Multiscale descriptors and metric learning for human body shape retrieval

The aim of this paper was to show the usefulness of applying feature projection or metric learning techniques to multiscale descriptor spaces for the effective retrieval of human bodies of labeled subjects. Using learned subspace projections it is possible to strongly improve the retrieval performance obtained with state-of-the-art global descriptors, and, in some cases, to perform an effective feature fusion. Results obtained on different human scan datasets show that Linear Discriminant Analysis, applied to Histograms of Area Projection Transform and Shape DNA features after a preliminary dimensionality reduction, creates compact descriptors that are quite effective in improving the subject retrieval scores both when class (subject) examples are available in the training set and when only examples of classes not included in the test set are used for training. Other mappings tested are less effective even if still able to improve the results. Retrieval scores obtained in the same experimental settings used in recent related papers show that the approach based on our mapped features largely outperforms the other methods proposed for the task, even those specifically designed for human body characterization.