Kwenzekeni
Abacwaningi bahlongoza i-MV2GF, isistimu ye-AI yokuthola abahamba ngezinyawo ekubukeni kwamakhamera amaningi futhi ibamele kumephu yokubuka iso lenyoni. Leli phepha lithi le ndlela isebenzisa imodeli yesisekelo esibonakalayo sejiyomethri ukuze ithuthukise ukusebenza lapho isetshenziswa ngokucushwa kwekhamera okungekho ekuqeqesheni.
Irekhodi le-arXiv likhomba i-MV2GF njengephepha lombono wekhompyutha elithunyelwe ngomhla ka-Aug. 21, 2026, futhi lithi lamukelwe i-ECCV 2026. Umsebenzi walo omaphakathi ukutholwa kwabahamba ngezinyawo kokubuka okuningi: ukuthatha izithombe kumakhamera amaningi nokulinganisa ukuthi abahamba ngezinyawo bakuphi ngokuboniswa kwezinyoni ngokuhlanganyela. Iphepha lichaza lokhu njengendlela yokuguqula ulwazi lusuka ekubukeni okuhlukene kwesithombe lube umfanekiso ojwayelekile wendawo yomhlaba we-3D.
Ngokuvumelana ne-abstract, izindlela ezikhona zokubona abahamba ngezinyawo abaningi zivame ukuveza izici zesithombe ezinezinhlangothi ezimbili esikhaleni se-3D futhi zizihlanganise zibe ukumelwa kwesici esisodwa. Ababhali bathi lezi zinhlelo zinobunzima bokujwayela ukucushwa kwamakhamera abengamelwe ngesikhathi sokuqeqeshwa. Zihlonza izimbangela ezimbili ezihlobene: ubunzima bokuthwebula ijiyomethri ebonakalayo enembile kuwo wonke ukubukwa ngamalungiselelo angajwayelekile, nokuncika kumaphethini okuhlanekezela adalwe inqubo yokuqagela isici.
I-MV2GF ibhekana nalezo zinkinga ngokwengeza izici zejiyomethri zenhloso evamile ukusuka kumodeli yesisekelo esibonakalayo sejiyomethri kuya kuzici zokutholwa eziqondene nomsebenzi othile. I-abstract ithi le modeli yesisekelo ingakwazi ukuhlanganisa konke ukubuka futhi ibikezele izibaluli ze-3D ngaphansi kokulungiselelwa okuhlukahlukene kwekhamera. I-MV2GF iphinda isebenzisa amamephu wamaphoyinti abikezelwe imodeli eyisisekelo ukuze iphrojekthi iphikseli ngayinye yesici sesithombe kulokho ababhali abakuchazayo njengendawo efanele ye-3D. Lokhu kuhloswe ngakho ukunciphisa ukuthembela emaphethini okuhlanekezela abonwa ngesikhathi sokuqeqeshwa.
Ababhali babika ukuthi ukuhlola kwabo kukhombisa ukuthi i-MV2GF iyasebenza ekutholweni kwabahamba ngezinyawo abanokubukwa okuningi futhi iba ngcono kunezindlela ezikhona. Umthombo onikeziwe awuqukethe imiphumela yezinombolo, amagama esethi yedatha, ukuhlukaniswa kwesisekelo, ukuhlaziya amaphutha, izilinganiso zokubala, noma imininingwane mayelana nokukhishwa kwesofthiwe. Lokho kweqiwe kusho ukuthi ukuthuthukiswa kwesihloko sephepha kungabikwa njengesimangalo ngababhali balo, hhayi njengomphumela wokusebenza osungulwe ngokuzimela.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Amasistimu okuthola ukubuka okuningi angancika ekuhlelweni kwekhamera namaphethini okuhlanekezela asetshenziswa phakathi nokuthuthukiswa. Uma izimangalo zephepha zibambelela ekuhlolweni okuzimele, indlela yalo egxile ku-geometry ingenza amasistimu anjalo avumelane nezimo ezishintshayo zokuhlelwa kwekhamera, nakuba umthombo ungakusunguli ukuthunyelwa noma umthelela womphakathi.
Inkinga engokoqobo esingathwa yi-MV2GF ukuguquguquka. Umtshina oqeqeshelwe ukuhlelwa kwamakhamera ungase ungagcini nje ngokufunda izinkomba mayelana nabahamba ngezinyawo kodwa nokufunda okujwayelekile okuhlobene nendlela lawo makhamera abheka ngayo indawo efanayo. Lapho amakhamera enyakaza, ukuma kwawo kuyashintsha, noma kwethulwa uhlelo olusha, lezo zimo ezifundiwe zingase zingabe zisadluliswa ngokuhlanzekile. Indlela yephepha iqondise lobo buthakathaka obuthile ngokugcizelela i-geometry engaphansi yesehlakalo.
Ukusetshenziswa kwemodeli yesisekelo se-geometric ebonakalayo kungumnikelo oyinhloko we-AI wephepha. Kunokuthembela kuphela kuzici ezifundiwe ukuze kutholwe abahamba ngezinyawo, i-MV2GF ihlanganisa ulwazi oluqondene nomsebenzi othile nolwazi lwejiyomethri ababhali abaluchaza njengoluvamile. Uma inzuzo yokwenza okuvamile ebikiwe iphindwaphindwa, lokhu kungase kunikeze indlela yokwenza amasistimu e-AI enamakhamera amaningi anganciki kusakhiwo sokuqeqeshwa esingashintshi. Lokho kungaba usizo ngempela konjiniyela abadinga ukulungisa isistimu njengoba kushintsha ukulungiselelwa kwekhamera, nakuba umthombo ungasilinganisi isikhathi sokuqeqesha kabusha, umzamo wokulinganisa, noma ukonga kokusebenza.
Umsebenzi uphinda ubonise isiqondiso esibanzi ku-AI esetshenziswayo: ukusebenzisa imodeli yesikhala esiqeqeshelwe kusengaphambili ukuthuthukisa umsebenzi wombono omncane. Kulokhu, imodeli yesisekelo ayivezwanga njengomtshina wokugcina. Inikeza ulwazi lwejiyomethri olusetshenziswa isistimu yokuhlonza ukubeka izici ezibonakalayo esikhaleni se-3D. Lokho kuhlukaniswa kungase kubaluleke uma kuthuthukisa ukuqina ngaphandle kokudinga imodeli yokuthola ukuthi ifunde wonke amalungiselelo ekhamera kusukela ekuqaleni.
Ukubaluleka komphakathi kusalokhu kunqunyelwe lokho umthombo ongakusunguli. Iphepha lichaza indlela yocwaningo, hhayi ukuthunyelwa okuqinisekisiwe kwezokuthutha, ukuphepha, amarobhothi, noma esinye isilungiselelo sokusebenza. Ayibiki ukuthi amaphutha abathinta kanjani abantu abatholwayo, ukuthi indlela isebenza kanjani ngokubuka okungaphelele noma okungalinganiswanga kahle, noma ukuthi ukucubungula kwayo kwemodeli yesisekelo eyengeziwe kuyathengeka ukuze kusetshenziswe isikhathi sangempela. Ukwamukelwa kwe-ECCV kubonisa ukuqashelwa kwezemfundo okubikwe ngumthombo, kodwa akubona ubufakazi bokuthi uhlelo lulungele ukuthunyelwa kabanzi.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Imibuzo engukhiye engaphenduliwe iwumthamo: yimaphi amasethi edatha nesisekelo esisetshenzisiwe, izinzuzo bezinkulu kangakanani, yiziphi izindleko zekhompuyutha ezengezayo imodeli yesisekelo, nokuthi ingabe indlela ihlala ithembekile ngaphansi kwezinguquko zekhamera zomhlaba wangempela. Umthombo futhi awunikezi ikhodi, imininingwane yokutholakala, noma ukuhlaziywa kwamacala okuhluleka.
Okubalulekile kokuqala kokuqinisekisa ubufakazi obusekela isimangalo esijwayelekile. Abafundi kufanele babheke ukusetha okuphelele kokuhlola kwephepha: amasethi edatha, amalungiselelo ekhamera asetshenziselwa ukuqeqeshwa nokuhlola, izindlela eziqhudelanayo, kanye nezinguquko zezinombolo zokunemba noma ezinye izindlela zokuthola. Kuzoba nendaba ukuthi "ukulungiselelwa kwekhamera okungabonakali" kusho izinguquko ezincane ngaphakathi kwebhentshimakhi elawulwayo noma izinguquko ezinkulu ezifana nezimo zokusebenzisa.
Indaba elandelayo izindleko kanye nesakhiwo sokuncika sesistimu. Umthombo awukhombi imodeli yesisekelo se-geometric ebonakalayo, uyachaza ukuthi yaqeqeshwa kanjani, noma usho ukuthi iyatholakala yini esidlangalaleni. Futhi ayibiki isivinini sokucatshangelwa, ukusetshenziswa kwenkumbulo, izidingo zehadiwe, ukuqagela kokulinganisa, noma ukuthi ukubikezela kwemephu yephoyinti kudala ibhodlela lokucubungula elengeziwe. Leyo mininingwane izonquma ukuthi i-MV2GF iwumphumela wocwaningo yini noma iyingxenye esebenzayo yezinhlelo zamakhamera amaningi.
Ukuphindaphinda okuzimele kufanele kuhlole ukuthi inzuzo ebikiwe ivela ekumelelweni kwejiyomethri ngokwayo nokuthi ingabe idlulela ngale kwezilungiselelo zokuhlola zababhali. Umsebenzi wokulandelela owusizo ungahlola izinguquko ekubekweni kwekhamera, ama-engeli wokubuka, nezinye izimo zokufaka, kuyilapho kuqhathaniswa i-MV2GF nezisekelo eziqinile zamanje ngaphansi kwemithetho yokuhlola efanayo. Umthombo awunikezi ubufakazi obunjalo obuzimele.
Okokugcina, izimangalo zokusatshalaliswa kufanele ziphathwe ngokwehlukana nezimangalo zebhentshimakhi. Isistimu ingasebenza kangcono ekuhlolweni kocwaningo kuyilapho idinga ukugada ngokucophelela komuntu nokuqinisekiswa ngaphambi kokusetshenziswa kuzilungiselelo lapho ukutholwa kwabahamba ngezinyawo okugejiwe noma okungalungile kunemiphumela. I-abstract yephepha ayixoxi ubumfihlo, ukubusa, izivikelo zokusebenza, noma ukuphatha ukwehluleka, ngakho-ke lokho kuhlala kungaziwa okubalulekile kunamandla abonisiwe noma ubuthakathaka.