Discriminative feature representation is significant for boosting the performance of computer vision tasks covering different levels. Traditional low-level feature representation exhibits good generalization and robustness while lacks of enough discriminant ability. In this paper, we focus on 3D local shape features, proposing a discriminative feature selection method, which is also closely related with mid-level 3D shape representation. We firstly design a histogram-signature hybrid 3D local shape descriptor using 3D geometrical information from the 3D point cloud of a tested object. Then, we propose a discrimination power metric to automatically select a collection of discriminative local shapes from a candidate set, resulting in a mid-level shape feature representation. The proposed algorithm is applied in the task of multi-view 2.5D scan registration. The performance was verified on public and popular instance-level 3D object datasets. Both qualitative and quantitative results demonstrate the effectiveness and robustness of the proposed algorithm on different 3D objects. Compared with low-level 3D object representation, the discriminative feature selection for 3D shape feature representation allows for superior performance with higher precision and recall rate.
|