search

yxgeee / SpCL

[NeurIPS-2020] Self-paced Contrastive Learning with Hybrid Memory for Domain Adaptive Object Re-ID.
https://yxgeee.github.io/projects/spcl
MIT License
318 stars 67 forks source link

Question about the paper #30

Closed Liam501 closed 3 years ago

Liam501 commented 3 years ago

Hi, In your paper “Self-paced Contrastive Learning with Hybrid Memory for Domain Adaptive Object Re-ID”, there was one particular place that gets me a little bit confuse. image In Table 5(a), there is a huge improvement in mAP and CMC when using self-paced learning strategy. However, there is still a large gap between Src. class + tgt. cluster (w/o self-paced) and Ours w/o R_comp & R_indep. Can you explain more about self-paced learning strategy, does that means using R_comp & R_indep?

yxgeee commented 3 years ago
  1. The difference between Src. class + tgt. cluster (w/o self-paced) and Ours w/o R_comp & R_indep is that, Src. class + tgt. cluster (w/o self-paced) does not use outliers for training while Ours w/o R_comp & R_indep uses outliers for joint training.
  2. Yes, self-paced means R_comp & R_indep.
Liam501 commented 3 years ago

Okey, now I understand it. Thank you for answering my question.