This repository contains checkpointed models of Google's Universal Video Quality (UVQ) model. UVQ is a no-reference perceptual video quality assessment model that is designed to work well on user-generated content, where there is no pristine reference.
Read this blog post for an overview of UVQ:
"UVQ: Measuring YouTube's Perceptual Video Quality", Google AI Blog 2022
More details are available in our paper:
Yilin Wang, Junjie Ke, Hossein Talebi, Joong Gon Yim, Neil Birkbeck, Balu Adsumilli, Peyman Milanfar, Feng Yang, "Rich features for perceptual quality assessment of UGC videos", CVPR 2021.
The corresponding data from the paper is available for download from: YouTube UGC Dataset
You must have FFmpeg installed and available on your path.
The models and code require Python 3.6 (or greater) and Tensorflow.
With virtualenv, you can install the requirements to a virtual environment:
virtualenv venv
source venv/bin/activate
pip3 install -r requirements.txt
You can grab some examples videos from the YouTube UGC Dataset. For example, you can get Gaming_1080P-0ce6_orig.mp4 using curl:
curl -o Gaming_1080P-0ce6_orig.mp4 https://storage.googleapis.com/ugc-dataset/vp9_compressed_videos/Gaming_1080P-0ce6_orig.mp4
You can then run the example:
mkdir -p results
python3 uvq_main.py --input_files='Gaming_1080P-0ce6_orig,20,Gaming_1080P-0ce6_orig.mp4' --output_dir results --model_dir models
The input files format is a line with the following fields:
id,video_length,filepath
The output_dir
will contain a csv file with the results for each model. For example,
cat /tmp/results/mos_ytugc20s_0_Gaming_1080P-0ce6_orig.mp4_orig.csv
Gives:
Gaming_1080P-0ce6,compression,3.927867603302002
Gaming_1080P-0ce6,content,3.945391607284546
Gaming_1080P-0ce6,distortion,4.267196607589722
Gaming_1080P-0ce6,compression_content,3.9505696296691895
Gaming_1080P-0ce6,compression_distortion,4.062019920349121
Gaming_1080P-0ce6,content_distortion,4.067790699005127
Gaming_1080P-0ce6,compression_content_distortion,4.058663845062256
We provide multiple predcited scores, using different combinations of UVQ features.
compression_content_distortion
(combining three features) is our default score for Mean Opinion Score (MOS) prediction.
The output features folder includes UVQ labels and raw features:
Gaming_1080P-0ce6_orig_feature_compression.binary
Gaming_1080P-0ce6_orig_feature_content.binary
Gaming_1080P-0ce6_orig_feature_distortion.binary
Gaming_1080P-0ce6_orig_label_compression.csv
Gaming_1080P-0ce6_orig_label_content.csv
Gaming_1080P-0ce6_orig_label_distortion.csv
UVQ labels (.csv, each row corresponding to 1s chunk):
compression: 16 compression levels per row, corresponding to 4x4 subregions of the entire frame.
distortion: 26 distortion types defined in KADID-10k for 2x2 subregions. The first element is the undefined type.
content: 3862 content labels defined in YouTube-8M.
UVQ raw features (in binary):
25600 float numbers per 1s chunk.