ECCV 2026 Papers

Recently, 3D Gaussian Splatting (3DGS), which succeeded in perspective image rendering, has been extended to handle the panoramic image rendering task. However, the performance of the existing methods in literature is generally limited due to the severe distortions involved in panoramic images. To address this problem, we construct a distortion map under the panoramic imaging model, where the value of each pixel reflects its corresponding distortion degree, and accordingly, we propose a novel Panoramic Gaussian Splatting method by utilizing this Distortion map, called PaD-GS. The proposed PaD-GS, consisting of a distortion-aware Gaussian decoding module and a distortion-aware opacity modulation module, employs a whole-to-partial strategy to impose the distortion map on Gaussian representation learning. Specifically, the distortion-aware Gaussian decoding module is designed to decode the ensemble of learnable scene features and the distortion map into a set of Gaussian representations, so that the intrinsic distortion information of the distortion map could be injected into the whole Gaussian representations. The distortion-aware opacity modulation module is designed to further impose the distortion map by adaptively modulating the partial attribute (opacity) of each Gaussian representation. Thanks to the introduced distortion map, the proposed PaD-GS could alleviate the negative influence of severe distortion involved in panoramic images. Extensive experimental results on two panoramic datasets demonstrate that PaD-GS significantly outperforms several state-of-the-art methods for panoramic image rendering in most cases. The code is available at https://github.com/CosyXu/PaD-GS.

The deployment of deep learning in digital pathology holds transformative potential for cancer diagnosis, yet its clinical adoption is hindered by the opaque and vulnerable nature of black-box models. In this work, we introduce a mechanistic examination framework that combines sparse autoencoders and activation patching to dissect adversarial vulnerabilities in pathology models as several learned features may correspond to non-morphological artifacts rather than clinically meaningful patterns. By isolating the neural circuits responsible for these vulnerabilities, we demonstrate that models often rely on spurious, nongeneralizable cues, particularly in middle layers of Vision Transformers and later convolutional blocks of ResNets. To address this, we propose Mechanism-Informed Adversarial Training, a novel approach that targets vulnerable circuits with precision regularization, achieving state-of-theart robustness while preserving clean performance. Crucially, our framework aligns with rapidly emerging clinical and regulatory demands for alignment, accountability, and safety in AI-driven diagnostics. By providing a mathematically rigorous, interpretable, and actionable method to audit and fortify deep learning models, we seek to bridge the gap between AI research and clinical deployment, offering a pathway to trustworthy, regulation-ready AI in healthcare. Our findings not only inform the field of adversarial machine learning but also set a precedent for mechanistically grounded artificial intelligence in real-world medical applications.

Generating high-fidelity and controllable 3D assets from a single image remains a significant challenge. We propose a novel framework that integrates preference optimization into a rectified flow-based diffusion model operating on a latent set representation. This approach enhances geometric fidelity and user alignment, enabling the efficient synthesis of diverse 3D formats. Our core contribution is a versatile reward-based optimization strategy for aligning generation with human-centric criteria like surface quality. We introduce Direct Reward Optimization (DRO), which fine-tunes the model using only absolute quality feedback, removing the need for pairwise data. The framework also seamlessly incorporates Direct Preference Optimization (DPO) when paired data is available. Parameter-efficient control is achieved via Low-Rank Adaptation (LoRA). To ensure data quality, we also develop a robust preprocessing pipeline using a neural SDF reconstruction method with Gaussian curvature constraints to produce high-quality watertight meshes. Extensive experiments show our framework achieves state-of-the-art performance. Quantitative evaluations demonstrate significant improvements in Chamfer Distance and F-Score, while qualitative results confirm superior geometric fidelity and user preference alignment compared to existing methods.

Recent advances in video world models enable interactive environments with free navigation, making translation between first-person (egocentric) and third-person (exocentric) perspectives increasingly important. However, existing studies focus on unidirectional exocentric-toegocentric translation, overlooking reference-guided exocentric perspective synthesis. This capability is crucial for gaming and embodied AI applications. Motivated by this, we present WorldWander, an in-context learning framework tailored for translating between egocentric and exocentric worlds in video generation. Building upon advanced video diffusion transformers, WorldWander integrates (i) In-Context Perspective Alignment and (ii) Collaborative Position Encoding to model cross-view synchronization and character consistency. To support our task, we curate EgoExo-8K, a dynamic and scene-rich dataset containing synchronized egocentric–exocentric triplets from both synthetic and real-world scenarios. Experiments demonstrate that WorldWander achieves superior perspective synchronization, character consistency, and generalization, setting a new benchmark for egocentric-exocentric video translation.

In this work, we propose a source-agnostic framework that dynamically refines a binary mask throughout the reverse diffusion process by computing the discrepancies of a pretrained diffusion model’s prediction for each latent time step. Rather than relying on a fixed threshold, our method introduces a time-dependent statistical thresholding scheme derived from the empirical mean and standard deviation of prediction discrepancies across the latent noisy images from the target distribution. This allows the mask to adapt to the model’s varying predictive confidence at different noise levels, effectively isolating domain-specific regions while preserving global structural coherence. Experimental results on the AFHQ and Celeba-HQ datasets demonstrate that our approach outperforms state-of-the-art unsupervised Image-to-Image methods in both realism (FID, KID) and faithfulness (SSIM, LPIPS). By requiring only a pretrained model of the target domain, our approach enables precise, automated localization and seamless translation across diverse source distributions without any specialized training. The project source code is available at: https://github.com/dtoma95/PM-Edit

Advanced Large Multimodal Models (LMMs) have demonstrated impressive performance in K-12 reasoning tasks, exhibiting great promise as intelligent tutors. Realizing this potential requires models to navigate real-world examinations effectively, yet most existing benchmarks fail to capture the complexity of authentic testing environments. Specifically, most datasets are static, prone to data contamination, and are often confined to restricted modalities, disciplines, and evaluation criteria. To address these issues, we introduce LiveK12Bench, a dynamic, holistic, multi-disciplinary benchmark designed to evaluate the reasoning abilities of LMMs in realistic examination scenarios. LiveK12Bench comprises 2K+ verified questions spanning Mathematics, Physics, Chemistry, and Biology, sourced from the latest real-world exam papers and designed to grow over time. Our framework features several core innovations: 1) featuring an automated pipeline that continuously ingests and parses latest examination papers to mitigate data leakage; and 2) proposing a novel ’Mock Exam’ evaluation scheme, which assesses the ability to complete end-to-end exams autonomously with accurate and efficient reasoning paths. Extensive experiments on 12 LMMs reveal that advanced models suffer substantial performance degradation under exam-realistic constraints: GPT-5’s score drops from 79 to 53 (out of 100) when process rigor and efficiency are jointly evaluated. Our findings expose critical vulnerabilities, such as sensitivity to complex visual layouts, highlighting the gap between idealized reasoning capabilities and true educational readiness. Both code and dataset are publicly available.

Cancer segmentation models can fail silently, generating plausible but incorrect masks that risk missed findings or unnecessary biopsies. A critical question arises: Do AI models “know” when they are wrong, and if so, can we use the signal to predict their own failures? Humans do have a “Feeling of Error” (FOE): a spontaneous sense of unease that flags a potential error during thinking. We investigate whether cancer segmentation models exhibit an analogous internal signal. Unlike output-level cues (e.g., prediction confidence or uncertainty), which offer no insight into why a failure occurs and suffer from a sensitivity–quality tradeoff where high detection sensitivity could degrade overall segmentation quality. We instead propose to capture the model’s FOE from its inner workings. Using mechanistic interpretability tools, specifically Sparse Autoencoders, we decompose internal neural activations into a dictionary of human-interpretable concepts and show that failure cases exhibit a distinct latent signature: fewer active concepts with lower activation magnitudes compared to successful segmentation. By training a classifier on these concept activations, we achieve accurate failure detection along with explanations for the model’s mistakes. Experiments on prostate, pancreatic, and brain cancer segmentation demonstrate that our approach outperforms output-based methods in failure detection while preserving segmentation quality. Code is abailable at https://github.com/deep-real/CancerSegFailure.

CLIP, known for its strong zero-shot generalization, has received significant attention in prompt tuning for its ability to adapt effectively to new tasks in few-shot settings. Recent studies have employed descriptions generated by Large Language Models (LLMs) as predefined prompts to obtain class-level semantics. However, because these predefined descriptions are not directly grounded in visual information, they often misalign with an image’s visual semantics. They tend to generate generic class-level descriptions rather than instance-specific ones. As a result, LLM-based prompts are structurally biased toward class-level semantics, making it difficult to capture and reflect instancelevel cues, which in turn limits generalization in few-shot settings where both levels of knowledge are crucial. To address this limitation, we propose Disentangled Instance-Class knowlEdge (DICE) prompt tuning, a framework that restructures LLM-derived class-level priors to capture instance-level semantics. Specifically, DICE decomposes LLM-derived priors into instance-level components using a Sparse Autoencoder (SAE), which selects instance-specific concept vectors. These concept vectors are then fused with class embeddings to form enriched representations. This synergy preserves class-level semantic coherence while capturing instance-level details, improving generalization to unseen classes. Our approach achieves competitive performance across 11 few-shot recognition benchmarks, while additionally offering a plug-and-play solution that enhances interpretability through SAE.

Large-scale Vision–Language Models (VLMs) such as CLIP learn powerful semantic representations but operate in Euclidean space, which fails to capture the inherent hierarchical structure of visual and linguistic concepts. Hyperbolic geometry, with its exponential volume growth, offers a principled alternative for embedding such hierarchies with low distortion. However, existing hyperbolic VLMs use entailment losses that are unstable: as parent embeddings contract toward the origin, their entailment cones widen toward a half-space, causing catastrophic cone collapse that destroys the intended hierarchy. Additionally, hierarchical evaluation of these models remains unreliable, being largely retrieval-based and correlation-based metrics and prone to taxonomy dependence and ambiguous negatives. To address these limitations, we propose an adaptive entailment loss paired with a norm regularizer that prevents cone collapse without heuristic aperture clipping. We further introduce an angle-based probabilistic entailment protocol (PEP) for evaluating hierarchical understanding, scored with AUC-ROC and Average Precision. This paper introduces a stronger hyperbolic VLM baseline ARGENT, Adaptive hieRarchical imaGe-tExt represeNTation. ARGENT improves the SOTA hyperbolic VLM by 0.7, 1.1, and 0.8 absolute points on image classification, text-to-image retrieval, and proposed hierarchical metrics, respectively.

Concept Bottleneck Models (CBMs) are a class of interpretable AI models that enable concept-level control over the decisionmaking process, allowing humans to directly intervene in model predictions. As a result, CBMs have been widely adopted in tasks requiring interpretability and controllability. Existing research has demonstrated that human-model interaction can signi(cid:28)cantly enhance CBM performance. However, in practical applications, each intervention requires manual human e(cid:27)ort and explicit data provision. Therefore, a central challenge is how to improve model performance while minimizing intervention e(cid:27)ort. In this paper, we show that trained CBM parameters encode usable concept-correlation structure. We leverage this structure during interaction by propagating user feedback through the shared inputto-concept pathway, thereby adjusting non-intervened concepts without training an additional realignment model. Experimental results demonstrate that our method improves low-intervention performance across multiple models and datasets. Compared with all tested baselines, our approach improves early-intervention performance by approximately 9%; compared with Concept Realignment, it improves AUC@10 and AUC@20 by 3.90% and 2.88%, respectively, while avoiding extra training. Code is available at https://github.com/LMBTough/CI. • •

Video temporal grounding (VTG) aims to localize relevant moments and predict saliency scores in untrimmed videos based on natural language queries. Recently, DETR-inspired methods have achieved instance-level moment predictions through query-based semantic evidence aggregation. However, unlike spatial objects with clear physical edges, video moments often exhibit ambiguous temporal boundaries due to the semantic gradient. Consequently, relying solely on global semantics can lead to severe boundary misalignment, yielding sub-optimal grounding results. To address this issue, we propose a novel Dual Evidence Detection Transformer (DE2TR) framework for better boundary localization. In particular, we introduce a dual-branch decoder that simultaneously captures global semantic evidence and fine-grained boundary cues. To prevent feature conflict between these distinct objectives, we channel-wise decouple the moment queries, facilitating independent evidence learning under the supervision of evidence alignment losses. Building upon this dual-branch design, we introduce a prior-guided refinement mechanism that employs explicit prior maps to steer queries toward potential regions of interest. Furthermore, a boundary augmentation strategy synthesizes ideal boundary patterns via simple temporal splicing, effectively enhancing the model’s boundary awareness. Extensive experiments on multiple popular benchmarks demonstrate that DE2TR significantly outperforms state-of-the-art approaches.

The strong generalization capability to diversified atmospheric conditions is crucial for autonomous driving models. However, models trained on existing datasets struggle with real-world complexity due to the homogeneous weather, limited maneuvers, and insufficient scale in training data. To overcome this issue, we introduce DiverseAD, a largescale driving dataset comprising 150K scenes that feature a great diversity in atmospheric condition, road types, and driving actions. Based on this dataset, we further propose a novel end-to-end autonomous driving model robust across diverse atmospheric conditions. More specifically, an atmospheric-invariant feature learning mechanism is proposed, which spots and disentangles atmospheric-agnostic features from visual inputs by using driving intent and scene structure cues as stable anchors. Our method thus allows the extracted features to be more invariant to changes in atmospheric conditions. Experiments show that DiverseAD is superior to existing public datasets in diversity, hence is valuable for training and benchmarking. Extensive comparisons also illustrate the superior performance of our proposed method.

Diffusion MRI of the fetal brain has potential to provide important insights into early development of white matter fibers. However, in vivo fetal acquisitions are severely corrupted by unpredictable motion, making the data unusable for downstream analysis. To address this challenge, we propose an implicit neural representation framework for fetal diffusion MRI reconstruction. The proposed method models the diffusion signal as a spatially and angularly continuous field by predicting direction-independent spherical harmonics (SH) coefficients from continuous spatial coordinates. Based on this representation, we formulate a physics-based forward slice acquisition model that links the underlying clean diffusion volume with the observed motion-corrupted slices, enabling joint optimization of the diffusion signal, fetal motion and acquisition artefacts. Furthermore, the single slice orientation acquisition typically used for each diffusion gradient introduces an ill-posed reconstruction problem along the slice-select direction. To mitigate this issue, we introduce a novel Through-Plane Continuity regularization loss that enforces smoothness in through-plane direction to suppress stripe artifacts, while preserving genuine anatomical details. Experiments on one in vivo fetal dataset and two simulated neonatal datasets demonstrate that the proposed framework achieves improved reconstruction fidelity, sharper anatomical features, and stronger robustness to acquisition artifacts. The code is available at: https://github.com/baby-MedIA/INR-dMRI-Recon

3D learning systems implicitly assume that objects occupy a coherent reference frame. Nonetheless, in practice, every asset arrives with an arbitrary global rotation, and models are left to resolve directional ambiguity on their own. This persistent misalignment suppresses pose-consistent generation, and blocks the emergence of stable directional semantics. To address this issue, we construct CanoVerse, a massive canonical 3D dataset of 320K objects over 1,156 categories – an order-ofmagnitude increase over prior work. At this scale, directional semantics become statistically learnable: CanoVerse improves 3D generation stability, enables precise cross-modal 3D shape retrieval, and strengthens zero-shot point-cloud orientation estimation even for out-of-distribution data. This is achieved by a new canonicalization framework that reduces alignment from minutes to seconds per object via compact hypothesis generation and lightweight human discrimination, transforming canonicalization from manual curation into a high-throughput data generation pipeline. Project page: https://github.com/123321456-gif/Canoverse

Recent generative models can produce high-fidelity videos, yet they often exhibit 3D spatial geometric inconsistencies. Existing evaluation methods fail to accurately characterize these inconsistencies: fidelity-centric metrics like FVD are insensitive to geometric distortions, while consistency-focused benchmarks often penalize valid foreground dynamics. To address this gap, we introduce SGC, a metric for evaluating 3D Spatial Geometric Consistency in dynamically generated videos. We quantify geometric consistency by measuring the divergence among multiple camera poses estimated from distinct local regions. Our approach first separates static from dynamic regions, then partitions the static background into spatially coherent sub-regions. We predict depth for each pixel, estimate a local camera pose for each subregion, and compute the divergence among these poses to quantify geometric consistency. Experiments on real and generative videos demonstrate that SGC robustly quantifies geometric inconsistencies, effectively identifying critical failures missed by existing metrics.

Text and formulas constitute the core informational components of many documents. Accurately and efficiently recognizing both is crucial for developing robust and generalizable document parsing systems. Recently, vision-language models (VLMs) have achieved impressive unified recognition of text and formulas. However, they are largesized and computationally demanding, restricting their usage in many applications. In this paper, we propose UniRec-0.1B, a unified recognition model with only 0.1B parameters. It is capable of performing text and formula recognition at multiple levels, including characters, words, lines, paragraphs, and documents. To implement this task, we first establish UniRec40M, a large-scale dataset comprises 40 million text, formula and mixed samples, enabling the training of a powerful yet lightweight model. Secondly, we identify two challenges when building such a lightweight but unified expert model. They are: structural variability across levels and semantic entanglement between textual and formulaic content. To tackle these, we introduce a hierarchical supervision training that explicitly guides structural comprehension, and a semantic-decoupled tokenizer that separates text and formula representations. Finally, we develop a comprehensive evaluation benchmark covering Chinese and English documents from multiple domains and with multiple levels. Experimental results on this and public benchmarks demonstrate that UniRec-0.1B outperforms both general-purpose VLMs and leading document parsing expert models, while achieving 2-9× speedup, validating its effectiveness and efficiency. Codebase and Dataset: https://github.com/Topdu/OpenOCR.

Understanding human emotions in spoken conversations is a key challenge in affective computing, with applications in empathetic AI, human computer interaction, and mental health monitoring. However, existing datasets vary in scale, emotion distribution, modality alignment, and data partitioning strategies, which can influence reliable crossdataset generalization and minority-emotion modeling. We introduce SpEmoC a Speaking segment Emotion for Conversations comprising 306,544 raw clips from 3,100 English language movies and TV series. From these, 30,000 high quality, class balanced clips are curated, featuring synchronized visual, audio, and textual modalities annotated for seven emotions through a hybrid pipeline that integrates pretrained models with human validation. SpEmoC uses strict movie and series-level splits to prevent content overlap between split sets, allowing more reliable evaluation of model generalization. The dataset also maintains a nearbalanced distribution across seven emotions, including minority classes such as Fear and Disgust, which supports more balanced learning across categories. Extensive experiments, including in-domain benchmarking, cross-dataset transfer, low-data training, class-imbalance analysis, and modality transfer show that balanced data and careful splitting lead to more stable performance across emotions when models are evaluated on other datasets. These results highlight the importance of dataset design for robust and transferable multimodal emotion recognition.

Self-Supervised Monocular Depth Estimation (MDE) has garnered attention in recent years due to its independence from ground truth. However, most existing models are limited to a single scale and exhibit considerable performance degradation in complex driving environments. Networks specifically designed to handle dynamic traffic participants tend to be overly complex, hindering their deployment on resource-constrained automotive edge devices. To address these limitations and move towards robust driving perception, we propose FlexDepth, a scale-driven and flexible family of self-supervised MDE models tailored for challenging road scenarios. FlexDepth employs a two-stage static-dynamic decoupled training strategy, enabling the independent assessment of confidence for both static backgrounds and dynamic road objects. Furthermore, it introduces a meticulously designed Scale-Driven Decoder (SDD) to dynamically select components based on scale size, facilitating efficient feature fusion and the output of high-precision depth maps. Extensive experiments on standard driving benchmarks demonstrate that without any auxiliary information, our model achieves state-of-the-art performance across arbitrary scales with minimal computational overhead. Our smallest model, Flex-Nano, requires only 0.7 GFLOPs and achieves 37.6 FPS on mobile platforms, ensuring reliable real-time perception while maintaining excellent zero-shot generalization. Our source code is avalible: https://github.com/startnew/flexdepth

We tackle controllable human grasp synthesis under sparse user-specified thumb-contact constraints. While existing methods use contact as conditioning, they lack explicit semantic grounding, yielding weak contact-intent coupling, poor constraint sensitivity, and pose drift. We propose TaxoGrasp, a framework threading Feix grasp taxonomy as a persistent intent bottleneck throughout: (1) intent inference via supervised taxonomy prediction from object geometry and thumb-contact constraint, (2) diffusion synthesis conditioned on taxonomy templates, and (3) anchor-optimized grounding preserving taxonomy structure. We augment HOGraspNet with patch-level thumb-contact annotations and introduce a standardized evaluation protocol with three metric categories. Experiments demonstrate improved intent consistency, physical plausibility, and constraint adherence.