Urban Scenes Dynamic Adaptation Semantic Segmentation for Autonomous Driving

Urban scene image semantic segmentation is an important task for autonomous driving. Deep learning is an effective approach for image semantic segmentation, but it has to adapt to the scene changes, that is domain adaptation semantic segmentation. Currently, self-training is an effective method to achieve domain adaptation semantic segmentation. However, the effectiveness of this method depends on the correctness of pseudo label prediction of target domain images. To this end, in this paper, we propose a target domain dynamic adaptation self training method for semantic segmentation. Firstly, fast Fourier transform (FFT) is applied to transfer target domain images into source domain images style based on which improves the confidence of the pseudo label prediction. Then, a dynamic threshold for the pseudo-label selection is proposed, which is determined by the previous prior probability and the current prior probability of the class in the target domain. In this way, the pseudo label selection can be more adapted to image changes in the target domain. Finally, an improved cross-entropy loss function is proposed to alleviates the long tail effect of semantic segmentation. To verify the effectiveness, we conduct experiments on the tasks of GTA5->Cityscapes and SYNTHIA->Cityscapes. The experimental results of mIOU of Cityscapes on the two tasks are 53.9% and $\mathbf{52.7 \%}$, which are outperform some of the state models.