Pytorch Implementation of Various Point Transformers
Recently, various methods applied transformers to point clouds: PCT: Point Cloud Transformer (Meng-Hao Guo et al.), Point Transformer (Nico Engel et al.), Point Transformer (Hengshuang Zhao et al.). This repo is a pytorch implementation for these methods and aims to compare them under a fair setting. Currently, all three methods are implemented, while tuning their hyperparameters.
Classification
Environment
conda create -n pointtransformer python=3.7
conda activate pointtransformer
pip install omegaconf tqdm -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install hydra-core --upgrade --pre
conda install pytorch=1.2.0 torchvision cudatoolkit=10.0 -c pytorch
Data Preparation
Download alignment ModelNet here and save in modelnet40_normal_resampled.
Run
Change which method to use in config/config.yaml and run
python train.py
Results
Using Adam with learning rate decay 0.3 for every 50 epochs, train for 200 epochs; data augmentation follows this repo. For Hengshuang and Nico, initial LR is 1e-3 (maybe fine-tuned later); for Menghao, initial LR is 1e-4, as suggested by the author. ModelNet40 classification results (instance average) are listed below:
| Model | Accuracy |
|---|---|
| Hengshuang | 89.6 |
| Menghao | 92.6 |
| Nico | 85.5 |
Miscellaneous
Some code and training settings are borrowed from https://github.com/yanx27/Pointnet_Pointnet2_pytorch. Code for PCT: Point Cloud Transformer (Meng-Hao Guo et al.) is adapted from the author's Jittor implementation https://github.com/MenghaoGuo/PCT.