I-LoGeR - Ukwakhiwa kabusha kwe-3D okuvela kumavidiyo amade kakhulu (DeepMind, UC Berkeley)

Ukuguqula Amahora Evidiyo Awenze Umhlaba We-3D Obumbene

Zibone ngeso lengqondo uthwebula ividiyo yawo wonke umcimbi—umcimbi womshado, umsebenzi wokwakha, noma uhambo lwemvelo ehlathini. Ugcina unamahora ezithombe, kodwa ukulandelana okuyisicaba, okuwumugqa. Kuthiwani uma ungaguqula leyo vidiyo ende, engalawuleki ibe imodeli ye-3D eyodwa, ekwazi ukuhambahamba yaso sonke isigcawu? Lona umgomo wokuvelela we-LoGeR, ukubambisana kocwaningo oluyisisekelo phakathi kwe-DeepMind ne-UC Berkeley. Lobu buchwepheshe abugcini nje ukuhlanganisa izithombe; yakha kabusha ngobuhlakani umhlaba we-3D ophikelelayo ovela ekusakazweni kwamavidiyo amade kukho kokubili ubude besikhathi nendlela ebonakalayo, ibhekana nenye yezinselele ezibaluleke kakhulu ekuboneni ikhompuyutha.

Inselele Eyinhloko: Ukungaguquguquki Ngaphezu Kwezikali Ezinkulu

Izindlela ezijwayelekile zokwakha kabusha i-3D zihamba phambili ngeziqeshana zevidiyo ezimfushane noma iqoqo lezithombe ezithathwe kuma-engeli ahlukene ngesikhathi esisodwa. Kodwa-ke, badonsa kanzima ngamavidiyo "amade". Ubunzima bukabili. Okokuqala, ubude besikhashana: njengoba ividiyo inweba emaminithini noma amahora, ukukhanya kuyashintsha, izinto ziyanyakaza, futhi abantu bayafika badlule. Okwesibili, isilinganiso sendawo: ikhamera ingase inqamule indawo enkulu, njengokuhamba epaki futhi iye esakhiweni, idale indawo enkulu neyinkimbinkimbi ukuze imephu. Amasistimu akhona ngokuvamile ahluleka ukugcina imephu yomhlaba wonke engaguquki, okuholela ekwakhiweni kabusha okungahlangani noma "okuntantayo"—ama-artifact ayisipoki angeyona ingxenye yanoma iyiphi indawo. I-LoGeR ibhekana nalokhu ngokugxila ekwakheni ukumelela okubumbene okuhlala kuhambisana kuzo zonke lezi zikali zesikhathi nendawo.

Indlela I-LoGeR Ifinyelela Ngayo Ukwakhiwa Kabusha Okubumbene

LoGeR, okusho ukuthi i-Long Generative Reconstruction, yethula indlela entsha egxile esu "lokuqalisa imbewu". Esikhundleni sokuzama ukwakha sonke isigcawu se-3D ngesikhathi esisodwa ngokusakazwa kwevidiyo okuyisiphithiphithi, isistimu iqala ngokuhlonza ingxenye encane, elawulekayo yevidiyo okulula ukuyakheka kabusha ngokuzethemba okukhulu. Lesi siqeshana sekhwalithi ephezulu se-3D sisebenza njengehange elizinzile noma "imbewu." Imodeli ibe isikhulisa ngokwandayo lokhu kumelwa kwe-3D, ifreyimu ngohlaka, ifake ngokucophelela ulwazi olusha olubonakalayo ngenkathi ibhekisela emuva kumbewu emisiwe ukuze kuqinisekiswe ukuvumelana komhlaba wonke. Le ndlela ivumela ngempumelelo imodeli ukuthi igweme izingibe ezivamile zesikali, idale imodeli ye-3D enembe kakhudlwana nethembekile kusukela kokokufaka okude kakhulu. Kusuka ekuzameni ukubona sonke isithombe ngesikhathi esisodwa kuya ekusenzeni usisuse kumongo owethembekile.

"Indlela yethu ivumela ukwakhiwa kabusha kwesigcawu se-3D esingaguquki emhlabeni wonke kusukela kuvidiyo ende, okuyisilungiselelo esiyinselele sezindlela ezikhona ezivame ukukhiqiza i-geometry enqanyuliwe." - LoGeR Research Ababhali

Izicelo Ezisebenzayo Zamabhizinisi Nabadali

Izinhlelo zokusebenza ezingaba khona zobuchwepheshe obufana ne-LoGeR ziningi. Kwabaklami bezakhiwo nabathuthukisi bezindlu ezithengiswayo, ingaguqula izinhlolovo zesayithi, ivumele ividiyo elula ukuthi ikhiqize imodeli enemininingwane ye-3D yendawo. Kwezokungcebeleka, abenzi befilimu bangakha amasethi edijithali kusukela kumavidiyo okuhlola indawo. Ngokuphathwa kwempahla nokuphathwa kwempahla, inganika amandla ukumepha okuguquguqukayo kwe-3D kwezindawo ezinkulu. Leli khono lokudala iwele ledijithali elihlangene kusuka kuvidiyo engahlelekile iyithuluzi elinamandla. E-Mewayz, sibona ukusebenzisana kwemvelo nalobu buchwepheshe. I-OS yethu ye-modular yebhizinisi yakhelwe ukuhlanganisa nokuhlela ukusakazwa kwedatha okuyinkimbinkimbi. Cabanga ngemojula yokuphatha iphrojekthi lapho ividiyo yokuhlola isayithi icutshungulwa ngokuzenzakalelayo ithuluzi elifana ne-LoGeR, futhi imodeli ye-3D ewumphumela ixhunyaniswa khona manjalo nohlu lwemisebenzi, i-inventory, nemigqa yesikhathi ngaphakathi kweplathifomu ye-Mewayz, ihlinzeka ngombono ogxilile ngempela futhi onedatha yenqubekelaphambili yephrojekthi.

Ukubheka Phambili: Ikusasa Lokuqonda Kwe-Spatiotemporal

I-LoGeR imele ukweqa okubalulekile okuya kumasistimu e-AI angaqonda umhlaba wethu hhayi nje njengochungechunge lwezithombe ezifinyeziwe, kodwa njengesikhala esiqhubekayo, esishintshayo se-4D (isikhathi se-3D +). Ukuphindaphinda kwesikhathi esizayo kungalandelela izinto nabantu ngaphandle komthungo emahoreni amaningi, kungaqondi nje ukuthi izinto zikuphi, kodwa ukuthi zishintsha futhi zihlanganyele kanjani isikhathi eside. Lokhu kuqonda kwe-spatiotemporal kungumngcele olandelayo. Ezisekelweni ezifana ne-Mewayz, ezihlose ukuba uhlelo olumaphakathi lwebhizinisi, ukuhlanganisa amakhono anjalo edatha yendawo kungase kuguqule indlela izinkampani ezihlela ngayo, ziqaphe, futhi zihlaziye ukusebenza komzimba. Kusisondeza ekusaseni lapho umhlaba wedijithali nowemvelo uxhumene ngaphandle komthungo ukuze kuthathwe izinqumo ezihlakaniphile.

Ngenkathi kuseyiphrojekthi yocwaningo, i-LoGeR ikhomba esikhathini esizayo lapho ukwakha ikhophi yedijithali ebanzi yanoma iyiphi indawo kulula njengokurekhoda ividiyo. Imithelela yokubhala, ukuhlaziya, nokusebenzisana okubonakalayo ijulile, iguqula ukurekhodwa okude kube imihlaba ehlala njalo, ehlolekayo.