Skip to content

Latest commit

 

History

History
73 lines (59 loc) · 3.56 KB

README.md

File metadata and controls

73 lines (59 loc) · 3.56 KB
Raw

ShapeNet Part Segmentation Experiment

This repo implements the ShapeNet part segmentation experiment presented in the paper. The experiment involves six architectures:

  1. PointNet: models/pointnet.py. This is adopted from Pointnet_Pointnet2_pytorch for comparison with Architecture 2.
  2. PointNet_VecKM: models/pointnet_VecKM.py. We replace the PointNetEncoder in the original architecture with our VecKM module.
  3. Pointnet2 models/pointnet2.py. This is adopted from Pointnet_Pointnet2_pytorch without changing the architecture. It is compared with Architecture 4.
  4. PointNet2_VecKM: models/pointnet2_VecKM.py. We replace the first set abstraction layer before inputing the point cloud to PointNet++.
  5. PCT: Point Cloud Transformer models/PCT.py. This is a reproduced version based on the descriptions in Point Cloud Transformer and their official codes.
  6. PCT_VecKM: models/PCT_VecKM.py We replace the initial point embedding module in PCT with our VecKM.

The data augmentation and training strategies are borrowed from PCT_Pytorch. Many thanks to their great codes! The data augmentation and training strategies are borrowed from Pointnet_Pointnet2_pytorch. Many thanks to their great codes!

Data Preparation

Please download the data here and unzip the file into ./data/shapenetcore_partanno_segmentation_benchmark_v0_normal/. The file structure shall look like:

├── data
│   └── shapenetcore_partanno_segmentation_benchmark_v0_normal
│       ├── 02691156
│       ├── 02773838
│       ├── 02954340
│       ├── 02958343
│       ├── 03001627
│       ├── ...
├── models
│   ├── pointnet.py
│   ├── pointnet_VecKM.py
│   ├── PCT.py
│   ├── PCT_VecKM.py
│   ├── pointnet2.py
│   ├── pointnet2_VecKM.py
│   └── utils.py
├── provider.py
├── README.md
└── main.py
└── data.py

Requirements

python >= 3.9
pytorch >= 1.13
scipy

Models

We get the following accuracies by setting random seed as 0. Different GPUs will produce different results. My result is given by an RTXA5000 GPU.

Instance mIoU Avg. Class mIoU Inference Time (ms) (1 batch) # parameters
PointNet 83.1% 77.6% 15.1 8.34M
VecKM -- PN 84.9% 81.8% 40.8 1.29M
PointNet++ 85.0% 81.9% 130.8 1.41M
VecKM -- PN++ 85.3% 82.0% 65.9 1.50M
PCT 85.7% 82.6% 145.2 1.63M
VecKM -- PCT 85.6% 82.3% 46.6 1.71M

Training

python main.py --model pointnet
python main.py --model pointnet_VecKM
python main.py --model pointnet2
python main.py --model pointnet2_VecKM
python main.py --model PCT
python main.py --model PCT_VecKM