Exploring Strategies for Minimizing Overlap Between Nodes in a Multimodal Metric Tree

Abstract

Slim2-tree is a multimodal metric tree which enables video indexing and retrieval by using information from multiple modalities. Experimental results have demonstrated its efficiency when compared to other multimodal solutions. This article explores different strategies related to the use of a post-processing algorithm for the Slim2-tree - named multimodal Slim-down, which tries to minimize the overlap between tree nodes. Experiments have also shown the performance improvement obtained by the policy, in which any element that presents the larger distance value to the pivot for any modality is selected as candidate to be moved. Moreover the results are better when that policy is repeatedly used during insertion.