Publications Repository - Gdańsk University of Technology

With the advances in systems-on-a-chip technologies, there is a growing demand to deploy intelligent vision systems on low-cost microcomputers. To address this challenge, much of the recent research has focused on reducing the model size and computational complexity of contemporary convolutional neural networks (CNNs). The state-of-the-art lightweight CNN is MobileNetV3. However, it was designed to achieve a good trade-off between accuracy and latency on a single large core of a Google Pixel 1 smartphone. Accordingly, MobileNetV3 is not optimized for platforms with different hardware characteristics and its predecessors may perform better for a given target platform. The aim of this paper is twofold: 1) to analyze the performance of different compact CNNs on Raspberry Pi 4; 2) to manually adapted the most promising models to better utilize the Raspberry Pi 4 hardware. After exploring a number of modifications, we present a new CNN architecture, namely MobileNetV3-Small-Pi, which is 36% faster and slightly more accurate on ImageNet classification compared to the baseline MobileNetV3-Small.

Authors

• Wojciech Glegoła link open in new tab ,
• dr inż. Aleksandra Karpus link open in new tab ,
• dr Adam Przybyłek link open in new tab