Active learning using a Variational Dirichlet Process model for pre-clustering and classification of underwater stereo imagery

This paper demonstrates an implementation of pool-based active learning through uncertainty sampling using a Variational Dirichlet Process (VDP) model. The VDP is used for both pre-clustering and classification, and is extended to incorporate fixed labels from an oracle (human annotator). Three different uncertainty sampling techniques are explored - least confident sampling, margin sampling and entropy based sampling. Clustering with the VDP is done in a completely unsupervised manner, without the need to specify the number of clusters. This appears particularly useful in improving the results when there are only few labelled samples, or if the cost of labelling is high. Results are shown for a toy dataset and the performance is compared to similar implementations using an Expectation Maximisation model (EM) and a Naive Bayes classifier (NB). The VDP active learning framework is tested on a stereo image dataset obtained by an autonomous underwater vehicle that covers several linear kilometres and consists of thousands of stereo image pairs. Our results show that combining an active learning strategy with the VDP significantly reduces the number of labelled images required to achieve a desired level of accuracy.