Training Instance Segmentation and Lane Detection Models in One Network Architecture

Abstract

A new network architecture with a novel training method is proposed in this paper which can achieve two tasks of road defects instance segmentation and lane detection.It is composed of a backbone and two independent output branches for instance segmentation and lane detection.The experiments are conducted on new datasets collected by us.Through our method of alternately training two network branches while continuously reducing the learning rate, it can be found that the accuracy of our two branches can be similar with the accuracy training with two different models.This shows the effectiveness of our training method.Furthermore, our method can reduce model memory.