The accuracy of the pre-trained model does not match the paper?

[why529913]

I use --headless, test with the pre-trained model on kitt6 dataset (image_kitti6_all_368.zip), and find that the accuracy does not match, I want to know which scene which dataset the pre-trained model is, whether to use headless, I want to reproduce the accuracy as much as possible. thanks.

[lhc991025]

This is the model trained by myself

Why is the author's pre trained model better, even better than what is written in the paper. Is my metric method written incorrectly?