Figure 5.

All elements in $L_{kq\left(\mathrm{ID}\right)}=\left(L_{1q\left(\mathrm{ID}\right)},L_{2q\left(\mathrm{ID}\right)},\dots ,L_{hq\left(\mathrm{ID}\right)}\right)$ form the input sequence to the ID encoder. The self-attention mechanism draws inter-identity relationships to produce the output sequence, $B=\left(b_1,b_2,\dots ,b_h\right)$. $b_1,b_2,\dots ,b_h$ are the final hidden states that corresponds to enrolled identity 1, enrolled identity 2,…, enrolled identity $h$.