[Paper Review] PolarMix: A General Data Augmentation Technique for LiDAR Point Clouds

Aoran Xiao, Jiaxing Huang+4 more|arXiv (Cornell University)|Jul 30, 2022

Robotics and Sensor-Based LocalizationEngineering38 citations

TL;DR

PolarMix proposes two LiDAR-tailored data augmentation strategies—scene-level azimuth swapping and instance-level rotate-paste—to enrich point clouds while preserving fidelity, improving semantic segmentation and object detection across architectures and datasets, and aiding unsupervised domain adaptation.

ABSTRACT

LiDAR point clouds, which are usually scanned by rotating LiDAR sensors continuously, capture precise geometry of the surrounding environment and are crucial to many autonomous detection and navigation tasks. Though many 3D deep architectures have been developed, efficient collection and annotation of large amounts of point clouds remain one major challenge in the analytic and understanding of point cloud data. This paper presents PolarMix, a point cloud augmentation technique that is simple and generic but can mitigate the data constraint effectively across different perception tasks and scenarios. PolarMix enriches point cloud distributions and preserves point cloud fidelity via two cross-scan augmentation strategies that cut, edit, and mix point clouds along the scanning direction. The first is scene-level swapping which exchanges point cloud sectors of two LiDAR scans that are cut along the azimuth axis. The second is instance-level rotation and paste which crops point instances from one LiDAR scan, rotates them by multiple angles (to create multiple copies), and paste the rotated point instances into other scans. Extensive experiments show that PolarMix achieves superior performance consistently across different perception tasks and scenarios. In addition, it can work as plug-and-play for various 3D deep architectures and also performs well for unsupervised domain adaptation.

Motivation & Objective

• Motivate data augmentation for LiDAR point clouds to address limited annotated data and domain gaps.
• Propose PolarMix as a simple, generic augmentation that respects LiDAR-specific properties such as partial visibility and depth-dependent density.
• Enable improvements across multiple perception tasks (semantic segmentation, object detection) and datasets.
• Demonstrate PolarMix as a plug-and-play module compatible with diverse 3D architectures and unsupervised domain adaptation.

Proposed method

• Define two cross-scan augmentation strategies operating in polar coordinates: scene-level swapping of azimuthal sectors between scans, and instance-level rotation and pasting of cropped semantic instances with multiple azimuth angles.
• Formulate operations in the input space using masks and simple geometric transforms to preserve LiDAR data fidelity.
• Allow concatenation of augmented scans and labels to form training samples without modifying the underlying loss.
• Provide implementation details and hyper-parameter choices (azimuth ranges, rotation angles, and probabilities) for practical use.

Experimental results

Research questions

• RQ1 Can PolarMix improve performance of LiDAR-based semantic segmentation and object detection across different architectures and datasets?
• RQ2 Does PolarMix reduce data requirements and enhance data efficiency for LiDAR learning?
• RQ3 Can PolarMix extend to unsupervised domain adaptation by mixing source and target domain data?
• RQ4 How do the two augmentation components (scene-level swapping and instance-level rotate-paste) contribute to performance gains?
• RQ5 Is PolarMix effective across varied LiDAR sensors and scanning configurations?

Key findings

• PolarMix yields consistent performance gains in semantic segmentation across MinkNet, SPVCNN, RandLA-Net, and Cylinder3D on SemanticKITTI, nuScenes-lidarseg, and SemanticPOSS.
• PolarMix improves data efficiency, achieving similar mIoU with about 75% of the annotated data on SemanticKITTI for two networks.
• PolarMix enhances object detection mAP and nuScenes detection score across PointPillar, Second, and CenterNet.
• PolarMix achieves state-of-the-art results in unsupervised domain adaptation for SynLiDAR to SemanticKITTI and SemanticPOSS.
• Ablation shows scene-level swapping and instance-level rotate-paste provide complementary gains, with the full PolarMix giving the largest improvement.