Not All Points Are Equal: Learning Highly Efficient Point-based Detectors for 3D LiDAR Point Clouds (CVPR 2022, Oral)

This is the official implementation of IA-SSD (CVPR 2022), a simple and highly efficient point-based detector for 3D LiDAR point clouds. For more details, please refer to:

Not All Points Are Equal: Learning Highly Efficient Point-based Detectors for 3D LiDAR Point Clouds
Yifan Zhang, Qingyong Hu*, Guoquan Xu, Yanxin Ma, Jianwei Wan, Yulan Guo

[Paper] [Video]

Getting Started

Installation

a. Clone this repository

git clone https://github.com/yifanzhang713/IA-SSD.git && cd IA-SSD

b. Configure the environment

We have tested this project with the following environments:

• Ubuntu18.04/20.04
• Python = 3.7
• PyTorch = 1.1
• CUDA = 10.0
• CMake >= 3.13
• spconv = 1.0

  # install spconv=1.0 library
  git clone https://github.com/yifanzhang713/spconv1.0.git
  cd spconv1.0
  sudo apt-get install libboostall-dev
  python setup.py bdist_wheel
  pip install ./dist/spconv-1.0*   # wheel file name may be different
  cd ..

*You are encouraged to try to install higher versions above, please refer to the official github repository for more information. Note that the maximum number of parallel frames during inference might be slightly decrease due to the larger initial GPU memory footprint with updated Pytorch version.

c. Install pcdet toolbox.

pip install -r requirements.txt
python setup.py develop

d. Prepare the datasets.

Download the official KITTI with road planes and Waymo datasets, then organize the unzipped files as follows:

IA-SSD
├── data
│   ├── kitti
│   │   ├── ImageSets
│   │   ├── training
│   │   │   ├──calib & velodyne & label_2 & image_2 & (optional: planes)
│   │   ├── testing
│   │   ├── calib & velodyne & image_2
│   ├── waymo
│   │   │── ImageSets
│   │   │── raw_data
│   │   │   │── segment-xxxxxxxx.tfrecord
|   |   |   |── ...
|   |   |── waymo_processed_data_v0_5_0
│   │   │   │── segment-xxxxxxxx/
|   |   |   |── ...
│   │   │── waymo_processed_data_v0_5_0_gt_database_train_sampled_1/
│   │   │── waymo_processed_data_v0_5_0_waymo_dbinfos_train_sampled_1.pkl
│   │   │── waymo_processed_data_v0_5_0_gt_database_train_sampled_1_global.npy (optional)
│   │   │── waymo_processed_data_v0_5_0_infos_train.pkl (optional)
│   │   │── waymo_processed_data_v0_5_0_infos_val.pkl (optional)
├── pcdet
├── tools

Generate the data infos by running the following commands:

# KITTI dataset
python -m pcdet.datasets.kitti.kitti_dataset create_kitti_infos tools/cfgs/dataset_configs/kitti_dataset.yaml

# Waymo dataset
python -m pcdet.datasets.waymo.waymo_dataset --func create_waymo_infos \
    --cfg_file tools/cfgs/dataset_configs/waymo_dataset.yaml

Quick Inference

We provide the pre-trained weight file so you can just run with that:

cd tools 
# To achieve fully GPU memory footprint (NVIDIA RTX2080Ti, 11GB).
python test.py --cfg_file cfgs/kitti_models/IA-SSD.yaml --batch_size 100 \
    --ckpt IA-SSD.pth --set MODEL.POST_PROCESSING.RECALL_MODE 'speed'

# To reduce the pressure on the CPU during preprocessing, a suitable batchsize is recommended, e.g. 16. (Over 5 batches per second on RTX2080Ti)
python test.py --cfg_file cfgs/kitti_models/IA-SSD.yaml --batch_size 16 \
    --ckpt IA-SSD.pth --set MODEL.POST_PROCESSING.RECALL_MODE 'speed' 

• Then detailed inference results can be found here.

Training

The configuration files are in tools/cfgs/kitti_models/IA-SSD.yaml and tools/cfgs/waymo_models/IA-SSD.yaml, and the training scripts are in tools/scripts.

Train with single or multiple GPUs: (e.g., KITTI dataset)

python train.py --cfg_file cfgs/kitti_models/IA-SSD.yaml

# or 

sh scripts/dist_train.sh ${NUM_GPUS} --cfg_file cfgs/kitti_models/IA-SSD.yaml

Evaluation

Evaluate with single or multiple GPUs: (e.g., KITTI dataset)

python test.py --cfg_file cfgs/kitti_models/IA-SSD.yaml  --batch_size ${BATCH_SIZE} --ckpt ${PTH_FILE}

# or

sh scripts/dist_test.sh ${NUM_GPUS} \
    --cfg_file cfgs/kitti_models/IA-SSD.yaml --batch_size ${BATCH_SIZE} --ckpt ${PTH_FILE}

Experimental results

KITTI dataset

Quantitative results of different approaches on KITTI dataset (test set):

Qualitative results of our IA-SSD on KITTI dataset:

Quantitative results of different approaches on Waymo dataset (validation set):

Qualitative results of our IA-SSD on Waymo dataset:

Quantitative results of different approaches on ONCE dataset (validation set):

Qualitative result of our IA-SSD on ONCE dataset:

Citation

If you find this project useful in your research, please consider citing:

@inproceedings{zhang2022not,
  title={Not All Points Are Equal: Learning Highly Efficient Point-based Detectors for 3D LiDAR Point Clouds},
  author={Zhang, Yifan and Hu, Qingyong and Xu, Guoquan and Ma, Yanxin and Wan, Jianwei and Guo, Yulan},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2022}
}

Acknowledgement

• This work is built upon the OpenPCDet (version 0.5), an open source toolbox for LiDAR-based 3D scene perception. Please refer to the official github repository for more information.

• Parts of our Code refer to 3DSSD-pytorch-openPCDet library and the the recent work SASA.

License

This project is released under the Apache 2.0 license.

Related Repos

1. RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds
2. SensatUrban: Learning Semantics from Urban-Scale Photogrammetric Point Clouds
3. 3D-BoNet: Learning Object Bounding Boxes for 3D Instance Segmentation on Point Clouds
4. SpinNet: Learning a General Surface Descriptor for 3D Point Cloud Registration
5. SQN: Weakly-Supervised Semantic Segmentation of Large-Scale 3D Point Clouds
6. SoTA-Point-Cloud: Deep Learning for 3D Point Clouds: A Survey