SELECT: A Submodular Approach for Active LiDAR Semantic Segmentation

TL;DR

The paper addresses the high annotation costs of LiDAR semantic segmentation by proposing SELECT, a three-stage voxel-centric active learning framework that jointly optimizes informativeness, uncertainty, and label diversity. It combines Stage 1 VLSSS for scalable voxel subset selection, Stage 2 VLMUE using MC dropout to quantify voxel uncertainty, and Stage 3 SMPCB to balance class representation via a submodular entropy objective, with explicit formulations for each stage. Across SemanticPOSS, SemanticKITTI, and nuScenes, SELECT achieves consistent, substantial improvements in mIoU under limited labeling budgets, demonstrating strong performance on both common and rare classes. The approach offers a practical path to high-accuracy 3D segmentation in autonomous driving settings, where labeling costs and class imbalance are critical constraints.

Abstract

LiDAR-based semantic segmentation plays a vital role in autonomous driving by enabling detailed understanding of 3D environments. However, annotating LiDAR point clouds is extremely costly and requires assigning semantic labels to millions of points with complex geometric structures. Active Learning (AL) has emerged as a promising approach to reduce labeling costs by querying only the most informative samples. Yet, existing AL methods face critical challenges when applied to large-scale 3D data: outdoor scenes contain an overwhelming number of points and suffer from severe class imbalance, where rare classes have far fewer points than dominant classes. To address these issues, we propose SELECT, a voxel-centric submodular approach tailored for active LiDAR semantic segmentation. Our method targets both scalability problems and class imbalance through three coordinated stages. First, we perform Voxel-Level Submodular Subset Selection, which efficiently identifies representative voxels without pairwise comparisons, ensuring scalability. Second, we estimate Voxel-Level Model Uncertainty using Monte Carlo dropout, aggregating point-wise uncertainties to identify informative voxels. Finally, we introduce Submodular Maximization for Point-Level Class Balancing, which selects a subset of points that enhances label diversity, explicitly mitigating class imbalance. Experiments on SemanticPOSS, SemanticKITTI, and nuScenes benchmarks demonstrate that SELECT achieves superior performance compared to prior active learning approaches for 3D semantic segmentation.

Figures (7)

• Figure 1: The proposed SELECT employs a unified submodular approach that ensures the selection of points that are both informative and balanced in label distribution. Here, informativeness refers to points that either belong to rare classes, where the model exhibits low confidence, or are situated along object boundaries, where semantic ambiguity is high. This stands in contrast to state-of-the-art (SOTA) methods xie2023annotator, which often struggle to jointly capture these critical aspects for active LiDAR semantic segmentation.
• Figure 2: Left: The active learning pipeline for LiDAR semantic segmentation. Right: The proposed SELECT, which consists of three key stages --- efficiently selecting points that are both informative and well-balanced in label distribution for training the LiDAR semantic segmentation model.
• Figure 3: We summarize the statistics of commonly used datasets for the LiDAR semantic segmentation task, including SemanticPOSS pan2020semanticposs and SemanticKITTI behley2019semantickitti. We observe that most points within the same voxel share the same semantic label. Therefore, a reasonable assumption for the voxel-level label is to assign it based on the majority predicted label among the points contained within the voxel.
• Figure 4: Label distribution of LiDAR data points is highly imbalanced, as shown in the SemanticPOSS pan2020semanticposs dataset.
• Figure 5: The mIoU results of AL baselines and SELECT across each active learning round for the SemanticPOSS, SemanticKITTI, and NuScenes datasets.