Evaluations of semantic segmentation algorithms in railway scenarios

As technology continuously developed, accurate and fast semantic segmentation algorithms are increasingly needed. In this paper, we concentrate on the railway traffic and use a typical dataset to test nine typical semantic segmentation algorithms. The model size, detection speed and detection accuracy were compared. We find the outcomes can be categorized into two types. In Type 1, ViT, STDC2 and MobileNetV3 exhibit higher speed, yet their accuracy is relatively lower. Conversely, in Type 2, the metrics of ANN, DeepLabv3+, FASTFCN, FCN, SWIN, and UPerNet show some similarities. Certain details regarding the distinctions between similar models are explored, including the comparison between FASTFCN and FCN, as well as that between SWIN and ViT, etc. Overall, considering the unique optimization of MobileNetV3 for lightweight scenarios, its performance is regarded as the most favorable for our research. In a word, this study provides a useful reference for future semantic segmentation work.