Combining Deep Feature and Handcrafted Features for Material Classification

Abstract

Material classification is a challenging problem in robot and computer vision. The deep learning methods have achieved major success in object classification, but they do not conquer in material classification. One of the main reasons is different materials may yield very similar appearance. In this paper, we propose a new method combining deep feature with geometry and texture information as handcrafted local features to improve classification accuracy. Experimental results on the datasets such as GTOS and FMD show that our proposed method achieves the best performance among the methods.