Okuqukethwe[Fihla][Bonisa]
- 1. CelebFaces Attributes Dataset
- 2. I-DOTA
- 3. Idathasethi yokuqhathanisa ye-Google Facial Expression
- 4. I-Visual Genome
- 5. LibriSpeech
- 6. Izindawo zedolobha
- 7. Isethi yedatha ye-Kinetics
- 8. CelebMask-HQ
- 9. I-Penn Treebank
- 10. VoxCeleb
- 11. I-SIXray
- 12. Izingozi zase-US
- 13. Ukuqashelwa Kwezifo Zamehlo
- I-14. Isifo senhliziyo
- 15. CLEVR
- 16. Ukuncika Kumhlaba Wonke
- 17. KITTI – 360
- 18. I-MOT(Ukulandelela Izinto Eziningi)
- 19. PASCAL 3D+
- 20. Amamodeli Ezilwane Ezikhubazeka Ebusweni
- 21. MPII Uhlu Lokuthunyelwe Kwabantu
- 22. UCF101
- 23. Isethi yomsindo
- 24. I-Stanford Natural Language Inference
- 25. Izimpendulo Zemibuzo Ebonakalayo
- Isiphetho
Namuhla, iningi lethu ligxile ekuthuthukiseni ukufunda komshini namamodeli e-AI nokubhekana nezinkinga kusetshenziswa amasethi edatha amanje. Kodwa okokuqala, kufanele sichaze idathasethi, ukubaluleka kwayo, kanye nendima yayo ekuthuthukiseni izixazululo eziqinile ze-AI ne-ML.
Namuhla, sinenqwaba yamasethi edatha omthombo ovulekile lapho senza khona ucwaningo noma sithuthukise izinhlelo zokusebenza zokubhekana nezinkinga zomhlaba wangempela emikhakheni eyahlukene.
Nokho, ukushoda kwedathasethi yobuningi bekhwalithi ephezulu kuwumthombo wokukhathazeka. Idatha ikhuphuke kakhulu futhi izoqhubeka nokukhula ngezinga elisheshayo esikhathini esizayo.
Kulokhu okuthunyelwe, sizofaka idathasethi etholakala mahhala ongayisebenzisa ukuthuthukisa iphrojekthi yakho elandelayo ye-AI.
1. I-CelebFaces Attributes Dataset
I-CelebFaces Attributes Dataset (CelebA) iqukethe izithombe zikasaziwayo ezingaphezu kuka-200K nezichasiselo ezingu-40 zesithombe ngasinye, okuyenza ibe isiqalo esihle kakhulu samaphrojekthi afana ukubona ubuso, ukutholwa kobuso, uphawu lwendawo (noma ingxenye yobuso) ukwenziwa kwasendaweni, nokuhlela ubuso nokuhlanganiswa. Ngaphezu kwalokho, izithombe ezikuleli qoqo ziqukethe ukwahlukahlukana okubanzi kwezikhundla kanye nemfuhlumfuhlu yangemuva.
2. I-DOTA
I-DOTA (Idatha ye- Ukutholwa Kwento kokuthi Izithombe zasemoyeni) iyidathasethi yesilinganiso esikhulu sokutholwa kwento ehlanganisa izigaba ezivamile ezingu-15 (isb, umkhumbi, indiza, imoto, njll.), izithombe ezingu-1411 zokuqeqeshwa, kanye nezithombe ezingu-458 zokuqinisekiswa.
3. Idathasethi yokuqhathanisa ye-Google Facial Expression
Idathasethi yokuqhathanisa yobuso be-Google iqukethe cishe ama-triplets ezithombe angu-500,000, okuhlanganisa nezithombe zobuso ezingu-156,000. Kuyaphawuleka ukuthi i-triplet ngayinye kule dathasethi yachazelwa okungenani abantu abayisithupha.
Le dathasethi iwusizo kumaphrojekthi ahilela ukuhlaziywa kwesimo sobuso, njengokubuyiswa kwesithombe esisekelwe ekuboniseni, ukuhlukanisa imizwelo ngezigaba, ukuhlanganisa inkulumo, njalo njalo. Ukuze uthole ukufinyelela kudathasethi, ifomu elifushane kufanele ligcwaliswe.
4. I-Visual Genome
Idatha yokuphendula imibuzo ebonakalayo endaweni yokuzikhethela okuningi iyatholakala ku-Visual Genome. Yakhiwe izithombe ze-MSCOCO ezingu-101,174 ezinamapheya e-QA ayizigidi ezingu-1.7, ngesilinganiso semibuzo engu-17 isithombe ngasinye.
Uma kuqhathaniswa nedathasethi Yempendulo Yemibuzo Ebonakalayo, isethi yedatha ye-Visual Genome inokusabalalisa okulungile kuzo zonke izinhlobo zemibuzo eyisithupha: Yini, Kuphi, Nini, Ubani, Kungani, Futhi Kanjani.
Ngaphezu kwalokho, isethi yedatha ye-Visual Genome ihlanganisa izithombe ezingu-108K ezimakwe kakhulu ngezinto, izakhiwo, nokuxhumana.
5. I-LibriSpeech
I-LibriSpeech corpus iqoqo lamahora angaba ngu-1,000 wama-audiobook avela kuphrojekthi ye-LibriVox. Iningi lama-audiobooks lisuka ku-Project Gutenberg.
Idatha yokuqeqeshwa ihlukaniswe yaba izingxenye ezintathu zamasethi angu-100hr, 360hr, kanye namasethi angu-500hr, kuyilapho idatha ye-dev neyokuhlola ilinganiselwa ku-5hr ngobude bomsindo.
6. Izindawo zokuhlala e- Cityspaces
Enye yedatha enkulu eyaziwa kakhulu yamavidiyo e-stereo anokubukwa kwasemadolobheni ibizwa ngokuthi i-Cityscapes.
Ngezichasiselo ezinembile ngephikseli ezihlanganisa izindawo ze-GPS, izinga lokushisa langaphandle, idatha ye-ego-motion, nemibono elungile ye-stereo, ihlanganisa ukurekhodwa okuvela emadolobheni ahlukene angama-50 aseJalimane.
7. Isethi yedatha ye-Kinetics
Enye yamasethi edatha evidiyo aziwa kakhulu yokubona imisebenzi yabantu ngezinga elikhulu futhi enekhwalithi enhle isethi yedatha ye-Kinetics. Okungenani kuneziqeshana zevidiyo ezingama-600 zekilasi ngalinye kwangama-600 okwenziwa abantu, afinyelela ku-500,000 esewonke.
Amafilimu akhishwe ku-YouTube; ngayinye ingamasekhondi ayi-10 ubude futhi inesigaba somsebenzi esisodwa kuphela esohlwini.
8. CelebMask-HQ
I-CelebAMask-HQ iqoqo lezithombe zobuso ezinesinqumo esiphezulu ezingu-30,000 ezinamamaski achazwe ngokucophelela namakilasi angu-19 afaka izingxenye zobuso njengesikhumba, ikhala, amehlo, izindlebe, izindlebe, umlomo, izindebe, izinwele, isigqoko, ingilazi yeso, icici, umgexo, intamo, impahla.
Idathasethi ingasetshenziselwa ukuhlola nokuqeqesha ukubonwa kobuso, ukucozulula ubuso, nama-GAN ekwenzeni ama-algorithms okudala ubuso nokuhlela.
9. I-Penn Treebank
Enye yezinkampani eziphawuleka kakhulu futhi evame ukusetshenziswa ekuhlolweni kwamamodeli wokumaka ngokulandelana yikhophasi ye-English Penn Treebank (PTB), ikakhulukazi ingxenye yekhophasi ehambisana nezindatshana ze-Wall Street Journal.
Igama ngalinye kufanele libe nengxenye yalo yenkulumo emakwe njengengxenye yomsebenzi. Izinga lomlingiswa kanye nezinga legama ukumodela ulimi futhi ngokuvamile isebenzisa ikhophasi.
10. VoxCeleb
I-VoxCeleb iyidathasethi enkulu yokuhlonza inkulumo ekhiqizwe ngokuzenzakalela kuyo imidiya yomthombo ovulekile. I-VoxCeleb inamazwi angaphezu kwesigidi avela kuzipikha ezingaphezu kuka-6k.
Njengoba idathasethi ihlanganisa okulalelwayo nokubukwayo, ingasetshenziselwa ezinye izinhlelo zokusebenza ezengeziwe, okuhlanganisa ukuhlanganisa kwenkulumo ebonakalayo, ukuhlukaniswa kwenkulumo, ukudluliswa kwezindlela ezihlukahlukene ukusuka ebusweni kuye ezwini noma okuphambene nalokho, nokuqeqesha ukuqashelwa kobuso kusukela kuvidiyo ukuze kwenezele ukuqaphela ubuso kwamanje. amasethi wedatha.
11. SIXray
Idathasethi ye-SIXray ihlanganisa izithombe ze-X-ray ezingu-1,059,231 eziqoqwe eziteshini zesitimela ezingaphansi futhi ezichasiswe abahloli bezokuphepha ukuze bathole izinhlobo eziyisithupha eziyinhloko zezinto ezingavunyelwe: izibhamu, imimese, izikliphu, amapliers, isikere, nezando. Ngaphezu kwalokho, amabhokisi okuhlanganisa ento ngayinye engavunyelwe engezwe mathupha kumasethi okuhlola ukuze kuhlolwe ukusebenza kokwenziwa kwasendaweni kwento.
12. Izingozi zase-US
Ingqikithi yephrojekthi isivele ivezwe ngegama ledathasethi, Izingozi zase-US. Le dathasethi yezingozi zezimoto ezweni lonke ihlanganisa ulwazi olusuka kuFebruwari 2016 kuya kuDisemba 2021 futhi ihlanganisa izifundazwe ezingu-49 e-USA.
Cishe amarekhodi ezingozi ayizigidi ezingu-1.5 akhona manje kuleli qoqo. Iqoqwe ngesikhathi sangempela ngokusebenzisa ama-API wethrafikhi ambalwa.
Lawa ma-API adlulisa ulwazi lwethrafikhi oluqoqwe kusuka emithonjeni eyahlukene, okuhlanganisa amakhamera omgwaqo, izinhlangano zomthetho, kanye ne-US kanye neminyango yezokuthutha yezwe.
13. Ukuqashelwa Kwezifo Zamehlo
Isizindalwazi esihleliwe se-ophthalmic Ocular Disease Intelligent Recognition (ODIR) siqukethe ulwazi ngeziguli ezingu-5,000, okuhlanganisa iminyaka yazo, umbala we-fundus emehlweni azo angakwesokunxele nangakwesokudla, kanye namagama angukhiye okuxilonga ochwepheshe bezokwelapha.
Le dathasethi iqoqo langempela ledatha yesiguli evela ezibhedlela nasezikhungweni zezokwelapha e-China etholwe yi-Shanggong Medical Technology Co., Ltd.. Nge ukuphathwa kokulawulwa kwekhwalithi, izichasiselo zimakwe abafundi abangabantu abanekhono.
14. Isifo senhliziyo
Le datha yesifo senhliziyo isiza ekuhlonzeni ukuba khona kwesifo senhliziyo esigulini okusekelwe emikhakheni engama-76 njengeminyaka yobudala, ubulili, uhlobo lwezinhlungu esifubeni, ukuphumula komfutho wegazi, nokunye.
Ngezimo ezingama-303, isizindalwazi sifuna ukumane sihlukanise ubukhona bokugula (inani elingu-1,2,3,4) nokungabikho kwaso (inani elingu-0).
15. I-CLEVR
Idathasethi ye-CLEVR (Ulimi Oludidiyelwe kanye Nokubonisana Okubalulekile Okubonakalayo) ilingisa Ukuphendulwa Kwemibuzo Ebonakalayo. Iqukethe izithombe zezinto ezihunyushwe nge-3D, nesithombe ngasinye siphelezelwa uchungechunge lwemibuzo eyakhiwe kakhulu ehlukaniswe izigaba ezimbalwa.
Kuzo zonke izithombe zesitimela nokuqinisekisa nemibuzo, isethi yedatha ihlanganisa izithombe ezingu-70,000 nemibuzo engu-700,000 yokuqeqeshwa, izithombe ezingu-15,000 nemibuzo engu-150,000 yokuqinisekiswa, kanye nezithombe ezingu-15,000 nemibuzo engu-150,000 yokuhlolwa ehilela izinto, izimpendulo, amagrafu esehlakalo, kanye namagrafu asebenzayo.
16. I-Universal Dependencies
Iphrojekthi ye-Universal Dependencies (UD) ihlose ukudala i-morphology efanayo yezilimi ezahlukene kanye nesichasiselo se-syntax treebank sezilimi eziningi. Inguqulo engu-2.7, eyakhululwa ngo-2020, inamabhange ezihlahla angu-183 ngezilimi ezingu-104.
Isichasiselo sakhiwe ngomaka be-POW bendawo yonke, amakhanda okuncika, namalebula okuncika kwendawo yonke.
17. I-KITTI - 360
Enye yedathasethi esetshenziswa kakhulu yamarobhothi eselula kanye ukushayela okuzimele iKITTI (Karlsruhe Institute of Technology kanye neToyota Technological Institute).
Yakhiwe izimo zethrafikhi ebiza amahora amaningi ezithathwe kusetshenziswa uhla lwezindlela zezinzwa, ezifana ne-high-resolution RGB, i-grayscale stereo, namakhamera we-3D laser scanner. Idathasethi iye yathuthukiswa ngokuhamba kwesikhathi ngabacwaningi abaningana abachasise ngokuzenzela izingxenye ezihlukahlukene zayo ukuze ivumelane nezidingo zabo.
18. I-MOT(Ukulandelela Izinto Eziningi)
I-MOT (Multiple Object Tracking) iyidathasethi yokulandelela izinto eziningi okufaka indawo yasendlini nengaphandle yezindawo zomphakathi ezihlanganisa abahamba ngezinyawo njengezinto ezithakaselwayo. Ividiyo yesigcawu ngasinye ihlukaniswe izingcezu ezimbili, eyodwa eyokuqeqeshelwa enye ngeyokuhlolwa.
Idathasethi ihlanganisa ukutholwa kwezinto kumafreyimu evidiyo asebenzisa izitholi ezintathu: i-SDP, i-Faster-RCNN, ne-DPM.
19. I-PASCAL 3D+
I-Pascal3D+ yokubukwa okuningi kwedathasethi yenziwe ngezithombe eziqoqwe endle, okungukuthi, izithombe zezigaba zezinto ezinokuhlukahluka okuphezulu, ezithwetshulwe ezimeni ezingalawuleki, ezindaweni eziminyene, nasezindaweni ezihlukahlukene. I-Pascal3D+ ihlanganisa izigaba zezinto eziqinile ezingu-12 ezithathwe kudathasethi ye-PASCAL VOC 2012.
Lezi zinto zinolwazi lokuma olumakwe kuzo (i-azimuth, ukuphakama, nebanga lekhamera). I-Pascal3D+ ihlanganisa nezithombe ezichasisekile ezivela eqoqweni le-ImageNet kulezi zigaba ezingu-12.
20. Amamodeli Ezilwane Ezikhubazekile Ebusweni
Umgomo wephrojekthi ye-Facial Deformable Models of Animals (FDMA) uwukubekela inselele izindlela zamanje zokuhlonza indawo eyingqopha-mlando yobuso bomuntu kanye nokusungula ama-algorithms amasha angabhekana nokuhlukahluka okukhulu okuwuphawu lwezimpawu zobuso bezilwane.
Ama-algorithms ephrojekthi abonise amandla okubona nokulandelela izimpawu zendawo ebusweni bomuntu kuyilapho kubhekwana nokuhlukahluka okubangelwa izinguquko emizweni yobuso noma ukuma, ukuvaleka okuncane, nokukhanya.
21. MPII Human Post Dataset
I-MPII Human Pose Dataset iqukethe cishe izithombe ezingu-25K, ezingu-15K zazo ezingamasampuli okuqeqesha, ezingu-3K zazo ezingamasampula okuqinisekisa, kanti ezingu-7K zazo ezingamasampula okuhlola.
Izikhundla zilebulwe ngokoqobo ngamajoyinti omzimba afinyelela kwangu-16, futhi izithombe zithathwe kumafilimu e-YouTube ahlanganisa izinto ezihlukahlukene zabantu ezingu-410.
22. UCF101
Isethi yedatha ye-UCF101 iqukethe iziqeshana zevidiyo eziyi-13,320 ezihlelwe ngezigaba eziyi-101. Lezi zigaba ezingu-101 zihlukaniswe izigaba ezinhlanu: ukunyakaza komzimba, ukusebenzisana komuntu nomuntu, ukuxhumana phakathi kwabantu, ukudlala izinsimbi zomculo, nezemidlalo.
Amavidiyo avela ku-YouTube futhi ahlanganisa amahora angama-27 ubude.
23. Isethi yomsindo
I-Audioset iyidathasethi yomcimbi womsindo eyakhiwe amasegimenti evidiyo ayimizuzwana eyi-2 achazwe ngabantu angaphezu kwezigidi ezimbili. Ukuchasisela le datha, i-ontology ye-hierarchical ehlanganisa izinhlobo zemicimbi engu-10 isetshenziswa, okusho ukuthi umsindo ofanayo ungase ulebulwe ngokuhlukile.
24. I-Stanford Natural Language Inference
Idathasethi ye-SNLI (I-Stanford Natural Language Inference) iqukethe ukubhanqa kwemisho engu-570k ehlelwe ngokuzenzela njengokubandakanya, ukungqubuzana, noma ukungathathi hlangothi.
Izakhiwo ziyizincazelo zezithombe ze-Flickr30k, kuyilapho imibono ecatshangelwayo yathuthukiswa izichasiselo ezivela emithonjeni yesixuku ezanikezwa isisekelo futhi zayalelwa ukuba zikhiqize izitatimende ezihlanganisayo, eziphikisanayo, nezingathathi hlangothi.
25. Ukuphendulwa Kwemibuzo Ebonakalayo
I-Visual Question Answering (VQA) iyidathasethi equkethe imibuzo evulekile mayelana nezithombe. Ukuze uphendule le mibuzo, udinga ukubamba umbono, ulimi, nokuqonda okuvamile.
Isiphetho
Njengoba ukufunda komshini nobuhlakani bokwenziwa (AI) kwanda kakhulu cishe kuwo wonke amabhizinisi nasezimpilweni zethu zansuku zonke, liyanda nenani lezisetshenziswa nolwazi olutholakala ngale ndaba.
Amasethi edatha asesidlangalaleni enziwe anikeza isiqalo esihle sokuthuthukisa amamodeli e-AI kuyilapho evumela abahleli bohlelo be-ML abanolwazi ukuba bonge isikhathi futhi bagxile kwezinye izici zamaphrojekthi abo.
shiya impendulo