A Learning-based Algorithm for Geometric Labeling of Indoor Images

This paper aims to use a large set of feature descriptions as geometric cues to build the structural knowledge of an indoor image. In this paper, a large quantity of training images are used to obtain the required information through learning. We apply a multi-class version of AdaBoost with weak learners based on the decision tree to label regions in an indoor image as “ground”, “wall” and “ceiling”. Through labeling, we can estimate the coarse geometric properties of an indoor scene, which can be used in a large number of applications, such as mobile robot navigation, object detection, automatic single-view or 3D reconstruction, virtual reality, video games, etc. Key–Words: Geometric cues, Indoor image, Multi-class, Adaboost, Weak-learner, Decision tree, Labeling, Learning.