Iqela labaphandi abavela kwiiyunivesithi ezininzi zaseMelika bapapashe iprojekthi (I-Pixel-Aligned Implicit Function), evumela ukuba usebenzise iindlela zokufunda ngomatshini ukwakha imodeli ye-3D yomntu ukusuka kumfanekiso omnye okanye ngaphezulu kwe-dimensional. Inkqubo ikuvumela ukuba uphinde uhlaziye iinketho zeempahla eziyinkimbinkimbi, ezifana neengubo ezinqabileyo kunye nezithende, kunye neentlobo zeenwele, ngokuzimeleyo ukubuyisela ukuthungwa kunye nokuma kwiindawo ezingabonakaliyo kwintelekelelo apho imodeli ye-3D yakhiwe. Ukwandisa umgangatho kunye neenkcukacha zemodeli yokugqibela ye-3D, imifanekiso emininzi evela kwii-engile ezahlukeneyo ingasetyenziswa. Ikhowudi yeprojekthi ibhalwe kwiPython usebenzisa isakhelo sePyTorch kunye phantsi kwelayisenisi ye-MIT.
Inethiwekhi ye-neural isetyenziswa njengomthombo wokwakha ngokutsha uyilo lwe-dimensional-dimensional, ekuvumela ukuba ukhethe eyona milo enokwenzeka kwaye uyile izinto ezifihlakeleyo, ukuqala kumzekelo oqeqeshwe kwiinguqulelo ezahlukeneyo zezinto ezikhoyo. Ngokunxuseneyo, iprojekthi ibonelela nge-algorithm yokuthelekisa isiphumo soyilo lwevolumetric kunye nokwenziwa kwimifanekiso ye-2D ebonelelweyo, elungelelanisa iipikseli zomfanekiso we-3D ngokwendawo yazo kwinto ye-XNUMXD kwaye ivelise ezona zinto zilahlekileyo. Nawuphi na umfanekiso unokukhowudwa ze
ukwakhiwa ngokutsha komphezulu kusetyenziswe uyilo "", a
Inethiwekhi ye-neural esekwe kwi-Architecture isetyenziselwa ukuthelekisa ukuthungwa .
Imodeli esele ilungisiwe esetyenziswa ngabaphandi iyafumaneka ngokukhuphela mahala, kodwa idatha ekrwada esetyenziselwa uqeqesho ihlala yabucala njengoko isekwe kurhwebo lwe-3D scans. Ingasetyenziswa njengomthombo wokuziqeqeshela imodeli abantu abasuka kwiprojekthi yeRenderpeople.
umthombo: opennet.ru