PadelTracker100: A dataset for intelligent player and ball tracking in padel sports.

Abstract

Advancements in computer vision and deep learning have revolutionized sports video analysis, enabling automated and precise data labeling. However, these technologies rely heavily on high-quality annotated datasets, which are essential for training supervised learning models. This article introduces PadelTracker100, a large-scale fully annotated dataset for padel, a rapidly growing sport, captured in a professional setting. The dataset, derived from two matches of the 2022 World Padel Tour (WPT) Finals at 1920 × 1080@30, contains nearly 100,000 frames with a single standard camera angle to reduce occlusions and ensure clarity. Annotations include ball trajectory tracking, real-world player positions, player pose estimation and shot event recognition, categorized into six classes: backhand, forehand, smash, serve, dropshot, and other. Ball trajectory annotations were generated using a semi-automatic pipeline with iterative YOLO training. Pose estimation was carried out with ViTPose-L, selected after comparing various state-of-the-art models. Shot events were annotated for 40,135 frames. A thorough manual refinement process was applied, ensuring annotation quality trough all annotation types. The lack of pre-annotated datasets has significantly restricted large-scale match analysis and the development of automated techniques in padel, hindering the progress of AI-driven solutions tailored for the sport. By addressing this gap, this dataset serves as a comprehensive benchmark for padel, fostering advancements in various applications such as cross-sport analysis, injury prevention, tactical evaluation, ball trajectory modelling, and real-time video processing.