Can not reproduce the provided result, 5 point gap here

[dwy927]

Using the provided config(completely same), get the result:

Same test on the provided checkpoint, get the result:

There is a 5 points gap between them. @chenxuluo @xiaodongyang

Is there anyone met the same problem like me?

[dwy927]

Training on RTX-6000(8card, 8 sample per gpu), syncbn=True.

[LouisShen233]

For me is self-training&test amota: 58.3; test with provided model amota: 61.1. Without changing any parameters.

[chenxuluo]

The provided checkpoint is finetuned from the detection task. I will update more experiments.

[turboxin]

The provided checkpoint is finetuned from the detection task. I will update more experiments.

Hi, I'm wondering how to finetune the model on detection task and why would that influence the result so much? Thanks a lot!