How do I import an Encord dataset?

šŸ¤ Community Help
3

Solved it. Hereā€™s the steps I took:

  1. Add a new supercategory to the categories array.
    Roboflow uses the ā€œAnnotation groupā€ name you input at the time of project creation as a supercategory for other categories, while Encord just uses a bounding-box supercategory. The former is added into the categories array with id 0, while the latter isnā€™t in the categories array to begin with. The Roboflow supercategory has "none" as a supercategory itself. This is how your categories array should look like:
  "categories": [
    { "supercategory": "none", "id": 0, "name": "stuff" }, // the actual name of your annotation group here
    { "supercategory": "stuff", "id": 1, "name": "foo" },
    { "supercategory": "stuff", "id": 2, "name": "bar" }
  ]
  1. Delete all "file_name" lines. Encordā€™s COCO JSON has the original file name in "image_title", which is useless, and a generated random filename in "file_name". Ctrl+F in VSCode, check Regular Expression search, input ^.*"file_name".* in Find, and Replace All with an empty string to delete all of the useless file names. Then, Find "image_title" and Replace All with "file_name", so you get the "file_name" pointing to the actual file name.
  2. Put this JSON next to all your images and upload by selecting the folder. The annotated ones should go into the Annotated tab. I named my JSON labels.coco.json, I donā€™t know if that holds any relevance.

I hope in the future Roboflow takes care of this automatically without user intervention. I also wish the documentation regarding annotation imports was more transparent. All they tell you is that Roboflow can magically import everything, but thereā€™s no concrete specification of each format and how you should set up files correctly for importing.

1 Like