MOD-NAS: Migration-Based Object Detection Neural Architecture Search

Abstract

Object detection is one of the most concerned issues in computer vision, especially in the real scene, its accuracy, model size, inference speed, IoU and other indicators have strict requirements, so the manual construction of neural network model often can not get good results. This work uses neural architecture search (NAS) as a tool to search the lightweight object detection network that meets the requirements of real scenes by combining manual and automatic methods. In practice, we take the intelligent driving as the background, using classification data sets with proxy, shared weight and migration-based algorithm to search object detection network. The first step to search out the sensitive to intelligent driving of backbone, the second step migrates the backbone to the object detection network, then using the actual data collected by the network further fine-tuning. Finally, a neural network model satisfying the requirements is obtained According to the experimental results, we can see that our work in the model size, precision, inference speed and so on comprehensive index made a balanced and excellent results, proves the effectiveness of this method, also represents this work for the practical application of neural network architecture search provides effective solution.