Actual value of the offset regression loss

[YangJiao1996]

Hello there! I recently implemented the side-refinement part of CTPN on TensorFlow. During training, I noticed that the regression loss of side-refinement offset part (L_o^re) is relatively quite smaller than other parts. For example, L_v^re is often 100 to 200 times bigger than L_o^re. I wonder whether my implementation of the side-refinement part is right or not...

[hcnhatnam]

hmm...in my implementation of me, Lv(re) is often 50 to 100 times smaller than Lo(re). My way is as follows: Example: the yellow bbox is ground truth( gt), the black box(bl) is considering anchor and I calculate xside(of bl)=(x_leftside,x_rightside):(wa=16)

• x_leftside=d1/16(dark green arrow)
• x_rightside=d2/16(bright blue arrow) (d1,d2 is distance(x-axis) of gt center of anchor to the side left and right, respectively )