Adding classes to default coco.names

[eliwilner]

I was wondering.

Let say I've manage to train on a new class which is not within coco.names and your model is able to predict it.

Is it possible to add it to the default coco classes?
That is -> 81 classes that includes my trained one such that when a new image/video goes into the network he will be able to recognize the new trained label along side the coco ones?

Thanks

[zzh8829]

that is indeed possible but required very advanced knowledge in deep learning
i would do something like this

1. create dataset set with your new class
2. load pretrained weight into model with 80 class
3. convert 80 class model to 81 class model
4. freeze all bottom layers and then train on your new dataset
5. unfreeze all layers and train with coco + your dataset mixed
6. predict with your newly trained 81 class model

it's definitely not as simple as adding one line to coco.names
if i am understanding your question wrong please correct me

[zzh8829]

I just added a tutorial on transfer learning hope that helps https://github.com/zzh8829/yolov3-tf2/blob/master/docs/training_voc.md