Unsupervised template learning for fine-grained object recognition

Shulin Yang, Liefeng Bo, Jue Wang, Linda Shapiro

Washington U

Two steps:

1) aligning image regions that contain the same object part

2) extracting image features within the aligned image regions

This model is learned using an alternative algorithm, which iterates between detecting aligned image regions, and updating the template model.

dataset: Caltech-UCSD Bird200 and Stanford Dogs     

Learn the template and relationships between patterns.->use LBP or other methods to extract features of each template->SVM

Template model:

Score function between templates and a give  image It :

衡量三个部分：

１）ｆｉｔｎｅｓｓ, 某一个模板和最为相似的图像区域的相似程度。

模板Ｔｉ在初始位置ｘi_l附近alpha范围内和该模板的匹配程度

对于同一ｉｍａｇｅ,对各模板的ｆｉｔｎｅｓｓ求和

２）ｃｏ-occurrence, 在同一张图像中同时出现两种模板的可能性。

３）　ｄｉｖｅｒｓｉｔｙ：任意两个模板之间的距离不应太短

Ｏｖｅｒａｌｌ score function between templates and images

Ｔｅｍplate learning

alternating algorithm to optimize the overall score function

找出模板子集和位置

学习模板特征

学习模板间关系，权重ｗ

总结：参数太多了，而且学习到的特征受训练集的种类影响很大，ｇｅｎｅｒａｌｉｚａｔｉｏｎ ability　尚待考察。

　　特征层次性抽象表示？好像做不到