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

Figure 2: Interest points detected by PreciseHarris on the same scene as in gure 1 under di erent imaging conditions. The rotation between the image on the left and the image of gure 1 is 155 . The repeatability rate is 92%. The scale change between the image on the right and the one in gure 1 is 1.4. The repeatability rate is 71%. Repeatability and informational content are the criteria used

for our comparison of di erent detectors. Quantitatively good results with respect to theses criteria explain why interest points have proved to be so successful in image matching. In addition, the small image support used for their detection makes them robust to occlusions. There exist other criteria to evaluate interest points. An example for another criterion is the precision of localization. An interest point has to be precisely located at the projection of a speci c 3D point in the scene. This criterion is important is the context of calibration, where the measured world points have to be detected precisely in the image. Repeatability and localization are contradictory criteria 4], as smoothing improves the repeatability and degrades the localization. The localization accuracy has been evaluated for detectors which t a template to the signal (cf. section 2.1.3). Deriche and Blaszka 7] show that their detector has a good precision in localization when synthetic noise is added. Brand and Mohr 3] measure the localization quality of their detector on real images. They use several di erent test criteria: alignment of the extracted points, precision of 3D reconstruction and a variation of the cross ratios. To our knowledge no other quantitative comparisons of interest point operators exist. In general, papers which present a new algorithm for interest point detectors just present results for a few test images. Yet visual evaluation of the results is not quantitative and is very likely to be subjective: results are compared with what the person thinks should be an interest point, and not with what is important in the context of interest points applications. This paper starts by giving an introduction on interest points as well as some implementation details for the detectors used in the comparison (cf. section 2). The repeatability criterion is de ned in section 3. We explain how it is determined experimentally and then compare the di erent detectors in the presence of di erent transformations. Section 4 explains the information content criterion. We rst show how to measure the information content of a distribution and then describe the characterization of local greyvalue patterns. The results for di erent detectors 3

Comparing and Evaluating Interest Points

Comparing and Evaluating Interest Points

上一页第3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