search

KovenYu / MAR

Pytorch code for our CVPR'19 (oral) work: Unsupervised person re-identification by soft multilabel learning
https://kovenyu.com/publication/2019-cvpr-mar/
315 stars 83 forks source link

A Question Concerning The Datasets Being Used #6

Closed PeixianChen closed 5 years ago

PeixianChen commented 5 years ago

Hello, Thanks for your work, as from which I have learned a lot. However, I found that most methods in the literature use datasets like DukeMTMC-reID and Market-1501 instead of MSMT-17. So I was wondering if you have done any experiments for D->M or M->D. If so, could you please release or send me the results? Thanks in advance.

KovenYu commented 5 years ago

@Pealing Hi, thanks for your attention.

we have not done experiments on D2M or M2D, since an underlying assumption of our model is that the reference persons (can be regarded as a set of bases) have to be diverse enough in terms of identity-wise visual appearance to represent the unseen target persons, and thus we adopted MSMT17 which has most identities among all datasets. If you are interested in testing D2M or M2D, please consider trying it by replacing the source dataset.