All the examples given in this paper are with rigid edges. I'm wondering will this work also be applied on the edges which are not straight or on the curves. One simple method is to divide the crooked contours into several small segment, and find the approximate line for each segment. But this way obviously will increase the error when doing the tracking, moreover, it may bring some bad effect on the model building steps. I agree that intertwining the tracking and model building is really a good idea, it utilizes the previous tracking result to provide a robust model, kind of information recycling, and this may also avoid dramatic drift when the initial model is not good enough.