Issa Loubani
← back

Real-time gameplay analysis

YOLOSegmentationReal-timePyTorch

Two models, same idea. One reads Dying Light 2 and identifies geometry the player can grab or climb, then suggests when to jump or use the rope. Segmentation rather than bounding boxes, because a ledge is a shape, not a rectangle.

The other reads Call of Duty and tracks enemy positions, then suggests cover.

Both were trained on footage I captured myself and run on a GTX 1050 Ti, which forced most of the interesting decisions. The constraint is the one I hit at work every day: a prediction that arrives after the moment has passed is worth nothing.