Segmentation mask generated contains strings and letters at count keyword

[akshitac8]

At evaluation time, The segmentation json file generated contains buffered strings and letters at count keyword.
Example:

{'size': [800, 800], 'counts': b'coa;<1E_h0b0M0O2O000O10000000000000001OO1000000000000000000000000000O1000000000000000000000000001O0000001N1000001O00001N101O0O1O2O2K]b^6'}

[akshitac8]

I followed the code given at #150 But then I got

if not isinstance(segms[0], (list, Polygons)):
IndexError: list index out of range

[fmassa]

Hi,

I suppose you managed to fix the issue?
You were checking the rle encoding of the mask, instead of the polygons. Those are generally the encodings for crowd categories

[Jinksi]

On this topic, how do I extract polygon data after eval?
segm.json contains counts but no polygon.
Am I missing something?

[shaayaansayed]

+1, does segm.json not store the polygon/mask prediction?

[shaayaansayed]

@Jinksi you can use this to decode RLE string to mask:
https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocotools/mask.py#L87

pass in the dict stored in the segmentation key

[Jinksi]

@shaayaansayed Good one, thanks.
I combined this with @jsbroks imantics to convert the decoded mask to a segmentation polygon:

from imantics import Mask
def maskToPolygon(mask):
    polygons = Mask(mask).polygons()
    return polygons.segmentation

It could be good to add this to segm.json by default, but it adds to filesize.

[jsbroks]

Another way to do this:

from imantics import Polygons
def maskToPolygon(mask):
    return Polygons.from_mask(mask).segmentation

both do the same thing, just though I would share.

@Jinksi If you would like RLE support added to the library, post a feature request under the issues

[AmberHen]

Is your problem solved? I have the same problem.

[wangg12]

Binary mask to polygons can not guarantee correctness by naively finding contours.

For example, a donut-like mask: binary mask --> polygons(contours) --> binary mask(decode), the hole in the mask will disappear.

Thus I would recommend not to convert any of binary mask or RLE back to polygons.

[AmberHen]

Binary mask to polygons can not guarantee correctness by naively finding contours.

For example, a donut-like mask: binary mask --> polygons(contours) --> binary mask(decode), the hole in the mask will disappear.

Thus I would recommend not to convert any of binary mask or RLE back to polygons.

你好 我在inference里输出来的counts和上边一样 我应该怎么把这些轮廓线画到原图上 这样直观地看到分割的效果。