DeepDarts

Note: This repository contains fine-tuned models based on the original DeepDarts project. Many thanks to wmcnally and the original authors for their excellent foundational work — this project wouldn't be possible without it.

Code for the CVSports 2021 paper: DeepDarts: Modeling Keypoints as Objects for Automatic Scorekeeping in Darts using a Single Camera

Prerequisites

Python 3.5-3.8, CUDA >= 10.1, cuDNN >= 7.6

Setup

1. Install Anaconda or Miniconda
2. Create a new conda environment with Python 3.7: $ conda create -n deep-darts python==3.7. Activate the environment: $ conda activate deep-darts
3. Clone this repo: $ git clone https://github.com/MikeMcGarry/deep-darts.git
4. Go into the directory and install the dependencies: $ cd deep-darts && pip install -r requirements.txt
5. Download images.zip from IEEE Dataport and extract in the dataset directory. Crop the images: $ python crop_images.py --size 800. This step could take a while. Alternatively, you can download the 800x800 cropped images directly from IEEE Dataport. If you choose this option, extract cropped_images.zip in the dataset directory.
6. Download models.zip from IEEE Dataport and extract in the main directory.

Validation / Testing

To test the Dataset 1 model:
$ python predict.py --cfg deepdarts_d1 --split test

To test the Dataset 2 model and write the prediction images:
$ python predict.py --cfg deepdarts_d2 --split test --write

Training

To train the Dataset 1 model:
$ python train.py --cfg deepdarts_d1

To train the Dataset 2 model:
$ python train.py --cfg deepdarts_d2

You may need to adjust the batch sizes to fit your total GPU memory. The default batch sizes are for 24 GB total GPU memory.

Sample Test Predictions

Dataset 1:

Dataset 2:

Pipeline

• Take photo
• Save to images/utrecht_{month}{day}{year}
• Run "annotate" on folder above using --img-folder
• Annotate images manually
• Run "combine_labels"
• Run "crop_images" with --size=800 and appropriate --labels-path