these script can be used to produce / verify specific encoders and scenarios according to TR26.955.
- video codec charcterization scripts
git clone https://github.com/haudiobe/5GVideo
cd 5GVideo
Please use a python virtual environment to install dependencies and run the scripts.
install python dependencies:
pip3 install -r src/requirements.txt
The scripts assume you have a local directory where content has been downloaded from: https://dash-large-files.akamaized.net/WAVE/3GPP/5GVideo/ or follows that structure.
No default value is provided for the path to the working directory, it must be supplied explicitely by one of these 2 means:
- the scripts accept a
--working-dirargument to supply the path. - the
VCC_WORKING_DIRenvironment variable is used to configure the location of that directory
to download anchors, bitstreams, metrics, encoder configs, & reference sequence list for a given scenario:
python3 src/download.py streams https://dash-large-files.akamaized.net/WAVE/3GPP/5GVideo/Bitstreams/Scenario-1-FHD/264/streams.csv $VCC_WORKING_DIR/Bitstreams/Scenario-1-FHD/264
to download anchors reference sequences for a csv list:
python3 src/download.py sequences $VCC_WORKING_DIR/Bitstreams/Scenario-1-FHD/reference-sequences.csv $VCC_WORKING_DIR/ReferenceSequences
src/vcc.py provides the commands to encode, decode, and compute metrics.
All the commands described below have a common usage pattern.
python3 src/vcc.py [--dry-run] -s S1-A01-264 decode
the scripts parses -s S1-A01-264, finds Scenario-1-FHD/265/streams.csv, and processes all QPs listed for S1-A01-264.
python3 src/vcc.py [--dry-run] -s S1-A01-264-22 decode
the scripts works as described above, but will only process a single QP - in this case: 22.
python3 src/metrics.py [--dry-run] -c S1-JM-01 decode
the scripts parses -c S1-JM-01, finds Scenario-1-FHD/264/streams.csv, then processes all QPs, for all anchors where the encoder config is S1-JM-01.
--dry-run: does not execute the subprocesses, but instead prints the corresponding commands to stdout.--queue: advanced - pushes the commands as tasks to a worker queue using the celery framework, instead of processing sequentialy. This option is meant to be used with thesample docker-compose.ymlstack. See Docker usage > Using docker-compose.
JM19_0_ENCODER=~/deps/JM/bin/lencod_static
python3 src/vcc.py -s S1-A01-264 encode
dry-run mode prints the command used to encode to stdout but doesn't run it, but it does compute existing bitstream md5.
JM19_0_DECODER=~/deps/JM/bin/ldecod_static
python3 src/vcc.py -s S1-A01-264 decode
dry-run mode prints the command used to encode to stdout but doesn't run it, but it does compute existing reconstruction md5.
HDRCONVERT_TOOL=~/deps/HDRTools/build/bin/HDRConvert
python3 src/vcc.py -s S1-A01-264 decode --reference
python3 src/vcc.py -s S1-A01-264 decode --reconstructions
Performs conversion needed before metrics computation using HDRTools' HDRConvert.
when --reconstructions is specfied, the reconstructions for all QPs are processed. when --reference is specified, the reference sequence is processed instead.
Conversion is either be 8 to 10 bit YUV, or YUV to EXR. When conversion is not needed for a given sequence, the script does nothing.
The converted sequences are stored in a tmp subfolder for each anchor.
dry-run mode prints the command used to convert to stdout but doesn't run it.
when running the metrics command, the result is stored in the json file.
HDRMETRICS_TOOL=~/deps/HDRTools/build/bin/HDRMetrics
VMAF_EXE=~/deps/vmaf/libvmaf/build/tools/vmafossexec
VMAF_MODEL=/home/deps/vmaf/model/vmaf_v0.6.1.json
python3 src/vcc.py -s S1-A01-264 metrics
Computes metrics for the given sequence. This step also computes bitrate after removing SEI.
HDRTOOLS or VMAF can be enabled / disabled through environment variables.
DISABLE_HDRMETRICS=1
DISABLE_VMAF=1
This step is meant to run after successfull conversion. It uses the converted sequence where appropriate, and raise an error if it doesn't exist.
dry-run mode doesn't run HDRTools or vmaf, but it parses existing logs if any.
Bitstream size / bitrate is always computed when running metrics. It can be computed explicitly as a separate step.
python3 src/vcc.py -s S1-A01-264-22 bitrate
First, SEI_REMOVAL_APP is used if implemented by the encoder class. Then bitrate is computed based on the resulting bitstream size, and the json metadata are updated with the result.
Metrics are saved on the anchors JSON file. A script can be used to export them to .csv:
python3 src/metrics.py -s S1-A01-264 csv-metrics
python3 src/metrics.py -c S1-HM-02 csv-metrics
For the above commands, the metrics .csv files can then be found in: $VCC_WORKING_DIR/Bitstreams/Scenario-1-FHD/265/Metrics
The characterization script is meant to compare test data to an anchor. For instance, to compare sequences JM and HM for Scenario-1-FHD, you should have the following data:
$VCC_WORKING_DIR/Bitstreams/Scenario-1-FHD/
$VCC_WORKING_DIR/Bitstreams/Scenario-1-FHD/264/S1-A01-264.json
$VCC_WORKING_DIR/Bitstreams/Scenario-1-FHD/264/streams.csv
$VCC_WORKING_DIR/Bitstreams/Scenario-1-FHD/264/Metrics/S1-A01-264.csv
[...]
$VCC_WORKING_DIR/Bitstreams/Scenario-1-FHD/
$VCC_WORKING_DIR/Bitstreams/Scenario-1-FHD/264/S1-A01-265.json
$VCC_WORKING_DIR/Bitstreams/Scenario-1-FHD/265/streams.csv
$VCC_WORKING_DIR/Bitstreams/Scenario-1-FHD/265/Metrics/S1-A01-264.csv
[...]
To compare and plot psnr, y_psnr, and ms_ssim for these 2 sequences:
python3 src/compare.py --plot -s S1-A01-264 S1-A01-265 psnr y_psnr ms_ssim
Assuming you have all the anchors for encoder config S1-JM-01 and S1-HM-01, you can process them at once:
python3 src/compare.py --plot -c S1-JM-01 S1-HM-01 psnr y_psnr ms_ssim
This guide provides step by step instructions to process with reconstruction and metrics verifications.
In this guide, /data will be the directory where reference content is downloaded.
All subsequent verification steps assume the VCC_WORKING_DIR environment variable is set to /data.
- download bitstreams and metadata for a given scenario
python3 src/download.py streams --bitstream https://dash-large-files.akamaized.net/WAVE/3GPP/5GVideo/Bitstreams/Scenario-1-FHD/VTM/streams.csv /data/Bitstreams/Scenario-1-FHD/VTM
/data
└ Bitstreams
└ Scenario-1-FHD
| reference-sequence.csv
└ VTM
| streams.csv
└── CFG
└── Metrics
└── S1-T11-VTM
| └── S1-T11-VTM-22
| S1-T11-VTM-22.json
| S1-T11-VTM-22.bin
| └── S1-T11-VTM-27
| └── S1-T11-VTM-32
| └── S1-T11-VTM-37
[...]
└ S1-T17-VTM
- download reference sequences for that scenario
python3 src/download.py sequences /data/Bitstreams/Scenario-1-FHD/reference-sequence.csv /data/ReferenceSequences
/data
└ ReferenceSequences
└ Life-Untouched
Life-Untouched-FHD.json
Life-Untouched-FHD.yuv
reconstruct QP 22 for S1-T11-VTM
python3 src/vcc.py decode -s S1-T11-VTM-22 decode
/data/Bitstreams/Scenario-1-FHD
└── S1-T11-VTM
└── S1-T11-VTM-22
S1-T11-VTM-22.yuv
S1-T11-VTM-22.dec.log
In some cases, must process conversion from one representation to another before. This 8 to 10 bit, 10 to 8 bit, and YUV to RGB444.
In the case of S1-T11-VTM, YUV to RGB444 needs to be performed on the reference sequence as well as on the reconstructions to compute HDR metrics using HDRTools.
Convert the reconstructions of the test scenario:
python3 src/vcc.py -s S1-T11-VTM-22 convert --reconstructions
/data/Bitstreams/Scenario-1-FHD
└── S1-T11-VTM
└── S1-T11-VTM-22
| S1-T11-VTM-22.yuv
| S1-T11-VTM-22.yuv.json
└── tmp
S1-T11-VTM-22_2020_444_%05d.json
S1-T11-VTM-22_2020_444_000000.exr
S1-T11-VTM-22_2020_444_000000.exr
S1-T11-VTM-22_2020_444_000001.exr
...
S1-T11-VTM-22_2020_444_%05d.hdrconvert.log
...
Convert the reference sequence used to encode S1-T11-VTM:
python3 src/vcc.py -s S1-T11-VTM convert --sequences
/data/ReferenceSequences
└── Life-Untouched
| Life-Untouched-FHD.yuv
| Life-Untouched-FHD.json
└── tmp
Life-Untouched-FHD_2020_444_%05d.json
Life-Untouched-FHD_2020_444_000000.exr
Life-Untouched-FHD_2020_444_000001.exr
...
Life-Untouched-FHD_2020_444_%05d.hdrconvert.log
...
Run metrics computations, and updates the json metadata.
python3 src/vcc.py -s S1-T11-VTM-22 metrics
/data/Bitstreams/Scenario-1-FHD
└── S1-T11-VTM
└── S1-T11-VTM-22
| S1-T11-VTM-22.metrics.log
| S1-T11-VTM-22.json
| ...
└── tmp
S1-T11-VTM-22_2020_444_%05d.metrics.log
...
When computing metrics, results are stored in json metadata. After computing metrics for all anchors of a given encoder configuration, the metrics are exported to the appropriate csv format.
this is conveniently done for a set of all anchors with a given encoder config:
python3 src/metrics.py -c S1-VTM-O2 csv-metrics
/data/Bitstreams/Scenario-1-FHD/Metrics
└── ...
S1-VTM-02.csv
but can also be done for individual anchor:
python3 src/metrics.py -s S1-T11-VTM csv-metrics
/data/Bitstreams/Scenario-1-FHD/Metrics
└── S1-T11-VTM.csv
to generate a verification report, we compare the local reconstruction and metrics computed in the working directory, to the original data downloaded from the server.
create the directory /data/Origin/Scenario-1-FHD/VTM and download the reference metrics there:
python3 src/download.py streams --metrics https://dash-large-files.akamaized.net/WAVE/3GPP/5GVideo/Bitstreams/Scenario-1-FHD/VTM/streams.csv /data/Origin/Scenario-1-FHD/VTM
to generate a verification verify-metrics command of the metrics.py script.
the following options are needed:
-o: the path to the directory containing the original data to which localy computed results will be compared-r: the resulting report path is specific with
the metrics key on which the verification report should run are specified as the last arguments. eg psnr y_psnr
the following metrics keys are implemented: psnr, y_psnr, u_psnr, v_psnr, ms_ssim, vmaf, bitrate, bitrate_log, wpsnr, y_wpsnr, u_wpsnr, v_wpsnr, psnr100, de100
python3 ./src/metrics.py -c S1-VTM-01 verify-metrics
-o /data/origin
-r /data/origin/s1-vtm-01-verification-report.csv
--doc - --email [email protected] --company corp
y_psnr psnr
All the dependencies in a single docker image.
on linux: ./build.sh (should work on mac as well)
on windows: .\build.ps1
the build script builds several docker images:
vcc:baseused as provides common build environment, and metrics dependencies : HDRTools, vmaf.vcc:jm,vcc:hm,vcc:etm,vcc:ctm,vcc:aombuilds the respective reference encoders, decoders, and sei removal executable.vcc:workercopies pre-built binaries from the above images, supplies the appropriate environment variables, adds the scripts & config folders.
build a specific image : docker image ls
list docker images on your machine: docker image ls
remove a specific docker image: docker image rm vcc:worker
remove all docker image: docker image rm vcc
The worker image is configured with /data as the root directory for all the ReferenceSequences and Bitstreams directories.
When running the worker image; bitstreams and reference sequences must be mounted to the container /data directroy, eg. using docker's --mount option.
docker run -it --mount type=bind,source=/path/to/local/vcc_working_dir,target=/data" vcc:workers vcc.py -s S1-A01-264 decode
local Bitstreams and ReferenceSequences data may be organized arbitrarily and mounted to the appropriate directory:
docker run -it \
--mount type=bind,source=/path/to/local/bitstreams_dir,target=/data/Bitstreams" \
--mount type=bind,source=/path/to/local/sequences_dir,target=/data/ReferenceSequences" \
vcc:workers vcc.py -s S1-A01-264 decode
A sample docker-compose is provided. It starts a background worker and a redis queue.
The --queue option of the vcc.py script can be used to add jobs to the queue.
The initial concurrency is set to 2 simultaneous jobs. Too high value for concurrency may not be suitable for all tasks (eg. 4K encodes).
to use it, first configure local directories to mount using environment variables: linux
export VCC_BITSTREAMS_DIR=/path/to/local/bitstreams_dir
export VCC_REFERENCES_DIR=/path/to/local/sequences_dir
windows
$env:VCC_BITSTREAMS_DIR="c:\path\to\local\bitstreams_dir"
$env:VCC_REFERENCES_DIR="c:\path\to\local\sequences_dir"
build workers:
docker-compose build workers
then start the stack:
docker-compose up -d
verify that all services are running:
docker-compose ps
the queue can be monitored with a web browser at http://localhost:8888/
queue some tasks, eg. decode all anchors encoded with s1-jm-01.cfg:
docker-compose exec worker python3 vcc.py --queue -c S1-JM-01 decode