RTMOSeg: Superior Segmentation Results for Railway Track Monitoring

Figure 5 showcases the visualization results of different methods utilizing dataset #1. It's evident that the segmentation output of RTMOSeg aligns most closely with the actual ground situation. In contrast, Swin-UNET and TransUNet exhibit slightly inferior segmentation effects compared to RTMOSeg, characterized by a few spikes, missing segments, and inadequate segmentation details. Furthermore, U-net, PSPnet, and DeepLabv3+ display numerous erroneous segmentation results. For example, U-net and DeepLabv3+ misclassify many background images as sleepers, while PSPnet inaccurately identifies many background images as rails, leading to subpar segmentation performance. Conversely, our proposed segmentation method, RTMOSeg, effectively distinguishes each category and handles segmentation details precisely.