论文专题笔记:cross-image segmentation

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

论文专题笔记:cross-image segmentation

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:

论文专题笔记:cross-image segmentation

特征图C*H*W, R为N*C  再用infoNCE计算

论文专题笔记:cross-image segmentation

3.Weakly Supervised 3D Semantic Segmentation Using Cross-Image Consensus and Inter-Voxel Affinity Relations --ICCV2021

论文笔记:ICCV2021 Weakly Supervised 3D Semantic Segmentation Using Cross-Image Consensus and Inter-Voxel_Blair_2的博客-CSDN博客 论文专题笔记:cross-image segmentation

 

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).

上一篇:halcon-polar_trans_image_inv将极坐标中的图像转换回笛卡尔坐标--直线变成圆环


下一篇:halcon-polar_trans_image_ext将图像中的圆弧变换为极坐标-->圆环拉成直线