How AI Traces a Golf Shot — Inside Shashot's Vision Pipeline

Drawing the trajectory of a golf shot onto a video is harder than it looks. The ball is small, moves fast, and tiny pixels easily get swallowed by camera shake and the texture of grass and sky. The reason Shashot can still draw a clean trajectory from a single take is that the vision pipeline is split into four stages.

1. Stabilization — taming the shake

Video analysis starts with consistency between frames. Phone footage always has a little jitter, so we first align the background using optical flow and feature matching. If this step is weak, we end up tracking the camera shake instead of the ball.

2. Detection — finding the ball, club, and hand

To find the small white ball, we combine three signals at once: color, shape, and motion. Single-frame detection often fails, so we fuse spatio-temporal cues across multiple frames to pick high-confidence candidates. Club and hand are handled by a separate keypoint model.

3. Tracking — stitching candidates across time

We connect frame-by-frame candidates into a single curve, while applying physical constraints (speed and acceleration limits) to reject noisy candidates. The result is a smooth curve rather than a jagged chain of points.

4. Rendering — making it look good

We composite the final curve back onto the video with thickness, color, and afterimage effects. Different filter styles produce different looks, and you can export straight to an Instagram-ready aspect ratio.

Why this shape?

You could try to do this end-to-end with a single model. But golf videos often have segments where the ball is just one or two pixels, and shooting conditions vary wildly per user. Separating each stage and validating its hypothesis is far more robust — and it lets the user tune each step individually if they want to.

In short, Shashot takes a single clip and runs it through stabilization → detection → tracking → rendering. Each stage leans on the previous one, and each leaves room for the user to nudge the result. It is not fully autonomous, but it is fast, predictable, and forgiving.