The performance gap in SBD.

[mt-cly]

Hi, thanks for your work.
I do some work base on your Intseg recently, but I get [email protected]=8.8 in reproduction. The thing is, as you did not provide the test code for generating next click, I implement that following the description of your paper as follows:

def get_next_anno_point_prob(pred, gt, seq_points):
    fndist_map = distance_transform_edt(np.pad((gt == 1) & (pred == 0), ((1, 1), (1, 1)), 'constant'))[1:-1, 1:-1]
    fpdist_map = distance_transform_edt(np.pad((gt == 0) & (pred == 1), ((1, 1), (1, 1)), 'constant'))[1:-1, 1:-1]
    fndist_map[seq_points[:, 1], seq_points[:, 0]], fpdist_map[seq_points[:, 1], seq_points[:, 0]] = 0, 0
    neg_prob_map, pos_prob_map = fndist_map.reshape(-1), fpdist_map.reshape(-1)
    prob_map = np.max((neg_prob_map, pos_prob_map), axis=0)
    prob_map = prob_map/prob_map.sum()
    pnt_index = np.random.choice(len(prob_map), replace=False, p=prob_map)
    if_pos = 1 if neg_prob_map[pnt_index]>0 else 0 if pos_prob_map[pnt_index]>0 else -1
    assert if_pos != -1
    h, w = fndist_map.shape
    pt_next = (pnt_index%w, int(pnt_index/w), if_pos)
    return pt_next

I do not know what results in this gap. Have you any suggestions or can you kindly provide the code of evaluation? Thank you.