多人部件解析--Towards Real World Human Parsing: Multiple-Human Parsing in the Wild
2017-06-06 14:25
555 查看
Towards Real World Human Parsing: Multiple-Human Parsing in the Wild
https://arxiv.org/abs/1705.07206
数据库没给出来啊!
本文针对当前 human parsing 数据库基本都是单人标记,而图像实际情况经常含有多人,这里我们提出了一个 Multiple-Human Parsing (MHP) 数据库,一般2-16人每张图像。接着我们提出了一个 Multiple-Human Parser (MH-Parser) 算法,在单人解析过程中同时考虑 global context and local cues,得到不错的效果。
先看数据库:
![](http://img.blog.csdn.net/20170606140020151?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
![](http://img.blog.csdn.net/20170606140027745?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
各个数据库规模:
![](http://img.blog.csdn.net/20170606140117714?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
Dataset statistics
![](http://img.blog.csdn.net/20170606140211931?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
MH-Parser:
![](http://img.blog.csdn.net/20170606140335418?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
MH-Parser 主要包含五个模块:
1)Representation learner: 是一个CNN特征器,它提取的特征由后面几个模块共享,这里使用全卷积网络,以保持 spatial 信息
2)Global parser : 获取整幅图像的全局信息,生成 a semantic parsing map of the whole image
3) Candidate nominator:包括三个子模块 Region Proposal Network (RPN), a bounding box classifier
and a bounding box regression,类似于 Faster RCNN,将每个人检测出来,得到矩形框
4)Local parser: 针对每个含有人的矩形框,进行 semantic labels 语义标记
5)Global-local aggregator :同时将 local parser and the global parser 网络中隐含的信息输入,用于单人矩形框的 semantic parsing predictions
4.2 Detect-and-parse baseline
检测阶段和解析阶段是分离的:
In the detection stage, we use the representation learner and the candidate nominator as the detection
model.
In the parsing stage, we use the representation learner and the local prediction as the
the parsing model.
![](http://img.blog.csdn.net/20170606141828863?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
![](http://img.blog.csdn.net/20170606141858707?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
![](http://img.blog.csdn.net/20170606141928629?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
![](http://img.blog.csdn.net/20170606142014127?watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvemhhbmdqdW5oaXQ=/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/SouthEast)
https://arxiv.org/abs/1705.07206
数据库没给出来啊!
本文针对当前 human parsing 数据库基本都是单人标记,而图像实际情况经常含有多人,这里我们提出了一个 Multiple-Human Parsing (MHP) 数据库,一般2-16人每张图像。接着我们提出了一个 Multiple-Human Parser (MH-Parser) 算法,在单人解析过程中同时考虑 global context and local cues,得到不错的效果。
先看数据库:
各个数据库规模:
Dataset statistics
MH-Parser:
MH-Parser 主要包含五个模块:
1)Representation learner: 是一个CNN特征器,它提取的特征由后面几个模块共享,这里使用全卷积网络,以保持 spatial 信息
2)Global parser : 获取整幅图像的全局信息,生成 a semantic parsing map of the whole image
3) Candidate nominator:包括三个子模块 Region Proposal Network (RPN), a bounding box classifier
and a bounding box regression,类似于 Faster RCNN,将每个人检测出来,得到矩形框
4)Local parser: 针对每个含有人的矩形框,进行 semantic labels 语义标记
5)Global-local aggregator :同时将 local parser and the global parser 网络中隐含的信息输入,用于单人矩形框的 semantic parsing predictions
4.2 Detect-and-parse baseline
检测阶段和解析阶段是分离的:
In the detection stage, we use the representation learner and the candidate nominator as the detection
model.
In the parsing stage, we use the representation learner and the local prediction as the
the parsing model.
相关文章推荐
- 多人部件解析--Towards Real World Human Parsing: Multiple-Human Parsing in the Wild
- 深度学习论文(十)---Multiple-Human Parsing in the Wild
- Bjarne新文章《Evolving a language in and for the real world: C++ 1991-2006》的读后感
- 微软《SOA in the Real World》笔记05——第一章
- 10841 - Lift Hopping in the Real World
- Spatial Influence - Measuring Followship in the Real World
- (intermediate) 最短路 UVA 10841 - Lift Hopping in the Real World
- 微软《SOA in the Real World》笔记01——目录
- Practices of an Agile Developer : Working in the Real World
- UVA- 10841 - Lift Hopping in the Real World(dijkstra)
- [Thought]What is the vital thing on the Internet or in real world?
- Joint Head Pose / Soft Label Estimation for Human Recognition In-The-Wild [2016]
- 微软《SOA in the Real World》笔记08——第一章
- Penetration Testing in the Real World
- (轉貼) Evolving a language in and for the real world C++ 1991-2006 (中文翻譯版) (By Bjarne Stroustrup) (C/C++)
- 微软《SOA in the Real World》笔记04——第一章
- 微软《SOA in the Real World》笔记07——第一章
- 微软《SOA in the Real World》笔记10——第二章
- WebRTC in the real world: STUN, TURN and signaling
- Paper Notes: On Community Detection in Real-world Networks and the Importance of Degree Assortativit