A comprehensive video coding dataset, namely MVCD, contains nearly 4 million records of encoding and decoding processes. It enables the usage of machine learning models on video streaming applications such as bitrate ladder prediction, resource allocation, rate-quality control, and energy-efficient streaming.
Content:
To create MVCD, 1000 video sequences are collected from Inter4K, which have a resolution of 4K and a frame rate of 60 fps. The input videos can be downloaded from: https://tinyurl.com/inter4KUHD
The provided dataset includes information on energy consumption, time complexity, video quality, and bitrate for various video codecs used on different devices.
- Energy Consumption: measured using CodeCarbon and Powermetrics tools and covers CPU, RAM, total energy consumption in kWh, and CO2 emissions in Kg.
- Time Complexity: measured using the time command in the Linux operating system and includes user-time and run-time values.
- Video Quality: reported video quality includes four different metrics, namely PSNR, VMAF, SSIM, and MS-SSIM.
- Bitrate: reported in Kbps.
The dataset is provided in four category files:
- Video Complexity: contains the value of SITI, Eh, and SCTC video complexity metrics for each video sequence.
- Video Encoding: contains information about encoding energy consumption, time complexity, video quality, and bitrate for all encoding processes.
- Video Decoding: contains decoding energy consumption, and time complexity for H.264/AVC, H.265/HEVC, AV1, H.266/VVC video codecs and encoding parameters across three different devices. Each decoding process is repeated five times for consistent results, so the aggregated values can be used here.
- Video Decoding and Upscaling to Original Resolutions: contains the same information as the decoding file provided, but here the impact of upscaling on the mentioned parameters is also considered.
To read CSV files and generate an output file, namely dataset_output.csv, with all the necessary information, use the following command.
python3 generate_output.py -a aggregation_method -d decoding_device -o dataset_output.csv
The aggregation method is utilized to combine decoding information that has been repeated five times.
Available aggregation methods:
- mean (Default)
- median
- min
- max
- first
- last
Available decoding devices:
- lenovo: ThinkPad P1 Gen2 laptop, Intel Core i7-9750H CPU @ 2.60 GHz, 16 GB of RAM, running Ubuntu 22.04.3 LTS.
- mini: Apple Mac Mini, Octa-core M1 processor, 16 GB of RAM, running macOS Ventura version 13.3.1.
- studio: Apple Mac Studio, 20-core M1 Ultra processor, 64 GB of RAM, running macOS Ventura version 13.5.
To reproduce the dataset, follow these steps:
Run the encoding script:
python3 run_encoding.py <path_to_the_input_videos>
Run the decoding script:
After the compressed files are generated, run the following command to obtain the decoding results:
python3 run_decoding.py
If this work is helpful for your research, please consider citing MVCD.
@inproceedings{amirpour_mvcd_2024,
title = {{MVCD}: {Multi-Dimensional} {Video} {Compression} {Dataset}},
volume = {2024},
shorttitle = {{MVCD}},
language = {English},
author = {Amirpour, H. and Ghasempour, M. and Tashtarian, F. and Afzal, S. and Hamidouche, W. and Timmerer, C.},
year = {2024},
keywords = {Video encoding, decoding, energy, complexity, quality},
}