SEL is a semantic line detection dataset containing 1,750 outdoor images, which are divided into 1,575 training images and 175 test images.

Each semantic line is annotated by the coordinates of its two endpoints on the image boundary. If an image has one dominant line, it is set as the ground truth dominant semantic line. If an image has multiple semantic lines, the line that the annotator considers to be the best ranked is set as the ground truth dominant line, and the others are set as additional ground truth semantic lines. In this dataset, 61% images contain multiple semantic lines.