Detecting Tennis Court Keypoints with Yolo

The goal of this project is to see the performance of YOLOv8 at detecting the 14 keypoints of a tennis court to map the lines and perspective.

Environment Setup

The project uses Python 3.12 and PyTorch. Setup the environment:

python3 -m venv venv
source venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

Dataset Preparation

The dataset consists of tennis court images with 14 annotated keypoints per image.

1. Ensure the dataset archive tennis_court_det_dataset.zip is located in the project root.
2. Run the preparation script to extract the dataset and convert it into the YOLO format required by ultralytics:

python prepare_dataset.py

This will extract the zip file and create a datasets/tennis_data folder inside your project directory containing the YOLO images and labels formatted according to datasets_config/tennis.yml.

Training the Model

To train the YOLOv8-pose model on the tennis court dataset:

python train.py --epochs 10 --batch_size 8

Hardware: Trained on a NVIDIA GeForce RTX 2070 SUPER, Python 3.12, torch 2.6.0+cu124

(Adjust --epochs and --batch_size according to your hardware capabilities. The model uses yolov8s-pose.pt as the base model and saves results to runs/pose/train/ or runs/pose/train2/.)

Evaluation

To evaluate the trained model:

• You can run the evaluate.ipynb Jupyter Notebook.
• This will output the mAP50 and mAP50-95 scores, as well as visualize the predicted keypoints vs ground truth on a test set. It also demonstrates mapping the court perspective to a 2D top-down reference court.

Training Results (Epoch 10, Batch Size 8, Imgsz 640)

After 10 epochs of training the yolov8s-pose.pt base model, the following validation metrics were achieved:

Metric	Score	Description
mAP50(B)	99.50%	Bounding box detection precision (court presence)
mAP50(P)	99.49%	Keypoint prediction accuracy (corners & intersections)
mAP50-95(P)	99.31%	Extremely strict keypoint mapping precision
Recall(B)	100.0%	Zero missed courts in the validation dataset

Video Inference

To test the line and point detection on a live tennis video:

1. Obtain a .mp4 video of a tennis match.
2. Run the inference script, specifying your input video and the path to your best trained weights:

   python inference_video.py --video_path your_video.mp4 --model_path runs/pose/train/weights/best.pt --output_path output.mp4

   This will output a video showing the tennis court lines and keypoints overlaid directly on the original footage.

Extra

Homography View of Court

Acknowledgments

Thanks to and inspired by https://github.com/yastrebksv/TennisCourtDetector. The dataset can also be found at this GitHub repository.