β-MuliVariational AutoEncoder for Entangled Representation Learning in Video Frames

This code is the official implementation of the following paper:

β-Multivariational Autoencoder (βMVAE) for Entangled Representation Learning in Video Frames

Exemplary result

These video sequences are from DAVIS16 dataset.(a.Image, b.Annotation, c.βMVUnet)

Install

To create the environment, follow these commands:

conda create -n bmvae python=3.8
conda activate bmvae
pip install -r requirements.txt

This code is developed and tested on Ubuntu OS 18.04.5.

Predict

• First, download the network weights from the Google Drive Link. and put them on the ckpts folder.
• Run the following command for testing βMVAE network:

python test_bmvae.py

• Run this command for testing βMVUnet network:

python test_bmvUnet.py

Dependencies

• U-Net: Semantic segmentation with PyTorch
• Samples are from DAVIS2016 dataset
• Training data are preprocessed sequences from YouTube-VOS dataset

Cite

Please cite our work as:

arXiv:2211.12627 (cs)

Contact

Fatemeh Nokabadi