Just use 2D interpolation techniques. If u and v are displacement fields along x (horizontal) and y (vertical) axes then x' = x+u and y' = y+v will be the resulting new pixels. Now, you can use warping functions without hassle with linear or cubic interpolation choices. I hope it will help you.
However, i did not understand how could you get pixel displacement (optical flow) in the order of thousands?
If the optical flow is an integer number of pixels per frame then that is of course easy. But if they are a non-integer number of pixels, then I'm afraid you will need to use a filter with a group delay equal to the computed optical flow velocity (kind of like a "local" image registration operation). If you are computing the optical flow in the frequency domain then this shouldn't be to difficult - just rotate the phase of each of the frequency components then compute the IFFT; otherwise, you will have a bit of work to do to design the filter.
Just use 2D interpolation techniques. If u and v are displacement fields along x (horizontal) and y (vertical) axes then x' = x+u and y' = y+v will be the resulting new pixels. Now, you can use warping functions without hassle with linear or cubic interpolation choices. I hope it will help you.
However, i did not understand how could you get pixel displacement (optical flow) in the order of thousands?
I agree with Sharib Ali. You should use a bi-cubic interpolation for the sub-pixel position. The integers are computed by re-adressing the pixel position in x- and y-direction. Bic-cubic interpolation is nothing else but a resampling technique with interpolates a sinc-function. This is more or less a textbook approach. But attention: Usually interpolation methods generate additional artefacts which should be reduced by an smoothing filter. A classical Gausian filter does a good job.
- Badges
- Science topic
- Similar topics
- Applied Mathematics
- Calculations