Extraction of Windows in Facade Using Kernel on Graph of Contours

2007

This work is about a novel methodology for window detection in urban environments and its multiple use in vision system applications. The presented method for window detection includes appropriate early image processing, provides a multi-scale Haar wavelet representation for the determination of image tiles which is then fed into a cascaded classifier for the task of window detection. The classifier is learned from a Gentle Adaboost driven cascaded decision tree [1] on masked information from training imagery and is tested towards window based ground truth information which is -together with the original building image databases -publicly available . The experimental results demonstrate that single window detection is to a sufficient degree successful, e.g., for the purpose of building recognition, and, furthermore, that the classifier is in general capable to provide a region of interest operator for the interpretation of urban environments. The extraction of this categorical information is beneficial to index into search spaces for urban object recognition as well as aiming towards providing a semantic focus for accurate post-processing in 3D information processing systems. Targeted applications are (i) mobile services on uncalibrated imagery, e.g. , for tourist guidance, (ii) sparse 3D city modeling, and (iii) deformation analysis from high resolution imagery.

Segmenting facades from buildings is important to a variety of vision applications operating in outdoor urban environments which include image-based modeling, landmark recognition, autonomous navigation and urban scene understanding. Existing method based on Structure-from-Motion relies on multiple images, and the cost of computation is high, while another method based on prior knowledge about arbitrarily shaped repetitive regions cannot handle non-repetitive facades segmentation. In order to segment facades simply, fast and accurately, we propose an automatic algorithm for facade segmen-tation, which segments building to a set of separate facades. By first detecting line segments and corresponding vanishing points from a single image, then rectifying the image to make the vertical direction be parallel to y-direction of the image, we convert the facade-segmentation problem to a simple 1D segmentation. The experiment results show that our method is accurate, robust and efficient to facades segmentation problem.

Computer Science Research Notes, 2021

For the purpose of analyzing a building according to its accessibility or structural resilience, printed 2D floor plans are not sufficient because of the missing link to semantic information. This paper tackles this issue and introduces a concept for clustering classified lines of a floor plan and for creating semantically enriched contour elements based on different image processing, computer vision and machine learning algorithms. Based on a general line clustering approach, we introduce type specific methods for walls, windows, doors and stairs. The resulting clusters are in turn used for a contour creation, which uses minimal rotated rectangles. Those rectangles are transformed to polygons that are refined using post processing steps. The approach is evaluated via positive testing using a pixel-based comparison of the process's result. For this, automatically generated as well as real world building plans are used. The final evaluation shows, that the concept reaches a confidence of >90% for door, stair and windows and only around 10% for stairs with the run-time linearly scaling with the size of the input.

Image and Vision Computing, 2007

We propose a novel and efficient method for recognition of buildings and locations with dominant man-made structures. The method exploits two different representations of model views, which differ in their complexity and discriminative capability. At the coarse level the model views are represented by spatially tuned color histograms computed over dominant orientations structures in the image. This representation enables fast and efficient retrieval of the closest match from the database. At the finer level the matching is accomplished using feature descriptors associated with local image regions. The probabilistic formulation of the matching as well as means of dealing with multiple matches due to repetitive structures is discussed. Extensive experiments on two building databases with varying degree of occlusions, viewpoint and illumination changes are presented.