Smart-YOLO: A Light-Weight Real-time Object Detection Network

Abstract

In the computer vision, YOLO has an important position in the field of object detection, but due to its speed limitation, it is not suitable for scenes that require extremely strict real-time performance, such as smart cameras or some mobile devices. However, inverted bottleneck layers, which are built upon depth-wise separable convolution, have been the predominant building blocks in state-of-the art object detection models on mobile devices. In this work, we propose a new lightweight algorithm Smart-YOLO based on the YOLO framework which uses inverted bottleneck blocks and deep-wise separable convolution. We also put forward a new loss function to make up for the loss of accuracy caused by the replacement of the backbone network. The results show that compared with YOLOv3, the accuracy of our model is reduced by about 21%, but achieved up to 4.5 speedup, and the model size is only about 1/8 of the original. This shows that our network is smaller, faster, and more suitable for scenarios that require higher speed and efficient.