
摘要
伪监督(pseudo supervision)被视为半监督语义分割中的核心思想,而如何在仅使用高质量伪标签与充分利用所有伪标签之间取得平衡,始终是一个关键的权衡问题。针对这一挑战,本文提出一种新颖的学习范式——保守-渐进协同学习(Conservative-Progressive Collaborative Learning, CPCL)。在该方法中,两个预测网络并行训练,伪监督机制基于两个网络预测结果的一致性与差异性进行构建。其中一个网络通过交集监督(intersection supervision)寻求共性,仅接受高质量伪标签的监督,以确保监督信号的可靠性;另一个网络则通过并集监督(union supervision)保留差异性,接受全部伪标签的监督,以维持探索过程中的好奇心与多样性。由此,实现了保守演化(conservative evolution)与渐进探索(progressive exploration)的协同。为进一步降低可疑伪标签带来的负面影响,损失函数根据预测置信度动态重加权。大量实验结果表明,CPCL在半监督语义分割任务上达到了当前最优(state-of-the-art)性能。
代码仓库
leofansq/CPCL
官方
pytorch
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| semi-supervised-semantic-segmentation-on-1 | CPCL (DeepLab v3+ with ResNet-50) | Validation mIoU: 76.98% |
| semi-supervised-semantic-segmentation-on-15 | CPCL (DeepLab v3+ with ResNet-101) | Validation mIoU: 77.67% |
| semi-supervised-semantic-segmentation-on-15 | CPCL (DeepLab v3+ with ResNet-50) | Validation mIoU: 75.3% |
| semi-supervised-semantic-segmentation-on-2 | CPCL (DeepLab v3+ with ResNet-50) | Validation mIoU: 74.6% |
| semi-supervised-semantic-segmentation-on-21 | CPCL (DeepLab v3+ with ResNet-101) | Validation mIoU: 73.44 |
| semi-supervised-semantic-segmentation-on-21 | CPCL (DeepLab v3+ with ResNet-50) | Validation mIoU: 71.66 |
| semi-supervised-semantic-segmentation-on-22 | CPCL (DeepLab v3+ with ResNet-50) | Validation mIoU: 69.92% |
| semi-supervised-semantic-segmentation-on-27 | CPCL (DeepLab v3+ with ResNet-50) | Validation mIoU: 61.88 |
| semi-supervised-semantic-segmentation-on-28 | CPCL (DeepLab v3+ with ResNet-50) | Validation mIoU: 67.02 |
| semi-supervised-semantic-segmentation-on-29 | CPCL (DeepLab v3+ with ResNet-50) | Validation mIoU: 72.14 |
| semi-supervised-semantic-segmentation-on-30 | CPCL (DeepLab v3+ with ResNet-50) | Validation mIoU: 74.25 |
| semi-supervised-semantic-segmentation-on-4 | CPCL (DeepLab v3+ with ResNet-50) | Validation mIoU: 73.74% |
| semi-supervised-semantic-segmentation-on-4 | CPCL (DeepLab v3+ with ResNet-101) | Validation mIoU: 76.4% |
| semi-supervised-semantic-segmentation-on-8 | CPCL (DeepLab v3+ with ResNet-50) | Validation mIoU: 78.17% |
| semi-supervised-semantic-segmentation-on-9 | CPCL (DeepLab v3+ with ResNet-50) | Validation mIoU: 74.58 |
| semi-supervised-semantic-segmentation-on-9 | CPCL (DeepLab v3+ with ResNet-101) | Validation mIoU: 77.16 |