NOTE; Instantiating Deformable Models with a Neural Net

is illustrated in Fig. 1. This work can be seen as a specific example of ‘‘caching’’ or ‘‘compiling down’’ the results of Deformable models are an attractive approach to recognizing objects which have considerable within-class variability such previous searches to speed up running time. as handwritten characters. However, there are severe search This paper is structured as follows: Section I describes problems associated with fitting the models to data which could our deformable models of handwritten digits and shows be reduced if a better starting point for the search were availhow they can be used for a digit recognition task. Section able. We show that by training a neural network to predict II shows how deformable models can be instantiated using how a deformable model should be instantiated from an input neural nets and reviews relevant previous work. Results image, such improved starting points can be obtained. This are presented showing that similar recognition performethod has been implemented for a system that recognizes mance can be obtained using the neural network instantiahandwritten digits using deformable models, and the results tions, but with more than a factor-of-2 reduction in the show that the search time can be significantly reduced without search time required. Section III discusses the results and compromising recognition performance.  1997 Academic Press outlines possible extensions to the method.