search

damo-cv / TransReID-SSL

Self-Supervised Pre-Training for Transformer-Based Person Re-Identification
MIT License
173 stars 20 forks source link

Performance on Domain Generalization #6

Closed liyuke65535 closed 2 years ago

liyuke65535 commented 2 years ago

Thanks for your works on ReID! I got several questions:

  1. I see you've done UDA experiments between market and msmt in TransReID-SSL. What if just train the model on A person-dataset then test on B person-dataset? How's the performance?
  2. What does 'patchify stem' mean? More specifically, what's the purpose of ICS component and where it works?
liyuke65535 commented 2 years ago

Looking forward to your generous help!

michuanhaohao commented 2 years ago
  1. You can evaluate the open-sourced models on B dataset. I consider it may be better than the baseline.
  2. 'patchify stem' means the 16*16 Conv2D layer before Transformer Blocks.