SUREAL is a toolbox developed by Netflix that includes a number of models for the recovery of mean opinion scores (MOS) from noisy measurements obtained in psychovisual subjective experiments. Read this paper and this latest paper for some background.
SUREAL also includes models to recover MOS from paired comparison (PC) subjective data, such as Thurstone (Case V) and Bradley-Terry.
SUREAL can be either installed through pip (available via PyPI), or locally.
To install SUREAL via pip, run:
pip install surealTo install locally, first, download the source. Under the root directory, (preferably in a virtualenv), install the requirements:
pip install -r requirements.txtUnder Ubuntu, you may also need to install the python-tk (Python 2) or python3-tk (Python 3) packages via apt.
To test the source code before installing, run:
python -m unittest discover --start test --pattern '*_test.py' --verbose --bufferLastly, install SUREAL by:
pip install .If you want to edit the source, use pip install --editable . or pip install -e . instead. Having --editable allows the changes made in the source to be picked up immediately without re-running pip install .
Run:
sureal --helpThis will print usage information:
usage: sureal [-h] --dataset DATASET --models MODELS [MODELS ...] [--output-dir OUTPUT_DIR]
[--plot-raw-data] [--plot-dis-videos] [--plot-observers]
optional arguments:
-h, --help show this help message and exit
--dataset DATASET Path to the dataset file.
--models MODELS [MODELS ...]
Subjective models to use (can specify more than one),
choosing from: MOS, P910, P913, BT500.
--output-dir OUTPUT_DIR
Path to the output directory (will force create is not existed).
If not specified, plots will be displayed and output will be printed.
--plot-raw-data Plot the raw data. This includes the raw opinion scores presented
in a video-subject matrix, counts per video and counts per subject.
--plot-dis-videos Plot the subjective scores of the distorted videos.
--plot-observers Plot the scores of the observers.Below are two example usages:
sureal --dataset resource/dataset/NFLX_dataset_public_raw_last4outliers.py --models MOS P910 \
--plot-raw-data --plot-dis-videos --plot-observers --output-dir ./output/NFLX_dataset_public_raw_last4outliers
sureal --dataset resource/dataset/VQEGHD3_dataset_raw.py --models MOS P910 \
--plot-raw-data --plot-dis-videos --plot-observers --output-dir ./output/VQEGHD3_dataset_rawHere --models are the available subjective models offered in the package, including:
- MOS - Standard mean opinion score.
- P910 - Model based on subject bias/inconsistency modeling and maximum likelihood estimation (MLE), newly standardized in ITU-T P.910 (11/21) Annex E (also in ITU-T P.913 (06/21) 12.6). The details of the algorithm is covered by the two papers aforementioned (paper 1 and paper 2).
- P913 - Model based on subject bias removal, standardized in ITU-T P.913 (06/21) 12.4.
- BT500 - Model based on subject rejection, standardized in ITU-R BT.500-14 (10/2019) A1-2.3.1.
The sureal command can also invoke subjective models for paired comparison (PC) subjective data. Below is one example:
sureal --dataset resource/dataset/lukas_pc_dataset.py --models THURSTONE_MLE BT_MLE \
--plot-raw-data --plot-dis-videos --output-dir ./output/lukas_pc_datasetHere --models are the available PC subjective models offered in the package:
- THURSTONE_MLE - Thurstone (Case V) model, with a MLE solver.
- BT_MLE - Bradley-Terry model, with a MLE solver.
Both models leverage MLE-based solvers. For the mathematics behind the implementation, refer to this document.
--dataset is the path to a dataset file. Dataset files may be .py or .json files. The following examples use .py files, but JSON-formatted files can be constructed in a similar fashion.
There are two ways to construct a dataset file. The first way is only useful when the subjective test is full sampling, i.e. every subject views every distorted video. For example:
ref_videos = [
{
'content_id': 0, 'content_name': 'checkerboard',
'path': 'checkerboard_1920_1080_10_3_0_0.yuv'
},
{
'content_id': 1, 'content_name': 'flat',
'path': 'flat_1920_1080_0.yuv'
},
]
dis_videos = [
{
'content_id': 0, 'asset_id': 0,
'os': [100, 100, 100, 100, 100],
'path': 'checkerboard_1920_1080_10_3_0_0.yuv'
},
{
'content_id': 0, 'asset_id': 1,
'os': [40, 45, 50, 55, 60],
'path': 'checkerboard_1920_1080_10_3_1_0.yuv'
},
{
'content_id': 1, 'asset_id': 2,
'os': [90, 90, 90, 90, 90],
'path': 'flat_1920_1080_0.yuv'
},
{
'content_id': 1, 'asset_id': 3,
'os': [70, 75, 80, 85, 90],
'path': 'flat_1920_1080_10.yuv'
},
]
ref_score = 100In this example, ref_videos is a list of reference videos. Each entry is a dictionary, and must have keys content_id, content_name and path (the path to the reference video file). dis_videos is a list of distorted videos. Each entry is a dictionary, and must have keys content_id (the same content ID as the distorted video's corresponding reference video), asset_id, os (stands for "opinion score"), and path (the path to the distorted video file). The value of os is a list of scores, reach voted by a subject, and must have the same length for all distorted videos (since it is full sampling). ref_score is the score assigned to a reference video, and is required when differential score is calculated, for example, in DMOS.
The second way is more general, and can be used when the test is full sampling or partial sampling (i.e. not every subject views every distorted video). The only difference from the first way is that, the value of os is now a dictionary, with the key being a subject ID, and the value being his/her voted score for particular distorted video. For example:
'os': {'Alice': 40, 'Bob': 45, 'Charlie': 50, 'David': 55, 'Elvis': 60}Since partial sampling is allowed, it is not required that every subject ID is present in every os dictionary.
In case a subject has voted a distorted video twice or more (repetitions), the votes can be logged by having a list in lieu of single vote. For example:
'os': {'Alice': 40, 'Bob': [45, 45], 'Charlie': [50, 60], 'David': 55, 'Elvis': 60}In case of a PC dataset, a distorted video is compared against another distorted video, and a vote is recorded. In this case, the key is a tuple of the subject name and the asset_id of the distorted video compared against. For example:
'os': {('Alice', 1): 40, ('Bob', 3): 45}where 1 and 3 are the asset_id of the distorted videos compared against. For an example PC dataset, refer to lukas_pc_dataset.py.
Note that for PC models, we current do not yet support repetitions.
The deprecated version of the command line can still be invoked by:
PYTHONPATH=. python ./sureal/cmd_deprecated.pySee here for an example script to use SUREAL in Google Collab notebook.
SUREAL uses tox to manage automatic testing and continuous integration with Travis CI on Github, and setupmeta for new version release, packaging and publishing. Refer to DEVELOPER.md for more details.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent