A Study of RobustNet, a Domain Generalization Method for Semantic Segmentation

Abstract

Domain Generalization alleviates the domain gap between training set and test set, improving the performance of deep neural networks on out-of-dataset data. This opens the possibility of deploying models on unlabelled data that were previously pretrained on other datasets. In this article, we study the ideas and performance of RobustNet [Choi et al. CVPR 2021], a recent method for Domain Generalization in Urban-Scene Semantic Segmentation. Instead of exposing the network to a wide range of domains, RobustNet tries to separate domain-variant from domain-invariant features via a whitening transformation. Then, only the domain invariant features are used for training, which allows to reduce training time since no combination of datasets is needed to achieve domain invariance. In addition, we provide an easy-to-use demo where users can quickly test their own data and compare the results of RobustNet against the state of the art for semantic segmentation. **This is an MLBriefs article, the source code has not been reviewed!**<br> **The original source code is [[available here|https://github.com/shachoi/RobustNet]] (last checked 2022/10/22).**