Research on SSD base network

Abstract

In this paper, eight neural networks are selected instead of VGG16 in the SSD model to construct the corresponding SSD model and were tested on ImageNet and VOC datasets. The accuracy, predict time and model size of the eight neural networks in object classification are analyzed. The mAP, FPS and model size of the eight neural networks as the SSD base network in the object detection are analyzed. The result can provide reference for object detection tasks under specific requirements.