Skip to content

Latest commit

 

History

History
20 lines (10 loc) · 1.14 KB

An unsupervised hierarchical feature learning framework for one-shot image recognition.md

File metadata and controls

20 lines (10 loc) · 1.14 KB

An unsupervised hierarchical feature learning framework for one-shot image recognition

1. Intro

One-shot recognition task includes data from

  • Target domain : 實際進行分類的domain,包含target categories的資料( each category只有一個training sample)  ->做監督式分類 (常用 Nearest neighbor、Naive Bayesian、SVM...)
  • Proir-knowledge domain : 不同於Target domain的data,此篇僅用"unlabled" images ( 模擬人腦的學習 ) -> 做feature learning

此篇的Proir-knowledge domain使用unlabled images,與以往論文有很大不同。

  • Proir-knowledge domain用unlabled images來學習的例子-分辨斑馬與馬: 從斑馬資料集中學習出"條紋"這個重要的特徵,此特徵未來也可運用在分辨老虎與豹等等不同類別。

此篇用Hierarchical Dirichlet Process (HDP,階層式狄氏程序)來學習Proir-knowledge domain,HDP-encoder can encode low-level特徵的histograms into higher-level特徵vector