Meridian's computer-vision ramp-safety initiative has years of raw security-camera footage of the tarmac but no structured records of past foreign-object-debris incidents. Before this footage can train a detection model, what must the project team do to extract usable value from it?

Large amounts of enterprise data, including video like Meridian's tarmac footage, are unstructured — they have no predefined schema. To extract value for a supervised computer-vision model, the team must apply annotation and feature-extraction techniques, such as labeling foreign-object debris, mishandled bags, and safety-zone intrusions, to convert the unstructured video into labeled, structured training examples. The choice to discard the footage because models can only be trained on structured, tabular data is wrong and reflects a common misconception CPMAI explicitly wants project managers to debunk: unstructured data is not unusable, it just requires preparation. The choice to feed the raw footage directly into the model without preprocessing is wrong because raw video pixels are not directly meaningful training labels — a model needs annotated ground truth to learn what foreign-object debris looks like. The choice to convert the footage to a spreadsheet of timestamps only, ignoring the visual content, over-corrects by discarding the visual signal entirely, which defeats the purpose of a computer-vision use case. The exam-relevant takeaway: extracting value from unstructured resources is a distinct, named skill in the Data for AI domain, and it's the necessary bridge between raw unstructured data and the labeled datasets Domain II's Data Preparation phase requires.