AH-SIFT: Augmented Histogram based SIFT descriptor
Hao Tang, Feng Tang · 2012
We propose Augmented Histogram (AH), a conceptually novel and systematic approach to enhancing the representational power of histogram-based local image descriptors such as SIFT. Our method takes a simple form that augments the histogram of local image patch features with a set of circular means and variances. We show that such augmentation is a natural result of modeling the distribution of local image patch features by a mixture of circular normal distributions learned through the expectation maximization algorithm. We show that the histogram is a degenerate case of this modeling. Extensive experiments indicate that our proposed AH-SIFT descriptor outperforms the original SIFT descriptor on the matching of real-world images that undergo various levels of geometric and photometric transformations, including blurring, zoom/rotation, lighting changes, and viewpoint changes.