Here are some example of papers describing research similar to the one mentioned in my text, ie linking face recognition algorithms on mobile device cameras with social media data bases. This was widely discussed also in the context of googles acquisition of viewdle and fb´s acquisition of face.com, apps like recongizr, name tag etc, and of course facebooks deep face technology Literature on different methods of denoising cell phone pictures abounds. On averaged probabilistic faces see discussions around eigenfaces vs fisherfaces.
4.3. Face Recognition using Social Context Recent research has shown that social context information can improve face recognition accuracy. The datacrawled from Facebook provides us with such context in-formation, and in order to include it, we follow an approachproposed by Stoneet al. . For each query, a pairwisey1y2y3p(y3|x)p(y1|x)p(y2|x)p(y1,y3|x)p(y1,y3|x)p(y1,y3|x)Figure 5. Example of a Conditional Random Field. The facesare represented through nodes and the edges reflect the relation-ships .conditional random fieldG= (V;E)is defined, where eachnodev2Vrepresents a face and the edgese2Erepresentthe relationship between each pair of faces. An illustrationof such a graphical model is shown in Figure 5. Also, thepossible discrete label spaceL=fl0;::;lNgis defined. Inour caseLcontains all friends of the user. Given an inputimage accompanied by meta dataxwe try to map face iden-titiesy=fyi2Lgto all nodesiin the graph. Stoneet althen represents the model asf(y;x) =Xii(yijx) +X(i;j);i6=jij(yi;yjjx)(5)withi(yijx) =Xmm(x)gm(yi;x)(6)i(yi;yjjx) =Xnn(x)g0n(yi;yj;x)(7)The mapping functionfis a linear combination of differentunivariate and bivariate feature functionsgmandg0n. Thelearned weightsmandnare used to combine the featurefunctions.In our implementation we use two different univariatefeature functions similar to [15, 26, 21]:appeareance(yijx) =1faceScore(i;x)(8)popularity(yijx) =2Pp2P(yi;p)jPj(9)(yi;p) =(1personyiappeare in imagep0else(10)where1and2represent two optional normalization con-stants andPis the total set of images available throw thenetwork. The functionfaceScorerepresents the confi-dence of the k-NN classifier described in the previous sec-tion. Naamanet al.  observed that some people tendto appear more often in private photo collections then otherpeople. They also observed that people that appear togetherin the same image, have a higher probability of appearingtogether in other images. The first observation we describewith the feature functionpopularityand the second onewith the bivariate feature functioncooccurrence.cooccurrence(yi;yjjx) =w3Pp2P(yi;yj;p)jPj(11)(yi;yj;p) =(1yiandyjappear together in imagep0