MLS-C01 Question 91
Single answerYou are working on a machine learning project to classify images of damaged and undamaged cars for an insurance company. Your dataset contains thousands of unlabeled images. To train your model effectively, you need to label these images as either 'Damaged' or 'Undamaged'. Which approach should you take to efficiently label this dataset using AWS services?
- A
Use Amazon Mechanical Turk to create a labeling task and recruit a workforce to label the images.
- B
Manually label the images by downloading them locally and adding labels using image editing software.
- C
Deploy an Amazon SageMaker Ground Truth labeling job and use Mechanical Turk as the workforce.
- D
Use Amazon Rekognition to automatically label the images without any human involvement.
Show answer and explanation
Correct answer: C
Explanation
The most efficient solution for labeling large datasets in this scenario is to use Amazon SageMaker Ground Truth. It allows you to create labeling jobs and supports integration with Amazon Mechanical Turk for crowd-sourced labeling. This approach is specifically designed for machine learning workflows, making it the correct choice. While Amazon Rekognition is powerful, it is not suitable for custom labeling tasks without prior training.
- A. Incorrect.
Amazon Mechanical Turk allows you to recruit a workforce for labeling tasks, but it doesn’t provide the built-in tools and integrations specific to machine learning workflows that Amazon SageMaker Ground Truth offers.
- B. Incorrect.
Manually labeling thousands of images is highly time-consuming and impractical for large datasets, making this approach inefficient.
- C. Correct.
Amazon SageMaker Ground Truth allows you to create labeling jobs efficiently, and you can integrate Mechanical Turk as a workforce to label the images. This is the most efficient and scalable solution for the given scenario.
- D. Incorrect.
Amazon Rekognition is a pre-trained computer vision service. While it can identify objects or features in images, it cannot automatically label custom categories like 'Damaged' or 'Undamaged' without prior training on your specific dataset.