LS Converter - COCO Dataset - KeyError

[wasson5e]

Im trying to retrain a YOLO model with label-studio (probably more than I want to do), but I don't want to lose the data the model was already trained on. Since YOLO uses a COCO dataset, Im trying to add my information to the COCO dataset. To do that, Im trying to get COCO into Label Studio (LS) but Im having trouble.

Ive pulled down COCO information from different locations:

1. Ultralytics will download the dataset during training of YOLO
2. From Kaggle - https://www.kaggle.com/datasets/awsaf49/coco-2017-dataset/data
3. From COCO - https://cocodataset.org/#download

When I try and run the conversion, Im hitting the same issue, typically a key error in the JSON data:

When using the json file from #3:

label-studio-converter import coco -i '/orin-robotics/datasets/annotations/captions_train2017.json' -o '/orin-robotics/datasets/coco_output' 
INFO:root:Reading COCO notes and categories from /orin-robotics/datasets/annotations/captions_train2017.json
Traceback (most recent call last):
  File "/root/Documents/label-studio/bin/label-studio-converter", line 8, in <module>
    sys.exit(main())
             ^^^^^^
  File "/root/Documents/label-studio/lib/python3.12/site-packages/label_studio_converter/main.py", line 191, in main
    imports(args)
  File "/root/Documents/label-studio/lib/python3.12/site-packages/label_studio_converter/main.py", line 173, in imports
    import_coco.convert_coco_to_ls(
  File "/root/Documents/label-studio/lib/python3.12/site-packages/label_studio_converter/imports/coco.py", line 139, in convert_coco_to_ls
    categories = {int(category['id']): category for category in coco['categories']}
                                                                ~~~~^^^^^^^^^^^^^^
KeyError: 'categories'

When using the data that Ultralytics pulled down in #1:

(label-studio) root@94c756140b69:/orin-robotics/datasets/coco/annotations# label-studio-converter import coco -i '/orin-robotics/datasets/coco/annotations/instances_val2017.json' -o '/orin-robotics/datasets/coco_output' 
INFO:root:Reading COCO notes and categories from /orin-robotics/datasets/coco/annotations/instances_val2017.json
INFO:root:Found 80 categories, 5000 images and 36781 annotations
WARNING:root:Segmentation in COCO is experimental
ERROR:root:RLE in segmentation is not yet supported in COCO
Traceback (most recent call last):
  File "/root/Documents/label-studio/bin/label-studio-converter", line 8, in <module>
    sys.exit(main())
             ^^^^^^
  File "/root/Documents/label-studio/lib/python3.12/site-packages/label_studio_converter/main.py", line 191, in main
    imports(args)
  File "/root/Documents/label-studio/lib/python3.12/site-packages/label_studio_converter/main.py", line 173, in imports
    import_coco.convert_coco_to_ls(
  File "/root/Documents/label-studio/lib/python3.12/site-packages/label_studio_converter/imports/coco.py", line 219, in convert_coco_to_ls
    item = create_segmentation(
           ^^^^^^^^^^^^^^^^^^^^
  File "/root/Documents/label-studio/lib/python3.12/site-packages/label_studio_converter/imports/coco.py", line 54, in create_segmentation
    segmentation = annotation['segmentation'][0]
                   ~~~~~~~~~~~~~~~~~~~~~~~~~~^^^
KeyError: 0

So my question is, am I using the correct data? Or am I doing something else wrong?

[heidi-humansignal]

Hello,

Thank you for contacting Label Studio,

Looking at the error message, it appears that the issue stems from how some of the annotations inside are formatted. The COCO dataset mixes two types of segmentation data: polygon-based and RLE-based (Run-Length Encoding). Label Studio’s converter currently supports only the polygon format in COCO, so when it encounters an annotation with an RLE mask, it raises the following error:

ERROR:root:RLE in segmentation is not yet supported in COCOKeyError: 0 This happens because the converter assumes every segmentation entry is a list (polygon coordinates), but RLE entries are dictionaries that store pixel-encoded masks.. Could you please look into this documentation for an alternative suggestion that is supported for RLE https://docs.humansignal.com/guide/predictions.html#Import-brush-segmentation-pre-annotations-in-RLE-format

Best Regards!

Comment by Oussama Assili
Workflow Run

[wasson5e]

Hello,

Thank you for contacting Label Studio,

Looking at the error message, it appears that the issue stems from how some of the annotations inside are formatted. The COCO dataset mixes two types of segmentation data: polygon-based and RLE-based (Run-Length Encoding). Label Studio’s converter currently supports only the polygon format in COCO, so when it encounters an annotation with an RLE mask, it raises the following error:

ERROR:root:RLE in segmentation is not yet supported in COCOKeyError: 0 This happens because the converter assumes every segmentation entry is a list (polygon coordinates), but RLE entries are dictionaries that store pixel-encoded masks.. Could you please look into this documentation for an alternative suggestion that is supported for RLE https://docs.humansignal.com/guide/predictions.html#Import-brush-segmentation-pre-annotations-in-RLE-format

Best Regards!

Comment by Oussama Assili
Workflow Run

Sweet, thanks for that link. I'll see what I can do on those segments.

[heidi-humansignal]

Sure! Please let us know if you have more questions!
Best Regards!

Comment by Oussama Assili
Workflow Run