Page 317 -

296 6 Feature-based alignment

overlapping images of a scene can both be used (Hartley and Kang 2007; Tardif, Sturm, and
Roy 2007). The same techniques used to calibrate for radial distortion can also be used to
reduce the amount of chromatic aberration by separately calibrating each color channel and
then warping the channels to put them back into alignment (Exercise 6.12).

6.4 Additional reading

Hartley and Zisserman (2004) provide a wonderful introduction to the topics of feature-based
alignment and optimal motion estimation, as well as an in-depth discussion of camera cali-
bration and pose estimation techniques.
Techniques for robust estimation are discussed in more detail in Appendix B.3 and in
monographs and review articles on this topic (Huber 1981; Hampel, Ronchetti, Rousseeuw et
al. 1986; Rousseeuw and Leroy 1987; Black and Rangarajan 1996; Stewart 1999). The most
commonly used robust initialization technique in computer vision is RANdom SAmple Con-
sensus (RANSAC) (Fischler and Bolles 1981), which has spawned a series of more efficient
variants (Nist´ er 2003; Chum and Matas 2005).
The topic of registering 3D point data sets is called absolute orientation (Horn 1987) and
3D pose estimation (Lorusso, Eggert, and Fisher 1995). A variety of techniques has been
developed for simultaneously computing 3D point correspondences and their corresponding
rigid transformations (Besl and McKay 1992; Zhang 1994; Szeliski and Lavall´ ee 1996; Gold,
Rangarajan, Lu et al. 1998; David, DeMenthon, Duraiswami et al. 2004; Li and Hartley 2007;
Enqvist, Josephson, and Kahl 2009).
Camera calibration was first studied in photogrammetry (Brown 1971; Slama 1980; Atkin-
son 1996; Kraus 1997) but it has also been widely studied in computer vision (Tsai 1987;
Gremban, Thorpe, and Kanade 1988; Champleboux, Lavall´ ee, Szeliski et al. 1992; Zhang
2000; Grossberg and Nayar 2001). Vanishing points observed either from rectahedral cali-
bration objects or man-made architecture are often used to perform rudimentary calibration
(Caprile and Torre 1990; Becker and Bove 1995; Liebowitz and Zisserman 1998; Cipolla,
Drummond, and Robertson 1999; Antone and Teller 2002; Criminisi, Reid, and Zisserman
2000; Hartley and Zisserman 2004; Pflugfelder 2008). Performing camera calibration without
using known targets is known as self-calibration and is discussed in textbooks and surveys on
structure from motion (Faugeras, Luong, and Maybank 1992; Hartley and Zisserman 2004;
Moons, Van Gool, and Vergauwen 2010). One popular subset of such techniques uses pure
rotational motion (Stein 1995; Hartley 1997b; Hartley, Hayman, de Agapito et al. 2000; de
Agapito, Hayman, and Reid 2001; Kang and Weiss 1999; Shum and Szeliski 2000; Frahm
and Koch 2003).

6.5 Exercises

Ex 6.1: Feature-based image alignment for flip-book animations Take a set of photos of
an action scene or portrait (preferably in motor-drive—continuous shooting—mode) and
align them to make a composite or flip-book animation.

1. Extract features and feature descriptors using some of the techniques described in Sec-
tions 4.1.1–4.1.2.