18. A. Pinz: Categorization-Driven Segmentation: Active Shape Models for Categories

The localization of objects in images is still a field of current categorization research.
It would be desirable to obtain improved object segmentation for existing categorization approaches, in which category recognition and approximate object delineation may already be satisfactoy. We build on our previous results in category detection using a Boundary-Fragment-Model (BFM, Opelt et al. ECCV06, CVPR06) and in segmentation of specific objects using shape priors in a level set framework (Fussenegger et al. ACCV06). Our novel segmentation approach starts by training standard BFMs for a number of categories.
In the recognition phase, these BFMs successfuly localize instances of the learnt categories in test images. The result is a collection of boundary fragments that vote for an object centroid. In the final segmentation phase, we use these boundary fragments, smooth them, calculate a gradient vector flow field and use it as a generic shape prior for a level set based segmentation of the test image. First experiments (Fussenegger et al. ICPR06) show very accurate segmentation results for bottles, cars side and cows side. Further results for our multiclass dataset are presented in this poster. Potential future extensions include incremental, online learning of generic shape priors and handling of significant amount of occlusion.

poster.pdf