/3D-Detection-Code

3D object detection code for autonomous / tele-operated vehicles.

Primary LanguagePython

3D-Detection-Code

General 3d object detection code for kitti dataset and nuscene dataset. Code to implement 3d object detection code for better understanding and cleaning.

Environment Setup

1.Install dependence python packages for environment setup

conda install scikit-image scipy numba pillow matplotlib
pip install fire tensorboardX protobuf opencv-python

Prepare Dataset

  • KITTI Dataset Preparation

Download KITTI dataset and create some directories first:

|── KITTI_DATASET_ROOT
    ├── Train    <-- 7481 train data
    |   ├── image_2 <-- for visualization
    |   ├── calib
    |   ├── label_2
    |   ├── velodyne
    |   └── velodyne_reduced <-- empty directory
    └── Test     <-- 7580 test data
        ├── image_2 <-- for visualization
        ├── calib
        ├── velodyne
        └── velodyne_reduced <-- empty directory

Download Nuscenes dataset and create some directories first:

# For nuScenes Dataset                                                                                                                                        │
└── NUSCENES_TRAINVAL_DATASET_ROOT                                                                                                                            
       ├── samples       <-- key frames                                                                                                                       
       ├── sweeps        <-- frames without annotation                                                                                                        
       ├── maps          <-- unused                                                                                                                           
       └── v1.0-trainval <-- metadata and annotations                                                                                                         
└── NUSCENES_TEST_DATASET_ROOT  
       ├── samples       <-- key frames                                                                                                                       
       ├── sweeps        <-- frames without annotation                                                                                                        
       ├── maps          <-- unused    
       └── v1.0-test     <-- metadata
  • convert to pickle file
# KITTI
python create_data.py kitti_data_prep --root_path=KITTI_DATASET_ROOT
# nuScenes
python create_data.py nuscenes_data_prep --root_path=NUSCENES_TRAINVAL_DATASET_ROOT --version="v1.0-trainval" --nsweeps=10
python create_data.py nuscenes_data_prep --root_path=NUSCENES_TEST_DATASET_ROOT --version="v1.0-test" --nsweeps=10

Performance

  • VoxelNet
Name db_sample subset mAP NDS
VoxelNet True True 45.33 58.12
SequenceVoxelNet True True