Object detection based on Yolov4-Tiny and Improved Bidirectional feature pyramid network

Abstract

In the field of small object detection, Yolov4-Tiny is inadequate in feature extraction and does not make best of multi-scale features. In this paper, an improved BiFPN framework is proposed based on Yolov4-Tiny to increase object detection precision. Moreover, the Yolov4-Tiny is taken as the backbone network and introduce spatial pyramid pooling (SPP) to connect and merge multi-scale regions. Finally, our method can achieve 79.53% map on Pascal VOC dataset, which is 2.12% higher than the original Yolov4-Tiny model.