Homeground App · Nov 2022 – Jan 2023
A lightweight MobileNet-V2 keypoint pipeline detecting human joints and cricket bats at about 81% accuracy and 0.73 F1, with shot classification, ball tracking and Kalman-filtered analytics for coaching insights. The visualisation below follows one delivery from frame to classified shot.
01 The scaffold
A video frame beside the MobileNet-V2 backbone's shrinking feature maps, a detection canvas with the joints barely showing, a Kalman panel with its predict and update legend, and an empty shot card. The pipeline is laid out end to end before anything runs.
02 One delivery
The frame is scanned, feature maps light up, keypoint heatmaps bloom and the skeleton draws on joint by joint with the bat boxed at 0.94. The ball rides its projectile arc while the covariance ellipse tightens with each measurement, and the shot card lands on cover drive at 0.91 confidence.
The pipeline drawn above: a MobileNet-V2 backbone with depthwise-separable convolutions downsampling 224×224 input through 112², 56² and 28² feature maps, keypoint heatmaps decoded into a 17-joint skeleton with a cricket-bat bounding box, and a Kalman filter tracking the ball, alternating the predict step x̂⁻ = Fx̂ with measurement updates that shrink the covariance ellipse along the projectile arc. A softmax over 12 shot classes reads the sequence and lands on cover drive.
The first scroll lays out the scaffold. The second animates the inference on one delivery, the same pass the app ran in real time at roughly 60 frames per second.