COCO number of classes

[Bala93]

Why does COCO have 91 classes, compared to 80 classes ?

[kazuto1011]

COCO has labels for only 80 classes, while the indices range from 0 to 90 (11 classes have no label).
To deal with this, the model simply outputs 91 class logits, and only 80 classes are supervised.
I understand that's kind of redundant, but I just followed the official Caffe implementation so that the pre-trained weights could be loaded as is.

[Bala93]

Thank you for the clarification.

[Bala93]

I would also like to know one more thing. How do I train for deeplab with 90 classes ? In the following file (https://github.com/kazuto1011/deeplab-pytorch/blob/master/libs/datasets/__init__.py), coco.yaml provision is not provided. I am working with train2014 dataset.

[Italy2006]

I would also like to know one more thing. How do I train for deeplab with 90 classes ? In the following file (https://github.com/kazuto1011/deeplab-pytorch/blob/master/libs/datasets/__init__.py), coco.yaml provision is not provided. I am working with train2014 dataset.

Sorry but do you solve this problem? I have the same question as you too, thanks for your reply