Object recognition using selective features extracted by predicting performance of discrimination

This paper introduces a reliable 3D position and pose recognition method for complicated scenes including randomly stacked objects. The proposed method achieves a reliable recognition by using a small number of effective features selected by evaluating performance of discrimination. The performance of discrimination is evaluated by two abilities. One is the stability of features, the other is separability of false features. For evaluating performance of discrimination, features are extracted from synthesized 3D scenes that are generated by using physics-based simulator and the 3D Computer Graphics (3D-CG) techniques. An object model's features that have high performance of discrimination in the feature space are selected by using synthesized 3D scenes' features, and selected features are used for the matching process. Experimental results using real scenes including randomly stacked objects show that the proposed method's recognition success rate is from 81.7% to 93.9% higher than that of the Vector Pair Matching (VPM) method.