Discontinuity-Preserving Optical Flow Computation by a Dynamic Overdetermined System

This paper presents an optical flow estimation technique that improves the accuracy of existing methods in the problematic case of motion discontinuity. An initial flow estimate at a pixel is calculated from selected "reliable" pixels in the spatial neighbourhood. This initial estimate is then used to distinguish smooth and discontinuous regions. The flow estimate in a smooth region is refined using data from the temporal neighbourhood. Flow in a discontinuous region is estimated by reasoning about the local motion boundary, which can result in accurate estimation even when the optical flow constraint does not hold. By carefully structuring this computation, the method runs at the same speed as existing methods in the literature, while preliminary experiments indicate it can produce more accurate results near motion boundaries.