# PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation

**Authors:** Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, Hassan Foroosh  
**arXiv:** [2003.14032](https://arxiv.org/abs/2003.14032)  
**Date:** 2020-03-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2003.14032v2/

## Key points

- Growing demand for fine-grained perception capabilities in autonomous driving systems
- Surge in research activity to enhance accuracy and efficiency of autonomous vehicles
- Challenges include near-real-time latency, long-tailed distribution of LiDAR points, and increasing complexity of semantic classes
- Introduction of PolarNet algorithm by researchers led by Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh
- PolarNet utilizes a polar bird's-eye-view representation to balance points across grid cells in a polar coordinate system
- Experimental results show improved mean Intersection over Union (mIoU) scores with near real-time throughput
- Potential of PolarNet to address challenges in online LiDAR point cloud semantic segmentation and advance autonomous driving systems

## Summary

In the realm of autonomous driving systems, there is a growing demand for fine-grained perception capabilities. This is particularly evident in the area of online semantic segmentation of single-scan LiDAR data. The surge in research activity has been fueled by the need to enhance the accuracy and efficiency of autonomous vehicles. However, despite the availability of emerging datasets and technological advancements, several challenges persist in this domain. One major challenge is the requirement for near-real-time latency while operating within hardware constraints. This necessitates the development of segmentation algorithms that can deliver rapid results without compromising on accuracy. Additionally, the uneven or long-tailed distribution of LiDAR points across space poses a significant obstacle to achieving precise semantic segmentation. Furthermore, as the number of semantic classes continues to expand, there is a pressing need for solutions that can effectively handle this increasing complexity. In response to these challenges, a team of researchers led by Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong and Hassan Foroosh have introduced a novel LiDAR-specific segmentation algorithm known as PolarNet. Unlike traditional spherical or bird's-eye-view projections, PolarNet utilizes a polar bird's-eye-view representation that strategically balances points across grid cells in a polar coordinate system. This innovative approach indirectly aligns the attention of segmentation networks with the long-tailed distribution of points along the radial axis. Through extensive experimentation with real urban LiDAR single scans across three diverse segmentation datasets, the researchers found that their encoding scheme significantly improved mean Intersection over Union (mIoU) scores while maintaining near real-time throughput. The success of PolarNet underscores its potential to address key challenges in online LiDAR point cloud semantic segmentation and advance the capabilities of autonomous driving systems. The research findings have been accepted by CVPR 2020 and are accompanied by code available on GitHub for further exploration and implementation.