4 个月前

重新审视半监督语义分割中的弱到强一致性

重新审视半监督语义分割中的弱到强一致性

摘要

在本研究中,我们重新审视了由FixMatch普及的弱到强一致性框架,该框架在半监督分类中被广泛应用,其中弱扰动图像的预测结果作为其强扰动版本的监督信号。令人惊讶的是,当我们将这一简单的管道应用于分割场景时,它已经能够取得与近期先进工作相当的竞争性结果。然而,其成功在很大程度上依赖于强数据增强的手动设计,这可能限制了对更广泛扰动空间的探索。基于此观察,我们提出了一种辅助特征扰动流作为补充手段,从而扩展了扰动空间。另一方面,为了充分探究原始图像级别的增强方法,我们引入了一种双流扰动技术,使得两个强视图可以同时由一个共同的弱视图引导。因此,我们的整体统一双流扰动方法(UniMatch)在Pascal、Cityscapes和COCO基准测试的所有评估协议中显著超越了所有现有方法。其优越性也在遥感解释和医学图像分析中得到了验证。我们希望我们的复现FixMatch以及我们的研究成果能够激发更多的未来工作。代码和日志可在https://github.com/LiheYoung/UniMatch 获取。

代码仓库

LiheYoung/UniMatch
官方
pytorch
GitHub 中提及

基准测试

基准方法指标
semi-supervised-change-detection-on-levir-cdUniMatch
IoU: 80.7
semi-supervised-change-detection-on-levir-cd-1UniMatch
IoU: 82
semi-supervised-change-detection-on-levir-cd-2UniMatch
IoU: 81.7
semi-supervised-change-detection-on-levir-cd-3UniMatch
IoU: 82.1
semi-supervised-change-detection-on-whu-10UniMatch
IoU: 81.7
semi-supervised-change-detection-on-whu-20UniMatch
IoU: 81.7
semi-supervised-change-detection-on-whu-40UniMatch
IoU: 85.1
semi-supervised-change-detection-on-whu-5UniMatch
IoU: 80.2
semi-supervised-medical-image-segmentation-on-2UniMatch
Dice (Average): 90.47
semi-supervised-medical-image-segmentation-on-3UniMatch
Dice (Average): 89.92
semi-supervised-medical-image-segmentation-on-4UniMatch
Dice (Average): 87.61
semi-supervised-semantic-segmentation-on-1UniMatch (DeepLab v3+ with ImageNet-pretrained ResNet-101, single scale inference)
Validation mIoU: 79.22%
semi-supervised-semantic-segmentation-on-10UniMatch (DeepLab v3 with ResNet-101)
Validation mIoU: 81.2
semi-supervised-semantic-segmentation-on-2UniMatch (DeepLab v3+ with ImageNet-pretrained ResNet-101, single scale inference)
Validation mIoU: 77.92%
semi-supervised-semantic-segmentation-on-21UniMatch (DeepLab v3+ with ResNet-101)
Validation mIoU: 80.94
semi-supervised-semantic-segmentation-on-22UniMatch (DeepLab v3+ with ResNet-101 pretraind on ImageNet-1K)
Validation mIoU: 76.59
semi-supervised-semantic-segmentation-on-27UniMatch (DeepLab v3+ with ResNet-101)
Validation mIoU: 75.20
semi-supervised-semantic-segmentation-on-28UniMatch (DeepLab v3+ with ResNet-101)
Validation mIoU: 77.20
semi-supervised-semantic-segmentation-on-29UniMatch (DeepLab v3+ with ResNet-101)
Validation mIoU: 78.80
semi-supervised-semantic-segmentation-on-3UniMatch (DeepLab v3+ with ResNet-101)
Validation mIoU: 73.0
semi-supervised-semantic-segmentation-on-30UniMatch (DeepLab v3+ with ResNet-101)
Validation mIoU: 79.90
semi-supervised-semantic-segmentation-on-4UniMatch
Validation mIoU: 81.92%
semi-supervised-semantic-segmentation-on-41UniMatch
Validation mIoU: 28.1
semi-supervised-semantic-segmentation-on-42UniMatch
Validation mIoU: 31.5
semi-supervised-semantic-segmentation-on-8UniMatch
Validation mIoU: 79.5%
semi-supervised-semantic-segmentation-on-9UniMatch (DeepLab v3+ with ResNet-101)
Validation mIoU: 80.43
semi-supervised-semantic-segmentation-on-cocoUniMatch
Validation mIoU: 31.9
semi-supervised-semantic-segmentation-on-coco-1UniMatch
Validation mIoU: 38.9
semi-supervised-semantic-segmentation-on-coco-2UniMatch
Validation mIoU: 44.5
semi-supervised-semantic-segmentation-on-coco-3UniMatch
Validation mIoU: 48.2
semi-supervised-semantic-segmentation-on-coco-4UniMatch
Validation mIoU: 49.8

用 AI 构建 AI

从想法到上线——通过免费 AI 协同编程、开箱即用的环境和市场最优价格的 GPU 加速您的 AI 开发

AI 协同编程
即用型 GPU
最优价格
立即开始

Hyper Newsletters

订阅我们的最新资讯
我们会在北京时间 每周一的上午九点 向您的邮箱投递本周内的最新更新
邮件发送服务由 MailChimp 提供
重新审视半监督语义分割中的弱到强一致性 | 论文 | HyperAI超神经