Deformable Template Estimation Implementation

By: Peeranat (ToTo) Tokaeo

This is an implementation of the deformable template model as described in Allassonnière et al. 2007. The paper is freely available here.

This program is an artifact produced as part of my final year capstone project at Yale-NUS in 2021.

Quick start

• Install dependencies using python -m pip install -r requirements.txt --user. This assumes that your computer has a GPU that supports at least CUDA 10.1.

• Run complete_workflow/main.py with Python.

How to use complete_workflow/main.py file

There are some options to configure at the top of main.py. Most of them should be self-explanatory.

Note that if the COINS variable is set to True, the inputs used will be in complete_workflow/input_coins. Else, complete_workflow/input_data is used. Currently, the data in input_data are images of digits from the MNIST dataset.

Data

Feel free to look through the ancient coins dataset in complete_workflow/input_coins. Only the *.png are used in training and classification.

Adding data

Currently, the data is organized in the following structure.

├───complete_workflow
│   ├───input_coins
│   │   ├───template1
│   │   │   ├───test
│   │   │   └───train
│   │   ├───template4
│   │   │   ├───test
│   │   │   └───train
│   │   ├───...
│   ├───input_data
│   │   ├───template0
│   │   │   ├───test
│   │   │   └───train
│   │   ├───template1
│   │   │   ├───test
│   │   │   └───train
│   │   ├───....
│   ├───...
├───....

The images to be used are in *.png format and are divided into the test and train for cross-validation. Images that are pre-classified into the same class should be in the same template* folder.

To use your own data, simply overwrite the images in the input_data folder and run the program with COINS variable set to FALSE. Note that they should still be a similar format to the default data, i.e. are in *.png, grayscale, etc.

Coin data info

Alexander tetradrachms from Damascus (Glenn 2018),

Glenn, S. (2018). Exploring localities. A die study of Alexanders from Damascus. In: Alexander the Great. A Linked Open World (Bordeaux), 91-126.

Thanks to Prof. Ernst Emanuel Mayer for providing the images.

Sample output data info

A sample output is available at complete_workflow/sample_train_output. Note that this is deliberately incomplete. I removed the saved covariance matrices since they are too large for a Git repo.

Outputs from running the actual script will have these covariance matrices.