Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/Pang-Yatian/Point-MAE
[ECCV2022] Masked Autoencoders for Point Cloud Self-supervised Learning
https://github.com/Pang-Yatian/Point-MAE
point-cloud self-supervised-learning
Last synced: 8 days ago
JSON representation
[ECCV2022] Masked Autoencoders for Point Cloud Self-supervised Learning
- Host: GitHub
- URL: https://github.com/Pang-Yatian/Point-MAE
- Owner: Pang-Yatian
- License: mit
- Created: 2022-03-09T06:26:31.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2023-06-26T01:51:27.000Z (about 1 year ago)
- Last Synced: 2024-03-09T21:35:42.455Z (4 months ago)
- Topics: point-cloud, self-supervised-learning
- Language: Python
- Homepage:
- Size: 1.34 MB
- Stars: 369
- Watchers: 6
- Forks: 50
- Open Issues: 3
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Lists
- awesome-3D-self-supervised-learning - [Code
- awesome-stars - Pang-Yatian/Point-MAE - [ECCV2022] Masked Autoencoders for Point Cloud Self-supervised Learning (Python)
README
# Point-MAE
## Masked Autoencoders for Point Cloud Self-supervised Learning, [ECCV 2022](https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136620591.pdf), [ArXiv](https://arxiv.org/abs/2203.06604)
[](https://paperswithcode.com/sota/3d-point-cloud-classification-on-scanobjectnn?p=masked-autoencoders-for-point-cloud-self)
[](https://paperswithcode.com/sota/3d-point-cloud-classification-on-modelnet40?p=masked-autoencoders-for-point-cloud-self)
In this work, we present a novel scheme of masked autoencoders for point cloud self-supervised learning, termed as Point-MAE. Our Point-MAE is neat and efficient, with minimal modifications based on the properties of the point cloud. In classification tasks, Point-MAE outperforms all the other self-supervised learning methods on ScanObjectNN and ModelNet40. Point-MAE also advances state-of-the-art accuracies by 1.5%-2.3% in the few-shot learning on ModelNet40.
## 1. Requirements
PyTorch >= 1.7.0 < 1.11.0;
python >= 3.7;
CUDA >= 9.0;
GCC >= 4.9;
torchvision;
```
pip install -r requirements.txt
```
```
# Chamfer Distance & emd
cd ./extensions/chamfer_dist
python setup.py install --user
cd ./extensions/emd
python setup.py install --user
# PointNet++
pip install "git+https://github.com/erikwijmans/Pointnet2_PyTorch.git#egg=pointnet2_ops&subdirectory=pointnet2_ops_lib"
# GPU kNN
pip install --upgrade https://github.com/unlimblue/KNN_CUDA/releases/download/0.2/KNN_CUDA-0.2-py3-none-any.whl
```
## 2. Datasets
We use ShapeNet, ScanObjectNN, ModelNet40 and ShapeNetPart in this work. See [DATASET.md](./DATASET.md) for details.
## 3. Point-MAE Models
| Task | Dataset | Config | Acc.| Download|
| ----- | ----- |-----| -----| -----|
| Pre-training | ShapeNet |[pretrain.yaml](./cfgs/pretrain.yaml)| N.A. | [here](https://github.com/Pang-Yatian/Point-MAE/releases/download/main/pretrain.pth) |
| Classification | ScanObjectNN |[finetune_scan_hardest.yaml](./cfgs/finetune_scan_hardest.yaml)| 85.18%| [here](https://github.com/Pang-Yatian/Point-MAE/releases/download/main/scan_hardest.pth) |
| Classification | ScanObjectNN |[finetune_scan_objbg.yaml](./cfgs/finetune_scan_objbg.yaml)|90.02% | [here](https://github.com/Pang-Yatian/Point-MAE/releases/download/main/scan_objbg.pth) |
| Classification | ScanObjectNN |[finetune_scan_objonly.yaml](./cfgs/finetune_scan_objonly.yaml)| 88.29%| [here](https://github.com/Pang-Yatian/Point-MAE/releases/download/main/scan_objonly.pth) |
| Classification | ModelNet40(1k) |[finetune_modelnet.yaml](./cfgs/finetune_modelnet.yaml)| 93.80%| [here](https://github.com/Pang-Yatian/Point-MAE/releases/download/main/modelnet_1k.pth) |
| Classification | ModelNet40(8k) |[finetune_modelnet_8k.yaml](./cfgs/finetune_modelnet_8k.yaml)| 94.04%| [here](https://github.com/Pang-Yatian/Point-MAE/releases/download/main/modelnet_8k.pth) |
| Part segmentation| ShapeNetPart| [segmentation](./segmentation)| 86.1% mIoU| [here](https://github.com/Pang-Yatian/Point-MAE/releases/download/main/part_seg.pth) |
| Task | Dataset | Config | 5w10s Acc. (%)| 5w20s Acc. (%)| 10w10s Acc. (%)| 10w20s Acc. (%)|
| ----- | ----- |-----| -----| -----|-----|-----|
| Few-shot learning | ModelNet40 |[fewshot.yaml](./cfgs/fewshot.yaml)| 96.3 ± 2.5| 97.8 ± 1.8| 92.6 ± 4.1| 95.0 ± 3.0|
## 4. Point-MAE Pre-training
To pretrain Point-MAE on ShapeNet training set, run the following command. If you want to try different models or masking ratios etc., first create a new config file, and pass its path to --config.
```
CUDA_VISIBLE_DEVICES= python main.py --config cfgs/pretrain.yaml --exp_name
```
## 5. Point-MAE Fine-tuning
Fine-tuning on ScanObjectNN, run:
```
CUDA_VISIBLE_DEVICES= python main.py --config cfgs/finetune_scan_hardest.yaml \
--finetune_model --exp_name --ckpts
```
Fine-tuning on ModelNet40, run:
```
CUDA_VISIBLE_DEVICES= python main.py --config cfgs/finetune_modelnet.yaml \
--finetune_model --exp_name --ckpts
```
Voting on ModelNet40, run:
```
CUDA_VISIBLE_DEVICES= python main.py --test --config cfgs/finetune_modelnet.yaml \
--exp_name --ckpts
```
Few-shot learning, run:
```
CUDA_VISIBLE_DEVICES= python main.py --config cfgs/fewshot.yaml --finetune_model \
--ckpts --exp_name --way <5 or 10> --shot <10 or 20> --fold <0-9>
```
Part segmentation on ShapeNetPart, run:
```
cd segmentation
python main.py --ckpts --root path/to/data --learning_rate 0.0002 --epoch 300
```
## 6. Visualization
Visulization of pre-trained model on ShapeNet validation set, run:
```
python main_vis.py --test --ckpts --config cfgs/pretrain.yaml --exp_name
```
## Acknowledgements
Our codes are built upon [Point-BERT](https://github.com/lulutang0608/Point-BERT), [Pointnet2_PyTorch](https://github.com/erikwijmans/Pointnet2_PyTorch) and [Pointnet_Pointnet2_pytorch](https://github.com/yanx27/Pointnet_Pointnet2_pytorch)
## Reference
```
@inproceedings{pang2022masked,
title={Masked autoencoders for point cloud self-supervised learning},
author={Pang, Yatian and Wang, Wenxiao and Tay, Francis EH and Liu, Wei and Tian, Yonghong and Yuan, Li},
booktitle={Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part II},
pages={604--621},
year={2022},
organization={Springer}
}
```