Inofficial PyTorch implementation of CNN-based Lidar Point Cloud De-Noising in Adverse Weather (Heinzler et al., 2019). pytorch-LiLaNet's repo is used as base code for this repo and necessary modifications are performed following the instructions in the original paper.
The Autolabeling process is currently not used. For better convergence we add batch normalization after each convolutional layer.
Information: Click here for registration and download.
| Clear | Rainy | Foggy | |
|---|---|---|---|
| WeatherNet | 88.1 | 83.1 | 70.5 |
- Install PyTorch (pytorch.org)
pip install -r requirements.txt- Download the DENSE CNN denoising dataset
Train model: train_dense.py trains the model on the given dataset/directory
Important: The dataset-dir must contain the train_01, test_01 and the val_01 folder.
python train_dense.pyTest model: test.py opens a saved model, classifies points in an HDF5 file, and saves the resulting classified points.
ROS Pipeline: ros_test.py runs a ros pipeline subscribing to a PointCloud2 topic and classifies/denoises the points beforepublishing a new PointCloud2 topic.