Novel concepts for recognition and representation of structure in spatio-temporal classes of images

Abstract

This paper discusses open problems and future research regarding the recognition and rep- resentation of structures in sequences of either 2D images or 3D data. All presented concepts aim at improving the recognition of structure in data (espe- cially by decreasing the influence of noise) and at extending the representational power of known de- scriptors (within the scope of this paper graphs and skeletons). For the recognition of structure critical points of a shape may be computed. We present an approach to derive such critical points based on a combination of skeletons and local features along a skeleton. We further consider classes of data (for example a temporal sequence of images of an ob- ject), instead of a single data sample only. This so called co-analysis reduces the sensitivity of analysis to noise in the data. Moreover, a representative for a whole class can be provided. Temporal sequences may not only be used as a class of data in a co- analysis process - focusing on the temporal aspect and changes of the data over time an analysis of these changes is needed. For this purpose we explore the possibility to analyse a shape over time and to derive a spatio-temporal representation. To extend the rep- resentational power of skeletons we further present an extension to skeletons using model fitting.