Zviri Mukati[Viga][Ratidza]
Imwe yemazano akareruka asi anonyanya kunakidza mukudzidza kwakadzama ndeyekuona chinhu. Pfungwa yakakosha ndeyekukamura chinhu chimwe nechimwe mumakirasi anoteedzana anomiririra hunhu hwakafanana uye wozodhirowa bhokisi rakaitenderedza.
Izvi zvinosiyanisa hunhu zvinogona kuve nyore sechimiro kana ruvara, izvo zvinobatsira mukugona kwedu kuzvironga.
Zvikumbiro zve Object Detection vanoshandirwa zvakanyanya musainzi yekurapa, kutyaira vakazvimirira, kudzivirira uye mauto, hutongi hweveruzhinji, uye nezvimwe zvakawanda minda nekuda kwekuvandudzwa kwakakura muComputer Vision uye Image Processing.
Pano isu tine MMDetection, inonakidza yakavhurika-sosi yekuona chinhu chishandiso chakavakirwa paPytorch. Muchinyorwa chino, isu tichaongorora MMDtection zvakadzama, enda maoko-mberi nayo, kurukura maitiro ayo, uye zvimwe zvakawanda.
Chii MMDetection?
The MMDetection bhokisi rematurusi rakagadzirwa sePython codebase yakanangana nezvinetso zvinosanganisira kuzivikanwa kwechinhu uye muenzaniso segmentation.
Kuitwa kwePyTorch kunoshandiswa, uye inogadzirwa mune modular fashoni. Nekuzivikanwa kwechinhu uye kupatsanurwa kwemuenzaniso, mhando dzakasiyana dzemhando dzinoshanda dzakaunganidzwa mumhando dzakasiyana dzemitoo.
Inobvumira kufungidzira kunoshanda uye nekukurumidza kudzidziswa. Nekune rimwe divi, bhokisi rematurusi rinosanganisira huremu hweanopfuura mazana maviri-asati adzidziswa network, zvichiita kuti igadzirise nekukurumidza mumunda wekuzivisa chinhu.
Nekugona kugadzirisa hunyanzvi hwazvino kana kugadzira detector nyowani uchishandisa mamodule aripo, MMDtection inoshanda sebhenji.
Chinhu chakakosha chebhokisi rezvishandiso kusanganisa kwayo kwakatwasuka, modular zvikamu kubva kune zvakajairika kuona kwechinhu chimiro chinogona kushandiswa kugadzira mapaipi akasiyana kana mhando dzakasiyana.
Kugona kwekuenzanisa kweiyi turusi rekushandisa kunoita kuti zvive nyore kuvaka detector framework pamusoro peiyo iripo uye kuenzanisa kuita kwayo.
Features
- Yakakurumbira uye yemazuva ano maitiro ekuona, akadai seFaster RCNN, Mask RCNN, RetinaNet, nezvimwewo, inotsigirwa zvakananga neturusi rekushandisa.
- Kushandiswa kwe360+ pre-yakadzidziswa modhi yekumisikidza (kana kudzidziswa patsva).
- Kune anonyatsozivikanwa ekuona datasets anosanganisira COCO, Cityscapes, LVIS, uye PASCAL VOC.
- PaGPUs, ese akakosha bbox uye mask mashandiro anoitwa. Mamwe macodebases, akadai seDetectron2, maskrcnn-benchmark, uye SimpleDet, anogona kudzidziswa nekukurumidza kupfuura kana padivi neiyi.
- Vatsvakurudzi vanoputsa kuona kwechinhu framework mumamodule akati wandei, ayo anogona kuzobatanidzwa kugadzira yakasarudzika yekuona chinhu system.
MMDetection Architecture
MMDetection inotsanangura dhizaini yegeneric inogona kuiswa kune chero modhi sezvo iri bhokisi rematurusi rine akasiyana-siyana ekare-akavakwa modhi, imwe neimwe ine dhizaini yayo. Izvi zvinotevera zvikamu zvinogadzira iyi yakazara architecture:
- Backbone: Backbone, senge ResNet-50 isina iyo yekupedzisira yakabatana zvizere layer, ndicho chikamu chinoshandura chifananidzo kuita mepu.
- Neck: Mutsipa ndicho chikamu chinobatanidza musana nemisoro. Pamamepu ekumusana ane mepu, inoita zvimwe zvigadziriso kana kugadzirisa zvakare. Feature Pyramid Network ndiyo imwe mufananidzo (FPN).
- DenseHead (AnchorHead/AnchorFreeHead): Ndicho chikamu chinoshanda munzvimbo dzakasimba dzemamepu ezvimiro, senge AnchorHead uye AnchorFreeHead, seRPNHead, RetinaHead, uye FCOSHead.
- RoIExtractor: Nekushandiswa kweRoIPooling-sevashandisi, ndicho chikamu chinodhonza RoIwise maficha kubva kune imwechete kana muunganidzwa wemamepu emhando. Iyo SingleRoIExtractor sampuli inobvisa RoI maficha kubva padanho rinoenderana remapiramidhi.
- RoIHead (BBoxHead/MaskHead): Icho chikamu chegadziriro inoshandisa maitiro eRoI seyekupinza uye inogadzira RoI-yakavakirwa basa-yakananga kufanotaura, sekusunga bhokisi kupatsanurwa/kudzoreredza uye kufanotaura kwemaski.
Kuvakwa kwe-single-stage uye maviri-stage detectors inoratidzirwa uchishandisa pfungwa dzambotaurwa. Isu tinokwanisa kugadzira yedu pachedu maitiro nekungogadzira mashoma mashoma ezvikamu uye nekubatanidza mamwe aripo.
Rondedzero yemamodheru anosanganisirwa muMMDetection
MMDetection inopa yepamusoro-notch codebases kune akati wandei anozivikanwa modhi uye basa-rinotungamirwa module. Mhando dzakambogadzirwa uye nzira dzinochinjika dzinogona kushandiswa neMMDetection bhokisi rekushandisa dzakanyorwa pazasi. Rondedzero inoramba ichikura sezvo mamwe mamodheru uye nzira dzinowedzerwa.
- Fast R-CNN
- Inokurumidza R-CNN
- Mask R-CNN
- RetinaNet
- DCN
- DCNv2
- Cascade R-CNN
- M2Det
- GHM
- ScratchDet
- Kaviri-Musoro R-CNN
- Grid R-CNN
- FSAF
- Libra R-CNN
- GCNet
- HRNet
- Mask Scoring R-CNN
- FCOS
- SSD
- R-FCN
- Mixed Precision Training
- Weight Standardization
- Hybrid Task Cascade
- Guided Anchoring
- Generalized Attention
Kuvaka chinhu chekuona modhi uchishandisa MMDtection
Muchidzidzo ichi, tichava Google collab notebook nekuti iri nyore kuseta nekushandisa.
kugadzwa
Kuti tiise zvese zvatinoda, isu tichatanga taisa maraibhurari anodiwa uye kutevedzera iyo MMdetection GitHub chirongwa.
Kutora kunze env
Zvakatipoteredza zvepurojekiti yedu zvino zvave kutengwa kubva kune inochengeterwa.
Kupinza maraibhurari uye MMdetection
Iye zvino tichaunza kunze maraibhurari anodiwa, pamwe neMMdetection hongu.
Dhawunirodha nzvimbo dzekutarisa dzakafanodzidziswa
Iwo pre-akadzidziswa modhi yekutarisa kubva kuMMdetection inofanirwa kudhawunirodha kuti iwedzere kugadziridzwa uye inference.
Kuvaka muenzaniso
Iye zvino tichavaka modhi uye toisa macheki ku dataset.
Inference the detector
Iye zvino iyo modhi yakanyatso kuvakwa uye kurodha, ngatitarisei kuti yakanaka sei. Isu tinoshandisa MMDetection's yepamusoro-level API inference detector. Iyi API yakagadzirirwa kuita kuti inference process iite nyore.
mugumisiro
Ngationei zvabuda.
mhedziso
Mukupedzisa, iyo MMDetection bhokisi rekushandisa inodarika ichangobva kuburitswa macodebases seSimpleDet, Detectron, uye Maskrcnn-benchmark. Nemuunganidzwa wakakura womuenzaniso,
MMDetection ikozvino yave-ye-the-art tekinoroji. MMDetection inokunda mamwe ese macodebases maererano nekubudirira uye kuita.
Chimwe chezvinhu zvakanakisa nezveMMdetection ndechekuti iwe unogona kungonongedza kune imwe faira yekumisikidza, dhawunirodha imwe yekutarisa, uye mhanyisa kodhi imwechete kana uchida kushandura mamodheru.
Ndinokurudzira kutarisa kwavo mirayiridzo kana iwe ukasangana nematambudziko nechero nhanho kana uchida kuita mamwe acho zvakasiyana.
Leave a Reply