/GCNs-Point-Cloud-Classification

Implementing 3D point cloud classification using a multi-layer GCN network.

Primary LanguagePython

3D Point Cloud Classification with Multi-Layer GCN

Implementing 3D point cloud classification on ModelNet40 using a multi-layer GCN network.

Overview

framework

4 main modules：Residual Block, EdgeConv Operation, Dynamic Graph Construction, and Dilation Convolution.

exp

Dataset Preparation

Due to network connectivity issues, it is recommended to download the ModelNet40 dataset locally and then upload it to the server using SCP or similar tools.

Link: https://shapenet.cs.stanford.edu/media/modelnet40_ply_hdf5_2048.zip

Extract it to the ./data directory to create the data/modelnet40_ply_hdf5_2048 folder.

Environment Setup

My experimental environment includes:

  • NVIDIA GeForce RTX 3090 GPUs
  • CUDA Version: 12.1

You can set up the environment using the following commands:

conda create -n pointcloud python=3.9
conda activate pointcloud
pip install torch scikit-learn numpy h5py
# for DeepGCNs depend on your own version
pip install torch-cluster -f https://data.pyg.org/whl/torch-2.3.0+cu121.html

For experiment visualization, I use wandb for real-time monitoring in main.py. Additionally, I use tensorboard for visualization in main_tensorboard.py.

pip install tensorboard  # TensorBoard
pip install wandb  # Weights and Biases

Command Line Arguments

This section provides detailed explanations for the command line arguments used in the script. Each argument has a specific role and can be adjusted to control various aspects of the model training and evaluation process.

Argument Name Type Description
--exp_name str Name of the experiment.
--model str Model to use. Options: pointnet, dgcnn, DeepGCN, GCN, EdgeGCN.
--dataset str Dataset to use. Options: modelnet40.
--batch_size int Size of the training batch.
--epochs int Number of training epochs.
--lr float Learning rate for the optimizer.
--eval bool Evaluate the model if set to True.
--model_path str Pretrained model path.
--k int Number of nearest neighbors to use.
--num_points int Number of points to use from each point cloud.
--n_blocks int Number of basic blocks in the backbone.
--n_filters int Number of channels of deep features.
--dynamic bool Use dynamic adjacency matrix if set to True.
--dilated bool Use dilated k-nearest neighbors if set to True.

The codes for models PointNet, DGCNN, and DeepGCNs are from the respective repositories. The GCN model is based on the GCN repository. EdgeGCN is my implementation.

Example Usage

Training

python main.py

Testing

python main.py --eval True

File Description

EdgeGCN/
├── imgs/ 					# images shown on README
├── data/ 					# dataset
├── deep_gcns_torch/		# implement of DeepGCNs
├── pretrained/				# pretrained model
├── data.py					# dataloader
├── util.py					# IO toolkit
├── model.py				# model PointNet&DGCNN
├── DeepGCN.py				# model DeepGCN
├── GCN.py					# model GCN
├── EdgeGCN.py				# model EdgeGCN
├── main.py					# main with wandb
├── main_tensorboard.py		# main with tensorboard
└── README.md				# README

License

This project is licensed under the MIT License.

Acknowledgement

The structure of this codebase is borrowed from DGCNN.