Does Icevision support segmentation masks in the RLE bitmask format?

[Robotatron]

I tried parsers.COCOMaskParser to load my COCO dataset but getting an error:

ValueError: invalid literal for int() with base 10: 'YXi61og0000og00QXO00000O1000000O10O100000O100000000O10000000000O10000000000O01000000000O100000000000000O100000000000000O10001O0000000000000000000O10000000001O000O100000000000001O00000000000O2O0000000

Does that mean icevision only supports polygons for COCO masks?

[gatordevin]

I just solves this problem myself but it required changing their source files. Essentially your COCO dataset already has encoded RLEs which is what icevision convert all polygons and RLEs to under the covers. They assume that your dataset consists of either polygons or just normal RLE rather than encoded RLEs. I changed the source to remove the second encoding call which is what is calling this issue and it solved my problem. Also ensure you have iscrowd set to 1 rather than set to 0.

icevision/icevision/core/mask.py

Lines 215 to 218 in ffc0815

	def to_erles(self, h, w) -> EncodedRLEs:
	return EncodedRLEs(
	mask_utils.frPyObjects([{"counts": self.to_coco(), "size": [h, w]}], h, w)
	)

Rather than call mask_utils I just took the [{"counts": self.to_coco(), "size": [h, w]}] and passed it into EncodedRLEs