This paper offers a way of automatically blending different motions together using a data structure called a registration curve.  A registration curve itself is a relational collection of other data structures: timewarp curves, coordinate alignment curves, and movement constraints matches.  The authors ouline how to construct and use registration curves for blending motions, demonstrate their techniques in practical applications, and discuss the benefits of such a system.

Registration curves were developed as an enhancement of the standard blending method called linear blending.  Linear blending fails to output decent results when motions have different timings or when linearly interpolating a path yields discontinuities in the motion orientation.  To correct these problems, registration curves implement a timewarp curve so corresponding events in two different motions occur simultaneously, and a coordinate alignment curve that allows for the blended path trajectory to be determined incrementally, yielding an orientation that is a weighted average of the input frames.  To guarantee that this averaging does not violate any constraints, registration curves also utilize a method for creating constraint matches where each constraint match is a set of related constraints from the separate motions.


Key Ideas:

1. The timewarp curve is created by fitting a spline curve to a set of frame correspondences so that the spline is strictly increasing.  The path of this spline should have the properties that it is continuous, causal, and has at each step a slope greater than some lower bound.

2. The alignment curve is constructed as a function of the timewarp curve.

3. Each constraint is grouped together with constraints from other motions into a constraint match by following two guidelines: each constraint match contains exactly one constraint from each motion, and each element of a constraint match must be connected.  To accomplish this, constraints are allowed to be eliminated or split into smaller-sized constraints.

4. Creating a single blended frame of motion involves four steps: determining a position on the timewarp curve, positioning and orienting the frames at that position, combining the frames based on the blending weights, and determining the constraints on the resulting frame.


Contributions:  Using timewarp curves, coordinate alignment curves, and constraint matches to overcome the limitations of linear blending.