Personnel and boat detection in unmanned aerial vehicles (UAVs) imagery plays a crucial role in open water search and rescue missions. The diverse perspectives and altitudes of UAV images often result in significant variations in the imagery’s appearance and dimensions of personnel and boats, and the false detections arising from water surface flares are acknowledged as a great challenge as well. The existing deep learning-based detection methods employ convolutional blocks with fixed kernel sizes to extract features from the imagery at a fixed spatial scale, which will lead to missed and false detections and severely affect detection accuracy when there are substantial differences in the appearance and size of the target objects. In this article, a spatial scale adaptive real-time object detection neural network, namely YoloOW, was proposed to tackle the challenge of personnel and boat detection amid the diverse UAV imagery, which comprises a feature extractor, a feature enhancer, and a postprocessor. The OaohRep convolutional block was proposed as a pivotal component in constructing the YoloOW and applied to the feature extractor and the feature enhancer. Compared with general convolution blocks, the OaohRep convolution block can extract image features across a wide range of spatial scales, show better scale adaptability, and achieve faster detection speed due to its unique merged convolution layer design. OaohRepBi-path aggregation network (PAN) was proposed in the feature enhancer, which imitated the architecture of the classic algorithm scale invariant feature transform (SIFT) and was successfully applied to deep learning models, showing better scale adaptability. A novel UAV detection box filter (UDBF) module was proposed in the postprocessor, which can effectively remove false detections caused by water surface flares. Experimental results demonstrate that our YoloOW model achieves 37.18% mean averaged precision (mAP) on the SeaDronesSee dataset, surpassing the baseline by 8.43%. This notable improvement positions our model at the first of the leaderboard. The code will be available at https://github.com/Xjh-UCAS/YoloOW .

Efficient extraction of building footprints from aerial and satellite imagery is essential for urban planning, infrastructure management, and large-scale geospatial analysis. Traditional raster-based approaches provide limited geometric precision, while existing polygon-generation methods often rely on detecting and ordering small-scale building vertices, which can lead to incomplete structures, distorted shapes, and high computational cost. To address these limitations, this study proposes an Edge-Attentive Dual-Branch Frame Field Network (EA-DBFFN) for automated and high-precision building polygon extraction. The method is built upon frame field learning and introduces a dual-branch architecture that separately predicts building masks and edges. A Dual-Task Decoder enlarges and adapts receptive fields while applying spatial attention to enhance the representation of structural details. Fixed Sobel and Laplacian filters are incorporated to strengthen boundary detection. In addition, a Dual-Task Mutual Guidance Module promotes the exchange of complementary information between the mask and edge branches, improving geometric consistency and reducing boundary errors. Experiments conducted on the Inria Aerial dataset and the CrowdAI dataset demonstrate that EA-DBFFN achieves superior performance in region-based metrics, with an AP75 of 72.9% on CrowdAI, representing a 2.3% improvement over competing methods. Furthermore, EA-DBFFN produces geometrically higher-quality polygons, with the Max Tangent Angle error reduced by 6.4%, the Invalid Polygon Ratio reduced by 66.3%, and Edge Smoothness improved by 72.7% compared to the best competing method. The results show that EA-DBFFN provides an effective and computationally efficient framework for generating high-quality vectorized building footprints suitable for large-scale urban analysis.

Accurate height estimation from single high-resolution satellite imagery remains a fundamental yet challenging problem in urban remote sensing. The primary difficulties stem from three aspects: insufficient cross-scale perception leads to blurred building boundaries; limited association between height interval modeling and global height semantics results in incomplete recovery of rooftop structural details; and uniform spatial modulation causes predictions in extreme-height regions to converge toward the mean, degrading local accuracy. To address these challenges, this paper proposes a novel monocular height estimation framework built upon a height-binned modeling paradigm and progressively optimized from global feature encoding to local spatial refinement. Specifically, a bidirectional multi-scale feature aggregation mechanism enhances cross-scale spatial perception; a semantic-guided height interval modeling module incorporates global height semantics into bin representations for more discriminative interval prediction; and a region-aware pixel-level modulation mechanism explicitly models height interval-spatial region relationships to suppress extreme-height prediction bias. Extensive experiments on DFC2023 and US3D benchmark datasets demonstrate consistently superior performance over all compared methods. Validation on GF-7 domestic satellite imagery further confirms the framework's generalization to distinct radiometric and imaging characteristics, demonstrating its practical value for large-scale domestic 3D urban mapping.

Local feature matching is a fundamental task in visual localization, simultaneous localization and mapping, and multi-view reconstruction. Recently, Mamba-based feature matchers have emerged as efficient alternatives to Transformer architectures due to their linear-complexity sequence modeling capability. However, existing methods uniformly propagate feature tokens using fixed scanning strategies, making them vulnerable to repetitive patterns, weak textures, and domain shifts.To address this limitation, we propose GamMa, a lightweight semi-dense feature matching framework based on a Gated Deformable Mamba architecture. The key insight is that coarse-level features contribute unequally to correspondence estimation. Instead of uniformly modeling all tokens, GamMa prioritizes informative matching features before sequence modeling. Specifically, a Gated Deformable Scanning mechanism is introduced to adaptively select reliable feature regions through deformable index learning, suppressing ambiguous and geometrically unstable responses. In addition, a Canny edge constraint is incorporated to improve feature localization robustness in multi-view and weak-texture scenarios.Extensive experiments demonstrate that GamMa achieves state-of-the-art performance among existing semi-dense matchers on MegaDepth, reaching 65.2/78.1/87.0 AUC at 5°/10°/20° while maintaining high efficiency. More importantly, GamMa significantly enhances the cross-domain generalization capability of Mamba-based matchers, outperforming JamMa by 3.7%/4.9%/3.1% on HPatches and 2.5%/2.8%/3.1% on ScanNet across all evaluation thresholds. The code will be available at https://github.com/Xjh-UCAS/GamMa.