Object recognition based on point cloud geometry construction and embeddable attention.

Abstract

A point cloud is a collection of disordered and discrete points with irregularity, and it lacks of topological structure. The number of discrete points in the point cloud is huge, and how to capture the key features from the large amount of points is crucial to improve the accuracy of model recognition. In this paper, based on point cloud geometry construction and embeddable attention, a 3D object recognition algorithm is proposed. By constructing triangular geometries between points, topological structure information to the point cloud is stored for points’ geometric construction module. The embeddable attention module uses an improved attention mechanism with feature bias and nonlinear mapping to enable focused attention to capture key features. In addition, a combination of max and average pooling to aggregate global feature has been applied to avoid situations when using only one method would ignore other key information. In comparison with other state-of-the-art methods using ModelNet40 and ScanObjectNN, the proposed method shows significant improvements in identifying both mAcc and OA. The experiments also demonstrate the effectiveness of the modules in this algorithm.