WACV 2026 论文列表

With the rapid growth of video content available, the ability to search for specific moments within videos using textual queries has become increasingly relevant. This is crucial in many scenarios, from surveillance cameras where it may be necessary to find specific events in extensive video streams to searching for exciting movie scenes. However, existing approaches for video Moment Retrieval and Highlight Detection often struggle to effectively align text and video features, limiting their performance. We argue that utilizing recent foundational video models designed for video-text alignment can overcome these limitations. We propose a novel architecture that utilizes such models to test this hypothesis. Combined with our novel Saliency-Guided Cross Attention mechanism and a hybrid DETR architecture, our approach provides significantly improved results. To further enhance our approach, we developed InterVid-MR - a large-scale, high-quality dataset specifically designed for a pretraining stage. Extensive experiments and comparisons with current state-of-the-art methods confirm the effectiveness of the approach, achieving 58.8 mAP on QVHighlights, 60.7 R@1 mIoU on Charades-STA, and 42.4 R@1 mIoU on TACoS. These results highlight the efficiency and scalability of the method for video-language tasks in both zero-shot and fine-tuning scenarios.

Zero-shot 3D object classification is crucial for real-world applications like autonomous driving, however it is often hindered by a significant domain gap between the synthetic data used for training and the sparse, noisy LiDAR scans encountered in the real world. Current methods trained solely on synthetic data fail to generalize to outdoor scenes, while those trained only on real data lack the semantic diversity to recognize rare or unseen objects. We introduce BlendCLIP, a multimodal pretraining framework that bridges this synthetic-to-real gap by strategically combining the strengths of both domains. We first propose a pipeline to generate a large-scale dataset of object-level triplets--consisting of a point cloud, image, and text description--mined directly from real-world driving data and human annotated 3D boxes. Our core contribution is a curriculum-based data mixing strategy that first grounds the model in the semantically rich synthetic CAD data before progressively adapting it to the specific characteristics of real-world scans. Our experiments show that our approach is highly label-efficient: introducing as few as 1.5% real-world samples per batch into training boosts zero-shot accuracy on the nuScenes benchmark by 27%. Consequently, our final model achieves state-of-the-art performance on challenging outdoor datasets like nuScenes and TruckScenes, improving over the best prior method by 19.3% on nuScenes, while maintaining strong generalization on diverse synthetic benchmarks. Our findings demonstrate that effective domain adaptation, not full-scale real-world annotation, is the key to unlocking robust open-vocabulary 3D perception. Our code and dataset will be released upon acceptance on https://github.com/kesu1/BlendCLIP

Test-Time Adaptation (TTA) enables deep neural networks to handle distribution shifts without requiring labels at inference. However, existing methods commonly assume complete class overlap between source and target domains, which rarely holds in practice. We study the challenging setting of Partial Class Imbalance, where the target domain contains only a subset of source classes. We show that entropy minimization-based TTA methods degrade over long test sequences because batch normalization updates bias feature representations toward visible classes, resulting in skewed predictions. To address this, we propose Logit-Adjusted Entropy Minimization, a simple yet effective strategy that integrates target class priors into the adaptation objective. Our method is model-agnostic and can be seamlessly applied to a wide range of TTA algorithms. Extensive experiments on CIFAR-100-C, ImageNet-C under diverse corruptions and severity levels, and the large-scale DomainNet-126 dataset demonstrate that our method consistently improves adaptation stability and accuracy for both CNNs and Vision Transformers. Compared to strong baselines, our approach reduces overfitting to visible classes and mitigates performance degradation in long-sequence adaptation. Code is available at https://github.com/thilinawee/latta

Three-dimensional geospatial analysis is critical to applications in urban planning, climate adaptation, and environmental assessment. Current methodologies depend on costly, specialized sensors (e.g., LiDAR and multispectral), which restrict global accessibility. Existing sensor-based and rule-driven methods further struggle with tasks requiring the integration of multiple 3D cues, handling diverse queries, and providing interpretable reasoning. We hereby present Geo3DVQA, a comprehensive benchmark for evaluating vision-language models (VLMs) in height-aware, 3D geospatial reasoning using RGB-only remote sensing imagery. Unlike conventional sensor-based frameworks, Geo3DVQA emphasizes realistic scenarios that integrate elevation, sky view factors, and land cover patterns. The benchmark encompasses 110k curated question-answer pairs spanning 16 task categories across three complexity levels: single-feature inference, multi-feature reasoning, and application-level spatial analysis. The evaluation of ten state-of-the-art VLMs highlights the difficulty of RGB-to-3D reasoning. GPT-4o and Gemini-2.5-Flash achieved only 28.6% and 33.0% accuracy respectively, while domain-specific fine-tuning of Qwen2.5-VL-7B achieved 49.6% (+24.8 points). These results reveal both the limitations of current VLMs and the effectiveness of domain adaptation. Geo3DVQA introduces a new frontier of challenges for scalable, accessible, and holistic 3D geospatial analysis. The dataset and code will be released upon publication at https://github.com/mm1129/Geo3DVQA.

Collisions on inland waters frequently occur due to the absence of clearly defined navigation guidance. To prevent such accidents, optical sensors such as stereo cameras can be employed to detect obstacles at an early stage, enabling timely warnings to vessel operators or the planning of collision avoidance trajectories. In this context, the paper presents a neural network for object detection on inland waters, designed to address the specific challenges of waterborne vehicles, including strong ego-motion and contextual cues like the shoreline appearing in the background. The proposed network leverages a plane sweep approach to integrate multiple views of a scene and predict object locations in bird's-eye view (BEV) coordinates. Its ability to incorporate more than two camera perspectives is demonstrated using the KITTI dataset. On real-world inland water data, the network is evaluated against both a traditional maritime stereo-based method and a learning-based stereo detector from autonomous driving, showing consistently higher mean average precision. The code is available at https://github.com/dionysos4/IWOD.

Generic Event Boundary Detection (GEBD) aims to identify moments in videos that humans perceive as event boundaries. This paper proposes a novel method for addressing this task, called Structured Context Learning, which introduces the Structured Partition of Sequence (SPoS) to provide structured context for temporal information learning. Our approach is end-to-end trainable and flexible, not being restricted to specific temporal models like GRU, LSTM, and Transformers. This flexibility enables our method to achieve a better speed-accuracy trade-off. Specifically, we apply SPoS to partition the input frame sequence and provide a structured context for the subsequent temporal model. Notably, SPoS's overall computation complexity is linear with respect to the video length. We next calculate group similarities to capture differences between frames, and a lightweight fully convolutional network is utilized to determine the event boundaries based on the grouped similarity maps. To remedy the ambiguities of boundary annotations, we adapt the Gaussian kernel to preprocess the ground-truth event boundaries. Our proposed method is extensively evaluated on the challenging Kinetics-GEBD, TAPOS, and shot transition detection datasets, demonstrating its superiority over existing state-of-the-art methods.

Matching between image modalities is a high-impact research area. Current state-of-the-art (SOTA) methods rely on extensive multi-million-scale training protocols, which demand significant computational resources. However, the learned cross-modal mapping remains largely opaque and locked within the trained matcher, with limited options for downstream use or transfer to other matchers. To enable such capabilities, we propose Patch Your Matcher (PYM) (https://xaf-cv.github.io/pym/), a highly adaptive method for leveraging pre-trained single-modality matchers for cross-modal matching by co-learning an explicit two-view geometrically consistent mapping. PYM learns image-to-image (I2I) translations that map new modalities into the original matcher's modality using a novel adversarial learning approach based on explicit evaluation of 6 DoF two-view correspondence plausibility. Trained with the semi-dense ELoFTR [80], our approach delivers substantially better cross-modal matching than classic I2I techniques, and recovers 97.05% of the matching precision of the extensively trained SOTA multi-modal MINIMA [62] variant. PYM also significantly boosts cross-modal matching performance of uni-modal sparse LightGlue [50] and dense RoMA [23] matchers, demonstrating high transferability of the learned mapping.

Adversarial attacks significantly challenge the safe deployment of deep learning models, particularly in real-world applications. Traditional defenses often rely on computationally intensive optimization (e.g., adversarial training or data augmentation) to improve robustness, whereas the human visual system achieves inherent robustness to adversarial perturbations through evolved biological mechanisms. We hypothesize that attention guided non-homogeneous sparse sampling and predictive coding plays a key role in this robustness. To test this hypothesis, we propose a novel defense framework incorporating three key biological mechanisms: foveal-peripheral processing, saccadic eye movements, and cortical filling-in. Our approach employs reinforcement learning-guided saccades to selectively capture multiple foveal-peripheral glimpses, which are integrated into a reconstructed image before classification. This biologically inspired preprocessing effectively mitigates adversarial noise, preserves semantic integrity, and notably requires no retraining or fine-tuning of downstream classifiers, enabling seamless integration with existing systems. Experiments on the ImageNet dataset demonstrate that our method improves system robustness across diverse classifiers and attack types, while significantly reducing training overhead compared to both biologically and non-biologically inspired defense techniques.

We propose a novel framework, named SimForce, for simultaneously estimating skeletal pose, ground reaction force and surface electromyography from an input video. Simforce predicts the more biomechanically accurate 3D human pose and shape of a given subject, along with their proposed muscle activations and resultant ground reaction force which leads to their input motion. Previous research has either focused on estimating these attributes singly and not treated them as a related task by taking into account the inherent shared motion between the three. In contrast, SimForce is designed to take advantage of the shared biological structure of the human body and its intrinsic connections to infer these attributes jointly using past, current, and future frames. SimForce features a newly introduced temporal and attention aware GCN-based architecture. To learn the subtle links between the body parts and how it affects the distribution of weight on the muscles over time, we introduce the Spatially Aware Attention Module.

Current autonomous driving (AD) simulations are critically limited by their inadequate representation of realistic and diverse human behavior, which is essential for ensuring safety and reliability. Existing benchmarks often simplify pedestrian interactions, failing to capture complex, dynamic intentions and varied responses critical for robust system deployment. To overcome this, we introduce HABIT (Human Action Benchmark for Interactive Traffic), a high-fidelity simulation benchmark. HABIT integrates real-world human motion, sourced from mocap and videos, into CARLA (Car Learning to Act, a full autonomous driving simulator) via a modular, extensible, and physically consistent motion retargeting pipeline. From an initial pool of approximately 30,000 retargeted motions, we curate 4,730 traffic-compatible pedestrian motions, standardized in SMPL format for physically consistent trajectories. HABIT seamlessly integrates with CARLA's Leaderboard, enabling automated scenario generation and rigorous agent evaluation. Our safety metrics, including Abbreviated Injury Scale (AIS) and False Positive Braking Rate (FPBR), reveal critical failure modes in state-of-the-art AD agents missed by prior evaluations. Evaluating three state-of-the-art autonomous driving agents, InterFuser, TransFuser, and BEVDriver, demonstrates how HABIT exposes planner weaknesses that remain hidden in scripted simulations. Despite achieving close or equal to zero collisions per kilometer on the CARLA Leaderboard, the autonomous agents perform notably worse on HABIT, with up to 7.43 collisions/km and a 12.94% AIS 3+ injury risk, and they brake unnecessarily in up to 33% of cases. All components are publicly released to support reproducible, pedestrian-aware AI research.

Recent advances in generative AI have enabled the creation of highly realistic digital content, raising concerns around authenticity, ownership, and misuse. While watermarking has become an increasingly important mechanism to trace and protect digital media, most existing image watermarking schemes operate as black boxes, producing global detection scores without offering any insight into how or where the watermark is present. This lack of transparency impacts user trust and makes it difficult to interpret the impact of tampering. In this paper, we present a post-hoc image watermarking method that combines localised embedding with region-level interpretability. Our approach embeds watermark signals in the discrete wavelet transform domain using a statistical block-wise strategy. This allows us to generate detection maps that reveal which regions of an image are likely watermarked or altered. We show that our method achieves strong robustness against common image transformations while remaining sensitive to semantic manipulations. At the same time, the watermark remains highly imperceptible. Compared to prior post-hoc methods, our approach offers more interpretable detection while retaining competitive robustness. For example, our watermarks are robust to cropping up to half the image.

3D Gaussian Splatting (3DGS) has enabled photorealistic and real-time rendering of 3D head avatars. Existing 3DGS-based avatars typically rely on tens of thousands of 3D Gaussian points (Gaussians), with the number of Gaussians fixed after training. However, many practical applications require adjustable levels of detail (LOD) to balance rendering efficiency and visual quality. In this work, we propose "ArchitectHead", the first framework for creating 3D Gaussian head avatars that support continuous control over LOD. Our key idea is to parameterize the Gaussians in a 2D UV feature space and propose a UV feature field composed of multi-level learnable feature maps to encode their latent features. A lightweight neural network-based decoder then transforms these latent features into 3D Gaussian attributes for rendering. ArchitectHead controls the number of Gaussians by dynamically resampling feature maps from the UV feature field at the desired resolutions. This method enables efficient and continuous control of LOD without retraining. Experimental results show that ArchitectHead achieves state-of-the-art (SOTA) quality in self and cross-identity reenactment tasks at the highest LOD, while maintaining near SOTA performance at lower LODs. At the lowest LOD, our method uses only 6.2% of the Gaussians while the quality degrades moderately (L1 Loss +7.9%, PSNR -0.97%, SSIM -0.6%, LPIPS Loss +24.1%), and the rendering speed nearly doubles. Project homepage: https://peizhiyan.github.io/docs/architect/.

Personalized Federated Learning (PFL) aims to deliver effective client-specific models under heterogeneous distributions, yet existing methods suffer from shallow prototype alignment and brittle server-side distillation. We propose HEART-PFL, a dual-sided framework that (i) performs depth-aware Hierarchical Directional Alignment (HDA) using cosine similarity in the early stage and MSE matching in the deep stage to preserve client specificity, and (ii) stabilizes global updates through Adversarial Knowledge Transfer (AKT) with symmetric KL distillation on clean and adversarial proxy data. Using lightweight adapters with only 1.46M trainable parameters, HEART-PFL achieves state-of-the-art personalized accuracy on CIFAR-100, Flowers-102, and Caltech-101 (63.42%, 84.23%, and 95.67%, respectively) under Dirichlet non-IID partitions, and remains robust to out-of-domain proxy data. Ablation studies further confirm that HDA and AKT provide complementary gains in alignment, robustness, and optimization stability, offering insights into how the two components mutually reinforce effective personalization. Overall, these results demonstrate that HEART-PFL simultaneously enhances personalization and global stability, highlighting its potential as a strong and scalable solution for PFL (code available at https://github.com/danny0628/HEART-PFL).

Conformal prediction is a model-agnostic approach to generating prediction sets that cover the true class with a high probability. Although its prediction set size is expected to capture aleatoric uncertainty, there is a lack of evidence regarding its effectiveness. The literature presents that prediction set size can upper-bound aleatoric uncertainty or that prediction sets are larger for difficult instances and smaller for easy ones, but a validation of this attribute of conformal predictors is missing. This work investigates how effectively conformal predictors quantify aleatoric uncertainty, specifically the inherent ambiguity in datasets caused by overlapping classes. We perform this by measuring the correlation between prediction set sizes and the number of distinct labels assigned by human annotators per instance. We further assess the similarity between prediction sets and human-provided annotations. We use three conformal prediction approaches to generate prediction sets for eight deep learning models trained on four datasets. The datasets contain annotations from multiple human annotators (ranging from five to fifty participants) per instance, enabling the identification of class overlap. We show that the vast majority of the conformal prediction outputs show a very weak to weak correlation with human annotations, with only a few showing moderate correlation. These findings underscore the necessity of critically reassessing the prediction sets generated using conformal predictors. While they can provide a higher coverage of the true classes, their capability in capturing aleatoric uncertainty remains limited.

We study the possibility of imitating the functionality of an image-to-image translation model by observing input-output pairs. We focus on cases where training the model from scratch is impossible, either because training data are unavailable or because the model architecture is unknown. This is the case, for example, with commercial models for biological applications. Since the development of these models requires large investments, their owners commonly keep them confidential and reveal only a few input-output examples on the company's website or in an academic paper. Surprisingly, we find that even a single example typically suffices for learning to imitate the model's functionality, and that this can be achieved using a simple distillation approach. We present an extensive ablation study encompassing a wide variety of model architectures, datasets and tasks, to characterize the factors affecting vulnerability to functionality imitation, and provide a preliminary theoretical discussion on the reasons for this unwanted behavior.

One-shot segmentation(OSS) aims to segment the target object in a query image using only one set of support image and mask. Therefore, having strong prior information for the target object using the support set is essential to guide the initial training of OSS, which leads to the success of one-shot segmentation in challenging cases, such as when the target object shows considerable variation in appearance, texture, or scale across the support and query images. To enrich this prior knowledge, we introduce FCC(Fully Connected Correlation) which integrates pixel-level correlations between support and query features, capturing associations that reveal target-specific patterns and correspondences in both same-layers and cross-layers. FCC captures previously inaccessible target information, effectively addressing the limitations of support mask. Our approach consistently demonstrates state-of-the-art performance in the PASCAL, COCO, and domain shift tests, while also notably accelerating model convergence. We conducted an ablation study and cross-layer correlation analysis to validate FCC's core methodology. These findings reveal the effectiveness of FCC in enhancing prior information and overall model performance for OSS.

We reinterpret NeRFs as a resource for extreme data augmentation to advance the field of camera relocalization. Our approach lets us automatically render a massive, densified dataset of novel views, given only sparse ground-truth viewpoints. We introduce a filtering strategy that, compared to existing novel-view-synthesis-focused relocalizers, does not rely on custom or specific NeRF backbones. This filtering strategy allows for significant spatial extrapolation within the scene, without compromising novel view quality. As a result, training a lightweight off-the-shelf vision backbone as a pose regressor on our expanded datasets significantly improves accuracy, uniquely enables relocalization of very spatially-novel views, and performs well on portable-scale hardware.

Diffusion models have demonstrated remarkable capabilities in image inpainting tasks, yet they often struggle to maintain semantic consistency and fine-grained details when filling large masked regions. Existing approaches typically require extensive fine-tuning or are trained from scratch, still losing the context, patterns, or realism. We introduce FuLLaMa, a novel training-free framework for diffusion-based removal that preserves semantic embeddings as well as the image quality throughout the infill process. FuLLaMa enhances traditional removal algorithms with the information manifold of LVLMs and generation capability of DMs. Through Adaptive Parameter Manifold Navigation (APMN), DM is guided to generate content that harmonizes with the existing context and structure of the image, without introducing new elements. FuLLaMa achieves a high and balanced performance compared to 6 SOTA object removal algorithms, on 2 datasets, for various tasks such as large area, small object, multi-instance, and patterned removal; using 9 visual and 5 contextual evaluation metrics. We conduct several ablation studies for the system and objective design, also showcasing mask-based, language-based, and point-and-click removal applications. Our work establishes context and quality co-preservation as a fundamental principle for diffusion-based removal.

Generating realistic food images for categories with multiple nouns is surprisingly challenging. For instance, the prompt "egg noodle" may result in images that incorrectly contain both eggs and noodles as separate entities. Multi-noun food categories are common in real-world datasets and account for a large portion of entries in benchmarks such as UEC-256. These compound names often cause generative models to misinterpret the semantics, producing unintended ingredients or objects. This is due to insufficient multi-noun category related knowledge in the text encoder and misinterpretation of multi-noun relationships, leading to incorrect spatial layouts. To overcome these challenges, we propose FoCULR (Food Category Understanding and Layout Refinement) which incorporates food domain knowledge and introduces core concepts early in the generation process. Experimental results demonstrate that the integration of these techniques improves image generation performance in the food domain.