ECCV 2026 Papers

Transformer-based object trackers are renowned for theirstrong performance, yet dense token processing often leads to prohibitivecomputational cost, limiting real-time deployment on edge devices. Whilerecent works explore token pruning to reduce computation, they oftenstop short of an end-to-end sparse pipeline, as early-layer token scores canbe noisy without a motion prior, and many trackers ultimately fall back todense reshaping to feed the dense prediction head that partially negatesthe savings. We introduce Motion-aware Sparse Tracker (MaST), a sparsetracking framework that makes sparsity effective from tokens to boxes.First, MaST injects a lightweight motion prior to refine cross-attention-based importance scores, enabling earlier and more stable token reductionin the search region. Second, we introduce a natively sparse predictionhead that operates directly on the retained unstructured tokens witha score-first, regress-once design, eliminating dense padding/reshapingand reducing redundant computation. Extensive experiments on multiplebenchmarks demonstrate that MaST establishes new state of the artamong lightweight trackers, where MaST-tiny attains 63.8 AUC on LaSOTand 80.1 SUC on TrackingNet, surpassing the prior best AsymTrack-Sby +1.0 AUC and +2.2 SUC while running at 152 FPS on Jetson Nano,nearly twice as fast as AsymTrack-S at 88 FPS. Code is available atgithub.com/TsingWei/MaST.

Deep vision models exploit shortcuts, relying on cues thatcorrelate with supervision signals. Prior work has focused on visiblebiases, such as object-background or texture correlations. We identifya different source of shortcut learning: invisible metadata traces em-bedded at the pixel level, for metadata such as image processing andphoto acquisition. We hypothesize that large-scale semantic supervi-sion, whether through categorical labels (ImageNet) or billion-scale cap-tions (LAION), naturally induces metadata-semantics correlations dur-ing pretraining, leading models to convert low-level signals into predic-tive features. By introducing controlled metadata-semantics correlations,we show that stronger ones produce systematically higher sensitivity tometadata traces and larger performance degradation under metadatadistribution shifts. We further explore mitigation strategies applied dur-ing and after pretraining that reduce sensitivity not only to targetedmetadata but also to unseen ones, without sacrificing performance ondownstream tasks. Metadata sensitivity also has a positive side: it partlyexplains the strong generated-image detection ability of some encoders,while its mitigation can improve out-of-distribution generalization. Code:

Deep learning based object detectors require trustworthinessbeyond competitive detection performance, but deep neural networks areprone to overconfident predictions, assigning high confidence scores topredictions that are likely to be inaccurate. To improve the alignmentbetween confidence scores and prediction accuracy, existing methodscalibrate confidence scores based on box-level localization, such as pre-cision or intersection over union with the ground truth bounding box.However, box-level localization reflects only a measure of agreement be-tween the predicted box and the ground truth, resulting in calibratedconfidence scores for box-level accuracy failing to capture the localizationaccuracy of coordinates of box. To tackle this issue, we propose a novelpost-hoc calibration framework, rethinking detection calibration (ReDC),which provides reliable coordinate-level confidence scores, including di-rectional information. The proposed framework defines coordinate-wisealignment and deviation direction between predictions and ground truth.Based on the alignment measure, confidence re-encoding produces reliablecoordinate-level confidence scores, while directional displacement estima-tion predicts coordinate-wise deviation directions. Extensive experimentsunder in-domain and out-domain scenarios demonstrate that the proposedapproach expresses the coordinate-wise localization of detected objectsmore precisely than existing methods. Furthermore, our method coversthe representational scope of prior calibration approaches by aggregatingcoordinate-level confidence scores into box-level localization.

Empowered by large-scale training, vision-language models (VLMs) achieve strong image and video understanding, yet their ability to perform spatial reasoning in both static scenes and dynamic videos remains limited. Recent advances try to handle this limitation by injecting geometry tokens from pretrained 3D foundation models into VLMs. Nevertheless, we observe that naive token fusion followed by standard finetuning in this line of work often leaves such geometric cues underutilized for spatial reasoning, as VLMs tend to rely heavily on 2D visual cues. In this paper, we propose GeoSR, a framework designed to make geometry matter by encouraging VLMs to actively reason with geometry tokens. GeoSR introduces two key components: (1) Geometry-Unleashing Masking, which strategically masks portions of 2D vision tokens during training to weaken non-geometric shortcuts and force the model to consult geometry tokens for spatial reasoning; and (2) Geometry-Guided Fusion, a gated routing mechanism that adaptively amplifies geometry token contributions in regions where geometric evidence is critical. Together, these designs unleash the potential of geometry tokens for spatial reasoning tasks. Extensive experiments on both static and dynamic spatial reasoning benchmarks demonstrate that GeoSR consistently outperforms prior methods and establishes new state-of-the-art performance by effectively leveraging geometric information. The project page is available at https://suhzhang.github.io/GeoSR/.

Vision-Language-Action (VLA) policies fine-tuned with RLand frozen VLM reward evaluators can produce observations that lookcorrect without physically completing the task—a phenomenon we for-malize as stage hallucination. To suppress it, we propose EvoVLA, aself-supervised framework with three synergistic modules: Stage-AlignedReward (SAR) penalizes near-miss states via counterfactual hard nega-tives in CLIP-based scoring; Pose-Based Object Exploration (POE)grounds curiosity in relative gripper–object pose; and Long-HorizonMemory uses selective context retrieval with gated fusion to stabi-lize credit assignment. We also introduce Discoverse-L, a long-horizonbenchmark (3 tasks, 18–74 stages) with a formal Hallucination Rate met-ric. Under matched backbones and budgets, EvoVLA achieves 69.2% suc-cess (+10.2 over OpenVLA-OFT), 1.5× sample efficiency, and reduceshallucination from 38.5% to 14.8%; real-robot deployment attains 54.6%(+11.0) across four tasks, confirming Sim2Real transfer.

Open-vocabulary scene understanding aims to recognize ar-bitrary semantic concepts in 3D beyond a fixed set of categories. Re-cent methods built upon 3D Gaussian Splatting (3DGS) enable sceneunderstanding by distilling features from vision–language models, butmost rely on per-scene optimization and dense observations, which limitscalability. More recent generalizable 3DGS methods show promise, butthey perform poorly on sparse-view satellite imagery due to the uniquepushbroom imaging mode, limited geometric constraints, and extremescale variations. To address these limitations, we propose SkySplat-OV,a feed-forward framework that integrates the rational polynomial co-efficient (RPC) model into a generalizable language 3DGS pipeline. AGeometry–Radiance–Semantics Fusion module, together with monoc-ular relative height supervision, improves Gaussian parameter predic-tion under limited geometric constraints. Moreover, SkySplat-OV em-ploys single-stage SAM3 distillation to handle extreme scale variationsin satellite imagery for accurate segmentation. Compared to per-sceneoptimization methods, SkySplat-OV achieves up to 1070 times speedupover LangSplatV2 with higher accuracy. It also outperforms generaliz-able 3DGS baselines, achieving 12.91%–25.75% mIoU improvements ontwo benchmarks, and demonstrates strong cross-dataset generalizationon the MVS3D test set. Additional qualitative results are available athttps://skysplat-ov.github.io/.

Image restoration aims to recover high-quality images fromdegraded observations, yet existing methods struggle with diverse degra-dation types. In this work, we propose SLAIR, an all-in-one image restora-tion framework integrating latent space separation and deterministictransport. We introduce a latent-space separation approach that ensuresstable, distinct representations of degraded and clean images, facilitat-ing effective flow matching. By learning continuous transport flow in alatent space via an ordinary differential equation (ODE), our methodenables deterministic restoration without relying on degradation labelsor stochastic sampling. While the restoration is performed in the la-tent space, a high-frequency refinement module is introduced into thelatent decoder to handle detail distortion during decoding. Experimentsdemonstrate that SLAIR outperforms existing approaches in terms ofboth perceptual quality and computational efficiency, making it a robustsolution for scalable, degradation-agnostic image restoration. Our projectpage is available at https://fightforql.github.io/SLAIR-page.

Camera traps are crucial for large-scale biodiversity monitor-ing, yet accurate automated analysis remains challenging due to diversedeployment environments. While the computer vision community haspredominantly framed this challenge as cross-domain (e.g., cross-site)generalization, this perspective overlooks a primary challenge faced byecological practitioners: maintaining reliable recognition at the fixed siteover time, where the dynamic nature of ecosystems introduces profoundtemporal shifts in both background and animal distributions. To bridgethis gap, we present the first unified study of camera-trap speciesrecognition over time. We introduce a realistic, large-scale bench-mark comprising 546 camera traps with a streaming protocol that evalu-ates models over chronologically ordered intervals. Our end-user-centricstudy yields four key findings. (1) Biological foundation models (e.g.,BioCLIP 2) underperform at numerous sites even in initial intervals,underscoring the necessity of site-specific adaptation. (2) Adaptation ischallenging under realistic evaluation: when models are updated usingpast data and evaluated on future intervals (mirrors real deployment life-cycles), naive adaptation can even degrade below zero-shot performance.(3) We identify two main drivers of this di!culty: severe class imbalanceand pronounced temporal shift in both species distribution and back-grounds between consecutive intervals. (4) We find that e"ective integra-tion of model-update and post-processing techniques can largely improveaccuracy, though a gap from the upper bounds remains. Finally, we high-light critical open questions, such as predicting when zero-shot modelswill succeed at a new site and determining whether/when model updatesare necessary. Together, our benchmark and analysis provide actionabledeployment guidelines for ecological practitioners while establishing newdirections for future research in vision and machine learning.

Localizing a font into new languages is a highly intricate task requiring precisedesign adaptation of glyphs, color/texture, and spacing/kerning, from source to target lan-guages. Most existing methods focus on single glyph generation with limited capability inhandling multilingual font rendering. In this work, we propose LoGAN, a VLM-based agenticframework for few-shot multilingual font localization, which takes in a small number of indi-vidual glyphs from a font or letters from a logo and uses them to generate complete charactersets in other languages. LoGAN breaks down this task into multiple components: a glyph-leveldiffusion model, a style finetuning module, a spacing and kerning transfer algorithm, and atexture expansion model, with a VLM agent coordinator. LoGAN achieves broad languagecoverage for font localization with various styles, including Chinese/Korean/Japanese (CJK).We evaluate our approach on both font and real-world logo datasets spanning more than27 languages and compare it against both specialized font generation and state-of-the-artimage editing models with strong text rendering capabilities (e.g., FLUX, Nano-Banana).Our approach yields higher glyph fidelity while maintaining better style, texture, and kerningconsistency according to both quantitative and qualitative evaluations.

Multimodal object tracking has received extensive attentiondue to its strong robustness and complementary collaboration. However,most existing methods only evaluate the cross-modal fusion at specificmulti-modal layers, and overlook the interaction among different fea-ture layers from single modal, which limits the utilization and propa-gation of cross-modal complementary information and consequently de-grades tracking performance. To alleviate this issue, we propose a pro-gressive multimodal tracker based on State Space Model (SSM) namedPSMTrack, which consists of three key modules. First, Difference-guidedBidirectional State-space Mamba Enhancement (DBME) is developedto leverage difference-aware cues and bidirectional SSM modeling toprogressively enhance modality representations for subsequent tracking,while suppressing background and similar distractions. Second, we adoptan expert routing mechanism to adaptive select a subset of layers andperform cross-layer interaction to alleviate the information loss causedby relying solely on the last layer. Finally, we design a Hybrid SpiralMamba Fusion (HSMF) module that performs dynamic spatial modelingwith spiral Mamba from both local and global scales for final cross-modalfusion, thereby capturing long-range dependencies with near-linear com-plexity and improving bounding-box regression stability. Extensive ex-periments on multiple mainstream RGB-X benchmarks validate that ourmodel consistently improves both accuracy and robustness.

Accurate dialogue description in audiovisual video captioning is crucial for downstream understanding and generation tasks. However, existing models generally struggle to produce captions that faithfully reflect spoken content and speaker dynamics. To mitigate this limitation, we propose DiaDem, a powerful audiovisual video captioning model capable of generating captions with more precise dialogue descriptions while maintaining strong overall performance. We first design a dedicated pipeline to synthesize high-quality dialogue-aware audiovisual captions for supervised fine-tuning (SFT), and then introduce a difficulty-partitioned two-stage reinforcement learning (RL) strategy to further enhance dialogue descriptions. To enable systematic evaluation of dialogue description capabilities, we present DiaDemBench, a comprehensive benchmark designed to evaluate models across diverse dialogue scenarios, focusing on both speaker attribution accuracy and utterance transcription fidelity in audiovisual captions. Extensive experiments on DiaDemBench reveal even commercial models still exhibit substantial room for improvement in dialogue-aware captioning. Notably, DiaDem not only outperforms the Gemini series in dialogue description accuracy but also achieves competitive performance on general audiovisual captioning benchmarks, demonstrating its overall effectiveness. Our project is available at https://diadem-captioner.github.io/.

Automating filament tracing in Cryo-Electron Microscopy(Cryo-EM) is essential for 3D helical reconstruction but challenged byintersecting topologies and extremely low Signal-to-Noise Ratios (SNR =σs2 /σn2 < 0.1 or -10 dB). Existing paradigms fail: pixel-wise segmenterssuffer from severe topological fracturing, box-based detectors face ghostcenter drift, sequential trackers derail due to error accumulation, and tra-ditional active contours collapse under artificial closed-curve constraints.To resolve these bottlenecks, we present FPicker, the first topology-guided framework reconciling these incompatibilities. It unifies percep-tion via a center-endpoint representation and an open-curve evolutionmodule to explicitly model non-cyclic connectivity. On simulated bench-marks, FPicker outperforms top baselines by over 40% relative gain inmean spatio-angular precision (mSAP) and reduces topological gap ratesby over 60% under extreme noise (−20 dB). By learning intrinsic phys-ical geometry rather than local texture, FPicker demonstrates strongpotential as a resilient geometric backbone. Its zero-shot performance onthe real-world EMPIAR dataset exhibits robust topological resistance,achieving a state-of-the-art 82.9% mSAP upon fine-tuning. Our resultsalso suggest modeling physical priors is a highly robust path towardbridging the sim-to-real gap in signal-starved scientific imaging. The codeis publicly available at: https://github.com/tomzhaosky/FPicker.

Due to the scarcity of large-scale in-the-wild triplet data and the improper use of masks, the performance of video virtual try-on models remains limited. In this paper, we first introduce TripVVT-10K, the largest and most diverse in-the-wild triplet dataset to date, providing explicit video-level cross-garment supervision that existing video datasets lack. Built upon this resource, we develop TripVVT, a Diffusion Transformer–based framework that replaces fragile garment masks with a simple, stable human-mask prior, enabling reliable background preservation while remaining robust to real-world motion, occlusion, and cluttered scenes. To support comprehensive evaluation, we further establish TripVVT-Bench, a 100-case benchmark covering diverse garments, complex environments, and multi-person scenarios, with metrics spanning video quality, try-on fidelity, background consistency, and temporal coherence. Compared to state-of-the-art academic and commercial systems, TripVVT achieves superior video quality and garment fidelity while markedly improving generalization to challenging in-thewild videos. We publicly release the dataset and benchmark at https://huggingface.co/datasets/TripVVT/TripVVT-10K, which we believe provide a solid foundation for advancing controllable, realistic, and temporally stable video virtual try-on.

Given the remarkable ability of 2D foundation image modelsto generate high-fidelity outputs, we investigate a fundamental question:do 2D foundation image models inherently possess 3D world model ca-pabilities? To answer this, we systematically evaluate multiple state-of-the-art image generation models and Vision-Language Models (VLMs)on the task of 3D world synthesis. To harness and benchmark their po-tential implicit 3D capability, we propose an agentic framing to facilitate3D world generation. Our approach employs a multi-agent architecture:a VLM-based director that formulates prompts to guide image synthesis,a generator that synthesizes new image views, and a VLM-backed two-step verifier that evaluates and selectively curates generated frames fromboth 2D image and 3D reconstruction space. Crucially, we demonstratethat our agentic approach provides coherent and robust 3D reconstruc-tion, producing output scenes that can be explored by rendering novelviews. Through extensive experiments across various foundation mod-els, we demonstrate that 2D models do indeed encapsulate a grasp of3D worlds. By exploiting this understanding, our method successfullysynthesizes expansive, realistic, and 3D-consistent worlds.

Hierarchical structure is common in image data, where finegrained clusters often merge into larger, coarser semantic clusters. In biological cell images, current self-supervised learning models tend to suppress the inherent hierarchical structure, as coarse factors (such as the imaging modality) systematically obscure finer attributes (e.g., morphological details) in the latent space. We propose a novel hierarchy-aware self-supervised training framework to address this problem. Our method integrates two components: we propose (i) a distillation framework by introducing a segmentation teacher that improves the latent space’s morphological awareness, and (ii) a hierarchy-aware contrastive loss based on HDBSCAN to improve decision boundaries between closely related yet distinct subtypes at each hierarchical level. Both components counteract the self-supervised learning’s tendency to overemphasize coarse factors by aligning embeddings with semantic and morphological cues, yielding biologically meaningful sub-clusters driven by fine morphological detail. We train and evaluate our method on a curated corpus of 2.3M single cells aggregated from 20 microscopy datasets (labeled and unlabeled), covering 208 cell classes. Our method improves upon the counterparts and baseline, with the average top-K accuracy up by +2.8%, top-9 retrieval on the dataset with a deeper hierarchy up by +6.3%, and an F1-score on biologically relevant classification of administered drug from perturbed cell morphology downstream up by +7.8%.

We present FFAvatar, a Transformer-based 3D Gaussian framework for fast construction of high-quality and animatable 4D head avatars from one or more reference portrait images. Unlike existing feed-forward approaches that require a fixed number of input views, FFAvatar supports incremental reconstruction, progressively refining the avatar representation as additional reference images become available. At the core of our method is an alternating attention mechanism that disentangles identity appearance from expression and viewpoint variations, enabling the reconstruction of a canonical 3D appearance that remains consistent across poses and facial expressions. To balance visual fidelity and computational efficiency, we introduce a sparse-to-dense learning paradigm. Coarse appearance features are first learned using sparse primitives anchored to the FLAME vertex level and are subsequently densified in the UV domain to capture fine-grained geometric and texture details. We further propose a plug-and-play motion refinement module that enables subject-specific dynamic personalization by modeling residual motion beyond parametric deformation. Extensive experiments demonstrate that FFAvatar efficiently produces high-fidelity and controllable 4D head avatars, achieving superior flexibility, driving efficiency, and identityconsistent rendering across diverse expressions and viewpoints. Project Page: https://jj-yao.github.io/ffavatar/

Instruction-guided image editing has a training-time blindspot. Generative editors are never required to semantically verify whethertheir outputs actually satisfy the instruction. Supervision stops at recon-struction and input textual-level conditioning. This produces incompleteedits, spatial spillover, and poor localization. We present IABEdit, amodel-agnostic framework that embeds differentiable semantic verifica-tion into training. A frozen vision-language model extracts spatially-aware descriptors from the ground-truth edit. A trainable aligner thenreproduces them from the generated output. The residual between thetwo becomes a gradient that teaches the generator both what to editand where, with no inference-time VLM cost. IABEdit is compatiblewith diverse backbones, including U-Net (Stable Diffusion) and MMDiT(FLUX), without altering their inference pipelines. On MagicBrush, itimproves structural fidelity by +3.49 DINO-I over the best diffusion base-line and +1.26 over the best overall baseline, while remaining competitiveon instruction alignment. It also achieves state-of-the-art instruction ad-herence performance on RealEdit and EMU Edit benchmarks based onembedding-based metrics. Most consequentially, on the D-LORD surveil-lance benchmark, it surpasses the proprietary Gemini agent by +5.13DINO-P under heavy occlusion, where preserving identity is hardest.This shows that gradient-aligned VLM distillation holds up under real-world-like surveillance and occlusion conditions. Human and GPT-4oevaluations confirm perceptually precise, well-localized edits.

Cross-view geo-localization matches ground-level observationsagainst geo-tagged satellite imagery. Recent methods show that sequen-tial queries such as video clips yield richer spatiotemporal cues thansingle images, yet they overlook a complementary sequential modality:route descriptions—which capture the same trajectory at a higher levelof abstraction and are often the only input available (e.g., a user direct-ing an autonomous vehicle to a pickup point). To bridge this gap, weintroduce SeqGeo-VL, a dataset of ∼39K video–text–satellite triplets,and TrajLoc, a unified framework capable of processing both video clipsand route descriptions. By leveraging both dense visual and abstract lin-guistic semantics, TrajLoc enables these modalities to mutually reinforcecross-view matching. We further propose TrajMod, a lightweight mod-ule that conditions query embeddings on trajectory geometry, yieldingspatially-aware representations. Experiments show that TrajLoc achievessubstantial gains over state-of-the-art methods on both video and textgeo-localization. Code, model weights, and the dataset are released athttps://humblegamer.github.io/trajloc/.

Significant progress has been achieved in subject-driven text-to-image (T2I) generation, which aims to synthesize new images depictingtarget subjects according to user instructions. However, evaluating thesemodels remains a significant challenge. Existing benchmarks exhibit criti-cal limitations: 1) insufficient diversity and comprehensiveness in subjectimages, 2) inadequate granularity in assessing model performance acrossdifferent subject difficulty levels and prompt scenarios, and 3) a pro-found lack of actionable insights and diagnostic guidance for subsequentmodel refinement. To address these limitations, we propose DSH-Bench,a comprehensive benchmark that enables systematic multi-perspectiveanalysis of subject-driven T2I models through four principal innovations:1) a hierarchical taxonomy sampling mechanism ensuring comprehensivesubject representation across 58 fine-grained categories, 2) an innovativeclassification scheme categorizing both subject difficulty level and promptscenario for granular capability assessment, 3) a novel Subject IdentityConsistency Score (SICS) metric demonstrating a 9.4% higher correlationwith human evaluation compared to existing measures in quantifyingsubject preservation, and 4) a comprehensive set of diagnostic insightsderived from the benchmark, offering critical guidance for optimizing fu-ture model training paradigms and data construction strategies. Throughan extensive empirical evaluation of 19 leading models, DSH-Bench un-covers previously obscured limitations in current approaches, establishingconcrete directions for future research and development.