PointNet in PyTorch

This is a PyTorch re-implementation of PointNet according to the specifications laid out in the paper with two minor differences:

• I exclude the adaptive batch normalization decay rate
• The trained model provided operates on pointclouds with 2000 points as opposed to 2048 (although you can re-train and change the pointcloud sizes)

Other Implementations

• The official TensorFlow implementation from the authors can be found here.
• Another PyTorch re-implementation can be found here.

If you use my re-implementation for your own work, please cite the original paper:

Qi, Charles R., et al. "Pointnet: Deep learning on point sets for 3d classification and segmentation." 
Proc. Computer Vision and Pattern Recognition (CVPR), IEEE 1.2 (2017): 4.

Repo TO-DO's

• Finish segmentation implementation
• Upload the sampled ModelNet40 data
• Write up how-to section

Classification Results

The pre-trained classifier model included in this repository was trained for 60 epochs with a batch size of 32 on a 2000-point-per-model sampling of ModelNet40.

Here is an graph showing the training loss over 60 epochs:

Below are the accuracy results for the included classifier model on the test set