No max_size for the conv_13/expand related prior box?
d-li14 opened this issue · comments
@chuanqi305 Thanks for your Caffe implementation.
I'm confused that in terms of the PriorBox layer relevant to conv_13/expand, there is only min_size=60, without max_size. Would you please explain why this setting is omitted? Thanks in advance.
@d-li14 the anchors of conv_13/expand is [(0.2, 1.0), (0.2, 2.0), (0.2, 0.5)],
0.2 = 60 / 300
1.0，2.0，0.5 is the aspect ratios.
Anchor number of the other layers is 6, that is:
min_size, aspect_ratio(1.0, 2.0, 0.5)
max_size, aspect_ratio(1.0, 2.0, 0.5)
Thanks @chuanqi305 ! Well, I see, in this scale, prior boxes with size \sqrt(s_k * s_{k+1}) is excluded, so max_size is useless, only the min_size is used to generate 3 prior boxes, unlike original VGG16-based network or other scales in MobileNetV2-based network with 4/6 prior boxes including aforementioned \sqrt(s_k * s_{k+1}), right?
@d-li14 right.
OK, thanks for your explanation. @chuanqi305
BTW, I'm interested in the script show_tensor.py that you mentioned in #1 , but haven't got what it is even after searching. Could you give some hint about that?