Neural network representation of continuous 3-D distance transform for invariant object recognition

Invariant 3-D object recognitions under detection noise and partial object viewing are difficult pattern recognition tasks. On the other hand, humans are extremely adept in performing these functions. It has been suggested by the studies of experimental psychology that the task of matching rotated and translated 3-D objects by humans is done by mentally rotating and translating gradually one of the objects into the orientation of the other and then testing for a match. Motivated by these studies, the authors present a novel and robust neural network solution for these tasks based on detected range data from 3-D objects. The method operates in two stages: the object is first parametrically represented by a continuous distance transform network (CDTN) which converts the range data (surface points) into a continuous distance transform representation of the 3-D object. When later presented with range data obtained from the distorted and partially-viewed object without point correspondence, this 3-D distance transform representation allows the mismatch information to backpropagate through the CDTN so as to gradually align the best similarity transform of the distorted object. Based on the use of distance transform representation, the distance measure between the exemplar object and the aligned distorted object can be easily computed