Object recognition in floor plans by graphs of white connected components

In this paper we describe the object location module of a complete floor plan understanding system. The approach relies on a Region Adjacency Graph (RAG) whose nodes correspond to connected components of the background. Two types of node attributes are proposed and are used to locate and recognize objects similar to the models or to generalize the recognition to different shapes. The tests are made on two benchmark datasets and on one collection of floor plans downloaded from Internet and containing heterogeneous images.