论文专题笔记:cross-image segmentation
作者:互联网
论文汇总:前两篇全监督,后两篇弱监督
1.Region-aware Contrastive Learning for Semantic Segmentation --ICCV2021
2.Exploring Cross-Image Pixel Contrast for Semantic Segmentation --ICCV2021
3.Weakly Supervised 3D Semantic Segmentation Using Cross-Image Consensus and Inter-Voxel Affinity Relations --ICCV2021
4.Cross-Image Region Mining with Region Prototypical Network for Weakly Supervised Segmentation --arxiv2108
1.Region-aware Contrastive Learning for Semantic Segmentation --ICCV2021
好的分割潜空间应具备:
1) address the categorization ability of individual pixel embeddings
2) be well structured to address intra-class compactness and inter-class dispersion.
受无监督对比学习启发,
pixel-wise cross entropy loss-->1)
pixel-wise contrastive loss-->2) -->idea : pixel-to-pixel contrast
算法创新点:
i)a region memory bank:更多具有代表性的数据样本,并充分探索像素和语义级段(区域)之间的结构关系 每个类别每张图随机选一个像素序列,含像素V远小于总像素
ii)different sampling strategies:难例挖掘
2.Exploring Cross-Image Pixel Contrast for Semantic Segmentation --ICCV2021
pixel-wise cross entropy loss + region-aware contrastive loss
Region Center:预测特征图上同类点的和/预测全图像素和。存在问题:构建的区域中心覆盖了像素的模糊特征,因为网络预测包含错误预测,这会误导区域中心的学习过程。 为了将更多的注意力分配给难以分类的像素,我们进一步提出了一种动态采样方法来构建区域中心。
提出dynamic sampling构造center:
特征图C*H*W, R为N*C 再用infoNCE计算
3.Weakly Supervised 3D Semantic Segmentation Using Cross-Image Consensus and Inter-Voxel Affinity Relations --ICCV2021
4.Cross-Image Region Mining with Region Prototypical Network for Weakly Supervised Segmentation --arxiv2108
提出了一个区域原型网络(RPNet)来探索训练集的跨图像对象多样性。 通过区域特征比较来识别图像中相似的对象部分。 对象置信度在区域之间传播以发现新的对象区域,同时抑制背景区域。
弱监督常用的CAM(分类热图)存在一定问题,
1)incomplete object regions (foreground)
2) falsely activated cluttered regions (background).
标签:segmentation,Semantic,--,Image,image,Segmentation,cross,Region,ICCV2021 来源: https://blog.csdn.net/Blair_2/article/details/123174472