Multiscale Meets Spatial Awareness: An Efficient Attention Guidance Network for Human Parsing

Figure 4

The structure of Attention SPP. Based on input features of resolution in AG-Net, we design three atrous branches with rates of 3, 7, and 13, respectively, and a global pooling branch to constitute our model. Besides, exploited from the input features, we propose a Spatial Attention branch to highlight the corresponding multiscale and spatial semantics. After the concatenation operation, we leverage a Channel Attention branch to optimize the feature-rich but redundant matrix in the channel level. The gray links indicate the flow of Mutual Attention operation imposed in SPP.