search

Klitter / Pyramidal_Person_ReID

A simple reproduction of Pyramidal Person Re-IDentification via Multi-Loss Dynamic Training (CVPR2019).
MIT License
34 stars 16 forks source link

how about inference time with the Pyramid #4

Open chunniunai220ml opened 4 years ago

chunniunai220ml commented 4 years ago

Hi ,thanks for your sharing! I want to know the change of the inference time with Pyramid module ,increased or not? I see the self.base=resnet101() in the Pyramid.py , so the models you shared with backbone resnet101? do you have trained resnet50? if you have,could you tell me the map and rank1 on 1501 and duke?

Klitter commented 4 years ago

Hi ,thanks for your sharing! I want to know the change of the inference time with Pyramid module ,increased or not? I see the self.base=resnet101() in the Pyramid.py , so the models you shared with backbone resnet101? do you have trained resnet50? if you have,could you tell me the map and rank1 on 1501 and duke?

Sorry for the late reply , it is because I am on Spring Festival holiday.

  1. Compared with simply using backbone, Pyramid increases the inference time because of extra convolution in Pyramid module but its cost is small.
  2. In my code, the backbone is resnet101.
  3. I do not know the performance with resnet50, but I will publish the training code as soon as possible so that you may be able to train and test it by yourself later.