计图点云库

已经实现的模型

Model	Classification	Segmentation
PointNet	√	√
PointNet ++	√	√
PointCNN	√	√
DGCNN	√	√
PointConv	√	√

使用方法

安装依赖

sudo apt install python3.7-dev libomp-dev
sudo python3.7 -m pip install git+https://github.com/Jittor/jittor.git
python3.7 -m pip install sklearn lmdb msgpack_numpy

安装点云库

git clone https://github.com/Jittor/PointCloudLib.git # 将库下载的本地
# 您需要将 ModelNet40 和 ShapeNet 数据集下载到 data_util/data/ 里面
ModelNet40 数据集链接 ： https://shapenet.cs.stanford.edu/media/modelnet40_normal_resampled.zip 
ShapeNet 数据集链接 ： https://shapenet.cs.stanford.edu/media/shapenet_part_seg_hdf5_data.zip 

sh train_cls.sh # 点云分类的训练和测试 
sh train_seg.sh # 点云分割的训练和测试 

所依赖的库

Python 3.7
Jittor 
Numpy
sklearn
lmdb
msgpack_numpy
...

实验结果

分类训练效果测试

Model	Input	overall accuracy
PointNet	1024 xyz	87.2
PointNet ++	4096 xyz + normal	92.3
PointCNN	1024 xyz	92.6
DGCNN	1024 xyz	92.9
PointConv	1024 xyz + normal	92.4

分类训练时间测试

Model	Speed up ratio (Compare with Pytorch)
PointNet	1.22
PointNet ++	2.72
PointCNN	2.41
DGCNN	1.22
PointConv

分割训练效果测试

Model	Input	pIoU
PointNet	2048 xyz + cls label	83.5
PointNet ++	2048 xyz + cls label + normal	85.0
PointCNN	2048 xyz + normal	86.0
DGCNN	2048 xyz + cls label	85.1
PointConv	2048 xyz	85.4

分割训练时间测试

Model	Speed up ratio (Compare with Pytorch)
PointNet	1.06
PointNet ++	1.85
PointCNN	None (No pytorch implementation)
DGCNN	1.05
PointConv	None (No pytorch implementation)

目录结构

.
├── data_utils                   # 数据相关工具
│   ├── data                     # 数据存放路径
│   ├── modelnet40_loader.py
│   └── shapenet_loader.py
├── misc
│   ├── layers.py
│   ├── ops.py
│   ├── pointconv_utils.py
│   └── utils.py
├── networks
│   ├── cls
│   │   ├── dgcnn.py
│   │   ├── pointcnn.py
│   │   ├── pointconv.py
│   │   ├── pointnet2.py
│   │   └── pointnet.py
│   └── seg
│       ├── dgcnn_partseg.py
│       ├── pointcnn_partseg.py
│       ├── pointconv_partseg.py
│       ├── pointnet2_partseg.py
│       └── pointnet_partseg.py

├── README.md
├── run_cls.sh
├── run_partseg.sh
├── train_cls.py
└── train_partseg.py

非常欢迎您使用计图的点云库进行相关的研究，如在使用中有问题，欢迎提交 issus。