SuperPixelSLIC accepts both 32F and 8U images but does not scale parameters accordingly

[oor001]

System information (version)

• OpenCV => 3.2.0
• Operating System / Platform => Ubuntu 18.04
• Compiler => GCC 7.3.0

Detailed description

When using the SuperPixelSLIC class to generate superpixels for an RGB image I have encountered the following phenomenon:

Code:

cv::Ptr<cv::ximgproc::SuperpixelSLIC> spixel = cv::ximgproc::createSuperpixelSLIC(rgb, cv::ximgproc::SLIC, 15, 50.0);
spixel->iterate();
cv::Mat cv_labels_32SC1;
spixel->getLabels(cv_labels_32SC1);    

Using the above code, I get two different results based on the format of the input image:

a) If 'rgb' is a CV8UC3 (0-255) then I get a desired result. Superpixels conforming to the shapes within the image. b) If 'rgb' is a CV32FC3 (0-1) then I get a slightly perturbed grid as a result.

This leads me to believe that whatever code is weighting the image structure vs compactness within the function is not taking into account the scale of the image. It is fine if this is the case, but possibly enforcing/documenting the incoming image format would reduce confusion here.

[cbalint13]

@oor001 ,

https://docs.opencv.org/trunk/df/d6c/group__ximgproc__superpixel.html

• Doc recommends to have LAB (even instead of RGB as input image & color space)

• CieLAB should be linear in perception (better than RGB), SLIC class superpixel use euclid distance on some sort of color features (or derived features like in case of LSC).

  • Dealing with float input (e.g. special imagery), or 16 bit (HDR) should be treated with care, i assume that in such special image cases user knows what he do and what he expect in terms of variance and spectra.

  Well, allowing any input format its a feature but can turn out as a "bug" if missunderstood :-)