e set the centerof each default box to ( i+0.5|fk | , j+0.5|fk | ), where |fk| is the size of the k-th square featuremap, i, j ∈ [0, |fk|).
0.5 is center of grid add the value of the grid num and using size of feature map(total number of grid) for normalize.
spect ratios for the default boxes, and denote them as ar ∈{1, 2, 3, 12 , 13 }.
the use of 2 and 1//2 is considered width and height are reciprocal
smin is 0.2 and smax is 0.9
in the code not include conv4_3 (using scale 0.1) it start at conv7 and end in conv11 (5 scale)
total 6 scale
m feature mapsfor prediction.
the number of using feature maps in downscale.(not channel)
Lloc(x, l, g) =N∑i∈P os∑m∈{cx,cy,w,h}xkij smoothL1(lmi − ˆgmj )ˆgcxj = (gcxj − dcxi )/dwi ˆgcyj = (gcyj − dcyi )/dhiˆgwj = log( gwjdwi)ˆghj = log( ghjdh
predict result is the how to move the default box to the ground truth. calculate the distance between d and g with unit of box size, and then calculate the difference between predict result and the distance.
(but i don't know why use log to reduce the ratio between g and d box)
3 × 3 × p small kernelthat produces either a score for a category
the conv: 3*3*(p*(classtypes+4))
out boxes: h*w*(p*4) out class: h*w*(p*classtypes)