This paper explores the problem of object recognition from multiple observers. The basic idea is to overcome the limitations of the recognition module by integrating information from multiple sources. Each observer is capable of performing appearance-based object recognition, and through knowledge of their relative positions and orientations, the observerrs can coordinate their hypotheses to make object recognition more robust.

A framework is proposed for appearance-based object recognition using Canny edge maps that are effectively normalized to be translation and scale invariant. Object matching is formulated as a non-parametric statistical similarity computation between two distribution functions, while information integration is performed in a Bayesian belief net framework. Such nets enable both a continuous and a cooperative consideration of recognition result. Experiments which are reported on two observers recognizing mobile robots show a significant improvent of the recognition results.