走边学从自中心上下文数据学习面部属性表征西北jing wangu walk and learn cvpr paperwang.pdfVIP

WalkandLearn:FacialAttributeRepresentationLearningfromEgocentric

andContextualData

JingWangYuChengRogerioSidtFeris

NorthwesternUniversityIBMT.J.WatsonIBMT.J.Watson

jing.wang@chengyu@rsferis@

oflearningrichvisualrepresentations(i.e.,“goodfeatures”)

formodelingattributeswithoutlabels,with

Thewaypeoplelookintermsoffacialattributes(ethnic-afocusonfacialattributeprediction.

ity,haircolor,facialhair,etc.)andtheclothesoracces-Thestateoftheartinfacialattributeclassification,as

soriestheywear(sunglasses,hat,hoodies,etc.)ishighlydemonstratedbystandardevaluationbenarks[31],has

dependentongeo-locationandweathercondition,respec-beenadvancedbymethodsthatusedeepconvolutionalneu-

tively.Thisworkexplores,forthefirsttime,theuseofthisraworks(CNNs)pre-trainedonmassiveamountsofim-

contextualinformation,aspeoplewithwearablecamerasagesthathavebeenlyannotatedwithidentitylabels.

walkacrossdifferentneighborhoodsofacity,inordertoInfact,ithasbeenshownthatidentity-relatedattributes

learnarichfeaturerepresentationforfacialattributeclas-suchasgender,haircolor,andageareimplicitlyencoded

sification,withoutthecostlyannotationrequiredbyinnodesofCNNsthataretrainedforidentitydiscrimina-

previousmethods.Bytrackingthefacesofcasualwalkerstion[31,42].