The University of Western Australia
Monday, January 27, 2020
2:00PM – 3:00PM – HEC 101
Convolutional Neural Networks (CNN) use rectangular kernels to learn features from data that follow grid like structures such as images. However, 3D point clouds from LiDARs are unstructured. We propose a spherical kernel to directly learn from unstructured point clouds. Our kernel systematically quantizes the local 3D space to identify distinctive geometric relationships in the data. Similar to regular CNN kernels, the spherical kernel maintains translation-invariance and asymmetry properties, where the former guarantees weight sharing among similar local structures and the latter facilitates fine geometric learning. The proposed kernel is combined with octree guided neural networks as well as graph neural networks without edge-dependent filter generation, making it computationally attractive for large point clouds. Inspired from fuzzy clustering, we further extend our spherical kernel to use fuzzy bins to overcome boundary effects and variations in point sampling. We also extend the network architecture to a more efficient graph convolutional network that exploits ResNet like blocks and separable convolutions in the encoder and 1×1 convolutions in the decoder. This network can segment over one million points per second with competitive performance on benchmark datasets such as S3DIS, ModelNet, ShapeNet, ScanNet, and RueMonge2014.
For more info, please follow this link.