Difference between original paper and code

[hope1337]

In original paper (on page 2, section 2.2) the authors said that they keep last 100 feature for each track to calculate the cosine distance metric :

But in their code here, they delete all feature every time a confirmed track has been update :

I don't understand why, I am interested in it because I want to use deepSORT in my own project and utilize the feature vector for some purpose.

[hope1337]

@nwojke I apologize for bothering you, but the purpose is to optimize for time and memory, isn't it?