CS129 / Project 6 / Automated Panorama Stitching
The Algorithm
Defined Correspondences between 2 Images (paper)
- ANMS or adaptive non-maximum suppression, is applied to the interest points generated by the Harris algorithm (which returns the locations of these interest points). ANMS ensures that we have a good spatial spread of interest points.
- Determine features of images. In this case, the descriptor is quite simple. 40x40 pixel patches are taken from the image for the interest points derived from Harris and then cleaned up by ANMS. These patches are pyramid reduced and then resized until they are of 8x8 dimensions. Then, the mean is subtracted from the patch, and then it is divided by the standard deviation.
- Reject False Matches using Lowe's Ratio Test: Let us note that the "the probability that a match is correct can be determined by taking the ratio of distance from the closest neighbor to the distance of the second closest." With this in mind, we can reject many of the points. The ideal ratio, as defined by Lowe in this paper, is approximately 0.8, although I modify this depending on the test case and how picky I want the algorithm to be.
Recover the Homography using RANSAC
- Choose 4 random feature pairs from each of the two images and compute the exact homography H.
- Then, compute the inliers where the SSD of Pi' and H*Pi is less than some epsilon value.
- Iterate a fixed number of times (1000) for steps 1 and 2.
- Keep the largest set of inliers
- Recompute the homography (least-squares H estimate) on all of the inliers.
Results
Own images. In sets of three (of paintings by Waterhouse).
Defined Correspondences between 2 Images (paper)
- ANMS or adaptive non-maximum suppression, is applied to the interest points generated by the Harris algorithm (which returns the locations of these interest points). ANMS ensures that we have a good spatial spread of interest points.
- Determine features of images. In this case, the descriptor is quite simple. 40x40 pixel patches are taken from the image for the interest points derived from Harris and then cleaned up by ANMS. These patches are pyramid reduced and then resized until they are of 8x8 dimensions. Then, the mean is subtracted from the patch, and then it is divided by the standard deviation.
- Reject False Matches using Lowe's Ratio Test: Let us note that the "the probability that a match is correct can be determined by taking the ratio of distance from the closest neighbor to the distance of the second closest." With this in mind, we can reject many of the points. The ideal ratio, as defined by Lowe in this paper, is approximately 0.8, although I modify this depending on the test case and how picky I want the algorithm to be.
Recover the Homography using RANSAC
- Choose 4 random feature pairs from each of the two images and compute the exact homography H.
- Then, compute the inliers where the SSD of Pi' and H*Pi is less than some epsilon value.
- Iterate a fixed number of times (1000) for steps 1 and 2.
- Keep the largest set of inliers
- Recompute the homography (least-squares H estimate) on all of the inliers.
Results
Own images. In sets of three (of paintings by Waterhouse).