CAGroup3D

This repo is the official implementation of the paper:

CAGroup3D: Class-Aware Grouping for 3D Object Detection on Point Clouds

PaperLink

NEWS

• Official implementation based on OpenPCDet is released.
• 🔥 CAGroup3D is accepted at NeurIPS 2022.

TODO

• [x] Implement CAGroup3D on OpenPCDet
• [ ] Fix the bug of slow inference (180ms->400ms) in our OpenPCDet version. Paper report is mmdet3d version.

Introduction

This paper presents a novel two-stage fully sparse convolutional 3D object detection framework, named CAGroup3D. The proposed method first generates some high-quality 3D proposals by leveraging the class-aware local group strategy on the object surface voxels with the same semantic predictions, which considers semantic consistency and diverse locality abandoned in previous bottom-up approaches. Then, to recover the features of missed voxels due to incorrect voxel-wise segmentation, we build a fully sparse convolutional RoI pooling module to directly aggregate fine-grained spatial information from backbone for further proposal refinement.

Requirements

The code is tested on the following environment:

• Unbuntu 18.04

• Python 3.7

• Pytorch 1.10

• CUDA 11.1

  Installation

• Clone this repo and install the pcdet library

  git clone https://github.com/Haiyang-W/CAGroup3D.git
  # install spconv
  pip install spconv-cu113
  cd CAGroup3D/
  python setup.py develop
  # if you meet some pakage not matched errors, just pip install them individually before install pcdet

• Compile additional CUDA ops

  # rotate iou ops
  cd CAGroup3D/pcdet/ops/rotated_iou/cuda_op
  python setup.py install
  # knn ops
  cd ../../knn
  python setup.py develop

• Install MinkowskiEngine

  apt-get install -y python3-dev libopenblas-dev
  pip install ninja==1.10.2.3
  pip install \
  -U git+https://github.com/NVIDIA/MinkowskiEngine@v0.5.4 \
  --install-option="--blas=openblas" \
  --install-option="--force_cuda" \
  -v \
  --no-deps
  # Or if you can not install MinkowskiEngine with pip sucessfully,
  git clone https://github.com/NVIDIA/MinkowskiEngine.git
  cd MinkowskiEngine
  python setup.py install --blas=openblas --force_cuda

Data preparation

We haven't achieved compatibility with the generated data of OpenPCDet yet and use the same data format as MMdeteciton3d for now. We will try to implement indoor data pre-processing based on OpenPCDet as soon as possible.

• follow MMdetection3D(0.15) to create data (ScanNetV2, SunRGBD). Note that the coordinate system adopted in MMdetection3D(0.15) is very different from MMdetection3D(>=1.0). We also provide processed data in GoogleDrive (highly recommended). For people from mainland China, we also provide BaiduDrive.

• remember to modify the DATA_PATH in tools/cfgs/dataset_configs/scannet_dataset.yaml, sunrgbd_dataset.yaml or link the generated data as follows:

  ln -s ${mmdet3d_scannet_dir} ./CAGroup3D/data/scannet
  ln -s ${mmdet3d_sunrgbd_dir} ./CAGroup3D/data/sunrgbd

Get started

ScanNetV2

• Training, num_gpus x batch_size can be set to 2x8 or 4x4.

  cd tools/
  CUDA_VISIBLE_DEVICES={} ./scripts/dist_train.sh {num_gpus} --cfg_file cfgs/scannet_models/CAGroup3D.yaml --ckpt_save_interval 1 --extra_tag {your name} --fix_random_seed

• Testing

  cd tools/
  # test single pth
  python test.py --cfg_file cfgs/scannet_models/CAGroup3D.yaml --ckpt {your pth}
  # or test all checkpoints
  python test.py --cfg_file cfgs/scannet_models/CAGroup3D.yaml --extra_tag {your name} --eval_all
  # dist test is also supported
  CUDA_VISIBLE_DEVICES={} ./scripts/dist_test.sh {num_gpus} --cfg_file cfgs/scannet_models/CAGroup3D.yaml --ckpt {your pth}

  Sun RGB-D

• Training, num_gpus x batch_size can be set to 2x8 or 4x4.

  cd tools/
  CUDA_VISIBLE_DEVICES={} ./scripts/dist_train.sh {num_gpus} --cfg_file cfgs/sunrgbd_models/CAGroup3D.yaml --ckpt_save_interval 1 --extra_tag {your name} --fix_random_seed

• Testing

  cd tools/
  # test single pth
  python test.py --cfg_file cfgs/sunrgbd_models/CAGroup3D.yaml --ckpt {your pth}
  # or test all checkpoints
  python test.py --cfg_file cfgs/sunrgbd_models/CAGroup3D.yaml --extra_tag {your name} --eval_all
  # dist test is also supported
  CUDA_VISIBLE_DEVICES={} ./scripts/dist_test.sh {num_gpus} --cfg_file cfgs/sunrgbd_models/CAGroup3D.yaml --ckpt {your pth}

Main Results

We reproduce CAGroup3D based on OpenPCDet(Paper reported is MMDet3D version). All models are trained with 4 3090 GPUs. The batch size of each card is 4 (2x8 is better). Please note that the pcdet version of scannet pretrained model has a higher mAP@50 (61.1 vs 60.3), but a lower mAP@25 (74.0 vs 74.5) than the original paper. Additionally, the sunrgbd pretrained model has a higher mAP@25 (67.1 vs 66.4) than the original paper. Since these two datasets fluctuate significantly, training more times should yield better results than the pre-trained models we provide.

Citation

Please consider citing our work as follows if it is helpful.

@inproceedings{
wang2022cagroupd,
title={{CAG}roup3D: Class-Aware Grouping for 3D Object Detection on Point Clouds},
author={Haiyang Wang and Lihe Ding and Shaocong Dong and Shaoshuai Shi and Aoxue Li and Jianan Li and Zhenguo Li and Liwei Wang},
booktitle={Advances in Neural Information Processing Systems},
editor={Alice H. Oh and Alekh Agarwal and Danielle Belgrave and Kyunghyun Cho},
year={2022},
url={https://openreview.net/forum?id=nLKkHwYP4Au}
}

Acknowledgments

This project is based on the following codebases.

• OpenPCDet
• FCAF3D
• RBGNet