Probabilistic Modeling for Structural Change Inference

We view the task of change detection as a problem of object recognition from learning. The object is defined in a 3D space where the time is the 3rd dimension. We propose two competitive probabilistic models. The first one has a traditional regard on change, characterized as a 'presence-absence' within two scenes. The model is based on a logistic function, embedded in a framework called 'cut-and-merge'. The second approach is inspired from the Discriminative Random Fields (DRF) approach proposed by Ma and Hebert [KUMA2003]. The energy function is defined as the sum of an association potential and an interaction potential. We formulate the latter as a 3D anisotropic term. A simplified implementation enables to achieve fast computation in the 2D image space. In conclusion, the main contributions of this paper rely on : 1) the extension of the DRF to a 3D manifold ; 2) the cut-and-merge algorithm. The application proposed in the paper is on remote sensing images, for building change detection. Results on synthetic and real scenes and comparative analysis demonstrate the effectiveness of the proposed approach