A quantitative metric of visual-words separability for a more discriminative visual vocabulary in an unsupervised manner

The task of visual vocabulary construction plays an important role in the bag-of-words based pattern analysis and robotic applications. A discriminative vocabulary generation in unsupervised case is an open issue for reducing perceptual aliasing in image matching based applications. In this paper, we present a scheme to evaluate the discriminative power of each visual word quantitatively in terms of Mahalanobis separability, and a discriminative visual vocabulary is obtained through adaptively updating the poor discriminative visual words in an unsupervised manner. The effectiveness of our metric is demonstrated in the experiment of loop-closure detection under strong perceptual aliasing condition in both indoor and outdoor image sequences.