Canot reproduce visdrone result even using pretrained weights?

[dragen1860]

Dear all: I found lots of people can not reproduce the visdrone result even using pretrained weights. Here is my result:

when i decrease the threshold from 0.4 to 0.2, the detectioin result is still wrose:

none of any human beings detected out. That's really strange. anyone solved this bugs? thank you very much.

[yuluzhong]

I think it's because the number of channels in the pre-trained model he provided was insufficient.