摘要
我们提出 ReCo，一种面向区域级别的对比学习框架，旨在辅助语义分割任务中的模型学习。ReCo 在稀疏的困难负样本像素上执行半监督或监督的像素级对比学习，且仅需极小的额外内存开销。ReCo 实现简单，可基于现成的分割网络进行构建，并在半监督与监督语义分割方法中均能持续提升性能，实现更平滑的分割边界与更快的收敛速度。在标注数据极少的半监督学习场景下，其提升效果最为显著。借助 ReCo，我们仅需每类语义类别5个样本即可训练出高质量的语义分割模型。代码已开源，地址为：https://github.com/lorenmt/reco。
代码仓库
lorenmt/reco
官方
pytorch
GitHub 中提及
dbash/zerowaste
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| semi-supervised-semantic-segmentation-on-1 | ReCo (DeepLab v3+ with ResNet-101 backbone, ImageNet pretrained) | Validation mIoU: 68.50% |
| semi-supervised-semantic-segmentation-on-1 | ReCo (DeepLab v2 with ResNet-101 backbone, ImageNet pretrained) | Validation mIoU: 67.53% |
| semi-supervised-semantic-segmentation-on-2 | ReCo (DeepLab v2 with ResNet-101 backbone, ImageNet pretrained) | Validation mIoU: 64.94% |
| semi-supervised-semantic-segmentation-on-2 | ReCo (DeepLab v3+ with ResNet-101 backbone, ImageNet pretrained) | Validation mIoU: 66.44% |
| semi-supervised-semantic-segmentation-on-3 | ReCo (DeepLab v2 with ResNet-101 backbone, ImageNet pretrained) | Validation mIoU: 56.53% |
| semi-supervised-semantic-segmentation-on-3 | ReCo (DeepLab v3+ with ResNet-101 backbone, ImageNet pretrained) | Validation mIoU: 60.28% |
| semi-supervised-semantic-segmentation-on-4 | ReCo | Validation mIoU: 71.00% |
| semi-supervised-semantic-segmentation-on-5 | ReCo (DeepLab v2 with ResNet-101 backbone, ImageNet pretrained) | Validation mIoU: 68.85% |
| semi-supervised-semantic-segmentation-on-5 | ReCo (DeepLab v3+ with ResNet-101 backbone, ImageNet pretrained) | Validation mIoU: 73.66% |
| semi-supervised-semantic-segmentation-on-6 | ReCo (DeepLab v3+ with ResNet-101 backbone, ImageNet pretrained) | Validation mIoU: 72.14% |
| semi-supervised-semantic-segmentation-on-6 | ReCo (DeepLab v2 with ResNet-101 backbone, ImageNet pretrained) | Validation mIoU: 66.41% |
| semi-supervised-semantic-segmentation-on-7 | ReCo (DeepLab v3+ with ResNet-101 backbone, ImageNet pre-trained) | Validation mIoU: 63.60% |
| semi-supervised-semantic-segmentation-on-7 | ReCo (DeepLab v2 with ResNet-101 backbone, ImageNet pre-trained) | Validation mIoU: 63.16% |
| semi-supervised-semantic-segmentation-on-8 | ReCo (DeepLab v2 with ResNet-101 backbone, ImageNet pretrained) | Validation mIoU: 68.69% |