Label matching, 182 or 183 labels?

[AlejandroHernandezMunuera]

Hi there,

I'm trying to use your model and migrate it to tensorflow using the caffe-tensorflow repository and your caffemodel and prototxt files. The problem I face now is that the Deeplab VGG-16 model trained on COCO-Stuff that you offer on thid repository only outputs 182 different labels but as I understand it should return 183 (91 for COCO, 91 for stuff and 1 for unlabeled).

Please let me know if I'm missing something,
thank you in advance

[nightrome]

Hi. That is correct. We do not output predictions for the unlabeled class, as it is (arguably) not a semantic class. Similar to PASCAL VOC it is treated as a void class and ignored from the loss. If you desperately need it you can retrain the network using the provided code or maybe you could just assign the unlabeled category to every pixel with a low confidence for all other classes.