🪖 ReCon: Contrast with Reconstruct

Contrast with Reconstruct: Contrastive 3D Representation Learning Guided by Generative Pretraining ICML 2023
Zekun Qi*, Runpei Dong*, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma and Li Yi

OpenReview | arXiv | Models

This repository contains the code release of ReCon: Contrast with Reconstruct: Contrastive 3D Representation Learning Guided by Generative Pretraining (ICML 2023). ReCon is also short for reconnaissance 🪖.

Contrast with Reconstruct

News

• 💥 Mar, 2024: Check out our latest work ShapeLLM, which achieves 95.25% fine-tuned accuracy and 65.4 zero-shot accuracy on ScanObjectNN
• 📌 Aug, 2023: Check out our exploration of efficient conditional 3D generation VPP
• 📌 Jun, 2023: Check out our exploration of pre-training in 3D scenes Point-GCC
• 🎉 Apr, 2023: ReCon accepted by ICML 2023
• 💥 Feb, 2023: Check out our previous work ACT, which has been accepted by ICLR 2023

1. Requirements

PyTorch >= 1.7.0; python >= 3.7; CUDA >= 9.0; GCC >= 4.9; torchvision;

# Quick Start
conda create -n recon python=3.8 -y
conda activate recon

conda install pytorch==1.10.0 torchvision==0.11.0 cudatoolkit=11.3 -c pytorch -c nvidia
# pip install torch==1.10.0+cu113 torchvision==0.11.0+cu113 torchaudio==0.10.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html

# Install basic required packages
pip install -r requirements.txt
# Chamfer Distance
cd ./extensions/chamfer_dist && python setup.py install --user
# PointNet++
pip install "git+https://github.com/erikwijmans/Pointnet2_PyTorch.git#egg=pointnet2_ops&subdirectory=pointnet2_ops_lib"

2. Datasets

We use ShapeNet, ScanObjectNN, ModelNet40 and ShapeNetPart in this work. See DATASET.md for details.

3. ReCon Models

Task	Dataset	Config	Acc.	Checkpoints Download
Pre-training	ShapeNet	pretrain_base.yaml	N.A.	ReCon
Classification	ScanObjectNN	finetune_scan_hardest.yaml	91.26%	PB_T50_RS
Classification	ScanObjectNN	finetune_scan_objbg.yaml	95.35%	OBJ_BG
Classification	ScanObjectNN	finetune_scan_objonly.yaml	93.80%	OBJ_ONLY
Classification	ModelNet40(1k)	finetune_modelnet.yaml	94.5%	ModelNet_1k
Classification	ModelNet40(8k)	finetune_modelnet_8k.yaml	94.7%	ModelNet_8k
Zero-Shot	ModelNet10	zeroshot_modelnet10.yaml	75.6%	ReCon zero-shot
Zero-Shot	ModelNet10*	zeroshot_modelnet10.yaml	81.6%	ReCon zero-shot
Zero-Shot	ModelNet40	zeroshot_modelnet40.yaml	61.7%	ReCon zero-shot
Zero-Shot	ModelNet40*	zeroshot_modelnet40.yaml	66.8%	ReCon zero-shot
Zero-Shot	ScanObjectNN	zeroshot_scan_objonly.yaml	43.7%	ReCon zero-shot
Linear SVM	ModelNet40	svm.yaml	93.4%	ReCon svm
Part Segmentation	ShapeNetPart	segmentation	86.4% mIoU	part seg

Task	Dataset	Config	5w10s (%)	5w20s (%)	10w10s (%)	10w20s (%)	Download
Few-shot learning	ModelNet40	fewshot.yaml	97.3 ± 1.9	98.9 ± 1.2	93.3 ± 3.9	95.8 ± 3.0	ReCon

The checkpoints and logs have been released on Google Drive. You can use the voting strategy in classification testing to reproduce the performance reported in the paper. For classification downstream tasks, we randomly select 8 seeds to obtain the best checkpoint. For zero-shot learning, * means that we use all the train/test data for zero-shot transfer.

4. ReCon Pre-training

Pre-training with the default configuration, run the script:

sh scripts/pretrain.sh <GPU> <exp_name>

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=<GPU> python main.py --config <config_path> --exp_name <exp_name>

5. ReCon Classification Fine-tuning

Fine-tuning with the default configuration, run the script:

bash scripts/cls.sh <GPU> <exp_name> <path/to/pre-trained/model>

Or, you can use the command.

Fine-tuning on ScanObjectNN, run:

CUDA_VISIBLE_DEVICES=<GPUs> python main.py --config cfgs/full/finetune_scan_hardest.yaml \
--finetune_model --exp_name <exp_name> --ckpts <path/to/pre-trained/model>

Fine-tuning on ModelNet40, run:

CUDA_VISIBLE_DEVICES=<GPUs> python main.py --config cfgs/full/finetune_modelnet.yaml \
--finetune_model --exp_name <exp_name> --ckpts <path/to/pre-trained/model>

6. ReCon Test&Voting

Test&Voting with the default configuration, run the script:

bash scripts/test.sh <GPU> <exp_name> <path/to/best/fine-tuned/model>

or:

CUDA_VISIBLE_DEVICES=<GPUs> python main.py --test --config cfgs/finetune_modelnet.yaml \
--exp_name <output_file_name> --ckpts <path/to/best/fine-tuned/model>

7. ReCon Few-Shot

Few-shot with the default configuration, run the script:

sh scripts/fewshot.sh <GPU> <exp_name> <path/to/pre-trained/model> <way> <shot> <fold>

or

CUDA_VISIBLE_DEVICES=<GPUs> python main.py --config cfgs/full/fewshot.yaml --finetune_model \
--ckpts <path/to/pre-trained/model> --exp_name <exp_name> --way <5 or 10> --shot <10 or 20> --fold <0-9>

8. ReCon Zero-Shot

Zero-shot with the default configuration, run the script:

bash scripts/zeroshot.sh <GPU> <exp_name> <path/to/pre-trained/model>

9. ReCon Part Segmentation

Part segmentation on ShapeNetPart, run:

cd segmentation
bash seg.sh <GPU> <exp_name> <path/to/pre-trained/model>

or

cd segmentation
python main.py --ckpts <path/to/pre-trained/model> --log_dir <path/to/log/dir> --learning_rate 0.0001 --epoch 300

Test part segmentation on ShapeNetPart, run:

cd segmentation
bash test.sh <GPU> <exp_name> <path/to/best/fine-tuned/model>

10. ReCon Linear SVM

Linear SVM on ModelNet40, run:

sh scripts/svm.sh <GPU> <exp_name> <path/to/pre-trained/model> 

11. Visualization

We use PointVisualizaiton repo to render beautiful point cloud image, including specified color rendering and attention distribution rendering.

Contact

If you have any questions related to the code or the paper, feel free to email Zekun ([email protected]) or Runpei ([email protected]).

License

ReCon is released under MIT License. See the LICENSE file for more details. Besides, the licensing information for pointnet2 modules is available here.

Acknowledgements

This codebase is built upon Point-MAE, Point-BERT, CLIP, Pointnet2_PyTorch and ACT

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{qi2023recon,
  title={Contrast with Reconstruct: Contrastive 3D Representation Learning Guided by Generative Pretraining},
  author={Qi, Zekun and Dong, Runpei and Fan, Guofan and Ge, Zheng and Zhang, Xiangyu and Ma, Kaisheng and Yi, Li},
  booktitle={International Conference on Machine Learning (ICML) },
  year={2023}
}

and closely related work ACT and ShapeLLM:

@inproceedings{dong2023act,
  title={Autoencoders as Cross-Modal Teachers: Can Pretrained 2D Image Transformers Help 3D Representation Learning?},
  author={Runpei Dong and Zekun Qi and Linfeng Zhang and Junbo Zhang and Jianjian Sun and Zheng Ge and Li Yi and Kaisheng Ma},
  booktitle={The Eleventh International Conference on Learning Representations (ICLR) },
  year={2023},
  url={https://openreview.net/forum?id=8Oun8ZUVe8N}
}
@article{qi2024shapellm,
  author = {Qi, Zekun and Dong, Runpei and Zhang, Shaochen and Geng, Haoran and Han, Chunrui and Ge, Zheng and Wang, He and Yi, Li and Ma, Kaisheng},
  title = {ShapeLLM: Universal 3D Object Understanding for Embodied Interaction},
  journal = {arXiv preprint arXiv:2402.17766},
  year = {2024}
}