Unmanned aerial vehicles (UAVs) have become essential platforms for 3D object detection. In contrast to DETR-style 2D detectors, the Euclidean distance between 3D predictions and the ground truth does not serve as an accurate proxy for sampling error due to complex projection geometry. To overcome this limitation, we propose the perpendicular distance from a predicted point to the ray connecting the camera optical center and the ground-truth, which serves as a precise geometric measure of sampling difficulty. Although this ray distance provides an ideal sampling reference, the isotropic circularity resulting from equal ray distances constrains the approach's ability to capture object shape and orientation. We therefore introduce an anisotropy-aware elliptical mapping that reformulates the circular projection into an ellipse aligned with the object's geometry. By integrating the ellipse-corrected ray distances into both the bipartite matching cost and the regression loss, we develop RayDETR-a geometry-consistent, multi-view 3D detection framework. Experiments on the UAV3D dataset demonstrate that RayDETR achieves state-of-the-art performance without compromising inference speed.

Oriented objects in images are typically embedded in complex backgrounds and exhibit arbitrary orientations. When using oriented bounding boxes (OBBs) to represent these objects, the periodicity of the angles and associated variations in side lengths lead to discontinuities in the angle loss. This paper fundamentally addresses this problem by proposing a trigonometric loss function in the complex plane. Moreover, a conformer RPN head is designed with convolution and multi-head self-attention, which can dynamically capture angular and classification information. The proposed loss function and conformer RPN head jointly generate high-quality oriented proposals. A category-aware dynamic label assignment based on predicted category feedback is proposed to address the limitations of solely relying on IoU for oriented proposal label assignment. This method makes negative sample selection more representative, ensuring consistency between classification and regression features. Experiments were conducted on five realistic oriented detection datasets, and the results demonstrate superior performance in oriented object detection with minimal parameter tuning and time costs. Specifically, mean average precision (mAP) scores of 82.02%, 71.99%, 69.87%, 46.45%, and 98.77% were achieved on the DOTA-v1.0, DOTA-v1.5, DIOR-R, STAR, and HRSC2016 datasets, respectively.

Feature point detection based on convolutional neural network (CNN) has been studied widely. The effective approaches for improving detection accuracy are building a deeper network or using a multi‐network cascade structure. However, some potential capacity of CNN has not been excavated. In this study, the authors mainly analyse several factors influencing CNN performance from two aspects: (i) the position relationships between feature points and (ii) the normalisation methods of coordinates. Whether the network can learn the position relationships is also studied. For extracting the deep features of images, a network containing three convolution layers is constructed. The specific geometric relationship constraints are applied during calibration to maximise the capability of the CNN for learning the position relationship between feature points. Considering that different feature points only appear in various local regions of an image, local normalisation is proposed, which increases the mapping scope of the feature points and decreases the mapping error. The experimental results prove that the specific position relationship and local normalisation obviously improve the feature point detection based on CNN. At the detection error of 5%, the average detection accuracy of eyelid feature points is improved by 7.1% and single‐point detection receives a high accuracy of 97.96%.