AI-102 Question 138
Select 3You are building an image classification solution using Azure Custom Vision. Your goal is to create a model that can classify images into different categories. During the data preparation phase, you need to ensure your dataset is properly labeled. Which steps should you take to correctly label your images and ensure optimal model performance?
- A
Assign a single label to each image, even if the image contains multiple objects.
- B
Ensure that all images are organized into separate folders corresponding to their labels.
- C
Use consistent and descriptive labels for all images in your dataset.
- D
Include a diverse set of images for each label to account for variations in real-world scenarios.
- E
Rely on automatic labeling without reviewing the results for accuracy.
Show answer and explanation
Correct answers: B, C, D
Explanation
Properly labeling and organizing your dataset is a critical step in building an image classification solution with Azure Custom Vision. Organizing images into labeled folders, using consistent and descriptive labels, and ensuring diversity in your dataset are essential practices to optimize model accuracy and generalization. Relying on automatic labeling without verification or oversimplifying labels can negatively impact your solution's performance.
- A. Incorrect.
Assigning a single label to images with multiple objects may oversimplify the dataset and result in poor model performance. It is important to use appropriate multi-label strategies if images contain multiple objects.
- B. Correct.
Organizing images into folders corresponding to their labels is a best practice for managing datasets and is required during upload to Azure Custom Vision.
- C. Correct.
Using consistent and descriptive labels ensures the model can correctly map input images to the appropriate categories during training.
- D. Correct.
Including diverse images for each label helps the model generalize better and improves its ability to handle real-world variations.
- E. Incorrect.
Relying solely on automatic labeling without reviewing the results can introduce errors into the dataset, leading to inaccurate model predictions.