A novel object categorization model with implicit local spatial relationship

Abstract

Object categorization is an important problem in computer vision. The bag-of-words approach has gained much research in object categorization, which has shown state-of-art performance. This bag-of-words(BOW) approach ignores spatial relationship between local features. But local features in most classes have spatial dependence in real world. So we propose a novel object categorization model with implicit local spatial relationship based on bag-of-words model(BOW with ILSR). The model use neighbor features of one local feature as its implicit local spatial relationship, which is integrated with its appearance feature to form two sources of information for categorization. The characteristic of the model can not only preserve some degree of flexibility, but also incorporate necessary spatial information. The algorithm is applied in Caltech-101 and Caltech-256 datasets to validate its efficiency. The experimental results show its good performance. © 2010 Springer-Verlag.