Scholarlink

<p>Deep learning models for LiDAR semantic segmentation face significant computational challenges when processing massive point clouds. Subsampling offers a practical solution to reduce these demands, yet its impact on segmentation accuracy remains insufficiently studied. This work presents a comprehensive benchmark of seven subsampling&#xa0;methods:&#xa0;Random Sampling (RS), Farthest Point Sampling (FPS), Poisson Disk Sampling (SB), Voxel Grid Sampling (VB), DBSCAN-based Sampling, DEPOCO (learned compression), and a novel Inverse Distance Importance Sampling (IDIS). Experiments are conducted using two architectures, RandLA-Net and Point Transformer V3 (PTv3), on the SemanticKITTI (mobile laser scanning) and DALES (aerial laser scanning) datasets. Models trained on subsampled data are evaluated against both subsampled and original full-resolution test data, revealing a critical generalization gap. At moderate subsampling (30% point loss), four methods (DEPOCO, FPS, IDIS, VB) match or exceed the baseline (mIoU: 0.672), with DEPOCO achieving the highest mIoU (0.686), indicating a beneficial regularization effect. At extreme subsampling (90% point loss), only RS, DBSCAN, and DEPOCO remain viable on full-resolution data, while geometry-preserving methods (FPS, VB, SB) suffer catastrophic generalization failure, collapsing to 0.11–0.21 mIoU. RS emerges as the most reliable method overall, ranking first or second across all subsampling levels in both architectures while achieving the highest preprocessing throughput. These findings provide practical guidance for selecting subsampling strategies that balance computational efficiency with segmentation robustness. Code: <a href="https://github.com/geoai4cities/LiDAR-Subsampling-Benchmark">https://github.com/geoai4cities/LiDAR-Subsampling-Benchmark</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Performance analysis of subsampled LiDAR point clouds using deep learning based semantic segmentation

  • 期刊论文

  • 发表日期：

    2026年5月25日

查看全文
  • Pyare Lal Chauhan,
  • Aakash Singh Bais,
  • Vaibhav Kumar

摘要

Deep learning models for LiDAR semantic segmentation face significant computational challenges when processing massive point clouds. Subsampling offers a practical solution to reduce these demands, yet its impact on segmentation accuracy remains insufficiently studied. This work presents a comprehensive benchmark of seven subsampling methods: Random Sampling (RS), Farthest Point Sampling (FPS), Poisson Disk Sampling (SB), Voxel Grid Sampling (VB), DBSCAN-based Sampling, DEPOCO (learned compression), and a novel Inverse Distance Importance Sampling (IDIS). Experiments are conducted using two architectures, RandLA-Net and Point Transformer V3 (PTv3), on the SemanticKITTI (mobile laser scanning) and DALES (aerial laser scanning) datasets. Models trained on subsampled data are evaluated against both subsampled and original full-resolution test data, revealing a critical generalization gap. At moderate subsampling (30% point loss), four methods (DEPOCO, FPS, IDIS, VB) match or exceed the baseline (mIoU: 0.672), with DEPOCO achieving the highest mIoU (0.686), indicating a beneficial regularization effect. At extreme subsampling (90% point loss), only RS, DBSCAN, and DEPOCO remain viable on full-resolution data, while geometry-preserving methods (FPS, VB, SB) suffer catastrophic generalization failure, collapsing to 0.11–0.21 mIoU. RS emerges as the most reliable method overall, ranking first or second across all subsampling levels in both architectures while achieving the highest preprocessing throughput. These findings provide practical guidance for selecting subsampling strategies that balance computational efficiency with segmentation robustness. Code: https://github.com/geoai4cities/LiDAR-Subsampling-Benchmark.