Affine Invariant Feature Detection Algorithm Performance Analysis

Feature detection is a fundamental problem in computer vision and pattern recognition. It is a key link in three-dimensional reconstruction, image classification and retrieval, object recognition and scene analysis. Because of this, the feature detection problem has been subject to much attention from researchers over the years and a large number of methods were proposed. In the large number of feature detection algorithms, performance evaluation of different methods undoubtedly has important research significance.This paper did some research addressed the position accuracy of three affine invariant region feature detection methods. In the previous literature, the main evaluation criterion is feature repetition rate and the position accuracy is measured by the regional overlap error. Based on this, this paper did further research on the position accuracy from the center of the regions, which is a complement for the traditional evaluation methods. Furthermore, the position accuracy of the regional center is a more appropriate performance indication for the application of three-dimensional reconstruction, image registration and so on. In addition, against the position accuracy of the matched region, we did further research. The main work is as follows:By programming, we achieved a prototype system of position accuracy evaluation for feature detection algorithm. This system is achieved by using Matlab and C mixed-language programming, which can be used for the evaluation of position accuracy for different features. Also, it can analysis the performance of matched features when combined with different descriptors.Based on the regional center position derivation, we did some research on the performance evaluation and analysis for three typical affine invariant feature regions. In the experiment, we use classic database, which contains five different image deformation and two kinds of different types of scenes. We compare different detection algorithm and analysis the characteristics. The conclusion has some guiding role for the reader.We did statistical analysis for the position accuracy of matched features. The features detected by the three affine invariant feature detection algorithms were matched by the well-know sift descriptors and the nearest neighbor matching strategy. Then, we did a statistical analysis for the position accuracy of the matched features. Results show that although most of the regional center derivation for the matched features is in low error range, but there are also a part of them is in large error range. This also shows that the range of the position accuracy of the matched features is larger.