Nvidia Hardware Optical flow costbuffer question

Python
1

I am using NvidiaOpticalFlow_2_0 and I am trying to access the cost buffer which shows how confident the motion tracking is


nvof = cv2.cuda.NvidiaOpticalFlow_2_0_create(

    (width, height), 

    cv2.cuda.NVIDIA_OF_PERF_LEVEL_SLOW, 

    enableCostBuffer=True

)

I have the current_frame, previous_frame, flow and cost as cv2.cuda.GpuMat

nvof.calc(current_frame, previous_frame, flow, cost)

However, the ‘cost’ array doesn’t seem to even get accessed at all, no values are being changed, I can even make it different dimensions and data types and it makes no difference

Is it possible to access the cost buffer from the Nvidia Hardware Optical flow?

2

might be a bug? the docs don’t state that were unused or anything.

you could submit an issue on the opencv_contrib repo:

3

Nvidia Hardware Optical Flow cost buffer · Issue #4093 · opencv/opencv_contrib · GitHub I opened an issue there

Some thoughts:

looking up instances of calc in nvidiaOpticalFlow.cpp, I find this:

   virtual void calc(InputArray inputImage, InputArray referenceImage,

        InputOutputArray flow, Stream& stream = Stream::Null(),

        InputArray hint = cv::noArray(), OutputArray cost = cv::noArray());

“noArray” appears nowhere else in this file, but appears in many other .hpp files under the CV_WRAP, CV_EXPORT and CV_EXPORT_W macros

In cudaoptflow.hpp , calcdoes not appear , whereas in other .hpp files, the class’s public methods with “noArray” in the initialization appear with one of those macros

4

Solved it, the docs say that the cost buffer format was CV_32SC1 , but that only applies to NvidiaOpticalFlow_1_0

for NvidiaOpticalFlow_2_0 , the cost buffer format has been changed to CV_8SC1, now its working properly for me

1 Like