ABSTRACT
Nowadays, cameras are ubiquitously available and billions of photos and videos are uploaded every year to photo and video sharing sites. Combined with the recent progress in computer vision, this has lead to the development of large-scale 3D modeling from these images with approaches proposed by Snavely et al. [Snavely et al. 06], Agarwal et al. [Agarwal et al. 11], and Frahm et al. [Frahm et al. 10]. Figure 7.1 illustrates a dense 3D model obtained from a photo collection modeling of Frahm et al. [Frahm et al. 10]. These reconstruction methods all rely on the registration of the cameras (determination of their relative/absolute poses at capture and their internal camera parameters) in a common coordinate system. Once the registration is known, the cameras can be leveraged within dense scene geometry estimation (see Chapter 8 for more details on dense depth estimation) to obtain a dense 3D model. Beyond dense depth estimation, camera registration is also required for a broad range of applications like sensor fusion with camera images (see Chapter 9), reconstruction of dynamic structure (see Chapters 11, 12), and more. The automated process of jointly esti-
Figure 7.1: An example dense 3D model from an Internet photo collection of Rome, Italy, computed by the method of Frahm et al. [Frahm et al. 10].
