Learning domain generalizable semantic segmentation model by transformation-invariant feature subspace projection

Semantic segmentation has made significant advances when the training and test data share an identical distribution. However, the performance of semantic segmentation models often degrades on test data different from the training domain. To tackle this challenge, this paper proposes a novel Transformation-Invariant Feature Subspace Projection (TIFSP) approach to learn semantic segmentation models that generalize to unseen target domains. Specifically, TIFSP constructs a transformation-invariant feature subspace by identifying the space less sensitive to data transformations, into which we project intermediate representations of deep neural networks, extracting domain-generalizable features. Moreover, we present an enhancement loss to enrich this subspace, facilitating transformation-invariant feature learning. TIFSP can be taken as a plug-and-play module, seamlessly incorporated into other convolutional neural network and transformer approaches to improve their cross-domain generalization abilities. Extensive experiments on five benchmarks demonstrate improved generalization performance when employing the proposed method for semantic segmentation, compared with the existing twelve baselines.