Reimplementation of VoteNet (Qi et al. ICCV 19') architecture.
3D object detection model for point cloud datatype.
Including implementation of backbone architectures: PointNet (Qi et al. CVPR 17'), PointNet++ (Qi et al. NIPS 17')
### python dataset/dataset.py
# OUTPUT:
Train dataset size: 9843
Test dataset size: 2468
airplane/test/airplane_0671.off
Class: airplane
Sampled points: 1024
Data samples:
tensor([[-0.4994, 0.0878, 0.0586],
[-0.1425, -0.2518, -0.0613],
[ 0.7644, -0.0099, 0.2678],
[ 0.1001, 0.1798, -0.1113],
[-0.4239, 0.0971, 0.0461]])
### python votenet/pointnet.py
# OUTPUT:
Test batch size: 10
Input transform matrix shape: torch.Size([10, 3, 3])
Matrix sample:
tensor([[1., 0., 0.],
[0., 1., 0.],
[0., 0., 1.]])
Feature transform matrix shape: torch.Size([10, 64, 64])
Global feature shape: torch.Size([10, 1024])
Class score shape: torch.Size([10, 40])