Okuqukethwe[Fihla][Bonisa]
Omunye wemibono elula kodwa ehlaba umxhwele kakhulu ekufundeni okujulile ukutholwa kwezinto. Umqondo obalulekile uwukuhlukanisa into ngayinye ibe amakilasi alandelanayo amelela izici ezifanayo bese udweba ibhokisi elizungezayo.
Lezi zici ezihlukanisayo zingaba lula njengefomu noma umbala, okusiza ekhonweni lethu lokuzihlukanisa ngokwezigaba.
Izicelo ze Ukutholwa Kwento zisetshenziswa kabanzi kusayensi yezokwelapha, ukushayela okuzimele, ezokuvikela nezempi, ezokuphatha umphakathi, kanye neminye imikhakha eminingi ngenxa yentuthuko enkulu ku-Computer Vision and Image Processing.
Lapha sine-MMDetection, ithuluzi lokuthola into evulekile elimangalisayo elakhelwe ku-Pytorch. Kulesi sihloko, sizohlola iMMDetection ngokuningiliziwe, sihambisane nayo, sixoxe ngezici zayo, nokunye okuningi.
Kuyini I-MMDetection?
The I-MMDetection ibhokisi lamathuluzi ladalwa njenge-codebase ye-Python eqondene ngqo nezinkinga ezibandakanya ukuhlonza into nokuhlukaniswa kwezibonelo.
Ukuqaliswa kwe-PyTorch kuyasetshenziswa, futhi kudalwa ngendlela eyimodular. Ukuze uthole ukuqashelwa kwento kanye nokuhlukaniswa kwezibonelo, izinhlobonhlobo eziningi zamamodeli asebenzayo zihlanganiswe zibe izindlela ezihlukahlukene.
Ivumela ukucabanga okusebenzayo nokuqeqeshwa okusheshayo. Ngakolunye uhlangothi, ibhokisi lamathuluzi lihlanganisa izisindo zamanethiwekhi angaphezu kuka-200 aqeqeshwe kusengaphambili, okulenza libe ukulungisa ngokushesha endaweni yokuhlonza into.
Ngekhono lokujwayela amasu amanje noma ukwakha umtshina omusha usebenzisa amamojula atholakalayo, i-MMDetection isebenza njenge-benchmark.
Isici esiyinhloko sebhokisi lamathuluzi ukufakwa kwalo kwezingxenye eziqondile, eziyi-modular kusuka kokujwayelekile ukutholwa kwento uhlaka olungasetshenziswa ukudala amapayipi ahlukile noma amamodeli ahlukile.
Amakhono okulinganisa ale khithi yamathuluzi akwenza kube lula ukwakha uhlaka lomtshina olusha phezu kohlaka olukhona bese uqhathanisa nokusebenza kwalo.
Izici
- Izinhlaka zokutholwa ezidumile nesimanje, ezifana ne-Faster RCNN, Mask RCNN, RetinaNet, njll., zisekelwa ngokuqondile ikhithi yamathuluzi.
- Ukusetshenziswa kwamamodeli angu-360+ aqeqeshwe kusengaphambili ukuze alungiswe kahle (noma ukuqeqeshwa kabusha).
- Okwamasethi edatha ombono awaziwayo afaka i-COCO, i-Cityscapes, i-LVIS, ne-PASCAL VOC.
- Kuma-GPU, yonke imisebenzi eyisisekelo ye-bbox ne-mask isetshenziswa. Amanye ama-codebases, afana ne-Detectron2, imaskrcnn-benchmark, kanye ne-SimpleDet, angaqeqeshwa ngenani elisheshayo kunaleli noma elihambisana naleli.
- Abacwaningi bahlukanisa ukutholwa kwento uhlaka lube amamojula amaningana, angase ahlanganiswe ukuze kwakhe isistimu yokuthola into eyingqayizivele.
I-MMDetection Architecture
I-MMDetection icacisa umklamo ojwayelekile ongasetshenziswa kunoma iyiphi imodeli njengoba iyibhokisi lamathuluzi elinamamodeli ahlukahlukene akhiwe ngaphambilini, ngayinye enezakhiwo zayo. Izingxenye ezilandelayo zakha lesi sakhiwo sezakhiwo:
- backbone: Umgogodla, njenge-ResNet-50 engenasendlalelo sokugcina esixhunywe ngokugcwele, ingxenye eguqula isithombe sibe namamephu.
- Neck: Intamo ingxenye ehlanganisa umgogodla namakhanda. Kumamephu wesici esingavuthiwe somgogodla, yenza izinguquko ezithile noma ukulungisa kabusha. I-Feature Pyramid Network ingumfanekiso owodwa (FPN).
- I-DenseHead (I-AnchorHead/AnchorFreeHead): Iyingxenye esebenza ezindaweni eziminyene zamamephu wesici, njenge-AnchorHead ne-AnchorFreeHead, njenge-RPNHead, i-RetinaHead, ne-FCOSHead.
- I-RoIExtractor: Ngokusetshenziswa kwama-opharetha afana ne-RoIPooling, yisigaba esidonsa izici ze-RoIwise kusukela kweyodwa noma iqoqo lesici samamephu. Isampula ye-SingleRoIExtractor ikhipha izici ze-RoI kusukela kuleveli yokufanisa yamaphiramidi esici.
- I-RoIHead (BBoxHead/MaskHead): Yingxenye yesistimu esebenzisa izici ze-RoI njengokufakwayo futhi ikhiqize izibikezelo eziqondene nomsebenzi othile ezisekelwe ku-RoI, njengokuhlukaniswa kwebhokisi elibophayo/ukuhlehla nokubikezela kwemaski.
Ukwakhiwa kwemitshina yesiteji esisodwa kanye nezigaba ezimbili kuboniswa kusetshenziswa imiqondo eshiwo ngenhla. Singathuthukisa ezethu izinqubo ngokwakha izingxenye ezimbalwa ezintsha futhi sihlanganise ezinye ezikhona kakade.
Uhlu lwamamodeli afakwe ku-MMDetection
I-MMDetection ihlinzeka ngezisekelo zekhodi ezisezingeni eliphezulu zamamodeli amaningana aziwayo namamojula agxile emsebenzini. Amamodeli enziwe ngaphambilini nezindlela eziguquguqukayo ezingase zisetshenziswe ngebhokisi lamathuluzi le-MMDetection zibalwe ngezansi. Uhlu lulokhu lukhula njengoba kunezelwa amamodeli nezindlela.
- I-R-CNN esheshayo
- I-R-CNN esheshayo
- Imaski R-CNN
- I-RetinaNet
- I-DCN
- I-DCNv2
- I-Cascade R-CNN
- I-M2Det
- I-GHM
- I-ScratchDet
- I-Double-Head R-CNN
- Igridi R-CNN
- I-FSF
- I-Libra R-CNN
- I-GCNet
- I-HRNet
- Imaski Amaphuzu R-CNN
- I-FCOS
- I-SSD
- I-R-FCN
- Ukuqeqeshwa Okunembayo Okuxubile
- Isisindo Esijwayelekile
- I-Hybrid Task Cascade
- I-Anchoring Eqondiswayo
- Ukunaka Okujwayelekile
Ukwakha imodeli yokuthola into usebenzisa i-MMDetection
Kulesi sifundo, sizoba incwadi yokubhalela ye-Google ngoba kulula ukusethwa nokusetshenziswa.
Ukufakwa
Ukufaka yonke into esiyidingayo, sizoqala ngokufaka imitapo yolwazi edingekayo futhi sihlanganise iphrojekthi ye-MMdetection GitHub.
Ingenisa i-env
Imvelo yephrojekthi yethu manje izongeniswa ivela endaweni yokugcina.
Ingenisa imitapo yolwazi kanye ne-MMdetection
Manje sizongenisa imitapo yolwazi edingekayo, kanye ne-MMdetection kunjalo.
Landa izindawo zokuhlola eziqeqeshwe kusengaphambili
Izindawo zokuhlola eziyimodeli eziqeqeshwe kusengaphambili ezivela kwa-MMdetection manje kufanele zilandwe ukuze zilungiswe futhi kucatshangwe.
Imodeli yokwakha
Manje sizokwakha imodeli futhi sisebenzise izindawo zokuhlola kudathasethi.
Inkomba umtshina
Manje njengoba imodeli yakhiwe ngendlela efanele futhi yalayishwa, ake sihlole ukuthi inhle kangakanani. Sisebenzisa i-MMDetection's high-level API detector. Le API yakhelwe ukwenza inqubo ye-inference ibe lula.
Result
Ake sibheke imiphumela.
Isiphetho
Sengiphetha, ibhokisi lamathuluzi le-MMDetection lisebenza kahle kakhulu kunama-codebase asanda kukhishwa afana ne-SimpleDet, i-Detectron, ne-Maskrcnn-benchmark. Ngeqoqo elikhulu lamamodeli,
I-MMDetection manje isiwubuchwepheshe besimanje. I-MMDetection idlula zonke ezinye izisekelo zekhodi ngokusebenza kahle nokusebenza.
Enye yezinto ezinhle kakhulu nge-MMdetection ukuthi ungakwazi manje ukukhomba ifayela elihlukile lokumisa, ulande indawo yokuhlola ehlukile, bese usebenzisa ikhodi efanayo uma ufisa ukushintsha amamodeli.
Ngiyeluleka ngibheke kwabo imiyalelo uma uhlangabezana nezinkinga nganoma yisiphi izigaba noma ufisa ukwenza ezinye zazo ngendlela ehlukile.
shiya impendulo