MLS-C01 Question 89
Select 4You are building a machine learning model to classify images of damaged cars into one of three categories: 'Minor Damage,' 'Moderate Damage,' and 'Severe Damage.' You have collected a dataset of 10,000 images but need labeled data to train your model. You are considering using Amazon SageMaker Ground Truth with Amazon Mechanical Turk to create the labeled dataset. Which of the following are steps you should take to implement this solution?
- A
Create a labeling job in Amazon SageMaker Ground Truth specifying the input dataset and the labeling workforce using Amazon Mechanical Turk.
- B
Define the labeling task and provide clear instructions for annotators, including category definitions and example images.
- C
Use Amazon Rekognition to automatically label the dataset and bypass human labeling entirely.
- D
Monitor the labeling job progress in the SageMaker console and review the labeled data for quality control.
- E
Set up a custom labeling UI template in Amazon SageMaker Ground Truth if the default template does not fit your use case.
Show answer and explanation
Correct answers: A, B, D, E
Explanation
Amazon SageMaker Ground Truth allows you to manage data labeling tasks efficiently. Using Amazon Mechanical Turk as the workforce enables scalable human labeling for your dataset. To ensure high-quality labels, you must define clear task instructions, monitor the labeling process, and review the results. Additionally, customizing the labeling UI can improve usability and accuracy for complex tasks. Automatic labeling tools like Amazon Rekognition can assist but are not a replacement for human labeling in tasks requiring nuanced judgment.
- A. Correct.
Correct. Creating a labeling job in Amazon SageMaker Ground Truth with Amazon Mechanical Turk as the workforce is a key step in enabling human annotators to label your dataset.
- B. Correct.
Correct. Providing clear instructions and examples ensures that annotators understand the labeling requirements, which is critical for obtaining high-quality labels.
- C. Incorrect.
Incorrect. While Amazon Rekognition can be used for some image labeling tasks, it cannot fully replace human labeling when specific or complex labels (e.g., damage categories) are required.
- D. Correct.
Correct. Monitoring the labeling job and reviewing the labeled data are best practices to ensure the quality of the annotations and address any issues promptly.
- E. Correct.
Correct. A custom labeling UI template can be created in SageMaker Ground Truth to meet the specific requirements of your labeling task if the default UI is not sufficient.