cuda.warpPerspective wrong matrix type

[nbackfisch]

Hopefully it is very easy, since there was a similiar, already solved issue: #132
#132

Error here:
When using cuda.warpPerspective function, there is an error raised:
You should explicitly call download method for cuda::GpuMat object

I copy the solution suggestion that seemed to work for that other issue:

From my findings this error is due to a wrong wrapping of transform matrix passed to openCV.
Currently, in file torch-opencv/src/cudacarping.cpp, lines 46, 52, 55, the transform matrix M is converted to a GPU matrix by: M.toGpuMat().
Converting M in a "CPU" matrix via M.toMat() leads to a working version.
(openCV uses only coefficients of that matrix - see openCV source code of warp.cpp)

@shrubb: Since you have/had no GPU to test this, it sounds probable for me, that it is the reason here as well.
Btw: CPU Version of warpPerspective is running fine.

Thanks in advance!

[shrubb]

Thanks for your help! It's awesome to have enthusiastic users! :)
Hope this fix will remove the bug.

[nbackfisch]

:)
Almost: now we end up in another error
OpenCV Error: Assertion failed (type == CV_32F || type == CV_64F) in invert, file /data/nbackfisch/opencv-3.2.0/modules/core/src/lapack.cpp, line 841

I called cv.cuda.warpPerspective{original:cuda(), nil, trafo:cuda()}

So maybe some wrong casting?
Also notice that I had issues installing lapack with cv if i remember right. So error can also be on my side.

Well but I need to go now, I have to catch a train and drive 700km home now ;)

[shrubb]

Try keeping M in main memory. Does this line work?

cv.cuda.warpPerspective{original:cuda(), nil, trafo}

[apvijay]

I too end up with the same error:

b2 = cv.cuda.warpPerspective{a1,nil,M:cuda()} -- a1 is CudaTensor
OpenCV Error: Assertion failed (type == CV_32F || type == CV_64F) in invert, file /home/user/opencv-3.2.0/modules/core/src/lapack.cpp, line 841
C++ exception

If M is in main memory, the following error occurs
b2 = cv.cuda.warpPerspective{a1,nil,M}
OpenCV Error: Gpu API call (invalid configuration argument) in call, file /home/user/opencv-3.2.0/modules/cudawarping/src/cuda/warp.cu, line 260
C++ exception

[nbackfisch]

Hi,
now I also ran in the Cuda error:
OpenCV Error: Gpu API call (invalid configuration argument) in call, file /data/nbackfisch/opencv-3.2.0/modules/cudawarping/src/cuda/warp.cu, line 258

Assume again, we use M:cuda() again. Can you check, whether the original and trafo are loaded to cv_8U and still need to be converted to cv_32F for some following method (e.g. from opencv)? If so, don't forget to cast back to cv_8U ;)
I found this post, where they also face such an Assertion Failure and loading the image to cv_8U sounds reasonable to me. http://stackoverflow.com/questions/26470149/c-convert-cvmat-from-cv-8u-to-cv-32f
What do you think?