Local descriptors robust to out-of-plane rotations

The process of matching local descriptors which are invariant to affine transformation is an interesting problem with many practical applications. Currently, many state-of-the-art approaches are robust to scale changes and in-plane rotations, but they are not dealing well with out-of-plane rotations. In order to solve this problem, we are proposing the method of simulated local deformations of the image patches, which improves the robustness of the local descriptors in the mentioned problem. By performing rotations locally and then storing descriptors of these rotated patches, one obtains a richer representation, which results in better matching. We tested our method on a publicly available dataset of deformable objects and obtained improved matching results. In future studies, we are aiming at optimising this approach through parallel computing.