Comparison of training methods for the binarized neural object detection network

Binarized neural networks are of great interest because they can dramatically reduce the amount of computation and memory, but they are still in the early stages of research and are still challenging. In this paper, we investigate three training methods of a binarized object detection network and measure their training speeds and detection accuracies based on Darknet platform.

1. Introduction

The object detection (classification/localization) technology using deep learning [1], [2] provides a higher detection rate than existing hand-craft feature based technologies [3], [4], but it is difficult to utilize it in real-time and mobile applications due to its large model size and high power consumption. In order to overcome the limitations, several small-sized and lightweight algorithms that operate with low power are proposed [5]–[7].