Vol. 8, Issue 1 (2019)

This research paper explores the performance of YOLOv3 on various object detection tasks. First, we give a summary of the YOLOv3 architecture and its main parts, such as the detection head, feature extraction layers, and Darknet-53 backbone. Next, we assess YOLOv3's performance on the COCO dataset, a popular object detection benchmark. Based on our findings, YOLOv3 achieves 57.9 percent mAP (mean average precision), which is state-of-the-art performance on this dataset.

Next, we examine how different parameters affect YOLOv3's performance. We specifically look into the impact of varying the confidence threshold, the number of anchor boxes, and the size of the input image. Our findings indicate that while lowering the confidence threshold may lead to more false positives, raising the size of the input image and the quantity of anchor boxes can enhance YOLOv3's performance. Lastly, we contrast YOLOv3's performance with that of several other cutting-edge object detection algorithms, such as RetinaNet and Faster R-CNN. Our findings show that, despite being much faster, YOLOv3 achieves performance that is either comparable to or better than these algorithms.

Overall, our research demonstrates the effectiveness of YOLOv3 for object detection and highlights the impact of various parameters on its performance. These findings can be valuable for researchers and practitioners working on computer vision applications that require accurate and efficient object detection.