[Paper Review] PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation

Yang Zhang, Zixiang Zhou+5 more|arXiv (Cornell University)|Mar 31, 2020

Advanced Neural Network Applications39 references31 citations

TL;DR

PolarNet introduces a polar BEV representation with ring-convolution CNNs and a learnable per-grid feature extractor to perform online LiDAR point cloud segmentation with higher mIoU and lower cost than prior methods.

ABSTRACT

The need for fine-grained perception in autonomous driving systems has resulted in recently increased research on online semantic segmentation of single-scan LiDAR. Despite the emerging datasets and technological advancements, it remains challenging due to three reasons: (1) the need for near-real-time latency with limited hardware; (2) uneven or even long-tailed distribution of LiDAR points across space; and (3) an increasing number of extremely fine-grained semantic classes. In an attempt to jointly tackle all the aforementioned challenges, we propose a new LiDAR-specific, nearest-neighbor-free segmentation algorithm - PolarNet. Instead of using common spherical or bird's-eye-view projection, our polar bird's-eye-view representation balances the points across grid cells in a polar coordinate system, indirectly aligning a segmentation network's attention with the long-tailed distribution of the points along the radial axis. We find that our encoding scheme greatly increases the mIoU in three drastically different segmentation datasets of real urban LiDAR single scans while retaining near real-time throughput.

Motivation & Objective

• Address the challenge of online fine-grained LiDAR segmentation with near-real-time latency, imbalanced point distributions, and many fine-grained classes.
• Propose a LiDAR-specific input representation that balances grid point distribution.
• Develop a learnable per-grid feature extractor and a ring-convolution CNN that operates on a polar BEV grid.
• Demonstrate performance gains across multiple urban LiDAR datasets with reduced computational cost.

Proposed method

• Quantize LiDAR points into a polar BEV grid that balances points across cells along radial and angular axes.
• Use a learnable simplified PointNet per grid cell to produce fixed-length local features.
• Attach a ring-convolution 2D CNN that respects polar grid connectivity by wrapping the angular axis, enabling end-to-end processing.
• Decode the polar-grid predictions back to Cartesian point-domain labels for evaluation.
• Train end-to-end on SemanticKITTI, A2D2, and Paris-Lille-3D using voxel-based segmentation loss.
• Compare against SqueezeSeg, SqueezeSegv2, PointNet, RangeNet++, and Unet variants with Cartesian BEV and spherical projections.

Experimental results

Research questions

• RQ1 Does a polar BEV grid improve point distribution balance and per-voxel label purity compared to Cartesian BEV or spherical projections?
• RQ2 Can a learnable per-grid representation combined with ring-convolutions improve segmentation accuracy (mIoU) while reducing MACs and latency?
• RQ3 Are performance gains consistent across diverse LiDAR datasets (SemanticKITTI, A2D2, Paris-Lille-3D)?
• RQ4 How does PolarNet compare to state-of-the-art online LiDAR segmentation methods in terms of accuracy and efficiency?

Key findings

• PolarNet outperforms state-of-the-art methods on SemanticKITTI, A2D2, and Paris-Lille-3D while using roughly one third of the parameters and MACs.
• On SemanticKITTI, PolarNet achieves mIoU of 54.3% and accuracy of 90.0% with 14M parameters and 105B MACs.
• On A2D2, PolarNet achieves a mIoU of 23.9% and per-class IoU improvements across several classes, with 60B parameters and 0.031s latency.
• On Paris-Lille-3D, PolarNet attains a mIoU of 43.7% and accuracy of 87.5%, outperforming DarkNet53 and Cartesian BEV baselines.
• Polar BEV achieves more balanced points per grid cell than Cartesian BEV (polar: 0.7±1.4 vs Cartesian: 0.7±3.2 points per cell).
• A majority of classes see IoU improvements, especially irregularly distributed or distant classes, due to the polar representation and ring-convolution design.