# Efficient Learning on Point Clouds with Basis Point Sets

**Authors:** Sergey Prokudin, Christoph Lassner, Javier Romero  
**arXiv:** [1908.09186](https://arxiv.org/abs/1908.09186)  
**Date:** 2019-08-24  
**Summary page:** https://www.summarizepaper.com/en/paper/1908.09186v1/

## Key points

- Authors Sergey Prokudin, Christoph Lassner, and Javier Romero address the challenge of handling point clouds in computer vision applications
- Point clouds are a valuable representation of everyday scenes but their unordered structure poses difficulties for machine learning algorithms
- Occupancy grid mapping is a common approach but can lead to loss of details and increased data storage
- The authors propose basis point sets (BPS) as an efficient and versatile method for processing point clouds with machine learning algorithms
- BPS representation is a residual representation that can be computed efficiently and integrated with standard neural network architectures
- Using BPS as input for a simple fully connected network shows comparable performance to PointNet on shape classification while reducing floating-point operations significantly
- BPS representation can be applied to register high-resolution meshes to noisy 3D scans, enabling single-pass high-resolution mesh registration without time-consuming per-scan optimization for real-time execution

## Summary

In their study titled "Efficient Learning on Point Clouds with Basis Point Sets," authors Sergey Prokudin, Christoph Lassner, and Javier Romero address the challenge of handling point clouds in computer vision applications. The increasing availability of 3D scanning technology has made point clouds a valuable representation of everyday scenes. However, their unordered structure poses difficulties for machine learning algorithms. One common approach to this issue is occupancy grid mapping, which can lead to loss of details through discretization and increased data storage. To overcome these challenges, the authors propose basis point sets (BPS) as an efficient and versatile method for processing point clouds with machine learning algorithms. The BPS representation is a residual representation that can be computed efficiently and integrated with standard neural network architectures and other machine learning algorithms. By using BPS as input for a simple fully connected network, the authors demonstrate comparable performance to PointNet on a shape classification task while significantly reducing floating-point operations. Additionally, the proposed BPS representation can be applied to register high-resolution meshes to noisy 3D scans. This novel approach enables single-pass high-resolution mesh registration without time-consuming per-scan optimization, allowing for real-time execution. Overall, this work introduces an innovative solution for efficiently processing point clouds in machine learning tasks and offers promising implications for various computer vision applications.