Abstract
Skeletal motion retargeting aims to transfer motion data from a source character to a target character with a different skeletal structure. This task is challenging due to the high coupling between motion data and specific skeletal topologies. We propose a novel approach using a high-order graph convolutional network (GCN) and a feature separation strategy. By decoupling global body orientation from local pose data and employing high-order skeletal convolutions to capture long-range joint dependencies, our method significantly improves retargeting accuracy. Experimental results on the Mixamo dataset demonstrate a 38.6% improvement in accuracy for heterogeneous retargeting and a 74.8% improvement for homomorphic retargeting compared to current state-of-the-art methods.
Introduction
Motion retargeting is a cornerstone of modern animation, allowing developers to reuse motion capture (mocap) data across various characters. However, motion data is inherently "tightly coupled" with a character's specific skeletal structure. When source and target characters differ in proportions, joint counts, or movement ranges—such as retargeting real human motion to stylized virtual characters—existing "skeleton-aware" networks often struggle to separate the essence of an action from the skeleton it was recorded on. This leads to artifacts like unnatural poses, incorrect body orientations, or "foot sliding."
Methodology
Feature Separation and Orientation Decoupling
A major challenge in retargeting is the simultaneous learning of global orientation and local joint poses. We introduce a feature separation stage that extracts the character's body orientation as an independent feature. By isolating global rotation from local joint poses, we reduce the complexity of the learning task, allowing the network to focus on the nuances of the movement itself. This decoupling significantly stabilizes training and improves the naturalness of the retargeted results.
High-Order Skeletal Convolution
Standard GCNs typically aggregate information only from immediate (1-hop) neighbors in the skeletal graph. To capture broader structural context, we propose a high-order skeletal convolution operator. By utilizing powers of the adjacency matrix, the network can capture information from a wider neighborhood, learning relationships between distant joints (e.g., the hand and the shoulder) without a massive increase in parameters.
Network Architecture
The overall architecture is based on a pair of autoencoders ($G^A$ and $G^B$) that share a latent space. This ensures that the motion's semantic meaning is preserved even as it is decoded onto a different skeletal structure.
Experiments and Results
We evaluated our method using two major benchmarks: Human3.6M (real human movements) and Mixamo (synthetic virtual characters). We specifically tested the "heterogeneous" scenario, retargeting 18-joint real-world data to virtual characters with higher joint counts (typically 22 or 28).
Quantitative Performance
Our method achieved significant improvements over the previous state-of-the-art (SAN):
- Heterogeneous Retargeting: Accuracy improved by 38.6% on the Mixamo dataset.
- Homomorphic Retargeting: For characters with the same joint count but different proportions, accuracy improved by 74.8%.
Qualitative Analysis
Visual results show that the proposed method maintains correct body orientation during complex turns and produces much more natural limb placements than previous techniques. By decoupling orientation, the model avoids common errors where characters fail to face the intended direction during dynamic movements.
Conclusion
By combining orientation decoupling with high-order graph convolutions, we have developed a more robust framework for skeletal motion retargeting. Our approach effectively handles structural discrepancies between heterogeneous characters, resulting in more accurate and visually natural animations. Future work will explore integrating Inverse Kinematics (IK) as a constraint to further reduce minor artifacts like foot sliding.
Summary updated