Support RLE format and binary masks.

[wangg12]

🚀 Feature

Currently the code is based on polygon masks. RLE and binary masks should be supported for general datasets.

[fmassa]

I totally agree about binary masks. Indeed, there is a draft implementation in here that supports some functionality.

Ideally, I'd want SegmentationMask to internally support both Polygons and Masks, but I didn't do it yet because it was not necessary for the experiments I was performing.

I'm less sure about RLE format, is it commonly used elsewhere, or only in COCO for the crowd category?

But contributions are more than welcome! :-)

[wangg12]

@fmassa For RLE format, we can store it in coco-style json file (Or maybe the paths of binary mask can be stored in json file). Then we can load the json file into memory and convert mask of any format to binary mask, and then do some transformations. I want this feature because sometimes polygons are not very convenient.

[fmassa]

My understanding is that RLE is great for storing data in disk, but many operations are hard to be performed directly on the RLE representation, and passing via a mask representation is generally easier / necessary.

Here is what I propose: adding full support for binary masks (i.e., finishing the Mask class), and if the user has data in RLE format, we can internally (in the constructor of the SegmentationMask class) convert it into Mask representation, or the user can himself convert from RLE to mask inside his __getitem__.

What do you think?

[wangg12]

@fmassa Agree with you. The internal converting to binary masks is exactly what I want.

[wangg12]

Solved by PR #150