Is there a way to automatically say how warped a pool table is, just by looking at a video of people playing? In this talk I will explain how I did that using OpenCV for pool balls tracking, and how more heuristic approaches may improve this model’s performance. I will also present a simple analysis of the effect of different levels of warp on players’ scores.
Imagine your workplace has a pool table and runs a pool tournament that you’re very excited about. You track the games, but it turns out there’s an extra factor at play: the table is pretty uneven. How would you quantify its ‘wonkiness’? Does it even matter for the scores?
In this talk I will describe how I used a combination of off-the-shelf computer vision tools like OpenCV’s multiple object trackers, trained models for object detection, and simple geometry to achieve a rating for a table, based on a video of gameplay. I will then explain why the standard models may not perform as well as may be expected, but heuristic approaches that incorporate some of the domain may help improve performance. Finally, I will talk about insights gained from comparing these ratings with scores from real games, and how we can (and why we probably shouldn’t) use it in a scoring model.