Reification through perceptual grouping

When humans perceive incomplete or ambiguous informations, they tend to instantaneously resolve them by creating meaningful completions. In psychological terms, this generative aspect of perception is called reification. In this paper, we present an approach to model reification by means of perceptual grouping. This technique plays an important role in human-robot interaction scenarios, because it allows technical systems to gain a similar believe state as humans when it comes to complete sparse informations, leading to more natural interactions. Employing a recurrent neural network, the Competitive Layer Model, we realize robust perceptual grouping of features and show how this model is extended to generate feature amendments to obtain a complete perceptual impression. After elucidating the theoretical principles, the approach is evaluated in a human-robot interaction scenario.