ImageDisplayerOCV - Githubissues

TorontoDeepLearning / convnet

A GPU implementation of Convolutional Neural Nets in C++

BSD 2-Clause "Simplified" License

506 stars 229 forks source link

ImageDisplayerOCV #16

Open avdmitry opened 10 years ago

avdmitry commented 10 years ago

Hello, @nitishsrivastava

I implemented ImageDisplayer using OpenCV and tested it on mnist example with display & localizer true. I also removed CImg dependency and updated install file.

nitishsrivastava commented 10 years ago

@avdmitry Thanks for the pull request. I tested it and there are some issues:

Normalization of float images: I used normalize(image, image, 0, 1, NORM_MINMAX); and 32FC3/32FC1 images.
waitKey(1) I had to use around WaitKey(30) for it to work, even then the display is not responsive and sometimes fails to update. There is also a distinctive dimming of the display window when it is being updated. This usually happens when a window becomes non-responsive. The whole need for WaitKey() seems a bit weird to me. I found the CImg GUI to be much better in this regard. I am considering keeping CImg for this purpose.

avdmitry commented 9 years ago

@nitishsrivastava,

I took a look on these issues:

yes, normalization needed. I used CV_8UC because it has less size and should work faster. Internally OpenCV anyway will convert floating 0..1 value to 0..255 char value. I also got strange behaviour when I tried to use 32FC today. May be there is some bug in my code, I'll test it more.
I don't have such issue on my PC even with waitKey(1). I also think that waitKey approach is little bit strange, but as it should be inside ImageDisplayer it should be transparent by using this class.

I'll test this code more and try implement convenient ImageDisplayer, if I can't, let keep current implementation.

avdmitry commented 9 years ago

@nitishsrivastava,

I switched to use 32FC3/32FC1 images and added normalization. There was bug, which, I think, was the cause of second issue. I fixed it, now code should work correct.