COCO has set a standard in Object Detection task dataset format. Is there a tool for translating this dataset format to CreateML format so that data can be used in CreateML for training and evaluation? I've found Roboflow but I would rather use a python script rather than this platform as it seems too complex for my needs.
You can download the COCO dataset in CreateML JSON format from here: https://universe.roboflow.com/jacob-solawetz/microsoft-coco/dataset/2
There are also over 90,000 other user-shared datasets there you can download for free in the CreateML format: https://roboflow.com/universe
Or use the pre-trained models users have shared directly with the Swift SDK (the model it uses behind the scenes is newer and more accurate than the one used by CreateML): https://blog.roboflow.com/roboflow-ios-sdk
I stand by my question as I already have a dataset that was annotated in coco format :)
Here's a brief Python script that would convert the COCO annotation format to CreateML.
import json
input_annotations = json.load(open('annotations.coco.json'))
output_annotations = []
images = input_annotations['images']
image_dict = {}
for image in images:
image_dict[image['id']] = image['file_name']
for annotation in input_annotations['annotations']:
coords_dict = {}
coords_dict['x'] = annotation['bbox'][0] + annotation['bbox'][2]/2
coords_dict['y'] = annotation['bbox'][1] + annotation['bbox'][3]/2
coords_dict['width'] = annotation['bbox'][2]
coords_dict['height'] = annotation['bbox'][3]
cur_dict = {}
cur_dict['annotation'] = [{'coordinates': coords_dict, 'label': annotation['category_id']}]
cur_dict['imagefilename'] = image_dict[annotation['image_id']]
output_annotations.append(cur_dict)
with open('annotations.createml.json', 'w') as f:
json.dump(output_annotations, f)