Module: tfg.nn.layer.pointnet

Implementation of the PointNet networks.

@inproceedings{qi2017pointnet, title={Pointnet: Deep learning on point sets for3d classification and segmentation}, author={Qi, Charles R and Su, Hao and Mo, Kaichun and Guibas, Leonidas J}, booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition}, pages={652--660}, year={2017} }

Note: scheduling of batchnorm momentum currently not available in keras. However experimentally, using the batch norm from Keras resulted in better test accuracy (+1.5%) than the author's custom batch norm version even when coupled with batchnorm momentum decay. Further, note the author's version is actually performing a "global normalization", as mentioned in the tf.nn.moments documentation.

This shorthand notation is used throughout this module: B: Number of elements in a batch. N: The number of points in the point set. D: Number of dimensions (e.g. 2 for 2D, 3 for 3D). C: The number of feature channels.

Classes

class ClassificationHead: The PointNet classification head.

class PointNetConv2Layer: The 2D convolution layer used by the feature encoder in PointNet.

class PointNetDenseLayer: The fully connected layer used by the classification head in pointnet.

class PointNetVanillaClassifier: The PointNet 'Vanilla' classifier (i.e. without spatial transformer).

class VanillaEncoder: The Vanilla PointNet feature encoder.