Copy existing AWS Rekognition datasets between AWS accounts

Applicable when

• There is already an existing AWS Rekognition dataset either on local or on another AWS account

Non-applicable when

• When there is no existing AWS Rekognition dataset and it needs to be built from scratch

Implementation

1. When creating 1st Rekognition project, AWS asks to auto-create a generic bucket for the classifier in the currently selected region. The bucket name will have the following naming format:
   custom-labels-console-<region>-<hash>
   e.g.: custom-labels-console-us-east-1-308ca889db
2. The structure that we care about from this bucket will be (as directories):
   • assets
     • holds all images used for training and testing grouped by <dataset_name>
     • those images were uploaded in batches of max 30 files
     • the batch directory name is just the timestamp of the upload time
   • datasets
     • holds multiple things (like: custom-labels, patches, etc.) but we mostly care about manifiest file, this describes the relation between asset, label, and bounding box
     • also grouped by <dataset_name> in folders
     • inside each group there is manifest/output/output.manifest file
3. Copy from the other AWS account either the whole bucket or just the directories you are interested to local:
   • aws s3 cp s3://<bucket_name> . --recursive (e.g.: aws s3 cp s3://custom-labels-console-eu-west-1-0e65f5a627 . --recursive)
4. Edit this manifest file and replace the bucket name in source-ref property value.
5. Copy from assets & datasets - directories we are interested to borrow to our Rekognition bucket from our AWS account:
   • Example directory tree:
     • assets:
       • patterns-training (copy this)
       • patterns-testing (copy this)
     • datasets:
       
       • patterns-training (copy this)
       • patterns-testing (copy this)
   • Upload to your AWS account Rekognition bucket
     • aws s3 cp <local_directory> s3://<bucket_name>/path/to/directory --recursive
     • e.g.: aws s3 cp patterns-training s3://custom-labels-console-us-east-1-308ca889db/assets/patterns-training --recursive
6. Go to AWS Rekognition and access `Datasets` and check that the new datasets are there and can be properly visualized and explored.