A Novel Wide-Baseline Stereo Matching Algorithm Combining MSER and DAISY

Abstract

Image1 matching is a fundamental problem in Computer Vision. In the context of feature-based matching, SIFT (Scale-Invariant Feature Transform) and its variants have long excelled in a wide array of applications. However, for wide-baseline images, it is more challenging due to large perspective distortions. In this paper, we introduce a novel affine invariant feature algorithm combining MSER (Maximum Stable Extreme Region) and DAISY descriptor for wide-baseline scenarios. Firstly, MSER algorithm is used to detect maximum stable extreme regions. Then the MSER region is fit to an elliptical shape. The ellipse area is normalized of direction and shape using the dominant orientation and the length of long and short axes of ellipse, respectively. And square feature region is resampled to a specified radius size. Finally, the DAISY descriptor is used to establish descriptor. The algorithm is validated by using benchmark images and real-world images. We demonstrate that our method has better reliability and is superior to SIFT and MSER+SIFT algorithm in the case of wide-baseline.