Appearance-based neural image processing for 3-D object recognition and localization

This paper presents an appearance-based neural image processing system for the recognition and localization of 3-D objects. All the objects are placed on the table and can be moved arbitrarily around, allowing both in-plane and out-of-plane rotations. Instead of doing object segmentation and object specific geometric modeling, objects are directly modeled by their appearance. First a principal component network is configured to generate a nonlinear filter. Then the objects are represented in a feature vector derived by hierarchical nonlinear filtering of the input image. With this compact feature vector and a small set of training samples, a neural classifier is configured for recognition purpose. Based on the same feature vector, object is localized by several neural pose estimators. Results for the recognition and localization of a large number of real images under heterogeneous background are shown.