Once I had reliable stroke trajectories, the new problem that appeared was that raw motion data is overwhelming.
A putting stroke contains thousands of data points: positions, rotations, and angles that change every frame. Looking at these individual variables over time could be difficult. The challenge was now how to represent this data.
I needed a way to compress an entire stroke into something smaller—something that still preserved its structure, timing, and smoothness, but removed irrelevant detail. Importantly, this representation had to be learned from the data itself, not imposed by assumptions about what “matters” in a putting stroke.
I wanted to compress all this data into something smaller that still preserved the important structure, timing, etc while removing irrelevant information. This ruled out traditional feature engineering. This is probably the turning point where the project wasn’t about physics anymore.