注册 登录  
 加关注
查看详情
   显示下一条  |  关闭
温馨提示!由于新浪微博认证机制调整,您的新浪微博帐号绑定已过期,请重新绑定!立即重新绑定新浪微博》  |  关闭

极夜.潜的博客

 
 
 

日志

 
 

Plink Art 是如何进行图像识别的?  

2011-01-10 11:52:29|  分类: Research |  标签: |举报 |字号 订阅

  下载LOFTER 我的照片书  |

How does Plink Art recognize paintings?

Plink Art is an app for your mobile phone that lets you identify almost any work of art just by taking a photo of it. See www.plinkart.com.

  Plink Art 是如何进行图像识别的? - 极夜.潜 - 极夜.潜的博客

Answer from Mark Cummins (Co-founder, PlinkArt)

The PhD-sized answer is here:
http://www.robots.ox.ac.uk/~mjc/...
http://www.robots.ox.ac.uk/~james/

James has a old demo which shows you some of the guts:
http://www.robots.ox.ac.uk/~vgg/...
(Scroll down to the images, click and drag a box to search. Look at the details page in the results)

We use something called a visual words architecture. It's the dominant approach in academia, and used by many visual search companies (e.g. SnapTell, Kooaba and others do this also, though there are a few companies who do things differently). We first detect invariant local features in the image (similar to SIFT features, but we use our own design), which can be recognized even after rotation, translation, scaling, some lighting and perspective change. For small databases you could stop there. But for searching large collections, we quantize SIFT space into "visual words". So Voronoi regions in a high-dimensional feature space get mapped to a discrete integer. This allows you to do very rapid retrieval using an inverted index, in just the same way that text search engines work. Finally the match gets verified for geometric consistency using an efficient RANSAC variant.
There is a fair bit of magic in refining the system to work well, e.g. picking good interest point detectors and descriptors, figuring out the best way to do the feature clustering and quantization, query expansion, learned ranking functions, etc. But the basic idea is not too complicated.

  评论这张
 
阅读(108)| 评论(0)
推荐 转载

历史上的今天

评论

<#--最新日志,群博日志--> <#--推荐日志--> <#--引用记录--> <#--博主推荐--> <#--随机阅读--> <#--首页推荐--> <#--历史上的今天--> <#--被推荐日志--> <#--上一篇,下一篇--> <#-- 热度 --> <#-- 网易新闻广告 --> <#--右边模块结构--> <#--评论模块结构--> <#--引用模块结构--> <#--博主发起的投票-->
 
 
 
 
 
 
 
 
 
 
 
 
 
 

页脚

网易公司版权所有 ©1997-2018