Going Deeper with Lean Point Networks

TL;DR

This paper introduces Lean Point Networks, a set of novel modules that enable training deeper, more accurate point processing networks with improved memory efficiency and inference speed, demonstrated across various segmentation benchmarks.

Contribution

The paper presents three new point processing blocks that enhance memory efficiency, enable deeper architectures, and improve accuracy in point cloud segmentation tasks.

Findings

Achieved systematic accuracy improvements on multiple segmentation benchmarks.

Reduced memory consumption compared to existing architectures.

Enabled deeper networks with faster inference times.

Abstract

In this work we introduce Lean Point Networks (LPNs) to train deeper and more accurate point processing networks by relying on three novel point processing blocks that improve memory consumption, inference time, and accuracy: a convolution-type block for point sets that blends neighborhood information in a memory-efficient manner; a crosslink block that efficiently shares information across low- and high-resolution processing branches; and a multiresolution point cloud processing block for faster diffusion of information. By combining these blocks, we design wider and deeper point-based architectures. We report systematic accuracy and memory consumption improvements on multiple publicly available segmentation tasks by using our generic modules as drop-in replacements for the blocks of multiple architectures (PointNet++, DGCNN, SpiderNet, PointCNN).