Abstract
본 연구에서는 SVG 포맷에 기반한 온라인 캐릭터의 표정 애니메이션을 자동으로 생성할 수 있는 방법을 제안한다. 기존의 극장용 애니메이션에 등장하는 캐릭터의 풍부한 표정 움직임을 수학적 계산에 의해 모델링하고 이를 SVG 포맷의 다른 캐릭터에 '라타게팅' 하면, 리타게팅된 캐릭터는 고유의 형태적 특정을 잃지 않으면서도 모델링된 표정의 움직임을 그대로 재현할 수 있다. 이러한 표정움직임을 리타게팅하는 방법은 하나의 표정 데이터를 여러 가지 캐릭터에 적용할 수 있기 때문에, 비록 애니메이터가 아닌 비전문가일지라도 캐릭터의 표정 애니메이션을 쉽고 빠르게 만들 수 있도록 한다. 본 연구에 의해서 제안되는 표정의 움직임은 비선형적이며 과장된 모션을 취하고 있어 보는 이에게 친근하며, 생동감이 있다. 뿐만 아니라 SVG 포맷에 기반 하므로 기존의 GIF 애니메이션 보다 적은 용량으로 부드러운 표정 움직임을 구현할 수 있다.
In this study, a method to automatically generate facial animations for SVG-based online characters was proposed. By mathematically modeling the expressive facial movements found in cinematic animations and 'retargeting' them onto other SVG characters, the target characters can reproduce complex movements while maintaining their unique formative properties. This approach allows non-specialists to create high-quality animations quickly. The suggested movements are non-linear and exaggerated, providing a lively and friendly aesthetic. Furthermore, the use of the SVG format ensures smoother motion with smaller file sizes compared to traditional GIF animations.
Introduction
Creating engaging character animations for web and mobile platforms requires a delicate balance between visual expressiveness and technical performance. While cinematic animations utilize sophisticated "expressive motion styles"—characterized by non-linear, exaggerated movements like "squash and stretch"—replicating these effects in online environments remains a challenge. Traditional web animations often rely on GIFs, which suffer from limited frame rates, large file sizes, and a lack of the fluidity inherent in vector graphics. Furthermore, high-quality facial animation typically requires professional artistry beyond the reach of non-specialists. This paper proposes a method to bridge this gap by automatically retargeting professional cinematic facial animations onto lightweight, user-defined SVG-based characters.
Methodology
The proposed pipeline translates complex cinematic motion into a mathematical model applicable to any 2D vector face through several key steps:
- Linear Half-Face Model: A standardized set of 47 points is defined across six facial components (face line, eyebrows, eyes, nose, mouth, and ears). By focusing on one side of the face, the system simplifies calculations while maintaining sufficient detail for complex expressions.
- Motion Extraction: Professional animations are analyzed to track the movement of these 47 points. Points are categorized into "Fixed Points," which provide structural stability, and "Moving Points," which drive the emotional expression.
- Relative Coordinate Mapping: To ensure compatibility across diverse character proportions, movement is defined using "Moving Point Change Rates." Instead of absolute pixel values, movement is calculated as a percentage of the character’s total face width ($W^0$) and height ($H^0$).
- SVG Implementation: The resulting data is applied to a target character in SVG format. SVG's support for vector interpolation allows for smooth, high-frame-rate transitions between keyframes with minimal data overhead.
Data and Evaluation
The study utilized a dataset derived from top-performing US-produced animations released between 1990 and 2005. Clips were captured as sequential images and categorized based on Paul Ekman’s six universal emotions (Joy, Sadness, Anger, Fear, Disgust, Surprise) using the Facial Action Coding System (FACS). To evaluate the retargeting effectiveness, the authors designed five distinct target characters based on the "Five Elements" (Wood, Fire, Earth, Metal, Water), ensuring a wide variety of facial structures for testing.
Key Insights: The "Squash and Stretch" Dynamics
A central contribution of this work is the mathematical modeling of traditional animation principles. The charm of high-end animation often stems from the ratio of deformation rather than the specific drawing itself. By normalizing these deformations against a bounding box, the "liveliness" of the motion becomes a portable data format.
By identifying "Fixed Points" statistically, the algorithm preserves the character's unique identity even during extreme exaggeration. Furthermore, using SVG for animation enables "in-betweening" (interpolation), making a small set of keyframes appear as fluid 60fps motion, a feat impossible with standard raster-based methods like GIFs.
Results
The system successfully mapped the six core emotions from professional reference models to five structurally diverse target models. Key findings include:
- Efficiency: Retargeted animations typically required only 19–21 frames. Stored as coordinate offsets in SVG, file sizes were significantly smaller than equivalent GIFs.
- Visual Fidelity: Qualitative tests confirmed that target characters maintained their unique design traits while accurately mimicking the non-linear, "bouncy" timing of the source professional animation.
Conclusion
This research demonstrates a viable path for non-experts to create high-quality web animations by retargeting professional motion data. While current implementation requires manual mapping of the 47 facial points, future work aims to automate feature detection. Expanding the library of reference models and exploring adaptations for 3D web characters represent promising directions for further development.
Summary updated