stijnh / scalable-crowd-analysis Goto Github PK

Hello, first of all thank you for sharing your work!

I was testing and was finding it was really sensitive to the parameters (and that ones would influence each other's choice severely). So I start looking at the code together with the paper and I found something different: the normalization

You seem to normalize the tracklets (both position and velocity) with:
normalize = alpha * 2 + beta * 2

then passing these normalized tracklets to prepare_quick_shift.
I experimented by doing the normalization individually for position and velocity inside this last function

dist_p = np.linalg.norm(p[i] - p[neighbors], axis=1) / alpha
dist_v = np.linalg.norm(v[i] - v[neighbors], axis=1) / beta

And found it to be much more robust, even when changing scenarios.

I'd like to know why you did such normalization, maybe I'm missing something... Thank you again!