Fig. 4
Download original image
Illustration of the cross-shaped window self-attention. A feature map of shape H W C is first linearly projected to K heads. This is then split into Group A and Group B to calculate horizontal and vertical self-attention in parallel. Finally, the outputs of the two groups are concatenated to obtain the cross-shaped window self-attention.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.