CV_16U support

[mlee980]

Hi, sorry to bring this up again. It was asked before: https://github.com/VisionLabs/torch-opencv/issues/94 but it seems that the fix in Torch7 (torch/torch7#566) has not been release yet, I couldn't find it in the latest release.

I am facing the same problem: write and read 16bit grayscale images. any help is appreciated.

[shrubb]

Hi! You can read 16-bit images. Wrappers convert them to 32-bit float with [0..1] values. However, indeed, you can't save into a 16-bit image via torch-opencv. That's because right now it's impossible to handle them in Torch (data type missing).
I can implement a temporary cv.imwrite16bit() workaround if you really need to have 16-bit images saved.

[mlee980]

thanks for the feed back. I have a work-around for now: use c/c++ code to write 16U 16-bit images.
But I also noted that for reading 16U images, you converted to 32F and NORMALIZED the values to [0...1.0]:

// See #94
if (matArg.depth() == CV_16U) {
    matPtr = new cv::Mat;
    matArg.convertTo(*matPtr, CV_32F, 1.0 / 65535);
    this->typeCode = CV_32F;
} else {
    matPtr = &matArg;
    this->typeCode = static_cast<char>(matArg.depth());
}

I think it would be better to keep the values as they are, just convert from 16U to 32F, at least for me application.

[mlee980]

one other issue related to this: for single channel images, like the one above, it returns a HxW image, instead a HxWxC(=1) image. that would break the image:size(3) code. wouldn't it be better to have the HxWxC(1) format? that would be more consistent.

[shrubb]

Hmm, yes, probably this is a good improvement, I (or someone else) should do this. Will try to work on it as possible, help appreciated.