Performance Characterization of Landmark Operators

Within the field of computer vision the development of algorithms with predictable performance is of major concern. Before using a certain algorithm we like to know how well it performs and whether it is suitable at all for a given task. To answer these questions, we need quantitative characterizations of the performance of algorithms. This includes a statement of obtained results for certain performance measures (e.g., the localization accuracy) as well as the specification for which image classes and under which assumptions this performance can be guaranteed. On the one hand, such a characterization is a major step towards a sound description of algorithms as well as the foundation of the field of computer vision. On the other hand, it serves a clear practical need. To solve a certain task it is generally required that we have to select certain existing algorithms which solve a subtask (e.g., edge detection). However, in computer vision there are often a large number of approaches and algorithms which have been designed to solve the same subtask. Since often only scarce information is given on the performance of a certain algorithm and also comprehensive comparisons with other schemes can hardly be found, the selection of a suitable algorithm is very difficult and can actually be compared with ‘playing dice’. In practical applications it is therefore often necessary to implement several algorithms and to find out on his own which algorithm is best suited.