
摘要
近期在点云领域的自监督学习方法取得了显著进展,但这些方法通常存在一些缺点,包括较长的预训练时间、需要在输入空间进行重建或需要额外模态。为了解决这些问题,我们提出了一种专门针对点云数据设计的联合嵌入预测架构——Point-JEPA。为此,我们引入了一个排序器,该排序器对点云补丁嵌入进行排序,以便在目标和上下文选择过程中高效地计算和利用它们之间的邻近关系。排序器还允许在上下文和目标选择之间共享补丁嵌入的邻近关系计算,进一步提高了效率。实验结果表明,我们的方法在避免输入空间重建或额外模态的情况下,仍能取得与现有最先进方法相当的结果。
代码仓库
Ayumu-J-S/Point-JEPA
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| 3d-part-segmentation-on-shapenet-part | Point-JEPA | Class Average IoU: 85.8±0.1 |
| 3d-point-cloud-classification-on-modelnet40 | Point-JEPA (voting) | Overall Accuracy: 94.1±0.1 |
| 3d-point-cloud-classification-on-modelnet40 | Point-JEPA (no voting) | Overall Accuracy: 93.8±0.2 |
| 3d-point-cloud-classification-on-scanobjectnn | Point-JEPA | OBJ-BG (OA): 92.9±0.4 |
| 3d-point-cloud-linear-classification-on | Point-JEPA | Overall Accuracy: 93.7±0.2 |
| few-shot-3d-point-cloud-classification-on-1 | Point-JEPA | Overall Accuracy: 97.4 Standard Deviation: 2.2 |
| few-shot-3d-point-cloud-classification-on-2 | Point-JEPA | Overall Accuracy: 99.2 Standard Deviation: 0.8 |
| few-shot-3d-point-cloud-classification-on-3 | Point-JEPA | Overall Accuracy: 95.0 Standard Deviation: 3.6 |
| few-shot-3d-point-cloud-classification-on-4 | Point-JEPA | Overall Accuracy: 96.4 Standard Deviation: 2.7 |