How do I train my own dataset on facebook/deepmask?

Question

I am currently trying to add test images to facebook/deepmask so that I can train these images. Currently, the deepmask program uses CoCo images and annotations but I was wondering if it was possible for me to add my own images. The only reason is that I want to only train one type of image, which in this case is grapes. I looked through the annotation file, but it is full of specs of the image. The only reason I am asking this question is because there were no tutorials/answers online for my specific problem.

Here is the github to deepmask: https://github.com/facebookresearch/deepmask

Also, to download the annotation file for CoCo image use this command in the terminal: wget http://msvocds.blob.core.windows.net/annotations-1-0-3/instances_train-val2014.zip

Thank you everybody for your time!

Stephen

Answer 1

My experience is the same: you should indeed dump your masks into a coco-style format. It's json, and it is available here: coco dataset . My only comment is that coco uses a "costum Run Length Encoding (RLE)", that is explained nowhere. Therefore it is easy to instead use a polygon encoding of the mask. Therefore, do:

1. Get the mask
2. encode it as a polygon
3. write it to a json file

Hope this helps :-)