Some questions about the results in Table 4 in the paper

[chensnathan]

Hi, thanks for your excellent work.

In the latest version of the paper, as illustrated in Table 4, the results of using the center-ness computed from the predicted regression vector are worse than those of using the proposed center-ness branch. If I understand the statement correctly, "using the center-ness computed from the predicted regression vector" means to predict a HxWx5 feature map in the regression branch, in which the first 4 channels are (l; t; r; b), and the last channel is the center-ness. So the parameters of the last conv layer in the regression branch should be a (5, 256, 3, 3) weight and a (5, ) bias.

In Table 3, you move the center-ness branch to the regression branch and obtain further improvements. Under the ctr. on reg assumption, in the regression branch, the parameters of the last layer contain two parts: 1. the params in the regression target branch, which should be a (4, 256, 3, 3) weight and a (4, ) bias; 2. the params in the center-ness target branch, which should be a (1, 256, 3, 3) weight and (1, ) bias.

In fact, the two settings mentioned above are equal from both the params view and the loss optimization view. I am confused about the difference between the results (33.5 vs 37.4) shown in the paper. Could you explain why there is a large gap between the results? Thanks in advance.

[tianzhi0549]

@chensnathan you misunderstand it. "using the center-ness computed from the predicted regression vector" means that we do not use a separate head for center-ness and compute it with the predicted (l, t, b, r) instead.

[chensnathan]

ok, I see. Thanks for the explanation. It solves my problem.