Graph matching in 3D space for structural seismic damage assessment

One common objective in computer vision and photogrammetry is to infer higher level object structure which is not directly observable in images or other sensing data. A practical problem field for such research is seismic building damage assessment. It is possible to observe objects such as façades, roofs, or rubble piles in oblique airborne images, but whether they are part of an actually intact or destroyed building is not observable directly: only the spatial relation between those directly observable objects allows conclusions about the structural integrity of a building. In this paper we present an approach to seismic building damage assessment, where a graph-based learning technique is employed to detect and to classify building damage levels, given instances of four object classes derived by supervised classification in object space. Results show that the vague building damage level description leads to relatively low classification score (52%), when a pre-defined building outline is assumed. However, if one is independent from such a pre-segmentation, the detection and classification rate is higher (70%).