Learning task-specific features for 3D pointcloud graph creation

TL;DR

This paper introduces a learnable, task-specific graph construction method for 3D pointclouds using a neural network transformation and regularization, improving graph quality and classification accuracy.

Contribution

It proposes a novel graph creation approach based on learned transformations and stress-based regularization, outperforming traditional k-NN methods in 3D pointcloud classification.

Findings

Outperformed baseline k-NN graphs by 0.3 accuracy on ModelNet40.

Introduced a learnable graph construction method optimized via backpropagation.

Demonstrated improved graph quality and classification performance.

Abstract

Processing 3D pointclouds with Deep Learning methods is not an easy task. A common choice is to do so with Graph Neural Networks, but this framework involves the creation of edges between points, which are explicitly not related between them. Historically, naive and handcrafted methods like k Nearest Neighbors (k-NN) or query ball point over xyz features have been proposed, focusing more attention on improving the network than improving the graph. In this work, we propose a more principled way of creating a graph from a 3D pointcloud. Our method is based on performing k-NN over a transformation of the input 3D pointcloud. This transformation is done by an Multi-Later Perceptron (MLP) with learnable parameters that is optimized through backpropagation jointly with the rest of the network. We also introduce a regularization method based on stress minimization, which allows to control how distant is the learnt graph from our baseline: k-NN over xyz space. This framework is tested on ModelNet40, where graphs generated by our network outperformed the baseline by 0.3 points in overall accuracy.