Optical flow computation in the presence of spatially-varying motion blur

Abstract

While most of the techniques for inferring optical flow are based on the brightness constancy assumption, various conditions including the presence of motion blur evidently violate this fundamental presumption. If the source image and the target image appear to be dissimilar due to different blur kernels, traditional methods will fail to achieve accurate results. We present a new method, MB-CLG, that considers constructing a new pair of blurred frames, followed by regular optical flow computation. The proposed method employs a coarse-tofine approach, in conjunction with a smoothness matrix to account for occluded regions. Rather than warping frames or precomputing a large grid of derivatives, MB-CLG warps the flows at each iteration. This leads to lower computational cost, and requires less data storage. Based on the results for various sequences, MB-CLG outperforms existing optical flow methods in the sense of AAE, AEP and MSE.