kazuto1011 / deeplab-pytorch

PyTorch re-implementation of DeepLab v2 on COCO-Stuff / PASCAL VOC datasets

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Inconsistency of the labels from VOC and COCO

JINGTING92 opened this issue · comments

Since there are 21 classes in PASCAL VOC datasets and 180 classes in COCO datasets, how to make these two datasets consistent with the same number of labels, say 21, to use together?

VOC is annotated with 20 foreground "thing" classes while other pixels are labeled as just background. In contrast, COCO-Stuff provides 182 "thing + stuff" classes on all pixels. I think you may have to find sharable VOC classes from COCO "thing" classes and to make all other pixels ignored by defining them as background. COCO-Stuff label hierarchy may help you.