The map of the datasets MSMT is different from that in the paper about apnet HOT 4 CLOSED

Thanks for your attention!
The config and code in this repo were only for Market and Duke. The config for MSMT17 is available in the repo, now. Please let me know if you have any additional questions.

from apnet.

Thanks for your reply!
Another question.In your paper, the baseline without any attention on Market is map 87.8% and rank1 95.0%. But the results on Market of paper bag of tricks are map 85.9% and rank1 94.5%. Why is your baseline so much higher than the paper bag of tricks? I modified the image size and lr in your code to the same as paper bag of tricks [256,128] and 0.00035, and your result is still much higher than the paper bag of tricks.I haven't found the part of your code that causes this result for the time being.

from apnet.

Please check the configs and paper for details. We modify some hyper-parameters and add some tricks for better performance.
I guess it may be because of the horizontal flip inference (may be useful for mAP), better batch size (may be useful for R1), and other hyper-parameters.
Hope you find it helpful.

from apnet.

Thanks! I will check the configs and paper carefully.
Thanks again for your reply.

from apnet.