Multi-instance multi-label learning based on Gaussian process with application to visual mobile robot navigation

Classification problems have been frequently encountered in visual mobile robot navigation. The studies reported so far are mainly focused on the single label problem; i.e., each sample (datum) is assigned to a single class. In the case when a sample belongs to multiple classes simultaneously, most existing approaches attempt to avoid handling this situation by labeling the samples subjectively with the base class, which is the most obvious to them, or by considering them as a new class. In this paper, a new multi-instance multi-label learning (MIML) algorithm, called MIMLGP, is proposed by using Gaussian process (GP) for solving the multiple labels problems in visual mobile robot navigation. Compared with the existing multi-label (ML) algorithms, the MIMLGP method represents each sample with multiple instances so that higher accuracy may be achieved. Moreover, correlations between the labels associated to the same sample, which are crucial to multi-label learning but rarely considered before, are analyzed by using a covariance matrix present in MIMLGP. Experiments in the area of place recognition and terrain classification have been conducted to substantiate the proposed algorithm. The experimental results show that the proposed algorithm can achieve better performance than the one produced by the existing algorithms. (C) 2011 Elsevier Inc. All rights reserved.