CNN-based Lidar Point Cloud De-Noising in Adverse Weather

Key Knowledgeable:

• Autolabeling for Noise Caused by Rain or Fog
  Judging the distance change of each pixel according to the range image to determine whether the pixel is noise caused by rain, fog and water droplets. Label the data:
  

Why the noise can be marked like this:
The author's understanding is that the rain and fog noise is often the reflection produced by the water droplets in the air, which is described in detail in the article "Fog Simulation on Real LiDAR Point Clouds".When the camera distance is at a certain value, the reflection intensity reaches its peak value. Therefore, there are often circular noise reflections next to the camera in rainy and foggy days. Therefore, when the scanning vehicle is moving, the distance from the non-noise point to the camera will change.There are dense water droplets in the middle, and there are always water droplets with the camera as the center of the circle reflecting noise, so the distance of the noise will not change much, so it can be preliminarily judged whether it is noise by judging whether the point has moved a threshold in each frame.
Of course, this is my personal opinion. The text does not describe this part too much. Corrections are welcome

• Network Architecture
  Annotated, augmented data for denoising with WeatherNet for semantic segmentation:
  
  
• Data Augmentation
  Simulation rain and fog weather augmentation training set data:
• Experiments
  Quantitative denoising results for each weather:
  
  
  Experiment1, 2, 3 represent three training sets respectivelyThe experimental results below:
  Visualize qualitative results:
  You can see that data enhancement and WeatherNet are denoising (semantic segmentation of rain and fog) work