摘要
我们提出了一种名为Dense FixMatch的简单方法，用于密集且结构化预测任务的在线半监督学习。该方法通过强数据增强结合伪标签生成与一致性正则化，实现高效学习。为将FixMatch应用于图像分类以外的半监督学习任务，我们在伪标签上引入了匹配操作，从而能够充分利用数据增强流水线的全部能力，包括几何变换等复杂操作。我们在Cityscapes和Pascal VOC数据集上的半监督语义分割任务中进行了评估，采用不同比例的标注数据，并对模型设计选择与超参数进行了消融分析。实验结果表明，Dense FixMatch在仅使用四分之一标注样本的情况下，显著优于仅依赖标注数据的监督学习方法，性能已接近全监督学习的水平。
代码仓库
miquelmarti/DenseFixMatch
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| semi-supervised-semantic-segmentation-on-15 | Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) | Validation mIoU: 74.73% |
| semi-supervised-semantic-segmentation-on-15 | Dense FixMatch (DeepLabv3+ ResNet-50, over-sampling, single pass eval) | Validation mIoU: 71.69% |
| semi-supervised-semantic-segmentation-on-2 | Dense FixMatch (DeepLabv3+ ResNet-101, uniform sampling, single pass eval) | Validation mIoU: 73.91% |
| semi-supervised-semantic-segmentation-on-2 | Dense FixMatch (DeepLabv3+ ResNet-50, uniform sampling, single pass eval) | Validation mIoU: 73.39% |
| semi-supervised-semantic-segmentation-on-21 | Dense FixMatch (DeepLabv3+ ResNet-50, over-sampling, single pass eval) | Validation mIoU: 52.15 |
| semi-supervised-semantic-segmentation-on-21 | Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) | Validation mIoU: 54.85 |
| semi-supervised-semantic-segmentation-on-22 | Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) | Validation mIoU: 71.1% |
| semi-supervised-semantic-segmentation-on-22 | Dense FixMatch (DeepLabv3+ ResNet-50, uniform sampling, single pass eval) | Validation mIoU: 70.65% |
| semi-supervised-semantic-segmentation-on-35 | Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) | Validation mIoU: 66.97 |
| semi-supervised-semantic-segmentation-on-35 | Dense FixMatch (DeepLabv3+ ResNet-50, over-sampling, single-pass eval) | Validation mIoU: 65.81 |
| semi-supervised-semantic-segmentation-on-4 | Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) | Validation mIoU: 65.82% |
| semi-supervised-semantic-segmentation-on-4 | Dense FixMatch (DeepLabv3+ ResNet-50, over-sampling, single pass eval) | Validation mIoU: 62.49% |
| semi-supervised-semantic-segmentation-on-40 | Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) | Validation mIoU: 80.82 |
| semi-supervised-semantic-segmentation-on-40 | Dense FixMatch (DeepLabv3+ ResNet-50, over-sampling, single pass eval) | Validation mIoU: 79.98 |
| semi-supervised-semantic-segmentation-on-9 | Dense FixMatch (DeepLabv3+ ResNet-101, over-sampling, single pass eval) | Validation mIoU: 72.04 |
| semi-supervised-semantic-segmentation-on-9 | Dense FixMatch (DeepLabv3+ ResNet-50, over-sampling, single pass eval) | Validation mIoU: 69.02 |