trainmatricx/cricket-bat-keypoint-detection

Cricket Bat Keypoint Detection Dataset (COCO format)

A small, human-labelled dataset for cricket bat keypoint detection: 177 consecutive video frames of a batting stroke, each with 4 keypoints along the bat and a bounding box, in COCO keypoint format. Labelled by Train Matricx, a sports data annotation team.

Annotations only. The frames come from a free Pexels stock video that is not ours to relicense, so images are not included. A script in this repo rebuilds them from the original video in one command (see How to get the images).

At a glance

Task	Keypoint detection (bat pose), object detection (bat box)
Frames	177, consecutive, from one stroke
Objects	1 cricket bat per frame
Keypoints	4 per bat: handle-top, handle-bottom, bat-middle, bat-bottom
Total keypoints	708, all marked visible
Format	COCO keypoints (annotations.coco.json)
Resolution	1080 × 1920 (portrait), native video coordinates
Scene	Outdoor net practice, one batter, fixed camera
Licence	Annotations and code: MIT. Video: Pexels License (not included)
Labelled by	Train Matricx sports-domain annotators

What can you use it for?

• Bat tracking and bat-angle estimation: follow the bat through backlift, downswing and follow-through.
• Batting stroke analysis: bat path and swing timing for coaching tools.
• Testing a keypoint model on a thin, fast-moving object: a cricket bat is narrow, often motion-blurred and partly hidden by the hands, which is hard for generic pose models.
• Checking an annotation vendor's quality: inspect real production labels before sending your own footage.

It is a sample, not a benchmark: one batter, one camera angle, one stroke. Use it to prototype, evaluate a pipeline or fine-tune on a tiny set, not to report accuracy on cricket in general.

Keypoint definitions

#	Keypoint	Where it is placed
1	handle-top	top end of the handle
2	handle-bottom	where the handle meets the blade (the splice)
3	bat-middle	middle of the blade
4	bat-bottom	toe of the bat

Skeleton: handle-top → handle-bottom → bat-middle → bat-bottom. Each annotation also includes a bounding box around the whole bat.

Files

File	Contents
annotations.coco.json	COCO keypoint annotations for 177 images
frame_map.csv	Maps each file_name to its 0-based frame number in the source video
rebuild_frames.py	Extracts the labelled frames from the source video

How to get the images

1. Download the 1080 × 1920 version of Dynamic Cricket Practice Session Outdoors by Gokuldham Nar on Pexels.
2. Run:

pip install av
python rebuild_frames.py path/to/video.mp4 images/

This writes images/frame_0000.jpg to images/frame_0183.jpg, matching the file_name entries in annotations.coco.json.

Load the annotations

import json

coco = json.load(open("annotations.coco.json"))
names = {img["id"]: img["file_name"] for img in coco["images"]}

for ann in coco["annotations"][:3]:
    k = ann["keypoints"]
    points = [(k[i], k[i + 1]) for i in range(0, len(k), 3)]
    print(names[ann["image_id"]], points)

The file also works with pycocotools and with any training framework that reads COCO keypoints.

How it was labelled

Keypoints were placed by Train Matricx annotators who know cricket. They were originally drawn on 640 × 640 resized frames and have been converted back to the video's native 1080 × 1920 coordinates for this release.

FAQ

What is cricket bat keypoint detection? It means locating fixed points on the bat (here the handle top, splice, blade middle and toe) in each frame, so a model can recover the bat's position and angle through a stroke. It is the bat equivalent of human pose estimation.

Why label keypoints instead of only a bounding box? A box says where the bat is but not which way it points. Four keypoints along the bat give its angle, length in frame and path, which is what stroke and swing analysis needs.

Can I use this commercially? Yes for the annotations and code (MIT). The video is under the Pexels License, so download it from Pexels and follow that licence.

Why are the images not included? The footage was filmed by a Pexels creator, not by Train Matricx. Publishing our labels with a rebuild script respects their rights while keeping the dataset fully usable.

Can I get more cricket or sports annotation like this? Yes. Train Matricx labels player and ball tracking, event tagging, skeletal keypoints and court registration across 20+ sports, with a free pilot on your own footage: trainmatricx.com/pricing.

License and credit

• Annotations (annotations.coco.json, frame_map.csv) and code (rebuild_frames.py): MIT, © 2026 Train Matricx.
• Video and frames: not included and not covered by this licence. Available under the Pexels License. Credit: Gokuldham Nar, Pexels.

Citation

@misc{trainmatricx2026cricketbat,
  title        = {Cricket Bat Keypoint Detection Dataset},
  author       = {{Train Matricx}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/trainmatricx/cricket-bat-keypoint-detection}}
}

About Train Matricx

Train Matricx produces human-verified training data for sports computer vision: player and ball tracking, event tagging, skeletal keypoints and court registration across 20+ sports.

• Website: trainmatricx.com
• Guide: How to build a sports computer vision dataset
• Hugging Face: huggingface.co/trainmatricx