[Paper Review] LaserMix for Semi-Supervised LiDAR Semantic Segmentation

Lingdong Kong, Jiawei Ren+2 more|arXiv (Cornell University)|Jun 30, 2022

Advanced Neural Network Applications4 citations

TL;DR

LaserMix proposes a novel semi-supervised learning framework for LiDAR semantic segmentation that leverages spatial priors by mixing laser beams from different scans to improve model generalization. It achieves state-of-the-art performance with 2× to 5× fewer labels, improving mIoU by up to 5.7% on nuScenes and 3.5% on SemanticKITTI, while being representation-agnostic and statistically grounded.

ABSTRACT

Densely annotating LiDAR point clouds is costly, which restrains the scalability of fully-supervised learning methods. In this work, we study the underexplored semi-supervised learning (SSL) in LiDAR segmentation. Our core idea is to leverage the strong spatial cues of LiDAR point clouds to better exploit unlabeled data. We propose LaserMix to mix laser beams from different LiDAR scans, and then encourage the model to make consistent and confident predictions before and after mixing. Our framework has three appealing properties: 1) Generic: LaserMix is agnostic to LiDAR representations (e.g., range view and voxel), and hence our SSL framework can be universally applied. 2) Statistically grounded: We provide a detailed analysis to theoretically explain the applicability of the proposed framework. 3) Effective: Comprehensive experimental analysis on popular LiDAR segmentation datasets (nuScenes, SemanticKITTI, and ScribbleKITTI) demonstrates our effectiveness and superiority. Notably, we achieve competitive results over fully-supervised counterparts with 2x to 5x fewer labels and improve the supervised-only baseline significantly by 10.8% on average. We hope this concise yet high-performing framework could facilitate future research in semi-supervised LiDAR segmentation. Code is publicly available.

Motivation & Objective

• Address the high cost of dense LiDAR annotation by enabling semi-supervised learning with limited labeled data.
• Exploit the strong spatial structure inherent in LiDAR point clouds—particularly the patterned distribution of laser beams across distances.
• Develop a representation-agnostic SSL method that works across range view, voxel, and point-based LiDAR representations.
• Improve model robustness and performance by enforcing consistent and confident predictions before and after beam mixing.
• Demonstrate effectiveness across low-data (10–50% labels) and high-data (full labels) regimes on benchmark datasets.

Proposed method

• Partition LiDAR scans into low-variation spatial regions based on beam-level distribution patterns, exploiting the spatial prior of laser beam distribution.
• Propose LaserMix: a beam-level mixing strategy that combines corresponding spatial regions from two different LiDAR scans in an intertwined manner.
• Apply a teacher-student training pipeline where the model is encouraged to make consistent and confident predictions on the same spatial region before and after mixing.
• Use a consistency regularization loss to enforce prediction agreement between original and mixed inputs, improving generalization.
• Operate directly on raw point clouds, making the method compatible with various LiDAR representations (e.g., range view, voxel, point-based) without architectural changes.
• Theoretical analysis supports the statistical validity of the mixing strategy under the assumption of spatial continuity and local homogeneity in LiDAR scans.

Experimental results

Research questions

• RQ1 Can spatial priors in LiDAR point clouds be effectively leveraged to improve semi-supervised learning performance?
• RQ2 Does beam-level mixing across scans lead to better generalization than standard data augmentation or contrastive learning in LiDAR segmentation?
• RQ3 How does LaserMix perform across different label budgets (e.g., 10%, 20%, 50%, 100%) on standard LiDAR benchmarks?
• RQ4 Can the proposed method achieve performance comparable to fully supervised SOTA models with significantly fewer annotations?
• RQ5 Is the framework robust and generalizable across different LiDAR representations (e.g., range view, voxel, point cloud) without retraining?

Key findings

• LaserMix achieves a 5.7% mIoU improvement over the supervised-only baseline on nuScenes using only 20% of the training labels.
• On SemanticKITTI, LaserMix improves mIoU by up to 3.5% compared to prior SoTA methods under the same training regime.
• With just 0.8% labeled data, LaserMix achieves 54.4% mIoU on SemanticKITTI, matching the performance of fully supervised models like PolarNet and RandLA-Net trained on 100% labels.
• The method achieves competitive results with 2× to 5× fewer labels than fully supervised SOTA models, demonstrating strong data efficiency.
• The ablation study confirms that spatial priors are critical: removing beam partitioning or mixing leads to significant performance drops.
• The framework generalizes across representations: it is effective when applied to range view, voxel, and point-based models without modification.