Oriented R-CNN and Beyond

Currently, two-stage oriented detectors are superior to single-stage competitors in accuracy, but the step of generating oriented proposals is still time-consuming, thus hindering the inference speed. This paper proposes an Oriented Region Proposal Network (Oriented RPN) to produce high-quality oriented proposals in a nearly cost-free manner. To this end, we present a novel representation manner of oriented objects, named midpoint offset representation, which avoids the complicated design of oriented proposal generation network. Built on Oriented RPN, we develop a simple yet effective oriented object detection framework, called Oriented R-CNN, which could accurately and efficiently detect oriented objects. Moreover, we extend Oriented R-CNN to the task of instance segmentation and realize a new proposal-based instance segmentation method, termed Oriented Mask R-CNN. Without bells and whistles, Oriented R-CNN achieves state-of-the-art accuracy on all seven commonly-used oriented object detection datasets. More importantly, our method has the fastest speed among all detectors. For instance segmentation, Oriented Mask R-CNN also achieves the top results on the large-scale aerial instance segmentation dataset, named iSAID. We hope our methods could serve as solid baselines for oriented object detection and instance segmentation. Code is available at https://github.com/jbwang1997/OBBDetection.