Bitrate-Ladder-Benchmark

Benchmarking Learning-based Bitrate Ladder Prediction Methods for Adaptive Video Streaming

This repository contains the code for our paper on Benchmarking Learning-based Bitrate Ladder Prediction Methods for Adaptive Video Streaming. If you use any of our code, please cite:

@article{Telili2022,
  title = {Benchmarking Learning-based Bitrate Ladder Prediction Methods for Adaptive Video Streaming},
  author = {Ahmed Telili, Wassim Hamidouche, Sid Ahmed Fezza, and Luce Morin},
  year = {2022}
}

• Requirements

• Features extraction

  • Handcrafted features
  • Deep features

• Model Training

  • Handcrafted features based models
  • Deep features based models

• Performance Benchmark

Requirements

pip install -r requirements.txt

Features extraction

a- Handcrafted features:

python features_extration  [-h] [-r 'path to raw videos directory']
                                   [-f 'path to meta-data csv file']
                                   [-o 'overlapping between patches']

b- Deep features:

python features_extration  [-h] [-v 'path to raw videos directory']
                                   [-f 'path to meta-data csv file']
                                   [-np 'number of patches']
                                   [-nf 'number of frames']
                                   [-m 'backbone model']
                                   [-o 'overlapping between patches']

Please note that we provide four pretrained backbone models for features extraction: resnet50, densenet169, vgg16 and inception_v3.

Model Training :

a- Handcrafted features:

Training can be started by importing Bitrate_Ladder.ipynb in Google Colab or Jupyter Notebook.

b- Deep features:

python train.py  [-h] [-v 'path to raw videos directory']
                                   [-np 'number of patches']
                                   [-nf 'number of frames']
                                   [-b 'batch_size (1)']

Performance Benchmark:

a-YPSNR quality metric:

Methods \ Scores	R2	SROCC	PLCC	ACCURACY	BD-BR vs GT	BD-BR vs AL	BD-BR vs RL
ExtraTrees Regressor	0.7635	0.8174	0.9000	0.8779	1.433%	-18.427%	-9.025%
XGBoost	0.6165	0.7560	0.8278	0.8578	2.320%	-18.099%	-8.706%
Gaussian Process	0.6390	0.7620	0.8473	0.8566	1.740%	-18.244%	-6.286%
Random Forest Regressor	0.6758	0.7993	0.8440	0.8671	1.535%	-18.324%	-8.879%
Densenet169	0.4725	0.6423	0.7756	0.8166	3.380%	-15.669%	-8.169%
VGG16	0.5172	0.5236	0.7652	0.8223	3.083%	-15.536%	-8.088%
ResNet-50	0.4564	0.5680	0.7457	0.8483	2.424%	-15.806%	-8.300%
EfficientNet B7	0.4237	0.5649	0.7159	0.8004	3.396%	-15.506%	-8.012%

b-VMAF quality metric:

Methods \ Scores	R2	SROCC	PLCC	ACCURACY	BD-BR vs GT	BD-BR vs AL	BD-BR vs RL
ExtraTrees Regressor	0.6420	0.6635	0.8277	0.8400	2.704%	-18.827%	-8.798%
XGBoost	0.5533	0.6470	0.7997	0.8347	3.444%	-18.650%	-8.608%
Gaussian Process	0.4292	0.4918	0.6983	0.8012	5.254%	-18.328%	-7.688%
Random Forest Regressor	0.5899	0.6564	0.8059	0.8300	3.052%	-18.887%	-8.616%
Densenet169	0.4216	0.6167	0.6433	0.7901	3.820%	-15.892%	-7.851%
VGG16	0.4992	0.5112	0.7601	0.8052	4.125%	-15.812%	-7.593%
ResNet-50	0.4045	0.5367	0.6962	0.8278	2.969%	-15.941%	-7.810%
EfficientNet B7	0.3920	0.5612	0.6905	0.7781	4.742%	-15.771%	-7.607%