TypeError: 'GeometryCollection' object is not subscriptable

[jetsonwork]

Hi I am getting the below error and I cannot complete the slicing, could you please advice how can I fix it?

TypeError: 'GeometryCollection' object is not subscriptable

[clairelin23]

I also get the same problem

[Preburk]

Same here, feels like a dependency has been updated and broke it

[fcakyon]

Have you tried with shapely==1.8.5? @Preburk @clairelin23 @jetsonwork

Moreover, all continuous integration tests for slicing operations are passing with the shapely==2.0.4

[PolyubinAI]

Hi, man!

shapely==1.8.5
slice_coco doesn't work with import error:
can't import MultiPolygon from shapely:

and also shapely==2.0.4 raises error:
TypeError: 'GeometryCollection' object is not subscriptable

[Alias-z]

I have the same problem here

[PolyubinAI]

I solve this problem with such case:

In my image, the bbox had a border right at the end of the snapshot. That is, I have a 384*384 snapshot and the bbox ends in a pixel with coordinates (30, 384).

As soon as I changed the end of the box to the value (30, 383.95) in the markup service, the year started working and everything went well

[Alias-z]

Hi Thanks. Follow your direction, I found the causation and potential solution:

https://github.com/obss/sahi/blob/main/sahi/utils/coco.py#L226-L240. The polygon must be valid in shapely. for example:

(https://shapely.readthedocs.io/en/stable/manual.html)

My solution is to pre-process my annotations before I convert them to COCO:

            for obj in data['objects']:
                polygon = Polygon(obj['segmentation'])  # convert segmentation to a Shapely Polygon and check validity
                if not polygon.is_valid:
                    invalid_reason = explain_validity(polygon)  # the reason for invalidity
                    print(f"Polygon from {json_path} is invalid: {invalid_reason}")  # print the reason for invalidity
                    polygon = polygon.buffer(0)  # correcting the polygon if it's invalid

I think the developer can do something similar with COCO annotation more efficiently. Otherwise I can find a way to do a PR

[fcakyon]

@Alias-z and @PolyubinAI, thanks a lot for explaining the core issue!

@Alias-z if you fix the errors in the tests, we are good to merge your PR and solve this issue 💯

[fcakyon]

@jetsonwork @clairelin23 @PolyubinAI @Preburk, this issue is fixed in the latest pr thanks to @Alias-z 🚀
#1047