The Video Intrinsic Integrity and Distortion Evaluation Oracle (VIIDEO) is a knowledge-based blind video quality assessment (BVQA) model designed to evaluate the perceptual quality of videos without requiring any reference videos, anticipated distortion types, or training on human opinion scores.

This repository is an unofficial Python/PyTorch implementation of the algorithm.

Installation

To set up the environment and install dependencies, run the following commands:

conda create --name viideo_env python=3.9
conda activate viideo_env

Optional: Install PyTorch with GPU support

For NVIDIA GPU (CUDA):

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

For Apple M-series GPU (MPS) or CPU-only:

conda install pytorch torchvision torchaudio -c pytorch

After installing PyTorch, install the remaining libraries.

conda install -c conda-forge scipy opencv

Usage

To analyze a video, run:

python run_analysis.py -v <video_path>

This repository includes two sample videos (from the KoNViD-1k VQA Database):

• good_video.mp4
• bad_video.mp4

Example output scores (on a MacBook Pro with M2 Pro):

The VIIDEO quality score for 'bad_video.mp4' is: 0.4580
The VIIDEO quality score for 'good_video.mp4' is: 0.6539

Licence

This repository contains an unofficial, third-party Python re-implementation of the VIIDEO algorithm.

The original MATLAB release was created by Anish Mittal et al., Laboratory for Image and Video Engineering (LIVE), UT Austin.
The original MATLAB source code is available on GitHub: https://github.com/utlive/VIIDEO/tree/main
It is distributed under the licence reproduced in LICENSE.

If you use this code in academic work, please cite:

1. A. Mittal, M. A. Saad and A. C. Bovik, “VIIDEO Software Release", URL: https://live.ece.utexas.edu/research/quality/VIIDEO_release.zip, 2014.

2. A. Mittal, M. A. Saad and A. C. Bovik, "A ‘Completely Blind’ Video Integrity Oracle", submitted to IEEE Transactions in Image Processing, 2014.