train_IDs.md

Train IDs

Regardless of the source data, the softmax output of the segmentation network follows Cityscapes class indexing:

ID	Class
0	road
1	sidewalk
2	building
3	wall
4	fence
5	pole
6	traffic light
7	traffic sign
8	vegetation
9	terrain
10	sky
11	person
12	rider
13	car
14	truck
15	bus
16	train
17	motorcycle
18	bicycle

Since the stored IDs in the segmentation masks in SYNTHIA, GTA5 and Cityscapes are inconsistent, one needs to convert them to the same indexing.

Converting ground truth to train IDs

It is possible to re-map the class indices of the segmentation masks directly in the dataloader and to load the original ground-truth maps. We pre-computed this mapping offline, however. The script tools/convert_train_ids.py reads in the original ground-truth masks, remaps the class IDs and saves the result on disk. To run the script, you can use the following template:

python tools/convert_train_ids.py --dataset [cs|gta|synthia]
                                  --ann-data [path/to/labels/]
                                  --ann-out [output/directory/]