PolarMix: A General Data Augmentation Technique for LiDAR Point Clouds

1 Introduction

• Collecting and annotating the point clouds needed for reliable DNN training is laborious, time-consuming, and poorly scalable across tasks and domains.The paper identifies this as a major constraint on LiDAR point-cloud analytics and understanding.Source: §1 · p. 1
• Global scaling, flipping, and rotation do not augment local structures or model relationships across scans, and local image cut-and-mix methods overlook LiDAR scanning properties.The paper notes that LiDAR-specific properties arise from continuous 360-degree sweeping.Source: §1 · pp. 1–2
• Partial visibility and depth-dependent point density are LiDAR-specific properties that effective augmentation must accommodate to preserve fidelity and usefulness.Only object sides facing the sensor are captured, and point density varies with depth.Source: §1 · p. 2
• PolarMix cuts, edits, and mixes points along the azimuth scanning direction through scene-level sector swapping and instance-level rotation and paste.These cross-scan strategies are designed to enrich point distributions while maintaining fidelity.Source: §1 · p. 2
• PolarMix is presented as applicable across architectures, tasks, datasets, and domains, including plug-and-play use and unsupervised domain adaptation.The paper also reports state-of-the-art performance for unsupervised domain adaptation.Source: §1 · p. 2

2 Related Works

• Image augmentation includes global transformations and local mix operations, including methods that cut and paste object instances for instance segmentation.Examples include random cropping, scaling, erasing, color jittering, mixup, and CutMix.Source: §2 · p. 3
• Point-cloud augmentation includes global 3D transformations and local methods, but several cited approaches focus on object-level augmentation unsuitable for autonomous-driving scenes.Local approaches include shape-wise and point-wise transformations, local-area augmentation, and weighted neighborhood transformations.Source: §2 · p. 3
• Earlier point-cloud mixing includes object-level methods and scene-level methods constrained to particular tasks, while PolarMix performs both object- and scene-level augmentation aligned with LiDAR properties.The paper identifies GT-Aug as requiring 3D boxes for object cutting and Mix3D as an out-of-context augmentation for semantic segmentation.Source: §2 · p. 3

3 PolarMix

• PolarMix combines scene-level swapping and instance-level rotate-paste to mix two LiDAR scans and their labels.The generated scan and labels are used for network training with the original loss function.Source: §3 · p. 3, Figure 2 · p. 4
• Scene-level swapping replaces a sector cut from scan A with the corresponding azimuth sector from scan B, with the cut angles within the sensor's horizontal field of view.The sector is defined by azimuth angles α and β.Source: §3 · p. 4
• Instance-level augmentation crops selected semantic classes from scan B, rotates the instances about the z-axis at multiple azimuth angles, and pastes the copies into scan A.The class mask selects instances, and the angle list determines their rotations.Source: §3 · p. 4
• Polar coordinates align cutting, rotating, and mixing with LiDAR scanning, helping preserve partial visibility and depth-dependent density while keeping computational overhead negligible.The described operations use point slicing and concatenation or dot products followed by concatenation; input-space augmentation is compatible with different architectures.Source: §3 · pp. 4–5
• The pipeline takes two scans and labels, a class list, an angle list, and a swapping azimuth range, then outputs a new scan and labels.Its algorithm initializes from scan A, conditionally performs sector swapping, and then conditionally adds rotated instances from scan B.Source: §3 · p. 5
• For unsupervised domain adaptation, PolarMix mixes labeled source scans with target scans carrying pseudo-labels produced by self-training.A supervised source-domain model predicts pseudo-labels for unlabeled target-domain data before the cross-domain mixing step.Source: §3 · p. 5

4.1.1 Experimental Settings

• The segmentation experiments use SemanticKITTI, nuScenes-lidarseg, and SemanticPOSS, four network architectures, and mIoU; RandLA-Net and Cylinder3D use a uniform 10% SemanticKITTI subsample.The reported settings also specify different rotation-angle choices by dataset.Source: §4.1.1 · pp. 5–6

4.1.2 Results

• PolarMix achieves the best semantic segmentation across MinkNet and SPVCNN on three benchmark datasets, with gains reported over competing augmentations.The experiments also compare conventional global augmentation and other local augmentation methods.Source: §4.1.2 · pp. 6–7, Table 2 · p. 7
• Both MinkNet and SPVCNN achieve similar segmentation mIoU using 75% of SemanticKITTI with PolarMix as using 100% without it.The paper reports this as saving around 25% of training-data collection and annotation effort.Source: §4.1.2 · p. 7, Figure 3 · p. 7
• With the same 10% of SemanticKITTI training data, PolarMix improves segmentation for both RandLA-Net and Cylinder3D.These architectures extend the experiments beyond MinkNet and SPVCNN.Source: §4.1.1 · p. 6, §4.1.2 · p. 7, Table 3 · p. 8

4.2 PolarMix helps learn effective representations for object detection

• Across PointPillar, Second, and CenterNet, incorporating PolarMix consistently improves both mAP and NDS on nuScenes.The authors attribute the generalization across tasks to cut-edit-mix augmentation in the input space.Source: Table 4 · p. 8, §4.2 · p. 8

4.3 PolarMix helps reduce domain gap

• For SynLiDAR-to-SemanticKITTI and SynLiDAR-to-SemanticPOSS adaptation, PolarMix achieves state-of-the-art mIoU in both setups.The procedure mixes labeled source data with pseudo-labeled target data for further training.Source: §4.3 · pp. 8–9, Table 5 · p. 9

4.4 Ablation study

• +5.9% mIoU is reported when scene-level swapping and full instance-level cut-and-mix are combined in the SPVCNN ablation.Scene-level swapping alone improves mIoU by 1.4%; basic instance cut-and-mix by 2.0%; and full instance cut-and-mix by 4.3%, each over the conventional global-augmentation baseline of 48.9%.Source: §4.4 · p. 9, Table 6 · p. 9

• PolarMix outperforms alternatives in semantic segmentation and object detection across multiple deep frameworks and public datasets.The conclusion also reports state-of-the-art performance on multiple synthetic-to-real LiDAR segmentation benchmarks.Source: §5 · p. 9
• Mixing labelled source points with unlabeled target points extends PolarMix to unsupervised domain adaptation, where it achieves state-of-the-art results on multiple synthetic-to-real segmentation benchmarks.Source: §5 · p. 9
• PolarMix creates training data through azimuth-based scene-level swapping and rotating-pasting instances from one scan into another.Source: §5 · p. 9
• The authors hope PolarMix will encourage future research to provide deeper insights into data augmentation for deep point-cloud learning.Source: §5 · p. 10