Scenario: A team is building a vision system that flags damaged suitcases on a baggage conveyor. The prototype works, but its error rate is still too high for deployment. In a design review, some teammates argue for gathering additional labeled images, while others suggest changing the model itself.

Question: What broad kinds of fixes should the team consider on the data side and on the model side, and what kinds of image variety would help the training set become more useful?

Sample answer: The team should look at two categories of changes.

Data-side changes:

• Add more labeled examples.
• Expand the dataset so it covers a wider range of real situations.

Examples of useful variety include suitcases photographed from different angles, in bright and dim lighting, partly blocked by other bags, and with different backgrounds or camera settings.

Model-side changes:

• Increase the model’s capacity by using a larger network.
• Decrease capacity if the model is overfitting.
• Add regularization.
• Try a different network design.

Key points:

• Separate improvements to the training set from improvements to the network itself.
• Give concrete examples of broader data coverage, such as angle changes, lighting changes, occlusion, backgrounds, and capture settings.
• Recognize common model adjustments, including changing network size, adding regularization, or switching architectures.

Rubric: The answer must identify both dataset-related options and model-related options, and it must include several concrete examples of training-data variety that would make the image set more representative.