Image registration is the process of coordinating images of the same scene taken from different views, at different times, in different locations, or using different sensors. This process greatly helps with image fusion in remote sensing, medical imaging, and computer vision applications. Feature detection and feature matching are the two most important steps in image registration. Scale-invariant feature transformation (SIFT) is a reliable method for image registration that yields robust results that are invariant to distortions from rotation, illumination, noise, and other scene effects. Belief propagation (BP) is used to calculate marginal distribution, and to make inferences on statistical models such as Markov models.

Even though SIFT is a reliable descriptor for image registration, it has only limited capability to represent geometric information. To address this, the authors of this paper propose the interesting idea of incorporating the geometrical information of SIFT using BP. They use BP, along with SIFT, to solve a global optimization problem iteratively.

I like the presentation style, particularly how the authors use mathematical equations to explain complicated processes. They chose challenging datasets to present their experimental results and took computational time into account. Overall, I recommend this very useful paper to researchers and practitioners who work in image fusion and image registration.