qianpeisheng / 3DIoUMatch

[CVPR 2021] PyTorch implementation of 3DIoUMatch: Leveraging IoU Prediction for Semi-Supervised 3D Object Detection.

3DIoUMatch

Updates

06/16/2022: Added support for Windows by @nicolgo. Thanks for the contribution!

10/15/2021: Fixed a bug which caused SUN RGB-D unlabeled data to go unaugmented, see this commit. This bug was introduced when we prepared the code for release so the results in the paper are unaffected. Credit to Bowen Cheng.

Introduction

News: Our paper has been accepted by CVPR 2021!

This is the code release of our paper 3DIoUMatch: Leveraging IoU Prediction for Semi-Supervised 3D Object Detection. (arXiv report here).

In this repository, we provide 3DIoUMatch implementation (with Pytorch) based on VoteNet and SESS, as well as the training and evaluation scripts on SUNRGB-D and ScanNet.

Please refer to our project page for more information.

The PV-RCNN based version is here.

Citation

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

@inproceedings{wang20213dioumatch,
  title={3DIoUMatch: Leveraging iou prediction for semi-supervised 3d object detection},
  author={Wang, He and Cong, Yezhen and Litany, Or and Gao, Yue and Guibas, Leonidas J},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={14615--14624},
  year={2021}
}

Installation

Preparation: A Ubuntu system with GPU.

Install Nvidia driver and CUDA Toolkit.

$ nvidia-smi  # check driver
$ nvcc --version # check toolkit

Install Python -- This repo is tested with Python 3.7.6.

Install NumPy -- This repo is tested with NumPy 1.18.5. Please make sure your NumPy version is at least 1.18.

Install PyTorch with CUDA -- This repo is tested with PyTorch 1.5.1, CUDA 10.1. It may work with newer versions, but that is not guaranteed. A lower version may be problematic.

pip install torch==1.5.1 torchvision==0.6.1

or

conda install pytorch==1.5.1 torchvision==0.6.1 cudatoolkit=10.1 -c pytorch

Install TensorFlow (for TensorBoard) -- This repo is tested with TensorFlow 2.2.0.

Compile the CUDA code for PointNet++, which is used in the backbone network:

cd pointnet2
python setup.py install

If there is a problem, please refer to Pointnet2/Pointnet++ PyTorch

Compile the CUDA code for general 3D IoU calculation in OpenPCDet:

cd OpenPCDet
python setup.py develop

I deleted the CUDA kernels except 3D IoU calculation in OpenPCDet for faster installation.

Install dependencies:

pip install -r requirements.txt

Installation on Windows

Preparation: GPU support CUDA 10.2, not use GeForce 30 series with Ampere micro-architecture, anaconda is recommended.

> nvidia-smi # check nvidia driver
> nvcc --version # check CUDA Toolkit version
> cl # check MSVC compiler, if no, add compile to environment path

1. create a virtual environment: conda create --name iou python=3.7
2. install numpy: conda install numpy -y
3. install pytorch and cudatoolkit: conda install pytorch==1.5.1 torchvision==0.6.1 cudatoolkit=10.2 -c pytorch
4. degrade setuptools: pip install setuptools=59.6
5. install pointnet2: cd pointnet2; python ./setup.py install
6. install OpenPCDet: cd OpenPCDet; python ./setup.py develop
7. install tensorflow: pip install tensorflow
8. install requirements: pip install -r ./requirements.txt

Datasets

ScanNet

Please follow the instructions in scannet/README.md. using the download script with -o $(pwd) --types _vh_clean_2.ply .aggregation.json _vh_clean_2.0.010000.segs.json .txt options to download data.

SUNRGB-D

Please follow the instructions in sunrgbd/README.md.

Download Pre-trained and Trained Models

We provide the pre-trained models of ScanNet 10% and SUNRGB-D 5%.

We also provide the trained models of ScanNet 10% and SUNRGB-D 5%.

You may download the models and put them into ckpts.

We provide 2 data splits of ScanNet 10% and SUNRGB-D 5% in scannet/meta_data and sunrgbd/sunrgbd_trainval, respectively.

Pre-training

Please run:

sh run_pretrain.sh <GPU_ID> <LOG_DIR> <DATASET> <LABELED_LIST>

For example:

sh run_pretrain.sh 0 pretrain_scannet scannet scannetv2_train_0.1.txt

sh run_pretrain.sh 0 pretrain_sunrgbd sunrgbd sunrgbd_v1_train_0.05.txt

Training

Please run:

sh run_train.sh <GPU_ID> <LOG_DIR> <DATASET> <LABELED_LIST> <PRETRAIN_CKPT>

For example, use the downloaded models:

sh run_train.sh 0 train_scannet scannet scannetv2_train_0.1.txt ckpts/scan_0.1_pretrain.tar

sh run_train.sh 0 train_sunrgbd sunrgbd sunrgbd_v1_train_0.05.txt ckpts/sun_0.05_pretrain.tar

You may modify the script by adding --view_stats to load labels on unlabeled data and view the statistics on the unlabeled data (e.g. average IoU, class prediction accuracy).

Evaluation

Please run:

sh run_eval.sh <GPU_ID> <LOG_DIR> <DATASET> <LABELED_LIST> <CKPT>

For example, use the downloaded models:

sh run_eval.sh 0 eval_scannet scannet scannetv2_train_0.1.txt ckpts/scan_0.1.tar

sh run_eval.sh 0 eval_sunrgbd sunrgbd sunrgbd_v1_train_0.05.txt ckpts/sun_0.05.tar

For evaluation with IoU optimization, please run:

Please run:

sh run_eval_opt.sh <GPU_ID> <LOG_DIR> <DATASET> <LABELED_LIST> <CKPT> <OPT_RATE>

The number of steps (of optimization) is by default 10.

Acknowledgements

Our implementation uses code from the following repositories:

• Deep Hough Voting for 3D Object Detection in Point Clouds
• SESS: Self-Ensembling Semi-Supervised 3D Object Detection
• Pointnet2/Pointnet++ PyTorch
• OpenPCDet