Objective Perceptual Analysis - Video Karma Predictor
This is a kit for testing codecs objectively through FFmpeg. It employs VMAF, SSIM, PSNR and also a Perceptual Hash metric. Multiple encoding tests can be ran comparing the encoders and the encodings, encoding techniques. The metrics and test harness allow quick feedback to test theories and new code in FFmpeg. There's objective metrics, graphs and easy comparisons historically.
Use OpenCV img_hash frame comparisons in FFmpeg libavfilter for per title encoding / Perceptual Quality comparisons
Documentation on OpenCV img_hash: https://docs.opencv.org/trunk/d4/d93/group__img__hash.html
This will use perceptual hashes from OpenCV's img_hash module which includes PHash and it the main algorithm used. Each video frame is compared to the last video frame then a hamming distance is derived from the two hashes. This values shows the perceptual similarity of the two images. The hamming distance is used to vary the encoders bitrate or CRF level. Currently only X264 is supported in this implementation.
Also research via bin/encode.py and bin/results.py script:
- Parallel Encoding / can set per test for comparisons
- Quality both Objective and setup for Subjective tests
- Easy encoding tests for H.265, VP9 and AV1
- Perceptual Hash research with encoding decisions and metrics
- Simple Objective metrics calculated
- Frame images and metrics burn in per frame via SRT
- Scene segmentation and analysis with objective metrics
Everything can be easily setup via setup.sh, it will install what is necessary for the most part. Please report back any issues so this can be improved for edge cases.
See the bin/readme.md file for information on bin/encode.py and bin/results.py. See the scripts/readme.md file for information on setting up tests.
Currenty works on Arch Linux (recommended), CentOS 7 (deprecated) and Mac OS X VMAF, libVPX, libAOM, libRav1e, svt-av1, libx264, libOpenCV build of FFmpeg
- rav1e support based off of work by Derek Buitenhuis https://github.com/dwbuiten/FFmpeg
type: make docker
Example using the docker image:
- sudo docker run --rm -v `pwd`/tests:/opaencoder/tests opaencoder sh scripts/run_example.sh
- sudo docker run --rm -v `pwd`/tests:/opaencoder/tests opaencoder bin/encode.py -m vmaf,psnr,phqm,cambi\
-n tests/test000 -p 2 -t "01000X264H264|ffmpeg|twopass|S|mp4||1080|-pix_fmt|yuv420p|-f|mp4|-movflags|+faststart|-profile:v|high|-preset|slow|-vcodec|libx264|-bf|0|-refs|4|-b:v|1000k|-maxrate:v|1500k|-bufsize:v|3000k|-minrate:v|1000k|-tune|animation|-x264opts|rc-lookahead=48:keyint=96|-keyint_min|48|-g|96|-force_key_frames|expr:eq(mod(n,48),0)|-hide_banner|-nostats"
Makefile will run the proper setup script and install mediainfo, opencv, libx264, libvmaf, nasm git, wget, freetype-devel... Everything should be done for you, although if not report it as a bug. Warning: Scripts will install / alter system packages via Sudo. Please keep this in mind
This uses an FFmpeg with an extra video filter which uses OpenCV to compute hamming distance values from each frames hash vs. the previous frames hash.
There is a ffmpeg_modifications.diff patch included...
(this is done for you via the make command which runs the proper setup* script)
git clone https://git.ffmpeg.org/ffmpeg.git FFmpeg
cd FFmpeg
git checkout tags/n5.1.2
cat ../ffmpeg_modifications.diff | patch -p1
You can run tests using the bin/encode.py script. See the /bin/readme.md for more details.
Perceptual Hash Quality Metric: (output a stats file with psnr/mse/phqm (perceptual hash quality metric)
./FFmpeg/ffmpeg -i <encode> -i <refvideo> -filter_complex "[0:v][1:v]phqm=hash_type=phash:stats_file=stats.log" -f null -
phqm AVOptions:
stats_file <string> ..FV..... Set file where to store per-frame difference information.
f <string> ..FV..... Set file where to store per-frame difference information.
scd_thresh <double> ..FV..... Scene Change Detection Threshold. (from 0 to 1) (default 0.5)
hash_type <int> ..FV..... Type of Image Hash to use from OpenCV. (from 0 to 6) (default phash)
average ..FV..... Average Hash
blockmean1 ..FV..... Block Mean Hash 1
blockmean2 ..FV..... Block Mean Hash 2
colormoment ..FV..... Color Moment Hash
marrhildreth ..FV..... Marr Hildreth Hash
phash ..FV..... Perceptual Hash (PHash)
radialvariance ..FV..... Radial Variance Hash
PHQM Scene Detection, frame ranges for each segmented scene with an avg hamming distance score per scene.
# (./FFmpeg/ffmpeg -loglevel warning -i encode.mp4 -i reference.mov -nostats -nostdin \
-threads 12 -filter_complex [0:v][1:v]phqm=stats_file=phqm.data -f null -)
[phqm @ 0x40def00] ImgHashScene: n:1-231 hd_avg:0.861 hd_min:0.000 hd_max:6.000 scd:0.80
[phqm @ 0x40def00] ImgHashScene: n:232-491 hd_avg:0.265 hd_min:0.000 hd_max:2.000 scd:0.57
[phqm @ 0x40def00] ImgHashScene: n:492-541 hd_avg:0.340 hd_min:0.000 hd_max:2.000 scd:0.57
[phqm @ 0x40def00] ImgHashScene: n:542-658 hd_avg:0.350 hd_min:0.000 hd_max:2.000 scd:0.82
[phqm @ 0x40def00] ImgHashScene: n:659-708 hd_avg:0.420 hd_min:0.000 hd_max:2.000 scd:0.92
[phqm @ 0x40def00] ImgHashScene: n:709-1057 hd_avg:1.009 hd_min:0.000 hd_max:6.000 scd:0.51
[phqm @ 0x40def00] ImgHashScene: n:1058-1266 hd_avg:0.708 hd_min:0.000 hd_max:4.000 scd:0.59
[Parsed_phqm_0 @ 0x40f1340] PHQM average:0.601282 min:0.000000 max:6.000000
This is implementing a Patent by Christopher Kennedy @ Ellation / Crunchyroll:
Patent for https://patents.justia.com/patent/10244234 Adaptive compression rate control
OPA can run VapourSynth filter chains on the mezzanines which can be associated to test labels. When using VapourSynth, new filtered mezzanines are being created which will be used as source files for the test encodes tagged with the same label. Metrics can either be calculated against the original mezzanine or the filtered mezzanine.
- Make sure that the filters you want to use are installed on your system. On Arch Linux distributions you can simply run
setupArch.shto setup a VapourSynth environment with most relevant filter libraries and wrapper scripts. - Create wrapper scripts in the
bin/includedirectory with your filter chains. They must havesrcandargsarguments. They also needreturn video(or whatever your clip object is called) at the bottom. Seebin/include/vs_example.pyfor an example. - Pass the
-vsparameter tobin/encode.pywith the following structure:Label1,Label2,...|wrapper_script.wrapper_function|arg1,arg2,...;Label3|wrapper_script.wrapper_function|arg1,arg2,...;... - Pass the
-vpparameter tobin/encode.pyto calculate encode metrics against the filtered mezzanines.
Docker example:
docker run --rm -v `pwd`/tests:/opaencoder/tests opaencoder bin/encode.py -m vmaf,psnr \
-n tests/test000 -p 2 \
-t "01000X264H264|ffmpeg|twopass|S|mp4||1080|-pix_fmt|yuv420p|-f|mp4|-movflags|+faststart|-profile:v|high|-preset|slow|-vcodec|libx264|-bf|0|-refs|4|-b:v|1000k|-maxrate:v|1500k|-bufsize:v|3000k|-minrate:v|1000k|-tune|animation|-x264opts|rc-lookahead=48:keyint=96|-keyint_min|48|-g|96|-force_key_frames|expr:eq(mod(n,48),0)|-hide_banner|-nostats" \
-vs "01000X264H264|vs_example.example_wrapper|640,360" \
-vp
Nov 18, 2016
Disclosed by way of example embodiments are a system and a computer implemented method for adaptively encoding a video by changing compression rates for different frames of the video. In one aspect, two frames of a video are compared to determine a compression rate for compressing one of the two frames. Hash images may be generated for corresponding frames for the comparison. By comparing two hash images, a number of stationary objects and a number of moving objects in the two frames may be determined. Moreover, a compression rate may be determined according to the number of stationary objects and the number of moving objects.
Patent number: 10244234 Filed: Nov 18, 2016 Date of Patent: Mar 26, 2019 Patent Publication Number: 20170150145 Assignee: Ellation, Inc. (San Francisco, CA) Inventor: Chris Kennedy (Alameda, CA) Primary Examiner: Dramos Kalapodas Application Number: 15/356,510