DeepLab: Semantic Images Segmentation with DCNNs,Atrous Convolution,and Fully connected CRFs
Abstract用DL做semanticimagesegmentation,本文有threemaincontributions:atrousconvolution:提高featuremap的分辨率,不增加参数的前提下增大感受野。AtrousSpatialPyramidPooling(ASPP):来对不同的尺度做分割.提高了效果。CRF:进一步提升localization的性能INTRODUCTIO