Hockey3D: A Public Field Hockey Shot Trajectory Dataset with 3D Reconstruction and Shot Type Classification

Abstract

While sports analytics has advanced significantly for popular sports, many domains remain underserved, particularly field hockey, which lacks comprehensive studies on shot statistics and trajectory analysis. We present a generic computer vision and machine learning pipeline applicable to any sports broadcast video, demonstrated specifically for field hockey shot analysis. Our contributions are threefold: (1) 3D Trajectory Reconstruction: We employ PnLCalib for camera calibration, conduct extensive ablations across detection methods (YOLO, RF-DETR, BlurBall, WASB) with Hough Circle, and formulate 3D localization as a Bayesian optimization problem—assuming parametric 3D trajectories, we project them to 2D image space and iteratively refine parameters to match observed detections. (2) Shot Classification with Comprehensive Ablations: We evaluate multiple ML/DL architectures including temporal models (TCN, LSTM, BiLSTM, GRU, Transformer) and classical methods (XGBoost, Random Forest, SVM, KNN) on seven hockey shot types (Hit/Drive, Slap Shot, Push Shot, Flick, Scoop, Tomahawk/Reverse Hit, Sweep Hit), achieving 96.4% accuracy with our TCN architecture. (3) Real-World Dataset: We make publicly available a dataset extracted from broadcast videos with annotated 3D trajectories and shot labels, addressing critical data scarcity in field hockey analytics. Our modular pipeline achieves sub-meter 3D localization accuracy and provides a reusable framework for trajectory analysis across different sports.

Ritabrata Chakraborty

ML Intern

Research Engineer specializing in robotics, computer vision, and autonomous systems. Currently developing automated data annotation solutions with foundation models for autonomous vehicles at Uber.