Training Amazon Rekognition Custom Labels model on adjusted manifest with corrections

Question

I have a set of images with bounding-box labels created using Ground Truth.  Some of the labels were adjusted using job-chaining as described here:

https://aws.amazon.com/blogs/machine-learning/chaining-amazon-sagemaker-ground-truth-jobs-to-label-progressively/

the output manifest from the chained job still has the original bounding-box information in it.  How does this affect training if I use this manifest to train an Amazon Rekognition Custom Labels model?  If the adjustment was correcting an error in the previous labeling job is it going to cause confusion in the training that the incorrect label information is still present?

Thanks.

Answer

Hello,

It sounds like your updated manifest contains both the correct and incorrect labels for that image, in which case the training would have varied accuracy assuming that the manifest line is still considered valid. Our training process will consider every label present in the manifest, and has no way of determining what is a correct or incorrect label.

Feel free to let us know if you have any further questions.

Training Amazon Rekognition Custom Labels model on adjusted manifest with corrections

Relevant content