CUDA-accelerated volumetric optical flow estimation using the Farneback algorithm
A CUDA implementation of the Farneback optical flow algorithm  for the calculation of dense volumetric flow fields. Since this algorithm is based on the approximation of the signal by polynomial expansion it is especial suited for the motion estimation in smooth signals without clear edges.
To know more about the implementation, have a look on this OpenCV class that was used as a template for this implementation.
Development takes place on Github.
The project uses pycuda to provide a pure-python package available on PyPi
pip install farneback3d
import farneback3d ... # create some numpy volumes vol0 and vol1 (can also be pycuda GPUArrays) # set parameters for optical flow (parameters have the same meaning as in the OpenCV implementation) optflow = farneback3d.Farneback( pyr_scale=0.8, # Scaling between multi-scale pyramid levels levels=6, # Number of multi-scale levels num_iterations=5, # Iterations on each multi-scale level winsize=9, # Window size for Gaussian filtering of polynomial coefficients poly_n=5, # Size of window for weighted least-square estimation of polynomial coefficients poly_sigma=1.2, # Sigma for Gaussian weighting of least-square estimation of polynomial coefficients ) # calculate frame-to-frame flow between vol0 and vol1 flow = optflow.calc_flow(vol0, vol1)
To be implemented…
The current implementation uses a naive approach to perform the necessary convolutions. The algorithm could be sped up drastically by performing separable convolutions along each coordinate axis.
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.