Temporal Tennis Ball Detection with Ball Diameter Prediction

Developed a temporal tennis ball detection model that uses multi-frame input to detect the ball and predict its diameter directly from the detection head.

The model extends a fixed-bounding-box ball detector into a unified single-stage detection pipeline that predicts class, bounding box, and ball diameter together. Previously, ball size estimation required a separate downstream cropped-ball model, but this project integrated diameter prediction directly into the main YOLO-based detection model.

I added a dedicated diameter regression branch using DFL-style supervision, updated the training and validation pipeline, added new loss components for diameter learning, and modified the prediction and NMS workflow so diameter outputs are handled separately from class predictions.

The project also included dataset generation from temporal ball annotations, triplet-frame visualization, model training, validation, inference testing, ONNX/TorchScript export, edge quantization, and active learning workflows for reviewing, correcting, and relabeling model outputs from real tennis videos.

This model improves the tennis perception pipeline by producing both ball location and ball size in a single inference pass, which is useful for downstream 3D trajectory estimation, camera-depth reasoning, and real-time tennis analytics.