Optical Flow Diffusion with Robustified Kernels

This paper provides a comparison study among a set of robust diffusion algorithms for processing optical flows. The proposed algorithms combine the smoothing ability of the heat kernel, modelled by the local Hessian, and the outlier rejection mechanisms of robust statistics algorithms. Smooth optical flow variation can be modelled very well using heat kernels. The diffusion kernel is considered Gaussian, where the covariance matrix implements the inverse of the local Hessian. Robust statistics operators improve the results provided by the heat kernel based diffusion, by rejecting outliers and by avoiding optical flow oversmoothing. Alpha-trimmed mean and median statistics are considered for robustifying diffusion kernels. The robust diffusion smoothing is applied onto multiple frames and is extended to 3D lattices.