The LSIOLD dataset stands for LSI (Laboratorio de Sistemas Inteligentes) Obstacles-Lanes Dataset, and consists of 19000 images, used for the semantic segmentation of Obstacles and Lanes, divided into two groups, the training group, with 15000 images, and the validation group with 4000 images.
Note: the link to the dataset will be published soon.
This dataset was generated by combining two groups of images belonging to the database BDD100K (https://bdd-data.berkeley.edu/), The BDD100K has 10 000 images focused on the segmentation of elements present in the streets, and another 100 000 images with driveables zones. In order to join these two classes of labels, a convolutional neurnal network was trained, based on the ERFNet architecture, in charge of carrying out the segmentation of obstacles. For this purpose, the classes considered obstacles were selected from the group of 10 000 images. Once the network has been trained, 20 000 images of the group of 100 000 were taken at random, and the network was used to generate weak labels of obstacles on these images. The badly segmented images were discarded in order to obtain 19 000 images, with which we have the weak labels of the obstacles and the labels of the driveables zones.
The next step is to overlap these two labels to obtain a single group of images with both groups of labels. proceed to add each of the pixels corresponding to the driveables zones, on the labels of the obstacles; then proceed to modify the the IDs of each class so that they are as follows:
Class | ID |
---|---|
Pedestrians | 0 |
Small vehicles | 1 |
Medium vehicles | 2 |
Large vehicles | 3 |
Current lane | 4 |
Parallel lanes | 5 |
Unlabelds | 6 |