Does this work support tracking every point in long-term? And handle new occurred points?

[luminousking]

1. For tracking initialization, we need to choose N initial points, can we choose every pixel in the video for tracking?
2. Does SceneTracker support adding new points as they get visible throughout the video?

Many thanks,

[wwsource]

1. Due to the limitation of memory size, our current method only supports long-term scene flow estimation of sparse points. In the future, we may develop new methods that support dense point tracking.
2. In fact, we learn to track points in both the first and middle frames during training. Therefore, theoretically, the weights we provide support the addition of new points, but further verification is needed. You can make relevant attempts.

[luminousking]

thx!