I am using VPI 3.0 Dense Optical Flow on the Orin Nx. The image below shows the current image being the largest image, the smaller image in the top left shows the current image but copied into a VPIImage to verify that the copy is correct, and the smaller image in the bottom right is the output of the optical flow using the visualization method provided in the example VPI - Vision Programming Interface: Dense Optical Flow . As you can see, the optical flow is really noisy and wrong. I have tried changing the block size to 1,2,4,8 with similar results. I have tried changing the quality to high, medium, and low with similar results. I have tried different number of pyramid levels with similar results. When I perform optical flow on the same image, the flow looks zero everywhere. When I play back the pedestrians.mp4 provided in the example, my results look like the ones online which makes me believe my implementation is correct and that the algorithm simply isn’t very robust. I have tried various suggestions and checks suggested by online forums and chat GPT, but nothing has reduced this noise. Is this noise typical of the algorithm? Is there a way to fix or reduce it? I have read that there is a way to get some form of confidence from the optical flow algorithm, but I haven’t been able to figure out where that is available.
The input image is gray image with size 1344x1008.
