*3.3. Parameter Refinement*

We also propose a very fast refinement algorithm inspired by Tsutake et al. [10]. It is very simplified to reduce the number of iterations. Instead of going over every 4 possible values for the affine parameter, the proposed method encoder only checks the neighbors with a step size of 2Δ, refines to Δ and then refines the quarter pixel translation parameter only once. In case the best value for the affine parameter is zero after the initial neighbor check, the refinement is aborted. In this case, only four affine prediction estimations had to be performed, much fewer than in Tsutake et al. even in the cases of an early abort.
