Detecting windows in city scenes

Abstract

In this paper we present an object detection system for city environments. We focus on the problem of automatically detecting windows on buildings. Several possible applications for the detection system are given, such as recognition of buildings, pose estimation, rectification and 3D reconstruction. Experimental validations on real images are also provided. The system is capable of detecting windows in images at several different orientations and scales. The approach is based on learning from examples using support vector machines. Since the system is trainable, the extension to detect other objects in the scene is straightforward. The performance of the system has been evaluated on an independent training set and the results show that the object category “window” can be reliably detected under various poses and lighting conditions.