liyuke65535
commented
8 months ago Hi, thank you for your attention.
Q1&Q2: Duke is not available any more.
A: Many of our experiments were conducted on duke, but we did notice the privacy issue. Therefore, in the version published on cvf, we deleted results that are related to duke. At the same time, we remained these results on arxiv. So they're actually two different versions. Check no-duke-version at https://openaccess.thecvf.com/content/ICCV2023/papers/Ni_Part-Aware_Transformer_for_Generalizable_Person_Re-identification_ICCV_2023_paper.pdf.
Q3: Why different hyper-parameter settings?
A: It's correct that DG requires a trained model to generalize to arbitrary unseen domain, so the hyper-parameters are supposed to be constant. However, the hyper-parameters only change when we use different source domains, which have nothing to do with target domains. In other words, if we use A as source, we will only train one model to generalize on E, F, G. But when applying B as source, we may try another training setting due to it's data amount and other factors. You can see similar modifications when applying different scales of training data: https://github.com/damo-cv/TransReID/issues/19#issuecomment-864552276.
Q4: Best or final results?
A: We exhibited the best ones on our paper since the differences between best ones and final ones are not apparent. But I guess it's not rigorous enough due to the defination of DG, and to show final ones is the better. So thank you for your advice again.
Zzhc3321
ds-giao
1.论文使用了duke数据集，这个数据集应该是不让使用的。 2.目前多源域行人重识别，大部分是基于新的协议展开（剔除duke）。 3.关于代码，请问为什么不同数据集要设置不同的参数，这是否符合领域泛化的要求。 4.还有代码里在保存模型参数时，有一个操作是选择最优mAP的模型保存，请问论文里使用是训练过程中最优的结果，还是最终epoch的结果。