Aerial Object Detection in Complex Scenes Based on Mask Guidance*

Object detection in aerial imagery presents distinct challenges compared to general object detection, notably false alarm interference. While general object detection methods primarily address scale variation in natural scenes, they often neglect two critical issues prevalent in large-area earth observation scenes. In this letter, we introduce a two-stage object detection approach, denoted as MG2ADet (Mask Guidance and Additive Angular Margin Loss). MG2ADet features a streamlined design yet achieves competitive performance. To enhance object feature discrimination and mitigate background noise interference, we incorporate a mask guidance module into the feature pyramid. This module directs network features towards the foreground area using pseudo mask labels. Additionally, the additive angular margin loss normalizes the original classification loss to cosine loss and introduces an angular margin penalty in the cosine angle space. This helps minimize intra-class distance and augment inter-class distance. Rigorous experiments conducted on public remote sensing datasets underscore the superior detection performance of our algorithm compared to numerous existing detectors, encompassing both one-stage and two-stage detectors.