improve model

[de-code]

The model is currently based on the pix2pix model.

One optional extension is to use separate channels per annotation (configured in color_map.conf).
However, the model needs to be amended further to make that work correctly (in progress).

I subsequently found this recent SegAN: Adversarial Network with Multi-scale $L_1$ Loss for Medical Image Segmentation paper which uses a similar approach. It appears to be using separate discriminators (which may not scale so well). Instead we may want to share weights.

A previous paper, Semi and Weakly Supervised Semantic Segmentation Using Generative Adversarial Network may also be interesting. Although it doesn't use the term discriminator in the traditional sense (as detailed in the paper itself).

[KayShenClarivate]

Is the current model available for download anywhere? Thanks!

[de-code]

Is the current model available for download anywhere? Thanks!

Hi, no, not currently. I would recommend re-training the model. In the previous experiments it was only training for about 30 minutes on a K80 GPU.