General_3D

A project to implement some state-of-the-art 3D network architecture in PyTorch. My research topics is Geometric learning, so I need to use some state-of-the art 3D network architecture in my work. However, most of the code provided by the paper authors is implemented by TensorFlow. I am more familiar with PyTorch, so I want to reproduce the paper results in PyTorch.

Existing network in this repo.

• Have been finished and tested

  • PointNet++ : 90.76% classification accuracy on ModelNet40.
  • DGCNN : 91.4% classification accuracy on ModelNet40(still not the result in the original paper). 85.1% mIoUs on ShapeNetpart.

• Have not finished.

  • PointCNN : I can only get 87% accuracy on ModelNet40 and I cannot find out what's wrong in my code. I really need some help.
  • SpiderNet : notfinished.

Dataset

• ModelNet40
• ShapeNetPart with normals

After download the dataset, you have to make a directory dataset in the root directory, and move the above dataset to that directory.

compile cuda kernel

To run the pointnet++ network, you have to first compile the Farthest sampling and query ball point module.

cd c_lib
sh make.sh

If you account for some problems in this stages, raise a issue.

Dependency

• PyTorch (above 4.0)
• python 3.6
• visdom visdom can visulize the training process and it's easy to use.

Training and evalution

Pointnet++ can only run in GPU and others should be able to run without GPU. It you run in a pulbic server, make sure to add CUDA_VISIBLE_DEVICES=1,2...

• train pointnet++

CUDA_VISIBLE_DEVICES=1 python train_supervised_cls.py --model-name 'pointnet' --visdom-name 'pointnet_cls'

for other network, you only have to replace the --model-name which can be pointnet, dgcnn, pointcnn, spidercnn.

Need some helps

I creat this repo to share my code and ask for some help. I can not reproduce PointCNN and cannot find out what's wrong in my code. So, if you are interested in this work, help me to test the PointCNN architecture.

Feel free to raise a issue to ask questions