Get problem on pixel weight matrix

[JiSheng]

Hi everyone, i have read the PixelLink paper, but in a part of loss on pixel. Author use instance-balance cross entropy loss, and he said that

In detail, for a given image with N text instances, all instances are treated equally by giving a same weight to everyone of them, denoted as B_i in Equ. 2. For the i-th instance with area = S_i, every positive pixels within it have a weight of w_i = B_i / S_i .

Can anyone help me understand this part? because i don't know that the given image with N text instances is the pixel_cls_groundtruth mask or the pixel_cls_predicted_score_map.

And the same question when author said
W_pos_link(i, j, k) = W(i, j) ∗ (Ylink(i, j, k) == 1)
W_neg_link(i, j, k) = W(i, j) ∗ (Ylink(i, j, k) == 0),
where Y_link is the label matrix of links

Y_link is groundtruth or just predicted score-maps for k-th neighbor

Thanks you!