Automatic CNN Compression System for Autonomous Driving
Daichi Murata, Toru Motoya, Hiroaki Ito · 2019
A system has been developed to realize autonomous driving (AD) with Convolutional Neural Networks (CNNs). To apply CNNs to AD, the amount of calculations must be compressed with little decline in accuracy because it is too large to implement in the in-vehicle processors. However, in the conventional methods, an optimal configuration for CNN compression takes a long time to explore. In this paper, we propose a technique to explore an optimal configuration within a short time by combining both training time reduction per configuration and the priority search only around the optimal solution. The results shown that the proposed system can reduce processing time and improve mean Average Precision (mAP) by 70.0% and 2.19 pt. compared with conventional methods, respectively.