Segment-based building detection with conditional random fields

In this paper we propose irregular graphs for building detection combining optical and interferometric SAR (InSAR) features. A Conditional Random Field (CRF) framework is first set up on a regular grid of square image patches and then compared to a CRF based on irregular image segments generated by a segmentation. We also investigate the potential benefits of an irregular graph structure in terms of a more sophisticated formulation of context within the CRF interaction potential. It is shown that computation time is significantly reduced while results are slightly improved with a graph based on image segments.