Channel-Spatial Dynamic Convolution : An Exquisite Omni-dimensional Dynamic Convolution
Zhilin Zhao, Ming Dong
Abstract
Most of the dynamic convolution algorithms adopted at this stage use the SE attention mechanism, but the attention mechanism, as a key part of dynamic convolution, has not attracted enough attention, and the relevant research is insufficient. In this paper, an exquisite ODConv which is called Channel-Spatial dynamic convolution is proposed. CSConv introduces the spatial attention module and the channel attention module into the ODConv in parallel, so that the convolution kernel pays more attention to the basic characteristics of the input and effectively improves the accuracy of the model and the efficiency of the convolution kernel. The experimental results show that CSConv achieves good results in the four datasets of ImageNet, COCO, HRRSD and DIOR.