The problem of the M in the paper

[ckqsars]

It is a nice work. Thank you for your share. I have some problems about the M in your heat map. You use the sum of N Gaussian function object center point to represent the M value. However when the xy is same as the x^i_c and y^i_c the sum of M_xy is bigger than 1, because sum others. In your loss of heatmap the positive is M=1, other is negative. whether use max replace sum is more reasonable? could you explain when the heatmap is same as groundtruth, but the value of M is not 1?

[ifzhang]

Yes, max is more reasonable. When using sum, we need to add a clip option to keep the value between 0 and 1. In fact, the code is using max instead of sum: https://github.com/ifzhang/FairMOT/blob/1851158a1bc025da7e6cb839ddef9d14e33b404a/src/lib/utils/image.py#L140

[ret-1]

@ckqsars In fact, in CenterNet on which the detection branch is built upon, the authors said that max is used