Case context: A logistics company is trying to build a vision system for a warehouse. The first version uses one end-to-end model to decide from raw images whether a parcel contains a fragile item, but the team only has a small labeled dataset and accuracy is poor.

Question: Using the idea of decomposing a pipeline, what should the team change to reduce the amount of labeled data they need and improve results?

Sample answer: The team should split the job into smaller stages. One stage can detect whether a parcel is present and properly framed, and a later stage can determine whether the parcel contains a fragile item. Each stage is easier than the full end-to-end problem, so each can usually be trained with less labeled data than a single model that tries to solve everything at once.

Key points:

• One end-to-end model is harder to train with limited data
• Splitting the problem creates simpler subproblems
• Each subproblem can be learned with less labeled data
• The pipeline directly addresses the data shortage

Rubric: Full credit for recommending a staged pipeline and explaining that the smaller tasks need less data. Partial credit for suggesting a split without explaining why it helps.