Why an extra 6 classes?

[brianoppenheim]

I saw that in a previous github issue, 6 classes had to be added to the 80 existing classes in the coco.names list (though they are only rewordings of objects already in the dataset):

motorcycle potted plant dining table tv couch airplane

This causes some of the .txt annotation files to have classes numbered up to 86, which causes errors when I try to run known implementations of YOLO (like the AlexeyAB repository, which expects classes only up to 80). Can somebody please explain?

[ssaru]

maybe.. it should add some logic that ignoring parse some class

before darknet(YOLO framework) has 80 classes list https://github.com/pjreddie/darknet/blob/master/data/coco.names

OR

you can modify the shape of tensor block of detector other implementation of detectors(like a YOLO, R-CNN, etc). in that case, you can't use pre-trained weights

Thanks

[eehoeskrap]

Hi !

This is not a matter of adding six classes. Just change the name of the class.

[coco.names] -> [origin_coco.names] motobike -> motorcycle pottedplant -> potted plant diningtable -> dining table tvmonitor -> tv couch -> sofa aeroplane -> airplane

This is because the name of the original yolo v3 class and coco.names in darknet are different.

Reference : https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/