error correctionFigure 5 shows the visualization results of each method based on dataset #1 We can see that the segmentation result of RTMOSeg is closest to ground truth while in other methods the seg

Figure 5 illustrates the visualization results of each method using dataset #1. It can be observed that the segmentation outcome of RTMOSeg closely resembles the ground truth. Conversely, Swin-Unet and TransuUNet display slightly inferior segmentation effects compared to RTMOSeg, with a few missing spikes and sleepers and inadequate segmentation details. Additionally, U-Net, PSPnet, and DeepLabv3+ exhibit numerous defects in their segmentation results, including incorrect segmentations. For instance, U-Net and PSPnet misclassify more background parts as sleepers, while DeepLabv3+ mistakenly identifies rail edges as rails, resulting in poor segmentation performance. In contrast, our proposed segmentation method, RTMOSeg, excels in accurately identifying each category and effectively handling segmentation details