Real-Time Semantic Segmentation on Edge Devices with Nvidia Jetson AGX Xavier

Semantic segmentation is the problem of classifying every pixel in the image from a predefined set of classes or categories. It is a vital part of an autonomous vehicle vision system to perform traffic scene parsing. In such application, having accurate and low-latency performance is necessary but also challenging due to the task's high computational complexity. This study presents our implementation and performance evaluation of recent state-of-the-art semantic segmentation methods aiming to operate in high processing speed while maintaining high accuracy. Our experimental results on Cityscapes dataset and on Nvidia Jetson AGX Xavier mobile embedded platform demonstrate that it is possible to achieve that goal on resource-limited devices.