WACV 2026 Papers

While there is rapid progress in video-LLMs with advanced reasoning capabilities, prior work shows that these models struggle on the challenging task of sports feedback generation and require expensive and difficult-to-collect finetuning feedback data for each sport. This limitation is evident from the poor generalization to sports unseen during finetuning. Furthermore, traditional text generation evaluation metrics (e.g., BLEU-4, METEOR, ROUGE-L, BERTScore), originally developed for machine translation and summarization, fail to capture the unique aspects of sports feedback quality. To address the first problem, using rock climbing as our case study, we propose using auxiliary freelyavailable web data from the target domain, such as competition videos and coaching manuals, in addition to existing sports feedback from a disjoint, source domain to improve sports feedback generation performance on the target domain. To improve evaluation, we propose two evaluation metrics: (1) specificity and (2) actionability. Together, our approach enables more meaningful and practical generation of sports feedback under limited annotations.

Domain generalisation aims to develop models that generalise from source domains to unseen target domains. However, most existing methods assume access to source domain data and require additional training, which may not always be practical. We focus on a more flexible and broadly applicable setting, zero-shot domain generalisation, where models generalise without access to source data, target data, or any additional training. In this work, we propose Prefer (prompt-driven feature refinement), a simple and effective approach which enhances the zero-shot domain generalisation ability of vision-language foundation models. Prefer generates a diverse set of textual prompts for each class by imagining domain-specific variations (e.g., "a painting of a cat under a golden sunset with thick brush strokes"), and uses them to probe the model. We evaluate how reliably each feature channel represents a class across domains by measuring two quantities: (1) how strongly the channel aligns with the original class prompt (e.g., "a photo of a cat") across the generated domain-specific prompts, and (2) how stable the channel remains across those prompts, quantified by its variance. Channels that exhibit both high alignment and low variability are selected at inference time to improve class prediction under domain shift. Without any model updates or external data, Prefer achieves consistent improvements across domain generalisation benchmarks, outperforming existing state-of-the-art methods. The source code is available at https://anonymous.4open.science/r/WACV26-Prefer.

Forecasting abnormal human behavior (AHB) in unconstrained real-world environments is critical for enabling proactive safety interventions. Unlike short-term anomaly detection, long-horizon forecasting offers a vital reaction window but remains underexplored due to three core challenges: (i) noisy, complex human-agent interactions; (ii) weak temporal coupling between normal observations and distant anomalies; and (iii) data scarcity limiting the scalability of autoregressive models. To address these, we propose \mathcal D ^3\mathcal P (Denoise, Divide, Distill, and Predict), a novel encoder-decoder framework that bridges denoised pasts with distilled autoregressive futures. Our Differential Past Encoder (DiPE) disentangles scene-level and object-level dynamics via differential attention, suppressing irrelevant interactions and enhancing discriminative cues. The Distilled Future Auto-Regressive Decoder (D-FAD) adopts a divide-and-conquer strategy, segmenting future queries into temporal chunks for sequential prediction, while leveraging distillation to balance robustness and latency. We validate our approach on the AHB-F benchmark, the only dataset dedicated to abnormal behavior forecasting, and further integrate D-FAD with several state-of-the-art methods. In all cases, our framework consistently outperforms prior work in both forecasting accuracy and computational efficiency.

Simultaneous Localization and Mapping (SLAM) is one of the most important environment-perception and navigation algorithms for computer vision, robotics, and autonomous cars/drones. Hence, high quality and fast mapping becomes a fundamental problem. With the advent of 3D GaussianSplatting (3DGS) as an explicit representation with excellent rendering quality and speed, state-of-the-art (SOTA) works introduce GS to SLAM. Compared to classical pointcloud-SLAM, GS-SLAM generates photometric information by learning from input camera views and synthesize unseen views with high-quality textures. However, these GS-SLAM fail when moving objects occupy the scene that violate the static assumption of bundle adjustment. The failed updates of moving GS affect the static GS and contaminates the full map over long frames. Although some efforts have been made by concurrent works to consider moving objects for GS-SLAM, they simply detect and remove the moving regions from GS rendering ("anti" dynamic GS-SLAM), where only the static background could benefit from GS. To this end, we propose the first real-time GS-SLAM, "DynaGSLAM", that achieves high-quality online GS rendering, tracking, motion predictions of moving objects in dynamic scenes while jointly estimating accurate ego motion. Our DynaGSLAM outperforms SOTA static & "Anti" dynamic GS-SLAM on three dynamic real datasets, while keeping speed and memory efficiency in practice.

Quantization-Aware Training (QAT) for vision transformers relies on expensive retraining to recover accuracy loss in non-linear layer quantization, which limits their use in resource-constrained environments. In contrast, existing Post-Training Quantization (PTQ) methods either partially quantize non-linear functions or adjust activation distributions to maintain accuracy but fail to achieve fully integer-only inference. In this paper, we introduce IPTQ-ViT, a PTQ framework for fully integer-only vision transformers without retraining. We present approximation functions: a polynomial-based GELU optimized for vision data and a bit-shifting-based Softmax designed to improve approximation accuracy in PTQ. In addition, we propose a unified metric integrating quantization sensitivity, perturbation, and computational complexity to select the optimal approximation function per activation layer. IPTQ-ViT consistently outperforms previous PTQ methods, achieving up to 6.44%p (avg. 1.78%p) top-1 accuracy improvement for image classification and 1.0 mAP for object detection under W8A8 and W4A8. IPTQ-ViT achieves accuracy and latency comparable to integer-only QAT methods.

Associating the in-vivo function of neurons with their ex-vivo anatomical structure is a central challenge in neuroscience. However, this field is constrained by a critical bottleneck: the extreme difficulty of acquiring paired cross-modal data, leading to a persistent scarcity of large-scale datasets. This inherent limitation frames the re-identification of the same neuron as a formidable few-shot, fine-grained visual recognition task. To address this challenge, we propose a novel deep metric learning framework designed to learn modality-invariant feature representations for single neurons under these data-scarce conditions. The core of this framework is a dual-channel network architecture that explicitly disentangles and fuses the local morphological information of the neuron's soma with the global topological context of the dendritic arbor, thereby capturing a more robust neural signature. To maximize data efficiency, we integrate a Circle Loss objective with a Multi-Similarity hard-sample mining strategy, which effectively optimizes the embedding space for better class separation. On a cross-modal neuron dataset that realistically reflects experimental data scarcity, our method demonstrates excellent performance, achieving a Recall of 77.4% and a Specificity of 90.1% on the test set. Extensive ablation studies and comparative analyses validate the effectiveness of our proposed method, establishing a new strong baseline for this critical yet data-limited biomedical application. To foster future research in this field, we will release our code, dataset, and pre-trained models.

Transformer architecture search (TAS) aims to automatically discover efficient vision transformers (ViTs), reducing the need for manual design. Existing TAS methods typically train an over-parameterized network (i.e., a supernet) that encompasses all candidate architectures (i.e., subnets). However, all subnets share the same set of weights, which leads to interference that degrades the smaller subnets severely. We have found that well-trained small subnets can serve as a good foundation for training larger ones. Motivated by this, we propose a progressive training framework, dubbed GrowTAS, that begins with training small subnets and incorporate larger ones gradually. This enables reducing the interference and stabilizing a training process. We also introduce GrowTAS+ that fine-tunes a subset of weights only to further enhance the performance of large subnets. Extensive experiments on ImageNet and several transfer learning benchmarks, including CIFAR-10/100, Flowers, CARS, and INAT-19, demonstrate the effectiveness of our approach over current TAS methods.

Novel Category Discovery (NCD) methods for LiDAR Semantic Segmentation (LSS) assume that labeled and unlabeled points coexist in every scan and that all unlabeled points belong solely to novel categories. We formalize this into a more practical task, Generalized Category Discovery (GCD) for LSS, in which the labeled and unlabeled subsets are disjoint, and the unlabeled data contain a mixture of known and novel categories. Existing 2D GCD methods fail under this setting, struggling to distinguish the two groups in sparse, imbalanced LiDAR data. To address this limitation, we present a unified framework that (i) employs a learnable adaptive threshold to obtain point-wise anomaly scores to capture candidates, (ii) refines these candidates through a clustering-based filtering mechanism, and (iii) stabilizes training with a novel-feature queue that supplies reliable novel features even when a scene lacks novel categories. This explicit modeling of novel categories preserves segmentation quality for known classes while markedly improving discovery performance--a direction not explored in prior GCD or NCD methods. Extensive experiments on SemanticKITTI and nuScenes demonstrate that our approach consistently surpasses adapted baselines, establishing a strong benchmark for future work in open-world LSS.

Existing video deraining methods are often trained on paired datasets, either synthetic, which limits their ability to generalize to real-world rain, or captured by static cameras, which restricts their effectiveness in dynamic scenes with background and camera motion. Furthermore, recent works in fine-tuning diffusion models have shown promising results, but the fine-tuning tends to weaken the generative prior, limiting generalization to unseen cases. In this paper, we introduce the first zero-shot video deraining method for complex dynamic scenes that does not require synthetic data nor model fine-tuning, by leveraging a pretrained text-to-video diffusion model that demonstrates strong generalization capabilities. By inverting an input video into the latent space of diffusion models, its reconstruction process can be intervened and pushed away from the model's concept of rain using negative prompting. At the core of our approach is an attention switching mechanism that we found is crucial for maintaining dynamic backgrounds as well as structural consistency between the input and the derained video, mitigating artifacts introduced by naive negative prompting. Our approach is validated through extensive experiments on real-world rain datasets, demonstrating substantial improvements over prior methods and showcasing robust generalization without the need for supervised training.

Intrinsic image decomposition aims at separating an image into its underlying albedo and shading components, isolating the base color from lighting effects to enable downstream applications such as virtual relighting and scene editing.Despite the rise and success of learning-based approaches, intrinsic image decomposition from real-world images remains a significantly challenging task due to the scarcity of labeled ground-truth data.Most existing solutions rely on synthetic data as supervised setups, limiting their ability to generalize to real-world scenes. Self-supervised methods, on the other hand, often produce albedo-like maps that contain reflections and lack consistency under different lighting conditions.To address this, we propose SAIL, an approach designed to estimate illumination-invariant representations from single-view real-world images to specifically target plausible relighting. We repurpose the prior knowledge of a latent diffusion model for unconditioned scene relighting as a surrogate objective for learning light-invariant estimates. To achieve this, we introduce a novel intrinsic image decomposition fully formulated in the latent space.To guide the training of our latent diffusion model, we introduce regularization terms that constrain both the lighting-dependent and -independent components of our latent image decomposition.Through our experiments, we demonstrate that SAIL produces stable albedo-like representations under varying lighting conditions and generalizes to multiple scenes, using only unlabeled multi-illumination data available online.

Aerial object detection is constrained by tiny targets, large scale variation, and strict real-time limits. It supports traffic monitoring, disaster response, and infrastructure inspection. Yet current detectors often ignore available platform metadata and process frames in isolation. This omission prevents receptive fields from adapting to scale variation and reduces accuracy. We propose Meta-YOLO to exploit platform metadata for scale-aware aerial object detection in real time. Meta-YOLO injects normalized telemetry into spatial sampling to guide feature extraction. It modulates deformable convolution offsets using a spatial metadata map aligned with the image. This links visual features with platform state and enables receptive fields to adapt to object scale. Built on YOLOX, Meta-YOLO adds two modules: feature modulation and offset correction. Evaluated on 327K aerial frames with metadata, Meta-YOLO achieves up to +8.7 AP gains over YOLOX in lightweight regimes and consistently outperforms other recent detectors. It preserves real-time throughput with negligible overhead and improves accuracy without extra visual processing.

We address the Federated source-Free Domain Adaptation (FFreeDA) problem, with clients holding unlabelled data with significant inter-client domain gaps. The FFreeDA setup constrains the FL frameworks to employ only a pre-trained server model as the setup restricts access to the source dataset during the training rounds. Often, this source domain dataset has a distinct distribution to the clients' domains. To address the challenges posed by the FFreeDA setup, adaptation of the Source-Free Domain Adaptation (SFDA) methods to FL struggles with client-drift in real-world scenarios due to extreme data heterogeneity caused by the aforementioned domain gaps, resulting in unreliable pseudo-labels. In this paper, we introduce FedSCAl, an FL framework leveraging our proposed Server-Client Alignment (SCAl) mechanism to regularize client updates by aligning the clients' and server model's predictions. We observe an improvement in the clients' pseudo-labeling accuracy post alignment, as the SCAl mechanism helps to mitigate the client-drift. Further, we present extensive experiments on benchmark vision datasets showcasing how FedSCAl consistently outperforms state-of-the-art FL methods in the FFreeDA setup for classification tasks.

Hyperspectral anomaly detection (HAD) aims to identify rare and irregular targets in high-dimensional hyperspectral images (HSIs), which are often noisy and unlabeled data. Existing deep learning methods either fail to capture long-range spectral dependencies (e.g., convolutional neural networks) or suffer from high computational cost (e.g., Transformers). To address these challenges, we propose DMS2F-HAD, a novel dual-branch Mamba-based model. Our architecture utilizes Mamba's linear-time modeling to efficiently learn distinct spatial and spectral features in specialized branches, which are then integrated by dynamic gated fusion mechanism to enhance anomaly localization. Across fourteen benchmark HSI datasets, our proposed DMS2F-HAD not only achieves a state-of-the-art average AUC of 98.78% but also demonstrates superior efficiency with an inference speed 4.6x faster than comparable deep learning methods. The results highlight DMS2F-HAD's strong generalization and scalability, positioning it as a strong candidate for practical HAD applications. The source code and models are available at https://github.com/Ayushma00/DMS2F-HAD.

Dynamic Neural Networks (DNNs) have emerged as a promising solution to improve the computational efficiency of deep neural networks by adaptively adjusting inference complexity based on input characteristics. Despite their advantages, the deployment of dynamic networks in real-world applications remains challenging because most methods are hard to adapt for practical use cases such as object detection, in combination with the lacking support of inference infrastructure.In this work, we present a dynamic neural network architecture specifically designed for object detection. Using our method, we build a variety of Pareto-optimal models for object detection on COCO for models in the 7-10 GFLOPs range.Additionally, to measure the routing efficacy, we introduce an evaluation metric that facilitates standardized benchmarking across different dynamic network approaches. Finally, we introduce an evaluation of a deployment pipeline utilizing the ONNX format, thus building a DNN that shows speedup in a realistic deployment scenario. Experimental results demonstrate the performance and practical viability of our approach for efficient object detection in resource-constrained scenarios.

Appearance-based gait recognition has achieved strong performance on controlled datasets, yet systematic evaluation of its robustness to real-world corruptions and silhouette variability remains lacking. We present RobustGait, a framework for fine-grained robustness evaluation of appearance-based gait recognition systems. RobustGait evaluation spans four dimensions: the type of perturbation (digital, environmental, temporal, occlusion), the silhouette extraction method (segmentation and parsing networks), the architectural capacities of gait recognition models, and various deployment scenarios. The benchmark introduces 15 corruption types at 5 severity levels across CASIA-B, CCPG, and SUSTech1K, with in-the-wild validation on MEVID, and evaluates six state-of-the-art gait systems. We came across several exciting insights. First, applying noise at the RGB level better reflects real-world degradation, and reveal how distortions propagate through silhouette extraction to the downstream gait recognition systems. Second, gait accuracy is highly sensitive to silhouette extractor biases, revealing an overlooked source of benchmark bias. Third, robustness is dependent on both the type of perturbation and the architectural design. Finally, we explore robustness-enhancing strategies, showing that noise-aware training and knowledge distillation improve performance and move toward deployment-ready systems.

This paper introduces a method and application for automatic detection of behavioral interaction between grazing cattle from a single image, which is essential for smart livestock management in the cattle industry, such as for detecting estrus. Although the interaction detection for humans has been actively studied, a non-trivial challenge lies in cattle interaction detection, i.e., the lack of a comprehensive behavioral dataset including interaction, since the interactions of grazing cattle are rare events. We, therefore, propose CattleAct, a data-efficient method for interaction detection by decomposing interactions into the combinations of actions by individual cattle. Specifically, we first learn an action latent space from a large-scale cattle action dataset, then embed rare interactions via the fine-tuning of the pre-trained latent space using contrastive learning, constructing a unified latent space of action and interactions. On top of the proposed method, we develop a practical working system integrating video and GPS inputs. Experiments in a commercial-scale pasture show the accurate interaction detection by our method compared to the baselines.

We present CalibBEV, a novel Bird's Eye View (BEV) alignment approach for LiDAR-camera calibration. Our method unifies LiDAR and camera data into a shared 3D spatial representation, enabling accurate and robust cross-modal calibration. CalibBEV extracts sensor-wise BEV features from each modality using domain-specific architectures and estimates the calibration matrix through a two-step alignment process. First, we perform an implicit alignment by regressing a coarse calibration matrix directly from the BEV features. To ease this alignment, we enforce semantic consistency between BEV representations across modalities using a contrastive loss inspired by CLIP, guiding both networks toward a unified feature space. In the second step, we leverage our BEV formulation to explicitly align the features of one modality with the other, refining the initial coarse estimate into a final, more accurate calibration matrix. CalibBEV significantly outperforms prior point-to-pixel matching methods, achieving state-of-the-art calibration accuracy. On the KITTI and nuScenes benchmarks, our method reduces the Relative Rotation Error (RRE) by 51% and 68%, and the Relative Translation Error (RTE) by 80% and 91%, respectively, compared to previous methods.

Joint compression of point cloud geometry and attributes is essential for efficient 3D data representation. Existing methods often rely on recoloring procedures and manually tuned bitrate allocation between geometry and attribute compressions in inference, which hinder end-to-end optimization and add system complexity. To overcome these limitations, we propose MEGA-PCC, a fully end-to-end, learning-based framework featuring two specialized models for joint compression. The main compression model employs a shared encoder that embeds both geometry and attribute information into a unified latent space, followed by dual decoders that sequentially reconstruct geometry and then attributes. Complementing this, the Mamba-based Entropy Model (MEM) enhances entropy coding by capturing spatial and channel-wise correlations to improve probability estimation. Both models are built on the Mamba architecture to effectively model long-range dependencies and rich contextual features. By eliminating the need for recoloring and heuristic bitrate tuning, MEGA-PCC enables data-driven bitrate allocation during training and simplifies the overall pipeline. Extensive experiments demonstrate that MEGA-PCC achieves superior rate-distortion performance and runtime efficiency compared to both traditional and learning-based baselines, offering a powerful solution for AI-driven point cloud compression.

With the rapid advancements in diffusion models and 3D generation techniques, dynamic 3D content generation has become a crucial research area. However, achieving high-fidelity 4D (dynamic 3D) generation with strong spatial-temporal consistency remains a challenging task. Inspired by recent findings that pretrained diffusion features capture rich correspondences, we propose FB-4D, a novel 4D generation framework that integrates a Feature Bank mechanism to enhance both spatial and temporal consistency in generated frames. In FB-4D, we store features extracted from previous frames and fuse them into the process of generating subsequent frames, ensuring consistent characteristics across both time and multiple views. To ensure a compact representation, the Feature Bank is updated by a proposed dynamic merging mechanism. Leveraging this Feature Bank, we demonstrate for the first time that generating additional reference sequences through multiple autoregressive iterations can continuously improve generation performance. Experimental results show that FB-4D significantly outperforms existing methods in terms of rendering quality, spatial-temporal consistency, and robustness. It surpasses all multi-view generation tuning-free approaches by a large margin and achieves performance on par with training-based methods. Our code and data will be publicly available to support future research.