Learning Rotation-Invariant Representations of Point Clouds Using Aligned Edge Convolutional Neural Networks
Junming Zhang · Ming-Yuan Yu · Ram Vasudevan · Matthew Johnson-Roberson
TL;DR
A point cloud network that learns features in local reference frames stays robust to rotation without needing rotation-based data augmentation.
BRIEF
Deep networks for 3D point clouds usually fail when an object appears in an orientation not seen during training. The standard fix is to augment the training data with rotations, which adds computation and often needs larger models. The Aligned Edge Convolutional Neural Network instead learns local features relative to Local Reference Frames, aligned to an automatically chosen reference point, so the representation does not depend on rotation. On classification and part segmentation, it was more robust to rotation than several state-of-the-art methods, including ones trained with augmentation.
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.