Optimal and Interactive Keyframe Selection for Motion Capture

How to read this

Category

technical brief, motion capture editing and keyframe reduction

Contributions

• Formulates optimal keyframe selection from dense motion capture as a shortest path problem over frames, solved exactly with dynamic programming, an all pairs cost table plus a DP recurrence, in place of the greedy or curvature based keyframe pickers used previously
• Computes, in one pass, optimal keyframe sets for every keyframe count up to a requested maximum, letting an artist interactively browse compression levels and pick a tradeoff between fidelity and editability
• Demonstrates the algorithm can reduce mocap to around 10 percent of its original frame count while remaining close to the source motion, letting studios re-enter mocap into conventional keyframe based editing tools instead of low level curve editing
• Shows the method strictly outperforms a prior greedy keyframe extraction algorithm, Lim and Thalmann 2001, and beats Maya's Simplify Curves at matched key counts, while coming within about 5 percent of naive PCA based compression

Context

This is less a descendant of a specific prior paper than a direct algorithmic answer to a long standing industry complaint, that dense mocap is hard to hand edit compared to sparse keyframe animation. The authors, spanning Weta Digital, EA SEED, and OLM Digital, validated the approach against working professional animators and motion editors at those studios rather than against a purely academic benchmark.

Correctness

The shortest path formulation is provably optimal for the stated error metric, maximum perpendicular distance in high dimensional pose space, so the optimality claim holds mathematically within that specific metric. The paper is upfront about its limits: the algorithm is roughly cubic in frame count, so very long clips need to be split for interactive use, and the approximation error does not decay smoothly, so an artist still has to browse solutions and use judgment to pick the right keyframe count.

Clarity

A short, clean SIGGRAPH Asia technical brief, four pages, easy reading for a working animator or TD, and the dynamic programming section is compact enough to implement directly from the paper.

How to read it

First pass: the abstract and Figure 1, which together state the whole idea, that good keyframes minimize approximation error between pairs of points rather than sitting at curvature extrema. Second pass: Section 2 in full, it is short and is effectively the entire method. Third pass: skim Section 3's benchmark comparisons for the headline numbers, then read the Limitations and Future Work in Section 4, which are candid about where artist judgment still has to fill the gaps.