A bootstrapped modular learning approach for scaling and generalisation of grey-level cornerdetection

Abstract

In this paper, we study a bootstrapped learning procedure applied to corner detection using synthetic training data generated from a grey-level model of a corner feature which permits sampling of the pattern space at arbitrary density as well as providing a self-consistent validation set to assess the classifier generalisation. Since adequate learning of the whole mapping by a single neural network is problematic we partition data across modules using bootstrapping and which we then combine by a meta-learning stage. We test the hierarchical classifier on real images and compare results with those obtained by a monolithic network.