search

phil-bergmann / tracking_wo_bnw

Implementation of "Tracking without bells and whistles” and the multi-object tracking "Tracktor"
https://arxiv.org/abs/1903.05625
GNU General Public License v3.0
818 stars 196 forks source link

Questions about the change of ReID model #154

Closed corentin87 closed 2 years ago

corentin87 commented 2 years ago

Hi, I was reading your paper and saw at that time, you were using a Siamese Network and getting a similarity score. But digging a bit into your repo I saw the commit when you switched to Torchreid and Resnet-50 to get features of an image and then measure the distance between two vectors. Why did you stop using your Siamese CNN? Was ReID with Resnet-50 more accurate and faster? Thank you

timmeinhardt commented 2 years ago

Hello, we are still using the concept of a Siamese network. It just means applying the same network to extract features from two inputs and then we compare their distance. The switch to torchreid did not change this as we were also using a ResNet50 before.

corentin87 commented 2 years ago

Ok thanks so you still train two ResNet50 as a Siamese Network. What was your idea to update your first implementation to TorchReid? Just easier to use? Did you retrain it compare to the model available in the Model-Zoo of Torchreid?

timmeinhardt commented 2 years ago

A Siamese network contains two identical subnetworks. This means in practice you only train one ResNet50 but apply it twice/in parallel.

corentin87 commented 2 years ago

Yes got it. Also last question your resnet50-fc512 trained on Market1501 using TorchReiD is 300MB while the same network in the TorchReid Model Zoo is only 100MB. Why is that? Thanks.

timmeinhardt commented 2 years ago

Our model file includes the state dictionary of the model, optimizer and learning rate scheduler. I assume the models from the official model zoo only include the first.

corentin87 commented 2 years ago

ok thanks for your replies!

timmeinhardt commented 2 years ago

No worries. Can we close this issue?