A Real-time Multipoint-based Object Detector

Abstract

A real-time multipoint-based object detector - EMPDet is proposed in this paper to improve the processing speed with reasonable sacrifice in accuracy. A lightweight neural network block is proposed and integrated into the compact hourglass networks to reduce the consumption in image feature extraction. The channel mechanism is used to enhance the performance of the convolutional neural network to screen shallow semantic information in high-resolution feature maps. Experiments results on the detection benchmark (Microsoft COCO) show that the proposed detector has superior performance compared to the current most popular YOLOv3 under reasonable overhead.