Insulator Detection via CNN for UAS Onboard Computers

This paper proposes the usage of single-stage CNN models for detecting insulators in aerial images and measures their applicability in low-power computing settings that often found in UAS onboard systems. In addition to methods in literature, we also design another network based on YOLOv2 modified with SPP (spatial pyramid pooling) block and CIoU loss as our baseline. Our results shows that while both using SPP block and optimizing the bounding box regression function increases the overall detection accuracy without significant cost, network architectures that is specifically designed for edge devices are much more suitable on said environments. One of such design is SF-YOLO, with computation cost of 3,842 BFLOP (29% lower than YOLOv3 tiny, 86% lower than ours) while retaining AP50 score higher than 0.9, and thus can be further used for autonomous navigation subsystems with proper edge devices.