Tafole ea likateng[Pata][Bontša]
E 'ngoe ea mehopolo e bonolo empa e khahlang haholo thutong e tebileng ke ho lemoha lintho. Taba ea mantlha ke ho arola ntho e 'ngoe le e 'ngoe ka lihlopha tse latellanang tse emelang litšobotsi tse tšoanang ebe joale o taka lebokose ho e potoloha.
Litšobotsi tsena tse khethollang li ka ba bonolo joaloka sebōpeho kapa 'mala, e leng se thusang bokhoni ba rona ba ho li arola.
Likopo tsa Ho tseba ntho ba hiriloe haholo mahlaleng a bongaka, ho khanna ba ikemetseng, ts'ireletso le sesole, tsamaiso ea sechaba, le mafapha a mang a mangata ka lebaka la ntlafatso e kholo ho Computer Vision le Image Processing.
Mona re na le MDetection, sesebelisoa sa ho lemoha lintho tse bulehileng se bulehileng se hahiloeng ho Pytorch. Sehloohong sena, re tla hlahloba MDetection ka botlalo, re tsamaee le eona, re buisane ka likarolo tsa eona, le tse ling tse ngata.
ke eng MMDetection?
The MMDetection Lebokose la lithulusi le entsoe e le Python codebase ka ho khetheha bakeng sa mathata a amanang le boitsebahatso ba ntho le karohano ea mohlala.
Ts'ebetso ea PyTorch e sebelisoa, 'me e entsoe ka mokhoa oa modular. Bakeng sa temoho ea ntho le karohano ea mohlala, mefuta e mengata e fapaneng e sebetsang e hlophisitsoe ka mekhoa e fapaneng.
E lumella tlhahiso e sebetsang le koetliso e potlakileng. Ka lehlakoreng le leng, lebokose la lithulusi le kenyelletsa litekanyo tsa marang-rang a fetang 200 a koetlisitsoeng pele, e leng se etsang hore e lokisoe ka potlako lebaleng la ho tsebahatsa ntho.
Ka bokhoni ba ho ikamahanya le mekhoa ea hajoale kapa ho theha sebui se secha se sebelisang li-module tse fumanehang, MMDtection e sebetsa joalo ka benchmark.
Karolo ea bohlokoa ea lebokose la lithulusi ke ho kenyelletsa likarolo tse otlolohileng, tse modular tse tsoang ho tse tloaelehileng ho lemoha ntho moralo o ka sebelisoang ho theha liphaephe tse ikhethang kapa mehlala e ikhethang.
Bokhoni ba ho benchmark ba sephutheloana sena sa lithulusi bo etsa hore ho be bonolo ho aha moralo o mocha oa detector holim'a moralo o teng le ho bapisa ts'ebetso ea ona.
Likaroloana
- Mekhahlelo e tsebahalang le ea sejoale-joale ea ho lemoha, joalo ka Faster RCNN, Mask RCNN, RetinaNet, joalo-joalo, li tšehetsoa ka kotloloho ke sephutheloana sa lisebelisoa.
- Tšebeliso ea mefuta e 360+ e koetlisitsoeng esale pele bakeng sa tokiso e ntle (kapa koetliso bocha).
- Bakeng sa li-database tse tsebahalang tsa pono tse kenyelletsang COCO, Cityscapes, LVIS, le PASCAL VOC.
- Ho li-GPU, lits'ebetso tsohle tsa mantlha tsa bbox le mask lia etsoa. Li-codebases tse ling, joalo ka Detectron2, maskrcnn-benchmark, le SimpleDet, li ka koetlisoa ka lebelo le potlakileng ho feta kapa ka tsela e lekanang le ena.
- Bafuputsi ba qhekella ho lemoha ntho moralo oa li-module tse 'maloa, tse ka kopanngoang ho theha sistimi e ikhethang ea ho lemoha lintho.
MMDetection Architecture
MMDetection e totobatsa moralo o akaretsang o ka sebelisoang mofuteng ofe kapa ofe kaha ke lebokose la lithulusi le nang le mefuta e mengata e hahiloeng esale pele, e 'ngoe le e 'ngoe e nang le meralo ea eona. Likarolo tse latelang li etsa moralo ona ka kakaretso:
- Backbone: Mokokotlo, joalo ka ResNet-50 ntle le lera la ho qetela le hokahaneng ka botlalo, ke karolo e fetolelang setšoantšo hore se be le limmapa.
- Nko: Molala ke karolo e kopanyang mokokotlo le lihlooho. Ho limmapa tse tala tsa mokokotlo, e etsa litokiso kapa litlhophiso tse ling. Feature Pyramid Network ke papiso e le 'ngoe (FPN).
- DenseHead (AnchorHead/AnchorFreeHead): Ke karolo e sebetsang libakeng tse teteaneng tsa limmapa tsa likarolo, joalo ka AnchorHead le AnchorFreeHead, joalo ka RPNHead, RetinaHead, le FCOSHead.
- RoIExtractor: Ka ts'ebeliso ea li-operators tse kang RooIPooling, ke karolo e hulang likarolo tsa RoIwise ho tsoa ho e le 'ngoe kapa pokello ea limmapa tsa likarolo. Mohlala oa SingleRoIExtractor o ntša likarolo tsa RoI ho tloha boemong bo tšoanang ba liphiramide.
- RoIHead (BBoxHead/MaskHead): Ke karolo ea sistimi e sebelisang litšoaneleho tsa RoI joalo ka ho kenya letsoho mme e hlahisa likhakanyo tse ikhethileng tsa mosebetsi tse thehiloeng ho RoI, joalo ka ho arola mabokose a tlamang / ho fokotseha le ho bolela esale pele ka mask.
Kaho ea li-detectors tsa mokhahlelo o le mong le mekhahlelo e 'meli li bontšoa ho sebelisoa likhopolo tse boletsoeng ka holimo. Re ka iketsetsa mekhoa ea rona ka ho etsa likarolo tse 'maloa tse ncha le ho kopanya tse seng li ntse li le teng.
Lethathamo la mefuta e kenyellelitsoeng ho MDetection
MMDetection e fana ka li-codebase tsa boemo bo holimo bakeng sa mefuta e mengata e tsebahalang le li-module tse shebaneng le mosebetsi. Mehlala e kileng ea etsoa le mekhoa e feto-fetohang e ka sebelisoang le lebokose la lithulusi la MMDtection li thathamisitsoe ka tlase. Lenane le ntse le eketseha ha mehlala le mekhoa e ntse e eketsoa.
- Ka potlako R-CNN
- Ka potlako R-CNN
- Mask R-CNN
- RetinaNet
- DCN
- DCNv2
- Cascade R-CNN
- M2Det
- GHM
- ScratchDet
- Habeli-Hlooho R-CNN
- Grid R-CNN
- FSAF
- Libra R-CNN
- GCNet
- HRNet
- Mask Scoring R-CNN
- FCOS
- SSD
- R-FCN
- Koetliso ea Phapang e Kopantsoeng
- Boima ba Boima
- Hybrid Task Cascade
- Tataiso ea ho Anchoring
- Tlhokomelo e Akaretsang
Ho aha mofuta oa ho lemoha lintho ka MMDtection
Thutong ena, re tla ba bukana ea ho sebelisana ea Google hobane ho bonolo ho e beha le ho e sebelisa.
tlhomamiso
Ho kenya ntho e 'ngoe le e' ngoe eo re e hlokang, re tla qala ka ho kenya lilaebrari tse hlokahalang ebe re kopanya morero oa MMdetection GitHub.
Ho kenya env
Tikoloho ea morero oa rona joale e tla rekoa ho tsoa sebakeng sa polokelo.
Ho kenya lilaeborari le MMdetection
Joale re tla tlisa lilaebrari tse hlokahalang, hammoho le MMdetection ehlile.
Khoasolla libaka tsa tlhahlobo tse koetlisitsoeng esale pele
Libaka tsa tlhahlobo tsa mohlala tse koetlisitsoeng esale pele ho tsoa ho MMdetection joale li lokela ho khoasolloa bakeng sa tokiso le boithuto bo bong.
Moetso oa mohaho
Hona joale re tla haha mohlala mme re sebelise li-checkpoints ho dataset.
Inference detector
Kaha joale mohlala o se o hahiloe hantle ebile o laetsoe, a re hlahlobeng hore na o motle hakae. Re sebelisa sesebelisoa sa API sa boemo bo holimo sa MDetection. API ena e etselitsoe ho nolofatsa ts'ebetso ea boithuto.
Phello
Ha re shebeng liphetho.
fihlela qeto e
Qetellong, lebokose la lithulusi la MDetection le feta li-codebase tse sa tsoa lokolloa tse kang SimpleDet, Detectron, le Maskrcnn-benchmark. Ka pokello e kholo ea mohlala,
Hona joale MMDetection ke theknoloji ea morao-rao. MMDetection e feta li-codebase tse ling kaofela mabapi le ts'ebetso le ts'ebetso.
E 'ngoe ea lintho tse ntle ka MMdetection ke hore joale u ka supa faele e fapaneng ea tlhophiso, u khoasolle sebaka se fapaneng sa tlhahlobo, 'me u tsamaise khoutu e ts'oanang haeba u lakatsa ho fetola mefuta.
Ke eletsa ho sheba bona fuoa litaelo haeba u thulana le mathata ka mekhahlelo efe kapa efe kapa u lakatsa ho etsa tse ling tsa tsona ka tsela e fapaneng.
Leave a Reply