摘要
从有限的标注数据中理解视觉信息是机器学习的一个重要方面。尽管在半监督设置下图像级别的分类已经得到了广泛研究，但基于有限数据的密集像素级别分类直到最近才引起关注。本文提出了一种半监督语义分割方法，该方法从有限的像素级标注样本中学习，并利用额外的无标注图像。该方法采用两个网络分支，将半监督分类与包含自训练在内的半监督分割联系起来。双分支方法减少了在少量标签训练时常见的低层次和高层次伪影。该方法在现有方法上取得了显著改进，特别是在使用非常少的标注样本进行训练时表现尤为突出。在多个标准基准测试集——PASCAL VOC 2012、PASCAL-Context 和 Cityscapes——上，该方法在半监督学习中达到了新的最先进水平。
代码仓库
sud0301/semisup-semseg
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| semi-supervised-semantic-segmentation-on-1 | s4GAN (DeepLab v2 ImageNet pre-trained) | Validation mIoU: 61.9% |
| semi-supervised-semantic-segmentation-on-11 | s4GAN+MLMT (DeepLab v2 ImageNet pre-trained) | Validation mIoU: 35.3 |
| semi-supervised-semantic-segmentation-on-12 | s4GAN+MLMT (DeepLab v2 ImageNet pre-trained) | Validation mIoU: 37.8 |
| semi-supervised-semantic-segmentation-on-18 | S4GAN (DeepLabv2 with ResNet101, MSCOCO pre-trained) | Validation mIoU: 50.48% |
| semi-supervised-semantic-segmentation-on-19 | S4GAN (DeepLabv2 with ResNet101, MSCOCO pre-trained) | Validation mIoU: 55.61% |
| semi-supervised-semantic-segmentation-on-2 | s4GAN (DeepLab v2 ImageNet pre-trained) | Validation mIoU: 59.3% |
| semi-supervised-semantic-segmentation-on-4 | s4GAN + MLMT | Validation mIoU: 71.4% |
| semi-supervised-semantic-segmentation-on-4 | s4GAN+MLMT | Validation mIoU: 70.4% |
| semi-supervised-semantic-segmentation-on-4 | s4GAN+MLMT | Validation mIoU: 67.3% |
| semi-supervised-semantic-segmentation-on-5 | s4GAN + MLMT (DeepLab v2 MSCOCO/ImageNet pre-trained) | Validation mIoU: 67.2% |
| semi-supervised-semantic-segmentation-on-5 | s4GAN+MLMT (DeepLab v3+ ImageNet pre-trained) | Validation mIoU: 66.6% |
| semi-supervised-semantic-segmentation-on-5 | s4GAN+MLMT (DeepLab v2 ImageNet pre-trained) | Validation mIoU: 62.9% |
| semi-supervised-semantic-segmentation-on-6 | s4GAN+MLMT (DeepLab v2 ImageNet pre-trained) | Validation mIoU: 60.4% |
| semi-supervised-semantic-segmentation-on-6 | s4GAN+MLMT (DeepLab v3+ ImageNet pre-trained) | Validation mIoU: 62.6% |
| semi-supervised-semantic-segmentation-on-6 | s4GAN + MLMT (DeepLab v2 MSCOCO/ImageNet pre-trained) | Validation mIoU: 63.3% |