论文阅读:BiSeg: Simultaneous Instance Segmentation and Semantic Segmentation
2017-08-08 10:42
549 查看
这篇文章,继承了InstanceFCN和FCIS方法的思想,进一步做了优化和补充。
在上一篇FCIS方法中,获得的score maps是2∗k2∗(C+1)维,在本文中,考虑以下两条线针对不同的feature map。At first set,针对conv5接1∗1卷积层,生成2∗k21∗(C+1)维score maps。the second set,和set1一样的操作,使用conv3替代conv5(conv3是conv5的两倍大),生成2∗k22∗(C+1)维score maps。(k1,k2)={7,9}。然后只需要将RoI按照不同大小映射到score maps上计算RoI likelihood map,最终得到2∗(C+1)维的RoI likelihood map(包含inside\outside)。示意图如下:
贝叶斯估计就是输入先验概率图,似然概率,获得后验概率图。本文就考虑使用一条semantic segmentation网络获得的分割图作为先验概率图,将求得的RoI likelihood map作为似然概率,优化获得最终的后验概率图作为分割结果。流程图如下:
得到结果图之后的操作和FCIS操作一致。
创新点1:
在InstanceFCN文章中描述的以7∗7作为每个单元格的分辨率时效果最好,本文考虑到在某些情况下9∗9效果也不错。所以提出多尺度融合的方法做进一步的优化。具体操作如下:在上一篇FCIS方法中,获得的score maps是2∗k2∗(C+1)维,在本文中,考虑以下两条线针对不同的feature map。At first set,针对conv5接1∗1卷积层,生成2∗k21∗(C+1)维score maps。the second set,和set1一样的操作,使用conv3替代conv5(conv3是conv5的两倍大),生成2∗k22∗(C+1)维score maps。(k1,k2)={7,9}。然后只需要将RoI按照不同大小映射到score maps上计算RoI likelihood map,最终得到2∗(C+1)维的RoI likelihood map(包含inside\outside)。示意图如下:
创新点2:
在FCIS方法中,作者直接使用RoI likelihood map使用Softmax和max操作分别获得了mask和label。本文进一步增加了贝叶斯估计,提高了分割和分类的准确性。贝叶斯估计就是输入先验概率图,似然概率,获得后验概率图。本文就考虑使用一条semantic segmentation网络获得的分割图作为先验概率图,将求得的RoI likelihood map作为似然概率,优化获得最终的后验概率图作为分割结果。流程图如下:
得到结果图之后的操作和FCIS操作一致。
相关文章推荐
- 论文阅读:Fully Convolutional Instance-aware Semantic Segmentation
- 论文阅读理解 - Fully Convolutional Instance-aware Semantic Segmentation
- 论文阅读:RCNN[Rich feature hierarchies for accurate object detection and semantic segmentation]
- 论文阅读笔记:R-CNN:Rich feature hierarchies for accurate object detection and semantic segmentation
- 【CV论文阅读】:Rich feature hierarchies for accurate object detection and semantic segmentation
- 论文阅读:Instance-aware Semantic Segmentation via Multi-task Network Cascades
- 论文阅读理解 - Semantic Image Segmentation With Deep Convolutional Nets and Fully Connected CRFs
- 论文阅读-《Semantic Segmentation with Reverse Attention》
- R-CNN:Rich feature hierarchies for accurate object detection and semantic segmentation(阅读)
- 论文阅读:ROBUST AND FULLY AUTOMATED SEGMENTATION OF MANDIBLE FROM CT SCANS
- Large Kernel Matters —— Improve Semantic Segmentation by Global Convolutional Network论文阅读
- [置顶] 实例分割初探,Fully Convolutional Instance-aware Semantic Segmentation论文解读
- 论文阅读:Fast and Accurate Semiautomatic Segmentation of Individual Teeth from Dental CT Images
- 论文笔记 《Rich feature hierarchies for accurate object detection and semantic segmentation》
- 论文阅读笔记:Fully Convolutional Networks for Semantic Segmentation
- 论文 Rich feature hierachies for accurate object detection and semantic segmentation(R-CNN)
- [深度学习论文笔记][Instance Segmentation] Simultaneous Detection and Segmentation
- Rich featureHierarchies for accurate object detection and semantic segmentation 阅读笔记
- Rich feature hierarchies for accurate object detection and semantic segmentation论文笔记
- 论文阅读:Deep Filter Banks for Texture Recognition and Segmentation