Page 318 -

6.5 Exercises 297

2. Match your features using nearest neighbor matching with a nearest neighbor distance
ratio test (4.18).
3. Compute an optimal 2D translation and rotation between the first image and all subse-
quent images, using least squares (Section 6.1.1) with optional RANSAC for robustness
(Section 6.1.4).
4. Resample all of the images onto the first image’s coordinate frame (Section 3.6.1) using
either bilinear or bicubic resampling and optionally crop them to their common area.

5. Convert the resulting images into an animated GIF (using software available from the
Web) or optionally implement cross-dissolves to turn them into a “slo-mo” video.

6. (Optional) Combine this technique with feature-based (Exercise 3.25) morphing.

Ex 6.2: Panography Create the kind of panograph discussed in Section 6.1.2 and com-
monly found on the Web.

1. Take a series of interesting overlapping photos.

2. Use the feature detector, descriptor, and matcher developed in Exercises 4.1–4.4 (or
existing software) to match features among the images.

3. Turn each connected component of matching features into a track, i.e., assign a unique
index i to each track, discarding any tracks that are inconsistent (contain two different
features in the same image).

4. Compute a global translation for each image using Equation (6.12).
5. Since your matches probably contain errors, turn the above least square metric into a
robust metric (6.25) and re-solve your system using iteratively reweighted least squares.

6. Compute the size of the resulting composite canvas and resample each image into its
final position on the canvas. (Keeping track of bounding boxes will make this more
efficient.)
7. Average all of the images, or choose some kind of ordering and implement translucent
over compositing (3.8).

8. (Optional) Extend your parametric motion model to include rotations and scale, i.e.,
the similarity transform given in Table 6.1. Discuss how you could handle the case of
translations and rotations only (no scale).

9. (Optional) Write a simple tool to let the user adjust the ordering and opacity, and add
or remove images.

10. (Optional) Write down a different least squares problem that involves pairwise match-
ing of images. Discuss why this might be better or worse than the global matching
formula given in (6.12).

Ex 6.3: 2D rigid/Euclidean matching Several alternative approaches are given in Section 6.1.3
for estimating a 2D rigid (Euclidean) alignment.