The culane dataset is a relatively general dataset for lane line detection, which was proposed by the SCNN paper. It was collected by cameras mounted on six different vehicles driven by different drivers in Beijing. Over 55 hours of video were collected and 133,235 frames were extracted. The data example is shown above. We split the dataset into 88880 training sets, 9675 validation sets and 34680 test sets. The test set is divided into normal categories and 8 challenging categories. Later, most of the lane line detection papers used this dataset as the target to show the performance of their methods. This article records the culane dataset and how the pictures are marked.
dataset content
After decompression, culane is divided into several files on the picture, among which:
1. Training and validation sets:
- driver_23_30frame
- driver_161_90frame
- driver_182_30frame
For each image, there will be a .txt annotation file. Taking 00120 as an example, the labeling result is shown in the figure. Every two numbers are a pixel coordinate (x, y), and culane is an equal vertical and equal interval label, every ten pixels are marked, and the lower half of the image is marked.
2. Test set:
- driver_37_30frame
- driver_100_30frame
- driver_193_90frame
3. Training/validation/test list:
- list
Among them, train.txt, test.txt, val.txt, these three folders record the path of the corresponding original image.
Among them, train_gt.txt val_gt.txt test_gt.txt, the three files record the path of the corresponding original image, the path of the segmentation map, and whether there are 4 lane lines (1 is 1, no is 0, 4 lane lines The order is the adjacent left lane line, the left lane line of this lane, the right lane line of this lane, and the adjacent right lane line)
4. Semantic segmentation information of the training and verification lanes:
-laneseg_label_w16
, which is generated from raw annotations.