Questions to ask: How to integrate this kind of coarse tracking within 2D image plane with Geometric estimation? Under what algorithm framework can this benefit Geometric estimation such as camera pose recover?
IDEA: grow multiple randomized decision trees on-line for classification
KEY, On-line Bagging, On-line Growing Trees, choose the best question(test), split the leaf only when needed and able to gain best benefit, On-line Adaptation, temporal knowledge weighting, ability to unlearn useless knowledge, ability to handle changing data distribution
APPLICATION, tracking object in real-time, able to handle, occlusion, fast movement, view point change
Line Feature Based, Moving in Stereo: Efficient Structure and Motion Using Lines
Point Feature Based