R. Wang-Manifold-Manifold Distance with Application to Face Recognition based on Image Set读后记

题目：Manifold-Manifold Distance with Application to Face Recognition based on Image Set

作者：Ruiping Wang 1, 2, 3 , Shiguang Shan 1, 2 , Xilin Chen 1, 2 , Wen Gao 4, 

主要思想：利用流形集合 与 流形集合的距离计算方法，识别人脸数据集。

比如：训练数据：n 个人 很多角度的人脸（每个人 不同角度的人脸图像形成一个流型）

           测试 某个人 很多角度的人脸（同样形成一个流形集合），计算流形集合的距离

关键：流形集合的距离度量，本文采用了 Principal angles，cos (theta_k) =  max max u_k^T *v_k，其中u_k,v_k是一对典型向量canonical vectors.

     

        而流形集的角度距离，使用了SVD矩阵分解，特征根就为所需的角度（PS:这个矩阵分解比较牛，虽然还没搞懂，为什么角度直接通过矩阵分解得到）。

       

         

原文摘要：In  this  paper,  we  address  the  problem  of  classifying image sets, each of which contains images belonging to the  same class but covering large variations in, for instance, viewpoint and illumination. We innovatively formulate the problem  as  the  computation  of  Manifold-Manifold Distance  (MMD),  i.e.,  calculating  the  distance  between nonlinear manifolds each representing one image set. To compute MMD, we also propose a novel manifold learning approach,  which expresses  a  manifold by  a collection  of local linear models, each depicted by a subspace. MMD is then converted to integrating the distances between pair of subspaces respectively from one of the involved manifolds. 

The proposed MMD method is evaluated on the task of Face  Recognition  based  on  Image  Set  (FRIS).  In  FRIS, each known subject is enrolled with a set of facial images and modeled as a gallery manifold, while a testing subject is  modeled  as  a  probe  manifold,  which  is  then  matched against all the gallery manifolds by MMD. Identification is achieved  by  seeking  the  minimum  MMD.  Experimental results  on  two  public  face  databases,  Honda/UCSD  and CMU MoBo, demonstrate that the proposed MMD method outperforms the competing methods. 

结果：