Point RCNN: An Angle-Free Framework for Rotated Object Detection

Rotated object detection in aerial images is still challenging due to arbitrary orientations, large scale and aspect ratio variations, extreme density of objects. Existing state-of-the-art rotated methods mainly rely on angle-based detectors. However, detectors can easily suffer from a long-standing boundary problem. To tackle this problem, we propose purely angle-free framework for detection, called Point RCNN. RCNN two-stage detector including both PointRPN PointReg which are angle-free. Given an input image, first, the backbone-FPN extracts hierarchical features, then, module generates accurate region interests (RRoIs) by converting learned representative points each using MinAreaRect function OpenCV. Motivated RepPoints, designed coarse-to-fine process regress refine more RRoIs. Next, based RRoIs PointRPN, learns corner RRoI perform detection. Finally, final bounding box be attained four points. In addition, often severely unbalanced categories, existing almost ignore dataset balanced strategy. We experimentally verified that re-sampling rare categories stabilize training procedure further improve performance. Specifically, performance was improved 80.37 mAP 80.71 DOTA-v1.0. Without unnecessary elaboration, our method achieved new multiple large-scale image datasets, DOTA-v1.0, DOTA-v1.5, HRSC2016, UCAS-AOD. better mAP. 79.31 mAP, significantly 2.86 (from ReDet’s 76.45 79.31). HRSC2016 UCAS-AOD, higher 90.53 90.04 respectively.