Effect of Loss Functions on Domain Adaptation in Semantic Segmentation

In this paper, it is analyzed how different loss functions affect the performance of domain adaptation in the field of semantic segmentation. Semantic segmentation is a pixel-wise classification problem of an image. Large amounts of annotated data are required to train successfully in multi-parameter deep learning architectures. In recent years, several works have demonstrated that synthetic datasets are a good alternative since they are automatically annotated in virtual environments. However, due to the different distribution of source and target datasets, there is a decrease in performance. Domain adaptation methods address this problem by decreasing gap between source and target data. In this study, it is investigated that the effect of Cross-Entropy, Lovasz-Softmax, Dice Coefficient, Tversky and mean Intersection-over-Union Loss functions on domain adaptation in semantic segmentation. For our study, KITTI and Virtual KITTI datasets are used for real and synthetic images respectively. By evaluating the quantitative results, it is observed that the Dice Coefficient is relatively more successful.