文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > 人文社科 > 文学研究 > Comparing and Evaluating Interest Points
侵权投诉

Comparing and Evaluating Interest Points

Comparing and Evaluating Interest Points Cordelia Schmid, Roger Mohr and Christian Bauckhage INRIA Rhone-Alpes 655 av. de l'Europe, 38330 Montbonnot, France Cordelia.Schmid@imag.fr Interest points are low-level features where the signal changes two-dimensionally. This paper shows that interest points are geometrically stable under di erent transformations and that the information contained in these points is high. To measure these two properties quantitatively, we de ne two evaluation criteria: repeatability rate and information content. These two properties make interest points very successful in the context of image matching. Repeatability signi es that we detect points with the same pre-image independently of changes in the imaging conditions: a 3D point should be detected in two images if it is detected in one of them. The repeatability rate measures the percentage of points which exist in both of two images. This is important in the context of image matching: if we want to match two images, at least a subset of the points have to be the same, otherwise no feature-based matching is possible. Information content measures the distinctiveness of the local greylevel patterns at interest points. If all the interest points are detected at very similar patterns, the matching based on signal information will fail. Any point might match with any other point. Information content is measured as the entropy of local greylevel patterns for a set of interest points. The form of local greylevel patterns is captured by rotationally invariant combinations of derivatives. The quality of the interest points depends on the detector used. Many such detectors exist in the literature. We therefore compare di erent detectors according to the criteria speci ed above. We determine which of them gives the best results and show that this one satis es the criteria well.

Abstract

Keywords: Interest Points, Evaluation, Comparison, Repeatability, Information Content

1 Introduction Many computer vision tasks rely one feature extraction. Features most commonly used are line segments, regions and interest points. This paper examines characteristics of interest points. Interest points are locations in the image where the signal changes two-dimensionally. Examples include corners1 and T-junctions, as well as locations where the texture varies signi cantly. Figure 1 shows an example of interest points detected on the sower painting of\Van Gogh". In the literature many di erent interest point detectors exist. However, there exists no performance evaluation of these di erent detectors. In the following, we establish two di erent evaluation criteria and compare several di erent detectors quantitatively according to these criteria. The evaluation criteria depend on the context the points are used in. We have chosen image The terms\interest points" and\corners" are often used equivalently in the literature. In this work\interest point" is more general than\corner", in other

words\corner" is a special subclass of\interest point". 1

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email