REAL-TIME ORIENTED SHIP DETECTION IN AERIAL IMAGES FOR EMBEDDED PLATFORMS

Ships in aerial images are slender objects and can face arbitrary directions, thus are more effectively described by oriented bounding boxes (OBBs) rather than horizontal bounding boxes (HBBs) in the aerial detection task. Existing oriented object detectors usually use two-stage anchor-based architecture, making them hard to be deployed to embedded platforms with neural processing units (NPUs). In this work, we propose a lightweight one-stage anchor-free oriented ship detector that uses center keypoints and edge offset vectors to capture the ship OBBs. Also, the detector is designed to be NPU hardware-friendly and can be efficiently processed on embedded NPUs. Experiments results show that our proposed detector can achieve real-time speed on embedded NPUs while maintaining accuracy similar or superior to existing methods. When deployed to VeriSilicon VIP8000OI embedded NPU, the proposed detector can achieve 88.1% AP on the HRSC2016 dataset at 27 FPS.