Ball tracking
Follow the ball.
Understand the play.
EyeBall follows the basketball across game film. A Kalman filter stitches noisy detections into a smooth trajectory; its shape tells you when passes happen. Layer in other signals -- dribbles, player position, clock state -- and you get actions, and from actions, plays: the coaching surface the box score never captures. The same primitive also drives a synthetic broadcast camera for amateur footage, no operator required.
Case 01 -- Tight tracking
Jordan, drive to the rim.
A classic crossover-and-finish. The ball changes direction and speed twice per second and briefly disappears behind MJ's body at the rim. The system holds the trajectory through all of it -- the kind of possession the box score never explains. That's the starting point for any downstream analytics: if you can't follow the ball, you can't measure the play.
Case 02 -- Possession breakdown
Spurs at Houston, half-court.
One half-court possession. The system watches the ball move between players and stamps a timestamp every time the trajectory changes direction sharply enough to register as a pass. It also over-counts here -- classical CV has no way to distinguish a real pass from a crisp dribble reversal. That's the natural limitation of a ball-only signal.
Passes are a starting point, not the answer. Combine them with dribble detection, player position, and shot-clock state and you get action detection -- pick-and-roll, hand-off, off-ball screen, isolation -- and from there, play detection. That's the coaching surface I'm building toward: the box score never tells you which actions actually got run.
Under the hood, no training data required: HSV + contour detection, Kalman smoothing, outlier rejection, and an angle-based pass detector with a cooldown gate.
Detection
HSV + contours
Smoothing
Two-pass Savgol + outlier rejection
Events
Angle-based pass detection
Output -- Houston half-court run
Passes detected
7
Detection rate
70%
294 / 420 frames
Clip duration
14.0s
Trajectory
4,796px
total path length
Pass timeline (seconds)
Downstream application
Virtual panning.
No operator, still a broadcast.
Almost no amateur game has a camera operator. A stationary wide panorama turns into a real broadcast by tracking the action and rendering a synthetic 16:9 pan over it. Motion inside a court ROI drives the pan target; Savitzky-Golay keeps the crop from jittering.
Source
Output
Engineering note
No training data? Generate it with NBA 2K.
Good labelled basketball film -- with the ball picked out on every frame -- is rare. Before classical CV got the current pipeline running, I trained the early ball-detection models on synthetic data generated from NBA 2K: realistic geometry, controllable camera, and a game engine I could produce labelled frames from on demand. Not a shortcut to production, but a legitimate proof of concept and a reliable test bed wherever a real labelled set doesn't exist yet.
Under the hood
Stack.
- DetectionHSV + contours (live); YOLO (production upgrade)
- State estimationKalman filter with outlier rejection on a speed gate
- SmoothingTwo-pass Savitzky-Golay with linear interpolation through gaps
- EventsAngle-based pass detection with cooldown + min-motion gate
- BroadcastingCourt-ROI motion envelope, Savgol-smoothed 16:9 crop
- StackPython 3.12, OpenCV 4.10, NumPy, SciPy, imutils