Rate and Distortion: A toolkit for video codec comparison

R&D is a Python library that faciliates multimedia codec comparison. It provides functions to perform rate-distortion analysis for a given codec with sparse samples in the rate-distortion space.

Features

• Efficiency: R&D provides the implementation of state-of-the-art rate-distortion function models, which enables efficient codec comparison with only sparse encoding samples.
• Accuracy: R&D is demonstrated to be more accuracy than the traditional BD-rate measure.
• Flexibility: R&D can be applied to analyze the encoder performance at different resolution, frame rate, and bit depth, with a unified interface.
• Robustness: R&D curve/surface fitting models are mathematically stable.
• Generalizability: R&D is compatible with most of the modern video quality assessment models such as PSNR, VMAF, and SSIMPlus.
• Extrapolation Capability: R&D can deliver reasonable codec performance analysis in the region where no encoding samples are provided.

Installation

R&D can be installed using pip or from source.

Install from source:

$ git clone git@github.com:UWIVC/randd.git
$ python setup.py install

Install from pip:

$ pip install randd

Sample Usage

1D Example

The following example illustrates how to use R&D to compare two video codecs with distortion measure being PSNR:

import randd as rd
import numpy as np

r1 = np.array([100., 400., 1500., 6000.])
d1 = np.array([23.83, 25.89, 31.28, 38.22])
r2 = np.array([100., 400., 1500., 6000.])
d2 = np.array([16.25, 29.75, 33.52, 39.12])
analyzer = rd.Analyzer(d_measure='psnr')
quality_gain, bitrate_saving, summary = analyzer(r1, d1, r2, d2, codec1='h264', codec2='vp9')

2D Example

The following example illustrates how to use R&D to compare two video codecs with distortion measure being VMAF. We can also specify the bitrate region [100, 3000] to compute the quality gain. Note that R&D can perform reasonable extrapolation to the region where no RD samples are given.

import randd as rd
import numpy as np

# 1st dim: bitrate, 2nd dim: diagonal size of spatial resolution
r1 = np.array([[2400.,  400.], [1700.,  865.], [ 300., 1469.], [2100., 2203.]])
d1 = np.array([28.39, 71.87, 26.98, 89.35])
r2 = np.array([[2400.,  400.], [1700.,  865.], [ 300., 1469.], [2100., 2203.]])
d2 = np.array([28.86, 75.03, 54.35, 96.49])
analyzer = rd.Analyzer(d_measure='vmaf', ndim=2, r_roi=[100, 3000])
quality_gain, bitrate_saving, summary = analyzer(r1, d1, r2, d2, codec1='h264', codec2='vp9')

Backward Compatibility

R&D also provides implementation of traditional codec comparison tools such as BD-rate:

import randd as rd
import numpy as np
from randd.model import LogCubic

r1 = np.array([100., 400., 1500., 6000.])
d1 = np.array([23.83, 25.89, 31.28, 38.22])
r2 = np.array([100., 400., 1500., 6000.])
d2 = np.array([16.25, 29.75, 33.52, 39.12])
analyzer = rd.Analyzer(d_measure='psnr', model1=LogCubic, model2=LogCubic)
quality_gain, bitrate_saving, summary = analyzer(r1, d1, r2, d2, codec1='h264', codec2='vp9')

Documentation

The full documentation is available at https://randd.readthedocs.io/.

Citation

We are making R&D publicly available. If you use R&D in your project, we kindly ask you to cite the following paper:

@article{duanmu2020egrd,
    title={Characterizing Generalized Rate-Distortion Performance of Video Coding: An Eigen Analysis Approach},
    author={Duanmu, Zhengfang and Liu, Wentao and Li, Zhuoran and Ma, Kede and Wang, Zhou},
    journal={IEEE Transactions on Image Processing},
    volume={29},
    number={},
    pages={6180-6193},
    year={2020}
}

Contacts

Wentao Liu - @w238liu - w238liu@uwaterloo.ca

Zhengfang Duanmu - @zduanmu - zduanmu@uwaterloo.ca