MLS-C01 · Question #109
A company uses camera images of the tops of items displayed on store shelves to determine which items were removed and which ones still remain. After several hours of data labeling, the company has…
The correct answer is D. Augment training data for each item using image variants like inversions and translations, build. With a limited dataset of 1,000 images for 10 distinct items, the initial training results for an image classification model were poor.
Question
A company uses camera images of the tops of items displayed on store shelves to determine which items were removed and which ones still remain. After several hours of data labeling, the company has a total of 1,000 hand-labeled images covering 10 distinct items. The training results were poor. Which machine learning approach fulfills the company's long-term needs?
Options
- AConvert the images to grayscale and retrain the model
- BReduce the number of distinct items from 10 to 2, build the model, and iterate
- CAttach different colored labels to each item, take the images again, and build the model
- DAugment training data for each item using image variants like inversions and translations, build
How the community answered
(28 responses)- A4% (1)
- B14% (4)
- C7% (2)
- D75% (21)
Why each option
With a limited dataset of 1,000 images for 10 distinct items, the initial training results for an image classification model were poor.
Converting images to grayscale reduces information (color channels) which is generally detrimental for image classification tasks and does not increase the amount of training data.
Reducing the number of distinct items simplifies the problem but does not address the fundamental issue of insufficient data per class, and it goes against the business need to distinguish 10 items.
Attaching different colored labels would require re-taking all images, which is an operational change rather than a machine learning solution to poor performance on existing data.
A common problem in deep learning with limited datasets is overfitting or poor generalization, and 1,000 images for 10 distinct items is a very small dataset for image classification. Image augmentation, which involves creating new training examples by applying transformations (e.g., rotations, flips, shifts, zooms) to existing images, significantly expands the effective size of the training dataset, helping the model learn more robust features and improve performance.
Concept tested: Overcoming limited data with image augmentation
Source: https://www.tensorflow.org/tutorials/images/data_augmentation
Topics
Community Discussion
No community discussion yet for this question.