder in existing 3D MAE frameworks tends to over-rely onpositional information, which weakens semantic representation learningand leads to suboptimal feature quality. To address this issue, we proposeMPL-MAE, a masked point learning framework that mitigates positionalover-reliance while enhancing the utilization of encoder features. Specifi-cally, we introduce a recalibrated positional embedding module that sup-presses metric-dominant coordinate signals while preserving geometrictopology, together with a gated positional interface module that dynam-ically regulates positional injection during reconstruction. These designspromote a more balanced interaction between spatial priors and seman-tic features, yielding robust and informative representations. Extensiveexperiments across downstream tasks demonstrate that MPL-MAE con-sistently achieves competitive performance, validating its effectiveness.Code is available at https://github.com/yanx57/MPL-MAE.

Recent advances in Virtual Try-On (VTON) and VirtualTry-Off (VTOFF) have greatly improved photo-realistic fashion synthesisand garment reconstruction. However, existing datasets remain static,lacking instruction-driven editing for controllable and interactive fash-ion generation. In this work, we introduce the Dress Editing Dataset(Dress-ED), the first large-scale benchmark that unifies VTON, VTOFF,and text-guided garment editing within a single framework. Each sam-ple in Dress-ED includes an in-shop garment image, the correspondingperson image wearing the garment, their edited counterparts, and anatural-language instruction of the desired modification. Built througha fully automated multimodal pipeline that integrates MLLM-basedgarment understanding, diffusion-based editing, and LLM-guided ver-ification, Dress-ED comprises over 146k verified quadruplets spanningthree garment categories and seven edit types, including both appear-ance (e.g., color, pattern, material) and structural (e.g., sleeve length,neckline) modifications. Based on this benchmark, we further propose aunified multimodal diffusion framework that jointly reasons over linguisticinstructions and visual garment cues, serving as a strong baseline forinstruction-driven VTON and VTOFF.

In this paper, we explore spatial-aware humanoid whole-body manipulation task. Compared with tabletop settings, this taskposes two key challenges: 1) Spatial understanding is challenging in com-plex 3D environments with diverse spatial relations. 2) Action generationis difficult to generalize, as limited and costly real-robot data restrictsdata-driven models generalization. To address these challenges, we pro-pose a generalizable humanoid loco-manipulation framework that lever-ages the spatial perception and action generation capabilities of multi-agent large models. Specifically, our framework includes two components:Active Spatial Brain for active spatial perception and decision-making,and Generalizable Action Cerebellum for executable robot action genera-tion. The first component actively perceives the spatial scene and makesdecisions on task planning and subtask decomposition. The second com-ponent generates executable robot actions based on the decisions madeby the first module without needs of task-specific real robot data. Tobenchmark our framework, we design a set of spatial manipulation tasksfrom two perspectives: evaluating spatial perception and understand-ing, and assessing real-robot task performance. The results demonstratestrong performance on both aspects across diverse tasks and environ-ments.

Standard pipelines for physics-ready 3D reconstruction relyon a decoupled two-stage paradigm: extracting surface geometry followedby an error-prone tetrahedralization process. While recent Lagrangianmethods like TetSphere Splatting attempt to bypass this by directlyoptimizing volumetric primitives, their homeomorphic constraints pre-vent topology-adaptive optimization. Consequently, they produce dis-joint tetrahedra rather than a single connected mesh, rendering thestructures unsuitable for further physical simulations. To address this,we propose a topology-adaptive framework for holistic tetrahedral meshreconstruction through end-to-end topological and geometric optimiza-tion. First, by coupling Gaussian spheres to tetrahedral elements andleveraging edge connections, we estimate a continuous opacity field fordifferentiable element pruning. Next, jointly minimizing mesh smoothingenergy and multi-view Gaussian rendering error drives alternating geo-metric refinement while preserving topological adaptivity. Consequently,our approach effectively constructs a unified and topologically coherenttetrahedral mesh. Extensive experiments demonstrate that our methodoutperforms state-of-the-art techniques by achieving superior geometricaccuracy and producing coherent, single-connected tetrahedral meshes,thereby effectively bypassing the error-prone conventional tetrahedraliza-tion step for reconstructed surface meshes and streamlining downstreamphysical simulation.

Multi-objective masked image modeling (MIM) combines com-plementary learning signals (token distillation, CLS alignment, and pixelreconstruction) but existing methods weight these objectives with globalscalars, ignoring spatial heterogeneity across patches. We present Ex-PLoRe (Expert Patch-Level Loss Routing), which repurposes Soft Mix-ture of Experts (MoE) dispatch weights as learned, per-patch loss coef-ficients. The key mechanism is loss-coupling: allowing loss gradients toflow through dispatch weights to the router enables content-dependentspecialization, where different patches receive different emphases acrossobjectives. A detach ablation confirms loss-coupling as the core mecha-nism, degrading performance by 1.6% when gradients are blocked. OnImageNet-1K with ViT-Base, ExPLoRe improves over non-MoE base-lines on two objective combinations (Token+CLS: +0.5% k-NN, +4.4%linear probe; Token+Pixel: +2.2% k-NN), achieving 80.6% linear probeand 85.3% finetuning accuracy, competitive with published methods. Fordownstream transfer, we develop adaptation recipes (Freeze Routing, Ex-pert Dropout, and Freeze Attention) that improve MoE finetuning by+1.5% over the vanilla MoE, and close a 2.5–2.9 mIoU segmentation gapso that MoE models match or exceed non-MoE baselines on ADE20K.

Recent generative approaches to geometry estimation adaptpretrained image diffusion models and treat the task as image-conditionedgeneration. Leveraging off-the-shelf image diffusion models, they either(i) train task-specific geometry models (for depth and surface normal es-timation) independently, losing the opportunity of exploring the intrinsiccorrelation of these geometric targets, or (ii) jointly fine-tune modifiedimage diffusion backbones (e.g., altered self-attention), which typicallydemands substantial labeled data. To overcome these limitations in aprincipled fashion, we repurpose pretrained video generative models as aunified and data-efficient framework for geometry estimation, formulatedinnovatively as a next-frames prediction task. Our method, GeoNeXt,inherits naturally structured knowledge and richer priors from the videomodel, while further adapting them for joint modeling of images and ge-ometry targets (

Towards an embodied generalist for real-world interaction,Multimodal Large Language Model (MLLM) agents still suffer fromchallenging latency, sparse feedback, and irreversible mistakes. Videogames offer an ideal testbed with rich visual observations and closed-loopinteraction, demanding fine-grained perception, long-horizon planning,and precise control. However, systematically evaluating these capabilitiesis currently hindered by heterogeneous action interfaces and heuristicverification. To this end, we introduce GameWorld, a benchmarkdesigned for standardized and verifiable evaluation of MLLMs as generalistgame agents in browser environments. Two game agent interfaces arestudied: (i) Computer-use agents that directly emit keyboard and mousecontrols, and (ii) Generalist multimodal agents that act in a semanticaction space via deterministic Semantic Action Parsing. GameWorldcontains 34 diverse games and 170 tasks, each paired with state-verifiable metrics for outcome-based evaluation. The results across 18model-interface pairs suggest that even the best-performing agent isfar from achieving human capabilities on video games. Further studieson real-time interaction, context-memory sensitivity, and action validityexpose more challenges ahead for game agents. Together, by offering astandardized and verifiable evaluation framework, GameWorld laysa foundation for advancing research on multimodal game agents andbeyond.