A Domain Adaptive Segmentation Label Generation Algorithm for Autonomous Driving Scenarios

Semantic segmentation algorithm is a cornerstone algorithm in the field of autonomous driving. The complex and variable data in the production environment makes the data domain in the production environment seriously offset, which causes significant performance degradation of the network model in the development environment data. In this paper, we propose a pseudo-label generation algorithm (EOPL) with smooth assignment of label expansion and erosion based on the adversarial learning algorithm and self-supervised learning strategy in domain adaptive learning. In this paper, experiments are conducted on GTAV and Cityscapes datasets, and the algorithm is validated using multiple network models. The experimental results show that under the condition that the amount of source domain data is reduced by half and the number of domain classification training iterations is shortened by half, the mIoU evaluation index of the FADA model with EOPL improves by 1.5% overall and 10 percentage points higher for individual class IoU evaluation index; while the ADVENT model with EOPL improves by 5% for most classes and 30% for individual classes, which not only shortens the running time, but also improves the efficiency of image segmentation.