Deep learning has shown remarkable success in medical image classification; however, model performance remains highly sensitive to data imbalance, annotation uncertainty, and subtle inter-class variability. Conventional loss functions such as cross-entropy often fail to capture these challenges, resulting in overconfident or poorly calibrated predictions. To address this, we propose the Probabilistic Margin-Aware Focal Loss (PMAF Loss), a unified objective designed to enhance discriminability, robustness, and reliability in medical image classification. PMAF integrates three key mechanisms: focal modulation to emphasize hard and informative samples, uncertainty weighting to adaptively suppress noisy or ambiguous predictions, and margin-based feature regularization to enforce interclass separation in the embedding space. Extensive experiments on multiple benchmark medical datasets demonstrate that our method consistently outperforms standard crossentropy loss in terms of accuracy, calibration, and interpretability.
This paper reviews the Robust Thermal-Image Object Detection (RTIOD) Challenge, organized as part of the 6th Real-World Surveillance (RWS) Workshop at the Winter Conference on Applications of Computer Vision (WACV) 2026. The challenge aims to stimulate the development and rigorous evaluation of object detectors that remain reliable under long-term appearance drift in real-world surveillance settings. To this end, RTIOD is built on the Long-term Thermal Drift v2 dataset and evaluates submissions using both global and month-wise metrics (e.g., mAP@0.5), enabling direct assessment of temporal stability over extended periods. In total, 60 participants registered and 14 teams submitted predictions for the final test phase. Beyond reporting the leaderboard and summarizing competing approaches, this paper provides an analysis of why top methods succeed. We distill recurring design patterns (e.g., thermal representation choices, data curation strategies, and deployment-informed priors), characterize month-dependent failure modes that reveal seasonal drift effects, and examine dominant error sources (notably localization and classification) that limit performance. The resulting insights offer practical guidance for building more temporally consistent thermal detectors and establish reference baselines and benchmarks for future research in robust thermalimage object detection.
Text-to-image generation has demonstrated impressive performance in various applications, thanks to the powerful generative models. However, the unleashed creativity of such models can lead to the abuse of undesired generative content, which undermines the social impact and sustainability of generative AI. Machine unlearning is proposed to eliminate specific content from AI models and can be used to erase key value representations associated with undesired prompts. Yet, complete removal of a concept is non-trivial due to semantic complexity, leading to both adversarial recovery and image quality degradation. On this basis, this paper introduces a Robust Erasing Method in Diffusion models (REMinD) to unlearn target concepts flexibly, which can handle hypernym concepts using various hyponym prompts. To enhance unlearning and image quality, a multi-level loss function is designed to explore local, adjacent, and global representations, removing target unlearning concepts while preserving generation quality for the remaining global concepts. Extensive experiments on four unlearning tasks demonstrate that the proposed method achieves more effective and robust unlearning, e.g., reducing the attack success rate by up to 28% against state-of-the-art style attacks.
The detection of foreign object debris (FOD) on railway transmission lines and airport runways is essential for maintaining operational safety and minimizing risks. However, accurately identifying small, occluded, or cluttered objects remains a significant challenge for existing object detection models due to limitations in feature representation and high computational overhead. This work introduces YOLO-OSA, a lightweight object detector enhanced with an Optimized ShuffleAttention (OSA) mechanism integrated into both the backbone and neck. The OSA module selectively emphasizes informative spatial and channel features, enabling improved multi-scale feature aggregation while reducing computational complexity. The model achieves high detection accuracy with fewer parameters and lower FLOPs, making it suitable for deployment on resourceconstrained edge devices. Experiments on the RailFOD23 and FOD-A datasets show that YOLO-OSA outperforms state-of-the-art models, particularly in detecting small and occluded FOD instances. On the MS COCO benchmark, YOLO-OSA achieves a 29.2% improvement in mAP@50-95 over YOLOv8n and a 7.35% gain over YOLOv8s, while reducing FLOPs by 3.45% and 70.6%, respectively. These results highlight the effectiveness and generalization capability of YOLO-OSA for real-time, safety-critical object detection applications.
Text detection and localization in entertainment artwork-such as movie posters, streaming thumbnails, and album covers-pose unique challenges absent in documents or natural scenes. Typography is often highly stylized, integrated with complex backgrounds, and presented in multiple scripts. Conventional OCR systems degrade under these conditions, while general-purpose vision-language models (VLMs) lack the pixel-level precision needed for quality control and metadata validation. We propose a two-stage reason-then-ground framework that bridges this gap by combining high-level semantic reasoning with finegrained spatial grounding. In the first stage, a VLM identifies visible text, classifies its semantic role (title, extraneous text, brand logo, or background), and detects its language or script. In the second stage, we adapt Florence-2 for text-specific grounding through multilingual tokenizer expansion and LoRA-based fine-tuning, enabling conditional word-level localization across diverse scripts. Evaluations on internal Prime Video benchmarks and the public Movie-Posters dataset demonstrate substantial gains in geometric precision and multilingual robustness. Our model achieves F1 scores of 0.993 (IoU=0.5) and 0.911 (IoU=0.9) for English title text, and 0.985 (IoU=0.5) and 0.837 (IoU=0.9) for non-English scripts, outperforming both OCR-based and general-purpose VLM baselines. These results establish a practical and scalable solution for production-grade text grounding in visually complex, multilingual artwork.
4D reconstruction recovers time-varying 3D geometry, yet accurate depth estimation in moving regions remains challenging. Most existing methods-from monocular predictors to 4D Gaussian Splatting-rely on single-view pri-ors prone to scale-shift ambiguity and geometric distortion. While stereo estimation offers metric accuracy through triangulation, NeRF-based stereo synthesis often introduces ghosting and motion blur in dynamic areas, degrading performance. To address this, we propose a hybrid framework combining NeRF-generated stereo depth with a confidence-aware diffusion module to restore dynamic geometry. Our approach eliminates affine uncertainty and monocular dependence, achieving temporally coherent sequences. Experimental results show our method reduces absolute relative error by over 50% and improves strict accuracy $(\delta< 1.05)$ by 10%, setting a new benchmark for high-fidelity 4D reconstruction.
Relativistic heavy-ion collisions create a quark-gluon plasma whose properties are typically inferred via costly Bayesian analysis. We introduce JetBench, a quality-aware benchmark for multi-parameter classification of heavy-ion events using the ML-JET dataset. Each event is encoded as a $32 \times 32$ jet image with three targets: energy-loss module, strong coupling constant, $\alpha_{s}$, and virtuality separation scale, $Q_{0}$. Jet image quality is controlled by a moment-based aggregation level that averages different numbers of events into each image. We benchmark modern CNN, Transformer, and state-space models under a unified training protocol. Our best model, ViT-CoMer, achieves ~100% accuracy on energy loss, $\sim 95 \%$ on $\alpha_{s}$, and 78% on $Q_{0}$, with 74.5% joint accuracy. Beyond accuracy, we assess prediction quality via loss-weight ablations, calibration-style probability summaries, and structured near-neighbor confusions consistent with Quantum Chromodynamics (QCD) continuity. JetBench thus provides a scalable, physics-informed complement to Bayesian approaches and a compact testbed for image and prediction quality assessment.
Deep neural networks have revolutionised computer vision with their powerful capabilities. However, they remain vulnerable to adversarial attacks, which seek to find small perturbations to input images that mislead them into making incorrect predictions. Traditional attacks often rely on gradient information, which is typically unavailable to attackers in real-world scenarios. To address this, black-box adversarial attacks have been developed that do not depend on full access to the network. In this study, we focused on the most challenging black-box setting, decision-based attacks, where the target model only returns output labels. While previous approaches for this setting exist, they either perform inconsistently across different networks, or are resourceintensive concerning running time. To address this, we developed a variant of the Covariance Matrix Adaptation Evolutionary Strategy (CMA-ES) with a selective query approach that enables optimisation for 80% more generations with the same query budget. Furthermore, we utilised state-ofthe-art automated algorithm configuration techniques to optimise our attack and achieve a success rate of 100% using an initial search strategy. We compared our novel Decision-based Attack via CMA-ES (DACES) against various well-known attacks, demonstrating the best average performance in creating minimal adversarial perturbations across several network architectures for ImageNet and CIFAR-100. Notably, DACES is more than three times as fast with respect to running time than any other attack on ImageNet, making it particularly applicable to real-world settings.
This paper presents K-Vehicles, a new dataset for vehicle detection in remote sensing aerial imagery. It is built from high-resolution RGB images captured by a Cessna aircraft over diverse real-world environments, including highways, agricultural fields, and industrial zones. The dataset comprises 15,168 cropped images of 1,024×1,024 pixels, annotated manually across seven vehicle categories: truck, fork-lift, machinery, pickup, tractor, car, and bus. It incorporates relevant challenges such as occlusion, scene clutter, intrascene variation, and variable lighting conditions, making it suitable for training and evaluating detection models in realistic scenarios. To assess its utility, we conduct a benchmarking study using four recent You Only Look Once (YOLO) architectures, from YOLOv9 to YOLOv12. The models demonstrate strong performance, with mAP@50:95 reaching approximately 70%, confirming the learnability of the dataset, while also revealing persistent challenges in complex cases. Visual analyses support the numerical results and highlight the presence of fine-grained classification errors and detection failures under visual ambiguity. It is released with annotations in YOLO format and includes predefined train, validation, and test splits, and can be further extended for broader remote sensing applications. The dataset is available at: https://github.com/Leo-Thomas/K-Vehicles
Predicting pedestrian crossing intentions is crucial for enhancing road safety in autonomous driving and Advanced Driver Assistance Systems (ADAS), particularly in complex urban environments where decision-making is imperative. To this end, we introduce GATEPose (Graph Attention Transformer Enhanced by Pose), a novel lightweight model designed to predict pedestrian crossing intention. GATEPose integrates pose feature through a novel ST-$\boldsymbol{G A N}^{+}$ block, while bounding box and orientation angles streams are modeled in parallel using GRU Conv1D blocks. The outputs are then fused and passed to transformer encoders to capture spatiotemporal patterns from multimodal inputs. Integrating body angles yields a significant performance improvement compared to similar stateof-the-art models that rely only on pose and bounding boxes. Angle-based representation provides a more stable cue for capturing the subtle human motions that precede crossing intent. This architecture enables the model to learn the dynamic movement of pedestrians while maintaining low latency. For the first time, the model is trained and evaluated on the largest portions of the JAAD and PIE benchmarks that are used to predict pedestrian trajectories. Two new datasets are proposed, JAAD ${}_{p}$ ose and PIE ${}_{p}$ ose. $J A A D_{p}$ ose contains $\mathbf{2 5 K}$ sequences derived from $\boldsymbol{J} \boldsymbol{A} \boldsymbol{A} \boldsymbol{D}_{\text{all}}$, while $P I E_{p}$ ose includes more than 72 K sequences from PIE. Both datasets provide high-quality pose annotations extracted using ViTPose, along with orientation angles computed using our method. GATEPose achieves stateof-the-art performance across multiple evaluation metrics while maintaining significantly lower inference latency than recent methods.
Optical Character Recognition (OCR) pipelines and vision-language models exhibit high accuracy on standard OCR and document-understanding benchmarks, yet they systematically underperform on document patterns that are operationally critical in financial workflows. These patterns include vertical or rotated text, logo-embedded vendor names, fine-print regulatory clauses, degraded scans, and complex multi-column layouts; while underrepresented in public datasets, they constitute a substantial portion of real-world failure cases. We introduce SAVIOR, a sampleefficient data curation methodology that targets such highimpact failure scenarios to adapt vision-language models for robust financial OCR. We further propose PaIRS, a structure-aware evaluation metric that measures layout fidelity by comparing pairwise spatial relationships between tokens. Fine-tuning Qwen2.5-VL-Instruct with SAVIOR-Train achieves a word recall of 0.93, and when the resulting layout-aware OCR output is used as context for document question answering, it attains an F1 score of 0.89, outperforming both open and closed-source models such as GPT-4o, Mistral-OCR, PaddleOCR-VL and DeepSeek-OCR.
Developing computer vision models for underwater environments is challenging due to the scarcity of high-quality annotated data, particularly for degraded conditions such as severe underexposure. In this work, we propose a progressive framework for synthetically generating realistic underexposed data to bridge this domain gap. We first introduce a rigorous multi-metric curation strategy to establish a ground-truth reference dataset of real underexposed images. We then propose three degradation modeling approaches, varying in complexity: RGB-Based Global Gamma Matching (GGM), Decorrelated Luminance Matching (DLM), and Perceptual Deep Feature Optimization (PDFO). The first two proposed methods are based on histogram alignment in different color spaces, while Our final proposed method, PDFO, leverages a pre-trained deep network to optimize for perceptual similarity. Qualitative and quantitative evaluations demonstrate that PDFO outperforms baseline methods generating the closest luminance distribution to real-world data. This framework provides a robust tool for augmenting underwater datasets, enabling more reliable object detection in low-light conditions.
Machine Learning (ML) holds significant potential for enhancing Remote Sensing (RS) data analysis. Today, many ML models have been introduced to process RS data, including a vast amount of Geospatial Foundation Models (GeoFM). However, selecting an effective (i.e., best performing) model is both time-consuming and technically challenging, especially for domain scientists without ML experience. Therefore, this work introduces a model recommendation method for a particular downstream task by using a lightweight approach. Our method is composed of a two-phase mechanism that leverages the similarity of feature vectors from provenance data within clustered representations. Thus, the recommendation is based on the accuracy and computational resources of each model. This allows us to yield a model recommendation and later selection for each sample point instead of the whole dataset. Importantly, high resource and energy gains are observed during the second phase of our method, leading to an efficient way to select ML-based solutions in RS applications.
Illegal waste dumping poses significant environmental and public health challenges worldwide, requiring automated surveillance systems for detection and prevention. This paper presents our solution for the IWDD 2026 Contest, addressing the dual challenge of detecting illegal dumping events in surveillance videos and localizing the exact moment of occurrence. We employ X3D-M, an efficient 3D convolutional network pretrained on Kinetics-400, combined with a sliding window inference strategy for temporal localization. Through systematic hyperparameter optimization across 96 configurations and ablation studies examining nine combinations of fine-tuning strategies and loss functions, we identify key design choices for this application domain. Our experiments reveal that differential learning rates-applying lower rates to the pretrained backbone while training the classifier more aggressively-outperform both frozen backbones and uniform fine-tuning. The optimal system achieves an F1-score of 0.8387 and a temporal F1score of 0.7742 on our test set, with 92.3% of correct detections within the temporal tolerance window. Operating at over 8 times real-time speed with only 2.97M parameters, our approach demonstrates that efficient video classification architectures can be effectively adapted for specialized surveillance applications through careful transfer learning and inference design.
Forms are a critical component of every day life for millions, whether filing for benefits, immigrating, taxes or healthcare. As such, building robust real-world benchmarks for forms and domain-specific model families is a critical bottleneck in integrating with legacy systems. However, no such dataset exists and existing vision-language models (VLMs) trained for document understanding are not well equipped to understand real-world faxes and forms. This paper presents (1) SynthForm-3k: the first dataset comprising synthetically perturbed real-world forms and (2) SynthForm-VL: a family of 2B, 4B and 8B VLM models fine-tuned for realistic form understanding. SynthForm-3k unifies 6 categories, across tax, immigration, finance, healthcare, dental and insurance, yielding 3417 samples. We furthermore evaluate our new model against open-weight VLMs, many built specifically for document under-standing tasks, finding all 3 of our family models outperforms their baselines. On the SynthForm-3k test set, we find our 2B model performs at 66.0% (+5.8 △ANLS from baseline), $4 B$ performs at 69.6% (+9.3 $\triangle A N L S$ from baseline) and $8 B$ performs $70.5 \%(+10.3 \Delta A N L S$ from baseline). In releasing this model & dataset, our goal is to push state-of-the-art forward for realistic real-world forms, encouraging integrations with legacy technologies.
Realtime clinical video analysis on mobile and edge devices demands models that are computationally efficient and reliably interpretable. Vision Transformers (ViTs) incur significant additional computational overhead due to the cost of attention computation and often rely on post-hoc explainability methods, which together limit their deployment in resource-constrained environments. The proposed Lightweight ViT with class activation mapping (LiteViT-CAM) addresses this challenge through a lightweight hybrid design that couples convolutional layers for local representation learning with vision transformer blocks for global context modeling. A dual-head architecture supports image level prediction and intrinsic explainability through native high resolution heatmaps generated during inference. A unified hybrid loss jointly optimizes recognition accuracy and interpretability, while a temporal consistency module stabilizes predictions and explanations across sequential frames. Quantitative evaluations across multiple endoscopic datasets show that LiteViT-CAM achieves the highest DSC, mIoU, and accuracy while maintaining realtime performance (14-15 FPS), and that complementary NanoNet variants deliver competitive segmentation quality with minimal parameters and the fastest inference rates. The framework is also deployed in a mobile application, demonstrating the feasibility of point-of-care use. Although validated on wireless capsule endoscopy videos, the method is generalized to more general medical and edge-based video understanding tasks.
Recent advances in computer vision and human pose estimation, together with the increasing accessibility of drone technology, have opened new opportunities for datadriven vision-based performance analysis in sports. However, their applications remain underexplored in cross-country (XC) skiing, where research has predominantly relied on wearable sensors in controlled environments. These approaches are difficult to generalize to field environments and mostly inaccessible to the general public, while video analysis models and unmanned aerial vehicles are rapidly improving. Hence, publicly available XC skiing datasets can enable novel and accessible technique feedback and insights for both professional and recreational athletes. In this work, we introduce the Nordic Skiing Dataset (NSD)1https://www.nordicskiingdataset.se/, a publicly available drone-based video data set designed to support pose estimation, gear classification, and feedback generation for skating-style cross-country skiing. The dataset features 149 annotated videos of 12 skiers across different skill levels, from beginners to elite national athletes. The dataset captures various real-world skiing conditions and includes annotated segments for gear classification and technique evaluation. Pose annotations were generated using AlphaPose with the Halpe26 keypoint format and refined. Developed in close collaboration with domain experts from the Swedish national skiing team, NSD was carefully designed to support robust and generalizable machine learning research. Although we present initial benchmark results for gear classification, our primary contribution is the dataset itself, an open resource designed to foster further research in athlete modeling, transfer learning, performance feedback, and sports biomechanics.
Face biometrics are widely used in security-based access control applications. However, Face Recognition Systems (FRS) remain vulnerable to Presentation Attack Instruments (PAIs), making Presentation Attack Detection (PAD) essential for biometric security. While PAD techniques have advanced across various conditions (e.g., smartphones, laptops, and surveillance), adapting to new operational scenarios remains challenging. This work introduces the first presentation attack scenario in an on-the-move setting, where individuals are verified while walking toward a camera without stopping. We present the Moving Mask Presentation Attack (MMPA) dataset, comprising 15 bona fide subjects and four custom silicone face masks with different types of disguise attacks, captured from a six-meter distance. To increase realism, accessories such as eyeglasses, wigs, caps, and hoodies are incorporated. Data were collected using five different cameras, resulting in 1,415,662 bona fide and 1,882,502 attack samples. We benchmark six PAD techniques on MMPA, highlighting the difficulty of detecting on-the-move attacks and identifying key challenges for future research. As a preliminary study, this work aims to establish a challenging benchmark and highlight open issues for future research on presentation attack detection in dynamic, on-the-move scenarios.
The widespread use of face data in computer vision raises significant concerns about privacy and identity leakage. Conventional anonymization techniques, such as blurring or masking, often degrade facial structure and expression, while existing generative methods may still retain identifiable cues when evaluated against strong face recognition systems. To address these limitations, we propose StructFormer, a structure-consistent face de-identification framework based on a Transformer-GAN generator. StructFormer adopts a dual-stream design in which facial landmarks and masks provide explicit structural priors that are fused with appearance features through a Structure-Aware Attention Fusion module. This enables the preservation of head pose, facial layout, and expression while modifying identity-related appearance. A privacy control coefficient further allows continuous adjustment of anonymization strength without architectural changes. Experiments on LFW, CelebA, and CelebA-HQ demonstrate that StructFormer achieves a favorable balance between visual fidelity and privacy protection, maintaining high face detection rates (approximately 99%-100%) and competitive FID scores, while substantially reducing re-identification performance under strong FaceNet and ArcFace attackers, with match rates as low as 0.03% on CelebA-HQ.
