SOLO2COCOConverter throws keyerror "instances" when no labeled objects in image

[TimSchoonbeek]

Hi,

Due to the settings used in my dataset generation, apparently not all images have a visible object (around 1/1000 images don't have 2D bounding boxes). This causes the SOLO2COCOConverter to crash with a keyerror "instances" error.

Perhaps the converter can be updated to just skip images if there is no label, rather than crash? As I don't know which images don't have any bounding boxes, I now have to write a separate script which deletes these instances this before feeding the data to the converter.

[StevenBorkman]

@TimSchoonbeek Thanks for the feedback. I'll definitely add this to the backlog because you are correct, we should handle this case more gracefully.

[nroadley]

Hi,

Due to the settings used in my dataset generation, apparently not all images have a visible object (around 1/1000 images don't have 2D bounding boxes). This causes the SOLO2COCOConverter to crash with a keyerror "instances" error.

Perhaps the converter can be updated to just skip images if there is no label, rather than crash? As I don't know which images don't have any bounding boxes, I now have to write a separate script which deletes these instances this before feeding the data to the converter.

@TimSchoonbeek you just saved me from giving up on the SOLO2COCO converter. Replacing one faulty label fixed it for me. So thank you!

@StevenBorkman I'd also recommend to add the same fix to your Pysolo Fiftyone as that also crashes when one label is missing on a sequence. I added a comment to an issue posted there too.

[eulercauchy]

I fixed this for myself by editing the dataclasses_json/core.py lines 170-175 to

        # The field should be skipped from being added
        # to init_kwargs as it's not intended as a constructor argument.
        if not field.init:
            continue
        if not field.name in kvs:
            field_value = None
            field_type = None
        else:
            field_value = kvs[field.name]
            field_type = types[field.name]
        if field_value is None:

Works for me, but I guess you could add this sort of error checking further up the chain and just dumping in None if there's no bounding box.