Closed vilemil closed 11 months ago
Hej, thanks for your interest! Re-identification can be very challenging for the point trackers. I'd suggest you first try replacing PIPS (i.e., the default point tracker) with the state-of-the-art CoTracker and see whether this helps. You can do so by adding the command line arguments model/point_tracker=cotracker model.positive_points_per_mask=16
.
Thank you! I will definitely check it out.
Hello,
First of all, congrats! Great work.
I run a few test and it has remarkable precision. Except one case: disappearing-reappearing objects, works really poorly (almost every case the tracking was lost) . I'm trying to track different, moving objects with moving camera. I haven't dug into your work in detail yet, but do you have any suggestions for that scenario? (more positive points for example?). I tested TAM also and it's working really well in this case(because of XMem), but has worse precision with small, blurry objects.