Masu bincike daga Microsoft da Jami'ar China ta Tsakiya sabuwar hanyar aiki mai girma don bin diddigin abubuwa da yawa a cikin bidiyo ta amfani da fasahar koyon injin - FairMOT (Fair Multi-Object Tracking). Lambobi tare da aiwatar da hanya bisa Pytorch da ƙwararrun ƙira ku GitHub.
Yawancin hanyoyin bin diddigin abubuwan da ake dasu suna amfani da matakai biyu, kowannensu ana aiwatar da su ta hanyar hanyar sadarwa ta daban. Mataki na farko yana gudanar da samfuri don gano abubuwan sha'awa, kuma mataki na biyu yana amfani da samfurin neman ƙungiyoyi don sake gano abubuwa da haɗa anka zuwa gare su.
FairMOT tana amfani da aiwatarwa mataki ɗaya bisa tushen hanyar sadarwa mai jujjuyawar juzu'i (, Deformable Convolutional Network), wanda ke ba ka damar cimma gagarumin karuwa a cikin saurin bin diddigin abu. FairMOT yana aiki ba tare da anga ba, ta amfani da hanyar sake ganowa don tantance ɓangarorin cibiyoyin abu akan taswirar abu mai tsayi. Hakazalika, ana aiwatar da na'ura mai sarrafa kwamfuta wanda ke tantance nau'ikan abubuwan da za'a iya amfani da su don hasashen ainihin su, kuma babban tsarin yana aiwatar da haɗakar waɗannan abubuwan don sarrafa abubuwa na ma'auni daban-daban.
Don horar da ƙirar a cikin FairMOT, an yi amfani da haɗin bayanan jama'a shida don ganowa da bincike (ETH, CityPerson, CalTech, MOT17, CUHK-SYSU). An gwada samfurin ta amfani da tsarin gwaji na bidiyo , , и wanda aikin ya bayar da rufe yanayi daban-daban, motsin kyamara ko juyawa, kusurwoyin kallo daban-daban. Gwajin ya nuna cewa
FairMOT mafi sauri gasa model и lokacin da aka gwada su akan firam 30 a kowane rafukan bidiyo na biyu, yana nuna aikin da ya isa don nazarin rafukan bidiyo na yau da kullun akan tashi.
source: budenet.ru