Unsupervised methods for size and shape

The aim of this article is to propose unsupervised classification methods for size-and-shape considering two-dimensional images (planar shapes). We present new methods based on hypothesis testing and the K-means algorithm. We also propose combinations of algorithms using ensemble methods: bagging and boosting. We consider simulated data in three scenarios in order to evaluate the performance of the proposed methods. The numerical results have indicated that for the data sets, when the centroid sizes change, the performance of the algorithms improves. In addition, bagging-based algorithms outperform their basic versions. Moreover, two real-world datasets have been considered: great ape skull and mice vertebrae references. These datasets have different configurations, such as multiple reference points and variability. Bagged K-means and boosted K-means methods achieved the best performance on the datasets. Lastly, considering the synthetic and real data, the bagged K-means proved to be the best method.