PC-FractalDB
收藏资源简介：
构建3D点云数据集需要付出大量的人力努力。因此，构建大规模3D点云数据集是困难的。为了解决这个问题，我们提出了一个新开发的点云分形数据库 (pc-fractaldb)，它是受自然3D结构中遇到的分形几何启发的公式驱动的监督学习的新颖族。我们的研究基于这样的假设，即通过学习分形几何，我们可以从比传统3D数据集更真实的3D模式中学习表示。我们展示了PC-FractalDB如何促进解决3D场景理解中最近与数据集相关的几个问题，例如3D模型收集和劳动密集型注释。实验部分显示了我们如何在当前的最高得分上分别为ScanNetV2和SUN rgb-d数据集实现高达61.9% 和59.4% 的性能速率。通过点对比，对比场景上下文 (CSC) 和随机室获得。此外，pc-fractaldb预训练模型在有限数据的训练中特别有效。例如，10% ScanNetV2上的训练数据，pc-fractaldb预训练的VoteNet在38.3% 执行，这比CSC 14.8% 高。特别值得注意的是，我们发现所提出的方法在有限的点云数据中进行3D对象检测预训练时达到了最高的结果。
Constructing 3D point cloud datasets requires substantial human labor, making the development of large-scale 3D point cloud datasets highly challenging. To address this issue, we propose a newly developed point cloud fractal database (pc-fractaldb), which is a novel family of formula-driven supervised learning frameworks inspired by fractal geometry observed in natural 3D structures. Our research is based on the hypothesis that learning fractal geometry allows us to acquire more realistic 3D pattern representations than those from traditional 3D datasets. We demonstrate how pc-fractaldb facilitates resolving several recent dataset-related challenges in 3D scene understanding, such as 3D model collection and labor-intensive annotation. The experimental section shows that we achieve up to 61.9% and 59.4% performance on the current state-of-the-art for the ScanNetV2 and SUN RGB-D datasets respectively, which is obtained via point contrast, contrasting scene context (CSC) and random room sampling. Furthermore, the pc-fractaldb pre-trained model is particularly effective for training with limited data. For example, when trained using only 10% of the ScanNetV2 training data, the VoteNet pre-trained with pc-fractaldb achieves 38.3% performance, which is 14.8% higher than that of the CSC approach. Notably, we find that the proposed method achieves the best results when pre-training for 3D object detection with limited point cloud data.
