Synthesising large, low cost and diverse datasets for robust semantic segmentation in self-driving tasks

Robust scene understanding algorithms are essential for the success of autonomous navigation. Unfortunately the supervised learning of semantic segmentation requires large and diverse datasets. For certain self-driving tasks like navigating a robot inside an industrial facility no datasets are freely available and the manual annotation of large datasets is impracticable for smaller development teams. Although approaches exist to automatically generate synthetic data, they are either too computational expensive, demand a huge preparation effort or miss a large variety of features. This paper presents a novel framework to generate synthetic datasets with great variance for low computing demand that are easily adaptable to different self-driving tasks (Available at https://github.com/cITIcar/SAD-Generator). As demonstration this approach was applied to a semantic segmentation task on a miniature road with random obstacles, lane markings and disturbing artefacts. Training thus synthesized data in a U-Net and and later fine-tuning it with a small amount of manually annotated data, improved pixel accuracy (PA) by 2.5 percentage points and mean intersection over union (mIoU) by 11.19 percentage points.