ECCV 2026 Papers

Vision-language models (VLMs) excel at zero-shot recogni-tion, yet their performance degrades under distribution shifts during in-ference. Unsupervised test-time adaptation (TTA) often relies on noisypseudo-labels or heuristics, offering limited guidance for where decisionboundaries should move; existing active TTA (ATTA) methods typi-cally treat annotations as local fixes rather than structured updates tothe shared semantics of VLMs. We introduce PAA (Probe, Anchor, andAmend), an ATTA framework that treats visual and textual prototypesas an evolving state and updates this state through a three-step cycle. Ineach cycle, the Probe step performs prototype-guided querying under aclass-balanced budget, selecting both central samples to preserve sourceknowledge and boundary samples to track distribution shifts. Using onlybinary verifications, the Anchor step applies a prototype-regularized up-date that couples cross-modal alignment with intra-class compactnessand inter-class separation, yielding a decision space that is stable yetplastic. As a slow controller, the Amend step revisits past hard casesstored in memory and converts temporal discrepancies into memory-driven self-refinement. Across VLM test-time benchmarks, PAA deliversconsistent gains over strong TTA and ATTA baselines under comparableannotation budgets and compute, turning sparse, delayed feedback intostructured evolution of the prototype-anchored decision space.

Recent advances in 3D Gaussian Splatting have enabled im-pressive photorealistic novel view synthesis. However, to transition froma pure rendering engine to a reliable spatial map for autonomous agentsand safety-critical applications, knowing where the representation is un-certain is as important as the rendering fidelity itself. We bridge thiscritical gap by introducing a lightweight, plug-and-play framework forpixel-wise, view-dependent predictive uncertainty estimation. Our post-hoc method formulates uncertainty as a Bayesian-regularized linear least-squares optimization over reconstruction residuals. This architecture-agnostic approach extracts a per-primitive uncertainty channel withoutmodifying the underlying scene representation or degrading baseline vi-sual fidelity. Crucially, we demonstrate that providing this actionablereliability signal successfully translates 3D Gaussian splatting into atrustworthy spatial map, further improving state-of-the-art performanceacross three critical downstream perception tasks: active view selection,pose-agnostic scene change detection, and pose-agnostic anomaly detec-tion. Code is available at github.io/3DGS-Uncertainty.

Recent work has shown that Chain-of-Thought (CoT) rea-soning improves text-to-motion generation, yet generating explicit rea-soning tokens at inference time introduces significant latency. We presentLaCT-Motion (Latent Chain-of-Thought for Motion), the first methodto transfer latent reasoning from language to LLM based human mo-tion generation, achieving explicit-CoT-level quality without its infer-ence overhead. Our three-stage recipe mirrors human skill acquisition:practice learns explicit CoT reasoning via supervised fine-tuning; in-ternalization progressively compresses variable-length reasoning stepsinto compact continuous latent tokens through a stage-based curricu-lum with stochastic scheduling; and perfection refines the fully latentpolicy with GRPO using format, motion similarity, and semantic simi-larity rewards. At inference, the model reasons entirely through latenttokens with no explicit text generated. Experiments on HumanML3Dshow that latent internalization not only matches but surpasses explicitCoT reasoning (R-Precision Top-1: 0.577 vs. 0.515 for Motion-R1, FID:0.167), suggesting that continuous latent tokens encode spatio-temporalreasoning more effectively than discrete text. Code will be released athttps://github.com/Erwin2233/LaCT-Motion.

Multimodal large language models (MLLMs) deliver impressive vision-language reasoning, but suffer steep inference latency because self-attention scales quadratically with sequence length and thousands of visual tokens contributed by high-resolution images. Naively pruning less-informative visual tokens reduces this burden, yet indiscriminate removal can erase contextual cues essential for background or fine-grained questions, undermining accuracy. In this paper, we present ParVTS (Parallel Vision Token Scheduling), a training-free scheduling framework that partitions visual tokens into subject and non-subject groups, processes them in parallel to transfer their semantics into question tokens, and discards the non-subject path mid-inference to reduce computation. This scheduling reduces computational complexity, requires no heuristics or additional modules, and is compatible with diverse existing MLLM architectures. Experiments across multiple MLLM backbones show that ParVTS prunes up to 88.9% of visual tokens with minimal performance drop, achieving 1.77× speedup and 70% FLOPs reduction. Our code is released at https://github.com/CrispyFeSo4/ParVTS.

Recent advances in foundation models (FMs) have led to aproliferation of Whole Slide Image (WSI) feature extractors, yet WSIanalysis still faces a practical challenge: no single FM consistently per-forms best across all tasks, and the performance gap among differentMultiple Instance Learning (MIL) aggregators built on a single strongFM can become limited in some settings. Recent research further showsthat simple ensembles of FMs can outperform the best individual FM,indicating that different FMs capture partially overlapping yet comple-mentary information. This motivates a shift from selecting a single FMtoward coordinating and exploiting multiple FMs jointly. However, exist-ing multi-FM approaches mostly rely on naive feature concatenation orpatch-level selection, operating on raw correlated FM features withoutexplicitly disentangling shared and view-specific information. To addressthis challenge, we reformulate FM ensembling as a Multi-View Learn-ing (MVL) problem. We propose Multi-View Multiple Instance Learning(MV-MIL), a plug-and-play framework that treats features from eachFM as a distinct view of a WSI. MV-MIL introduces an information-theoretic disentanglement module to exploit cross-view complementar-ity while suppressing redundancy, together with a Multi-Branch Predic-tion Head that integrates shared and view-specific representations withinMIL pipelines. Extensive experiments on 14 public benchmarks showthat MV-MIL consistently achieves state-of-the-art performance, withits plug-and-play design yielding robust gains across diverse MIL meth-ods and FM combinations.

Category-Agnostic Pose Estimation (CAPE) aims to local-ize category-specific keypoints on a query image from only a few anno-tated support examples. Existing CAPE methods are largely based ondirect coordinate regression or local heatmap matching, which becomeunreliable under severe occlusion because missing keypoints cannot berecovered from local evidence alone. We present FlowCape, a condi-tional flow-matching framework for occlusion-resilient CAPE. FlowCapefirst extracts query-image features and combines them with keypoint-description embeddings in a Heatmap-Guided Initialization module, pro-ducing a stable initial pose for transport. A shared Riemannian PoseHead then predicts the conditional velocity field, while its internal GraphFlow Encoder fuses query image, textual keypoint semantics, and skele-ton topology to enforce structured pose evolution. The final predictionis obtained by probability-flow ODE rollout, and training is driven byhybrid flow-matching supervision together with initialization and roll-out consistency constraints. To better evaluate robustness under severeocclusion, we further introduce Occ80, an occlusion-focused benchmarkspanning 80 categories. Experiments on MP-100 and Occ80 show thatFlowCape consistently outperforms strong CAPE baselines, reach thestate-of-the-art under occluded scenario.

The incorporation of additional modalities into action recog-nition models increases their performance across a wide range of settings.However, how this additional information can contribute to making themodels more robust remains underexplored, particularly for the case ofmulti-modal out-of-distribution (OOD) detection. While methods existthat regularize the multi-modal training process with OOD detection inmind, they still apply off-the-shelf OOD detectors designed for the uni-modal case during inference, discarding important information. Basedon an interesting relationship we find between the multi-modal and uni-modal predictions, we propose to use this signal to build a post-hocdetector explicitly designed for the multi-modal scenario. We combinethis new source of information with a feature-space score, which detectsoff-manifold samples in the multi-modal space, and normalize them bythe multi-modal logits. In doing so, the proposed hybrid detector is com-patible with existing training-time approaches and consistently improvesperformance. Experiments on a wide range of established datasets fromthe MultiOOD benchmark show that, on average, our approach outper-forms the state of the art. Our results show the importance of explicitlyconsidering the different modalities at inference time for multi-modalOOD detection.

The recent large video foundation model, SAM2, enablessegment anything in both images and videos, serving as a powerful basemodel for various applications. However, many of such use cases requireto operate on resource-constrained devices like mobile phones and laptops.In this work, we aim to make SAM2 more mobile-friendly by distillingthe heavyweight SAM2 into a lightweight model, facilitating segmentanything in both images and videos on mobile devices. To this end, wepropose Hypergraphical Knowledge Distill (HyperKD), which introducesthe idea of hypergraph into knowledge distillation, aiming to effectivelymodel and transfer SAM2’s generalizable and comprehensive knowledge.HyperKD consists of Temporal HyperKD and Granularity HyperKD thatconstruct hypergraphs to explicitly model and extract the generalizabletemporal knowledge and the comprehensive multi-granularity knowledgefrom SAM2 respectively, which are then distilled into the lightweightstudent model by aligning it with the constructed hypergraphs. Besides,we present MobileSAM2, a new family of lightweight SAM2 that balancesefficiency and effectiveness via searching the best model architectures withHyperKD during model size reduction. Extensive experiments validateMobileSAM2 across multiple benchmarks and show promising generaliza-tion performance on embodied AI tasks.

Pre-trained Vision Foundation Models (VFMs) provide strongvisual representations for diverse downstream tasks. The key challenge ofVFM adaptation stems from the prohibitive costs of full fine-tuning andcatastrophic forgetting. To address this, Low-Rank Adaptation (LoRA)has emerged as the prevailing paradigm for Parameter-Efficient Fine-Tuning (PEFT). However, LoRA is typically designed for transformerself-attention layers parameterized by 2D matrices. Since convolutionalkernels inherently couple spatial and channel information within a 4Dtensor, forcing them into a monolithic 2D matrix disrupts the inher-ent spatial topology. In this paper, we propose Low-Rank ConvolutionalAdaptation (LoCA), a convolution-aware PEFT framework that addressesspatial-channel entanglement by decoupling channel and spatial adapta-tion. LoCA introduces a low-rank channel adaptation for dense cross-channel mixing and refines spatial bases extracted from pre-trained ker-nels via Singular Value Decomposition (SVD). Experimental results showthat LoCA preserves pre-trained spatial priors and achieves competitiveor state-of-the-art performance across fine-grained classification, domain-generalized semantic segmentation, and generative benchmarks.

Multiresolution hash encodings have recently enabled fast and high-fidelity implicit neural representations by storing multi-scale features in fixed-size hash tables along a geometric resolution schedule. However, the standard design is data-agnostic: different resolution levels receive identical hash-table capacity despite large differences in image frequency content. As a result, some levels experience severe hash collisions while others underutilize parameters, leading to inefficient capacity allocation. To address this issue, we propose Collision-Aware Resolution Adaptation (CARA), a method that assigns per-level resolutions by balancing the effective information load across hash levels. This adaptive allocation reduces capacity bottlenecks and improves parameter efficiency. In addition, we introduce an invertible pixel-shuffle transform that reduces hash load factors by redistributing spatial information, thereby mitigating collision-induced information loss without enlarging the hash tables. To support evaluation on extremely high-resolution data, we also curate, to the best of our knowledge, the first uncompressed wholeslide image dataset for academic research. Experiments on Kodak images, gigapixel natural images, and raw whole-slide images demonstrate that CARA consistently improves the fidelity–parameter trade-off. Our method matches state-of-the-art performance while using only 27.76% of the parameters, and achieves up to 6.11dB PSNR improvement at comparable parameter counts. Code is provided in the supplementary.

Vision-Language Navigation in Continuous Environments(VLN-CE) requires agents to learn complex reasoning from long-horizonhuman interactions. While multi-modal large language models (MLLMs)have driven recent progress, current training paradigms struggle to bal-ance generalization capability, error recovery, and training stability. Specif-ically, (i) policies derived from supervised fine-tuning (SFT) suffer fromcompounding errors, struggling to recover from out-of-distribution states,and (ii) Reinforcement fine-tuning (RFT) methods such as GRPO arebottlenecked by sparse outcome rewards. Their binary feedback fails toassign credit to individual steps, leading to gradient signal collapse infailure-dominant batches. To address these challenges, we introduce Step-Aware Contrastive Alignment (SACA), a framework designed to extractdense supervision from imperfect trajectories. At its core, the Perception-Grounded Step-Aware auditor evaluates progress step-by-step, disentan-gling failed trajectories into valid prefixes and exact divergence points.Leveraging these signals, the Scenario-Conditioned Group Constructionmechanism dynamically routes batches to specialized resampling andoptimization strategies. Extensive experiments on VLN-CE benchmarksdemonstrate that SACA achieves state-of-the-art performance.

We present KineticGS, a novel approach for reconstructingdynamic scenes from monocular videos through momentum-driven 4DGaussian Splatting. While recent advances in motion modeling have im-proved dynamic reconstruction, they often produce temporal inconsis-tencies and non-physical deformations due to the lack of physical pri-ors constraining motion trajectories. KineticGS addresses this with aphysically grounded momentum hierarchy that models Gaussian parti-cles as momentum-carrying entities, thereby establishing a hierarchicalmotion representation from pixel observations to object-level coherence.Specifically, Gaussian particles serve as momentum carriers, with dy-namics driven by local motion energy. We perform momentum-guidedparticle sampling, concentrating representation in high-energy regionsto capture fine-grained non-rigid motions. A learnable energy-flow mod-ule then predicts per-particle temporal activation, ensuring continuousand physically plausible motion propagation. Further, a holistic kineticsynchronization mechanism enforces consistent motion evolution amongcorrelated particles, preserving structural integrity without explicit disin-tegration. Experiments on dynamic scene datasets show that KineticGSimproves reconstruction fidelity, temporal coherence, and physical real-ism, achieving +0.91 dB average PSNR gain in dynamic-object regionsover the second-best.

Self-supervised learning (SSL) has revolutionized representation learning, with Joint-Embedding Architectures (JEAs) emerging as an effective approach for capturing semantic features. Existing JEAs rely on implicit or explicit batch interaction – via negative sampling or statistical regularization – to prevent representation collapse. This reliance becomes problematic in regimes where batch sizes must be small, such as high-dimensional scientific data, where memory constraints and class imbalance make large, well-balanced batches infeasible. We introduce IConE (Instance-Contrasted Embeddings), a framework that decouples collapse prevention from the training batch size. Rather than enforcing diversity through batch statistics, IConE maintains a global set of learnable auxiliary instance embeddings regularized by an explicit diversity objective. This transfers the anti-collapse mechanism from the transient batch to a dataset-level embedding space, allowing stable training even when batch statistics are unreliable, down to batch size 1. Across diverse 2D and 3D biomedical modalities, IConE outperforms strong contrastive and non-contrastive baselines throughout the small-batch regime (from B = 1 to B = 64) and demonstrates marked robustness to severe class imbalance. Geometric analysis shows that IConE preserves high intrinsic dimensionality in the learned representations, preventing the collapse observed in existing JEAs as batch sizes shrink.

We present a dense prediction system for ice cloud geom-etry retrieval from extremely sparse supervision, where active satellitesensors (EarthCARE radar/lidar) provide accurate but spatially narrowtraining labels (∼1.6% of pixels) and passive imagers (VIIRS, VisibleInfrared Imaging Radiometer Suite) observe the full globe. A ConvNex-tUNet trained on co-located tracks predicts eight targets at every pixelwith calibrated 90% prediction intervals using conformalized quantileregression (CQR). The best configuration, a five-member quantile en-semble, achieves R2= 0.742 with prediction quality constant regardlessof distance to the supervision track, confirming that the model learnsa per-pixel spectral retrieval rather than interpolating from nearby la-bels. Calibration is robust across latitudes and cloud types, with deepconvective clouds as the main failure mode. Deployed globally, the sys-tem produces 609M predictions from one day of VIIRS data in 2.4 hourson a single GPU, enabling dense 3D cloud characterization at planetaryscale.1

We address multiview self-calibration from image correspon-dences when camera intrinsic parameters vary across views. While Kruppaequations constrain an implicit matrix encoding of the camera intrinsics(called the dual image of the absolute conic; DIAC), practical calibrationpriors are expressed directly in intrinsic-parameter space. We formulatesuch priors algebraically and map them into DIAC space, yielding explicitconstraints that integrate directly with Kruppa relations. Unlike priorself-calibration methods that focus on specific varying-intrinsic regimes(e.g. shared focal length, fixed aspect, or zero skew), we provide a unifiedalgebraic framework that handles arbitrary intrinsic priors expressed aspolynomial constraints in the space of intrinsics, yielding a more flex-ible formulation of varying-intrinsic self-calibration. Since Kruppa sys-tems are projective and often algebraically dependent, we algorithmicallyconstruct locally independent square subsystems via Jacobian analysis toassess solvability and algebraic complexity under varying intrinsics. Ex-periments demonstrate parity with a few previous classical approachesin shared-focal settings as well as shared-principal-point settings andenable stable estimation for varying-intrinsic configurations previouslyconsidered unsolved, validated on synthetic and real dynamic-intrinsicsequences. Code is available at https://github.com/Rowing0914/Self-Calibration-Varying-Intrinsics.

Prompt learning has gained significant attention as a parameter-efficient approach for adapting large pre-trained vision-language modelsto downstream tasks. However, when only partial labels are available, itsperformance is often limited by label ambiguity and insufficient supervi-sory information. To address this issue, we propose Holistic Optimal La-bel Selection (HopS), leveraging the generalization ability of pre-trainedfeature encoders through two complementary strategies. First, we designa local density-based filter that selects the top frequent labels from thenearest neighbors’ candidate sets and uses the softmax scores to identifythe most plausible label, capturing structural regularities in the featurespace. Second, we introduce a global selection objective based on optimaltransport that maps the uniform sampling distribution to the candidatelabel distributions across a batch. By minimizing the expected trans-port cost, it can determine the most likely label assignments. These twostrategies work together to provide robust label selection from both lo-cal and global perspectives. Extensive experiments on eight benchmarkdatasets show that HopS consistently improves performance under par-tial supervision and outperforms all baselines. Those results highlightthe merit of holistic label selection and offer a practical solution forprompt learning in weakly supervised settings. The code is available athttps://github.com/Qizhoay/HopS.

Multi-animal tracking (MAT) is critical for wildlife monitor-ing and behavioral analysis, yet remains challenging due to uniform ap-pearance, high density, and irregular motion. Existing methods typicallyfollow heuristic- or query-based paradigms: the former relies on hand-crafted geometric associations without end-to-end optimization, whereasthe latter enables joint optimization but relies heavily on appearanceembeddings. In such conditions, continuous geometric embeddings canbe unstable, as small coordinate perturbations may disproportionatelyalter cross-frame attention weights, degrading identity association perfor-mance. To address this limitation, we propose HieDG, a HierarchicalDiscrete Geometry-guided tracking framework that reformulates geo-metric dynamics as structured discrete representations within a query-based tracker. Instead of directly using raw geometric signals, HieDGemploys a two-stage residual codebook to discretize position, scale, andvelocity cues, transforming unstable continuous geometry into struc-tured, stable discrete tokens. These tokens are aligned with visual em-beddings and integrated into the tracking queries to enhance identityconsistency. Extensive experiments on animal-specific benchmarks (Ani-malTrack, BFT, and BuckTales) demonstrate state-of-the-art associationperformance with significant improvements in HOTA, AssA, and IDF1.Additional evaluations on generic multi-object tracking benchmarks, in-cluding DanceTrack and SportsMOT, show competitive performance,indicating the broader applicability of discretized geometric modelingbeyond animal-specific scenarios.

Domain-incremental object detection (DIOD) requires mod-els to continually adapt to new domains while preserving prior knowl-edge. Recently, parameter-efficient fine-tuning offers a promising avenue,wherein a pre-trained model is frozen and a small number of learnable pa-rameters are injected for downstream tasks. However, these methods riskoverwriting critical past knowledge, triggering inter-domain interferenceand performance degradation. To address this challenge, we propose Or-thogonal Knowledge Refreshing (OKR), a simple yet effective frameworkfor DIOD. OKR incrementally constructs independent domain-specificsubspaces via dedicated low-rank branches for each domain, which areseamlessly fused for a holistic decision, enabling conflict-free capacity ex-pansion without domain selection during inference. To minimize knowl-edge interference during fusion, we present a gradient-based orthogo-nal refreshing strategy that projects gradient updates of new domainsonto the orthogonal complement of the fused historical subspace, sup-porting continual adaptation without forgetting. Moreover, to mitigatesemantic fragmentation across domains, we enforce topology-aware con-sistency, aligning the semantic structures of old and new domains. Ex-tensive experiments validate the superiority of OKR, outperforming thebest exemplar-free method by significant margins of +5.6% and +6.5%mAP on the Pascal VOC and BDD100K series, respectively.

Video large language models (Video-LLMs) have achieved re-markable progress in video understanding tasks, but their heavy compu-tational overhead during inference severely limits deployment in resource-constrained scenarios. This highlights the urgent need for efficient infer-ence acceleration frameworks tailored to Video-LLMs. Existing accelera-tion methods typically apply a homogeneous processing approach to allvideo frames, disregarding the fact that each frame within a video clipconveys distinct visual information. Additionally, many of them under-utilize text queries, or add text guidance only through costly pre-trainingor intrusive LLM modifications, underexploiting the text-video align-ment that is critical in long-form video understanding where redundantvisual content is common. To address this limitation, we propose Guide-Then-Refine (GTR), a plug-and-play inference acceleration frameworkfor Video-LLMs that realizes token compression through a two-stagemechanism. First, for all tokens in each video frame, we compute bothglobal scores and text-aware scores. Based on the sum of these scores forall tokens in a frame, we dynamically adjust the token retention ratio foreach frame. Subsequently, according to the retention ratio, we calculatethe local score of each token to evaluate its importance within the currentframe, thereby selecting critical visual tokens. Extensive evaluations onmainstream video benchmarks and various Video-LLMs demonstrate theeffectiveness of our approach. With only 15% of visual tokens retained,our method maintains an average of 95.8% of the original performanceacross four benchmarks.