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

Figure 2

An overview of the Attention Guidance Network (AG-Net). Our network is established upon an Encoder-Decoder architecture extracting features from four resolution layers (i.e., , , , ), and the Attention RefineNet learns the attention score maps to optimize the predicted label results. We impose the Attention SPP into the end of the encoder and leverage a multiscale supervision strategy to refine our model.