Skip to content

Latest commit

 

History

History

obj_cls

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
config
config
 
 
cuda_lib
cuda_lib
 
 
model
model
 
 
util
util
 
 
README.md
README.md
 
 
eval_voting.py
eval_voting.py
 
 
main.py
main.py
 
 
main_ddp.py
main_ddp.py
 
 

README.md

3D Object Classification

Installation

Requirements

  • Hardware: GPU to hold 6000M. (Better with two gpus or higher-level gpu to satisfy the need of paralleled cuda_kernels.)
  • Software: Linux (tested on Ubuntu 18.04) PyTorch>=1.5.0, Python>=3, CUDA>=10.1, tensorboardX, h5py, pyYaml, scikit-learn

Dataset

Download and unzip ModelNet40 (415M). Then symlink the paths to it as follows (you can alternatively modify the path here):

mkdir -p data
ln -s /path to modelnet40/modelnet40_ply_hdf5_2048 data

Usage

  • Build the CUDA kernel:

    When you run the program for the first time, please wait a few moments for compiling the cuda_lib automatically. Once the CUDA kernel is built, the program will skip this in the future running.

  • Train:

    • Multi-thread training (nn.DataParallel) :

      • python main.py --config config/dgcnn_paconv_train.yaml (Embed PAConv into DGCNN)

      • python main.py --config config/pointnet_paconv_train.yaml (Embed PAConv into PointNet)

    • We also provide a fast multi-process training (nn.parallel.DistributedDataParallel, recommended) with official nn.SyncBatchNorm. Please also remind to specify the GPU ID:

      • CUDA_VISIBLE_DEVICES=x,x python main_ddp.py --config config/dgcnn_paconv_train.yaml (Embed PAConv into DGCNN)
      • CUDA_VISIBLE_DEVICES=x,x python main_ddp.py --config config/pointnet_paconv_train.yaml (Embed PAConv into PointNet)

  • Test:

    • Download our pretrained model and put it under the obj_cls folder.

    • Run the voting evaluation script to test our pretrained model, after this voting you will get an accuracy of 93.9% if all things go right:

      python eval_voting.py --config config/dgcnn_paconv_test.yaml

    • You can also directly test our pretrained model without voting to get an accuracy of 93.6%:

      python main.py --config config/dgcnn_paconv_test.yaml

    • For full test after training the model:

      • Specify the eval to True in your config file.

      • Make sure to use main.py (main_ddp.py may lead to wrong result due to the repeating problem of all_reduce function in multi-process training) :

        python main.py --config config/your config file.yaml

  • Visualization: tensorboardX incorporated for better visualization.

    tensorboard --logdir=checkpoints/exp_name

Citation

If you find the code or trained models useful, please consider citing:

@inproceedings{xu2021paconv,
  title={PAConv: Position Adaptive Convolution with Dynamic Kernel Assembling on Point Clouds},
  author={Xu, Mutian and Ding, Runyu and Zhao, Hengshuang and Qi, Xiaojuan},
  booktitle={CVPR},
  year={2021}
}

Contact

You are welcome to send pull requests or share some ideas with us. Contact information: Mutian Xu ([email protected]) or Runyu Ding ([email protected]).

Acknowledgement

This code is is partially borrowed from DGCNN.