Using layer-wise training for Road Semantic Segmentation in Autonomous Cars

A recently developed application of computer vision is pathfinding in self-driving cars. Semantic scene understanding and semantic segmentation, as subfields vision, are widely used autonomous driving. segmentation for uses deep learning methods various large sample datasets to train a proper model. Due the importance this task, accurate robust models should be trained perform properly different lighting weather conditions presence noisy input data. In paper, we propose novel method called layer-wise training evaluate it on light efficient structure an neural network (ENet). The results proposed compared with classic approaches, including mIoU performance, robustness noise, possibility reducing size two RGB image road (CamVid) off-road (Freiburg Forest) paths. Using partially eliminates need Transfer Learning. It also improves performance when noisy.