On improving cross-domain performance of semantic segmentation in the case of limited data

This thesis focuses on the problem of semantic segmentation in computer vision. Semantic segmentation aims to assign a class label to each pixel in an image. Accurately segmenting object boundaries and preserving their integrity improves segmentation performance. In addition, the model successfully learns low-weight classes in the dataset, which improves the performance. Domain adaptation aims to make a model trained in the source domain work well in the target domain with a different distribution. In the first section, a difference of Gaussian loss function is proposed to improve semantic segmentation accuracy. This loss function helps the model better segment object boundaries and preserve object integrity. It also enhances segmentation performance for classes that are less represented in the dataset compared to other classes. The second section introduces blind domain adaptation for semantic segmentation. Blind domain adaptation refers to the scenario where only access to the source domain is available during training, and the target domain is unknown. To achieve well performance on a target domain with real data when the model is trained on a synthetic dataset, an edge attention module is proposed. Edges are adopted as a common base feature between the source and unknown target domains. Edges have been shown to contribute to model learning through the attention mechanism in blind domain adaptation for semantic segmentation.