Computer Vision
A computer-vision pipeline that tracks a VR table-tennis ball and reconstructs each rally from a stabilized top-down view.
View repository1,100+
labels per dataset
2
YOLO pipelines
2D
court projection
TableTennisAnalysisVR turns first-person footage from Eleven Table Tennis into an analysable overhead replay. Separate detectors locate the fast-moving ball and the table keypoints, then geometric reconstruction maps the observations onto a clean two-dimensional court.
VR footage combines rapid ball motion, motion blur, occlusion, viewpoint changes, and table corners that frequently leave the frame. The pipeline had to recover a stable playing surface while keeping a tiny, high-speed target visible across the rally.
The team produced two custom annotated datasets with more than 1,100 images each, augmented them with Roboflow, and trained YOLOv8l for ball detection and YOLO11l for table-keypoint detection. The final notebook accepts match video and produces a top-view trajectory reconstruction.
Technical approach
Extract and label VR match frames for independent ball and table-keypoint datasets.
Detect the ball with YOLOv8l and the court geometry with a YOLO11l keypoint model.
Reconstruct missing table corners and apply a perspective transform to project the rally into a top-down court.