
摘要
一致性正则化描述了一类在半监督分类问题中取得突破性成果的方法。先前的研究已经确立了聚类假设——即数据分布由样本组成的均匀类别簇,并且这些簇之间由低密度区域分隔——对其成功至关重要。我们分析了语义分割问题,发现其分布并不表现出分隔类别的低密度区域,并将其作为解释为什么半监督分割是一个具有挑战性的问题的原因之一,仅有少数成功的报告。随后,我们确定了增强方法的选择是在没有这种低密度区域的情况下获得可靠性能的关键。我们发现,最近提出的CutOut和CutMix增强技术的适应变体在标准数据集上产生了最先进的半监督语义分割结果。此外,鉴于其挑战性,我们认为语义分割可以作为一个有效的酸性测试(acid test),用于评估半监督正则化器的效果。实现代码位于:https://github.com/Britefury/cutmix-semisup-seg。
代码仓库
Britefury/cutmix-semisup-seg
官方
pytorch
GitHub 中提及
moucheng2017/learning_morphological_perturbation_ssl
pytorch
GitHub 中提及
lorenmt/reco
pytorch
GitHub 中提及
moucheng2017/mismatchssl
pytorch
GitHub 中提及
ZHKKKe/PixelSSL
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| semi-supervised-semantic-segmentation-on-1 | CutMix (DeepLab v2, ImageNet pre-trained) | Validation mIoU: 63.87% |
| semi-supervised-semantic-segmentation-on-2 | CutMix (DeepLab v2, ImageNet pre-trained) | Validation mIoU: 60.34% |
| semi-supervised-semantic-segmentation-on-23 | CutMix-Seg (Range View) | mIoU (1% Labels): 36.7 mIoU (10% Labels): 50.7 mIoU (20% Labels): 52.9 mIoU (50% Labels): 54.3 |
| semi-supervised-semantic-segmentation-on-24 | CutMix-Seg (Range View) | mIoU (1% Labels): 37.4 mIoU (10% Labels): 54.3 mIoU (20% Labels): 56.6 mIoU (50% Labels): 57.6 |
| semi-supervised-semantic-segmentation-on-25 | CutMix-Seg (Range View) | mIoU (1% Labels): 43.8 mIoU (10% Labels): 63.9 mIoU (20% Labels): 64.8 mIoU (50% Labels): 69.8 |
| semi-supervised-semantic-segmentation-on-3 | CutMix (DeepLab v2, ImageNet pre-trained) | Validation mIoU: 51.2 |
| semi-supervised-semantic-segmentation-on-4 | CutMix | Validation mIoU: 72.45% |
| semi-supervised-semantic-segmentation-on-4 | CutMix | Validation mIoU: 67.6% |
| semi-supervised-semantic-segmentation-on-41 | CutMix | Validation mIoU: 26.2 |
| semi-supervised-semantic-segmentation-on-42 | CutMix | Validation mIoU: 29.8 |
| semi-supervised-semantic-segmentation-on-5 | CutMix (DeepLab v2 ImageNet pre-trained) | Validation mIoU: 66.48% |
| semi-supervised-semantic-segmentation-on-5 | CutMix (DeepLab v3+ ImageNet pre-trained) | Validation mIoU: 69.57% |
| semi-supervised-semantic-segmentation-on-6 | CutMix (DeepLab v3+ ImageNet pre-trained) | Validation mIoU: 67.05% |
| semi-supervised-semantic-segmentation-on-6 | CutMix (DeepLab v2 ImageNet pre-trained) | Validation mIoU: 64.81% |
| semi-supervised-semantic-segmentation-on-7 | CutMix (DeepLab v3+ ImageNet pre-trained) | Validation mIoU: 59.52% |
| semi-supervised-semantic-segmentation-on-7 | CutMix (DeepLab v2 ImageNet pre-trained) | Validation mIoU: 53.79% |