Chimiro Chitsva neBeeble Vanotsvaga Vanovimbisa Kuunza Kupenya Kwechokwadi kune Digital Portraits Uchishandisa AI.

Table of Links

• Abstract uye 1. Nhanganyaya
• 2. Basa Rinoenderana
• 3. SwitchLight uye 3.1. Preliminaries
• 3.2. Dambudziko Kugadzira
• 3.3. Architecture
• 3.4. Zvinangwa
• 4. Multi-Masked Autoencoder Pre-kudzidziswa
• 5. Data
• 6. Miedzo
• 7. Mhedziso

Appendikisi

• A. Tsanangudzo Yekuita
• B. Mushandisi WeChidzidzo Chekushandisa
• C. Vhidhiyo Kuratidzira
• D. Yekuwedzera Hunhu Mibairo & References

Abstract

Isu tinosuma nzira yakabatanidzwa yekuvheneka mufananidzo wemunhu unosanganisa chivakwa chinotungamirwa nefizikisi ne pre-training framework. Tichidhirowa paCook-Torrance reflectance modhi, isu takagadzirisa zvine hungwaru dhizaini yekuvaka kuti inyatso kutevedzera kupindirana kwechiedza-pamusoro. Pamusoro pezvo, kuti tikunde kudzikisira kwekushomeka kwemhando yepamusoro-yemhando yemwenje data, isu takagadzira nzira yekuzvitarisira yega pre-training. Iyi inovhenganiswa musanganiswa weiyo chaiyo yemuviri modhi uye yakawedzera dzidziso dhataset inogadza bhenji nyowani mukujekesa chokwadi.

1. Nhanganyaya

Relighting inopfuura chimbo chekunakisa; inovhura mikana isingaperi yekurondedzera uye inogonesa kubatanidzwa kwezvidzidzo munzvimbo dzakasiyana siyana (ona Fig. 1). Kufambira mberi uku kunoenderana nechishuwo chedu chemukati chekupfuura zvipingaidzo zvenyama zvenzvimbo nenguva, tichipawo mhinduro dzinobatika kumatambudziko anoshanda mukugadzirwa kwedhijitari. Iyo inonyanya kushandura mune chaiyo (VR) uye augmented real (AR) zvikumbiro, uko kuvhenekesa kunofambisa iyo chaiyo-nguva kuchinjika kwemwenje, kuve nechokwadi chekuti vashandisi uye zvinhu zvedhijitari zvinogarisana mukati mechero nharaunda, zvichipa inotevera nhanho ye telepresence.

Muchiitiko ichi, isu tinotarisa pane relighting yemunhu. Nepo basa rekuvheneka richida kunzwisisa kwakadzama kwejometri, zvinhu zvenyama, uye kuvhenekerwa, dambudziko rinowedzera kana uchitaura nevanhu, nekuda kwehunhu hwakasiyana hwenzvimbo yeganda pamwe nekusiyana kwakasiyana uye kutaridzika kwezvimiro zvemhando dzakasiyana siyana. zvipfeko, vhudzi uye zvishongedzo. Zvinhu izvi zvinodyidzana nenzira dzakaoma kunzwisisa, zvinoda maalgorithms epamusoro anokwanisa kutevedzera kupindirana kusinganzwisisike kwechiedza nenzvimbo dzakasiyana siyana idzi.

Parizvino, nzira inovimbisa zvikuru inosanganisira kushandiswa kweyakadzika neural network yakadzidziswa pazviviri zvemhando yepamusoro relit mifananidzo yemifananidzo uye hunhu hwayo hunowirirana, hunotorwa kubva padanho rechiedza [10]. Kuedza kwekutanga kwakasvika pakuvhenekazve se 'bhokisi dema' [45, 48], pasina kuongorora mukati memaitiro epasi. Gare gare kufambira mberi kwakatora dhizaini inotungamirwa nefizikisi, ichibatanidza kuenzanisira kwakajeka kwemifananidzo yemukati uye fizikisi yekugadzira mifananidzo [32]. Pandey et al. [34] yakaronga iyo Yese Relight (TR) dhizaini, zvakare inotungamirwa nefizikisi, iyo inobvisa mufananidzo wekuisa mumisikidzo yepamusoro uye mepu dzealbedo, uye inoita kuvhenekera zvichienderana nePhong specular reflectance modhi. Iyo TR dhizaini yave yekutanga modhi yekuvhenekesa mufananidzo, neazvino uye epamberi ekuvaka ekuvaka pamusoro pemusimboti wayo [23, 31, 52].

Tichitevera nzira inotungamirwa nefizikisi, mupiro wedu uri mumubatanidzwa wedhizaini yekuvaka ine inozvitarisira yega pre-training framework. Chekutanga, dhizaini yedu inoshanduka ichienda kune yakanyatsojeka yemuviri modhi nekubatanidza iyo Cook-Torrance specular reflectance modhi [8], inomiririra kufambira mberi kunoshamisa kubva kune empirical Phong specular modhi [37] yakashandiswa muiyo Yese Relight architecture. Iyo Cook-Torrance modhi inotevedzera zvine hungwaru kupindirana kwechiedza nepamusoro mamicrofacets, achiverengera kune nzvimbo dzakasiyana roughness uye kutaridzika. Chechipiri, yedu pretraining framework inoyera nzira yekudzidza kupfuura iyo inowanzo kuoma-ku-kuwana-lightstage data. Nekudzokorodza iyo masked autoencoder (MAE) chimiro [19], isu tinoibata iyo basa rekuvheneka zvakare. Aya magadzirirwo akagadzirwa kuti agadzirise matambudziko akasarudzika akaunzwa nebasa iri, zvichiita kuti modhi yedu idzidze kubva kune isina kunyorwa data uye kunatsa kugona kwayo kuburitsa echokwadi relit mapikicha panguva yekugadzirisa zvakanaka. Sekuziva kwedu, ino inguva yekutanga kushandisa yekuzvitarisira-yega-pre-training zvakanangana nebasa rekuvheneka.

Kupfupisa, mupiro wedu wakapetwa kaviri. Chekutanga, nekusimudzira chimiro chekuratidzira chemuviri, taunza nhanho nyowani yechokwadi mune zvinobuda. Chechipiri, nekutora kudzidza kwekuzvitarisira, takawedzera chiyero chedhata rekudzidziswa uye nekunatsiridza mataurirwo emwenje mumamiriro akasiyana epasirese. Pamwe chete, kufambira mberi uku kwakatungamira SwitchLight chimiro kuti uwane hutsva hwemazuva ano mukuvhenekera kwemunhu.