Skip to yearly menu bar Skip to main content


Poster

ShiftwiseConv: Small Convolutional Kernel with Large Kernel Effect

Dachong Li · li li · zhuangzhuang chen · Jianqiang Li

ExHall D Poster #405
[ ] [ Paper PDF ]
Sun 15 Jun 8:30 a.m. PDT — 10:30 a.m. PDT

Abstract: Largekernelsplayacrucialroleinenhancingtheperformanceofstandardconvolutionalneuralnetworks(CNNs),enablingCNNstooutperformtransformerarchitecturesincomputervision.ScalingupkernelsizehassignificantlycontributedtotheadvancementofCNNmodelslikeRepLKNet,SLaKandUniRepLKNet.However,therelationshipbetweenkernelsizeandmodelperformancevariesacrossthesework.Itimpliesthatlargekernelconvolutionmayinvolvehiddenfactorsthataffectmodelperformance.Insteadofmerelyincreasingthekernelsize,wereassesstheroleoflargeconvolutionsanddecomposethemintotwoseparatecomponents:extractingfeaturesatacertaingranularityandfusingfeaturesbymultiplepathways.Inthispaper,wecontributefromtwoaspects.1)Wedemonstratethat$3×3$convolutionscanreplacelargeconvolutionsinexistinglargekernelCNNstoachievecomparableeffects.2)Wedevelopamultipathlongdistancesparsedependencyrelationshiptoenhancefeatureutilization.Specifically,weintroducetheShiftwise(SW)convolutionoperator,apureCNNarchitecture.Inawiderangeofvisiontaskssuchasclassification,segmentationanddetection,SWsurpassesstateofthearttransformersandCNNarchitectures,includingSLaKandUniRepLKNet.Codeandallthemodelsat\urlhttps://anonymous.4open.science/r/shiftwiseConv8978.

Live content is unavailable. Log in and register to view live content