Algorithm 1: Cyclic Attention $\mathit{CycAttn}\mathbf{\left(}{\mathit{S}}_{\mathit{l}}\mathbf{\right)}$

Input: Partial limb motion features $S_l=\{S_l^1,S_l^2,\cdots ,S_l^K\}$

for each $i\in [1,K]$ do:

1. Calculate: $S_{rest}=Concatenate(S_l-\left\{S_l^i\right\})$;

2. Calculate Query, Key and Value: $S_{query}=W_q{S}_l^i$, $S_{key}=W_k{S}_{rest}$, $S_{value}=W_v{S}_{rest}$;

3. Calculate $f_l^i=CrossAttn(S_{query},S_{key},S_{value})=SoftMax\left({\displaystyle \frac{S_{query}{S}_{key}^{\mathrm{T}}}{\sqrt{d}}}\right)S_{value}$;

4. $i\leftarrow i+1$.

end

where $Concatenate$ means splice partial limb motion features other than $S_l^i$, $W_q$, $W_k$,

and $W_v$ denote the projection weights, d is the channel dimension of $S_{query}$.

Output: Interacted local features $f_l=\{f_l^1,f_l^2,\cdots ,f_l^K\}$.