Calculating the median at each pixel for multiple frames OpenCV (C++)

Asked Jul 21 '20 at 10:49

Active Sep 06 '22 at 04:51

Viewed 394 times

The task is to extract stable background out of the video, the idea is to choose n random frames from the video and take median for each pixel. I was doing this task in python using numpy

medianFrame = np.median(frames, axis=0).astype(dtype=np.uint8)

But now i need to perform the same task in C++. I have tried the naive way of splitting channels and going through rows*cols number of pixels for n frames to calculate the median frame but its not efficient at all and takes way more time than what np.median was taking. I also try to use xtensor for performing the task but wasn`t able to worlk with it. Any suggestions or direction about how to approach this task would greatly help me, Thanks!

edited Sep 06 '22 at 04:51

Christoph Rackwitz

11,317
4
27
36

asked Jul 21 '20 at 10:49

Syed Zohaib Hassan

I think for the task above, your way is correct and it is true that it will take more time than python. Basically numpy library is built on C++ but it deals with bits of the data and thus is fast. – Rahul Kedia Jul 21 '20 at 10:53
So you wanna calculate median for each pixel coordinate in n different frames? If so, you can calculate it by reaching each frame's pixel by one by – Yunus Temurlenk Jul 21 '20 at 10:53
Have you tried using matrix addition? – Rahul Kedia Jul 21 '20 at 10:54
@YunusTemurlenk thats exactly what i am doing right now and its extremely expensive computationally – Syed Zohaib Hassan Jul 21 '20 at 11:02
see [here](https://stackoverflow.com/a/38916002/5008845) – Miki Jul 21 '20 at 11:03
@RahulKedia No, i have not matrix addition, how can i exploit that for calculating median? – Syed Zohaib Hassan Jul 21 '20 at 11:03
I'd try 'cv::Mat::forEach' combined with 'std::nth_element'. ForEach is multithreaded in the background so I'd expect faster calculation. I assume that accessing (read-only) the corresponding elements in sequence of images will be safe in this context (it has to be done in foreach body so it can be invoked on other thread). – michelson Jul 21 '20 at 11:58
What is the number of frames you are talking about ? – ZdaR Jul 21 '20 at 14:09
Are you using pointer arithmetic for image traversal? – stateMachine Jul 21 '20 at 23:25
@Zdar there will be 25 at least FHD frames. – Syed Zohaib Hassan Jul 22 '20 at 12:50
@eldesgraciado yeah but now i am trying to work with cv::Mat::forEach, suggested by michelson which is much faster – Syed Zohaib Hassan Jul 22 '20 at 12:52
any of you that suggested `std::nth_element`, perhaps turn that into an answer – Christoph Rackwitz Sep 06 '22 at 04:50

Calculating the median at each pixel for multiple frames OpenCV (C++)

0 Answers0