I'm getting strange results using the GPU Brox Optical Flow calculation through Emgu. Has anybody done this recently?
When the optical flow values are consistent across the image, the resulting flow values are approximately correct. When the optical flow changes in segments across the image (moving objects), the resulting optical flow values look incorrect. The optical flow image is inconsistent and scattered in areas where it should have a single value due to a moving object.
Here are the results from OpenCV Brox optical flow on GPU:
And here are the results from my own CPU-based implementation:
I've tried varying different parameters in the Brox calculation but nothing seems to help. Any idea what could be going wrong?