CODEWITHSUNDEEP
amazon-web-servicesobject-detectionbounding-boxmechanicalturkamazon-sagemaker
Mia
Mia

Reputation: 579

Converting Mechanical Turk bounding boxes to JSON for object detection

I am looking to have several 1000 images annotated with bounding boxes on MTurk. When reading the following tutorial (https://blog.mturk.com/tutorial-annotating-images-with-bounding-boxes-using-amazon-mechanical-turk-42ab71e5068a), it seems like the output of a task will look like the following:

The Worker with ID A39ECJ12CY7TE9 gave the answer 
[{'left': 369, 'top': 47, 'width': 151, 'height': 160, 'label': 'stop sign'}]

For my object detection model in Sagemaker, I need a dataset of training images in jpg in one folder, and the correct annotations in json format for each image in another folder. An example of a proper json file: https://docs.aws.amazon.com/sagemaker/latest/dg/object-detection.html

So the question is, how can I use MTurk to get a dataset in the format I need? Is there a way to easily convert their output to JSON files? I have no experience with MTurk whatsoever, so any suggestions are appreciated!

Upvotes: 1

Views: 619

Answers (1)

Stretch
Stretch

Reputation: 3759

AWS have just released a new feature called "Sagemaker GroundTruth" - where you can send data to Mechanical Turk to get them to create the annotations for you.

https://aws.amazon.com/sagemaker/groundtruth/

I believe this would fix your issue - you will still be able to use MT, but will no longer need to worry about how the result set comes back. It will integrate straight into your Sagemaker model.

Upvotes: 2

Related Questions