search

silverbulletmdc / PVEN

Parsing based vehicle ReID
108 stars 25 forks source link

about your baseline #25

Open abcdvzz opened 3 years ago

abcdvzz commented 3 years ago

Could you please tell me how you did the experiments of baseline? Could you release the code of baseline? I notice that your baseline method achieves 77.2% on veri according to Table 6 in your paper which is a very high baseline beating down the other sotas.

silverbulletmdc commented 3 years ago

You can change the veri776_b64_pven.yml file and delete the local-triplet loss. The PVEN without local loss can be regarded as a strong-reid-baseline.

abcdvzz commented 3 years ago

No local-triplet loss means no vehicle part parser, no view-aware feature alignment, no common visible feature enhancement. So using a plain resnet50 with some tricks can get 77.2% on veri776. Am I right?

silverbulletmdc commented 3 years ago

Yes

gutengzczy commented 3 years ago

Hi, sir, is it the same as baseline when the lambda_ was set to 0? There is no localdistmat when lambda=0, and uses the global distmat to calculate cmc and mAP, am I right? Hope for your reply, thanks!

silverbulletmdc commented 3 years ago

Hi, sir, is it the same as baseline when the lambda_ was set to 0? There is no localdistmat when lambda=0, and uses the global distmat to calculate cmc and mAP, am I right? Hope for your reply, thanks!

Yes