Single class object detection model with Amazon Rekognition Custom Labels throws too few usable labels error

Question

I'm using AWS Rekognition to perform single-class object detection. I'm assigning image-level labels with only one label in my entire dataset.

This is based on a new feature released by AWS,

https://aws.amazon.com/about-aws/whats-new/2020/06/amazon-rekognition-custom-labels-now-supports-single-object-training/

I have created my dataset with the following configuration.

When I Train my model, my model fails with the following status message,

The manifest file has too few usable labels.

Any ideas on what I might be missing?

Answer 1

The minimum unique label count for the object location (bounding box / detection) use case is 1 label, but the minimum label count for "Objects, Scenes, and Concepts (classification)" i.e. image-level data is 2. If you were auto-assigning image level labels and there was only 1 label assigned, this is likely why you were getting the "manifest file has too few usable labels" error.

Source: https://docs.aws.amazon.com/rekognition/latest/customlabels-dg/limits.html

We would also like to invite users of Amazon Rekognition Custom Labels with questions about the service to consider asking your questions on AWS re:Post: https://repost.aws/

Thank you for using Amazon Rekognition Custom Labels.

Christian Dunn