When generating anchor boxes with $n$ scales $s_1, \ldots, s_n$ and $m$ aspect ratios $r_1, \ldots, r_m$, using every possible $(s_i, r_j)$ combination at each pixel would produce $whnm$ total anchor boxes, which is computationally prohibitive. In practice, only those pairings that include either the first scale $s_1$ or the first aspect ratio $r_1$ are retained:

$(s_1, r_1), (s_1, r_2), \ldots, (s_1, r_m), (s_2, r_1), (s_3, r_1), \ldots, (s_n, r_1)$

This yields $n + m - 1$ distinct anchor boxes per pixel and $wh(n + m - 1)$ anchor boxes for the entire image, dramatically reducing the computational burden while still providing sufficient shape diversity to cover most ground-truth objects.