Domain Adaptation for Lane Marking: An Unsupervised Approach

A major roadblock for designing deep learning based supervised solutions to any problem is the requirement of huge amount of labelled data for training. Domain adaptation techniques are designed to alleviate this problem. The aim of domain adaptation is to ensure that a model trained over available labelled data of source domain generalizes well over unlabelled data of target domain. Data distribution among source and target domains are different but related. Various techniques have been successfully applied to adapt deep learning based image classifiers from source to target domain data. In this paper, we expand the scope of domain adaptation by designing it for a much more complex image processing system, i.e., lane marking that involves instance segmentation. Automatic system to mark lanes in images is a crucial component of self-driving cars. Such systems have been trained over data from very specific domains such as highway traffic images. It is crucial for lane marking systems to be able to adapt to different domains as traffic scenarios vary across countries, road types, etc. We designed a lane marking model that successfully generalizes from highway-traffic images to images of traffic data from Lahore (provincial capital of Pakistan). Our system shows that domain adaptation significantly improves the performance of lane marking system for the unlabelled data of the target domain.