PyTorch implementation of "HVPR: Hybrid Voxel-Point Representation for Single-stage 3D Object Detection"

This is the implementation of the paper "HVPR: Hybrid Voxel-Point Representation for Single-stage 3D Object Detection (CVPR 2021)".

Our code is mainly based on OpenPCDet. We also plan to release the code based on PointPillars. For more information, checkout the project site [website] and the paper [PDF].

Dependencies

• Python >= 3.6
• PyTorch >= 1.4.0

Update

• 20/06/21: First update

Installation

• Clone this repo, and follow the steps below (or you can follow the installation steps in OpenPCDet).

1. Clone this repository:

   git clone https://github.com/cvlab-yonsei/HVPR.git

2. Install the dependent libraries:

   pip install -r requirements.txt

3. Install the SparseConv library from spconv.

4. Install pcdet library:

   python setup.py develop

Datasets

• KITTI 3D Object Detection

1. Please download the official KITTI 3D object detection dataset and organize the downloaded files as follows (the road planes could be downloaded from [road plane], which are optional for data augmentation in the training):

   HVPR
   ├── data
   │   ├── kitti
   │   │   │── ImageSets
   │   │   │── training
   │   │   │   ├──calib & velodyne & label_2 & image_2 & (optional: planes)
   │   │   │── testing
   │   │   │   ├──calib & velodyne & image_2
   ├── pcdet
   ├── tools

2. Generate the data infos by running the following command:

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

Training

• The config files is in tools/cfgs/kitti_models, and you can easily train your own model like:

  cd tools
  sh scripts/train_hvpr.sh 

• You can freely define parameters with your own settings like:

  cd tools
  sh scripts train_hvpr.sh --gpus 1 --result_path 'your_dataset_directory' --exp_dir 'your_log_directory'

Evaluation

• Test your own model:

  cd tools
  sh scripts/eval_hvpr.sh

Pre-trained model

• Download our pre-trained model.
  [KITTI 3D Car]

Bibtex

@article{noh2021hvpr,
  title={HVPR: Hybrid Voxel-Point Representation for Single-stage 3D Object Detection},
  author={Noh, Jongyoun and Lee, Sanghoon and Ham, Bumsub},
  journal={arXiv preprint arXiv:2104.00902},
  year={2021}
}

References

Our work is mainly built on OpenPCDet codebase. Portions of our code are also borrowed from spconv, MemAE, and CBAM. Thanks to the authors!