SCOTLAND

This paper presents a novel approach to the recovery of generic solid parts of objects from real 2D images. The part vocabulary chosen is the one of geons, which are qualitative volumetric part primitives that are dened by simple but perceptu-ally relevant properties which are viewpoint quasi-invariant. Most previous works on detection and recognition of geons from 2D images relied on quasi-perfect line drawings. The use of aspects has also been proposed for matching xed templates of synthetic images. Here we use parametrically deformable aspects as 2D models to be matched to real images of geons in the framework of Model-Based Optimisation. The use of parametric models allows us to eciently represent geons, whereas the use of topologically dierent aspects yields more robustness in the optimisation process we use, which is Adaptive Simulated Annealing. A simple control strategy is developed that generates initial aspect hypotheses followed by a maximum a posteriori choice of the best one. Experiments are shown that demonstrate the validity of the approach. The proposed method is general, in the sense that it could be easily applicable to other parametrically dened part vocabularies