dmlc--dgl
esang
4fb50be4d9
* some modifications for pointnet2 * temporarily save changes * move files to new directory point_transformer * implement point transformer for classification * restore train_cls in pointnet * implement point transformer for partseg * fix point transformer for nan loss * modify point transformer for cls * modify training setting * update transformer for cls * update code * update code for latest performance * update the example index * some minor changes Co-authored-by: Tong He <hetong007@gmail.com>
PointNet and PointNet++ for Point Cloud Classification and Segmentation
This is a reproduction of the papers
- PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation.
- PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space.
Performance
Classification
| Model | Dataset | Metric | Score - PyTorch | Score - DGL | Time(s) - PyTorch | Time(s) - DGL |
|---|---|---|---|---|---|---|
| PointNet | ModelNet40 | Accuracy | 89.2(Official) | 89.3 | 181.8 | 95.0 |
| PointNet++(SSG) | ModelNet40 | Accuracy | 92.4 | 93.3 | 182.6 | 133.7 |
| PointNet++(MSG) | ModelNet40 | Accuracy | 92.8 | 93.3 | 383.6 | 240.5 |
Part Segmentation
| Model | Dataset | Metric | Score - PyTorch | Score - DGL | Time(s) - PyTorch | Time(s) - DGL |
|---|---|---|---|---|---|---|
| PointNet | ShapeNet | mIoU | 84.3 | 83.6 | 251.6 | 234.0 |
| PointNet++(SSG) | ShapeNet | mIoU | 84.9 | 84.5 | 361.7 | 240.1 |
| PointNet++(MSG) | ShapeNet | mIoU | 85.4 | 84.6 | 817.3 | 821.8 |
- Score - PyTorch are collected from this repo.
- Time(s) are the average training time per epoch, measured on EC2 g4dn.4xlarge instance w/ Tesla T4 GPU.
How to Run
For point cloud classification, run with
python train_cls.py
For point cloud part-segmentation, run with
python train_partseg.py
To Visualize Part Segmentation in Tensorboard
First pip install tensorboard
then run
python train_partseg.py --tensorboard
To display in Tensorboard, run
tensorboard --logdir=runs