caffe
caffe copied to clipboard
weiliu89
Calculation of default boxes min_sizes / max_sizes is not consistent with paper?
According to the paper, the naive approach to do the calculation of min_sizes and max_sizes in models/VGGNet/VOC0712Plus/SSD_512x512/ssd_pascal_512.py would be something similar to
min_dim = 512
mbox_source_layers = ['conv4_3', 'fc7', 'conv6_2', 'conv7_2', 'conv8_2', 'conv9_2', 'conv10_2']
min_ratio = 15
max_ratio = 90
n = len(mbox_source_layers)
np.linspace(min_ratio, max_ratio, n+1) * min_dim / 100.
array([ 76.8 , 131.6571, 186.5143, 241.3714, 296.2286, 351.0857, 405.9429, 460.8 ])
But how the calculation is done in the implementation leads to different values and code equivalent to
a = np.linspace(min_ratio, max_ratio, n-1)
a = np.concatenate((np.array([7]), a ,np.array([a[-1]+min_ratio])))
a * min_dim / 100.
array([ 35.84, 76.8 , 153.6 , 230.4 , 307.2 , 384. , 460.8 , 537.6 ])
Is there a reason for the different values?
@mvoelk Did you solve the problem? I can't find any theory suport the min_size setting. @weiliu89 is it min_size just a experience value or there are tricks about setting.Thanks!
