The　dance　generation　task　is　an　interesting　task.　The　general　task　is　to　predict　the　next　dance　sequence　through　a　given　initial　frame,　or　to　convert　from　music　to　dance　through　audio　information.　However,　these　works　cannot　control　the　direction　of　dance　generation　and　lack　suitable　ground　truth　to　evaluate　the　generated　results.　In　our　work,　we　give　the　start　frame　and　the　end　frame,　which　can　effectively　solve　the　above　problems,　and　can　also　be　understood　as　a　standardized　new　task.　Our　experimental　results　show　that　through　a　two-stream　CNN　network,　we　can　well　complete　the　new　task　that　we　proposed.　©　2020　Published　under　licence　by　IOP　Publishing　Ltd.