Abacwaningi abavela eMicrosoft naseCentral China University indlela entsha esebenza kahle kakhulu yokulandelela izinto eziningi kuvidiyo kusetshenziswa ubuchwepheshe bokufunda komshini - i-FairMOT (I-Fair Multi-Object Tracking). Ikhodi esebenzisa indlela esekelwe kuPytorch namamodeli aqeqeshiwe ku-GitHub.
Izindlela eziningi ezikhona zokulandelela into zisebenzisa izigaba ezimbili, ngasinye senziwe inethiwekhi ye-neural ehlukile. Isigaba sokuqala sisebenzisa imodeli yokunquma indawo yezinto ezithakaselwayo, futhi isigaba sesibili sisebenzisa imodeli yokusesha esetshenziselwa ukuhlonza kabusha izinto nokunamathisela amahange kuzo.
I-FairMOT isebenzisa ukuqaliswa kwesiteji esisodwa okusekelwe kunethiwekhi ye-neural convolutional deformable (, Inethiwekhi ye-Deformable Convolutional), ekuvumela ukuthi ufinyelele ukwanda okubonakalayo kwejubane lokulandela into. I-FairMOT isebenza ngaphandle kwamahange, isebenzisa indlela yokuphinda ihlonze ukuze inqume ukuhlukana kwezindawo zento kumephu yento enembayo ephezulu. Ngokuhambisanayo, kwenziwa iphrosesa ehlola izici ngazinye zezinto ezingasetshenziswa ukubikezela ubunikazi bazo, futhi imojuli eyinhloko yenza ukuhlangana kwalezi zici ukuze kusetshenziswe izinto zezikali ezihlukene.
Ukuze uqeqeshe imodeli ku-FairMOT, inhlanganisela yamadathasethi ayisithupha asesidlangalaleni okutholwa nokusesha kwabantu yasetshenziswa (ETH, CityPerson, CalTech, MOT17, CUHK-SYSU). Imodeli ihlolwe kusetshenziswa amasethi okuhlola amavidiyo , , ΠΈ ehlinzekwa iphrojekthi nokumboza izimo ezihlukene, ukunyakaza kwekhamera noma ukuzungezisa, ama-engeli ahlukene wokubuka. Ukuhlola kwabonisa lokho
I-FairMOT amamodeli ancintisana asheshayo ΠΈ uma ihlolwa kozimele abangu-30 ngesekhondi ngayinye yokusakaza kwevidiyo, ebonisa ukusebenza okwanele ukuze kuhlaziywe ukusakazwa kwamavidiyo avamile ngokushesha.
Source: opennet.ru