tsagkas / capsules

Inference for Generative Capsule Models

Inference and Learning for Generative Capsule Models

This code is the implementation of the algorithm described in the following paper: Inference and Learning for Generative Capsule Models .

This repo contains the code for the "face experiments". The original constellation experiments can be found at: anazabal/GenerativeCapsules.

Figure 1: Reconstruction demo of randomly transformed faces in a given scene.

Instructions

1. Setup

a. Install manually the following dependecies:

• opencv-python>=4.5.5.62
• scikit-learn>=1.0.2
• matplotlib>=3.5.1
• pandas>=1.3.5
• monty>=2022.1.19
• numpy>=1.21.5

b. To generate the dataset run the following script:

./dataset/create_dataset.sh

NOTE: this step will take several hours since the code has to:

• generate 100,842 synthetic face images,
• train 5 PPCA models,
• train the FA model,
• generate 100,842 x 5 appearance labels.

2. Running Experiments

To run the algorithm, select the number of faces that will exist in the scene and execute the following command:

python -m main --num_faces=3

Results for our variational inference algorithm and the RANSAC-based approach will be saved in the ./results directory.

3. Citation

4. Notes

The code was tested on Ubuntu 20.04.4 with python 3.7.10. All the part-images were downloaded from the PhotoFitMe project page.