Comparing one-stage and two-stage learning strategy in object detection

Abstract

Object detection plays a vital role in computer social perception and computer vision. It could be applied to computer navigation, video monitoring, industrial detection, and so on. It greatly reduces the human labours by automatically locate and identify objects. Nowadays, the mainstream methods of object detection could be separated into the one- and two-stage method. The one-stage method leverages Convolutional Neural Network (CNN) for obtaining features and directly locate the target objects and their corresponding category probabilities. Different from the two-stage solutions, its accuracy is lower and the recognition speed is higher. The two-stage method is a straight forward solution, which process is mainly completed through a complete CNN, so CNN features will be leveraged to extract the feature description of the target among candidate regions through a CNN. The accuracy of the two-step method has been greatly improved, but the running speed is much slower than the one-step method. While the one-step method is less accurate, it is much faster. In this work, representative works for object detection are conducted and compared. The results could further demonstrate their respective advantages.