Learning Rotation-Invariant Representations of Point Clouds Using Aligned Edge Convolutional Neural Networks

TL;DR

This paper introduces AECNN, a neural network that learns rotation-invariant features of point clouds by aligning local features with reference frames, improving robustness without data augmentation.

Contribution

The paper presents a novel rotation-invariant neural network for point cloud analysis that eliminates the need for data augmentation by aligning features with local reference frames.

Findings

Outperforms state-of-the-art methods in rotation robustness

Achieves better accuracy without data augmentation

Effective in classification and segmentation tasks

Abstract

Point cloud analysis is an area of increasing interest due to the development of 3D sensors that are able to rapidly measure the depth of scenes accurately. Unfortunately, applying deep learning techniques to perform point cloud analysis is non-trivial due to the inability of these methods to generalize to unseen rotations. To address this limitation, one usually has to augment the training data, which can lead to extra computation and require larger model complexity. This paper proposes a new neural network called the Aligned Edge Convolutional Neural Network (AECNN) that learns a feature representation of point clouds relative to Local Reference Frames (LRFs) to ensure invariance to rotation. In particular, features are learned locally and aligned with respect to the LRF of an automatically computed reference point. The proposed approach is evaluated on point cloud classification and part segmentation tasks. This paper illustrates that the proposed technique outperforms a variety of state of the art approaches (even those trained on augmented datasets) in terms of robustness to rotation without requiring any additional data augmentation.