search

wosugi / compressive-bilateral-filter

This is an official sample code of the compressive bilateral filter, which is published on IEEE Transactions on Image Processing in 2015. It is open to public under the MIT license.
MIT License
17 stars 6 forks source link

Incorrect PSNR reporting from the c++ code #3

Open seannz opened 4 years ago

seannz commented 4 years ago

Sugimoto-san,

There seems to be an issue with the PSNR calculation in the C++ version. In the main.cpp::test_bilateral_filter function, cv::imread returns a 4-channel matrix even when a 3-channel (RGB) PNG is loaded (the 4th channel seems to be an empty alpha channel). As a result, the MSE is under-reported by a factor of ~4/3=1.333, making PSNR also seem higher.

Best, Sean Young,

seannz commented 4 years ago

I just checked again, and cv::imread returns a matrix with the correct number of channels. However, if you check from within calc_psnr(), image1 and image2 are 4-channel matrices (this is true even for grayscale images).

wosugi commented 4 years ago

Hello, seannz-san,

Thanks for your comments! I've checked the latest rep with a slight modification of setting to OpenCV4.2 and VS2017. The results seem to be good as far as I trace. I mean a matrix after imread() and two matrices in calc_psnr() are 3ch images when loading a 3ch PNG image (I check the bit-depth is 24 from file property dialog). May I ask you to send me an image file you are using?

seannz commented 4 years ago

Here's an image that I used for testing: http://r0k.us/graphics/kodak/kodak/kodim01.png

Thank you!

wosugi commented 4 years ago

I've tested the image and got the correct results (all the matrices mentioned above are 3-channel). I'm not sure why your code produces 4-channel images, but I'm happy if you correctly modify the code when seeing this strange behavior. Thanks!