Fixing imread() function 16 bit reading png problem with libspng #27113
Conversation
640563f to
bc40042
Compare
|
@asmorkalov the PR is still WIP. no need to start CI yet. |
This reverts commit 4e15215.
|
The purpose of the PR was to load bit-exact compatible results with libspng and libpng. To test this, |
|
I added the motivation to the PR description to save it in Git history after merge. |
| #ifdef HAVE_PNG | ||
| EXPECT_EQ(19519, img.at<ushort>(0, 0)); | ||
| #endif | ||
| #ifdef HAVE_SPNG | ||
| EXPECT_EQ(19517, img.at<ushort>(0, 0)); | ||
| #endif |
There was a problem hiding this comment.
Sounds like wrong rounding somewhere in libspng.
There was a problem hiding this comment.
last commit is temporarily to show that the spng conversion code is the same as libpng. i temporarily used if (m_frame_count > 5) that means the difference caused by cvtColor which i used in apng decoding.
[ RUN ] Imgcodecs_APNG.imread_animation_16u
/build/precommit_linux64/4.x/opencv/modules/imgcodecs/test/test_animation.cpp:680: Failure
Expected equality of these values:
19519
img.at<ushort>(0, 0)
Which is: 19517
[ FAILED ] Imgcodecs_APNG.imread_animation_16u (2 ms)
There was a problem hiding this comment.
If you find it appropriate, I can add a to-do note and let's merge this. I'll work on apng and try to find a solution for this difference later.
This reverts commit 1d6b6f8.
To show that the spng conversion code is the same as libpng
|
The last commit looks suspicious. |
now the value is |
|
in conclusion, I think this PR is correct. I suggest leaving a to-do note about libpng and postponing the solution. |
This reverts commit 58aac2d.
|
|
||
| // Uncomment this block to write out the decoded images for visual/manual inspection | ||
| // or for regenerating expected ground truth PNGs (for example, after changing decoder logic). | ||
| #if 0 |
There was a problem hiding this comment.
I forgot to mention that i used this commented blocks to be sure libspng and libpng loads exact same data. first i set 1 the upper block with OpenCV compiled with libpng. and for each pngsuite file saved what loaded with each flag. then set 0 the upper block and set 1 the lower block with OpenCV compiled with libspng...
The purpose of the PR was to load bit-exact compatible results with libspng and libpng. To test this,
Imgcodecs_Png_PngSuitewas improved. Files containing gamma correction were moved to a separate test calledImgcodecs_Png_PngSuite_Gammabecause the logic created for the other files did not apply to those with gamma correction. As a result, libspng now works in bit-exact compatibility with libpng. The code can be refactored later.Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
Patch to opencv_extra has the same branch name.