Going Deeper with Lean Point Networks

Eric-Tuan Le, Iasonas Kokkinos, Niloy J. Mitra

TL;DR: We introduce three generic point cloud processing blocks that improve both accuracy and memory consumption of multiple state-of-the-art networks, thus allowing to design deeper and more accurate networks.

Abstract:

We introduce three generic point cloud processing blocks that improve both accuracy and memory consumption of multiple state-of-the-art networks, thus allowing to design deeper and more accurate networks.

The novel processing blocks that facilitate efficient information flow are a convolution-type operation block for point sets that blends neighborhood information in a memory-efficient manner; a multi-resolution point cloud processing block; and a crosslink block that efficiently shares information across low- and high-resolution processing branches. Combining these blocks, we design significantly wider and deeper architectures.

We extensively evaluate the proposed architectures on multiple point segmentation benchmarks (ShapeNetPart, ScanNet, PartNet) and report systematic improvements in terms of both accuracy and memory consumption by using our generic modules in conjunction with multiple recent architectures (PointNet++, DGCNN, SpiderCNN, PointCNN). We report a 9.7% increase in IoU on the PartNet dataset, which is the most complex, while decreasing memory footprint by 57%.

Keywords: point cloud processing, point convolutions, memory-efficient training, deep neural network design