How to implement specific segmentation with classification in code.

[Cohesion97]

Hi,

Congrats on your solid work which is impressive and seems promising.

I am interested in the classification part of your proposed method. As you mentioned in your paper, " a classifier with a simple segmentation head on top of the proposals". Is this classifier added directly after the upsampler layer of the first stage VAE?

Understanding the exact placement and integration of the classifier within the architecture would be very helpful. Could you let me know if you've included these parts of the codes?

Thank you for your work and for any assistance you can provide.

[wvangansbeke]

Hi,

Thank you for your interest. Many configurations worked.
For the numbers in the paper, I trained a classifier on top of the decoder of the VAE. I implemented this with 4 ResNet blocks + a standard DeepLab head + CE loss. If I find some time, I will add this and also include different dataset(s).

[Cohesion97]

Thanks for your detailed explanation, I have one additional question. Is the classifier trained separately (after the 1st stage vae training) or co-trained with the vae?

I will continue to follow your repo and look forward to more interesting features.

[wvangansbeke]

I tried multiple configurations. Separately training the classifier worked best.