Automatic Panoramas

Daniel Geng

Feature Detection

We use Harris corners for our feature detection with adaptive non-maximal suppresion. We will be using the following images as example:

The points of interest extracted from ANMS look like:

Correspondences

Now that we have interest points, we have to somehow find correspondences between the feature points. We do this by sampling the neighborhood of an interest point and finding a “best match” in the other image. Instead of comparing absolute squared errors, we threshold on the ratio between the first best and second best squared errors. This makes the algorithm robust.

The features for each point of interest:

And the correspondences after thresholding on the ratios:

Homography

Now with the correspondences, we can finally compute a homography using RANSAC:

Panorama using best homorgraphy from RANSAC

More Panoramas

Telegraph

Points of interest:

Features:

And the correspondences after thresholding on the ratios:

Final panorama:

Panorama using best homorgraphy from RANSAC (notice the two suns...)

Vs manual panorama:

Panorama using manual correspondence points (notice the lack of multiple suns...)

Room

Points of interest:

Features:

And the correspondences after thresholding on the ratios:

Final panorama:

Panorama using best homorgraphy from RANSAC