真实色彩合成点云可显著提升语义分割性能
Impact of color and mixing proportion of synthetic point clouds on semantic segmentation
- 用BIM生成带真实色与均匀色的合成点云
- 真实色点云使准确率和交并比提升8.2%
- 合成数据占比超70%时效果更优，适合训练大型模型
基于深度学习的点云分割对理解建筑环境至关重要。尽管合成点云（SPC）有弥补数据不足的潜力，但其颜色与混合比例如何影响深度学习分割效果仍是长期未解问题。本文通过大量实验，提出：1）利用BIM生成带有真实色彩和均匀色彩的合成点云的方法；2）改进的基准测试以更好评估性能。在PointNet、PointNet++和DGCNN等模型上的实验表明，使用真实色彩的合成点云在总体准确率（OA）和平均交并比（mIoU）上比均匀色彩的合成点云高出8.2%。此外，当合成点云占比超过70%时，模型性能通常更优。研究还发现，合成点云可替代真实点云用于训练检测大型平坦建筑构件的深度学习模型。本研究揭示了合成点云提升性能的机制，为构建大规模点云模型提供了新思路。
原文摘要 · Abstract (English)
Deep learning (DL)-based point cloud segmentation is essential for understanding built environment. Despite synthetic point clouds (SPC) having the potential to compensate for data shortage, how synthetic color and mixing proportion impact DL-based segmentation remains a long-standing question. Therefore, this paper addresses this question with extensive experiments by introducing: 1) method to generate SPC with real colors and uniform colors from BIM, and 2) enhanced benchmarks for better performance evaluation. Experiments on DL models including PointNet, PointNet++, and DGCNN show that model performance on SPC with real colors outperforms that on SPC with uniform colors by 8.2 % + on both OA and mIoU. Furthermore, a higher than 70 % mixing proportion of SPC usually leads to better performance. And SPC can replace real ones to train a DL model for detecting large and flat building elements. Overall, this paper unveils the performance-improving mechanism of SPC and brings new insights to boost SPC's value (for building large models for point clouds).
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。