Align paper results by myself

[LuPaoPao]

Thank you for your outstanding contributions. Will more training models and code be made public in the future? For example, a comprehensive comparison with UniAD.

I tried using a 1/8 pre-trained model in 1/4 fine-tune, according to the config you provided, and then pre-trained on 8*A100, but got results that were different from the return results. mAP：35.92 vs 36.90 NDS: 45.43 vs 45.77。Is this result a normal range of random fluctuations?

20240317_074726.log.json

It would be helpful to provide more pre-trained models and fine-tuned code. Because I want to continue to develop this paper, then I need a Baseline of repeatability.

Thank you again for your excellent paper.

[tomztyang]

Hi,

The results are weird. I have run the provided code 3 times, and their results are (NDS/mAP): 0.4583 / 0.3677; 0.4563 / 0.3704; and 0.4577 / 0.3690, respectively. The variance is around +-0.3 in my cases.

I have no idea what happens in your case for now. Maybe you can try the provided checkpoint as pre-training first?

Thanks, Zetong

[LuPaoPao]

Hi,

The results are weird. I have run the provided code 3 times, and their results are (NDS/mAP): 0.4583 / 0.3677; 0.4563 / 0.3704; and 0.4577 / 0.3690, respectively. The variance is around +-0.3 in my cases.

I have no idea what happens in your case for now. Maybe you can try the provided checkpoint as pre-training first?

Thanks, Zetong

Thank you very much

[LuPaoPao]

Hi, The results are weird. I have run the provided code 3 times, and their results are (NDS/mAP): 0.4583 / 0.3677; 0.4563 / 0.3704; and 0.4577 / 0.3690, respectively. The variance is around +-0.3 in my cases. I have no idea what happens in your case for now. Maybe you can try the provided checkpoint as pre-training first? Thanks, Zetong

Thank you very much

I'll try again

[tomztyang]

@LuPaoPao

Ping me if you need the logs for all these 3 attempts, as well as any further updates.

Thanks, Zetong