Table of Contents[Hide][Show]
Ib qho ntawm cov tswv yim yooj yim tshaj plaws tab sis xav paub ntau tshaj plaws hauv kev kawm tob yog kev tshawb nrhiav khoom. Lub tswv yim tseem ceeb yog faib txhua yam khoom mus rau hauv cov chav kawm ua tiav uas sawv cev rau qhov zoo sib xws thiab tom qab ntawd kos ib lub thawv nyob ib puag ncig nws.
Cov yam ntxwv txawv no tuaj yeem yooj yim li daim ntawv lossis xim, uas pab peb muaj peev xwm los faib lawv.
Cov kev siv ntawm Kev Tshawb Fawb yog dav siv nyob rau hauv kev kho mob sciences, autonomous tsav tsheb, tiv thaiv thiab tub rog, pej xeem kev tswj hwm, thiab ntau lwm yam teb ua tsaug rau ntau yam kev txhim kho hauv Computer Vision thiab Image Processing.
Ntawm no peb muaj MMDetection, qhov zoo heev qhib qhov khoom tshawb nrhiav cov cuab yeej tsim los ntawm Pytorg. Hauv tsab xov xwm no, peb yuav tshuaj xyuas MMDetection kom ntxaws, mus koom tes nrog nws, sib tham txog nws cov yam ntxwv, thiab ntau ntxiv.
Yuav ua li cas yog MMDetection?
cov MMDetection toolbox tau tsim los ua Python codebase tshwj xeeb rau cov teeb meem cuam tshuam nrog kev txheeb xyuas cov khoom thiab piv txwv segmentation.
Kev siv PyTorch yog siv, thiab nws yog tsim nyob rau hauv ib tug modular zam. Rau kev paub txog cov khoom thiab cov piv txwv segmentation, ntau yam qauv zoo tau muab tso ua ke rau hauv ntau txoj kev.
Nws tso cai rau kev siv tau zoo thiab kev cob qhia sai. Ntawm qhov tod tes, lub thawv cuab yeej suav nrog qhov hnyav rau ntau tshaj 200 qhov kev cob qhia ua ntej, ua rau nws kho sai hauv cov khoom txheeb xyuas qhov chaw.
Nrog lub peev xwm los hloov kho cov txheej txheem tam sim no lossis tsim cov cuab yeej tshiab siv cov qauv muaj, MMDetection ua haujlwm raws li qhov ntsuas.
Lub toolbox tus yam ntxwv tseem ceeb yog nws suav nrog ntawm ncaj nraim, modular qhov chaw los ntawm ib txwm nrhiav pom khoom lub moj khaum uas tuaj yeem siv los tsim cov kav dej tshwj xeeb lossis cov qauv tshwj xeeb.
Lub peev xwm ua tau zoo ntawm cov cuab yeej no ua rau nws yooj yim los tsim cov txheej txheem ntsuas tshiab nyob rau sab saum toj ntawm cov txheej txheem uas twb muaj lawm thiab sib piv nws cov kev ua tau zoo.
nta
- Nrov thiab niaj hnub nrhiav kom pom lub hauv paus, xws li Faster RCNN, Mask RCNN, RetinaNet, thiab lwm yam, tau txais kev txhawb ncaj qha los ntawm cov khoom siv.
- Siv 360+ cov qauv ua ntej kev cob qhia rau kev kho kom zoo (lossis kev cob qhia tshiab).
- Rau cov ntaub ntawv tsis pom kev paub zoo xws li COCO, Cityscapes, LVIS, thiab PASCAL VOC.
- Ntawm GPUs, txhua qhov tseem ceeb bbox thiab daim npog ntsej muag raug tua. Lwm cov codebases, xws li Detectron2, maskrcnn-benchmark, thiab SimpleDet, tuaj yeem raug cob qhia ntawm tus nqi nrawm dua lossis sib npaug nrog qhov no.
- Cov kws tshawb fawb rhuav tshem cov nrhiav pom khoom lub moj khaum rau hauv ntau lub modules, uas tuaj yeem muab sib xyaw ua ke los tsim ib qho khoom tshwj xeeb nrhiav kom tau qhov system.
MMDetection Architecture
MMDetection qhia txog cov qauv tsim uas tuaj yeem siv rau txhua tus qauv vim nws yog lub thawv uas muaj ntau yam qauv ua ntej, txhua tus muaj nws tus kheej architecture. Cov nram qab no Cheebtsam ua li no tag nrho architecture:
- Tshaj pob: Cov pob txha, xws li ResNet-50 yam tsis muaj txheej txheem txuas nrog kawg, yog cov khoom siv uas hloov cov duab rau cov duab qhia chaw.
- caj dab: Lub caj dab yog ntu uas txuas lub nraub qaum mus rau lub taub hau. Ntawm cov pob txha pob txha daim duab qhia chaw, nws ua qee yam kev hloov kho lossis kho dua tshiab. Feature Pyramid Network yog ib qho piv txwv (FPN).
- DenseHead (AnchorHead/AnchorFreeHead): Nws yog cov khoom siv uas ua haujlwm ntawm qhov chaw ntom ntom ntawm cov duab qhia chaw, xws li AnchorHead thiab AnchorFreeHead, xws li RPNHead, RetinaHead, thiab FCOSHead.
- RoIextractor: Nrog rau kev siv RoIPooling-zoo li cov neeg ua haujlwm, nws yog ntu uas rub RoIwise nta los ntawm ib qho lossis ib qho ntawm cov ntawv qhia tshwj xeeb. Cov qauv SingleRoIextractor rho tawm RoI nta los ntawm qib sib piv ntawm cov pyramids feature.
- RoIHead (BBoxHead/MaskHead): Nws yog ib feem ntawm lub kaw lus uas siv RoI cov yam ntxwv raws li kev tawm tswv yim thiab tsim RoI-raws li kev ua haujlwm tshwj xeeb kev kwv yees, xws li bounding box classification/regression and mask prediction.
Kev tsim kho ntawm ib-theem thiab ob-theem detectors yog qhia siv cov ntsiab lus hais saum toj no. Peb tuaj yeem tsim peb tus kheej cov txheej txheem yooj yim los ntawm kev tsim ob peb qhov tshiab thiab sib txuas qee qhov uas twb muaj lawm.
Cov npe ntawm cov qauv suav nrog hauv MMDetection
MMDetection muab cov codebases saum toj kawg nkaus rau ntau tus qauv paub zoo thiab cov haujlwm ua haujlwm. Cov qauv uas tau ua yav dhau los thiab hloov tau txoj hauv kev uas yuav siv tau nrog MMDetection toolbox muaj npe hauv qab no. Daim ntawv yuav loj hlob raws li cov qauv thiab cov txheej txheem ntxiv.
- Ceev R-CNN
- Faster R-CNN
- Mask R-CNN
- RetinaNet
- DCN
- DCNv 2
- Cascade R-CNN
- M2 Det
- GHM
- ScratchDet
- Ob lub taub hau R-CNN
- Grid R-CNN
- FSAF
- Libra R-CNN
- GCNet
- HRNet
- Mask Scoring R-CNN
- FCOS
- SSD
- R-FCN
- Mixed Precision Training
- Qhov hnyav Standardization
- Hybrid Task Cascade
- Taw Qhia Anchoring
- Generalized Attention
Lub tsev kuaj pom tus qauv siv MMDetection
Hauv phau ntawv qhia no, peb yuav yog Google collab phau ntawv vim nws yooj yim rau kev teeb tsa thiab siv.
installation
Txhawm rau nruab txhua yam peb xav tau, peb yuav xub nruab cov tsev qiv ntawv tsim nyog thiab clone MMdetection GitHub qhov project.
Importing env
Ib puag ncig rau peb qhov project tam sim no yuav raug xa tawm los ntawm qhov chaw cia khoom.
Ntshuam cov tsev qiv ntawv thiab MMdetection
Tam sim no peb yuav import cov tsev qiv ntawv xav tau, nrog rau MMdetection tau kawg.
Download tau cov chaw kuaj xyuas ua ntej
Cov qauv kev cob qhia ua ntej los ntawm MMdetection yuav tsum tau rub tawm tam sim no rau kev hloov kho ntxiv thiab kev xav.
Lub tsev qauv
Tam sim no peb yuav tsim cov qauv thiab siv cov checkpoints rau cov dataset.
Inference tus ntes
Tam sim no hais tias tus qauv tau tsim kom zoo thiab thauj khoom, cia saib seb nws zoo npaum li cas. Peb siv MMDetection's high-level API inference detector. Qhov API no tau tsim los ua kom cov txheej txheem inference yooj yim dua.
tshwm sim
Cia peb saib cov txiaj ntsig.
xaus
Hauv kev xaus, MMDetection toolbox outperforms nyuam qhuav tso tawm codebases zoo li SimpleDet, Detectron, thiab Maskrcnn-benchmark. Nrog ib tug loj qauv sau,
MMDetection yog tam sim no thev naus laus zis thev naus laus zis. MMDetection outperforms tag nrho lwm cov codebases nyob rau hauv cov nqe lus ntawm efficiency thiab kev ua tau zoo.
Ib qho zoo tshaj plaws ntawm MMdetection yog tias tam sim no koj tuaj yeem taw tes rau cov ntaub ntawv teeb tsa sib txawv, rub tawm qhov chaw sib txawv, thiab khiav tib txoj cai yog tias koj xav hloov cov qauv.
Kuv qhia saib lawv cov lus qhia yog tias koj khiav mus rau hauv cov teeb meem nrog ib qho ntawm cov theem lossis xav ua qee yam ntawm lawv txawv.
Sau ntawv cia Ncua