Project 6: Automated Panorama Stitching (Writeup)

Jason Pacheco (pachecoj)

April 17, 2010

Algorithm Description

The implementation is a variant of that proposed in Brown et al.. We run a Harris detector on each of the images to derive a set of interest points. For each point we perform adaptive non-maximal suppression as outlined in the paper to reduce the number of interest points to roughly 200.

Features and RANSAC

At each interest point we compute discriminative features which consist of a color histogram in an 8x8 grid around the interest point and on all three color channels. These features are obviously not invariant to rotation or scale, but they seem to work well enough.

We compute the pairwise distance between each feature in each of the images. We compute a set of candidate matches by using Lowe's method e1-nn / e2-nn. These candidates are provided to the RANSAC algorithm. We run RANSAC iterations until at least 10 "outlier" features are matched within a distance of 0.5 pixels. The resulting homography is applied to image 1 and projected into the space of image 2.

Results