We propose a fully automatic annotation scheme which takes a raw 3D point cloud with a set of fitted CAD models as input, and outputs convincing point-wise labels which can be used as cheap training data for point cloud segmentation. Compared to manual annotations, we show that our automatic labels are accurate while drastically reducing the annotation time, and eliminating the need for manual intervention or dataset-specific parameters. Our labeling pipeline outputs semantic classes and soft point-wise object scores which can either be binarized into standard one-hot-encoded labels, thresholded into weak labels with ambiguous points left unlabeled, or used directly as soft labels during training. We evaluate the label quality and segmentation performance of PointNet++ on a dataset of real industrial point clouds and Scan2CAD, a public dataset of indoor scenes. Our results indicate that reducing supervision in areas which are more difficult to label automatically is beneficial, compared to the conventional approach of naively assigning a hard "best guess" label to every point.

翻译：我们提出了一种全自动标注方案，该方案以原始三维点云和一组拟合的CAD模型作为输入，输出可用的逐点标签，这些标签可作为点云分割的低成本训练数据。与人工标注相比，我们证明自动标签在保持准确性的同时显著缩短了标注时间，并消除了人工干预或特定数据集参数的需求。我们的标注流程可输出语义类别和逐点软目标分数，这些分数既可以二值化为标准独热编码标签，也可以设定阈值生成保留模糊点的弱标签，或直接作为软标签用于训练。我们在真实工业点云数据集和室内场景公开数据集Scan2CAD上评估了标签质量及PointNet++的分割性能。结果表明：相较于对每个点朴素分配硬性"最佳猜测"标签的传统方法，在自动标注难度较高的区域降低监督强度更具优势。