Whenever possible, we now use the positions of the predecessors at t of a merger at t+1 to initialize the GMM fitting when resolving mergers. This runs from t=0 to t=Tmax such that we can even propagate this information when mergers remain connected for more than one frame.
It seems to help quite a lot, here's for instance the result on my example dataset for this problem:
Whenever possible, we now use the positions of the predecessors at
tof a merger att+1to initialize the GMM fitting when resolving mergers. This runs fromt=0tot=Tmaxsuch that we can even propagate this information when mergers remain connected for more than one frame.It seems to help quite a lot, here's for instance the result on my example dataset for this problem: