Detecting scale-space consistent corners based on corner attributes

Corner points are widely used as control points in computer vision, computer graphics, etc., because they are discrete and invariant to the scale and rotational change, which is an asset in many applications. Unfortunately, almost all the corner detectors failed to give consistent results over smoothing scale, which makes the use of the corners as control points unreliable for multi-resolution applications. To solve this, either adaptive smoothing or corner detection in multiple scales are applied. Both methods are very expensive computationally, and virtually impossible for real-time or time-constrained applications. From their study, the authors have reasons to believe that scale space persistent corners can be identified based on the the scale information. A new concept of significant value associated with a corner´s detectability is introduced in this paper. A simple smoothing function is used to obtain analytical results. Such obtained results proved to be representative for general smoothing functions as well. The authors show that this value is strongly correlated with the scale-space behavior of corners and can be used to predict the corner behavior over the space scale. In other words, the authors are able to use this value to identify the scale space consistent corners based on the fine scale features only. The result is consistent with the studies of other literatures on this issue, and more important the scale space behavior of corners are quantified and measured by the corner attributes and neighboring features