Visual artifacts

[olivier-gillet]

Hi,
Thank you for this great implementation!
I am using it to generate simple "maps", like a segmentation task.
I am trying your pix2pix unet256 implementation of the network with my own loader,
I get these artifacts : sample image link.
It is lots of single or dual pixels of red green or blue "dots" spread over the prediction where it should predict no data (ie. black).
Do you know where could it come from? Is it due to the network ? Thank you

[junyanz]

I recommend that you use a semantic segmentation network for your task. GANs sometimes struggle when the task's output is discrete,

[olivier-gillet]

Thank you for your answer. Sure. But I want to try the power of GANs for this task. I was just surprised that theses dots are exactly (255, 0, 0) , (0,255, 0) and (0, 0, 255). So I was wondering if it was possible to know from which layer it comes. This dots are only present on no data (ie. black). It is like the network cannot learn to map "noise" to "black" (no data).

[junyanz]

Have you checked if the data loader reads these maps correctly? Another sanity check is to disable GAN loss and only use L1 loss, and see if you can reasonable results.

[kunyao2015]

@olivier-gillet Have you solved this problem?