Benchmarking domain adaptation for semantic segmentation

Abstract

Deep Learning (DL) requires a massive, labeled dataset for supervised semantic segmentation. Getting massive labeled data under a new setting (target domain) to perform semantic segmentation requires huge efforts in time and resources. One possible solution is domain adaptation (DA) where researchers transform the data distribution of existent annotated public data (source domain) to resemble the target domain. We develop a model on this transformed data. Nevertheless, this poses the questions of what source domain/s to utilize, and what types of transformation to perform on that domain/s. In this research work, we study those answers by benchmarking different data transformation approaches on source-only and single-source domain adaptation setups. We provide a new well-suited dataset using unmanned ground vehicle Husarion ROSbot 2.0 to analyze and demonstrate the relative performance of different DA approaches.