Implementation of SECOND in PyTorch for KITTI 3D Object Detetcion

• This repository references open-mmlab's work.

• Clone this repository

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

  second
  ├── 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 second/data/kitti.
  cd second
  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 --show_boxes
  
  # Display the testing dataset
  python dataset_player.py --show_boxes
  

• Run the demo with a pretrained model (Download second_7862.pth and save it in second/weights.)

  # Run on the testing dataset
  python demo.py --ckpt=weights/second_7862.pth
  
  # Run on a single sample from the testing dataset
  python demo.py --ckpt=weights/second_7862.pth --sample_idx=000008
  

• Run the command below to train

  python train.py --batch_size=2
  

• Run the command below to evaluate

  python test.py --ckpt=weights/second_7862.pth
  

• The 3D detection performance on KITTI should be

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