YOLOv8 Oriented Bounding Box (OBB) Model for Waymo Open Dataset

Atanasko Boris Mitrev, Georgina Mirceva

Abstract: This paper explores the new YOLOv8 oriented bounding boxes object detection capabilities in Bird’s Eye View (BEV) images using Waymo Open Dataset. The Waymo Open Dataset provides high-quality real-world driving data, making it ideal dataset for training object detection models. The motivation behind this research arise from the possibility to explore oriented bounding boxes object detection capabilities of the YOLOv8 model for accurate and robust object detection in autonomous driving applications. Traditional bounding box methods often struggle with objects that have complex orientations, such as vehicles and pedestrians, particularly in dynamic and cluttered environments. By leveraging BEV images, which offer a top-down view of the scene, we analyse bounding boxes object detection for objects with complex orientations, such as vehicles and pedestrians. We trained a model for generating oriented bounding boxes in the BEV domain and demonstrated its effectiveness in improving object detection performance. The experimental results from this research gave very good results in terms of detection accuracy and robustness, for objects with non-axis-aligned orientations. This research contributes with its exploration of practical use of object detection techniques using Waymo Open Dataset, used for autonomous driving applications research.

External IDs:dblp:conf/ictinnovations/MitrevM24