Learning the Consensus of Multiple Correspondences between Data Structures

Identifier:  TDX:2292

Handle: https://hdl.handle.net/20.500.11797/TDX2292

Authors:  Moreno García, Carlos Francisco

Abstract:
In this work, we present a framework to learn the consensus given multiple correspondences. It is assumed that the several parties involved have generated separately these correspondences, and our system acts as a mechanism that gauges several characteristics and considers different parameters to learn the best mappings and thus, conform a correspondence with the highest possible accuracy at the expense of a reasonable computational cost. The consensus framework is presented in a gradual form, starting from the most basic approaches that used exclusively well-known concepts or only two correspondences, until the final model which is able to consider multiple correspondences, with the capability of automatically learning some weighting parameters. Each step of the framework is evaluated using databases of varied nature to effectively demonstrate that it is capable to address different matching scenarios. In addition, two supplementary advances related on correspondences are presented in this work. Firstly, a new distance metric for correspondences has been developed, which lead to a new strategy for the weighted mean correspondence search. Secondly, a framework specifically designed for correspondence generation in the image registration field has been established, where it is considered that one of the images is a full image, and the other one is a small sample of it. The conclusion presents insights of how our consensus framework can be enhanced, and how these two parallel developments can converge with it.