Confusion about vpi calculation of optical flow

[22160115]

Hello, when I use vpi to calculate the optical flow between two images of size (320 * 180), vpi will output an optical flow image of (80 * 45), which is right. In order to obtain the optical flow image with the size of (320*180), I up-sampled the input image 4 times before calculating the optical flow. I ended up with an optical flow image of (320 * 180), but there was a problem. The optical flow image looked like a series of massed-blocks. I counted them carefully with the windows drawing tool, and the massed-blocks were exactly (4 * 4). This is quite puzzling to me as to why the output optical flow graph is a (4 * 4) Masseille block rather than a single optical flow pixel. Even if I don’t upsample, the output optical flow diagram still has (4 * 4) Masseille blocks.

It was also very inaccurate when I use optical flow for image alignment.

But using opencv is fine:

[AastaLLL]

Hi,

You can find the algorithm details below:
The algorithm splits input images into 4x4 pixel blocks.

https://docs.nvidia.com/vpi/algo_optflow_dense.html

Instead of scaling up the input image, have you tried to use the 320x180 input and scale up the 80x45 flow?

Thanks.