PointNet and PointNet++ are typical deep learning architectures that directly take point clouds as input and output class of labels for the entire input. We introduce the constructions of PointNet and PointNet++ and discuss the mathematical reason behind them. In addition, we present a recent method using quaternions to modify these networks so that the new network obtained can achieve rotation equivariant properties. Open problems regarding the analysis of the new network will be given.
Một số tài liệu tham khảo:
1. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space
stanford.edu/~r...
2. Quaternion Product Units for Deep Learning on 3D Rotation Groups
arXiv preprint arXiv:1912.07791, 2019
Негізгі бет Thieu Vo: A deep network with point cloud dataset: a mathematical point of view
No video
Пікірлер: 3