Uncertain Reasoning and Learning for Feature Grouping

摘要

Hierarchical perceptual organization is needed for 3-D object detection and description. The hypothesize and verify paradigm offers one approach to this task. Hypotheses are generated from simpler features satisfying some possibly task-dependent properties. More global evidence is used to verify and assign a confidence measure to the hypotheses. This evidence may consists of several components which may be of very different natures. How to combine these components in an effective and efficient manner becomes of critical interest. We describe formal methods that use neural networks and Bayesian approaches for verification. These approaches also allow automatical learning of parameters from some examples. We illustrate these methods using a system for building detection and description from aerial images, but the techniques themselves are not specific to this domain. Experimental results indicate that substantial improvements in performance can be obtained by the use of these methods without extensive hand-tuning of parameters.