Geometry-Aware Detection Transformer for Aerial Oriented Object Detection

Résumé fourni par la source

Oriented object detection in aerial images is a challenging task due to the complexity of object orientation and varying aspect ratios. While existing approaches have shown promise, they often overlook the impact of aspect ratio on angle prediction, which can negatively affect detection accuracy. Recently, the emerging transformer-based approaches have achieved efficient object detection due to their global modeling capability and the advantage of not requiring manually designed anchor boxes. In this paper, we propose a novel Geometry-Aware DEtection TRansformer framework, termed GA-DETR, which integrates a geometry-aware mechanism for oriented object detection. Specifically, we introduce a new angle classification method, called power-adjusted circular smooth label (PA-CSL), which smooths angle labels via a power-law function, thereby adapting to the target’s geometry. Then, we design a geometry-aware (GA) component that adapts based on the aspect ratio of the object, ensuring better feature alignment and further enhances the model’s feature extraction ability. This framework is further optimized with a denoising (DN) training strategy to accelerate convergence. Experiments on the DOTAv1.0 dataset show that our method demonstrates competitive performance in terms of detection accuracy, especially for objects with varying orientations and aspect ratios. The proposed approach provides a promising direction for improving oriented object detection in aerial images.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.