### Summary
This paper introduces Snap-Together Motion, a system to precompute transitions in a large body of motion capture data by identifying key poses and generating smooth transitions between clips of motion connected by these poses.
### Problem
Often, a character is required to perform a variety of actions from a starting pose and return to that pose when finished.  Previously, this was done manually; Snap-Together Motion is an automatic method for generating these sorts of motions.
### Methods Used
* The system finds sets of frames that are close to each other according to the distance metric given in the paper (the same as for Motion Graphs)
* The user selects which frames to use as hub nodes in the graph to be created and provides a threshold for the distance metric to decide which frames should connect to it
* If the poses to be connected are unconstrained:
  * A 'median' pose is determined (the pose with the least distance from all other poses)
  * All connecting motions are modified using displacement maps to pass through this median pose.
* If the one or more of the poses are constrained:
  * All the poses to be connected are averaged to approximate the hub pose
  * The poses are modified as above to pass through the hub pose
  * Each pose's constraints are applied individually
  * Using this modified set of poses, the algorithm returns to the first step and continues until the poses have sufficiently converged

To reproduce these motions, a program need only walk the graph starting and ending at a hub pose

### Key Ideas
Using semi-automatically generated motion graphs is a reasonable alternative to move trees
Good hub poses can be found by determining numerically similar poses
The problem of connecting multiple constrained motions can be solved by the above method