Abstract
Simulating believable facial animation is a topic of increasing interest in computer graphics and visual effects. In this paper, we present a hybrid technique for the generation of facial animation based on motion capture data. After capturing a range of facial expressions defined by the Facial Action Coding System (FACS), the radial basis function (RBF) is used to transfer the motion data onto two facial models: one realistic and one stylized. The calculations of the distances for the RBF technique are approached in three variants: Euclidean-based, geodesic mesh-based, and hybrid-based. The hybrid approach leverages the advantages of the first two methods. To increase efficiency, calculations are aided by preprocessed distance data. Our findings demonstrate the efficiency and quality of the hybrid technique when generating facial animation with motion capture.
Introduction
Transferring motion capture (mocap) data from a live performer to a digital character remains a fundamental challenge in computer graphics. Radial Basis Function (RBF) interpolation is a widely used tool for this task, determining vertex displacement based on proximity to mocap markers. However, the choice of distance metric—Euclidean versus Geodesic—significantly impacts both visual fidelity and computational performance.
The Distance Metric Challenge
Traditional RBF systems rely on Euclidean distance, which measures the straight-line path through 3D space. While computationally efficient, Euclidean metrics are "surface-blind." For instance, when a performer opens their mouth, a Euclidean system may incorrectly perceive the upper and lower lips as being close together, leading to "gluing" artifacts where the lips stick or pull unnaturally.
Conversely, Geodesic distance measures the shortest path along the mesh surface. While anatomically accurate and effective at preventing artifacts near "holes" like the mouth and eyes, calculating geodesics across complex meshes is computationally intensive, often proving too slow for real-time or high-iteration workflows. The authors sought a "Goldilocks" solution: a method that is anatomically correct where it matters but fast enough for practical use.
Proposed Hybrid Methodology
This paper introduces a hybrid RBF interpolation framework designed to balance anatomical accuracy with speed. The methodology consists of four key pillars:
- FACS-Based Capture: 24 facial expressions were captured using the Facial Action Coding System (FACS) with 41 reflective markers.
- Hybrid Distance Metric: The system selectively applies geodesic distances (calculated via an A* search algorithm) to critical regions—specifically the mouth and eyes—while utilizing standard Euclidean distances for the rest of the face.
- Gaussian Blending: To ensure smooth transitions between geodesic and Euclidean zones, a Gaussian decay function weights the influence of each metric, preventing visible seams or jumps.
- Distance Matrix Optimization: To maximize efficiency, distances between every vertex and marker are pre-calculated and stored in a matrix, bypassing redundant calculations during the animation phase.
Experimental Setup
The framework was tested on two distinct 3D models to evaluate its robustness across different topologies: a Realistic model (2,746 vertices) and a Stylized model (3,774 vertices).
Results and Discussion
The researchers compared Euclidean, Geodesic, and Hybrid variants across quantitative and qualitative benchmarks.
Quantitative Performance
Without optimization, the Geodesic approach was significantly slower (~462.51s) compared to the Euclidean (~9.38s) and Hybrid (~288.82s) methods. However, with the implementation of pre-calculated distance matrices, the runtime for all three methods dropped to approximately 4.6s to 5.0s. This optimization makes the superior quality of the Hybrid approach computationally "free" during the animation loop, as the bottleneck shifts from calculation to simple lookup.
Qualitative Quality
- Euclidean: Suffered from significant artifacts and "gluing" around the mouth due to its inability to respect surface topology.
- Geodesic: Provided accurate surface awareness but introduced minor distortions due to A* heuristics.
- Hybrid: Achieved the highest fidelity, offering smooth deformations by combining Euclidean stability with Geodesic precision in high-risk areas.
Conclusion and Future Work
The hybrid RBF technique effectively addresses the limitations of pure Euclidean or Geodesic methods. By focusing computational resources on anatomically complex regions and utilizing pre-calculated matrices, the system delivers high-quality facial animation suitable for production environments.
Future work will explore integration with blendshape rigs and the application of machine learning to further refine surface deformations on highly stylized characters, where differing facial proportions still present unique challenges for automated mapping.
Summary updated