ECCV 2026 Papers

Thermal infrared (TIR) imaging is indispensable for all-weatherperception and energy management, yet existing 3D thermal modelingmethods remain primarily confined to static environmental snapshots.These approaches typically struggle to capture complex temporal tem-perature variations and lack the capacity to simulate scenes under unseenenvironmental conditions. To bridge this gap, we propose ThermoGS,a unified framework that couples 4D Gaussian Splatting (4DGS) withfundamental thermodynamic principles to achieve physically-consistentthermal field emulation. Unlike previous data-driven methods that di-rectly regress radiance, ThermoGS decouples intrinsic material proper-ties, including emissivity (e), convective heat transfer coefficient (h), vol-umetric heat capacity (c), and absorptivity (α)—from transient temper-ature states. By integrating Physics-Informed Neural Networks (PINNs),our framework ensures that the predicted spatiotemporal thermal distri-bution strictly adheres to energy balance equations, enabling generalizedsimulation across arbitrary viewpoints and time periods. Furthermore,we introduce the Thermo Dataset, a comprehensive, drone-captured col-lection featuring synchronized multi-modal data and diverse environmen-tal parameters. Extensive experiments demonstrate that ThermoGS sig-nificantly outperforms comparative methods in thermal simulation accu-racy, maintaining a predicted temperature error within 1 degree Celsius.The code is available at https://github.com/NPU-CVPG/ThermoGS.

Vision-Language-Action (VLA) models have emerged as apromising paradigm for end-to-end autonomous driving. However, exist-ing VLAs’ traffic relies heavily on text-centric visual question answeringand chain-of-thought reasoning data, which emphasizes linguistic rea-soning rather than action-grounded planning. As a result, the learnedrepresentations capture semantic knowledge but lack spatial dependenciescrucial for reliable trajectory prediction. We propose DriveTeach-VLA,a framework that explicitly teaches VLAs what to see and where tolook. Driving-aware Vision Distillation (DVD) injects driving-specificperceptual priors into the vision encoder, while 2D Trajectory-GuidedPrompts (2D-TGP) provide spatial conditioning aligned with feasibledriving trajectories. Together they form a vision-guided learning pipeline:what to see (DVD pretraining) → where to look (TGP-guided SFT) →how to act (TGP-guided GRPO). DriveTeach-VLA achieves the state-of-the-art performance on NAVSIM and nuScenes. Our code is available at:https://github.com/ShivaTeam/DriveTeach-VLA.

Video Large Language Models (Video LLMs) aim to under-stand and reason over dynamic visual content. However, videos inher-ently exhibit low task-relevant information density because only a smallfraction of frames and regions provide meaningful cues. As a result, theuseful signals are sparse across space and time. We present STVFocus, atraining-free framework that unifies spatio-temporal focusing to capturesalient contexts and build informative video representations. STVFocusconsists of three components. First, Hierarchical Temporal Focus appliesa coarse-to-fine frame selection strategy that captures global temporalstructure while selectively emphasizing informative moments without ex-cessive query bias. Second, Query-informed Spatial Focus uses spatialsaliency to highlight query-relevant intra-frame regions. Finally, Tem-poral Reallocation reinvests the saved spatial capacity into additionalhigh-value frames, enabling richer temporal coverage. By jointly leverag-ing temporal and spatial saliency, STVFocus produces compact yet se-mantically enriched video representations. Experiments on VideoMME,LongVideoBench, and MLVU show consistent gains with meaningfulmargins, confirming the effectiveness of unified spatio-temporal focus-ing.

Image registration is an ill-posed dense vision task, wheremultiple solutions achieve similar loss values, motivating probabilisticinference. Variational inference has previously been employed to capturethese distributions, however restrictive assumptions about the posteriorform can lead to poor characterisation, overconfidence and low-qualitysamples. More flexible posteriors are typically bottlenecked by the com-plexity of high-dimensional covariance matrices required for dense 3Dimage registration.In this work, we present a memory and computationally e!cient in-ference method, Structured SIR, that enables expressive, multi-modal,characterisation of uncertainty with high quality samples. We proposethe use of a Sampled Importance Resampling (SIR) algorithm witha novel memory-e!cient high-dimensional covariance parameterisationas the sum of a low-rank covariance and a sparse, spatially structuredCholesky precision factor. This structure enables capturing complex spa-tial correlations while remaining computationally tractable.We evaluate the e!cacy of this approach in 3D dense image registrationof brain MRI data, which is a very high-dimensional problem. We demon-strate that our proposed method produces uncertainty estimates that aresignificantly better calibrated than those produced by variational meth-ods, achieving equivalent or better accuracy. Crucially, we show thatthe model yields highly structured multi-modal posterior distributions,enable e"ective and e!cient uncertainty quantification.

Deep learning models exhibit fairness concerns when pre-dictions are inadvertently influenced by sensitive attributes. However,existing attempts to make Path-Specific Counterfactual Fairness opti-mizable rely on estimating marginal potential outcome probabilities—anapproach that fundamentally requires high-dimensional conditional den-sity estimation and breaks down in modalities such as medical images,where the curse of dimensionality renders reliable estimation infeasible.To address this limitation, we reduce the problem of enforcing Path-Specific Counterfactual Fairness to a causal conditional independenceconstraint and prove that satisfying this constraint is sufficient to elim-inate the unfair causal effect. This reduction replaces intractable coun-terfactual estimation with a discriminative optimization objective thatremains scalable in high-dimensional settings.

Vision-language models (VLMs) such as CLIP enable zero-shot classification by comparing image features with text prompts in ashared embedding space. A fundamental property underlying this ca-pability is the global comparability of logits across arbitrary candidateclasses. However, VLMs are often adapted to fine-grained domains usingtechniques such as LoRA. While this improves in-domain accuracy, out-of-domain accuracy degrades. This leads to a higly fragmented modelecosystem, with thousands of specialized models. Multi-Expert-Domain(MED) classification seeks to address this problem, by merging LoRAstrained independently on specialized domains. However, due to the inde-pendent training, the various domain experts no longer produce globallycalibrated logits. As a result, when evaluating over the union of multipledomain-specific class sets, heterogeneous logit scales induce cross-domaininterference and artificially high confidence for out-of-domain classes,inducing prediction errors. In this work, we identify domain supervi-sion and cross-domain logit miscalibration as the key issue to scalablemulti-domain zero-shot recognition. We propose a mixture-of-expertsMED architecture, MED-DSLC, combining domain supervised training anddomain-wise logit scaling, to explicitly restore global logit comparability.MED-DSLC is a lightweight solution for MED classification, which is shownto preserve within-domain discrimination while reducing cross-domainlogit interference with minimal data. Extensive experiments across di-verse fine-grained benchmarks demonstrate that it substantially improvesmean accuracy (+15%), cross-domain robustness, and scalability in thesize of MED classification problem. Our results show that restoring output-level calibration is essential under highly data imbalanced settings forachieving a truly zero-shot VLM under multi-domain specialization. Codeis publicly available at MED-DSLC.

Multimodal Large Language Models (MLLMs) have achievedremarkable success in visual understanding but remain constrained invisual generation due to the fundamental feature discrepancy betweensemantic perception and pixel-level reconstruction. Bridging this gap re-quires overcoming two core challenges: endowing semantic encoders withhigh-fidelity reconstruction capabilities, and effectively aligning gener-ative models with semantic spaces without relying on external teach-ers. To this end, we propose a novel unified multimodal framework fea-turing Semantic-Pixel self-alignment and Adaptive Routing (SPAR).First, to reconcile semantic perception with pixel-level reconstruction, weintroduce an asymmetric dual-stream unified tokenizer. A lightweightsemantic stream anchors discriminative features, while a Transformer-augmented pixel stream recovers fine-grained visual details into a uni-fied compact latent space. Second, to eliminate external dependencies, wepropose a self-aligned generation paradigm that natively leverages thisoptimized tokenizer as an internal alignment teacher for the diffusionmodel. Furthermore, to facilitate flexible multimodal interaction withinthis unified space, we introduce Dynamic Token Routing, which enableseach token to adaptively aggregate multi-layer MLLM features based onits distinct semantic demands. Extensive experiments demonstrate thatSPAR establishes the state-of-the-art for unified architectures, achiev-ing exceptional generation and reconstruction quality while preservingfoundational visual understanding capabilities.

Visual analogy learning enables image editing via demonstration rather than textual description, allowing users to specify complex transformations difficult to articulate in words. Given a triplet {a, a′, b}, the goal is to generate b′ such that a : a′ :: b : b′. Recent methods adapt text-to-image models with a single Low-Rank Adaptation (LoRA) module, but they face a fundamental limitation: attempting to capture the diverse space of visual transformations within a fixed module constrains generalization. Inspired by recent work showing that LoRAs in constrained domains span meaningful, interpolatable semantic spaces, we propose LoRWeB, which specializes the model for each analogy task in a single inference pass. LoRWeB dynamically composes learned transformation primitives, informally, choosing a point in a “space of LoRAs”. We introduce two key components: (1) a learnable basis of LoRAs to span the space of different visual transformations, and (2) a lightweight encoder that dynamically weighs these basis LoRAs given the input analogy pair. Comprehensive evaluations demonstrate state-of-the-art performance and significantly improved generalization to unseen transformations. Our findings suggest LoRA basis decompositions are a promising direction for flexible visual manipulation tasks. See our website for code.

Incremental Vision-Language Object Detection (IVLOD) em-powers pre-trained open-vocabulary detectors to continuously learn newvisual concepts from sequential tasks without forgetting foundationalknowledge. However, existing methods predominantly rely on oracle taskidentities during inference, severely limiting their open-world practical-ity. In this paper, we formalize Task-Agnostic IVLOD (TA-IVLOD), amore realistic setting where such priors are unavailable, requiring themodel to simultaneously recognize classes from all learned tasks. Thisunconstrained setting exposes severe performance degradation due tocross-task semantic interference and parameter conflicts. To tackle thesechallenges, we propose TADA, a novel modular framework. Specifically,TADA introduces Stochastic Prompt Augmentation to mitigate semanticinterference via training-time noise injection, and Test-Time Distribution-Aware Fusion to dynamically weight class-specific LoRA experts, effec-tively resolving parameter conflicts. Extensive evaluations demonstratethat our method significantly outperforms baselines in standard IVLODand yields substantial improvements in the rigorous TA-IVLOD set-ting on the ODinW-13 benchmark, while effectively preserving zero-shot generalization on the MS COCO dataset. The code is available athttps://github.com/yonghanjiang/TADA.

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.