Effective Rotation-invariant Point CNN with Spherical Harmonics kernels
This addresses the need for robust shape analysis in computer vision, particularly for non-rigid objects, though it is incremental as it builds on existing point-based architectures.
The paper tackles the problem of achieving rotation invariance in point cloud neural networks by injecting spherical harmonics kernels into all layers, enabling invariance to both global and local rotations without data augmentation, and reports accurate results on classification and segmentation tasks.
We present a novel rotation invariant architecture operating directly on point cloud data. We demonstrate how rotation invariance can be injected into a recently proposed point-based PCNN architecture, at all layers of the network, achieving invariance to both global shape transformations, and to local rotations on the level of patches or parts, useful when dealing with non-rigid objects. We achieve this by employing a spherical harmonics based kernel at different layers of the network, which is guaranteed to be invariant to rigid motions. We also introduce a more efficient pooling operation for PCNN using space-partitioning data-structures. This results in a flexible, simple and efficient architecture that achieves accurate results on challenging shape analysis tasks including classification and segmentation, without requiring data-augmentation, typically employed by non-invariant approaches.