LiSnowNet: Real-time Snow Removal for LiDAR Point Cloud

This is the offical implementation of LiSnowNet: Real-time Snow Removal for LiDAR Point Clouds.

• Results for CADC

• Results for WADS

Requirements

• Ubuntu 20.04+
• NVIDIA driver >= 515
• Docker with the post-installation steps
• NVIDIA Container Runtime

Installation

• (Only for WSL2) Install x11-xserver-utils

  $ sudo apt install -y x11-xserver-utils
  

• Build the docker image

  $ docker build --tag lisnownet -f docker/Dockerfile .
  

• Launch a container

  $ DATA_PATH=/path/to/datasets     # the dataset path to be mounted to the container
  $ ./docker/run.sh       # use all GPUs
  $ ./docker/run.sh 0     # use GPU #0
  $ ./docker/run.sh 2,3   # use GPU #2 and #3
  

Datasets

Download the Canadian Adverse Driving Conditions (CADC) Dataset and the Winter Adverse Driving dataSet (WADS), and create symlinks to them under the data folder:

./data
├── cadcd
|   └── {DATE}/{DRIVE_ID}/raw/lidar_points_corrected/data/{FRAME_ID}.bin
└── wads
    └── {DRIVE_ID}
        ├── labels/{FRAME_ID}.label
        └── velodyne/{FRAME_ID}.bin

Train

To train the model, run

$ ./train.py [--batch_size BATCH_SIZE] [--dataset DATASET] [--alpha ALPHA] [--tag TAG] [...]

For example:

$ ./train.py --dataset cadc --tag cadc_alpha=5.0 --lr_decay -1 --alpha 5.0

Evaluate

$ ./eval.py [--batch_size BATCH_SIZE] [--dataset DATASET] [--tag TAG] [--threshold THRESHOLD] [...]

To reproduce the results using pretrained weights, run

$ ./eval.py --tag wads_alpha=5.5 --batch_size 8 --dataset wads --threshold 8e-3
$ ./eval.py --tag cadc_alpha=5.0 --batch_size 8 --dataset cadc --threshold 1.2e-2

Citation

@INPROCEEDINGS{9982248,
    author={Yu, Ming-Yuan and Vasudevan, Ram and Johnson-Roberson, Matthew},
    booktitle={2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
    title={LiSnowNet: Real-time Snow Removal for LiDAR Point Clouds},
    year={2022},
    volume={},
    number={},
    pages={6820-6826},
    doi={10.1109/IROS47612.2022.9982248}}