use provided checkpoint but get low MOTP

[liz2020]

I download the jde.1088x608.uncertainty.pt and use it to evaluate on the MOT16 train set. While the MOTA looks good, the MOTP is strangely low. Did anyone encounter the same issue?

[IwenLeeO]

hi @liz2020 do you train you own data ? and get the results as paper ?

[liz2020]

I was using the pre-trained model for this issue. I latter trained on other datasets and get pretty good results.

[IwenLeeO]

when training the model using author's data , the box loss, score loss is quite well , but the Id loss is really high. what's the reason ? thanks .

[IwenLeeO]

@liz2020 ,sorry to bother, when training the model , the id loss is always 19.6 after 20 epoch , however the score loss 0.006, the box loss is 0.007 , so how to fix the Id loss ？thanks.