Constant Tangential Angle Elected Interest Points
Résumé
This paper presents a new interest points detector whose goal is to better catch visual attention than standard detectors. It is based on the generalization of Loy and Zelinsky transform which was originally developed to detect radial symmetry centers. The voting algorithm is here extended to a set of voting orientations which allows the construction of a more complete analyze space. The specificity of this detector is that it enables the detection of points with different topological natures with the same detector. Centers of geometric forms or curves as well as junctions or vanishing points can be detected depending on the image content. The paper describes and illustrates the principle of this new algorithm and discusses some properties and improvements. Preliminary experiments are also given and shows that the detected points are stable under several typical transformations and that they are promising in terms of objects generic subparts detection.