多输出回归监督描述符学习supervised descriptor learning cvpr paperzhen.pdf

SupervisedDescriptorLearningforMulti-OutputRegression

XiantongZhenZhijieWang

UniversityofWesternOntario,London,ON,CanadaGEHealthcare,London,ON,Canada

xzhen7@uwo.cazhijie@ualberta.ca

MengyangYuShuoLi

NorthumbriaUniversity,Newcastle,UnitedKingdomGEHealthcare,London,ON,Canada

mengyang.yu@northumbria.ac.ukshuo.li@

ousmultivariatetargetsratherthandiscreteclasslabels[4].

However,multi-outputregressionhasrecentlyemergedand

Descriptorlearninghasrecentlydrawnincreasingatten-extensivelystudiedformanycomputervisiontasks,e.g.,

tionincomputervision,Existingalgorithmsaremainlyde-headposeestimation[15],humanbodyposeestimation

velopedforclassificationratherthanforregressionwhich[29]andviewpointestimation[28].Moreover,manyre-

howeverhasrecentlyemergedasapowerfultooltosolveasearchershavefoundtheirapplications,e.g.,camerarelo-

broadrangeofproblems,e.g.,headposeestimation.Inthiscalization[24,13]andcardiacvolumeestimation[1,38],

paper,weproposeanovelsuperviseddescriptorlearningcanbeelaboratelysolvedbytransferringtheoriginalprob-

(SDL)algorithmtoestablishadiscriminativeandcompactlemintoamulti-outputregressiontask,whichnotonlysub-

featurerepresentationformulti-outputregression.Byfor-