From descriptor to boosting: Optimizing the k-NN classification rule

The k-nearest neighbor (K-NN) framework was successfully used for tasks of computer vision. In image categorization, k-NN is an important and significant rule. However, two major problems usually affect this rule: the NN classifier vote and the metric employed to compute the distance between neighbors. This paper deals with both. First, a boosting k-NN algorithm learns the coefficients of weak classifiers, hence allowing to assign weights for k-NN votes. Second, we have recourse to metric learning: a function is trained on sets of similar and dissimilar samples to increase inter-class distances and reduce intra-class ones.