demalenk/caltennis
CalTennis: Large Multi-View Tennis Video Dataset CalTennis is a large-scale video benchmark designed for evaluating monocular-to-3D human pose estimation in the wild. The dataset comprises over 11 million frames (51 hours) of tennis practice and match play from 40 players, captured with 2–6 synchronized cameras at 60Hz. It is 10x larger than existing in-the-wild human motion video datasets and offers the first large-scale benchmark for synchronized multi-view recordings of… See the full description on the dataset page: https://huggingface.co/datasets/demalenk/caltennis.
CalTennis: Large Multi-View Tennis Video Dataset
CalTennis is a large-scale video benchmark designed for evaluating monocular-to-3D human pose estimation in the wild.
The dataset comprises over 11 million frames (51 hours) of tennis practice and match play from 40 players, captured with 2–6 synchronized cameras at 60Hz. It is 10x larger than existing in-the-wild human motion video datasets and offers the first large-scale benchmark for synchronized multi-view recordings of expert athletic motion.
Dataset Highlights
- Scale: 11.03M frames across 51 hours of unscripted motion.
- Multi-View: 2–6 synchronized views per session, enabling label-free evaluation via multi-view consistency.
- Expert Motion: Focuses on high-speed, articulated tennis strokes (serves, volleys, sprints).
- Long Range: 90% of poses are 13.4–16.7m from the camera, testing depth estimation at greater distances than laboratory benchmarks like Human3.6M.
- Privacy: All faces are blurred to protect player privacy, and all data was collected with IRB approval and informed consent.
Repository Structure
This repository contains the "full" dataset and a "mini" subset for faster iteration.
.
├── 01_23_2026_17_00_court2/ # Session folders/videos
├── camera_calibration/ # Intrinsics and extrinsics
├── metadata_mini.jsonl # Index for mini split
├── metadata_mid.jsonl # Index for mid split
└── README.md