Towards Efficient Semi-Supervised Semantic Segmentation for Solid-State LiDAR Point Clouds
Mardanjan Abla, Eksan Firkat, Bangquan Xie, Eliyas Suleyman, Jiazhan Gao, Bin Zhu, Askar Hamdulla
AI summary
Problem
Existing semi-supervised LiDAR segmentation methods are tailored for mechanical spinning LiDAR and fail to generalize to solid-state LiDAR due to differing scanning patterns and dense point cloud distributions, leading to unreliable pseudo-labels and poor performance in low-label regimes.
Approach
The authors propose SSLiMix, which replaces beam-based mixing with 2D grid partitioning and checkerboard-style data augmentation to preserve spatial consistency in dense solid-state LiDAR point clouds. They also introduce HCAP, a hierarchical pseudo-labeling mechanism that dynamically adjusts confidence thresholds per semantic class to filter noise and leverage uncertain labels.
Key results
- Novel 2D grid-based checkerboard mixing strategy tailored for solid-state LiDAR
- Hierarchical confidence-aware pseudo-labeling with class-wise adaptive thresholds
- 11.3% mIoU improvement over fully-supervised baseline using only 1% labeled data on PandaSet
- Consistently outperforms state-of-the-art semi-supervised LiDAR segmentation methods across varying annotation ratios
Why it matters
Reduces annotation costs for autonomous driving by enabling highly accurate 3D scene understanding with minimal labeled data, providing a new benchmark for solid-state LiDAR processing.
Abstract
LiDAR-based 3D semantic segmentation is a crit- ical task in autonomous driving, but its scalability is limited by the reliance on large-scale labeled datasets. Semi-supervised learning (SSL) offers a potential solution by leveraging unla- beled data. However, most existing SSL segmentation methods are designed for mechanical spinning LiDAR (MSLR) and fail to generalize well to solid-state LiDAR (SSLR) due to different scanning patterns and point cloud distributions. To address this challenge, we propose SSLiMix, a novel semi-supervised segmentation method with checkerboard mixing for solid-state LiDAR. Unlike prior MSLR-oriented methods, SSLiMix em- ploys 2D grid partitioning with checkerboard mixing to adapt to SSLR’s dense and uniform point clouds, thereby preserving spatial consistency even when beam-based augmentations fail. Additionally, we introduce a hierarchical confidence-aware pseudo-labeling mechanism (HCAP), which classifies pseudo- labels by confidence and applies targeted processing to enhance pseudo-label reliability. Experiments on the PandaSet dataset show that SSLiMix improves mIoU by 11.3% over the fully- supervised baseline using only 1% labeled data, demonstrating its effectiveness in low-label regimes and providing a strong benchmark for semi-supervised SSLR segmentation.