Optical Flow Estimation Using Tensor-Product Smoothing Splines.

It is well known that optical flow computation suffers from high noise sensitivity. There have been several attempts to compensate for this by invoking "regularization" on the resulting motion field data. These techniques provide some improvement, but are artificial in the sense that they minimize some cost criterion based on heuristics. We present a new method based on constructing a local tensor-product spline approximation to the intensity data. This technique allows us to obtain dense motion field and depth data without performing spatial regularization on the motion field.