Lee Seungho ; Lee Hwijeong ; Shim Hyunjung

Abstract

We address the challenges of the semi-supervised LiDAR segmentation (SSLS)problem, particularly in low-budget scenarios. The two main issues inlow-budget SSLS are the poor-quality pseudo-labels for unlabeled data, and theperformance drops due to the significant imbalance between ground-truth andpseudo-labels. This imbalance leads to a vicious training cycle. To overcomethese challenges, we leverage the spatio-temporal prior by recognizing thesubstantial overlap between temporally adjacent LiDAR scans. We propose aproximity-based label estimation, which generates highly accurate pseudo-labelsfor unlabeled data by utilizing semantic consistency with adjacent labeleddata. Additionally, we enhance this method by progressively expanding thepseudo-labels from the nearest unlabeled scans, which helps significantlyreduce errors linked to dynamic classes. Additionally, we employ a dual-branchstructure to mitigate performance degradation caused by data imbalance.Experimental results demonstrate remarkable performance in low-budget settings(i.e., <= 5%) and meaningful improvements in normal budget settings (i.e., 5 -50%). Finally, our method has achieved new state-of-the-art results onSemanticKITTI and nuScenes in semi-supervised LiDAR segmentation. With only 5%labeled data, it offers competitive results against fully-supervisedcounterparts. Moreover, it surpasses the performance of the previousstate-of-the-art at 100% labeled data (75.2%) using only 20% of labeled data(76.0%) on nuScenes. The code is available on https://github.com/halbielee/PLE.