Regarding Distance Threshold Optimization about light-reid HOT 4 CLOSED

DTO is only used in the validation stage, not the eval stage.

in the training stage, we first split training data into training and validation sets.
then,

1. train AiO with the training set.
2. process DTO with the validation set.
3. train AiO with the whole training set.
4. eval our CtF model (CtF = AiO + DTO's outputs).

So, the comparison is fair with other state-of-the-arts.

from light-reid.

@wangguanan OK, I see. But I have another question:

If I am going to use CtF in a generalized ReID settings, that is, data distribution of the training and testing sets are quite different, is the current DTO applicable? Or how to modify it to fit such settings?

Thanks for your suggestions!

from light-reid.

yes, the current DTO is applicable.
no need to change.

generalizable Re-ID is a very good setting, glad to see you apply CtF to it.

from light-reid.

Thank you very much.

from light-reid.