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

ResNet seems to give different embeddings across different runs #86

Closed suxuann closed 4 years ago

suxuann commented 4 years ago

I tried to run the ReID network on two different sets of boxes, as shown in the following code.

dets = torch.tensor([[1337., 417., 1503., 795.]]) embeddings = self.reid_network.test_rois(image, dets).data

and,

dets = torch.tensor([[1337., 417., 1503., 795.], [585., 446., 669., 708.]]) embeddings = self.reid_network.test_rois(image, dets).data

It seems like the resulting embeddings for the box [1337., 417., 1503., 795.] are different in the two runs. Could you kindly help me verify if this is the case?

timmeinhardt commented 4 years ago

I just did a quick check and the embeddings seemed to be the same for different runs. You can also see how we set the random seeds and CUDNN backend to deterministic in test_tracktor.py. How do you measure the similarity of embeddings?

suxuann commented 4 years ago

Hello, I checked again. The embeddings are the same. Thank you.

timmeinhardt commented 4 years ago

Okay. I will close this issue.