0

I have a set of 2D Points at time $T=0$, and another set at time $T=1$. Each point $P$ provides (X,Y,$\dot{X}$,$\dot{Y}$).

I need to basically match points at $T=0$ and $T=1$. A simple L2 metric of $\left \| P^{t=0}, P^{t=1} \right \|$ won't work, as I could have two points that are close, but travelling in opposite directions.

Taking the dot product $dot(P^{t=0}, P^{t=1})$ of a pair of points won't work too, as I could have two points that are far apart but traveling in the same direction.

I could combine them somehow and weigh them accordingly, but we could also include a vector $V=P^{t=0}-P^{t=1}$ and now we could compute a dot product of $dot(V, P^{t=1})$ or $dot(V, P^{t=0})$ and try to minimize those.

Just to note, the velocity vector's magnitude also ranges from 0 to 1. Generally, if the magnitude is low, the probability/stability of that vector is considered low as well, so we need to incorporate that as well.

How can I proceed? The assumption is that a pair of matching points across frames are generally close to each other

  • If at time t=0 you know position and velocity why not extrapolate the position at time t=1 and compare that against the given set of points at t=1 by L-2 norm – Attack68 Jun 26 '18 at 21:38
  • The velocity is calculated not directly from the prior positions, but rather from a neural network as a flow field. – user1436508 Jun 26 '18 at 21:57

0 Answers0