search

mcordts / cityscapesScripts

README and scripts for the Cityscapes Dataset
MIT License
2.12k stars 607 forks source link

[BUG]: Instance Train ID Images #184

Open collinmccarthy opened 2 months ago

collinmccarthy commented 2 months ago

I'll keep this brief. Based on the semantics of json2labelImg.py, the json2instanceImg.py has a typo which will lead to labeling "caravan" and "trailer" as instance classes instead of void, when using "trainIds" encoding, for some images.

In https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/preparation/json2instanceImg.py#L141-L143:

if labelTuple.hasInstances and not isGroup and id != 255:
    id = id * 1000 + nbInstances[label]
    nbInstances[label] += 1

should be as follows, were we use labelTuple.trainId instead of id when comparing to 255.

if labelTuple.hasInstances and not isGroup and labelTuple.trainId != 255:
    id = id * 1000 + nbInstances[label]
    nbInstances[label] += 1

If someone doesn't want to fix the code, one "solution" would be to just not use train IDs, but use original label IDs which don't have this issue, and then use the mapping in cityscapesscripts when loading in the data. Also I'm not entirely sure this matters if one is ignoring these classes when loading in the data. But depending on how the data is used, it could lead to issues.

collinmccarthy commented 2 months ago

I spent some more time on this and determined this script to generate trainId images is not pixel-level accurate either, as described in #136. I instead wrote my own script to input the _instanceIds.png images and directly create _instanceTrainIds.png images using numpy and the label mappings in helpers/label.py.