摘要
本文研究了通过利用标注数据与额外未标注数据来解决半监督语义分割问题的方法。我们提出了一种新颖的一致性正则化方法，称为交叉伪监督（Cross Pseudo Supervision, CPS）。该方法在对同一输入图像进行不同初始化扰动的两个分割网络之间施加一致性约束。其中一个扰动分割网络输出的伪独热标签图（pseudo one-hot label map），将用于以标准交叉熵损失监督另一个分割网络，反之亦然。CPS 一致性机制具有双重作用：一方面促使同一输入图像在两个扰动网络间的预测结果高度相似；另一方面通过为未标注数据赋予伪标签，有效扩展了训练数据规模。实验结果表明，该方法在 Cityscapes 和 PASCAL VOC 2012 数据集上均达到了当前最优的半监督语义分割性能。代码已开源，地址为：https://git.io/CPS。
代码仓库
yhuang1997/3D-CPS
pytorch
GitHub 中提及
charlesCXK/TorchSemiSeg
官方
pytorch
GitHub 中提及
harshm121/m3l
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| semi-supervised-semantic-segmentation-on-1 | CPS (DeepLab v3+ with ImageNet-pretrained ResNet-101, single scale inference) | Validation mIoU: 79.21% |
| semi-supervised-semantic-segmentation-on-2 | CPS (DeepLab v3+ with ImageNet-pretrained ResNet-101, single scale inference) | Validation mIoU: 77.62% |
| semi-supervised-semantic-segmentation-on-22 | CPS (DeepLab v3+ with ResNet-101) | Validation mIoU: 69.8 |
| semi-supervised-semantic-segmentation-on-23 | CPS (Range View) | mIoU (1% Labels): 33.7 mIoU (10% Labels): 50.0 mIoU (20% Labels): 52.8 mIoU (50% Labels): 54.6 |
| semi-supervised-semantic-segmentation-on-24 | CPS (Range View) | mIoU (1% Labels): 36.5 mIoU (10% Labels): 52.3 mIoU (20% Labels): 56.3 mIoU (50% Labels): 57.4 |
| semi-supervised-semantic-segmentation-on-25 | CPS (Range View) | mIoU (1% Labels): 40.7 mIoU (10% Labels): 60.8 mIoU (20% Labels): 64.9 mIoU (50% Labels): 68.0 |
| semi-supervised-semantic-segmentation-on-27 | CPS (DeepLab v3+ with ResNet-101) | Validation mIoU: 64.1 |
| semi-supervised-semantic-segmentation-on-28 | CPS (DeepLab v3+ with ResNet-101) | Validation mIoU: 67.4 |
| semi-supervised-semantic-segmentation-on-29 | CPS (DeepLab v3+ with ResNet-101) | Validation mIoU: 71.7 |
| semi-supervised-semantic-segmentation-on-30 | CPS (DeepLab v3+ with ResNet-101) | Validation mIoU: 75.9 |
| semi-supervised-semantic-segmentation-on-4 | CPS | Validation mIoU: 76.44% |
| semi-supervised-semantic-segmentation-on-45 | CPS | Mean IoU: 62.87 |
| semi-supervised-semantic-segmentation-on-8 | CPS (DeepLab v3+ with ImageNet-pretrained ResNet-101, single scale inference) | Validation mIoU: 80.21% |
| semi-supervised-semantic-segmentation-on-9 | CPS (DeepLab v3+ with ImageNet-pretrained ResNet-101, single scale inference) | Validation mIoU: 77.68% |