Panorama

Panoramas are not just planar, it can also be spherical and cylindrical.

• Capture images
• Detection and matching
• Warping => Aligning images
• Blending, Fading, Cutting
• Cropping

Given x number of views, it is possible to generate any synthetic camera view as long as it has the same center of projection

RANSAC approach

Loop till you find a convergence/popular H

1. select four feature pairs (at random)
2. Compute homography H (exact)
3. Compute inliers where:
4. Keep largest set of inliers
5. Recompute least-squares H estimate on all of the inliers

Not that there are more inliers than outliers, but that the outliers are wrong in different ways.

Panorama - From Scratch

• Load images
• Pass each image through a feature detector like cv2.ORB/cv2.SIFT that extracts the keypoints and compute descriptors
• Optionally, you can draw circles around the keypoints by passing them over the image through cv2.drawKeyPoints to visualize
• Match the keypoints using matcher like the brute force matcher cv2.BFMatcher
• Create two sequences of corresponding matched points based on what was returned by the matcher above
• Compute homography (since our objects are planar) using RANSAC to reject outliers
• Create panorama by applying the homography matrix as a warp to the images (look in lecture video for clarity)
• Apply blending and cuts for more sleek panorama