cv::optflow::DISOpticalFlow: got different result between cpu version and opencl version

[changingivan]

System information (version)

• OpenCV => 3.3
• Operating System / Platform => Windows 64 Bit
• Compiler => Visual Studio 2015
• OpenCL => 1.2
• GPU => Intel(R)HD Graphics 4600

Detailed description

I use the cv::optflow::DISOpticalFlow , both cpu version and opencl version to calculate optical, when the input two images is similar , the two vesions result have litter difference ,
rubberwhale1:

rubberwhale2:

cpu-opticalflow:

opencl-opticalflow:

However , when input images is larger different , the cpu version and opencl version got different result ,
left:

right:

cpu-opticalflow:

opencl-opticalflow:

I think there is a bug.

Steps to reproduce

[mshabunin]

CC @sbokov , @pengli

[yvo2m]

Well, your second dataset is not appropriate at all for OpticalFlow algorithm.
Both cpu-opticalflow and opencl-opticalflow give ridiculus/randomness results.
Optical Flow algorithm are not designed to handle such displacements between the 2 images point of view.
Better use ORB keypoints matching.

[changingivan]

@yvo2m ,I know it is not appropriate,but i think the calculate result from cpu version and opencl version should be similar even through inappropriateness.

[weidingerc]

I have the same question, is there a fix in this matter?