摘要
我们描述了一种用于点云的简单预训练方法。该方法分为三个步骤：1. 遮蔽在相机视图中被遮挡的所有点；2. 学习一个编码器-解码器模型以重建这些被遮挡的点；3. 使用编码器的权重作为下游点云任务的初始化。我们发现，即使仅从ModelNet40构建单一的预训练数据集，这种方法也能在不同的数据集和编码器上提高多种下游任务的准确性。具体而言，我们的方法在物体分类以及基于部件和语义分割任务中均优于先前的预训练方法。我们对预训练特征进行了研究，发现它们导致了广泛的下游最小值，具有较高的变换不变性，并且其激活与部件标签高度相关。代码和数据可在以下地址获取：https://github.com/hansen7/OcCo
代码仓库
hansen7/OcCo
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| 3d-point-cloud-linear-classification-on | OcCo | Overall Accuracy: 89.2 |
| 3d-point-cloud-linear-classification-on-1 | OcCo | Overall Accuracy: 78.3 |
| few-shot-3d-point-cloud-classification-on-1 | OcCo+DGCNN | Overall Accuracy: 90.6 Standard Deviation: 2.8 |
| few-shot-3d-point-cloud-classification-on-1 | OcCo+PointNet | Overall Accuracy: 89.7 Standard Deviation: 1.9 |
| few-shot-3d-point-cloud-classification-on-2 | OcCo+PointNet | Overall Accuracy: 92.4 Standard Deviation: 1.6 |
| few-shot-3d-point-cloud-classification-on-2 | OcCo+DGCNN | Overall Accuracy: 92.5 Standard Deviation: 1.9 |
| few-shot-3d-point-cloud-classification-on-3 | OcCo+DGCNN | Overall Accuracy: 82.9 Standard Deviation: 1.3 |
| few-shot-3d-point-cloud-classification-on-3 | OcCo+PointNet | Overall Accuracy: 83.9 Standard Deviation: 1.8 |
| few-shot-3d-point-cloud-classification-on-4 | OcCo+PointNet | Overall Accuracy: 89.7 Standard Deviation: 1.5 |
| few-shot-3d-point-cloud-classification-on-4 | OcCo+DGCNN | Overall Accuracy: 86.5 Standard Deviation: 2.2 |
| few-shot-3d-point-cloud-classification-on-6 | OcCo | Overall Accuracy: 57.0 |
| point-cloud-classification-on-pointcloud-c | OcCo-DGCNN | mean Corruption Error (mCE): 1.047 |
| point-cloud-segmentation-on-pointcloud-c | OcCo-PointNet | mean Corruption Error (mCE): 1.130 |
| point-cloud-segmentation-on-pointcloud-c | OcCo-DGCNN | mean Corruption Error (mCE): 0.977 |
| point-cloud-segmentation-on-pointcloud-c | OcCo-PCN | mean Corruption Error (mCE): 1.173 |