I want to knonw MOTR's performance on MOT17 train dataset because I got a very high MOTA

[Soulmate7]

Hi! Congratulations for your nice work.
When I try to use the supported model to run eval.py, I get a very shocking results:

Then I use my trained model which runs 70+ epochs on crowdhuman and mot17 to eval and get the following result:

I am shocked by it's performance on the train dataset, so I wonder know is it normal?

[YannX1e]

Hello. I try to follow the MOTR work, and I also get the same shocking MOTA, have you find the reason?

Hi! Congratulations for your nice work. When I try to use the supported model to run eval.py, I get a very shocking results:

Then I use my trained model which runs 70+ epochs on crowdhuman and mot17 to eval and get the following result:

I am shocked by it's performance on the train dataset, so I wonder know is it normal?

[Soulmate7]

Not yet, exactly I don’t follow MOTR now. If you find the reason, we could have a discussion maybe.

…

--------------原始邮件-------------- 发件人："YannX1e ***@***.***>; 发送时间：2023年4月27日(星期四) 晚上8:54 收件人："megvii-research/MOTR" ***@***.***>; 抄送："Rui ***@***.***>;"Author ***@***.***>; 主题：Re: [megvii-research/MOTR] I want to knonw MOTR's performance on MOT17 train dataset because I got a very high MOTA (Issue #60) ----------------------------------- Hello. I try to follow the MOTR work, and I also get the same shocking MOTA, have you find the reason? Hi! Congratulations for your nice work. When I try to use the supported model to run eval.py, I get a very shocking results: Then I use my trained model which runs 70+ epochs on crowdhuman and mot17 to eval and get the following result: I am shocked by it's performance on the train dataset, so I wonder know is it normal? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[TerranLord]

The high performance on MOT17 training set may be attributed to overfitting, as MOT17 is a small dataset with only 5k frames. Therefore, many works use additional data such as CrowdHuman to mitigate this issue.

A possible reason for the performance gap is the short training iterations (only 70 epochs).

Hi! Congratulations for your nice work. When I try to use the supported model to run eval.py, I get a very shocking results:

Then I use my trained model which runs 70+ epochs on crowdhuman and mot17 to eval and get the following result:

I am shocked by it's performance on the train dataset, so I wonder know is it normal?