Research on semantic segmentation method of stepless convolution based on DDRNet
This paper refers to the SPD-Conv module,which replaces the strided convolution layer by the module of spatial conversion depth and the stepless convolution layer,and preserves all the information in the channel dimension.Applying the SPD-Conv module in the DDRNet structure achieves better performance.A single RTX3090 GPU achieves 79.09%MIoU on the city dataset without pre-training.
DDRNetSPD-Convconvolution without stridemulti-resolution path
NETL
NSTL
万方数据
