Improvement of YOLOv3 by Using Object Detection Layers of SSD

Convolutional Neural Networks are being widely used in computer vision tasks including object detection. YOLOv3 is a preferred choice in resource-limited environments because of its high detection capability and low complexity. YOLOv3 has an object detection layer different from that of its competitor, SSD. In this paper, the object detection layers of SSD will be applied to YOLOv3. The object detection layers of YOLOv3 output the class information, the confidence value, and the localization information at the same time. In the proposed architecture, the localization information is produced by a separate layer, and the confidence information is absorbed into the class information as a backgroud class. The experimental results shows that YOLOv3 can reach 6.4% point higher MS COCO AP with the variation.