Closed kornelski closed 7 years ago
@kornelski do you still have the results txt
?
No, but that's the script I use to run the test on macOS via Wine. It uses gnu parallel. Expects reference BMP images converted to PNG.
set -e
set -o pipefail
{
for i in reference_images/*png; do
for dis in distorted_images/`basename -s .png "$i"`_*.png; do
echo 'printf "%s\t%s\n" $(../butteraugli/bazel-bin/butteraugli "'"$i"'" "'"$dis"'") "'"$dis"'"'
done
done
} | time parallel | sort -k2 > testtmp
/Applications/Wine.app/Contents/Resources/bin/wine metrics_values/spearman.exe mos.txt testtmp
Latest version from current repo: https://github.com/google/butteraugli/commit/71b18b636b9c7d1ae0c1d3730b85b3c127eb4511
TID2013 Category | Spearman correlation |
---|---|
Noise | -0.871 |
Actual | -0.872 |
Simple | -0.849 |
Exotic | -0.711 |
New | -0.852 |
Color | -0.858 |
Full | -0.760 |
Latest version from pik: https://github.com/google/pik/commit/b4866ff9332fe13b7f7f70e55de02459f5fbb3b3 As suggested in issue https://github.com/google/butteraugli/issues/46
butteraugli_main_pik_b4866ff.txt
TID2013 Category | Spearman correlation |
---|---|
Noise | -0.798 |
Actual | -0.814 |
Simple | -0.825 |
Exotic | -0.584 |
New | -0.782 |
Color | -0.831 |
Full | -0.643 |
It would be interesting to take a p-norm over the butteraugli field and use that for the spearman correlation. In JPEG XL development we take an average of 3, 6 and 12-norms to support decision making and mostly use the maximum based butteraugli score for reference only.
@jyrkialakuijala how can I force butteraugli to use p-norm?
@Ser-Gen the Butteraugli inside JPEG XL supports this - would it be an option to use https://gitlab.com/wg1/jpeg-xl/-/blob/master/tools/butteraugli_main.cc ? To build it, we recommend Docker (documentation).
@jan-wassenberg can't run butteraugli_main from host system.
./butteraugli_main: error while loading shared libraries: libIlmImf-2_2.so.22: cannot open shared object file: No such file or directory
I will try to use it from docker, if fail to find workaround.
Hi @Ser-Gen, did you successfully install the libopenexr-dev and other packages mentioned in doc/developing_in_debian.md?
@jan-wassenberg initially no, I used docker to build tools as you suggested. https://gitlab.com/wg1/jpeg-xl/-/blob/master/doc/developing_in_docker.md
After installing packages from doc/developing_in_debian.md
same error is appeared.
Current master
of JPEG-XL
https://gitlab.com/wg1/jpeg-xl/-/tree/0709f3a494c901c3a9f93159312b1177e3fd4fe4
butteraugli-jpegxl-0709f3a4.txt
TID2013 Category | Spearman correlation |
---|---|
Noise | -0.809 |
Actual | -0.835 |
Simple | -0.858 |
Exotic | -0.607 |
New | -0.777 |
Color | -0.843 |
Full | -0.658 |
butteraugli-jpegxl-3-norm-0709f3a4.txt
TID2013 Category | Spearman correlation |
---|---|
Noise | -0.865 |
Actual | -0.884 |
Simple | -0.899 |
Exotic | -0.517 |
New | -0.813 |
Color | -0.862 |
Full | -0.655 |
Looks like you got it working, nice! Are any changes needed to our documentation? I just verified that the Docker image has /usr/lib/x86_64-linux-gnu/libIlmImf-2_2.so.22, which is referenced from butteraugli_main.
FYI, scores for 290e39552e213ed352761114561f92b93c902f55 measured against TID2013 database: