meaning of "counts" and "size" in "iscrowd: 1" segmentation

[waspinator]

These example were taken from instances_val2017.json

When "iscrowd": 0 segmentation looks like

"segmentation": [
    [
        289.74,
        443.39
    ],
    [
        289.74,
        443.39
    ]
]

but when "iscrowd: 1" then there are additional "counts" and "size" fields. What do they mean?

"segmentation": {
    "counts": [
        272,
        2,
        4,
        4
    ],
    "size": [
        240,
        320
    ]
}

[alwansm]

"iscrowd": 0 if your segmentation based on polygon (object instance)
"iscrowd": 1 if your segmentation based uncompressed RLE (crowd)
check: facebookresearch/Detectron#100

[waspinator]

@alwansm Thanks, its as bit more clear now. I'm still confused on how I convert a compressed RLE to an uncompressed RLE version.

When I run

a = np.asfortranarray(np.zeros([5, 10])).astype(np.uint8)
mask.encode(a)

produces:

{'counts': b'b1', 'size': [5, 10]}

which is the encoded RLE instead of the list of something that coco expects.

[waspinator]

def binary_mask_to_rle(binary_mask):
    rle = {'counts': [], 'size': list(binary_mask.shape)}
    counts = rle.get('counts')
    for i, (value, elements) in enumerate(groupby(binary_mask.ravel(order='F'))):
        if i == 0 and value == 1:
            counts.append(0)
        counts.append(len(list(elements)))
    return rle

[andics]

def binary_mask_to_rle(binary_mask):
    rle = {'counts': [], 'size': list(binary_mask.shape)}
    counts = rle.get('counts')
    for i, (value, elements) in enumerate(groupby(binary_mask.ravel(order='F'))):
        if i == 0 and value == 1:
            counts.append(0)
        counts.append(len(list(elements)))
    return rle

from itertools import groupby
Is missing from this code snippet