WACV 2026 Papers

Semantic Scene Completion (SSC) aims to predict the semantic occupancy of each voxel within a 3D scene using sensor data, a critical task for autonomous driving and robotics. Despite recent progress, camera-based SSC remains challenging due to various difficulties, including voxel class imbalance, occlusion, and depth ambiguity. This paper introduces FairScene, a novel approach that learns class-disentangled 2D/3D representations to improve SSC. By ensuring balanced representations across classes, FairScene mitigates the dominance of majority classes and promotes fairer voxel categorization. Additionally, FairScene explicitly models spatial dependencies between different classes through a novel inter-class occupancy reasoning mechanism. Such explicit modeling helps alleviate occlusion and depth ambiguities in SSC. To address the scarcity of SSC training data, we propose OccMix, a novel augmentation strategy that generalizes MixUp from 2D to 2.5D and 3D metric spaces while maintaining geometric consistency. Extensive quantitative and qualitative experiments demonstrate that FairScene outperforms prior methods on both the SemanticKITTI and SSCBench-KITTI-360 benchmarks. The code is available at https://github.com/DianJJ/FairScene.

This paper aims to bring fine-grained expression control while maintaining high-fidelity identity in portrait generation. This is challenging due to the mutual interference between expression and identity. On one hand, fine expression control signals inevitably introduce appearance-related semantics (e.g., facial contours, and ratio), which impact the identity of the generated portrait. On the other hand, even coarse-grained expression control can cause facial changes that compromise identity, since they all act on the face. Here, we introduce EmojiDiff, the first end-to-end solution that enables simultaneous control of extremely detailed expression (RGB-level) and high-fidelity identity in portrait generation. To address the above challenges, EmojiDiff adopts a two-stage scheme involving decoupled training and fine-tuning. For decoupled training, we innovate ID-irrelevant Data Iteration (IDI) to synthesize high-quality cross-identity expression pairs by separating and optimizing the processes of maintaining expression and altering identity. Training the model with this data, we effectively disentangle fine expression features in the expression template from other extraneous information (e.g., identity, skin). Subsequently, we present ID-enhanced Contrast Aignment (ICA) for further fine-tuning. ICA achieves rapid reconstruction and joint supervision of identity and expression information, thus aligning identity representations of images with and without expression control. Experimental results demonstrate that our method significantly outperforms its counterparts, achieving precise expression control with highly maintained identity, and generalizing well to various diffusion models. Project page: https://emojidiff.github.io.

The trend in sign language generation is centered around data-driven generative methods that require vast amounts of precise 2D and 3D human pose data to achieve an acceptable generation quality. However, currently, most sign language datasets are video-based and limited to automatically reconstructed 2D human poses (i.e., keypoints) and lack accurate 3D information. Furthermore, existing state-of-the-art for automatic 3D human pose estimation from sign language videos is prone to self-occlusion, noise, and motion blur effects, resulting in poor reconstruction quality. In response to this, we introduce DexAvatar, a novel framework to reconstruct bio-mechanically accurate fine-grained hand articulations and body movements from in-the-wild monocular sign language videos, guided by learned 3D hand and body priors. DexAvatar achieves strong performance in the SGNify motion capture dataset, the only benchmark available for this task, reaching an improvement of 35.11% in the estimation of body and hand poses compared to the state-of-the-art. The official website of this work is: https://github.com/kaustesseract/DexAvatar.

Compact models can be effectively trained via Knowledge Distillation (KD), where a lightweight student model learns to replicate the behavior of a larger, high-performing teacher. A persistent challenge in KD lies in the misalignment between the representational spaces of teacher and student networks, especially when they differ in architecture or capacity. To address this, we propose Subspace-Driven Knowledge Distillation (SDMD), a novel framework that mitigates representational disparity by projecting features into an indefinite inner product space. This relaxation from traditional Hilbert spaces enables more flexible geometric alignment, capturing transformations such as rotations and reflections that are often necessary for accurate knowledge transfer. By learning a subspace that bridges the semantic gap between teacher and student, SDMD facilitates more effective distillation without increasing model complexity. We validate SDMD through extensive experiments on large-scale image classification (ImageNet-1K) and object detection (COCO), where it consistently outperforms existing distillation methods. Notably, SDMD-trained models not only achieve state-of-the-art results in distilled settings but also surpass the performance of equivalent models trained from scratch, highlighting the strength of our subspace-based alignment strategy.

Accurate and low-latency 3D object detection is essential for autonomous driving, where safety hinges on both rapid response and reliable perception. While rotating LiDAR sensors are widely adopted for their robustness and fidelity, current detectors face a trade-off: streaming methods process partial polar sectors on the fly for fast updates but suffer from limited visibility, cross-sector dependencies, and distortions from retrofitted Cartesian designs, whereas full-scan methods achieve higher accuracy but are bottlenecked by the inherent latency of a LiDAR revolution. We propose Polar-Fast-Cartesian-Full (PFCF), a hybrid detector that combines fast polar processing for intra-sector feature extraction with accurate Cartesian reasoning for full-scene understanding. Central to PFCF is a custom Mamba SSM-based streaming backbone with dimensionally-decomposed convolutions that avoids distortion-heavy planes, enabling parameter-efficient, translation-invariant, and distortion-robust polar representation learning. Local sector features are extracted via this backbone, then accumulated into a sector feature buffer to enable efficient inter-sector communication through a full-scan backbone. PFCF establishes a new Pareto frontier on the Waymo Open dataset, surpassing prior streaming baselines by 10% mAP and matching full-scan accuracy at twice the update rate.

Once-for-All (OFA) training enables a single super-net to generate multiple sub-nets tailored to diverse deployment scenarios, supporting flexible trade-offs among accuracy, robustness, and model-size without retraining. However, as the number of supported sub-nets increases, excessive parameter sharing in the backbone limits representational capacity, leading to degraded calibration and reduced overall performance. To address this, we propose SOLAR (Switchable Output Layer for Accuracy and Robustness in Once-for-All Training), a simple yet effective technique that assigns each sub-net a separate classification head. By decoupling the logit learning process across sub-nets, the Switchable Output Layer (SOL) reduces representational interference and improves optimization, without altering the shared backbone. We evaluate SOLAR on five datasets (SVHN, CIFAR-10, STL-10, CIFAR-100, and TinyImageNet) using four super-net backbones (ResNet-34, WideResNet-16-8, WideResNet-40-2, and MobileNetV2) for two OFA training frameworks (OATS and SNNs). Experiments show that SOLAR outperforms the baseline methods: compared to OATS, it improves accuracy of sub-nets up to 1.26 %, 4.71 %, 1.67 %, and 1.76 %, and robustness up to 9.01 %, 7.71 %, 2.72 %, and 1.26 % on SVHN, CIFAR-10, STL-10, and CIFAR-100, respectively. Compared to SNNs, it improves TinyImageNet accuracy by up to 2.93 %, 2.34 %, and 1.35 % using ResNet-34, WideResNet-16-8, and MobileNetV2 backbones (with 8 sub-nets), respectively. The code of SOLAR is publicly available at: https://github.com/NAIL-UH/SOLAR and its website can be accessed at https://saktx.github.io/solar.github.io/.

Driven by advancements in deep learning, computer-aided diagnoses have made remarkable progress. However, outside controlled laboratory settings, algorithms may encounter several challenges. In the medical domain, these difficulties often stem from limited data availability due to ethical and legal restrictions, as well as the high cost and time required for expert annotations--especially in the face of emerging or rare diseases. In this context, open-set recognition plays a vital role by identifying whether a sample belongs to one of the known classes seen during training or should be rejected as an unknown. Recent studies have shown that features learned in the later stages of deep neural networks are observed to cluster around their class means, which themselves are arranged as individual vertices of a regular simplex [32]. The proposed method introduces a loss function designed to reject samples of unknown classes effectively by penalizing open space regions using auxiliary datasets. This approach achieves significant performance gain across four MedMNIST datasets--BloodMNIST, OCTMNIST, DermaMNIST, TissueMNIST and a publicly available skin dataset[29] outperforming state-of-the-art techniques.

Large vision-language models (LVLMs) achieve impressive performance on multimodal tasks but often suffer from hallucination and confidently describe objects or attributes not present in the image. Current training-free interventions struggle to maintain accuracy in open-ended and long-form generation scenarios. We introduce the Confidence-Aware Attention Calibration (CAAC) framework to address this challenge by targeting two key biases: spatial perception bias, which distributes attention disproportionately across image tokens, and modality bias, which shifts focus from visual to textual inputs over time. CAAC employs a two-step approach: Visual-Token Calibration (VTC) to balance attention across visual tokens, and Adaptive Attention Re-Scaling (AAR) to reinforce visual grounding guided by the model's confidence. This confidence-driven adjustment ensures consistent visual alignment during generation. Experiments on CHAIR, AMBER, and POPE benchmarks demonstrate that CAAC outperforms baselines, particularly in long-form generations, effectively reducing hallucination.

Continuous eye tracking is critical for applications in human-computer interaction, including biometric authentication, gaze-based systems, and affective-cognitive modeling. Recent interest in neuromorphic event cameras has grown due to their sub-microsecond latency in capturing eye movement dynamics. However, existing event-based eye-tracking methods face challenges such as limited labels, sub-optimal event accumulation, and a lack of frameworks that fail to capture fine-grained temporal relationships within event volumes. To address these, we propose a directed acyclic graph-based semi-supervised approach with a framework that is adaptable across multiple closely related tasks, including gaze tracking, pupil tracking, and eye-based emotion recognition. Our approach enables efficient spatiotemporal learning with 95.5% parameter reduction compared to existing methods, achieving significant performance improvements: 38.75% improvement in pupil tracking accuracy, 68% and 63% reductions in gaze angle error on EV-Eye and EBV-Eye datasets respectively, and 3.3% improvement in emotion recognition across all evaluated datasets. Project page is at https://eye-tracking-for-physiological- sensing.github.io/SaccadeX/.

In semi-supervised semantic segmentation, the Mean Teacher- and co-training-based approaches are employed to mitigate confirmation bias and coupling problems. However, despite their high performance, these approaches frequently involve complex training pipelines and a substantial computational burden, limiting the scalability and compatibility of these methods. In this paper, we propose a PrevMatch framework that effectively mitigates the aforementioned limitations by maximizing the utilization of the temporal knowledge obtained during the training process. The PrevMatch framework relies on two core strategies: (1) we reconsider the use of temporal knowledge and thus directly utilize previous models obtained during training to generate additional pseudo-label guidance, referred to as previous guidance. (2) we design a highly randomized ensemble strategy to maximize the effectiveness of the previous guidance. PrevMatch, a simple yet effective plug-in method, can be seamlessly integrated into existing semi-supervised learning frameworks with minimal computational overhead. Experimental results on three benchmark semantic segmentation datasets show that incorporating PrevMatch into existing methods significantly improves their performance. Furthermore, our analysis indicates that PrevMatch facilitates stable optimization during training, resulting in improved generalization performance.

Adversarial patch attacks critically endanger object detection systems by causing severe mispredictions with small, easily realizable perturbations in both digital and physical environments. Existing defenses such as certified methods or patch detection suffer from high latency, while conventional adversarial training often overfits to specific patches and lacks generalization, particularly in multi-object scenarios. To overcome high latency and poor generalization, we introduce TRACE (Tuning Robustness by Adversarial-patch Confounder Elimination), an adversarial fine-tuning framework that leverages Instrumental Variable Regression in the feature space. TRACE treats patch-related variations--including location, rotation, and brightness--as confounders, thereby eliminating spurious correlations and guiding the model toward causal features that sustain robust detection. Evaluations on YOLOv5 and YOLOv8 show that TRACE consistently outperforms conventional defense methods in both efficiency and robustness under adaptive and unseen patch attacks. Moreover, physical testbed experiments confirm its effectiveness beyond digital settings, highlighting TRACE as a practical solution for achieving generalized robustness in object detection.

Despite the strong generalization capabilities of pre-trained vision-language models (VLMs) like CLIP, adapting these models for few-shot generalization tasks still presents a fundamental challenge. This challenge, often termed the discrimination-generalization dilemma, highlights the need to fine-tune task-specific knowledge while simultaneously preserving the model's general, pre-trained knowledge. Prompt learning offers a partial solution but often struggles to capture rich visual-textual interactions. Adapter-based methods like MMA improve alignment by adding learnable modules, but their use of multiple independent adapters increases parameter overhead and can limit transferability due to naive fusion of adapted and frozen features.To address these limitations, we introduce Recurrent Multi-Modal Adapter (R-MMA), a lightweight and efficient extension of MMA that enhances both performance and generalization. R-MMA employs a recurrent adapter module with shared weights across multiple layers of the image and text encoders. This design substantially reduces the parameter count while maintaining high expressive capacity. Additionally, R-MMA integrates an attention-based alignment mechanism to harmonize the adapter outputs with the frozen encoder features before fusion. This ensures better preservation of the pre-trained representations and enhances cross-modal consistency. Extensive experiments across 15 datasets on diverse tasks, including few-shot learning, generalization to novel classes and domains, and dataset transfer scenarios, demonstrate that R-MMA consistently surpasses state-of-the-art baselines, achieving strong performance with improved efficiency and a better balance between adaptation and generalization. Our work achieves one of the highest forms of parameter efficiency with only three trainable weight matrices for the whole network, regardless of the network depth. Our code is available at: https://github.com/farhanishmam/R-MMA.

The effectiveness of Multimodal Large Language Models (MLLMs) demonstrates a profound capability in multimodal understanding. However, the simultaneous generation of images with coherent texts is still underdeveloped. Addressing this, we introduce a novel interleaved vision-and-language generation method, centered around the concept of "generative vokens". These vokens serve as pivotal elements contributing to coherent image-text outputs. Our method is marked by a unique two-stage training strategy for description-free multimodal generation, which does not necessitate extensive descriptions of images. We integrate classifier-free guidance to enhance the alignment of generated images and texts, ensuring more seamless and contextually relevant multimodal interactions. Our model, MiniGPT-5, exhibits substantial improvement over the baseline models on multimodal generation datasets, including MMDialog and VIST. The human evaluation shows MiniGPT-5 is better than the baseline model on more than 56% of cases for multimodal generation, highlighting its efficacy across diverse benchmarks.

Due to their increased field of view, wide-angle lenses are increasingly used in applications such as VR, security, or in autonomous driving. Typically, existing models either ignore wide-angle distortions or "undistort" images to a perspective projection, often resulting in severe stretching. More recent distortion-aware architectures address these issues, yet they impose substantial computational burdens and limit the use of powerful pre-trained vision backbones. In this work, we revisit the undistortion strategy by exploring alternative projection functions beyond the conventional perspective model.Specifically, we investigate square-to-disc mapping functions, most notably, the elliptical grid map (EGM) projection, which minimizes stretching. We show how EGM projection can be combined with the known lens distortion curves to achieve distortion invariance directly in image space. This network-agnostic approach seamlessly integrates with existing deep learning architectures, allowing fine-tuning on pre-trained models trained on large perspective datasets, while adapting to both seen and unseen wide-angle lenses without re-training each time the lens changes during evaluation. We perform experiments on the semantic segmentation task, comparing methods on zero-shot adaptation to unseen lenses from different wide-angle lenses. Our extensive experiments show that using the EGM projection with existing segmentation models significantly outperforms baselines when trained on bounded distortion levels and tested across both seen and out-of-distribution distortions. Furthermore, EGM projection achieves improved performance on real-world datasets, highlighting the robustness and practicality of our approach in real-world applications.

Satellite missions provide valuable optical data for monitoring rivers at diverse spatial and temporal scales. However, accessibility remains a challenge: high-resolution imagery is ideal for fine-grained monitoring but is typically scarce and expensive compared to low-resolution imagery. To address this gap, we introduce SuperRivolution, a framework that improves river segmentation resolution by leveraging information from time series of low-resolution satellite images. We contribute a new benchmark dataset of 9,810 low-resolution temporal images paired with high-resolution labels from an existing river monitoring dataset. Using this benchmark, we investigate multiple strategies for river segmentation, including ensembling single-image models, applying image super-resolution, and developing end-to-end models trained on temporal sequences. SuperRivolution significantly outperforms single-image methods and baseline temporal approaches, narrowing the gap with supervised high-resolution models. For example, the F1 score for river segmentation improves from 60.9% to 80.5%, while the state-of-the-art model operating on high-resolution images achieves 94.1%. Similar improvements are also observed in river width estimation tasks. Our results highlight the potential of publicly available low-resolution satellite archives for fine-scale river monitoring.

Anomaly segmentation is a critical capability for safety-critical robotics applications that must be aware of unexpected events. Normalizing flows (NFs), a class of generative models, are a promising approach for this task due to their ability to model the inlier data distribution efficiently. However, their performance falters in dynamic scenes, where complex, multi-modal data distributions cause them to struggle with out-of-distribution samples, leaving a performance gap to leading discriminative methods.To address this limitation, we introduce FlowCLAS, a hybrid framework that enhances the traditional maximum likelihood objective of NFs with a discriminative, contrastive loss. Leveraging Outlier Exposure, this objective explicitly enforces a separation between normal and anomalous features in the latent space, retaining the probabilistic foundation of NFs while embedding the discriminative power they lack.The strength of this approach is demonstrated by FlowCLAS establishing new state-of-the-art (SOTA) performance across multiple challenging anomaly segmentation benchmarks for robotics, including Fishyscapes Lost & Found, Road Anomaly, SegmentMeIfYouCan-ObstacleTrack, and ALLO. Our experiments also show that this contrastive approach is more effective than other outlier-based training strategies for NFs, successfully bridging the performance gap to leading discriminative methods.

Sports field calibration is critical for mapping image coordinates to standardized coordinates, enabling precise analysis of player trajectories and tactical formations. However, traditional methods designed for TV broadcast footage rely on sparse field features (corners, lines) that are susceptible to occlusion and viewpoint variations, limiting their effectiveness for tactical camera calibration. To address these limitations, we propose a novel pose-based calibration framework that directly regresses the 6-DoF camera pose from tactical view images. The proposed framework consists of three novel components: (1) a Lanczos-based spatial encoding module that preserves fine-grained geometric structures in the field representation, (2) an offset-based subpixel localization strategy that enhances occlusion robustness by refining keypoints' position to sub-pixel accuracy, and (3) a query-driven pose regression head with self-attention mechanisms that directly estimates camera pose without requiring additional calibration metadata. Extensive experiments on the SoccerNet-2023 and World Cup 2014 benchmarks demonstrate that our method achieves state-of-the-art performance in terms of Jaccard Index at threshold t = 5 (JaC@t), establishing superior accuracy and cross-dataset generalization capabilities.

We propose KFS-Bench, the first benchmark for key frame sampling in long video question answering (QA), featuring multi-scene annotations to enable direct and robust evaluation of sampling strategies. Key frame sampling is crucial for efficient long-form video understanding. In long video QA, selecting informative frames enables multimodal large language models (MLLMs) to improve both accuracy and efficiency. KFS-Bench addresses the limitation of prior works that only indirectly assess frame selection quality via QA accuracy. By providing ground-truth annotations of multiple disjoint scenes required per question, KFS-Bench allows us to directly analyze how different sampling approaches capture essential content across an entire long video. Using KFS-Bench, we conduct a comprehensive study of key frame sampling methods and identify that not only sampling precision but also scene coverage and sampling balance are the key factors influencing QA performance. Regarding all the factors, we design a novel sampling quality metric that correlates with QA accuracy. Furthermore, we develop a novel key frame sampling method that leverages question-video relevance to balance sampling diversity against question-frame similarity, thereby improving coverage of relevant scenes. Our adaptively balanced sampling approach achieves superior performance in both key frame sampling and QA performance. The benchmark is available at https://github.com/NEC-VID/KFS-Bench.

Distilled diffusion models generate images in far fewer timesteps but suffer from reduced sample diversity when generating multiple outputs from the same prompt. To understand this phenomenon, we first investigate whether distillation damages concept representations by examining if the required diversity is properly learned. Surprisingly, distilled models retain the base model's representational structure: control mechanisms like Concept Sliders and LoRAs transfer seamlessly without retraining, and SliderSpace analysis reveals distilled models possess variational directions needed for diversity yet fail to activate them. This redirects our investigation to understanding how the generation dynamics differ between base and distilled models. Using x0 trajectory visualization, we discover distilled models commit to their final image structure almost immediately at the first timestep, while base models distribute structural decisions across many steps. To test whether this first-step commitment causes the diversity loss, we introduce diversity distillation, a hybrid approach using the base model for only the first critical timestep before switching to the distilled model. This single intervention restores sample diversity while maintaining computational efficiency. We provide both causal validation and theoretical support showing why the very first timestep concentrates the diversity bottleneck in distilled models. Our code and data are available at https://distillation.baulab.info/