Introduction

This implementation of VoxelNet is based on the work of @jeasinema and @qianguih. It uses the TensorFlow framework and the KITTI dataset. Several bug fixes make it run on both Windows and Linux systems. The code is reworked for better functionality:

• it is not longer possible to show image visualizations while training
• usage of the KITTI testing dataset is now possible without any labels

Dependencies

• python3 (tested on 3.6.8)
• tensorflow
• numpy
• opencv
• shapely
• numba
• easydict

KITTI Data

Download the KITTI dataset. You need the following data:

• Velodyne point clouds
• Left color images of object data set
• Training labels of object data set
• Camera calibration matrices of object data set (not needed for training and validation data)

File Structure

Split the KITTI dataset for training into training and validation.

└── data    <-- KITTI data directory 
    └── object 
        ├── training       <-- training data
        |   ├── image_2   
        |   ├── label_2   
        |   └── velodyne  
        ├── validation     <-- validation data
        |   ├── image_2   
        |   ├── label_2   
        |   └── velodyne 
        └── testing        <-- testing data
            ├── image_2   
            ├── calib   
            └── velodyne

config.py must be updated if the data folder is not located in root.

Compilation

Compile box_overlaps.pyx with running setup.py.

python setup.py build_ext --inplace

Usage

0. Update config.py if you want to use multiple GPUs.
1. Run train.py

python train.py

2. Run test.py

python test.py

See the default parameter settings in train.py and test.py. Set parameters in command line if other settings are needed.

python train.py --lr 0.01

python test.py --vis False

Predict single sample by running test_single.py. Set data tag with command line parameter -t or --data-tag (default 000000).

python test_single.py -t 000001