Sampling of Web Images with Dictionary Coherence for Cross-Domain Concept Detection

Abstract

Due to the existence of cross-domain incoherence resulting from the mismatch of data distributions, how to select sufficient positive training samples from scattered and diffused web resources is a challenging problem in the training of effective concept detectors. In this paper, we propose a novel sampling approach to select coherent positive samples from web images for further concept learning based on the degree of image coherence with a given concept. We propose to measure the coherence in terms of how dictionary atoms are shared since shared atoms represent common features with regard to a given concept and are robust to occlusion and corruption. Thus, two kinds of dictionaries are learned through online dictionary learning methods: one is the concept dictionary learned from key-point features of all the positive training samples while the other is the image dictionary learned from those of web images. Intuitively, the coherence degree is then calculated by the Frobenius norm of the product matrix of the two dictionaries. Experimental results show that the proposed approach can achieve constant overall improvement despite cross-domain incoherence.