Implementation of PointPillars in PyTorch for KITTI 3D Object Detetcion

• This repository references open-mmlab's work.

• Clone this repository

  git clone [email protected]:shangjie-li/pointpillars.git
  

• Install PyTorch environment with Anaconda (Tested on Ubuntu 16.04 & CUDA 10.2)

  conda create -n pcdet.v0.5.0 python=3.6
  conda activate pcdet.v0.5.0
  cd pointpillars
  pip install -r requirements.txt
  

• Install spconv

  # Try the command below:
  pip install spconv-cu102
  
  # If there is `ERROR: Cannot uninstall 'certifi'.`, try:
  pip install spconv-cu102 --ignore-installed
  

• Compile external modules

  cd pointpillars
  python setup.py develop
  

• Install visualization tools

  pip install mayavi
  pip install pyqt5
  
  # If you want import opencv, run:
  pip install opencv-python
  
  # If you want import open3d, run:
  pip install open3d-python
  

• Download KITTI3D dataset: calib, velodyne, label_2 and image_2.
• Download road plane for data augmentation.
• Organize the downloaded files as follows

  pointpillars
  ├── data
  │   ├── kitti
  │   │   │── ImageSets
  │   │   │── training
  │   │   │   ├──calib & velodyne & label_2 & image_2 & planes
  │   │   │── testing
  │   │   │   ├──calib & velodyne & image_2
  ├── layers
  ├── utils
  

• Generate the ground truth database and data infos by running the following command

  # This will create gt_database dir and info files in pointpillars/data/kitti.
  cd pointpillars
  python -m data.kitti_dataset create_kitti_infos data/config.yaml
  

• Display the dataset

  # Display the training dataset with data augmentation
  python dataset_player.py --training --data_augmentation
  
  # Display the testing dataset
  python dataset_player.py
  

• Run the demo with a pretrained model (Download pointpillar_7728.pth and save it in pointpillars/weights.)

  # Run on a single file
  python demo.py --ckpt=weights/pointpillar_7728.pth --data_path=data/kitti/training/velodyne/000008.bin
  
  # Run on a folder
  python demo.py --ckpt=weights/pointpillar_7728.pth --data_path=data/kitti/training/velodyne
  

• Run the command below to train

  python train.py --batch_size=2
  

• Run the command below to evaluate

  python test.py --ckpt=weights/pointpillar_7728.pth
  

• The 3D detection performance on KITTI should be

  Class	AP (R11) BEV	AP (R11) 3D
  Car (Iou=0.7)	89.6590, 87.1725, 84.3762	86.4617, 77.2839, 74.6530
  Pedestrian (Iou=0.5)	61.6348, 56.2747, 52.6007	57.7500, 52.2916, 47.9072
  Cyclist (Iou=0.5)	82.2593, 66.1110, 62.5585	80.0483, 62.6080, 59.5260

  • Report in different difficulties, which are Easy, Moderate and Hard.