WACV 2026 Papers

Accurate disease interpretation from radiology remains challenging due to imaging heterogeneity. Achieving expert-level diagnostic decisions requires integration of subtle image features with clinical knowledge. Yet major vision-language models (VLMs) treat images as holistic entities and overlook fine-grained image details that are vital for disease diagnosis. Clinicians analyze images by utilizing their prior medical knowledge and identify anatomical structures as important region of interests (ROIs). Inspired from this human-centric workflow, we introduce Anatomy-VLM, a fine-grained, vision-language model that incorporates multi-scale information. First, we design a model encoder to localize key anatomical features from entire medical images. Second, these regions are enriched with structured knowledge for contextually-aware interpretation. Finally, the model encoder aligns multi-scale medical information to generate clinically-interpretable disease prediction. Anatomy-VLM achieves outstanding performance on both in- and out-of-distribution datasets. We also validate the performance of Anatomy-VLM on downstream image segmentation tasks, suggesting that its fine-grained alignment captures anatomical and pathology-related knowledge. Furthermore, the Anatomy-VLM's encoder facilitates zero-shot anatomy-wise interpretation, providing its strong expert-level clinical interpretation capabilities.

We present V-HPOT, a novel approach for improving the zero-shot performance of 3D hand pose estimation from egocentric images across diverse, unseen domains. State-of-the-art methods demonstrate strong performance when trained and tested within the same domain. However, they struggle to generalise to new environments due to limited training data and depth perception -- overfitting to specific camera intrinsics. Our method addresses this by estimating keypoint z-coordinates in a virtual camera space, normalised by focal length and image size, enabling camera-agnostic depth prediction. We further leverage this invariance to camera intrinsics to propose a self-supervised test-time optimisation strategy that refines the model's depth perception during inference. This is achieved by applying a 3D consistency loss between predicted and in-space scale-transformed hand poses, allowing the model to adapt to target domain characteristics without requiring ground truth annotations. V-HPOT significantly improves 3D hand pose estimation performance in zero-shot scenarios, achieving a 71% reduction in mean pose error on the H2O dataset and a 41% reduction on the AssemblyHands dataset. Compared to state-of-the-art methods, V-HPOT outperforms all single-stage approaches across all datasets and competes closely with two-stage methods, despite needing x3.5 to x14 less data. The code is planned to be released under https://github.com/tobereleased.

Cross-view pose estimation entails predicting the relative 3 Degrees-of-Freedom (3DoF) pose of an image within an aerial view. Existing work focuses on imagery in controlled settings featuring highly constrained parameters. In contrast, a wide variety of camera parameterizations are seen in-the-wild across tasks where such estimation is useful. To address this gap, we propose a method capable of performing cross-view pose estimation in these less constrained scenarios with ground-view images of unknown FoV, pitch, roll, and projection type (panoramic or rectilinear). Namely, our method avoids common assumptions--such as gravity/horizon alignment needed for geometric-based projections--and purely relies on a transformer to learn the cross-view relationships in a data-driven manner, paired with prediction modules to enable continuous querying of the pose search space. Evaluations of our approach demonstrates it's ability to perform competitively with the state-of-the-art over the VIGOR benchmark, while maintaining performance in those harder less constrained scenarios. This supports our work as the first generalized approach to this task that is capable of operating with less-constrained imagery.

Explainable AI (XAI) is essential for trust and transparencyin deep learning, especially in medical imaging.Existing local explanation methods provide per-instance insightsbut fail to show whether similar explanations holdacross samples of the same class. This limits global interpretabilityand demands time-consuming manual reviewby clinicians to trust models in practice. We introduce theConsensus Alignment Score (CAS), a novel metric thatquantifies consistency of explanations at the class level.We also present ConsensusXAI, an open-source, modelandmethod-agnostic framework that evaluates explanationagreement quantitatively (via CAS) and qualitatively(through consensus heatmaps) per class. Unlike priorbenchmarks, ConsensusXAI uses a latent-space clusteringapproach, Latent Consensus, to identify dominant explanationpatterns, exposing biases and inconsistencies towardscertain classes. Evaluated across four benchmark datasetsand two imaging modalities, our method consistently revealsmeaningful class-level insights, outperforming traditionalmetrics like SSIM and IoU, and enabling faster, moreconfident clinical adoption of AI models.

Traditional object detection models in medical imaging operate within a closed-set paradigm, limiting their ability to detect objects of novel labels. Open-vocabulary object detection (OVOD) addresses this limitation but remains underexplored in medical imaging due to dataset scarcity and weak text-image alignment. To bridge this gap, we introduce MedROV, the first Real-time Open Vocabulary detection model for medical imaging. To enable open-vocabulary learning, we curate a large-scale dataset, Omnis, with 600K detection samples across nine imaging modalities and introduce a pseudo-labeling strategy to handle missing annotations from multi-source datasets. Additionally, we enhance generalization by incorporating knowledge from a large pre-trained foundation model. By leveraging contrastive learning and cross-modal representations, MedROV effectively detects both known and novel structures. Experimental results demonstrate that MedROV outperforms the previous state-of-the-art foundation model for medical image detection with an average absolute improvement of 40 mAP50, and surpasses closed-set detectors by more than 3 mAP50, while running at 70 FPS, setting a new benchmark in medical detection. Our source code, dataset, and trained model are available at https://github.com/toobatehreem/MedROV.

Self-supervised pretraining methods with masked prediction demonstrate remarkable within-dataset performance in skeleton-based action recognition. However, we show that, unlike contrastive learning approaches, they do not produce well-separated clusters. Additionally, these methods struggle with generalization in few-shot settings. To address these issues, we propose Self-supervised Tuning for 3D Action Recognition in Skeleton sequences (STARS). Specifically, STARS first uses a masked prediction stage using an encoder-decoder architecture. It then employs nearest-neighbor contrastive learning to partially tune the weights of the encoder, enhancing the formation of semantic clusters for different actions. By tuning the encoder for a few epochs, and without using hand-crafted data augmentations, STARS achieves state-of-the-art self-supervised results in various benchmarks, including NTU-60, NTU-120, and PKU-MMD. In addition, STARS exhibits significantly better results than masked prediction models in few-shot settings, where the model has not seen the actions throughout pretraining.

Unsupervised optical flow methods have become more popular in the last decade, enabling the training of models across domains without ground truth data. Although RAFT and its successors have achieved significant success in the supervised settings, many unsupervised approaches continue to use older backbones such as PWC-Net. One reason for this architectural stagnation is that the current RAFT-based SOTA approach has proven challenging for the community to reproduce. In this paper, we revive and advance unsupervised optical flow: First, we introduce Sun-RAFT: a simple unsupervised RAFT. Second, building on Sun-RAFT, we present Muun-RAFT: a novel multi-scale unsupervised RAFT, where we propose a gradual context-based upsampling to refine the flow, further improving both accuracy and preservation of details. Third, we reexamine previously advised unsupervised strategies to identify effective training settings. In terms of results, both our methods demonstrate strong generalization capabilities and set a new SOTA for unsupervised two-frame approaches on MPI-Sintel, with Muun-RAFT surpassing even the current multi-frame SOTA by up to 28%. Finally, we open-source our PyTorch code, enabling further developments in the field: https://cv-stuttgart.github.io/Reviving-Unsupervised-OpticalFlow.

Instance-level image retrieval aims to find images containing the same object as a given query, despite variations in size, position, or appearance. To address this challenging task, we propose Patchify, a simple yet effective patch-wise retrieval framework that offers high performance, scalability, and interpretability without requiring fine-tuning. Patchify divides each database image into a small number of structured patches and performs retrieval by comparing these local features with a global query descriptor, enabling accurate and spatially grounded matching. To assess not just retrieval accuracy but also spatial correctness, we introduce LocScore, a localization-aware metric that quantifies whether the retrieved region aligns with the target object. This makes LocScore a valuable diagnostic tool for understanding and improving retrieval behavior. We conduct extensive experiments across multiple benchmarks, backbones, and region selection strategies, showing that Patchify outperforms global methods and complements state-of-the-art reranking pipelines. Furthermore, we apply Product Quantization for efficient large-scale retrieval and highlight the importance of using informative features during compression, which significantly boosts performance.

Federated Learning (FL) enables collaborative model training without sharing raw data, but remains highly vulnerable to gradient manipulation and backdoor attacks, particularly under heterogeneous client distributions. Most existing defenses either target a narrow class of attacks, rely on client data, or fail to adapt in heterogeneous settings. We propose SD-CSFL (Synthetic Data-Driven Conformity Scoring for Federated Learning), a unified and privacy-preserving defense algorithm. SD-CSFL leverages a synthetic calibration dataset, independent of client data, to compute entropy-based nonconformity scores that capture irregularities in client updates. An adaptive percentile thresholding mechanism with stratified calibration dynamically distinguishes benign from malicious updates across training rounds. We establish a conformal-prediction-based guarantee showing that percentile thresholds bound false positives under arbitrary score distributions. Experiments on CIFAR-10 and Birds-525 demonstrate up to 35% higher detection of gradient manipulation and an 80% reduction in backdoor success rates, outperforming recent defenses in heterogeneous environments. Our implementations and synthetic datasets are available at https://github.com/EbtisaamCS/SD-CSFL

Segmenting texture-rich images is challenging due to the presence of various local ambiguities. These are common in domains like metallography, where scanning electron microscopy (SEM) images must be segmented to quantify different elements that determine material properties. Conventional encoder-decoder models often struggle in these scenarios because they rely solely on local features and lack global distributional awareness. We propose a guided segmentation framework that introduces coordinate-aware class-ratio mapping. It is a mechanism that transforms expected class proportions into spatial maps and integrates them with encoder features via an adaptive gate fusion module. This enforces consistency between global class ratios and local pixel predictions, allowing the model to resolve ambiguous textures more effectively. Unlike traditional methods, we condition the model on image-specific global ratios, which can be obtained from experts or estimated in an auto-regressive method. Extensive experiments on metallographic SEM benchmarks demonstrate that our framework consistently improves performance across diverse backbones, achieving up to +6.1% Dice score improvements with minimal parameter overhead (<2%). Ablation studies confirm that both coordinate-aware class-ratio mapping and the adaptive gate fusion module contribute complementary benefits. In addition, we demonstrate that segmentation performance improves provided that the accuracy of the input class ratios is at least 60%.

Precise manipulation of dermoscopic attributes is essential for augmenting long-tailed datasets and enhancing diagnostic interpretability. However, generic diffusion-based editing methods often induce global distortions or fail to preserve the peri-lesional context critical for diagnosis. To address these limitations, we propose Context-Preserving Dermoscopic Editing (CPDE), a framework tailored for lesion-aware attribute modification. CPDE introduces a dual-branch diffusion pipeline that disentangles lesion editing from background reconstruction. Attribute changes are predicted by a Spatial--channel Transformer at the U-Net bottleneck and are trained with a lesion-aware objective that enforces semantic directionality only inside pathology regions. On the ISIC 2017 and 2018 datasets, CPDE produces spatially localized, clinically coherent edits that preserve lesion extent and surrounding skin. Our method achieves superior image fidelity with an FID of 0.274, while significantly outperforming existing approaches in terms of semantic alignment and background preservation.

Detecting objects in 3D space using a monocular image is inherently a highly ill-posed problem: multiple plausible 3D bounding boxes can explain the same 2D observation of an object. Existing approaches typically follow a single-point prediction paradigm, failing to capture this multimodal nature and often regressing to an implausible mean solution. This paper introduces MonoMH, a novel multi-hypothesis framework for monocular 3D object detection. By explicitly modeling and learning the multimodal distribution of plausible 3D object configurations, MonoMH not only significantly improves detection performance but also provides richer information to support downstream decision-making. MonoMH introduces three key innovations: (1) a novel multi-hypothesis predictor that leverages spatially diverse features across different windows within an RoI to generate a rich variety of hypotheses without increasing model complexity; (2) a new multi-hypothesis learning approach that derives diverse and relevant hypotheses from single-modal ground truth by integrating uncertainty modeling with best-of-many learning; and (3) a novel adaptive hypothesis filtering mechanism that enhances detection capability by dynamically retaining a variable number of plausible hypotheses based on each object's uncertainty. Experimental results demonstrate the effectiveness of our approach. Notably, MonoMH achieves 29.12 (easy), 20.88 (mod.), and 17.93 (hard) Car AP3D on the KITTI test set, a significant boost over the previous state-of-the-art methods. The code can be found at https://github.com/HyeonjeongPark37/MonoMH.

While deep learning-based models like transformers, have revolutionized time-series and vision tasks, they remain highly susceptible to noise and often overfit on noisy patterns rather than robust features. This issue is exacerbated in vision transformers, which rely on pixel-level details that can easily be corrupt. To address this, we leverage the discrete wavelet transform (DWT) for its ability to decompose into multi-resolution layers, isolating noise primarily in the high frequency domain while preserving essential low-frequency information for resilient feature learning. Conventional DWT-based methods, however, struggle with computational inefficiencies due to the requirement for a subsequent inverse discrete wavelet transform (IDWT) step. In this work, we introduce RobustFormer, a novel framework that enables noise-robust masked autoencoder (MAE) pre-training for both images and videos by using DWT for efficient downsampling, eliminating the need for expensive IDWT reconstruction and simplifying the attention mechanism to focus on noise-resilient multi-scale representations. To our knowledge, RobustFormer is the first DWT-based method fully compatible with video inputs and MAE-style pre-training. Extensive experiments on noisy image and video datasets demonstrate that our approach achieves up to 8% increase in Top-1 classification accuracy under severe noise conditions in Imagenet-C and up to 2.7% in Imagenet-P standard benchmarks compared to the baseline and up to 13% higher Top-1 accuracy on UCF-101 under severe custom noise perturbations while maintaining similar accuracy scores for clean datasets. We also observe the reduction of computation complexity by up to 4.4% through IDWT removal compared to VideoMAE baseline without significant performance drop.

Foundation models (FM) in digital pathology have revolutionized the field of whole slide image (WSI) analysis, with models such as UNI, Virchow, Prov-GigaPath, and many more outperforming the previously established benchmarks set by the ImageNet-based backbones. However, despite several benchmarking studies, there has been no clear consensus on the choice of a single FM that is best suited for a variety of histopathology datasets and/or tasks. With more than 25 pathology FMs in the literature so far, the challenge of model selection is a growing concern. Although an ensemble of FMs can circumvent this issue, given the bulky nature of individual FMs, the inference time and computational cost drastically increase with the addition of each FM into the ensemble. To this end, we propose HistoMILKD, the first multi-teacher knowledge distillation (MKD) framework for WSI classification. To handle the gigapixel resolution of WSIs, we use multiple instance learning (MIL), making this also the first work to integrate MIL and MKD frameworks into a single model. Our approach leverages the complementary representations of different FMs to distill collective task-specific knowledge into a single trainable MIL adapter on top of the student FM, which is utilized during inference. Evaluated on five public datasets, the proposed approach significantly (p<0.05) outperforms the individual FMs, their ensemble, and previous MKD approaches in WSI classification. Codes are available at https://github.com/AIMLab-UBC/HistoMILKD.

In this work, we present MUSE (Model-based Uncertainty-aware Similarity Estimation), a training-free framework for model-based zero-shot 2D object detection and segmentation. First, MUSE incorporates 2D multi-view templates from 3D unseen objects and 2D object proposals from the input query image, respectively. In the embedding stage, we propose a new feature embedding scheme which integrates class and patch embeddings. Specifically, the patch embeddings are normalized using the generalized mean pooling (GeM). In the matching stage, a joint similarity score is introduced, which integrates an absolute score and a relative score. Finally, we update the similarity score using an uncertainty-aware object prior. MUSE achieves state-of-the-art performance on the BOP Challenge 2025, ranking first in the Classic Core, H3, and Industrial tracks--without any additional training or fine-tuning. Therefore, we believe that MUSE is a promising framework for zero-shot 2D object detection and segmentation.

Open-set driving in complex real-world environments requires reliable identification of out-of-distribution (OOD) objects to avoid overconfident predictions on unseen categories. However, the sparsity and limited semantic richness of LiDAR point clouds make separating known and unknown classes difficult. This work proposes DOODLE, a diffusion model-based OOD learning framework for open-set 3D semantic segmentation. DOODLE trains a diffusion model to reconstruct in-distribution semantic features; feature-level reconstruction discrepancies then serve as OOD evidence. The resulting OOD scores are used to enhance backbone semantic features, improving discrimination of unknown regions during segmentation. To further reduce false positives arising from nonuniform measurements, a density-aware post-processing (DAP) module incorporates spatial variation in LiDAR point density when refining OOD predictions. DOODLE integrates seamlessly with existing open-set models and does not constrain backbone design. Experiments on SemanticKITTI and nuScenes demonstrate state-of-the-art OOD performance. On SemanticKITTI, DOODLE improves area under the precision-recall curve (AUPR) by 1.85%p and area under the receiver operating characteristic (AUROC) by 1.29%p over prior methods. Ablation studies confirm complementary benefits from diffusion-based reconstruction and DAP. Code is available at https://github.com/chang9711/DOODLE.

Video anomaly detection (VAD) has long been studied as a crucial problem in public security and crime prevention. In recent years, weakly-supervised VAD (WVAD) has attracted considerable attention due to its easy annotation process and promising research results. While existing WVAD methods mainly tackle static datasets, the possibility that the domain of data can vary has been neglected. To adapt such domain shift, the continual learning (CL) perspective is required because otherwise additional training using only newly incoming data could easily cause performance degradation for previous data, i.e., forgetting. Therefore, we propose a brand-new approach, called Continual Anomaly Detection with Ensembles (CADE) that is the first work combining CL and WVAD viewpoints. Specifically, CADE uses the Dual-Generator (DG) to address data imbalance and label uncertainty in WVAD. We also found that forgetting exacerbates the "incompleteness" where the model becomes biased towards certain anomaly modes, leading to missed detections of various anomalies. To address this, we propose a Multi-Discriminator (MD) ensemble that captures missed anomalies in past scenes due to forgetting, using multiple models. Extensive experiments show that CADE significantly outperforms existing VAD methods on the common multi-scene VAD datasets, such as the ShanghaiTech, UCF-Crime, and Charlotte Anomaly Dataset.

Recent research has demonstrated the potential of deep neural networks (DNNs) to accurately predict wildfire spread on a given day based upon high-dimensional explanatory data from a single preceding day, or from a time series of T preceding days. For the first time, we investigate a large number of existing data-driven wildfire modeling strategies under controlled conditions, revealing the best modeling strategies and resulting in models that achieve state-of-the-art (SOTA) accuracy for both single-day and multi-day input scenarios, as evaluated on a large public benchmark for next-day wildfire spread, termed the WildfireSpreadTS (WSTS) benchmark. Consistent with prior work, we found that models using time-series input obtained the best overall accuracy, suggesting this is an important future area of research. Furthermore, we create a new benchmark, WSTS+, by incorporating four additional years of historical wildfire data into the WSTS benchmark. Our benchmark doubles the number of unique years of historical data, expands its geographic scope, and, to our knowledge, represents the largest public benchmark for time-series-based wildfire spread prediction.

Despite recent text-to-image models achieving highfidelity text rendering, they still struggle with long or multiple texts due to diluted global attention. We propose DCText, a training-free visual text generation method that adopts a divide-and-conquer strategy, leveraging the reliable short-text generation of Multi-Modal Diffusion Transformers. Our method first decomposes a prompt by extracting and dividing the target text, then assigns each to a designated region. To accurately render each segment within their regions while preserving overall image coherence, we introduce two attention masks--Text-Focus and Context-Expansion--applied sequentially during denoising. Additionally, Localized Noise Initialization further improves text accuracy and region alignment without increasing computational cost. Extensive experiments on single- and multisentence benchmarks show that DCText achieves the best text accuracy without compromising image quality while also delivering the lowest generation latency.