Lab Course: Building 3D Models from image sequences

Description

Detailed and realistic 3D models are key-features in producing high-quality computer generated images. However, creating such models from scratch still remains a non-trivial, time-consuming task and even an artistic challenge. Therefore, there is a constantly increasing need for 3D laser range scanning devices and methods, however, the accessibility to this technology is limited by ist extreme expensiveness. Thus, it would be desirable to have a method capable of generating visually satisfying reconstructions of a scene's geometry from an input of a sequence of images taken for example with a common digital camera.

To achieve this, we theoretically only need the projective camera mappings. Having these, by utilizing the inverse camera mapping, we project the two corresponding image-points back into the 3D space and thus obtain two rays intersecting at the location of the genuine 3D point (the so-called triangulation algorithm). By conducting this procedure for all located pairs of corresponding image points we generate a more or less dense point-cloud in the 3D space. In order to obtain a renderable model, this point-cloud has to be further processed.

In order to design and implement such a system, the following general problems have to be addressed:

• Determination of corresponding points in an image sequence
• Finding the camera matrix for each image
• Converting the retrieved data into a renderable model

After a discussion of existing and novel solutions, the goal of this practical course will be the implementation and integration of these methods into a modular system for "3D-reconstruction from images".