Model Differences From The Reference Paper

[ghost]

I know that you made some modifications on the vanilla deep sort and your repo is working just fine. But as i checked the paper for deep sort, saw that there are some differences that really fed my curiosity. Like Relus instead of Elus, a classifier block with 751 output neurons at the end of the net, even before that the net yields a tensor with 256 elements while the net in the paper was ended just with a 128 element tensor out of a BatchNorm. I was wondering if you could give some intuition about those.

[joaquinCaceres]

I have the same question. Does anyone know?

[SuperbTUM]

Same question here... I also don't know why we need a classifier after having n-dim features for matching.