in this commit, 0385000 by some reason, was changed padding in last convolution. After this, convolutions, conv8_2_h, conv9_2_h has output blob size 5x5. But, last detection levels should be 3x3 and 1x1, like in original SSD. When it detect big face 100 pixels side or more, i get strange effect:

When i remove this padding, and set to 0, it fine working:

for what, was this commit made?