Skip to content

Latest commit

 

History

History
46 lines (30 loc) · 1.83 KB

README.md

File metadata and controls

46 lines (30 loc) · 1.83 KB

DeepDarts

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/wmcnally/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:
alt text

Dataset 2:
alt text