Questions about focal loss used, with comparison to CenterNet about onenet HOT 4 CLOSED

Hi~
The main difference of OneNet and CenterNet are:
(1) As you mentions, OneNet is labeling without Gaussian pre-processing.
(2) CenterNet labels the positive sample as the feature point closest to center of gt box (only location cost), OneNet labels the positive samples by MinCost(sum of classification cost and location cost).

from onenet.

Hi,

Thank you. I see that in Fig. 5 of your paper.
Sorry that I am still unclear about that

OneNet labels the positive samples by MinCost(sum of classification cost and location cost).

because you should prepare the positive samples before getting MinCost.

Since your positive samples are more discriminative than center of gt box, how did you decide their positions?
Did you used technical methods like attention map, or just labeled them by human decisions ?

from onenet.

Hi,

Thank you. I see that in Fig. 5 of your paper.
Sorry that I am still unclear about that

OneNet labels the positive samples by MinCost(sum of classification cost and location cost).

because you should prepare the positive samples before getting MinCost.

Since your positive samples are more discriminative than center of gt box, how did you decide their positions?
Did you used technical methods like attention map, or just labeled them by human decisions ?

The positive samples are decided by MinCost, instead of human annotations.

from onenet.

Hi,
I still cannot understand how to calculate the sum of classification cost and location cost of each pixel ?

from onenet.