PMHA-Net: Positional Multi-Head Attention Network for Point-Cloud Part Segmentation and Classification

Abstract

For understanding unordered sets of point clouds, the positional information of each point must be effectively used. To this end, the existing models use the absolute position of the point and its relative position within a local group. However, accurately capturing the positional information is challenging because the relative position within a local group typically has a considerably smaller value than that of the feature information. Moreover, in terms of the data characteristics of point clouds, closer points are more strongly correlated, but their relative position approaches zero. To address these problems, we process the relative position within a local group by normalizing it within the overall object range and local range according to the data characteristics. This transformation helps maintain the meaning and pattern of the relative position while facilitating its learning. The transformed data are combined with the absolute position to encode the position vector, which serves as the positional encoding in multi-head attention across multiple resolutions. Extensive experiments are conducted on benchmark point cloud datasets to demonstrate that the proposed model exhibits competitive performance in part segmentation and classification tasks.

Keywords

• Point cloud
• deep learning
• multi-head attention
• positional encoding
• up-sampling
• part segmentation