Abstract:
Robustness of Augmented Reality (AR) applications depends heavily on image registration procedures. The registration process in AR either make use of manually placed markers in the scene or use natural features extracted from the scene. These markers or detected features estimate the correct position of virtual objects that are to be integrated with the view of real environment. Use of markers for this purpose have only limited applicability. Therefore, for incorporating AR in a wide variety of applications, there is a need for detecting affine invariant and stable natural features from an image.
This paper presents a comparative study of six feature detectors, which could be used for the 3D registration process in AR. These feature detectors are applied on 48 images (eight image-sets with six images in each set) varying in terms of change of viewpoint, scale, blur, illumination and compression ratio. Detectors chosen for comparison are Harris-Affine, Hessian-Affine, MSER, SIFT, ASIFT and SURF. The novelty of the work done is the usage of image quality metrics and Pearson Coefficient to study the traits of number of detected keypoints in an image and number of matches between two images with respect to image quality. RANSAC algorithm is used for determining correct number of correspondences between two images by eliminating outliers.
Experimental results show that in most cases, performance of feature detectors could be correlated with the quality of images used for experimentation. The comparative evaluation of feature detectors is also done with respect to computational complexity to reason detector's applicability or inapplicability in an AR system.