Landmark-Guided Elastic Shape Analysis

Motions of virtual characters in movies or video games are typically generated by recording actors using motion capturing methods. Animations generated this way often need postprocessing, such as improving the periodicity of cyclic animations or generating entirely new motions by interpolation of existing ones. Furthermore, search and classification of recorded motions becomes more and more important as the amount of recorded motion data grows.

In a recent paper, we (Martin Bauer, Markus Eslitzbichler, and Markus Grasmair) have applied methods from shape analysis, augmented by the incorporation of additional feature point information, to the processing of animations. On this page you find a short overview of the numerical results we have obtained with this method.

hands.zip: This video shows a comparison of the energy minimizing paths between two hand shapes. On the left, the geodesic was computed by using only the elastic metric without any feature point information. In the middle, feature points were added on the tips of the index and ring finger of both the starting shape and the end shape. On the right, the feature points were incorrectly set, which leads to the thumb slowly deforming into the index finger of the final shape.

In addition, this figure shows selected intermediate steps of the transformation:

walk.zip: These videos show the matching of two different walking animations using the following methods:

• Linear interpolation of the Euler angles describing the joints.
• Elastic shape matching without reparametrization (i.e.: local speed changes of the animation) and without feature point information.
• Elastic shape matching with reparametrization but without feature point information.
• Elastic shape matching with reparametrization and feature point information. The feature points were defined as the first three times when the left knee in the walking animation passes the right knee.

All movies show the initial animation, the final animation, and their mean according to the distance measure used. Note that the first animation consists of four steps, while the second only of three steps. Without feature point information, this leads to the artificial insertion of one or even more small steps in the intermediate animations.

This figure shows the trajectories of the feet in the different intermediate animations:

step.zip: Here we try to match two animations that show a person stepping over an obstacle of varying height. Again, we used the following methods:

• Linear interpolation of the Euler angles describing the joints.
• Elastic shape matching without reparametrization (i.e.: local speed changes of the animation) and without feature point information.
• Elastic shape matching with reparametrization but without feature point information.
• Elastic shape matching with reparametrization and feature point information. The feature points were defined as the first three times when the left knee in the walking animation passes the right knee.

This figure shows the trajectories of the feet in the different intermediate animations: