Abstract
The problem of searching for a model-based scene interpre- tation is analyzed within a probabilistic framework. Ob ject models are formulated as generative models for range data of the scene. A new statis- tical criterion, the truncated ob ject probability, is introduced to infer an optimal sequence of ob ject hypotheses to be evaluated for their match to the data. The truncated probability is partly determined by prior knowl- edge of the ob jects and partly learned from data. Some experiments on sequence quality and ob ject segmentation and recognition from stereo data are presented. The article recovers classic concepts from ob ject recognition (grouping, geometric hashing, alignment) from the proba- bilistic perspective and adds insight into the optimal ordering of ob ject hypotheses for evaluation. Moreover, it introduces point-relation densi- ties, a key component of the truncated probability, as statistical models of local surface shape.