CamVid like Mask

Part 1 (2019)
1

Newbie alert
I want to make camvid like dataset for a problem. I am getting transparent images of those “objects” and mixing it other backgrounds , so that I can try with Unet or Tiramansu to do segmentation. Is it possible to can create mask from transparent images programmatically than manually annotating ?
Any code will be helpful.
Also ,any recommendation on the architecture will be great,

Thanks in advance
Rajeev

2

You should be able to load the image with the alpha channel and make the mask be 1 where alpha is not 0. Fastai does not support tiramisu out of the box so would recomment unet.