3D machine learning has received growing attention from the industries due to the potential to offer significant time gain over the classical numerical approaches. We introduce a sampling approach to construct PointNet using point clouds, with orders of magnitude better efficiency compared to the conventional iterative gradient descent, at the expense of a small performance loss. This sampling is an extension of Sampling Where It Matters (SWIM). We propose two approaches to preserve the geometric properties in point cloud, namely KDTree and Recursive Sampling. For the latter, we leverage quantile and the length-squared distribution of point coordinates to enhance the sampled representations. The research also tackles space complexity by implementing batch-wise weight and bias updates. Our approach is proven effective and efficient under large-scale settings, specifically, sampled PointNet preserves more than 90% of PointNet’s accuracy on classifying the standard 3D benchmark ModelNet40 at the costs of less than 10% of PointNet’s GPU training time on a CPU, assuming that the point clouds share a standard orientation.     «

3D machine learning has received growing attention from the industries due to the potential to offer significant time gain over the classical numerical approaches. We introduce a sampling approach to construct PointNet using point clouds, with orders of magnitude better efficiency compared to the conventional iterative gradient descent, at the expense of a small performance loss. This sampling is an extension of Sampling Where It Matters (SWIM). We propose two approaches to preserve the geometri...     »