Ukulahlekelwa Okugxile Kokutholwa Okungalingani
Ukulahlekelwa okugxilwe kukho kuwumsebenzi wokulahlekelwa okushintshiwe okwenza izibonelo ezilula zokunciphisa isisindo ukuze umtshina ukwazi ukugxila kokuqinile, okungajwayelekile.
Uhlolojikelele
It solved the extreme background-versus-object imbalance that crippled one-stage object detectors.
I-Deep Dive
Ekutholeni into, isithombe singaqukatha izinto ezimbalwa kuphela zangempela kodwa amashumi ezinkulungwane ezindawo zamakhandidethi, cishe zonke ezingemuva okulula. Nge-cross-entropy evamile, lesi sikhukhula sezinto ezingezinhle ezilula silawula ukuhleleka futhi kuminza okuhle okungandile. Ukulahlekelwa okugxilwe kukho, okwethulwe ephepheni le-RetinaNet lango-2017 ngu-Lin kanye nozakwabo ku-Facebook AI, kulungisa lokhu ngokuphindaphinda i-cross-entropy nge-factory (1 - p_t)^gamma. Uma isampula ihlukaniswa ngokuzethemba nangokulungile, i-p_t iseduze noku-1, ngakho isici sincipha siye ku-zero futhi isibonelo esihlukaniswe kahle asinikeleli kangako. Izibonelo eziqinile, ezingahlukaniswanga kahle zigcina isisindo siseduze. Nge-gamma cishe engu-2, i-RetinaNet ifaniswe noma yashaya izitholi ezinezigaba ezimbili ezihamba kancane njenge-Faster R-CNN kuyilapho ihlala inethiwekhi elula ye-single-pass.
I-Technical Insight
Ipharamitha egxilile ye-gamma ilawula ukuthi izibonelo ezilula zicindezelwa kanjani: ku-gamma 0 ukulahlekelwa kwefocal kulingana ne-cross-entropy evamile, futhi i-gamma ephakeme ilola ukugxila kumakesi aqinile. I-alpha yesisindo sokulinganisa (imvamisa engu-0.25 yesigaba esingavamile) ivamise ukuhlanganiswa nayo. Okubalulekile ukuthi into elungisayo ibumba kabusha ama-gradients, hhayi nje inani lokulahlekelwa, ngakho-ke ukusakazwa kwe-backpropagation ngokwemvelo kugcizelela amasampula angacacile ngaphandle kokumbiwa kwezimayini okuyisibonelo esiqinile noma ukusampula kabusha.
I-Strategic Impact
Izindleko kanye nesabelomali
Izinqumo zezakhiwo ziqhuba ukusebenza kanye nezindleko zokusebenza iminyaka.
Izinqumo ezicacile
Imfundo yobuchwepheshe isiza amaqembu ukuthi akhethe isitaki esifanele, hhayi nje esisha.
Ukulawulwa kwekhwalithi
Izinketho ezingcono zobunjiniyela zinciphisa izehlakalo ezinokwethenjelwa ekukhiqizeni.
Ikusasa Lokulahleka Okugxile Kokutholwa Okungalingani
Ukulahleka kwe-focal sekuphenduke isithako esizenzakalelayo ngale kwe-RetinaNet, okuvela kuzitholi ezifana ne-FCOS, ngezigaba, kanye nokuhlukaniswa ngezigaba ezinomsila omude. Izinhlobonhlobo ezifana nokulahlekelwa kwekhwalithi, ukulahlekelwa okugxilwe kuyo, kanye nokulahlekelwa kwe-varifocal kuyayilungisela izitholi zesimanje ezingenazo ihange nezisekelwe ku-transformer. Njengoba ukutholwa kushintsha kuya kumamodeli okuqagela okusethiwe afana ne-DETR asebenzisa ukufanisa kabili, ukukala kabusha kwesitayela sokugxilile kuhlala kuyithuluzi elisebenzayo nomaphi lapho amafrikhwensi ekilasi egwegwe khona kakhulu.
Ukuqaliswa Komhlaba Wangempela
Ithola izimpawu zomgwaqo ezincane noma abahamba ngezinyawo abakude kumafreyimu azishayelayo lapho amaphikseli amaningi angemuva.
Ukuthola izimila ezingavamile noma izilonda ekuhlolweni kwezokwelapha okubuswa izicubu ezinempilo.
Ukubona amaphutha kulayini wokukhiqiza lapho ingxenye enkulu yezingxenye ezihloliwe zijwayelekile.
Ukuhlonza imikhumbi emincane noma izimoto ngesathelayithi enkulu nesithombe sasemoyeni.
Izingozi & Guardrails
Ukuthuthukisa ibhentshimakhi eyodwa kungafihla ubuthakathaka obubanzi besistimu.
Izindleko zengqalasizinda nezokulungisa zivame ukubukelwa phansi.
Izikhala zokuphepha nokubonakala zingakhula njengoba izinhlelo ziba nzima kakhulu.
Ukuqalisa Umhlahlandlela
Chaza ukubambezeleka, ikhwalithi, nezindleko ezihlosiwe ngaphambi kokuqaliswa.
Ibhentshimakhi ngaphansi komthwalo wangempela nezimo zedatha.
Ukuqapha amathuluzi amaphutha, ukukhukhuleka, nomthelela wabasebenzisi.
Lungiselela izindlela zokuhlehlisa nezigameko ngaphambi kokukala.
Qhubeka Uhlole
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Focal Loss for Imbalanced Detection quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Umhlahlandlela olandelayo
Amanethiwekhi we-Siamese kanye Nokulahlekelwa Okuthathu
Imibuzo evame ukubuzwa
What is Focal Loss for Imbalanced Detection?
Ukulahlekelwa okugxilwe kukho kuwumsebenzi wokulahlekelwa okushintshiwe okwenza izibonelo ezilula zokunciphisa isisindo ukuze umtshina ukwazi ukugxila kokuqinile, okungajwayelekile. Ixazulule ukungalingani okwedlulele kwengemuva-kuqhathaniswa-kwento okukhubaze izitholi zento zesiteji esisodwa.
Iyiphi inkinga eyinhloko ekutholweni kwento yesigaba esisodwa ukulahlekelwa okugxilwe kuyo okuklanyelwe ukubhekana nayo?
Ekutholeni okuminyene, izibonelo ezilula zangemuva zidlula kakhulu izinto zangempela futhi zibusa igradient evamile ewela i-entropy. I-focal loss reweights eyivelakancane, izibonelo eziqinile zibala kakhulu.
Iyini i-modulating factor that focal loss iphindaphindeka ku-cross-entropy?
Ukulahlekelwa okugxilwe kukho ngu-(1 - p_t)^gamma * log(p_t). I-(1 - p_t)^itemu le-gamma lihlehla liye kuziro ngokubikezela okulungile ngokuzethemba.
Kwenzekani ekulahlekelweni kokugxilile lapho ipharamitha ye-gamma yokugxilisa isethelwe ku-0?
Ku-gamma = 0 isici (1 - p_t)^0 ilingana no-1, ngakho ukulahlekelwa okugxilwe kunciphisa ncamashi ku-cross-entropy evamile.
Kwenzani ukwanda kwe-gamma ezibonelweni ezilula, ezihlukaniswe kahle?
I-gamma ephakeme yenza (1 - p_t)^itemu le-gamma linciphe ngokushesha ku-p_t ephezulu, ngakho izibonelo ezilula zifaka isandla esincane nakakhulu ekulahlekelweni nasekukhanyeni.
Iyiphi i-architecture yomtshina eyethule ukulahleka okugxile?
Ukulahlekelwa okugxilwe kukho kwethulwe ephepheni le-RetinaNet lango-2017, okuvumela umtshina wesiteji esisodwa ukubambana nezindlela ezinembile zezigaba ezimbili.