# LiDAR Snowfall Simulation for Robust 3D Object Detection

**Authors:** Martin Hahner, Christos Sakaridis, Mario Bijelic, Felix Heide, Fisher Yu, Dengxin Dai, Luc Van Gool  
**arXiv:** [2203.15118](https://arxiv.org/abs/2203.15118)  
**Date:** 2022-03-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2203.15118v2/

## Key points

- The paper addresses the challenge of detecting and localizing objects in adverse weather conditions, specifically under snowfall.
- The authors propose a physically based simulation method to replicate the effect of snowfall on LiDAR point clouds captured during clear weather.
- The simulation involves sampling snow particles in 2D space for each LiDAR line and modifying the measurements of each LiDAR beam accordingly.
- Ground wetness on LiDAR point clouds is also simulated due to snowfall.
- Partially synthetic snowy LiDAR data is generated using this simulation technique for training robust 3D object detection models.
- Extensive evaluation shows that the proposed simulation consistently improves performance on real snowy datasets compared to clear-weather baselines and competing simulation approaches.
- The research provides a valuable contribution by addressing the challenge of training robust object detection models under adverse weather conditions such as snowfall.
- The proposed physically based simulation method enhances the performance of LiDAR-based systems in accurately detecting and classifying objects even during snowfall.
- Availability of code facilitates reproducibility and encourages further research in this area.

## Summary

The paper titled "LiDAR Snowfall Simulation for Robust 3D Object Detection" addresses the challenge of detecting and localizing objects in adverse weather conditions, specifically under snowfall. This is a crucial task for applications like autonomous driving, where the system needs to accurately identify surrounding traffic agents regardless of weather conditions. To overcome the limitation of collecting and annotating training data for LiDAR-based object detection in snowy conditions, the authors propose a physically based simulation method that replicates the effect of snowfall on real LiDAR point clouds captured during clear weather. The simulation involves sampling snow particles in 2D space for each LiDAR line and using the resulting geometry to modify the measurements of each LiDAR beam accordingly. Additionally, since snowfall often leads to wetness on the ground, they also simulate ground wetness on LiDAR point clouds. By leveraging this simulation technique, partially synthetic snowy LiDAR data is generated which can be used to train 3D object detection models that are robust to snowfall. The authors evaluate their approach extensively using various state-of-the-art 3D object detection methods and demonstrate that their simulation consistently improves performance on real snowy datasets compared to clear-weather baselines and competing simulation approaches without sacrificing performance in clear weather conditions. Overall, this research provides a valuable contribution by addressing the challenge of training robust object detection models under adverse weather conditions such as snowfall. The proposed physically based simulation method enables the generation of synthetic data that can enhance the performance of LiDAR-based systems in detecting and classifying objects accurately even when it is snowing. Furthermore, availability of code facilitates reproducibility and encourages further research in this area.