ECCV 2026 论文列表

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

Monocular 3D object understanding has largely been cast asa 2D RoI-to-3D box lifting problem. However, emerging downstream ap-plications require image-plane geometry (e.g., projected 3D box corners)which cannot be easily obtained without known intrinsics, a problem forobject detection in the wild. We introduce MoCA3D, a Monocular,Class-Agnostic 3D model that predicts projected 3D bounding box cor-ners and per-corner depths without requiring camera intrinsics at in-ference time. MoCA3D formulates pixel-space localization and depthassignment as dense prediction via corner heatmaps and depth maps.To evaluate image-plane geometric fidelity, we propose Pixel-AlignedGeometry (PAG), which directly measures image-plane corner anddepth consistency. Extensive experiments under an oracle-2D box pro-tocol demonstrate that MoCA3D achieves state-of-the-art performance,improving image-plane corner PAG by 22.8% while remaining compa-rable on 3D IoU, using up to 57× fewer trainable parameters. Finally,we apply MoCA3D to downstream tasks which were previously imprac-tical under unknown intrinsics, highlighting its utility beyond standardbaseline models. Project page: https://jeoncwcw.github.io/moca3d/.

Driven by Artificial Intelligence-Generated Content (AIGC),the authenticity of audio-visual content is facing severe challenges. Tem-poral Forgery Localization (TFL) aims to precisely identify manipulatedsegments within untrimmed sequences. However, existing methods arelimited by CNNs’ local receptive fields or Transformers’ quadratic com-plexity, while emerging linear models often struggle to balance globalauthentic context compression with local abrupt forgery perception. Toaddress this, we propose MG-RWKV, a multi-granularity framework thatleverages the data-dependent state evolution of RWKV to achieve ef-ficient full-sequence processing with O(T ) complexity. Our frameworkfeatures three core innovations: (1) a Bidirectional RWKV architec-ture that captures bidirectional temporal contexts without quadraticoverhead; (2) a Multi-Granularity Mixture of Experts (MG-MoE) that performs dynamic routing over explicit temporal recep-tive fields, adaptively selecting granularities based on forgery dura-tion to significantly enhance decision interpretability; and (3) Cross-Granularity Consistency (CGC), which aligns adjacent feature pyra-mid levels through hierarchical scale-wise pairing and spatial boundary-aware weighting, effectively reducing false positives in authentic regions.Extensive experiments on Lav-DF, TVIL, and Psynd datasets demon-strate that MG-RWKV achieves state-of-the-art performance with lowcomputational cost.

Event cameras have recently shown promising capabilities ininstantaneous motion estimation due to their robustness to low light andfast motions. However, computing wide-baseline correspondence betweentwo arbitrary views remains a significant challenge, since event appear-ance changes substantially with motion, and learning-based approachesare constrained by both scalability and limited wide-baseline supervi-sion. We therefore introduce the first event matching model that achievescross-dataset wide-baseline correspondence in a zero-shot manner:a single model trained once is deployed on unseen datasets withoutany target-domain fine-tuning or adaptation. To enable this capabil-ity, we introduce a motion-robust and computationally efficient atten-tion backbone that learns multi-timescale features from event streams,augmented with sparsity-aware event token selection, making large-scaletraining on diverse wide-baseline supervision computationally feasible.To provide the supervision needed for wide-baseline generalization, wedevelop a robust event motion synthesis framework to generate large-scale event-matching datasets with augmented viewpoints, modalities,and motions. Extensive experiments across multiple benchmarks showthat our framework achieves a 37.7% improvement over the previousbest event feature matching methods. Code and data are available at:https: // github. com/ spikelab-jhu/ Match-Any-Events .

Generating talking avatars is a fundamental task in videogeneration. Although existing methods can generate full-body talkingavatars with simple human motion, extending this task to groundedhuman-object interaction (GHOI) remains an open challenge, requiringthe avatar to perform text-aligned interactions with surrounding objects.This challenge stems from the need for environmental perception and thecontrol-quality dilemma in GHOI generation. To address this, we proposea novel dual-stream framework, InteractAvatar, which decouples per-ception and planning from video synthesis for grounded human-object in-teraction. Leveraging detection to enhance environmental perception, weintroduce a Perception and Interaction Module (PIM) to generate text-aligned interaction motions. Additionally, an Audio-Interaction Aware† ⋆Equal contribution. Corresponding author.Generation Module (AIM) is proposed to synthesize vivid talking avatarsperforming object interactions. With a specially designed motion-to-video aligner, PIM and AIM share a similar network structure and enableparallel co-generation of motions and plausible videos, effectively miti-gating the control-quality dilemma. Finally, we establish a benchmark,GroundInter, for evaluating GHOI video generation. Extensive exper-iments and comparisons demonstrate the effectiveness of our method ingenerating grounded human-object interactions for talking avatars.

Egocentric videos capture rich and diverse human–object in-teractions and have emerged as a fundamental resource for understand-ing human activities related to objects. In this context, Video ReferringExpression Comprehension (Video REC), the task of localizing the tem-poral and spatial extent of a referred object in video frames given anatural language query, plays a key role in linking textual descriptionsto observed objects in untrimmed egocentric recordings. However, exist-ing egocentric Video REC benchmarks primarily focus on short videoclips, where some target object appears densely within frames. Suchsettings do not reflect real-world egocentric recordings, which are long-form, untrimmed, and characterized by sparse object occurrences andcomplex activity transitions. To address this limitation, we introduceLongEgoRefer, a novel and challenging benchmark constructed fromlong-form videos in the Ego4D dataset. LongEgoRefer contains 1,498referring expressions with an average video duration of 45 minutes. Thebenchmark exhibits extreme target sparsity, detailed linguistic descrip-tions, and complex human–object interactions embedded in long, dy-namic egocentric narratives. Consequently, it defines a demanding spatio-temporal grounding problem that requires models to identify both whenan event occurs and where the referred object appears within extendedvideo sequences. We evaluate existing Video REC approaches, includ-ing training-free baselines based on vision–language models combinedwith Grounded SAM2. Extensive experiments show that even advancedbaselines and current state-of-the-art models struggle significantly onLongEgoRefer. These results highlight the intrinsic difficulty of long-form egocentric spatio-temporal grounding and emphasize the need formore robust video understanding models. Our benchmark and code areavailable at https://github.com/shunya-kato/LongEgoRefer.LongEgoRefer0s 4345sA transparent, round bowl containing a bakedvegetable dish is shown being removed from anoven. The bowl, encased in a blue protective cloth,is then placed on the Induction stove. Its contents,a creamy vegetable and mushroom mixture, aresubsequently subjected to a brief external inspection.Human-Object Interaction3067s 3111sRefEgo Long-form video Sparse appearance Linguistic complexity0s 5sThere have white colored hot box in theshelf of the room.Fig. 1: Data comparison of LongEgoRefer and RefEgo. (a) Video durations in LongE-goRefer are orders of magnitude longer than those in RefEgo. (b) The appearance rateis significantly lower and sparser in LongEgoRefer compared to RefEgo. (c) To handlethe complexity of long-form videos, captions in LongEgoRefer are substantially longerand more descriptive than those in RefEgo. Dashed lines indicate the mean value foreach distribution.

Despite rapid progress in learning-based stereo matching,high accuracy is often achieved at the cost of heavy backbones andcomputationally intensive 3D cost volume processing, resulting in sub-stantial memory and runtime overhead. More critically, these methodsfrequently struggle to generalize across domains, limiting their practi-cal deployment. We present LiteMatch, a lightweight stereo matchingframework that achieves strong zero-shot generalization through costvolume stabilization-without expensive 3D convolutions. LiteMatch em-ploys two complementary encoders: a Cross-View Correspondence En-coder (CVCE) to capture global cross-view interactions, and a High-Frequency Encoder (HFE) that enhances fine structural details via FFT-based frequency cues. To stabilize the cost volume, we introduce the CostVolume Consistency Loss (CVC-Loss), a voxel-wise binary cross-entropyobjective applied to softmax-normalized cost distributions. By encour-aging sharp and unimodal disparity probabilities, CVC-Loss promotesstable cost distributions and enables rapid convergence. A lightweightrefinement module further produces sharp full-resolution disparities withlow-iteration updates, avoiding heavy recurrent refinement. With a flexi-ble design ranging from 3.36M to 9.58M parameters, LiteMatch achievesexceptional zero-shot generalization, delivering competitive EPE and D1performance across Scene Flow, KITTI, Middlebury, ETH3D, and Driv-ingStereo. Our results establish that lightweight architectures can indeedgeneralize across domains without sacrificing accuracy. Code

General Salient Object Detection (SOD) aims to identify andsegment visually interesting objects from uni-modality or multi-modalityscenes, recently advanced by cutting-edge State Space Models (SSMs).However, a critical limitation of current approaches is their neglect of theinherent spectral biases exhibited by different neural network paradigms.By digging to the dataset-level spectral analysis of Convolutional Neu-ral Networks (CNNs) and SSMs, their semantic representations are in-herently complementary based on their complementary frequency prefer-ences. Inspired by this, we harmonize heterogeneous representations fromSSMs and CNNs to bridge their spectral biases for general salient objectdetection. To this end, inspired by the dynamic information propaga-tion of Liquid Neural Networks (LNNs), we introduce a liquid fusion todynamically integrates features from two backbones, including VMambaand ConvNeXt, referred to Liquid Fusion Network (LFNet). Concretely,by treating the continuous VMamba features and ConvNeXt features asevolving states and exogenous stimulus, respectively, LFNet employs adynamic gating mechanism for content-aware feature aggregation. Cru-cially, this state-stimulus paradigm enables to scale to multi-modal cues,resulting in flexibility in general SOD. Besides, a Saliency-Guided Up-sampling (SGU) operator to propagate the features to the shallow layer,which leverages a spectral-spatial co-design to suppress upsampling ar-tifacts while preserving semantics. Extensive experiments across five di-verse tasks (RGB, RGB-D, RGB-T, VSOD, and VDT) demonstrate thatLFNet achieves state-of-the-art performance, offering a superior trade-offbetween detection accuracy and model efficiency. Code has been releasedat https://github.com/cke520/LFNet.

With the advancement of video generation foundation mod-els (VGFMs), customized generation, particularly subject-to-video (S2V),has attracted growing attention. However, a key challenge lies in balanc-ing the intrinsic priors of a VGFM, such as motion coherence, visualaesthetics, and prompt alignment, with its newly derived S2V capa-bility. Existing methods often neglect this balance by enhancing oneaspect at the expense of others. To address this, we propose Libra-Gen, a novel framework that views extending foundation models forS2V generation as a balance game between intrinsic VGFM strengthsand S2V capability. Specifically, guided by the core philosophy of “Rais-ing the Fulcrum, Tuning to Balance,” we identify data quality as thefulcrum and advocate a quality-over-quantity approach. We constructa hybrid pipeline that combines automated and manual data filteringto improve overall data quality. To further harmonize the VGFM’s na-tive capabilities with its S2V extension, we introduce a Tune-to-Balancepost-training paradigm. During supervised fine-tuning, both cross-pairand in-pair data are incorporated, and model merging is employed toachieve an effective trade-off. Subsequently, two tailored direct prefer-ence optimization (DPO) pipelines, namely Consis-DPO and Real-FakeDPO, are designed and merged to consolidate this balance. During in-ference, we introduce a time-dependent dynamic classifier-free guidancescheme to enable flexible and fine-grained control. Experimental resultsdemonstrate that LibraGen outperforms both open-source and commer-cial S2V models using only thousand-scale training samples. ProjectPage: https://github.com/Phantom-video/LibraGen

Streaming autoregressive (AR) video generators combined with few-step distillation achieve low-latency, high-quality synthesis, yet remain difficult to align via reinforcement learning from human feedback (RLHF). Existing SDE-based GRPO methods face challenges in this setting: few-step ODEs and consistency model samplers deviate from standard flow-matching ODEs, and their short, low-stochasticity trajectories are highly sensitive to initialization noise, rendering intermediate SDE exploration ineffective. We propose AR-CoPO (AutoRegressive Contrastive Policy Optimization), a framework that adapts the Neighbor GRPO contrastive perspective to streaming AR generation. ARCoPO introduces chunk-level alignment via a forking mechanism that constructs neighborhood candidates at a randomly selected chunk, assigns sequence-level rewards, and performs localized GRPO updates. We further propose a semi-on-policy training strategy that complements onpolicy exploration with exploitation over a replay buffer of reference rollouts, improving generation quality across domains. Experiments on SelfForcing demonstrate that AR-CoPO improves both out-of-domain generalization and in-domain human preference alignment over the baseline, providing evidence of genuine alignment rather than reward hacking.

Achieving both global semantic coherence and precise in-stance level control in wide-aspect-ratio panorama generation is an un-resolved challenge. Existing methods that synchronize independent viewsto generate panoramas often lack semantic coherence and struggle withfine-grained multi-instance placement, resulting in contextual artifactsand fragmented objects. We introduce InstaPano, a training-free frame-work for zero-shot instance-controlled panorama generation. InstaPanointegrates a Global Attention Fusion mechanism into a pre-trained layout-to-image model. Through Sync-Fuse-Dispatch workflow, it periodicallyaggregates latent features from all local views to construct a unifiedglobal context, performs multi-level attention computation over this con-text to achieve true fusion, and then dispatches the enriched global fea-tures back to each view. This enables the coherent rendering of complexpanoramas with multiple specified instances. Furthermore, we introducea conditional positional mask to resolve object repetition artifacts thatmay arise in large bounding boxes. On a newly constructed yet challeng-ing benchmark for panoramic instance layout control, InstaPano achievessuperior performance in both layout fidelity and semantic coherence,faithfully generating complex panoramic scenes.

Multimodal learning robust to missing modalities is essen-tial for real-world applications. Existing methods mainly focus on inter-modality missing, where entire modalities are absent, while overlookingintra-modality degradation, where modalities are present but severelycorrupted. In practice, these two types of missing often coexist, mak-ing existing approaches ineffective. To address this limitation, we pro-pose General Incomplete Multimodal Learning (GIML), a unified frame-work that simultaneously handles both inter-modality missing and intra-modality degradation through dynamic quality perception. Specifically,GIML models heterogeneous missing patterns as continuous modality in-formation degradation, enabling degradation-aware adaptive fusion. Toachieve reliable quality perception, we introduce a Noise-aware QualityEstimator that learns the mapping from corrupted features to noise in-tensity through controlled noise injection. Furthermore, we propose aNoise–Semantic Decoupled module that separates semantic informationfrom noise interference. This improves robustness and generalization tounseen corruption patterns. Extensive experiments across datasets withdiverse modality types demonstrate the effectiveness and generality ofGIML. Code is available at: https://github.com/Yu-Five/GIML.