Open perrywu1989 opened 5 years ago
Actually the visualization tool is not running the tracker itself. The tracker is a meta algorithm on top of the network, that uses the previous estimated position and passes multiple scales of the image to get a robust estimation of the target. The application just focuses on the network itself, as a similarity metric. The purpose is to have a peek into the "Black box" that is the network, in order to explore its behavior and limitations.
It wouldn't be that hard to implement the option of running the full tracker though. I don't have that much time right now to work on that, but if anyone wants to work on that, please let me know.
After we got the score map, how can we get the bounding rect box of the object?
And we only have a video and one reference box in the first frame, can we keep tracking this object? I think the problem is how we can get the reference box of next frame correctly.