Would it be possible to publish the Python implementations of ꟻLIP to PyPI? It would be really convenient to directly pip install the package from command-line.

Hello and thank you for this great tool. I was trying to generate FLIP metrics and images for a big set of images and thought I could speed things up by computing multiple comparisons at once. So instead of running flip-cuda.exe -r Reference.png -t TestX.png for every image in my set, I tried flip-cuda.exe -r Reference.png -t Test1.png -t Test2.png [...].

I attached a screenshot of the result:

While the pairwise execution produces the result I expected (hard to see in the thumbnail), the combined execution produces completely unexpected results. Did I misunderstand the purpose of supplying multiple test images or is it a bug in the tool?

Thanks for your help and the great software!

• Make solve2degree() a bit more robust by using a better 2nd-degree solver.
• Renaming of, e.g., float2color3() → floatToColor3() etc, because in this case it is a bit confusing. When doing this, we should rename so no function uses a 2 instead of To
• Remove image() constructors that are not needed?
• Test if LDR images are in [0,1] and return warning or bool or something.
• Overlapping histograms for C++ is missing.
• Fix constructor for FLIP::filename("tmp.png"); so that we do not need tmp but can set the extension (png).
• Rename saveHDROutputLDRImages() to something better, perhaps saveIntermediateLDRImages(). Same goes for corresponding variable names: returnLDRFLIPImages, hdrOutputFlipLDRImages, returnLDRImages, and hdrOutputLDRImages.
• Add more test/ref images + test of histograms.
• Handle the case when the median = 0
• Publish to PyPI
• Make it possible for Python version to run on GPU.
• Make a fast CUDA version of std::nth_element() (would make the HDR-version faster)
• Better handling of num channels and dimensions (if they differ). Do it in a single place for pyhton/cpp. If file does not exist, then we should never write any image.

Hi,

I noticed that FLIP currently crashes when OpenEXR images which contain more channels than just RGB, for example AOVs.

Steps to reproduce:

1. Download the attached example reference and test images images.zip
   The reference image has the following channels: R, G, B, sampleCount.R, sampleCount.G, sampleCount.B
   The test image has the following channels: R, G, B, absoluteRenderTime, relativeRenderTime.R, relativeRenderTime.G, relativeRenderTime.B, sampleCount.R, sampleCount.G, sampleCount.B
2. Run flip.exe --reference reference.exr --test test.exr
   FLIP outputs the following and crashes:

Undefined EXR channel name: absoluteRenderTime
Undefined EXR channel name: relativeRenderTime.B
Undefined EXR channel name: relativeRenderTime.G
Undefined EXR channel name: relativeRenderTime.R
Undefined EXR channel name: sampleCount.B
Undefined EXR channel name: sampleCount.G
Undefined EXR channel name: sampleCount.R
EXR channels may be loaded in the wrong order.
Insufficient target channels when loading EXR: need 10
Undefined EXR channel name: sampleCount.B
Undefined EXR channel name: sampleCount.G
Undefined EXR channel name: sampleCount.R
EXR channels may be loaded in the wrong order.
Insufficient target channels when loading EXR: need 6

The differing additional channels do not matter, FLIP also crashes when running flip.exe --reference reference.exr --test reference.exr

Expected behavior:
FLIP should not crash. It should load the OpenEXR images, recognize the RGB channels and compute the metrics, ignoring the other channels. I think it is not uncommon to have other channels than RGB in an OpenEXR image.

I tested on Windows 10

I filed a PR with a proposed fix: #31

I stumbled across this very minor nuisance when using flip:
Specifying an output directory for the error maps using the -d parameter, flip fails silently if the directory does not exist. This results in no error maps being saved.

Example:
Running flip-cuda.exe -d output -r reference.png -t test.png will not create the output directory, but instead flip will print the metrics and then exit.

Expected behavior:
Either create the directory or print a warning that the error maps will not be saved.

Hey, forgive me for a newbie question, image filters are not my strongest skill :). I have been evaluating your work a bit and it does seem to fit my case really well! I am going to continue evaluating, but at the moment my main concern is performance, as I plan to run this for a hundreds of thousands of image pairs. At least in the C++ implementation, most of the time is spent in convolution, so I wonder if there is a way to separate the "spatial", "point" and "edge" filters? Unfortunately GPU acceleration is not available in the instances I plan to run this on :(.

Hello, I tried to get this repository to work with Python venv on macOS. But having issue installing OpenEXR. Is there a method to get this repository working with Python venv on macOS?

Hello and thank you for sharing your work!
Do you thing it would be possible to apply your algorithm to a 3D image? Les's say an MRI for example.
Would you recommend any other tool for this if it weren't possible?
Thank you in advance!

Hi,

I noticed that if no inputParameters dict is supplied to flip.evalute via the Python bindings, the default argument is used. Since the default argument only exists one time in memory and is mutated, this leads to unexpected results for subsequent calls.

Steps to reproduce:

1. Download the attached images and Python Script flip-reproduce.zip
2. Execute python reproduce.py. The script calls flip.evaluate from Python with no custom input parameters. For the first call, Flip will calculate start/stop exposure (in this case nan) and will store them in the inputParameters before returning them. The second flip does not supply its own inputParameters, but since the default argument was mutated in the previous call, the values from the previous call are used. This leads to different results:

Two flip calls:

python reproduce.py
0.0 {'ppd': 67.02064514160156, 'startExposure': nan, 'stopExposure': nan, 'numExposures': 2, 'tonemapper': 'aces'}
0.0 {'ppd': 67.02064514160156, 'startExposure': nan, 'stopExposure': nan, 'numExposures': 2, 'tonemapper': 'aces'}

Skipping the first flip call:

python reproduce.py 1
0.07767405360937119 {'ppd': 67.02064514160156, 'startExposure': -4.310622692108154, 'stopExposure': 8.824731826782227, 'numExposures': 14, 'tonemapper': 'aces'}

Expected behavior
When not providing own inputParamaters, all calls to flip.evaluate should use defaults.
Flip should not reuse the computed start/stop exposure and other parameters.

Workaround
There is a workaround by passing an empty parameters dict to flip.evaluate in Python (see also workaround.py). This works as expected:

python reproduce_with_workaround.py
0.0 {'ppd': 67.02064514160156, 'startExposure': nan, 'stopExposure': nan, 'numExposures': 2, 'tonemapper': 'aces'}
0.07767405360937119 {'ppd': 67.02064514160156, 'startExposure': -4.310622692108154, 'stopExposure': 8.824731826782227, 'numExposures': 14, 'tonemapper': 'aces'}

I filed a PR with a proposed fix: #33

Hi, I'm trying to get the python wrapper of this working and everything is fine except that in flip/python/flip/main.py the first import of import pbflip errors as therre's no module called that available.

This is clearly from pybind11 in main.cpp. I've installed pybind with conda and added the directory to CMakeLists.txt but it still can't resolve the imports.

Sorry if this is a dumb question but it's not clear from the documentation in the READMEs where this module's supposed to come from