Real-Time Semantic Segmentation Benchmarking Framework

Semantic segmentation has major benefits in autonomous driving and robotics related applications, where scene understanding is a necessity. Most of the research on semantic segmentation is focused on increasing the accuracy of segmentation models with few research on real-time performance. The few work conducted in this direction does not also provide principled methods to evaluate the different design choices for segmentation. In this paper, we address this gap by presenting the first real-time semantic segmentation benchmarking framework 2. The framework is comprised of different network architectures for feature extraction such as VGG16, MobileNet, and ResNet-18. It is also comprised of multiple metaarchitectures for segmentation that define the decoding methodology. These include Skip architecture, UNet, and Dilation Frontend. Experimental results on cityscapes with a case study using MobileNet architecture and two meta-architectures are presented.