You work for a scientific research company where you need to gather data on tree specimens. You have scientist peers who go out in the field across the globe and photograph tree species. The images that they gather need to be classified and labeled to use them in your training datasets in your machine learning models. What is the best way to label your image data most accurately and in the most cost-efficient manner?

Answer options:

A.Hire human image labelers to process all of your images and label them.
B.Use Amazon Rekognition to analyze all of your images. For the ones that the Rekognition cannot label, have human labelers that you hire attempt to label them.
C.Use an open-source labeling tool such as BBox-Label-Tool to process all of your images. For the ones that the tool cannot label, have human labelers that you hire attempt to label them.
D.Use AWS SageMaker Ground Truth to automatically label your images and use the AWS Ground Truth human labelers to label the images that the automatic labeling cannot label.

Answer: D
Option A is correct. Human labelers may be able to label all of your images correctly. But they will be slow and expensive.
Option B is incorrect. While the Amazon Rekognition service analyzes image data, it does not have the human labeler to active learning model loop that trains an automatic labeling model that Amazon SageMaker Ground Truth has. Therefore, a labeling process based on Rekognition will be more costly and less accurate than a process based on Amazon SageMaker Ground Truth. (See the Amazon Rekognition overview and the Amazon SageMaker Ground Truth overview)
Option C is incorrect. An open-source image labeling solution may label some images automatically, and a human labeling team that you hire can label the ones the open-source software cannot label. This process lacks the human labeler to active learning model loop that trains an automatic labeling model that Amazon SageMaker Ground Truth has. Therefore, a labeling process based on an open-source image labeling solution will be less accurate than a process based on Amazon SageMaker Ground Truth.
Option D is correct. As documented in the Amazon SageMaker Ground Truth overview: “Amazon SageMaker Ground Truth uses a process that starts with an active learning model that is trained from human-labeled data. Any image that it understands is automatically labeled. Ambiguous data is sent to human labelers for annotation. Then the human-labeled images are sent back to the active learning model to retrain the model to improve its accuracy incrementally. (See the Amazon SageMaker Ground Truth service overview)
Reference:
See the Amazon SageMaker Ground Truth service overview) and the Amazon Rekognition overview