PIFu imuchina wekudzidza sisitimu yekugadzira iyo 3D modhi yemunhu yakavakirwa pamifananidzo ye2D

Boka revatsvagiri kubva kumayunivhesiti akati wandei eAmerica rakaburitsa chirongwa PIFu (Pixel-Aligned Implicit Function), iyo inokutendera kuti ushandise nzira dzekudzidza muchina kugadzira 3D modhi yemunhu kubva kune mumwe kana anopfuura maviri-dimensional mufananidzo. Iyo sisitimu inobvumidza iwe kudzoreredza sarudzo dzakaoma dzezvipfeko, senge masiketi ane pleated uye zvitsitsinho, uye akasiyana mabvudzi ebvudzi, akazvimiririra kudzoreredza magadzirirwo uye chimiro munzvimbo dzisingaonekwe mune fungidziro kubva iyo 3D modhi inovakwa. Kuti uwedzere kunaka uye ruzivo rweyekupedzisira 3D modhi, akati wandei mifananidzo kubva kumakona akasiyana anogona kushandiswa. Iyo kodhi yeprojekiti yakanyorwa muPython uchishandisa iyo PyTorch chimiro uye inoparadzirwa ne pasi peMIT rezinesi.

Neural network inoshandiswa sesosi yekuvakazve matatu-dimensional dhizaini, iyo inokutendera iwe kuti usarudze iyo ingangoita chimiro uye kugadzira yakavanzika zvinhu, kutanga kubva pamuenzaniso wakadzidziswa pane dzakasiyana siyana dzezvinhu zviripo. Mukufambirana, purojekiti inopa algorithm yekufananidza mhedzisiro yevhoriyamu marongero nemaumbirwo mumifananidzo yakapihwa ye2D, inoyananisa mapixels emufananidzo we3D zvichienderana nenzvimbo yavo pachinhu cheXNUMXD uye inogadzira iwo angangoshaikwa maumbirwo. Chero mufananidzo unogona kukodha convolutional neural networknokuti
pamusoro pekuvakazve yakashandiswa architecture "Yakaturikidzana hourglass", a
architecture-based neural network inoshandiswa kuenderana nemavara CycleGAN.

Iyo yakagadzirira-yakagadzirirwa yakadzidziswa muenzaniso inoshandiswa nevatsvakurudzi inowanikwa inowanikwa kudhawunirodha mahara, asi iyo mbishi data inoshandiswa pakudzidzisa inoramba yakavanzika sezvo yakavakirwa pakutengesa 3D scans. Inogona kushandiswa senzira yekuzvidzidzisa yega yemuenzaniso 3D modhi database vanhu vanobva kuRenderpeople project.

Source: opennet.ru