Clustering-TinyPointNet for fast large-scale point cloud semantic segmentation

Efficient processing of massive point cloud datasets is crucial for achieving fast semantic segmentation in various applications. While PointNet++ has demonstrated excellent performance in point cloud segmentation, its processing speed may not meet the requirements of real-time applications. This paper investigates the PointNet++ approach and proposes a novel deep learning architecture, termed Clustering-TinyPointNet, which aims to perform coarse-to-fine point cloud data segmentation using a combination of a fast K-means++ clustering algorithm and a lightweight neural network. The Clustering-TinyPointNet method is derived from PointNet++ by reducing the total number of layers from 79 to 55 and the number of learnable parameters from 892.8K to 638.2K. As a result, the proposed method achieves a reduction in running time by over 30% while maintaining comparable segmentation accuracy. This advantage positions Clustering-TinyPointNet as a promising solution for efficient point cloud semantic segmentation in real-time applications.