UltraLiDAR

Unofficial implementation of UltraLiDAR. We add the code to support training and generation on the nuScenes and Waymo datasets.

Installation instructions

Following https://mmdetection3d.readthedocs.io/en/v1.0.0rc1/getting_started.html

a. Create a conda virtual environment and activate it.

conda create -n ultralidar python=3.8 -y
conda activate ultralidar

b. Install PyTorch and torchvision following the official instructions.

pip install torch==1.10.0+cu113 torchvision==0.11.1+cu113 torchaudio==0.10.0+cu113 --extra-index-url https://download.pytorch.org/whl/cu113
# Recommended torch==1.10

c. Install mmcv-full, mmdet and mmseg.

pip3 install openmim
mim install mmcv-full==1.5.0
mim install mmdet==2.28.2
mim install mmsegmentation==0.30.0

d. Install mmdet3d from source code and other packages.

git clone https://github.com/open-mmlab/mmdetection3d.git
cd mmdetection3d
git checkout v1.0.0rc1 # Other versions may not be compatible.
pip install -v -e .

e. Clone UltraLIDAR.

git clone git@github.com:myc634/UltraLiDAR_nusc_waymo.git
cd UltraLiDAR_nusc_waymo
pip install -r requirements.txt

Preparing Dataset

Please follow the official MMDetection3D dataset preparation guidance

The final data structure should be:

UltraLiDAR_nusc_waymo
├── datasets
│   ├── kitti-360
│   │── nuScenes
│   │── waymo

For Kitti360 dataset, please run

python tools/data_converter/kitti360_converter.py

For nuScenes dataset, please run

python tools/data_converter/nuscenes_converter.py --data-root datasets/nuScenes/

to generate the pkl file which fits the MMDetection3D framework.

For Waymo Open Dataset, please run: cd mmdetection3d and follow the official instruction of MMDetection3D Data Preparation pipeline to convert waymo format into kitti format for training

Training

For Stage 1 training:

./tools/dist_train.sh configs/ultralidar_kitti360.py 8

For Stage 2 training, please select the best result in Stage 1, copy the dir and paste it on configs/ultralidar_kitti360_gene.py: 238 and run:

./tools/dist_train.sh configs/ultralidar_kitti360_gene.py 8

Eval

Step0:

./tools/dist_test.sh configs/ultralidar_kitti360_static_blank_code.py ${PATH_TO_WEIGHTS} --eval "mIoU"

Run this command to calculate the blank code, this will generate a blank_code.pkl file which contains 20 blank codes.

Step1:

./tools/dist_test.sh configs/ultralidar_kitti360_gene.py ${PATH_TO_WEIGHTS} --eval "mIoU"

After running this command, there will be a file named ultralidar_samples containing 2000 generated samples with .ply format.

Step2:

First:

export KITTI360_DATASET=datasets/kitti-360

For evaluation MMD, please run:

python mmd.py

For evaluation JSD, please run:

python jsd.py

Here, we provide a set of 2k and 10k generated LiIDAR point cloud samples, utilizing the weightes from our implementated UltraLiDAR model.

	MMD	JSD
KITTI360	2.25e-4	0.0989