An anchor-free single-stage LiDAR-based 3D object detector purely based on the range view representation.

This is the official implementation of RangeDet (ICCV 2021).

Update the KITTI results, which is also reported at our CVF access version.

RangeDet:In Defense of Range View for LiDAR-based 3D Object Detection [arXiv] [CVF] [Supplementary materials]

Lue Fan*, Xuan Xiong*, Feng Wang, Naiyan Wang , Zhaoxiang Zhang

*Equal contribution

We propose a pure range-view-based framework - RangeDet, which includes range conditioned pyramid, Meta-Kernel, weighted non-maximum suppression, and a iou-awared classification loss to overcome a couple of challenges in range-view-based 3D detector and achieves comparable performance with state-of-the-art multi-view-based detectors.

datasets
    create_range_image_roidb.py
rangedet
    symbol
        backbone
            dla_backbone.py
            meta_kernel.py
        head
            builder.py
            loss.py
    core
        detection_metric.py
        input.py
        util_func.py
config 
    ...
scripts 
     distributed_train_and_test.sh
tools
    train.py
    test.py
    create_prediction_bin_3d.py

Some extra points and reminders：

• operator_cxx/contrib contain MXNet's Custom Op written in C++, which need to be compiled with MXNxt.
• operator_cxx/src_cxx contain wnms Op written in C++.
• pybind_cxx contain bbox assigner function written in C++.
• operator_py contain MXNet's Custom Op written in Python.
• mxnext is a foundational library, we have modified some functions in this library.
• utils contain modular python files, most of them copy from SimpleDet.

KITTI results

Experiments were performed using uniformly sample 25% training data (∼40k frames)

We will further add the experiment results here, for now please refer to our overall performance on Waymo leaderboard in this link .

Installing and compiling rangedet in the following command.

pip install -v -e .

We will further add requirements and installation instructions on how to install Horovod , OpenMPI and MXNet.

To train on the Waymo Open Dataset:

• Download the Waymo Open Dataset
• Unzip all *.tar files and merge them into three directories
• Make sure to put the files as the following structure:

datasets
    waymo
        training/*.tfrecord
        validation/*.tfrecord
        testing/*.tfrecord

• Next you need to extract the data from tfrecord; We provide the python script in datasets/create_range_image_roidb.py
• Store the training data in npz format and the label information in pickle format respectively

To run the python script datasets/create_range_image_roidb.py, do the following settings in the file:

python datasets/create_range_image_roidb.py --data_path  "**/datasets/waymo" --save_path "**/datasets/waymo-range" --dataset-split 'training' --save_dir 'npz_data'

• After that, make sure to put the files as the following structure:

datasets
    waymo-range
        training/*.roidb
        validation/*.roidb
        testing/*.roidb
        npz_data/**/*.npz

• Note: you need set DatasetParam.data_root = 'path/to/datasets/waymo-range' in the config/*.py file.

Most of the configuration files that we provide assume that we are running on 8 GPUs. In order to be able to run it on fewer GPUs, there are a few possibilities:

1. Run the following

python tools/train.py --config "/path/to/config/file.py"

./scripts/horovodrun.sh 8 "python tools/train.py --config /path/to/config/file.py"

Note you don't need to change any config between single GPU training and multi-GPU training.

To enable, just do Single-GPU or Multi-GPU training and set General.fp16 = True in the config/*.py file.

We use Hough Transformation to obtain scanning parameters following RCD. Based on the scanning parameters, it easy to create high-quality range images.

To create range images on KITTI, just simply run:

python ./datasets/create_range_image_in_kitti.py --source-dir your_source_kitti_folder --target-dir your_save_directory

We assume that the source KITTI data is in MMDetection3D format. In other words, the folder your_source_kitti_path is supposed to contains kitti_infos_trainval.pkl and kitti_infos_test.pkl. It is easy to create these files following the instruction of MMDetection3D.

You can test your model on single gpus. Here is an example:

python tools/test.py --config "/path/to/config/file.py"

You need to create a bin file so that you can use the official waymo evaluation metrics code to get the evaluation results.

python tools/create_prediction_bin_3d.py --p "/path/to/experiments" --c "config_file" -e num_epoch --s "/path/to/save_bin_file"

This project is based on SimpleDet. Thanks Yuntao Chen and his colleagues for their great work!

Please consider citing our paper in your publications if it helps your research.

@InProceedings{Fan_2021_ICCV,
    author    = {Fan, Lue and Xiong, Xuan and Wang, Feng and Wang, Naiyan and Zhang, ZhaoXiang},
    title     = {RangeDet: In Defense of Range View for LiDAR-Based 3D Object Detection},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
}

RangeDet is released under the Apache license. See LICENSE for additional details.