A Transformation-Invariant Relaxation Scheme for Feature Mapping

Abstract

截至目前為止已有不少供幾何特徵匹配用之鬆弛法架構被提出來，這些架構雖然宣稱可以對空間轉移具有不變性，但是實際上，大部份只能處理和旋轉及平移有關的轉移，對於具有尺度因素的轉換則儘量避免，因此便有各種不同的假設被加諸於所考慮的問題，例如假設我們已知景深值，因此可以先將物件的尺度正規化後再比對，或者假設物形是完整的，於是物形和模型問的尺度比率事先可以推知。本篇文章提出一種新的架構，它能夠同時對旋轉，平移，和尺度具有不變性，此外，新架構也能處理變形物件及不完整物形。我們的實驗結果顯示新的架構確具有可行性。

A large number of relaxation schemes for feature mapping, claimed to be invariantto transformation, have been reported. However, most of them can deal with transformations involving only rotation and translation, but not scaling. To stay away from theissue of scaling, unrealistic assumptions have to be imposed, such as the conjectures thatrange data are available so that objects can be rescaled before mapping, and that objectshapes are complete so that ratios between object shapes and prototypes can be figuredout beforehand. In this paper, we propose a relaxation scheme which is able to be invariant at a time to rotation, translation, as well as scaling. In addition, the proposedscheme can also cope with shapes that may be distorted and incomplete. Our schemehas been tested on both synthetic and real data. Experimental results manifest that theproposed scheme is applicable.