All projects

Frame to Game

A seven-class image classifier that identifies games from a single gameplay frame.

Built a classifier from scratch to identify six video games or other content. A data pipeline processed 683 gameplay clips into 11,055 cleaned frames, with recordings split by person and session.

On 9,879 frames from 14 held-out players, the model achieved 85.1% accuracy and 0.725 macro-F1, compared with 39.7% accuracy and 0.319 macro-F1 for a colour-histogram baseline.

Source