Reusable Implementations are Necessary to Characterize and Compare Vision Algorithms

This paper argues that the difficulties in implementing computer vision algorithms are a major reason for the lack of research into algorithm comparison. We conclude that it is important to represent algorithms in form of reusable code. Since current vision systems do not fulfill all requirements we must pose on reusable implementations, we propose to solve the reuse problem by applying generic programming. We define two dimensional iterators, which mediate between image processing algorithms and their underlying data structures, so that the same algorithm implementation can be applied to any number of different image formats. The elegance and efficiency of this approach is illustrated by a number of useful examples. Keywords: reusable algorithms, low-level image processing 1 Introduction If you want to test an algorithm, you must first implement it. This sounds trivial, but unfortunately it isn't -- implementing vision algorithms correctly and efficiently is often difficult and tim..