Semantic segmentation with inexpensive simulated data
Jukka Peltomäki, Mengyang Chen, Heikki J. Huttunen · 2019
This paper studies the benefits of adding inexpensively gathered simulated data to improve the training of semantic segmentation models. We introduce our implementation to gather simulated datasets with minimal effort. In our implementation, we utilize a commonly available game engine (Unity) and aux-illiary graphical assets to assemble an environment to generate simulated data inexpensively. We also demonstrate that even the usage of spartan simulated data mixed with real-life data can increase the performance of the trained model slightly, given that the ratio of simulated data is suitable for the datasets.