Grouping with Bias

Abstract

We present a graph partitioning method to integrate prior knowledge in data grouping. We consider priors represented by three types of constraints: unitary constraints on labelling of groups, partial a priori grouping information, external influence on binary constraints. They are modelled as biases in the grouping process. We incorporate these biases into graph partitioning criteria. Computationally this formulation leads to a constrained eigenproblem. We demonstrate the effectiveness of this algorithm on image segmentation with priors and object detection with spatial attention.