摘要
我们提出n-CPS——一种对近期最先进的交叉伪监督（Cross Pseudo Supervision, CPS）方法的泛化，用于半监督语义分割任务。在n-CPS中，存在n个同时训练的子网络，它们通过one-hot编码扰动和一致性正则化机制相互学习。此外，我们证明了将集成技术应用于子网络输出，可显著提升模型性能。据我们所知，n-CPS结合CutMix方法在Pascal VOC 2012数据集（在1/16、1/8、1/4和1/2的监督比例下）以及Cityscapes数据集（1/16监督比例下）上均优于CPS，达到了新的最优性能，刷新了当前的最先进水平。
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| semi-supervised-semantic-segmentation-on-1 | n-CPS (ResNet-50) | Validation mIoU: 78.41% |
| semi-supervised-semantic-segmentation-on-15 | n-CPS (ResNet-50) | Validation mIoU: 77.07% |
| semi-supervised-semantic-segmentation-on-15 | n-CPS (ResNet-101) | Validation mIoU: 80.26% |
| semi-supervised-semantic-segmentation-on-2 | n-CPS (ResNet-50) | Validation mIoU: 77.61% |
| semi-supervised-semantic-segmentation-on-21 | n-CPS (ResNet-50) | Validation mIoU: 72.03 |
| semi-supervised-semantic-segmentation-on-21 | n-CPS (ResNet-101) | Validation mIoU: 75.86 |
| semi-supervised-semantic-segmentation-on-22 | n-CPS (ResNet-50) | Validation mIoU: 76.08 |
| semi-supervised-semantic-segmentation-on-4 | n-CPS (ResNet-101) | Validation mIoU: 77.99% |
| semi-supervised-semantic-segmentation-on-4 | n-CPS | Validation mIoU: 74.21% |
| semi-supervised-semantic-segmentation-on-8 | n-CPS (ResNet-50) | Validation mIoU: 79.29% |
| semi-supervised-semantic-segmentation-on-9 | n-CPS (ResNet-50) | Validation mIoU: 75.85 |
| semi-supervised-semantic-segmentation-on-9 | n-CPS (ResNet-101) | Validation mIoU: 78.97 |