Road Segmentation

Abstract

This project aims to discuss the nature of segmentation with respect to the requirements of its application on Road Segmentation. Road Segmentation are popularly discussed nowadays because it is served as an important role of the autonomous car. One important requirement of road segmentation is the inference time. A model has to be able to inference road scene fast enough so that it can react to the road scene in real-time. The discussion between leveraging model accuracy and running time become important if we want to use it in an autonomous car. However, semantic segmentation nowadays are in nature not designed for segmentation tasks. We discuss a newly design architecture (FRRN) which separate the model into two streams. This architecture gives better training efficiency while impose large burden on memory use, which results in slow inference time. we discuss the potential of two streams architecture and proposed a new architecture that has even better learning capacity with multi branch of streams. Together with our newly invented modularized unit, we proposed a system that speed up the inference time without suffering from accuracy loss. Our system achieved accuracy of 65.4% on Cityscapes datasets with inference time 3 times faster than FRRN, which adopted similar two streams architecture.