Although skeleton-based action recognition has achieved great success in
recent years, most of the existing methods may suffer from a large model size
and slow execution speed. To alleviate this issue, we analyze skeleton sequence
properties to propose a Double-feature Double-motion Ne