GreenLeaf Farms runs a Rekognition Custom Labels project to classify images of plant diseases. The engineering team has 12,000 labeled images in an S3 bucket organized by prefix (s3://greenleaf-data/images/) and a SageMaker Ground Truth augmented manifest (JSON Lines) that references each image and its label. They need to start a training job in Rekognition Custom Labels that uses the labeled images and enforces a 80/20 train/test split. What is the most appropriate next step to ensure Rekognition can train the custom image-classification model?

The correct approach is to create a Rekognition Custom Labels project and import the Ground Truth augmented manifest. Rekognition Custom Labels is designed to directly consume augmented manifests (JSON Lines) from S3, which contain both image locations and their corresponding labels. The service role needs read access to the S3 bucket to access the images. Rekognition Custom Labels can then automatically handle the 80/20 train/test split when creating the dataset from the manifest. Incorrect options: Zipping images and inferring labels from folder names is not the standard or most efficient way to use Rekognition Custom Labels with pre-labeled data, especially when an augmented manifest already exists. Rekognition Custom Labels is a fully managed service; launching a SageMaker training job with a Rekognition built-in Docker image is not how you interact with Rekognition Custom Labels for training. It can consume augmented manifests. Rekognition Custom Labels does not require EFS mounts or S3 object tags for labeling when an augmented manifest is available. It operates on S3 data referenced by manifests.