深度学习及其应用_ 课件 21-卷积网络视觉应用上.pptxVIP

ComputervisiontasksPARTiDeepLearninganditsApplicationSJTUDeepLearningLecture.1

Practices:DataAugmentationSJTUDeepLearningLecture.2

Practices:DataAugmentationChangethepixelswithoutchangingthelabelTrainontransformeddataVerywidelyusedHorizontalflipsRandomcrops/scalesColorjitter…EspeciallyusefulforsmalldatasetsSJTUDeepLearningLecture.3

Practices:DataAugmentationHorizontalflipsSJTUDeepLearningLecture.4

Practices:DataAugmentationRandomcrops/scalesTakeResNetasanexample:Training:samplerandomcrops/scalesPickrandomLinrange[256,480]Resizetrainingimage,shortside=LSamplerandom224*224patchesTesting:averageafixedsetofcropsResizeimageat5scales:{224,256,384,480,640}Foreachsize,use10224*224crops:4corners+center,and+flipsSJTUDeepLearningLecture.5

Practices:DataAugmentationColorjitterRandomlyjittercontrastSJTUDeepLearningLecture.6

Practices:TransferLearningSJTUDeepLearningLecture.7

Practices:TransferLearningSJTUDeepLearningLecture.8

Practices:TransferLearningSJTUDeepLearningLecture.9

Practices:TransferLearningSJTUDeepLearningLecture.10

ComputerVisionTasksClassificationClassification+LocalizationDetectionSegmentationSJTUDeepLearningLecture.11

SJTUDeepLearningLecture.12CowGrassSkyTreesThisimageisCC0publicdomainGrassCatSkyTreesLabeleachpixelintheimagewithacategorylabelDon’tdifferentiateinstances,onlycareaboutpixelsSemanticSegmentation

SJTUDeepLearningLecture.13FullimageExtractpatchClassifycenterpixelwithCNNCowCowGrassProblem:Veryinefficient!NotreusingsharedfeaturesbetweenoverlappingpatchesFarabetetal,“LearningHierarchicalFeaturesforSceneLabeling,”TPAMI2013PinheiroandCollobert,“RecurrentConvolutionalNeuralNetworksforSceneLabeling”,ICML2014SemanticSegmentation:

SlidingWindow

SJTUDeepLearningLecture.14Input:3xHxWConvolutions:DxHxWConvConvConvConvScores:CxHxWargma