Official implementation of 'Learning 3D Representations from 2D Pre-trained Models via Image-to-Point Masked Autoencoders'.

The paper has been accepted by CVPR 2023 🔥.

• The pre-training and fine-tuning code of I2P-MAE has been released.
• The 3D-only variant of I2P-MAE is our previous work, Point-M2AE, accepted by NeurIPS 2022 and open-sourced. We have released its pre-training and fine-tuning code.
• 📣 Please check our latest work Point-NN, Parameter is Not All You Need accepted by CVPR 2023, which, for the first time, acheives 3D understanding with $\color{darkorange}{No\ Parameter\ or\ Training.}$ 💥
• 📣 Please check our latest work PiMAE accepted by CVPR 2023, which promotes 3D and 2D interaction to improve 3D object detection performance.

Comparison with existing MAE-based 3D models on the three spilts of ScanObjectNN:

We propose an alternative to obtain superior 3D representations from 2D pre-trained models via Image-to-Point Masked Autoencoders, named as I2P-MAE. By self-supervised pre-training, we leverage the well learned 2D knowledge to guide 3D masked autoencoding, which reconstructs the masked point tokens with an encoder-decoder architecture. Specifically, we conduct two types of image-to-point learning schemes: 2D-guided masking and 2D-semantic reconstruction. In this way, the 3D network can effectively inherit high-level 2D semantics learned from rich image data for discriminative 3D modeling.

Guided by pre-trained CLIP on ShapeNet, I2P-MAE is evaluated by Linear SVM on ModelNet40 and ScanObjectNN (OBJ-BG split) datasets, without downstream fine-tuning:

Synthetic shape classification on ModelNet40 with 1k points:

Real-world shape classification on ScanObjectNN:

Create a conda environment and install basic dependencies:

git clone https://github.com/ZrrSkywalker/I2P-MAE.git
cd I2P-MAE

conda create -n i2pmae python=3.7
conda activate i2pmae

# Install the according versions of torch and torchvision
conda install pytorch torchvision cudatoolkit
# e.g., conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3

pip install -r requirements.txt
pip install torch-scatter -f https://data.pyg.org/whl/torch-1.11.0+cu113.html

Install GPU-related packages:

# Chamfer Distance and EMD
cd ./extensions/chamfer_dist
python setup.py install --user
cd ../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

For pre-training and fine-tuning, please follow DATASET.md to install ShapeNet, ModelNet40, ScanObjectNN, and ShapeNetPart datasets, referring to Point-BERT. Specially for Linear SVM evaluation, download the official ModelNet40 dataset and put the unzip folder under data/.

The final directory structure should be:

│I2P-MAE/
├──cfgs/
├──datasets/
├──data/
│   ├──ModelNet/
│   ├──ModelNetFewshot/
│   ├──modelnet40_ply_hdf5_2048/  # Specially for Linear SVM
│   ├──ScanObjectNN/
│   ├──ShapeNet55-34/
│   ├──shapenetcore_partanno_segmentation_benchmark_v0_normal/
├──...

I2P-MAE is pre-trained on ShapeNet dataset with the config file cfgs/pre-training/i2p-mae.yaml. Run:

CUDA_VISIBLE_DEVICES=0 python main.py --config cfgs/pre-training/i2p-mae.yaml --exp_name pre-train

To evaluate the pre-trained I2P-MAE by Linear SVM, create a folder ckpts/ and download the pre-train.pth into it. Use the configs in cfgs/linear-svm/ and indicate the evaluation dataset by --test_svm.

For ModelNet40, run:

CUDA_VISIBLE_DEVICES=0 python main.py --config cfgs/linear-svm/modelnet40.yaml --test_svm modelnet40 --exp_name test_svm --ckpts ./ckpts/pre-train.pth

For ScanObjectNN (OBJ-BG split), run:

CUDA_VISIBLE_DEVICES=0 python main.py --config cfgs/linear-svm/scan_obj-bg.yaml --test_svm scan --exp_name test_svm --ckpts ./ckpts/pre-train.pth

Please create a folder ckpts/ and download the pre-train.pth into it. The fine-tuning configs are in cfgs/fine-tuning/.

For ModelNet40, run:

CUDA_VISIBLE_DEVICES=0 python main.py --config cfgs/fine-tuning/modelnet40.yaml --finetune_model --exp_name finetune --ckpts ckpts/pre-train.pth

For the three splits of ScanObjectNN, run:

CUDA_VISIBLE_DEVICES=0 python main.py --config cfgs/fine-tuning/scan_pb.yaml --finetune_model --exp_name finetune --ckpts ckpts/pre-train.pth

CUDA_VISIBLE_DEVICES=0 python main.py --config cfgs/fine-tuning/scan_obj.yaml --finetune_model --exp_name finetune --ckpts ckpts/pre-train.pth

CUDA_VISIBLE_DEVICES=0 python main.py --config cfgs/fine-tuning/scan_obj-bg.yaml --finetune_model --exp_name finetune --ckpts ckpts/pre-train.pth

This repo benefits from Point-M2AE, Point-BERT, Point-MAE, and CLIP. Thanks for their wonderful works.

If you have any question about this project, please feel free to contact zhangrenrui@pjlab.org.cn.