A multiple-view approach to 3D recognition based on complex model neurons1

Recent psychophysical studies support the conjecture that 3D objects are represented as multiple viewpoint-specific 2D representations. We show that a multiple-view approach to 3D recognition directly based upon the activity of model neurons similar to complex cortical cells can be both straight and tractable. For simple geometric objects 5-10 views are required to enable recognition from arbitrary viewpoints; 20-30 views are sufficient to recognize objects with higher structural complexity (e.g. aeroplanes).