In this work we explore causal relationships in visual tracking. There are obvious cases: the cameraman following the object or moving around a static scene. In the left video the object stays static throughout the sequence. This also holds for the right video, after compensating for camera motion.
There are less obvious cases, where the camera motion and its relationship with the object can be beneficial to tracking. In this sequence, there are several challenging moments. The first one is when the camera suddenly shakes. The second moment is after a full occlusion. Using the camera motion information and causal relationships, prior information can be supplied to a tracker to prevent failure in both these cases.
We propose a method to discover these causal links. Our entropy based approach can measure complex non-linear relationships.
These are then used to predict future object motions, outperforming non-causal methods, such as auto-regression or Kalman Filter.
For more information see:
Lebeda, Hadfield, Bowden: Exploring Causal Relationships in Visual Object Tracking. In Proc. of ICCV, 2015.
or visit http://cvssp.org/Personal/KarelLebeda/publications.html#ICCV2015