Semantic segmentation using TAG label and transformer

Currently, research on artificial intelligence that autonomously interacts with surroundings without human management has attracted research attention in vehicle and robot-related fields. The recognition of the surrounding environment is the basis for artificial intelligence that requires interaction with the surroundings, which means that research on object detection is necessary. In general, object detection proceeds in the same way as detection and segmentation. Among them, in the case of segmentation, the size of the model is smaller, and more information can be obtained than detection using anchors. However, the inferior detection performance and generalization ability of this method for small objects has limited its further application. In this paper, a modified transformer structure with different configuration of training data from the existing label data is presented to improve the performance of segmentation.