Seeing Through the Blur

Coordinated Science Laboratory was formerly known as Control Systems LaboratoryThis paper addresses the problem of image alignment using models such as affine and homography and by directly using pixel intensity values. Coarse-to-fine scheme has become a standard for direct intensity-based alignment. It is believed that such coarse-to-fine scale sampling (Gaussian blur) can improve region of convergence of the alignment optimization. Although, it has been proposed that such isotropic blur may not be optimal for some motion models, no rigorous derivation for such kernels has been known to date. In this work, we derive kernels for some of the common motion models such as affine and homography, which are able to smooth the alignment objective function. This is appealing because the smoothing process often removes poor local minima and thus reaches deeper solutions. Our derivation shows that these kernels coincide with Gaussian blur of the image only for displacement motion.National Science Foundation / NSF IIS 11-1601