WACV 2026 论文列表

Despite the increasing popularity of avatar systems such as Snapchat Bitmojis, existing production avatar platforms face several limitations, such as a limited number of predefined assets, tedious customization processes, and inefficient rendering requirements. Addressing these shortcomings, we introduce Snapmoji, an avatar generation system that instantly creates 3D avatars, and enables customization in a process we call dual-stylization. Snapmoji first maps a selfie of a user to a primary avatar (e.g., Bitmoji style) using a new technique we name Gaussian Domain Adaptation (GDA), then applies a secondary style (e.g., skeleton, yarn, toy) to the primary avatar, all while preserving the user's identity. The generated 3D avatars can then be rendered an animated on mobile devices at 30--40 FPS.

Multilingual capability is a crucial requirement for large multimodal models, which are increasingly deployed across diverse countries and languages. However, most existing benchmarks for multilingual multimodal reasoning fail to effectively distinguish models of different strengths; in fact, even text-only language models without visual capabilities can often achieve high scores. As a result, the comprehensive evaluation of state-of-the-art multilingual multimodal models remains underexplored. In this work, we present M4U, a novel and challenging benchmark designed to evaluate multilingual, multi-discipline multimodal understanding and reasoning. M4U comprises 10k samples spanning 64 disciplines across 16 subfields in Science, Engineering, and Healthcare, covering six languages. Using this benchmark, we conduct extensive evaluations of leading Large Multimodal Models (LMMs) and Large Language Models (LLMs) augmented with external tools. Our results reveal that even the strongest LMMs exhibit pronounced language preferences and struggle with reasoning tasks that require integrating multilingual information across visual and textual modalities. In particular, performance drops markedly when models are prompted with cross-lingual multimodal questions, highlighting significant gaps in current multilingual multimodal reasoning capabilities.

The explosive growth of online video content has heightened the need for efficient text-to-video retrieval systems. These systems heavily rely on accurate video-text representations for effective retrieval. However, challenges such as the multi-thematic nature of videos and inconsistent caption quality hinder performance by making it difficult to capture comprehensive video-text relationships. To address these issues, we propose CoreCaption, a novel framework that extracts and leverages core captions--the most representative captions capturing essential video themes. Our approach includes a unique core caption extraction method based on similarity-based density estimation and clustering, and introduces the Core Caption Guided Attention (CCGA) mechanism to integrate video-specific semantic information into text queries while preserving their original intent. Furthermore, we employ a teacher-student architecture for efficient inference without reliance on core captions during deployment. Extensive experiments on benchmark datasets like MSR-VTT, VATEX, and MSVD demonstrate that CoreCaption outperforms state-of-the-art methods. These results validate the effectiveness and robustness of our framework across diverse video-text datasets.

In recent years, artificial intelligence (AI) has significantly impacted digital forensics, yet its broader deployment remains limited due to the difficulty of explaining AI decisions. Explainable AI (XAI) presents a promising solution to increase transparency and trust, but its application in digital forensics is still underexplored. In this work, we present a practical and structured explainable digital forensics AI (xDFAI) approach tailored to the forensic task of video source camera identification (VSCI). Our method enables forensic examiners to interpret the behavior of AI models, assess whether decisions are driven by intended logic or arise from random or content-dependent artifacts, and establish the integrity and reliability of explanations. We implement and evaluate this approach on two state-of-the-art VSCI models, providing step-by-step analysis of explanation quality, spatial consistency of high-impact features, and content dependence. Our results reveal that although models achieve strong classification accuracy, their explanations lack spatial stability and are impacted by video content, raising concerns about forensic reliability. To support reproducibility and future research, we provide an open-source implementation. This work underscores the potential of XAI to improve transparency in digital forensics and highlights the challenges of interpreting and presenting results. Our study takes an important step toward the operational deployment of xDFAI in multimedia forensics.

The advancement of general large multi-modal models (LMMs) has transformed many computer vision tasks, shifting image quality assessment (IQA) from specialized algorithms to models built on pre-trained LMM backbones. This evolution raises the question of whether dedicated IQA metrics remain relevant or are becoming obsolete in the age of LMMs. In this paper, we address this challenge by introducing ScoreNet, a novel framework that fuses the strengths of traditional metrics to elevate the IQA capabilities of LMMs. ScoreNet employs a soft prompting mechanism, learning prompts from a curated set of lightweight IQA scores and image embeddings. This context-driven learning strategy enhances the adaptability of LMMs for IQA tasks with a small additional computation cost. We show that ScoreNet serves as a general-purpose extension applicable to modern LMM-based IQA models. We integrate ScoreNet into two high-performing methods--CLIP-IQA and Q-Align--and observe consistent improvements. Experimental results show that ScoreNet not only boosts both models but also surpasses other state-of-the-art IQA approaches. The code is available at https://github.com/AISmartPerception/ScoreNet

Text-to-image diffusion models can generate stunning visuals, yet they often fail at tasks children find trivial--like placing a dog to the right of a teddy bear rather than to the left. When combinations get more unusual--a giraffe above an airplane--these failures become even more pronounced. Existing methods attempt to fix these spatial reasoning failures through model fine-tuning or test-time optimization with handcrafted losses that are suboptimal. Rather than imposing our assumptions about spatial encoding, we propose learning these objectives directly from the model's internal representations. We introduce Learn-to-Steer, a novel framework that learns data-driven objectives for test-time optimization rather than handcrafting them. Our key insight is to train a lightweight classifier that decodes spatial relationships from the diffusion model's cross-attention maps, then deploy this classifier as a learned loss function during inference. Training such classifiers poses a surprising challenge: they can take shortcuts by detecting linguistic traces in the cross-attention maps, rather than learning true spatial patterns. We solve this by augmenting our training data with samples generated using prompts with incorrect relation words, which encourages the classifier to avoid linguistic shortcuts and learn spatial patterns from the attention maps. Our method dramatically improves spatial accuracy: from 20% to 61% on FLUX.1-dev and from 7% to 54% on SD2.1 across standard benchmarks. It also generalizes to multiple relations with significantly improved accuracy.

The role of temporal information in predicting human gaze in dynamic scenes remains a critical open question, underscored by the paradoxical finding that strong static models can outperform complex video-based models. This suggests that the true contribution of temporal cues has been obscured by confounding architectural variables. To resolve this, we present a rigorous, controlled experiment centered on a minimal architectural pair: a spatio-temporal saliency model (UniformerSal-ST) and its identical spatial-only counterpart (UniformerSal-S), designed to unambiguously isolate the impact of temporal feature integration. Our results demonstrate that principled temporal fusion yields a substantial Information Gain (IG) of +0.20 bits on temporally coherent datasets like LEDOV. Crucially, our controlled comparison also uncovers a key failure mode: on datasets with frequent hard cuts like DIEM, the same mechanism degrades performance, incurring a 0.07 bits IG deficit. We provide a mechanistic explanation for this dichotomy, revealing how certain visual scenarios (scene discontinuities, rapid camera zooms) can disrupt current temporal fusion approaches. By precisely quantifying both the benefits and drawbacks of temporal processing, our work provides the community with clear, actionable insights into when and why temporal information should be modeled for more robust and accurate video saliency prediction.

Change detection in remote sensing imagery is essential for applications such as urban planning, environmental monitoring, and disaster management. Traditional change detection methods typically identify all changes between two temporal images without distinguishing the types of transitions, which can lead to results that may not align with specific user needs. Although semantic change detection methods have attempted to address this by categorizing changes into predefined classes, these methods rely on rigid class definitions and fixed model architectures, making it difficult to mix datasets with different label sets or reuse models across tasks, as the output channels are tightly coupled with the number and type of semantic classes. To overcome these limitations, we introduce Referring Change Detection (RCD), which leverages natural language prompts to detect specific classes of changes in remote sensing images. By integrating language understanding with visual analysis, our approach allows users to specify the exact type of change they are interested in. However, training models for RCD is challenging due to the limited availability of annotated data and severe class imbalance in existing datasets. To address this, we propose a two-stage framework consisting of (I) RCDNet, a cross-modal fusion network designed for referring change detection, and (II) RCDGen, a diffusion-based synthetic data generation pipeline that produces realistic post-change images and change maps for a specified category using only pre-change image, without relying on semantic segmentation masks and thereby significantly lowering the barrier to scalable data creation. Experiments across multiple datasets show that our framework enables scalable and targeted change detection. Code will be made publicly available on \href https://github.com/yilmazkorkmaz1/referring_change_detection Github .

The six-degree-of-freedom (6-DoF) pose and metric size estimation of multiple objects from RGB images alone remains a challenging task, particularly due to significant variations in object shape, appearance, and frequent occlusions in complex scenes. To address these challenges, we introduce KMOPS, a Keypoint-driven method tailored specifically for Multi-Object Pose and metric Size estimation from a single calibrated stereo image pair. Leveraging the stereo input, our approach first extracts the 2D keypoints of the enclosing bounding boxes of the objects across both views, and subsequently triangulates them to acquire metric 3D positions. Then, we obtain each object's rotation, translation, and dimensions by aligning the triangulated 3D keypoints to the canonical ones using a closed-form solution. Our formulation eliminates the need for predefined 3D search spaces or volumetric anchors, which are often required by other methods to constrain the vast 3D solution space. With extensive experiments on the challenging dataset Transparent Object Dataset (TOD) and StereOBJ-1M, we show that our method outperforms all competing methods with a simple and effective architecture.

Synthetic data has played a pivotal role in developing large-scale 3D vision models due to its high-quality annotations and ease of curation. In domains where labeled data collection is difficult, such as endoscopy, synthetic data holds promise as a means to generate the large-scale annotated datasets required to train modern neural networks. In this work, we address a core question for data-scarce applications in 3D vision: how can we generate synthetic labeled data, and how useful would the data be for training downstream vision models? First, we introduce a novel data generation module that takes images with known geometry and camera poses as input and estimates the material and lighting conditions of the scene. To stabilize training, we leverage domain-specific properties like non-stationary lighting and anatomical material priors. We model the material properties as a bidirectional reflectance distribution function, parameterized by a neural network. Via the rendering equation, we can generate photorealistic images at arbitrary camera poses. We demonstrate that this method produces competitive novel view synthesis results compared to previous work while being more lightweight, flexible, and efficient. Second, we use our synthetic data to train models on various downstream 3D vision tasks and find that models trained solely on our synthetic data generally outperform those trained on real data across various metrics and tasks. Our experiments show that synthetic data is a promising avenue towards robust 3D vision in surgical scenes.

Current network training paradigms primarily focus on either centralized or decentralized data regimes. However, in practice, data availability often exhibits a hybrid nature, where both regimes coexist. This hybrid setting presents new opportunities for model training, as the two regimes offer complementary trade-offs: decentralized data is abundant but subject to heterogeneity and communication constraints, while centralized data--though limited in volume and potentially unrepresentative--enables better curation and high-throughput access. Despite its potential, effectively combining these paradigms remains challenging, and few frameworks are tailored to hybrid data regimes. To address this, we propose a novel framework that constructs a model atlas from decentralized models and leverages centralized data to refine a global model within this structured space. The refined model is then used to reinitialize the decentralized models. Our method synergizes federated learning (to exploit decentralized data) and model merging (to utilize centralized data), enabling effective training under hybrid data availability. Theoretically, we show that our approach achieves faster convergence than methods relying solely on decentralized data, due to variance reduction in the merging process. Extensive experiments demonstrate that our framework consistently outperforms purely centralized, purely decentralized, and existing hybrid-adaptable methods. Notably, our method remains robust even when the centralized and decentralized data domains differ or when decentralized data contains noise, significantly broadening its applicability.

Existing Open-Set Recognition (OSR) methods struggle in remote sensing (RS) as their reliance on unimodal visual features fails to resolve the severe inter-class similarity inherent in overhead imagery. To address this, we propose ViSTA-RS, a novel multimodal framework that leverages semantic context from language to disambiguate visually similar scenes. Our approach first constructs semantically-rich class prototypes by jointly encoding images with generated text captions using a Vision-Language Model. We then introduce a reconstruction-based mechanism where an image's visual embedding is expressed as a weighted combination of these semantic prototypes. The magnitude of the reconstruction error serves as a robust novelty score, with a statistically principled threshold determined by Extreme Value Theory (EVT). This alignment of multimodal semantics with prototype reconstruction is uniquely suited for the fine-grained nature of RS data. On four challenging benchmarks, ViSTA-RS sets a new state-of-the-art, improving the AUROC for unknown detection by a significant 6.7% over leading baselines while maintaining high accuracy on known classes.

When we train models on biased datasets, they not only reproduce data biases, but can worsen them at test time --- a phenomenon called bias amplification. Many of the current bias amplification metrics (e.g., BA, DPA) measure bias amplification only in classification datasets. These metrics are ineffective for image captioning datasets, as they cannot capture the language semantics of a caption. Recent work introduced Leakage in Captioning (LIC), a language-aware bias amplification metric that understands caption semantics. However, LIC has a crucial limitation: it cannot identify the source of bias amplification in captioning models. We propose Directional Bias Amplification in Captioning (DBAC), a language-aware and directional metric that can identify when captioning models amplify biases. DBAC has two more improvements over LIC: (1) it is less sensitive to sentence encoders (a hyperparameter in language-aware metrics), and (2) it provides a more accurate estimate of bias amplification in captions. Our experiments on gender and race attributes in the COCO captions dataset show that DBAC is the only reliable metric to measure bias amplification in captions.

This paper addresses a key challenge in Handwritten Text Recognition (HTR): the dependence on large volumes of labeled data. To overcome this, we propose a self-supervised learning (SSL) framework, LoGo-HTR, that minimizes labeling requirements while achieving strong recognition performance. We introduce a large-scale dataset, SSL-HWD of 10 million word-level handwritten images from diverse scanned documents, partitioned into a small labeled subset and a much larger unlabeled subset. LoGo-HTR combines a local contrastive loss for spatial consistency and a global decorrelation loss to enhance feature diversity. This dual objective enables robust, invariant, and spatially discriminative feature learning. After self-supervised pretraining, we fine-tune a transformer-based decoder using limited labeled data. Extensive experiments on standard HTR benchmarks, which include multilingual and historical data, demonstrate that, after SSL pretraining on our unlabeled dataset, our method consistently outperforms state-of-the-art approaches, even when fine-tuned using only 80% and 20% of the available labeled training data from the respective benchmarks. Ablation studies highlight the effectiveness of our dual loss design and demonstrate the potential of scalable, label-efficient handwritten text recognition. The SSL-HWD dataset and LoGo-HTR model with code are publicly available at https://logo-ssl.github.io/.

Action Quality Assessment (AQA) predicts fine-grained execution scores from action videos and is widely applied in sports, rehabilitation, and skill evaluation. Long-term AQA, as in figure skating or rhythmic gymnastics, is especially challenging since it requires modeling extended temporal dynamics while remaining robust to contextual confounders. Existing approaches either depend on costly annotations or rely on unidirectional temporal modeling, making them vulnerable to spurious correlations and unstable long-term representations. To this end, we propose CaFlow, a unified framework that integrates counterfactual de-confounding with bidirectional time-conditioned flow. To address the issue of spurious correlations, the Causal Counterfactual Regularization (CCR) module disentangles causal and confounding features in a self-supervised manner and enforces causal robustness through counterfactual interventions. To tackle unstable temporal refinement in long sequences, the BiT-Flow module explicitly models forward and backward dynamics with a cycle-consistency constraint, producing smoother and more coherent representations. Extensive experiments on multiple long-term AQA benchmarks demonstrate that CaFlow achieves state-of-the-art performance, validating its effectiveness.

Accurate annotations are critical to providing ground truth to supervised learning and to evaluating the performance of autonomous vehicle systems. These vehicles are typically equipped with active sensors, such as LiDAR, which scan the environment in predefined patterns. 3D box annotation based on data from such sensors is challenging in dynamic scenarios, where objects are observed at different timestamps, hence different positions. Without proper handling of this phenomenon, systematic errors are prone to being introduced in the annotations.Our work is the first one to describe why such annotation errors occur and illustrates them using examples from widely used, publicly available datasets. Through our novel estimation method, we correct the annotations so that they follow physically feasible trajectories and achieve spatial and temporal consistency with the sensor data. For the first time, we define metrics for this problem; and we evaluate our method on the Argoverse 2 and MAN TruckScenes datasets, as well as in our proprietary dataset. Our approach demonstrates robust performance in increasing the quality of ground truth by more than 17% in these datasets. Finally, we quantify the annotation errors in them and find that the original annotations are misplaced by up to 2.5 m, with highly dynamic objects being the most affected. Our code is available at https://github.com/alexandre-justo-miro/annotation-correction-3D-boxes.

Vision-Language Models (VLMs) effectively integrate visual and textual information, often relying on shared embedding spaces to align modalities. However, the extent to which these spaces capture complex, subjective human judgments, such as perceived facial trustworthiness and attractiveness, and whether they replicate associated human social biases, remains underexplored. This paper investigates the representation of subjective face attributes within Multimodal LLM and VLM embedding spaces, examining whether these representations encode human-like biases and assessing their interpretability. Using probing techniques on face datasets annotated with human judgments, we analyze the structure of VLM embeddings (e.g., from CLIP-like models). Our findings demonstrate that similarity scores between face image and textual description in the VLM embedding space align with human ratings of subjective attributes like trustworthiness and attractiveness, and crucially, these representations exhibit correlations and demographic disparities mirroring known biases in human social perception. Furthermore, we show that the use of variable context via face and attribute-specific captions can significantly improve the alignment of the VLM embedding space with human impressions. Interpreting the embedded social biases highlights the need for critical evaluation and bias-aware development of VLMs to mitigate the risk of perpetuating harmful stereotypes in downstream applications that involve Human-AI interaction.

Recent generative models can synthesize high-quality images but often fail to generate humans interacting with objects using their hands. This arises mostly from the model's misunderstanding of such interactions, and the hardships of synthesizing intricate regions of the body. In this paper, we propose GraspDiffusion, a novel generative method that creates realistic scenes of human-object interaction. Given a 3D object mesh, GraspDiffusion first constructs life-like whole-body poses with control over the object's location relative to the human body. This is achieved by separately leveraging the generative priors for 3D body and hand poses, optimizing them into a joint grasping pose. The resulting pose guides the image synthesis to correctly reflect the intended interaction, allowing the creation of realistic and diverse human-object interaction scenes. We demonstrate that GraspDiffusion can successfully tackle the relatively uninvestigated problem of generating full-bodied human-object interactions while outperforming previous methods.

Human Action Recognition (HAR) in real-world scenarios is significantly challenged by unseen domain shifts, such as variations in the camera viewpoint, illumination, lighting, or background. Although recent advancements in video domain generalization are promising for HAR by introducing models that are robust to these shifts, existing methods often fall short. They typically depend on a single modality or employ static frame-level fusion approaches, which inherently limit the capture of multi-scale temporal dependencies and the alignment of the asynchronous modalities frequently present in video data. To address these limitations, we propose Multimodal Alignment and Distillation for Domain Generalization (MAD-DG), a novel framework that synchronizes asynchronous modalities through a segment-label aligned temporal binding window with a contrastive alignment mechanism. We further incorporate an online self-distillation temporal module that captures multi-scale temporal relationships and learns robust, domain-invariant representations. Extensive experiments demonstrate that MAD-DG achieves state-of-the-art performance and exhibits better generalization capabilities across both single- and multi-source domain generalization settings. Our source code is available at https://github.com/dxlabskku/MAD-DG.