Awesome

Spatio-temporal Self-Supervised Representation Learning for 3D Point Clouds

This is the official code implementation for the paper "Spatio-temporal Self-Supervised Representation Learning for 3D Point Clouds" (ICCV 2021) paper

Checklist

Self-supervised Pre-training Framework

• BYOL
• SimCLR

Downstream Tasks

• Shape Classification
• Semantic Segmentation
• Indoor Object Detection
• Outdoor Object Detection

Installation

The code was tested with the following environment: Ubuntu 18.04, python 3.7, pytorch 1.7.1, torchvision 0.8.2 and CUDA 11.1.

For self-supervised pre-training, run the following command:

git clone https://github.com/yichen928/STRL.git
cd STRL
pip install -r requirements.txt

For downstream tasks, please refer to the Downstream Tasks section.

Datasets

Please download the used dataset with the following links:

• ShapeNet: https://drive.google.com/uc?id=1sJd5bdCg9eOo3-FYtchUVlwDgpVdsbXB

• ModelNet40: https://shapenet.cs.stanford.edu/media/modelnet40_normal_resampled.zip

• ScanNet (subset): Please follow the instruction in their official website. The 25k frames subset is enough for our model. You may also need to download the preprocessed data for evaluation here.

Make sure to put the files in the following structure:

|-- ROOT
|	|-- BYOL
|		|-- data
|			|-- modelnet40_normal_resampled_cache
|			|-- shapenet57448xyzonly.npz
|			|-- scannet
|				|-- scannet_frames_25k

Pre-training

BYOL framework

Please run the following command:

python BYOL/train.py

You need to edit the config file BYOL/config/config.yaml to switch different backbone architectures (currently including BYOL-pointnet-cls, BYOL-dgcnn-cls, BYOL-dgcnn-semseg, BYOL-votenet-detection).

Pre-trained Models

You can find the checkpoints of the pre-training and downstream tasks in our Google Drive.

Linear Evaluation

For PointNet or DGCNN classification backbones, you may evaluate the learnt representation with linear SVM classifier by running the following command:

For PointNet:

python BYOL/evaluate_pointnet.py -w /path/to/your/pre-trained/checkpoints

For DGCNN:

python BYOL/evaluate_dgcnn.py -w /path/to/your/pre-trained/checkpoints

Downstream Tasks

Checkpoints Transformation

You can transform the pre-trained checkpoints to different downstream tasks by running:

For VoteNet:

python BYOL/transform_ckpt_votenet.py --input_path /path/to/your/pre-trained/checkpoints --output_path /path/to/the/transformed/checkpoints

For other backbones:

python BYOL/transform_ckpt.py --input_path /path/to/your/pre-trained/checkpoints --output_path /path/to/the/transformed/checkpoints

Fine-tuning and Evaluation for Downstream Tasks

For the fine-tuning and evaluation of downstream tasks, please refer to other corresponding repos. We sincerely thank all these authors for their nice work!

• Classification: WangYueFt/dgcnn
• Semantic Segmentation: AnTao97/dgcnn.pytorch
• Indoor Object Detection: facebookresearch/votenet

Citation

If you found our paper or code useful for your research, please cite the following paper:

@article{huang2021spatio,
  title={Spatio-temporal Self-Supervised Representation Learning for 3D Point Clouds},
  author={Huang, Siyuan and Xie, Yichen and Zhu, Song-Chun and Zhu, Yixin},
  journal={arXiv preprint arXiv:2109.00179},
  year={2021}
}