Automatic acquisition of object models by relational learning

Abstract

An intelligent visual information processing system should have the ability to understand its visual inputs. The input contents may be texts, drawings, or images. To recognise such inputs successfully, the system usually contains a priori knowledge about the class of possible inputs. This knowledge is normally hand-coded by experts. Hence, the approach is error prone, time-consuming, and requires considerable expertise. To solve these problems, researchers have proposed the use of learning methods to acquire this knowledge. This paper introduces a methodology to automatically acquire (learn) this prior knowledge (models) for a system which has the capability to recognise objects in images. Recent efforts to learn such models suffer from drawbacks. They construct models of two-dimensional objects, or use CAD designs of the object to build the model. Some have used attribute-value learners as their learning tool. Moreover, models have been often represented as graphs. Our system has the capability to learn three-dimensional object models from real images by using a relational learning system. Object features are first extracted, and the relations between them are found. These relations are then converted to symbolic form, and fed to FOIL, a relational learning system. FOIL produces definitions of objects which may be used during the object recognition phase.