We focus on two aspects of image understanding with broad recent interest and bright scientific perspective, but also limitations which need to be overcome. These are particularly appearance based approaches for object extraction and classification as well as statistical and combinatorial modeling. We illustrate the paper with examples of our work on building façade interpretation resting on developments in three-dimensional (3D) reconstruction from uncalibrated image sequences. Particularly, we introduce implicit shape models as well as Random Sample Consensus (RANSAC) and Markov Chain Monte Carlo (MCMC). We finally note, that in spite of the large progress in image understanding in recent years, the gap to what people would like to have still seems to widen.

Artikelauszug / Extract:

1. Introduction
The title given by the organizers issues a challenge which we were happy to take up on one hand, but which is also hard to fulfill on the other hand. We therefore decided to limit our scope to two areas with broad recent interest, namely appearance based approaches for object extraction and classification as well as statistical and combinatorial modeling. Appearance based approaches, i.e., approaches, where image information is directly used to model an object, have been around for a while. Yet, only with recent work they have become a main stream of research focusing on the extraction of objects from images, e.g., (Agarwal et al. 2004, Leibe et al. 2004) and to some extent (Lowe 2004), but also on the classification of whole images. For the latter, Fei-Fei et al. (2004) have shown how to incrementally learn and discriminate 101 object classes. For all the above approaches, the basic idea is to combine the comparison of small patches of the image around salient points with the modeling of the spatial arrangement of these points. The big advantage of doing so is, that the model can be learnt automatically from images or their parts tagged to be examples of a certain class.