Recognition of shape-changes in 3D-objects by GRBF network

从图像中自动获取三维模型是一个非常困难的问题。克服三维模型需要的一种方法是利用2D视图集合（2D投影）旋转3D对象来表示对象。Poggio et al.和Maruyama et al.提出了这样一种基于视图的物体识别方法，称为GRBF （Generalized Radial Basis Function）网络，它依赖于多个2D视图而不是3D模型。该网络类似于传统的人工神经网络，然而，隐藏层中的每个单元由径向基函数（如高斯函数）表示。本文将GRBF网络应用于手部抓握、襟翼点、张开等形状变化的识别，并设计了利用GRBF捕获手部运动的系统。我们通过计算机模拟和数据手套展示了它们的性能。由于序列运动由很多帧组成，我们可以挑选出典型的几个典型的运动帧，这些帧可以应用到GRBF网络中进行学习。训练后的GRBF网络识别率达到84%以上。

It is usually a very difficult problem to acquisition of 3D models from images automatically. One way to overcome the need of 3D model is to exploit methods for representing objects by a collection of 2D views (2D projection) rotating 3D object. Poggio et. al. and Maruyama et al. have proposed such a view-based object-recognition method named GRBF (Generalized Radial Basis Function) network which relies on multiple 2D views instead of 3D models. This network resembles to the conventional artificial neural network, however, each unit in a hidden layer is represented by radial basis function such as Gaussian. In this paper, we have applied GRBF network to the recognition of hand shape-changes such as grasp, flap point and open, and have designed the system to capture motions of the hand with the GRBF. We show their performance by computer simulations and by using data glove. Since sequential motion consists of a lot of frames, we can pick up typical several canonical frames of the motion and these frames can be applied to the GRBF network for learning. The trained GRBF network could achieve a recognition over 84%.