PointNet++ RandAugment

This repo is implementation for PointNet and PointNet++ with Novel data augmentation methods and RandAugment for point cloud data to handle with 3D classification task
Baseline code for PointNet2 mostly borrowed from erikwijmans/Pointnet2_PyTorch

This repo is 3D pointcloud version of RandAugment: Practical automated data augmentation with a reduced search space.

• Supports Multi-GPU via nn.DataParallel <https://pytorch.org/docs/stable/nn.html#torch.nn.DataParallel>_.

• Supports PyTorch version >= 1.0.0. Use v1.0 <https://github.com/erikwijmans/Pointnet2_PyTorch/releases/tag/v1.0>_ for support of older versions of PyTorch.

Setup

---

• Install python -- This repo is tested with 2.7, 3.5, and 3.6

• Install dependencies

  pip install -r requirements.txt

• Install with: pip install -e .
  

If you rush into error message RuntimeError: Ninja is required to load C++ extension, please refer to zhanghang1989/PyTorch-Encoding#167 for error shooting.

Training

---

Two training examples are provided by pointnet2/train/train_sem_seg.py and pointnet2/train/train_cls.py. The datasets for both will be downloaded automatically by default.

They can be run via

  python pointnet2/train.py task=cls

  # Or with model=msg for multi-scale grouping

  python pointnet2/train.py task=cls model=msg

Both scripts will print training progress after every epoch to the command line. Use the --visdom flag to enable logging to visdom and more detailed logging of training progress.

Visualization

• Multiple augmentations were deployed.(Check below image that shows some examples of augmentation methods)
  python ./data/pointnet2/ModelNet40Loader

Classification Result

classfication	acc
PointNet++(Official, w/o normal)	90.7
PointNet++(Official, with normal)	91.9
Ours	92.8

Segmentation Result

Testing

Citation

  @article{pytorchpointnet++,
        Author = {Seungjun Lee},
        Title = {Pointnet++ RandAugment},
        Journal = {https://github.com/seungjunlee96/PointNet2_RandAugment},
        Year = {2020}
  }

  @inproceedings{qi2017pointnet++,
      title={Pointnet++: Deep hierarchical feature learning on point sets in a metric space},
      author={Qi, Charles Ruizhongtai and Yi, Li and Su, Hao and Guibas, Leonidas J},
      booktitle={Advances in Neural Information Processing Systems},
      pages={5099--5108},
      year={2017}
  }

Brief summary of RandAugment

The primary goal of RandAugment is to remove the need for a separate search phase on a proxy task.
There are only two parameters to tune the RandAugment

• N, Number of augmentation transformation to apply sequentially.
• M, Magnitude for all the transformations Below code shows applying RandAugment.(code from original paper of RandAugment)

transforms = [
’Identity’, ’AutoContrast’, ’Equalize’,
’Rotate’, ’Solarize’, ’Color’, ’Posterize’,
’Contrast’, ’Brightness’, ’Sharpness’,
’ShearX’, ’ShearY’, ’TranslateX’, ’TranslateY’]
def randaugment(N, M):
"""Generate a set of distortions.
Args:
N: Number of augmentation transformations to
apply sequentially.
M: Magnitude for all the transformations.
"""
sampled_ops = np.random.choice(transforms, N)
return [(op, M) for op in sampled_ops]

To Do

• Training and evaluation on Segmentation
• SSG vs MSG
• More Data augmentation methods

Reference By

yanx27/Pointnet_Pointnet2_pytorch
erikwijmans/Pointnet2_PyTorch
ildoonet's RandAugment
halimacc/pointnet3
fxia22/pointnet.pytorch
charlesq34/PointNet
charlesq34/PointNet++