<p>Due to the expensive and time-consuming nature of obtaining pixel-level annotations for real-world images in semantic segmentation, using readily available synthetic data to train models is practical. This allows the model to adapt to real-world images without needing additional annotations. This process has been explored extensively in the context of Unsupervised Domain Adaptation (UDA). Despite numerous studies proposing novel adaptation strategies, many have largely overlooked the role of the student network and the crucial impact of spatial relationships between pixels on pseudo-label generation. To address pseudo-label noise and enhance the quality of pseudo-labels, three simple yet crucial modification modules are employed: (1) Student Network Reverse-Guiding The Teacher Network: Replace low-confidence pseudo-labels generated by the teacher network at the current position with high-confidence pseudo-labels generated by the student network. (2) Pixel-Space Self-Modification: Leverage the spatial distribution characteristics between pixels by replacing unreliable low-probability labels within a specified range with high-probability labels identified by different categories of pixels within the same range. (3) Small Connected Component Elimination: Identify small connected domains within a specified range that are below a certain threshold and replace the labels within these domains with the label having the highest number of pixels in the surrounding area. In summary, the improvements introduced by these three modules increased the mIoU from GTA to Cityscapes to 76.8 and from Synthia to Cityscapes to 67.9, demonstrating the effectiveness of the proposed modules in enhancing model segmentation performance.</p>