WACV 2026 论文列表

Object detection has advanced significantly with Detection Transformers (DETRs). However, these models are computationally demanding, posing challenges for deployment in resource-constrained environments (e.g., self-driving cars). Knowledge distillation (KD) is an effective compression method widely applied to CNN detectors, but its application to DETR models has been limited. Most KD methods for DETRs fail to distill transformer-specific global context. Also, they blindly believe in the teacher model, which can sometimes be misleading. To bridge the gaps, this paper proposes Consistent Location-and-Context-aware Knowledge Distillation (CLoCKDistill) for DETR detectors, which includes both feature distillation and logit distillation components. For feature distillation, instead of distilling backbone features like many existing KD methods, we distill the transformer encoder output (i.e., memory) that contains valuable global context and long-range dependencies. Also, we enrich this memory with object location details during feature distillation so that the student model can prioritize relevant regions. To facilitate logit distillation, we create target-aware queries based on the ground truth, allowing both the student and teacher decoders to attend to consistent and accurate parts of encoder memory. Experiments on KITTI and COCO show our CLoCKDistill method's efficacy across various DETRs. Our method boosts student detector performance by 2.2% to 6.4%. Our code is available at: https://github.com/lanqz7766/CLoCKDistill.

Recognizability, a key perceptual factor in human face processing, strongly affects the performance of face recognition (FR) systems in both verification and identification tasks. Effectively using recognizability to enhance feature representation remains challenging. In deep FR, the loss function plays a crucial role in shaping how features are embedded. However, current methods have two main drawbacks: (i) recognizability is only partially captured through soft margin constraints, resulting in weaker quality representation and lower discrimination, especially for low-quality or ambiguous faces; (ii) mutual overlapping gradients between feature direction and magnitude introduce undesirable interactions during optimization, causing instability and confusion in hypersphere planning, which may lead to poor generalization, and entangled representations where recognizability and identity are not cleanly separated. To address these issues, we introduce a hard margin strategy - Quality Control Face (QCFace), which overcomes the mutual overlapping gradient problem and enables the clear decoupling of recognizability from identity representation. Based on this strategy, a novel hard-margin-based loss function employs a guidance factor for hypersphere planning, simultaneously optimizing for recognition ability and explicit recognizability representation. Extensive experiments confirm that QCFace not only provides robust and quantifiable recognizability encoding but also achieves state-of-the-art performance in both verification and identification benchmarks compared to existing recognizability-based losses.

Monitoring critically endangered Western Lowland Gorillas is currently hampered by the immense manual effort required to re-identify individuals from vast archives of camera trap footage. The primary obstacle to automating this process has been the lack of large-scale, "in-the-wild" video datasets suitable for training robust deep learning models. To address this gap, we introduce a comprehensive benchmark with three novel datasets: Gorilla-SPAC-Wild, the largest video dataset for wild primate re-identification to date; Gorilla-Berlin-Zoo, for assessing cross-domain re-identification generalization; and Gorilla-SPAC-MoT, for evaluating multi-object tracking in camera trap footage. Building on these datasets, we present GorillaWatch, an end-to-end pipeline integrating detection, tracking, and re-identification. To exploit temporal information, we introduce a multi-frame self-supervised pretraining strategy that leverages consistency in tracklets to learn domain-specific features without manual labels. To ensure scientific validity, a differentiable adaptation of AttnLRP verifies that our model relies on discriminative biometric traits rather than background correlations. Extensive benchmarking subsequently demonstrates that aggregating features from large-scale image backbones outperforms specialized video architectures. Finally, we address unsupervised population counting by integrating spatiotemporal constraints into standard clustering to mitigate over-segmentation. We publicly release all code and datasets to facilitate scalable, non-invasive monitoring of endangered species.

Recent advances in customizing Text-to-Image models allow users to generate personalized images with just a few samples. As demand for multi-concept generation grows, methods using weight fusion and test-time optimization have emerged, integrating multiple concepts within a single image. However, these approaches inject concept knowledge into the parametric space, leading to high overhead in multi-concept generation. We introduce DreamCatcher, an efficient framework based on representation finetuning. Our key innovation embeds conceptual information into the feature space, achieving up to 5x faster multi-concept generation while reducing learnable storage per concept by 88%, all without quality loss. Besides, our method is highly versatile, enabling personalized inpainting without additional training.

Annotating 3D data remains a costly bottleneck for 3D object detection, motivating the development of weakly supervised annotation methods that rely on more accessible 2D box annotations. However, relying solely on 2D boxes introduces projection ambiguities since a single 2D box can correspond to multiple valid 3D poses. Furthermore, partial object visibility under a single viewpoint setting makes accurate 3D box estimation difficult. We propose MVAT, a novel framework that leverages temporal multi-view present in sequential data to address these challenges. Our approach aggregates object-centric point clouds across time to build 3D object representations as dense and complete as possible. A Teacher-Student distillation paradigm is employed: The Teacher network learns from single viewpoints but targets are derived from temporally aggregated static objects. Then the Teacher generates high quality pseudo-labels that the Student learns to predict from a single viewpoint for both static and moving objects. The whole framework incorporates a multi-view 2D projection loss to enforce consistency between predicted 3D boxes and all available 2D annotations. Experiments on the nuScenes and Waymo Open datasets demonstrate that MVAT achieves state-of-the-art performance for weakly supervised 3D object detection, significantly narrowing the gap with fully supervised methods without requiring any 3D box annotations.

Wildfires are an escalating global concern due to the devastating impacts on the environment, economy, and human health, with notable incidents such as the 2019-2020 Australian bushfires and the 2025 California wildfires underscoring the severity of these events. AI-enabled camera-based smoke detection has emerged as a promising approach for the rapid detection of wildfires. However, existing wildfire smoke segmentation datasets that are used for training detection and segmentation models are limited in scale, geographically constrained, and often rely on synthetic imagery, which hinders effective training and generalization. To overcome these limitations, we present AusSmoke, a new smoke segmentation dataset collected from Australia to address the data scarcity in this region. Furthermore, we introduce a MultiNational geographically diverse and substantially larger fully-labelled benchmark, called MultiNatSmoke, that consolidates publicly available international datasets with the newly collected Australian imagery, expanding the scale by an order of magnitude over previous collections. Finally, we benchmark smoke segmentation models, demonstrating improved performance and enhanced generalization across diverse geographical contexts.

OpenLVLM-MIA is a new benchmark that highlights fundamental challenges in evaluating membership inference attacks (MIA) against large vision-language models (LVLMs). While prior work has reported high attack success rates, our analysis suggests that these results often arise from detecting distributional bias introduced during dataset construction rather than from identifying true membership status. To address this issue, we introduce a controlled benchmark of 6,000 images where the distributions of member and non-member samples are carefully balanced, and ground-truth membership labels are provided across three distinct training stages. Experiments using OpenLVLM-MIA demonstrated that the performance of state-of-the-art MIA methods approached chance-level. OpenLVLM-MIA, designed to be transparent and unbiased benchmark, clarifies certain limitations of MIA research on LVLMs and provides a solid foundation for developing stronger privacy-preserving techniques.

This paper introduces CommonForms, a web-scale dataset for form field detection. It casts the problem of form field detection as object detection: given an image of a page, predict the location and type (Text Input, Choice Button, Signature) of form fields. The dataset is constructed by filtering Common Crawl to find PDFs that have fillable elements. Starting with 8 million documents, the filtering process is used to arrive at a final dataset of roughly 55k documents that have over 450k pages. Analysis shows that the dataset contains a diverse mixture of languages and domains; one third of the pages are non-English, and among the 14 classified domains, no domain makes up more than 25% of the dataset.In addition, this paper presents a family of form field detectors, FFDNet-Small and FFDNet-Large, which attain a very high average precision on the CommonForms test set. Each model cost less than 500 to train. Ablation results show that high-resolution inputs are crucial for high-quality form field detection, and that the cleaning process improves data efficiency over using all PDFs that have fillable fields in Common Crawl. A qualitative analysis shows that they outperform a popular, commercially available PDF reader that can prepare forms. Unlike the most popular commercially available solutions, FFDNet can predict checkboxes in addition to text and signature fields. This is, to our knowledge, the first large scale dataset released for form field detection, as well as the first open source models. The dataset, models, and code will be released after review.

Multimodal Large Language Models (MLLMs) combine visual and textual representations to enable rich reasoning capabilities. However, the high computational cost of processing dense visual tokens remains a major bottleneck. A critical component in this pipeline is the visual projector, which bridges the vision encoder and the language model. Standard designs often employ a simple multi-layer perceptron for direct token mapping, but this approach scales poorly with high-resolution inputs, introducing significant redundancy. We present Delta-LLaVA, a token-efficient projector that employs a low-rank DeltaProjection to align multi-level vision features into a compact subspace before further interaction. On top of this base alignment, lightweight Transformer blocks act as specialization layers, capturing both global and local structure under constrained token budgets. Extensive experiments and ablations demonstrate that this base-then-specialize design yields consistent gains across multiple benchmarks with only 144 tokens, highlighting the importance of token formation prior to scaling interaction capacity. With Delta-LLaVA, inference throughput improves by up to 55%, while end-to-end training accelerates by nearly 4-5x in pretraining and over 1.5x in finetuning, highlighting the dual benefits of our design in both efficiency and scalability.

Existing for audio- and pose-driven human animation methods often struggle with stiff head movements and blurry hands, primarily due to the weak correlation between audio and head movements and the structural complexity of hands.To address these issues, we propose VividAnimator, an end-to-end framework for generating high-quality, half-body human animations driven by audio and sparse hand pose conditions. Our framework introduces three key innovations. First, to overcome the instability and high cost of online codebook training, we pre-train a Hand Clarity Codebook (HCC) that encodes rich, high-fidelity hand texture priors, significantly mitigating hand degradation. Second, we design a Dual-Stream Audio-Aware Module (DSAA) to model lip synchronization and natural head pose dynamics separately while enabling interaction. Third, we introduce a Pose Calibration Trick (PCT) that refines and aligns pose conditions by relaxing rigid constraints, ensuring smooth and natural gesture transitions. Extensive experiments demonstrate that Vivid Animator achieves state-of-the-art performance, producing videos with superior hand detail, gesture realism, and identity consistency, validated by both quantitative metrics and qualitative evaluations.

Remote photoplethysmography (rPPG) is a crucial technique for non-contact heart rate (HR) estimation using facial videos, gaining significance in driver monitoring systems where contact-based measurements are impractical. Existing rPPG methods often rely on either RGB or NIR data, each susceptible to limitations under motion artifacts and varying illumination in real-world driving scenarios. To address these challenges, we introduce a novel RGB-NIR fusion model tailored for robust rPPG and HR estimation in dynamic vehicle environments. Our approach features two main contributions, an NIR-specific decoder that facilitates effective cross-modal knowledge transfer from RGB to NIR, enhancing model adaptability, and a dual autoencoder architecture for efficient feature disentanglement and reconstruction, mitigating noise from driver motion and changing lighting conditions. Comprehensive evaluations, including inter- and cross-dataset testing and ablation studies across various driving and garage conditions, demonstrate that our model achieves superior performance on the MR-NIRP car dataset, showcasing significant robustness in complex vehicular environments.

Occlusion remains a significant challenge for existing human pose estimation algorithms, often resulting in inaccurate and anatomically implausible predictions. Although recent occlusion-robust methods report strong performance, they typically rely heavily on supervised learning and privileged information, such as multiview data or temporal sequences. Furthermore, these models often fail under domain changes. Domain-adaptive human pose estimation seeks to mitigate this issue; however, when occlusions are present in the target domain, a common occurrence in real-world applications, performance of these algorithms deteriorates significantly. To address these challenges, we propose VisOR, a novel Visibility guided Self-Supervised algorithm for Occlusion-Resilient Human Pose Estimation. VisOR achieves robustness to both domain shifts and occlusions by integrating contextual reasoning with iterative pseudo-label refinement. It mitigates the overfitting to noisy labels from occluded regions via a visibility-driven curriculum learning strategy, which progressively introduces the model to increasingly occluded training samples. Additionally, VisOR is regularized by a learned human pose prior that maintains anatomical plausibility throughout the adaptation process. Recognizing the scarcity of human pose datasets with realistic occlusions, we introduce BOW: Blended Occlusions in-the-Wild, a rigorously constructed context-aware synthetic benchmark designed to evaluate the occlusion resilience of human pose estimation algorithms. BOW offers a diverse range of context-aware occlusions across both indoor and outdoor environments, simulating real-world conditions. Through extensive experiments, we demonstrate that VisOR outperforms current state-of-the-art methods by 7% in challenging occluded human pose estimation benchmarks and provides a baseline performance on BOW, against existing algorithms.

Backdoor attacks are among the most prominent security threats to deep learning models. Traditional backdoors rely on fixed trigger patterns (e.g., a red square) that existing defenses can often effectively remove. However, recent attacks embed semantic triggers that vary with the input and blend with meaningful features, rendering prior defenses ineffective. We propose MARTINI, a novel backdoor mitigation framework that addresses both traditional and semantic backdoors. MARTINI reconstructs backdoor samples via a dedicated trigger reconstruction procedure, producing malicious inputs that replicate the injected attack effect across a spectrum of attacks. Using these reconstructed samples paired with their correct labels, MARTINI then hardens the model through retraining to neutralize the targeted misclassification. Our evaluation on 14 types of backdoor attacks in image classification shows that MARTINI can reduce the attack success rate (ASR) from 96.56% to 5.17% on average, outperforming 12 state-of-the-art backdoor removal approaches, which at best reduce the ASR to 26.56%. It can also mitigate backdoors in self-supervised learning, object detection and NLP sentiment analysis.

This work addresses action segmentation in videos under sparse timestamp supervision, where only a single frame per action segment---referred to as a timestamp---is labeled during training. We propose the Timestamp Query Transformer (TQT) that treats timestamps as learnable class query tokens. While existing approaches rely on iterative, multi-step generation of framewise pseudo-labels, TQT directly predicts temporal segmentation masks by leveraging query-feature cross-attention. This design enables fully end-to-end learning and maximizes the utility of sparse labels from the entire training dataset, rather than relying on only a few local timestamps within each training video as in prior work. Experiments on the GTEA, 50Salads, and Breakfast datasets demonstrate that TQT outperforms SOTA methods by up to 5.8% in accuracy and 7.7% in F1@50. The model and code will be released.

Existing RGB-IR (infrared) bi-modal 3D reconstruction methods generally have difficulty in simultaneously processing non-aligned multi-modal data with significant differences in resolution and spectral characteristics and achieving high-precision pixel-level reconstruction. Non-aligned RGB-IR 3D reconstruction and rendering represents a new domain. To this end, we propose BiNAR, a bi-modal framework that can directly process non-aligned data collected by conventional RGB and IR cameras and generate high-resolution, pixel-level aligned renderings. BiNAR first uses cross-modal multi-camera joint calibration to estimate the internal and external parameters of the RGB-IR camera and unify the coordinate system; then, it fuses the features of different modalities in the Unified Gaussian Field and jointly optimizes the Gaussians to achieve cross-modal consistent 3D scene expression. Experimental results show that BiNAR significantly outperforms traditional single-modal and bi-modal Gaussian splatting methods in rendering quality, achieving a sub-pixel average reprojection error of 0.242 px and improves IR PSNR by 8.00 dB. We also build a pixel-level aligned RGB-IR dataset covering a variety of indoor and outdoor scenes and including real temperature data, providing a reliable benchmark for subsequent multi-modal research. The code and dataset are available at https://github.com/jankin-wang/BiNAR.

Object detection in dynamic construction environments presents significant challenges due to vast scale variations, occlusions, and clutter. Conventional deep learning models struggle to balance the semantic information needed for classification with the spatial detail required for localization. This paper introduces a novel framework that systematically fuses features from different network depths to resolve this trade-off. Our primary contribution is a Hierarchical Feature Adjustment architecture that employs a coarse-to-fine strategy, progressively adjusting detections. We enhance robustness with an Efficient RoI Aggregation module for contextual aggregation and improve localization with a Modified IoU loss. Furthermore, a proposed Overlap Discriminating Module aids non-maximum suppression in dense scenes. Extensive experiments on the SODA, COD, and Small Tools datasets show our integrated approach significantly outperforms state-of-the-art methods, establishing a new benchmark for this critical application. We also demonstrate the model's generalizability to other domains with strong results on the BDD100K and COCO datasets.

Ensembling fine-tuned models initialized from powerful pre-trained weights is a common strategy to improve robustness under distribution shifts, but it comes with substantial computational costs due to the need to train and store multiple models. Dropout offers a lightweight alternative by simulating ensembles through random neuron deactivation; however, when applied to pre-trained models, it tends to over-regularize and disrupt critical representations necessary for generalization. In this work, we investigate Mixout, a stochastic regularization technique that provides an alternative to Dropout for domain generalization. Rather than deactivating neurons, Mixout mitigates overfitting by probabilistically swapping a subset of fine-tuned weights with their pre-trained counterparts during training, thereby maintaining a balance between adaptation and retention of prior knowledge. Our study reveals that achieving strong performance with Mixout on domain generalization benchmarks requires a notably high masking probability of 0.9 for ViTs and 0.8 for ResNets. While this may seem like a simple adjustment, it yields two key advantages for domain generalization: (1) higher masking rates more strongly penalize deviations from the pre-trained parameters, promoting better generalization to unseen domains; and (2) high-rate masking substantially reduces computational overhead, cutting gradient computation by up to 45% and gradient memory usage by up to 90%. Experiments across five domain generalization benchmarks--PACS, VLCS, OfficeHome, TerraIncognita, and DomainNet--using ResNet and ViT architectures show that our approach, High-rate Mixout, achieves out-of-domain accuracy comparable to ensemble-based methods while significantly reducing training costs.

The increasing demand for on-device AI, driven by privacy concerns and the need for real-time processing, poses new challenges for fundamental computer vision tasks. This paper addresses one such task, person clustering in photo galleries, which has traditionally relied on server-side computation or simplistic on-device models. We introduce the Multimodal Person Clustering Architecture (MPCA), a research framework designed to explore the feasibility of a high-performance, multimodal clustering pipeline operating entirely under mobile constraints. Our framework makes three principal contributions: (1) Multimodal Appearance-Assisted Identity Recovery (MAIR), a late-fusion strategy that leverages temporal consistency to recover identities when facial data is unreliable; (2) Language-Guided Appearance Extractor (LGAE), which adapts a vision-language paradigm to construct robust appearance representations efficiently; and (3) Sequential Graph-Density Clustering (SGDC), a novel algorithm that synergistically combines graph-based and density-based methods to handle the high variance of appearance data. We demonstrate through extensive experiments that our on-device framework achieves an unprecedented 87.97% average recall, significantly outperforming leading cloud-based commercial systems like Google Photos (77.74%) and on-device systems like Apple Photos (67.84%) and Samsung Gallery (83.39%). This work provides a blueprint for future research in privacy-preserving, efficient, and robust person clustering, highlighting a viable path for deploying next-generation computer vision applications directly on mobile devices.

While contrast-enhanced CT (CECT) is standard for assessing abdominal aortic aneurysms (AAA) , the required iodinated contrast agents pose significant risks, including nephrotoxicity, patient allergies, and environmental harm. To reduce contrast agent use, recent deep learning methods have focused on generating synthetic CECT from non-contrast CT (NCCT) scans. However, most adopt a multi-stage pipeline that first generates images and then performs segmentation, which leads to error accumulation and fails to leverage shared semantic and anatomical structures.To address this, we propose a unified deep learning framework that generates synthetic CECT images from NCCT scans while simultaneously segmenting the aortic lumen and thrombus. Our approach integrates conditional diffusion models (CDM) with multi-task learning, enabling end-to-end joint optimization of image synthesis and anatomical segmentation. Unlike previous multitask diffusion models, our approach requires no initial predictions (e.g., a coarse segmentation mask), shares both encoder and decoder parameters across tasks, and employs a semi-supervised training strategy to learn from scans with missing segmentation labels, a common constraint in real-world clinical data.We evaluated our method on a cohort of 264 patients, where it consistently outperformed state-of-the-art single-task and multi-stage models. For image synthesis, our model achieved a PSNR of 25.61 dB, compared to 23.80 dB from a single-task CDM. For anatomical segmentation, it improved the lumen Dice score to 0.89 from 0.87 and the challenging thrombus Dice score to 0.53 from 0.48 (nnU-Net). These segmentation enhancements led to more accurate clinical measurements, reducing the lumen diameter MAE to 4.19 mm from 5.78 mm and the thrombus area error to 33.85% from 41.45% when compared to nnU-Net.