This is an unofficial inplementation of VoxelNet in TensorFlow.
Switch branches/tags
Nothing to show
Clone or download
Latest commit b74823d Apr 14, 2018


This is an unofficial inplementation of VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection in TensorFlow. A large part of this project is based on the work here. Thanks to @jeasinema. This work is a modified version with bugs fixed and better experimental settings to chase the results reported in the paper (still ongoing).


  • python3.5+
  • TensorFlow (tested on 1.4.1)
  • opencv
  • shapely
  • numba
  • easydict


  1. Clone this repository.
  2. Compile the Cython module
$ python3 build_ext --inplace
  1. Compile the evaluation code
$ cd kitti_eval
$ g++ -o evaluate_object_3d_offline evaluate_object_3d_offline.cpp
  1. grant the execution permission to evaluation script
$ cd kitti_eval
$ chmod +x

Data Preparation

  1. Download the 3D KITTI detection dataset from here. Data to download include:

    • Velodyne point clouds (29 GB): input data to VoxelNet
    • Training labels of object data set (5 MB): input label to VoxelNet
    • Camera calibration matrices of object data set (16 MB): for visualization of predictions
    • Left color images of object data set (12 GB): for visualization of predictions
  2. In this project, we use the cropped point cloud data for training and validation. Point clouds outside the image coordinates are removed. Update the directories in data/ and run data/ to generate cropped data. Note that cropped point cloud data will overwrite raw point cloud data.

  3. Split the training set into training and validation set according to the protocol here. And rearrange the folders to have the following structure:

       ├── training   <-- training data
       |   ├── image_2
       |   ├── label_2
       |   └── velodyne
       └── validation  <--- evaluation data
       |   ├── image_2
       |   ├── label_2
       |   └── velodyne
  1. Update the dataset directory in and kitti_eval/


  1. Specify the GPUs to use in
  2. run with desired hyper-parameters to start training:
$ python3 --alpha 1 --beta 10

Note that the hyper-parameter settings introduced in the paper are not able to produce high quality results. So, a different setting is specified here.

Training on two Nvidia 1080 Ti GPUs takes around 3 days (160 epochs as reported in the paper). During training, training statistics are recorded in log/default, which can be monitored by tensorboard. And models are saved in save_model/default. Intermediate validation results will be dumped into the folder predictions/XXX/data with XXX as the epoch number. And metrics will be calculated and saved in predictions/XXX/log. If the --vis flag is set to be True, visualizations of intermediate results will be dumped in the folder predictions/XXX/vis.

  1. When the training is done, executing will generate the learning curve.
$ python3 predictions
  1. There is a pre-trained model for car in save_model/pre_trained_car.


  1. run -n default to produce final predictions on the validation set after training is done. Change -n flag to pre_trained_car will start testing for the pre-trained model (only car model provided for now).
$ python3

results will be dumped into predictions/data. Set the --vis flag to True if dumping visualizations and they will be saved into predictions/vis.

  1. run the following command to measure quantitative performances of predictions:
$ ./kitti_eval/evaluate_object_3d_offline [DATA_DIR]/validation/label_2 ./predictions


The current implementation and training scheme are able to produce results in the tables below.

Bird's eye view detection performance: AP on KITTI validation set
Car Easy Moderate Hard
Reported 89.60 84.81 78.57
Reproduced 85.41 83.16 77.10
3D detection performance: AP on KITTI validation set
Car Easy Moderate Hard
Reported 81.97 65.46 62.85
Reproduced 53.43 48.78 48.06


  • improve the performances
  • reproduce results for Pedestrian and Cyclist
  • fix the deadlock problem in multi-thread processing in training
  • fix the infinite loop problem in
  • replace averaged calibration matrices with correct ones