Author(s): Shijun Pan; Yuki Yamada; Daichi Shimoe; Keisuke Yoshida

Linked Author(s): Shijun Pan, Keisuke Yoshida

Keywords: Nnotation Approach; Instance Segmentation; Riparian Waste Detection; UAV

Abstract: This paper presents a comprehensive analysis of the impact of distinct annotation approaches, namely bounding boxes (B-Box) and masks, on the accuracy of YOLOv8-seg, a state-of-the-art instance segmentation model, when applied to unmanned aerial vehicle (UAV) imagery. Our investigation centres on the UAV-BD dataset, a challenging and diverse dataset designed for UAV applications. Through meticulous comparisons, we assess the performance of YOLOv8-seg using mAP50 and mAP50-95 metrics under different annotation approaches, providing an understanding of the efficacy of each annotation method. Our study yields intriguing findings: while mAP50 scores exhibit comparability between bounding box and mask annotations, a distinctive divergence surfaces in the mAP50-95 metric. The mask annotation approach consistently outperforms bounding boxes, suggesting its dominance in effectively handling instances with higher IoU thresholds under all the backgrounds (i. e., sand, lawn, bush, land, step, ground and playground). The enhanced performance of mask-based annotations, particularly at higher IoU thresholds. This research provides valuable insights that can inform decision-making in digital twin applications and other domains reliant on accurate UAV-based object detection, fostering advancements in technology and application domains.

DOI: https://doi.org/10.3850/iahr-hic2483430201-307

Year: 2024