欧式距离的结果One-shotSimilarity的结果.PPT

系统描述 One-shot Similarity 实验 LFW(Labeled Face in the Wild) a database of face photographs designed for studying the problem of unconstrained face recognition. 13,000 images,1680 of the people pictured have two or more distinct photos in the database. 2 views view1： Model selection and algorithm development view2： Performance reporting Training Image-restricted： only know whether a pair of images is matched or mismatched. Image-unrestricted：image’s label is provided 实验结果 ITML：Information Theoretic Metric Leaning technique 实验结果 Single descriptor: LBP 提纲 作者信息 文章信息 One-Shot Similarity One-Shot Similarity (ECCV 08 workshop) One-Shot Similarity Kernel (ICCV 09) One-Shot Similarity 扩展1 Multiple One-Shot Similarity (BMVC 09) One-Shot Similarity 扩展2 Similarity Based on Background Samples (ACCV 09) 扩展2：Similarity Scores based on Background Samples* Background Samples called Negative Sample Sets before(‘A’) 为什么 background samples 对定义相似度函数是有用的呢？ 样本是位于一个向量空间的，这个空间可以使用多种metric，为了确定哪种metric是最适合当前任务的，我们需要对样本所在流形的内在结构进行分析。 有监督学习的方法需要很多label的信息。对于大量的没有label信息的background sample也可以提供一些信息 *Lior Wolf, Tal Hassner, and Yaniv Taigman. Similarity Scores based on Background Samples. ACCV, 2009. 扩展2 扩展2： Similarity Scores based on Background Samples 对于大量的没有label信息的background sample也可以提供一些信息如： 样本更接近某个样本还是更接近background set ？ (one-shot) 两个样本可以很好的与background set分开吗？(two-shot) 两个样本在background set中有相似的近邻吗？(ranking similarity) 扩展2.1: Two-Shot Similarity Function One-Shot-Similarity (I, J, A) Model1 = train (I, A) Score1 = classify (J, Model1) Model2 = train (J, A) Score2 = classify (I, Model2) return ? * (Score1 + Score2) Function Two -Shot-Similarity (I, J, A) score = train ({ I, J }, A) return score One-Shot Similarity Two-Shot Similarity 算法 扩展2.2: Ranking Based Background Similarity 给定background sample set A，计算图像I和J的ranking similarity 计算I(J)与A中每幅图像的相似度，得到两组排序向量rI (rJ)和πI(πJ) rI (rJ) rI(k) – A中第k副图像与图像I的相似度排在第几位 πI(πJ) πI (k) – A中与图像I第k个最相似的图像的序号 图像I和J的ranking相似度： 实验结果 Featur