README.md

Card Segmentation

Develop Guide

1. Coding style
   • PEP-8
   • black
   • flake8
2. Pre-commit
   • We integrate pre-commit into our framework to make sure the consistency of coding style. If you use git commit inside docker container, it will check whether coding style is compliant to our requirement or not. Instead of docker env, if you use git commit in another env, you may need to run the following command first:

     cd card-segmentation/
     pip3 install pre-commit==2.6.0
     pre-commit install --install-hooks

Setup

1. Install docker, docker-compose and nvidia-docker2
2. Clone repo

   git clone https://github.com/NickLi0605/card_segmentation.git

3. Run docker-compose

   cd card_segmentation
   docker-compose up

   Note:
   • This docker image will start jupyter notebook automatically.
   • You may need to replace the volume by your self in docker-compose.yml
4. Install pyenv
5. Create virtual env for python with pyenv

   pyenv install 3.8.3
   pyenv virtualenv 3.8.3 <venv_name> # create virtual env
   pyenv local <venv_name>  # apply virtualenv for the project

6. Install dependencies

   pip install -r requirement.txt

7. Enable pre-commit hooks

   pre-commit install

Action item

• Must to
  • Construct environment with dockers
  • Download dataset
    • Download scripts for midv2019 and midv500
    • Convert to coco format
    • How to donlowad:

      git clone https://github.com/AlexLi0605/midv500
      python run.py --dataset_dir "path to store dataset" --convert_to_coco

      Note: this repo is forked from here and add download links for midv2019
  • Split dataset into training / validation / testing with reasons
  • Basic training & inference code with pre-trained model
  • Apply data augumentation with albumentations
  • Benchmark for different models with
    • Inference time
    • Model size
    • Memory usage
  • Demo video
• Nice to have
  • Refactor others' codes if we use them
  • Analysis the pros and cons for detection2
  • Consider the issues that model may face and how to solve it
  • Idea about model deployment, data collection, data sacing and automatic model-tuning