Can we use 16-bit RGB image as input without converting into 8-bit RGB(by libpng)?

[Starry-lei]

Hi, I think the code here:
" image->SetSize(width, height);
png_bytep* row_pointers = new png_bytep[height];
for (u32 y = 0; y < height; ++ y) {
row_pointers[y] = reinterpret_cast<png_bytep>(image->row(y));
}
png_read_image(png_ptr, row_pointers);
"

will convert 16-bit RGB image into 8-bit RGB right?

The libpng api: void PNGAPI
png_read_image(png_structrp png_ptr, png_bytepp image) require png_bytepp( uchar pointer pointer).

Is there any easy way to change the above code so that it supports 16-bit RGB image as input?

[puzzlepaint]

The PNG loading code is written such that it always tries to convert the file's format to the format of the Image object which the file is loaded into. So, if that works as intended, with a 3-channel 16-bit Image used as output, loading the images might already work.

However, to actually use 16-bit RGB images in the SLAM algorithm would require adjusting other places in the code. As a start, this definition has Vec3u8 (a 3-channel 8-bit unsigned char vector) hard-coded as color type and would need to be adjusted to something like Eigen::Matrix<u16, 3, 1>:

badslam/applications/badslam/src/badslam/main.cc

Line 508 in c66bdc8

RGBDVideo<Vec3u8, u16> rgbd_video;

And then any potential other places that might assume that the color is 8 bits per channel need to be adjusted. For example, the color storage in keyframes:

badslam/applications/badslam/src/badslam/keyframe.h

Line 230 in c66bdc8

CUDABuffer<uchar4> color_buffer_;

[Starry-lei]

I adjusted a lot of places that assume the color image is 8-bit RGB along the dirction of data flow, but I got stuck here:

badslam/libvis/src/libvis/cuda/cuda_buffer.cu

Line 154 in c66bdc8

template class CUDABuffer_<uchar4>;

Once I extend the "template class CUDABuffer_<uchar4>;" to "template class CUDABuffer_<ushort4>;", an error

""Invalid operands to binary expression ('ushort4' and 'float')" "

occurs at :

badslam/libvis/src/libvis/cuda/cuda_buffer.cu

Line 89 in c66bdc8

buffer(y, x) = factor * tex2D<float>(texture, x + 0.5f, y + 0.5f);

How do you think I should adjust this?
Hmm, does that mean I should override operator "*"? Does it make sense for next steps?

[puzzlepaint]

No, look here instead:

badslam/libvis/src/libvis/cuda/cuda_buffer.cu

Line 124 in c66bdc8

// Avoid compilation of some functions for some types by declaring but not

This list contains an entry for uchar4:

badslam/libvis/src/libvis/cuda/cuda_buffer.cu

Line 132 in c66bdc8

template<> void CUDABuffer_<uchar4>::SetToReadModeNormalized(cudaTextureObject_t texture, cudaStream_t stream);

So it probably makes sense to add an entry for ushort4 there as well. This should cause the function containing line 89 not to be compiled for that type.