Merry Christmas!
As I’m done with the research paper, I thought I’d organize the process into a few organized posts, as my previous ones were all over the place. This is part one.
I started this project because I kept running into the same limitation whenever I looked at golf putting analysis: most methods reduce a stroke to a small set of measurements, usually at impact, and then try to explain performance from there.
That approach works reasonably well for large, fast movements like a full swing, but putting is different. The motion is slow, controlled, and very small in amplitude. What actually matters is not just where the club face is at impact, but how consistently and smoothly the entire stroke unfolds over time.
When I looked into existing biomechanics research, I noticed two recurring issues. First, many systems rely on specialized sensor setups or laboratory-grade motion capture. These are accurate, but expensive, restrictive, and difficult to scale. Second, a lot of analysis, especially AI-based analysis, depends on supervised learning, where strokes are labeled as correct or incorrect based on predefined ideas of ideal technique.
For putting, that felt limiting. Two strokes can look mechanically similar and still produce different outcomes due to randomness, surface conditions, or very small disturbances. I wanted to separate how a stroke moves from what happens to the ball.
That led me to frame the problem differently: instead of evaluating strokes using outcome-based labels, I wanted to see whether motion patterns could be discovered directly from the data itself. In other words, can we let the motion speak first, and interpret performance later?
This idea was to treat a putting stroke as a continuous time-series motion with an underlying structure. This became the foundation of the entire project.