Manifold multi-view learning for cartoon alignment

Abstract

Cartoon alignment is a key to retrieve cartoon characters and synthesise new cartoon clips. To successfully achieve the tasks, it is necessary to extract visual features that comprehensively denote cartoon characters and to align the feature points accurately between cartoon characters. In this paper, Speed Up Robust Feature (SURF) and Shape Context (SC) are introduced to characterise the cartoon character from multi-view. To increase accuracy rate of cartoon character alignment, semi-supervised alignment and Procrustes alignment require predetermining the correspondence. To overcome the flaw, we propose a Manifold Multi-View Learning (MML) to align cartoon characters. MML learns a projection that maps data instance (from cartoon characters with different dimensionality) to a lower-dimensional space, which simultaneously matches the local geometry and preserves the neighbourhood relationship within each cartoon character. The matching relationship can be obtained from local geometry structure. Experimental results show the good performance.