Automatic scene interpretation of aerial images is a major purpose of photogrammetry. Therefore, we want to improve building detection by exploring the "life-time " of stable and relevant image features in scale space. We use watersheds for feature extraction to gain a topologically consistent map. We will show that characteristic features for building detection can be found in all considered scales, so that no optimal scale can be selected for building recognition. Nevertheless, many of these features "live " in a wide scale interval, so that a combination of a small number of scales can be used for automatic building detection. 1

The image trees described in this paper hierarchically organize image seg-ments according to scale, with the coarsest scale, the scale of the image itself, as the root of the tree and the finest scales as the leaves. The segmentation algorithm used to form the tree nodes is the sieve, a nonlinear morphological scale-space operator. The trees are a transform so it is possible to reconstruct the associated image without loss. Scale trees may have more nodes than are needed but the trees may be simplified using a standard statistical test to reduce the number of nodes without affecting the reconstructed image significantly. These simplified trees may be used to generate regions for a stereo al-gorithm that reduces the errors in the resulting disparity map particularly within sharp-edged regions with low texture – conditions where conventional methods often fail. 1

The paper addresses the problem of feature selection during classification of image regions within the context of interpreting images showing highly structured objects such as buildings. We present a feature selection scheme that is connected with the classification framework Adaboost, cf. (Schapire and Singer, 1999). We constricted our weak learners on threshold classification on a single feature. Our experiments showed that the classification with Adaboost is based on relatively small subsets of features. Thus, we are able to find sets of appropriate features. We present our results on manually annotated and automatically segmented regions from facade images of the eTRIMS data base, where our focus were the object classes facade, roof, windows and window panes. 1

The sieve is a morphological scale-space operator that filters an input signal by removing intensity extrema at a specific scale. In images, this processing can be carried out along a path-the 1D sieve- or over a connected graph-the 2D sieve. The 2D version of the sieve generally performs better; it is however much more complex to implement. In this paper we present the 1.5D sieve, a Hamiltonian path-based version of the sieve algorithm that behaves “in between ” the 1D or 2D sieve algorithms, depending on the number of paths used. Experiments show that its robustness to the presence of noise and its performance in texture classification are similar to the original 2D sieve formulation, while being much faster and simpler to implement. 1