Multiple masks area

[crapthings]

can i split the cloth masking to tops and bottoms? or should training multiple models for such task?

[plemeri]

Hi @crapthings, you need to change our model into multi-class segmentation model while this model can be regarded as technically a binary segmentation, foreground and background. It is not a hard work, just change the output channel from 1 to the number of classes (in your case, 2), change the activation of the final output logit from sigmoid to softmax, and change the loss function. You might want cross entropy loss which supports multi class logit with softmax activation. Also, don't forget to modify dataloder to encode the annotation into one-hot or integer mask.

Or, you can just use other semantic segmentation models such as mmsegmentation. Good luck on your project!