Improving the matching precision of SIFT

Zhongwei Tang (1) , Pascal Monasse (2, 3, 4) , Jean-Michel Morel (1)

1 CMLA - Centre de Mathématiques et de Leurs Applications
2 imagine [Marne-la-Vallée]
3 ENPC - École des Ponts ParisTech
4 LIGM - Laboratoire d'Informatique Gaspard-Monge

Matching precision of scale-invariant feature transform (SIFT) is evaluated and improved in this paper. The aim of the paper is not to invent a new feature detector more invariant than the others. Instead, we focus on SIFT method and evaluate and improve the matching precision, defined as the root mean square error (RMSE) under ground truth geometric trans-form. Matching precision reflects to some extent the average relative localization precision between two images. For scale invariant feature detectors like SIFT, the matching precision decreases with the scale of features due to the sub-sampling in the scale space. We propose to cancel the sub-sampling to improve the matching precision. But in case of scale change, the improvement is marginal due to the coarse scale quanti-zation in the scale space. One more sophisticated method is also proposed to improve the matching precision in case of scale change. These modifications can be easily extended to other scale invariant feature detectors.

Keywords: Matching precision; localization precision; scale space; scale-invariant feature transform (SIFT)

Domains: Computer Vision and Pattern Recognition [cs.CV]

Zhongwei Tang, Pascal Monasse, Jean-Michel Morel. Improving the matching precision of SIFT. International Conference on Image Processing, Oct 2014, Paris, France. ⟨10.1109/ICIP.2014.7026164⟩. ⟨hal-01075813⟩