ECCV 2026 论文列表

Recent depth foundation models trained on perspective im-agery achieve strong performance, yet generalize poorly to 360∘ imagesdue to the substantial geometric discrepancy between perspective andpanoramic domains. Moreover, fully fine-tuning these models typicallyrequires large amounts of panoramic data. To address this issue, wepropose RePer-360, a distortion-aware self-modulation framework formonocular panoramic depth estimation that adapts depth foundationmodels while preserving powerful pretrained perspective priors. Specifi-cally, we design a lightweight geometry-aligned guidance module to derivea modulation signal from two complementary projections (i.e., ERP andCP) and use it to guide the model toward the panoramic domain withoutoverwriting its pretrained perspective knowledge. We further introduce aSelf-Conditioned AdaLN-Zero mechanism that produces pixel-wise scalingfactors to reduce the feature distribution gap between the perspectiveand panoramic domains. In addition, a cubemap-domain consistency lossfurther improves training stability and cross-projection alignment. Byshifting the focus from complementary-projection fusion to panoramic do-main adaptation under preserved pretrained perspective priors, RePer-360surpasses standard fine-tuning methods while using only 1% of the train-ing data. Under the same in-domain training setting, it further achievesan approximately 20% improvement in RMSE. The code is available athttps://github.com/munimo/RePer360.

Recent advances in online reconstruction of dynamic scenesdemonstrate impressive visual quality, showing potential for real-worldapplications such as VR content streaming. Yet, existing online recon-struction methods still require a large number of input views and time-consuming iterative optimization. Moreover, reliable pose estimation,a prerequisite for NVS, introduces additional delay. In this paper, wepresent REON-NVS, a feedforward online NVS framework for sparse-view input video streams, which goes from pose estimation to novel-viewreconstruction in real time. REON-NVS comprises two main compo-nents. First, our method leverages a feedforward pose estimator andmapper that replace conventional camera pose optimization pipelines.Second, we design a scene reconstructor built upon a state-space model(SSM), which efficiently synthesizes temporally consistent novel-view im-ages by exploiting information from previous frames. To train and eval-uate our approach in realistic in-the-wild streaming scenarios, we intro-duce a new multi-view dynamic scene dataset of 150 dynamic scenescaptured with moving cameras. Extensive experiments demonstrate thatREON-NVS achieves high visual quality while operating in real time (32FPS), validating its applicability in real-world scenarios.

Cross-modal image registration is essential for multi-sensorperception but remains fundamentally challenging due to severe non-linear radiometric discrepancies and geometric distortions. Existing de-terministic matching methods lack uncertainty awareness, struggling tonavigate the resulting highly non-convex optimization landscape and fre-quently accumulating errors in ambiguous regions. In this paper, wepropose RBE-Flow, a novel framework that reformulates dense cross-modal flow estimation as a closed-loop recurrent Bayesian estimationproblem on learned feature manifolds. Diverging from standard feed-forward regression, RBE-Flow establishes a robust self-correcting mech-anism by deeply coupling feature-metric non-linear optimization withprobabilistic state updates. Specifically, a Recurrent Manifold Optimiza-tion (RMO) block iteratively generates flow observations and their as-sociated uncertainties, which are then optimally assimilated into theprior state via an Uncertainty-Adaptive Probabilistic Update (UAPU)using deterministic sigma-point projection. Crucially, the resulting cal-ibrated posterior covariance is fed back to adaptively regularize thedamping of subsequent optimization steps, allowing the system to mod-ulate its convergence based on predictive confidence. To ensure stableprobabilistic training, we introduce a hybrid supervision scheme featur-ing a geometry-aware rectified NLL loss that structurally prevents vari-ance collapse. Extensive experiments on challenging OSdataset, WHU-OPT-SAR, and RoadScene benchmarks demonstrate that RBE-Flowconsistently achieves state-of-the-art performance, outperforming exist-ing methods by a significant margin, particularly under strict sub-pixelcriteria. Project page: https://github.com/NEU-Liuxuecong/RBE-Flow

Multimodal Sentiment Analysis (MSA) seeks to infer humanemotions by integrating textual, acoustic, and visual cues. However, ex-isting approaches often rely on all modalities are completeness, whereasreal-world applications frequently encounter noise, hardware failures, orprivacy restrictions that result in missing modalities. There exists a sig-nificant feature misalignment between incomplete and complete modali-ties, and directly fusing them may even distort the well-learned represen-tations of the intact modalities. To this end, we propose PRLF, a Pro-gressive Representation Learning Framework designed for MSA underuncertain missing-modality conditions. PRLF introduces an AdaptiveModality Reliability Estimator (AMRE), which dynamically quantifiesthe reliability of each modality using recognition confidence and Fisherinformation to determine the dominant modality. In addition, the Pro-gressive Interaction (ProgInteract) module iteratively aligns the othermodalities with the dominant one, thereby enhancing cross-modal con-sistency while suppressing noise. Extensive experiments on CMU-MOSI,CMU-MOSEI, and SIMS verify that PRLF outperforms state-of-the-artmethods across both inter- and intra-modality missing scenarios, demon-strating its robustness and generalization capability.

Estimating room layouts from multi-view imagery is a coretask for indoor scene understanding. Existing methods are typically lim-ited either by poor generalization to new datasets or restrictive geometricassumptions of the room shape or camera configuration. Most also es-timate rooms independently, failing to exploit shared building structuresuch as dominant directions, ground plane or ceiling height.We propose PolyLayout, a multi-room layout estimation method that pa-rameterizes room layouts as Manhattan 3D polygons and optimizes themjointly across multiple rooms. The optimization objective is predicted bya neural network on top of robust pre-trained visual features and trainedend-to-end with supervision only on output room layouts. At the sametime, camera projection and polygon updates remain explicit and model-based. This separation between learned scoring and geometry improvesgeneralization to new datasets and camera parameters. During optimiza-tion, PolyLayout adaptively refines the polygon topology through iter-ative wall split and merge operations while jointly utilizing structuralcues across rooms. We introduce two new multi-view multi-room layoutbenchmarks by providing layout annotations to existing datasets, andexperiments show that PolyLayout outperforms prior approaches, bothin terms of accuracy and robustness.Project page: https://ghanning.github.io/PolyLayout

Autoregressive (AR) models are highly effective for imagegeneration, yet their standard maximum-likelihood estimation traininglacks direct optimization for sample quality and diversity. While rein-forcement learning (RL) has been used to align diffusion models, thesemethods typically suffer from output diversity collapse. Similarly, concur-rent RL methods for AR models rely strictly on instance-level rewards,often trading off distributional coverage for quality. To address these lim-itations, we propose a lightweight RL framework that casts token-basedAR synthesis as a Markov Decision Process, optimized via Group Rela-tive Policy Optimization (GRPO). Our core contribution is the introduc-tion of a novel distribution-level Leave-One-Out FID (LOO-FID) reward;by leveraging an exponential moving average of feature moments, it ex-plicitly encourages sample diversity and prevents mode collapse duringpolicy updates. We integrate this with composite instance-level rewards(CLIP and HPSv2) for strict semantic and perceptual fidelity, and stabi-lize the multi-objective learning with an adaptive entropy regularizationterm. Extensive experiments on LlamaGen and VQGAN architecturesdemonstrate clear improvements across standard quality and diversitymetrics within only a few hundred tuning iterations. The results alsoshow that the model can be updated to produce competitive sampleseven without Classifier-Free Guidance, and bypass its 2x inference cost.

Real-time object detection has achieved substantial progress through meticulously designed architectures and optimization strategies. However, the pursuit of high-speed inference via lightweight network designs often leads to degraded feature representation, which hinders further performance improvements and practical on-device deployment. In this paper, we propose a cost-effective and highly adaptable distillation framework that harnesses the rapidly evolving capabilities of Vision Foundation Models (VFMs) to enhance lightweight object detectors. Given the significant architectural and learning objective disparities between VFMs and resource-constrained detectors, achieving stable and task-aligned semantic transfer is challenging. To address this, on one hand, we introduce a Deep Semantic Injector (DSI) module that facilitates the integration of high-level representations from VFMs into the deep layers of the detector. On the other hand, we devise a Gradient-guided Adaptive Modulation (GAM) strategy, which dynamically adjusts the intensity of semantic transfer based on gradient norm ratios. Without increasing deployment and inference overhead, our approach painlessly delivers striking and consistent performance gains across diverse DETR-based models, underscoring its practical utility for real-time detection. Our new model family, RT-DETRv4, achieves state-of-the-art results on COCO, attaining AP scores of 49.8/53.7/55.4/57.0 at corresponding speeds of 273/169/124/78 FPS.

We propose OTCache, a training-free framework for ac-celerating diffusion sampling via caching schedule prediction. Existinggraph-based caching methods reduce redundant computation by opti-mizing shortest-path objectives, but rely on an additive independenceassumption, which often breaks down in the low NFE regime. To addressthis issue, OTCache models caching schedules across inference budgetsas a smooth evolution in policy space, inspired by Optimal Transport(OT). The framework consists of three stages: (1) obtaining a high-fidelity reference schedule using a graph-based caching method un-der a conservative budget; (2) performing a lightweight anchor searchunder an extreme low-budget setting via Optuna optimization with anend-to-end perceptual objective; and (3) predicting schedules for tar-get budgets via quantile interpolation between the reference and an-chor policies using continuous warping representations. Experiments onFLUX.1 [dev], Qwen-Image, and HunyuanVideo show that OTCacheachieves 4.5×, 4.7×, and 3.66× acceleration, respectively, while consis-tently improving generation fidelity over state-of-the-art caching base-lines. This work provides a new perspective on accelerating diffusionmodels through Optimal-Transport-inspired schedule modeling. Code:https://github.com/UnicomAI/OTCache

Recently, open-vocabulary 3D object detection (3D-OVD)has gained increasing attention for its ability to detect unseen objects in3D scenes. Existing approaches typically adopt a two-stage pipeline thatfirst discovers novel objects using foundation models and then trains a3D-OVD model based on these discovered objects. Although effective,this pipeline often suffers from inaccurate localization and mismatchedclassification during the discovery stage, which subsequently limits theperformance of the model training stage. To address these limitations,we advocate for improving both the reliability of novel object discov-ery and the robustness of model training, and propose an innovativeframework. Specifically, for reliable discovery, our co-distillation strategydistills high-quality novel objects by applying Hungarian matching overa comprehensive score that incorporates geometric consistency, struc-tural objectness, and semantic certainty. To enhance robust model train-ing, we further propose a dual-guidance learning scheme, incorporatinga scene-awareness-guided uncertainty regularization for the regressionhead and an LLM-guided hierarchical alignment for the classificationhead, effectively mitigating the negative effects of imprecise 3D boundingboxes and semantic ambiguity. Extensive experiments on SUN RGB-Dand ScanNetV2 demonstrate that our method achieves significant per-formance gains over state-of-the-art approaches. Code is available athttps://github.com/shangboyuan/Co-3DGT.

Machine unlearning enables deep neural networks to selec-tively remove the influence of specific data in response to privacy andregulatory requirements. While prior work largely studies single-shot un-learning, real-world systems must accommodate continual unlearning,where multiple unlearning requests occur sequentially over time. In thiswork, we identify a fundamental limitation of this setting: plasticity col-lapse, a progressive breakdown in a model’s ability to effectively forget.Through theoretical analysis of continual unlearning dynamics, we showthat continual unlearning operations accumulate geometric constraintsin parameter space, leading to saturated subspaces that restrict futureupdates. This structural effect induces two distinct failure modes: (1)Forward failure – diminishing forgetting quality for subsequent tasks,and (2) Backward failure – spontaneous re-memorization of previouslyforgotten information. Extensive experiments across multiple architec-tures, datasets, and methods in image classification confirm that plastic-ity collapse is not an artifact of specific implementations, but a pervasivephenomenon inherent to continual unlearning. Our findings reveal a criti-cal barrier to the long-term reliability of machine unlearning systems andmotivate the development of plasticity-preserving unlearning algorithms.Our code is available at https://github.com/TIML-Group/Continual-Machine-Unlearning-Plasticity-Collapse.

Omnidirectional depth estimation from multi-fisheye camerarigs is complicated by visibility conflicts: wide baselines cause differentcameras to observe different portions, or even different faces, of the sameobject, so aggregating their features into a unified equirectangular (ERP)representation under fixed projection produces ambiguous matching ev-idence near occlusion boundaries and thin structures. Although existingmethods mitigate this by down-weighting unreliable views, they do notresolve the underlying discrepancy because context formation and cross-view fusion remain tied to rigid fisheye-to-ERP sampling. We presentOmniDS, an iterative depth refinement framework that replaces rigidaggregation by combining dynamic context fusion with consensus-awaremulti-view similarity. A dual-stream encoder pairs a lightweight CNNfor geometric detail with a frozen DINOv3 for semantic priors; their fea-tures are reprojected into ERP space at each refinement step via learnedview weighting and deformable cross-attention with geometric distortionbias. In parallel, a multi-view consensus volume captures global cross-camera agreement through group-wise correlation and feature variance,regularized by a 3D U-Net. For efficient deployment, we distill the dual-stream representation into a single MobileNet-based encoder. OmniDSachieves state-of-the-art performance on the OmniThings, OmniHouse,and Sunny benchmarks while maintaining competitive inference speed.Project page and codes are available here.

Facial Expression Recognition (FER) in the wild is inher-ently affected by annotation ambiguity and noisy label. Recent label dis-tribution learning (LDL) methods alleviate this issue by exploiting neigh-borhood information. However, existing approaches typically rely onstatic or heuristic mechanisms to estimate neighbor contribution weights,limiting their capacity to model complex dependencies among neighbor-ing samples. In this paper, we propose Mamba-driven Neighbor WeightRefinement (Mamba-NWR), a robust FER framework that integratescross-attention routing (CAR) for initial neighbor weight estimation andMamba-based iterative refinement (MIR) for distribution aggregation,effectively handling noisy labels. Specifically, our CAR mechanism intro-duces a multi-route local feature projection for both the input and neigh-bor samples. The subsequent cross-attention-based interaction amongthese features serves as the foundation for computing the initial neigh-bor weights. Then, we employ a Mamba-based state space model to iter-atively refine the contribution weights of neighboring label distributions,enabling dynamic balance between preserving historical states and ab-sorbing new neighborhood information. The optimized neighbor distribu-tions are then aggregated with the original logical label through a learn-able fusion factor to construct the final supervision information. Exten-sive experiments on multiple benchmark datasets demonstrate that ourapproach consistently improves robustness under ambiguous and noisyannotations, achieving state-of-the-art performance.

As real-world prediction systems often face missing modali-ties at inference, incomplete multimodal learning (IML) remains a prac-tical challenge. While prior methods aim to learn representations robustto missing inputs, representations from incomplete modalities inevitablydeviate from their full-modality counterparts due to missing evidence. Toexplicitly leverage these deviations, we propose MARS (Missingness-Aware Residual-guided Specialization), a mixture-of-experts frameworkthat guides expert specialization based on how representations are re-shaped by missingness. By contrasting task representations derived fromincomplete inputs with their complete counterparts during training, wederive a privileged residual signal that captures this representational gap.The residual signal guides a residual router to assign samples to the ex-perts specialized for the corresponding deviation patterns. In parallel, afeature router learns to imitate this routing behavior using only incom-plete inputs, enabling deployment without access to full modalities. Tomitigate this train–test router gap, we develop a discrepancy-aware noiseregularization that adaptively perturbs the residual router’s decisionswhen the feature router deviates, enhancing the expert robustness underimperfect imitation. Experiments on multimodal classification (CASIA-SURF, CREMA-D, UPMC Food-101) and segmentation (MCubeS) un-der missing scenarios show that MARS consistently surpasses baselines,while remaining efficient and extensible to diverse backbones and tasks.

On-policy exploration is a crucial component for trainingrobust Vision-Language Navigation (VLN) agents, as it exposes the pol-icy to a broader state distribution. However, such exploration inevitablyleads to trajectories that deviate from expert demonstrations, resulting ina semantic mismatch between the executed visual stream and the originallanguage instruction. In this work, we address this challenge by introduc-ing Φ-Nav, a unified on-policy framework that leverages hindsight rea-soning to align instructions with the agent’s actual exploratory journey.Specifically, Φ-Nav operates through a three-stage dual-supervision cy-cle: 1) the agent performs oracle-guided on-policy exploration, samplinga trajectory while learning from expert action feedback, 2) a hindsightspeaker synthesizes a path-level hindsight instruction grounded in thecollected visual observations, and 3) the agent conducts a second imita-tion pass, treating the synthesized trajectory–instruction pair as an ad-ditional expert demonstration. Through this process, Φ-Nav bridges thecritical semantic supervision gap inherent in on-policy methods, trans-forming semantically unlabeled movement into dense training signals.Evaluations on the R2R-CE and RxR-CE benchmarks show that Φ-Navyields competitive performance while requiring only a fraction of the ex-pert demonstrations used by current baselines. These results underscorethe necessity of semantic exploration in VLN, positioning Φ-Nav as aeffective solution for training embodied agents with limited data.

Single Image Super-Resolution (SISR) is a fundamental com-puter vision task that aims to reconstruct a high-resolution (HR) im-age from a low-resolution (LR) input. Transformer-based methods haveachieved remarkable performance by modeling long-range dependenciesin degraded images. However, their feature-intensive attention compu-tation incurs high computational cost. To improve efficiency, most ex-isting approaches partition images into fixed groups and restrict atten-tion within each group. Such group-wise attention overlooks the inher-ent asymmetry in token similarities, thereby failing to enable flexibleand token-adaptive attention computation. To address this limitation,we propose the Individualized Exploratory Transformer (IET), which in-troduces a novel Individualized Exploratory Attention (IEA) mechanismthat allows each token to adaptively select its own content-aware andindependent attention candidates. This token-adaptive and asymmetricdesign enables more precise information aggregation while maintainingcomputational efficiency. Extensive experiments on standard SR bench-marks demonstrate that IET achieves state-of-the-art performance undercomparable computational complexity. The code is available at here.

Primitive-based methods such as 3D Gaussian Splatting haverecently become the state-of-the-art for novel-view synthesis and relatedreconstruction tasks. Compared to neural fields, these representationsare more flexible, adaptive, and scale better to large scenes. However,the limited expressivity of individual primitives makes modeling high-frequency detail challenging. We introduce Neural Harmonic Textures, aneural representation approach that anchors latent feature vectors on avirtual scaffold surrounding each primitive. These features are interpo-lated within the primitive at ray intersection points. Inspired by Fourieranalysis, we apply periodic activations to the interpolated features, turn-ing alpha blending into a weighted sum of harmonic components. Theresulting signal is then decoded in a single deferred pass using a smallneural network, significantly reducing computational cost. Neural Har-monic Textures yield state-of-the-art results in real-time novel view syn-thesis while bridging the gap between primitive- and neural-field-basedreconstruction. Our method integrates seamlessly into existing primitive-based pipelines such as 3DGUT, Triangle Splatting, and 2DGS. We fur-ther demonstrate its generality with applications to 2D image fitting andsemantic reconstruction.

Editing dynamic 3D scenes reconstructed from monocularvideos is important for applications such as content creation and dataaugmentation, but remains challenging. Prior works adapt image diffu-sion models across frames while enforcing spatial–temporal consistency,yet they often produce uniform, appearance-only edits. We instead lever-age a video diffusion model to enable photorealistic 4D edits and control-lable motion changes. In monocular settings, however, fitting dynamic3D Gaussians with pure photometric loss is under-constrained, lead-ing to shallow geometric reconstructions and temporal artifacts in theedited scene. We address this with a novel 4D editing pipeline that usescompact latents distilled from DINO features as additional supervisionduring scene optimization and preserves them during editing as a se-mantic identity signal. Such semantic consistency regularization providesstrong geometric and motion cues, enabling robust scene editing evenunder large or complex edits from the video diffusion model. Further-more, to promote semantically novel content generation during editing,we apply a gradient-based feature lifting method based on accumulatedfeature gradients during Gaussian densification. Experiments show im-proved spatial-temporal realism and text alignment over image editingbaselines, and demonstrate compositional motion-appearance edits withsignificant geometry changes that prior methods cannot achieve.

Federated learning (FL) protects sensitive (Vision-LanguageNavigation) VLN data without centralizing trajectories or instructions,but severe non-IID environments make personalized FL (pFL) necessary.Moreover, VLN poses several coupled challenges for personalized feder-ated learning, including environment heterogeneity, multimodal language-vision fusion, and long-horizon navigation with time-varying decisioncontexts. To address these challenges, we propose FedNASP, a step-wisepersonalized federated learning framework for VLN. The key idea is todynamically calibrate personalization strength along a navigation trajec-tory. Specifically, we introduce a lightweight Step-wise Personalized Mod-ulator (SPM) that predicts personalization strength at each navigationstep. We further design a structure-aware adapter-based personalizedprefix injection mechanism that enables client-specific grounding whilekeeping the backbone shared across clients. Experiments on three rep-resentative datasets show that FedNASP consistently outperforms state-of-the-art federated VLN methods under substantial cross-client hetero-geneity. Compared with the non-centralized baselines, FedNASP improvesRemote Grounding Success on REVERIE by 13.0% and Success Rateon CVDN by 22.6%. Extensive ablation studies and visualizations fur-ther validate the effectiveness of adaptive step-wise personalization forfederated VLN. Code is available at: https://github.com/IntelliSys-Lab/FedNASP.git

Data augmentation is known to improve generalization ofdeep visual models. Recent methods favor mixup strategies that gener-ate interpolated samples to improve model performance. However, thesetechniques not only incur significant computational overhead, they alsolead to semantic disruption of augmentation data due to cross-samplemixing. We first propose Self-Saliency (S 2 ) Mixup, which constructschallenging yet label-consistent samples by extracting multi-scale salientpatches and reinserting them into non-salient regions of the same im-age. This promotes scale-invariant feature learning while avoiding cross-sample interference. To further enhance model robustness, we intro-duce FracMix, a mixing scheme that injects self-similarity patterns intosalient regions using adaptive ratios. Collectively, our unified framework,S 2 -FracMix, enables simultaneous learning from fractal and non-fractalstructures within a single image, yielding a targeted and structurally co-herent augmentation strategy. We theoretically analyze the advantageof our technique, and empirically establish its superiority over the ex-isting methods by achieving state-of-the-art performance in extensiveevaluation with seven benchmarks across classification (coarse and fine-grained), robustness, calibration, object detection, and transfer learningtasks. Project page is available at fracmix-data-augmentation.github.io