Fine Tuning of Quasi Linear Feature Descriptors by the Means of SSD Error Measure

This paper focuses on the optimal weighting of the components of a rotation invariant feature vector. This feature descriptor is not expected to be outstanding in performance, it is published here to illustrate the mathematics of the tuning problem. Theoretical tools are used to find proper distance functions. It is assumed, that the resulting error is properly expressed as the sum of squared pixel differences of the corresponding images. This leads to closed formulae for the elements of the weighting matrix. The deductions are intended to be general enough, enabling the application to any linear feature vector. Test results are presented to show differences between uniform and weighted distance measures. Application of the results to the Self Affine Feature Transform (SAFT) is shown briefly.