Abstract
Estimating the pose and anatomy of 3D-scanned archaeological statues presents unique challenges due to non-realistic artistic proportions and the lack of training data for stylized or imaginary creatures. We present FAKIR (Forward And backward Iterative Registration), a kinematic algorithm that fits a flexible "sphere-mesh" model to raw point clouds. By treating limb length and thickness as variables alongside joint angles, FAKIR adapts to a wide range of morphologies—from classical sculptures to non-humanoid creatures—without the need for deep learning priors or clean surface meshes.
Introduction
In cultural heritage preservation, 3D scanning provides high-fidelity digital replicas of historical artifacts. However, extracting structural information—such as skeletal pose and anatomical proportions—from these raw point clouds remains difficult. Most state-of-the-art pose estimation tools are built on deep learning models like SMPL, which are trained on databases of realistic human scans. These models often fail when applied to art, where artists historically prioritized aesthetic expression over anatomical accuracy, leading to elongated limbs or stylized features. Furthermore, "imaginary" creatures like centaurs or mermaids lack the training data required for supervised learning approaches.
FAKIR addresses these gaps by providing a fast, iterative, and flexible method that works directly on raw point sets. It avoids the "deep learning baggage" of requiring massive datasets and instead relies on a robust geometric optimization process.
Methodology
The FAKIR framework is built upon two primary innovations: a simplified geometric representation and a specialized iterative registration algorithm.
The Sphere-Mesh Model
Rather than using a complex, high-resolution mesh, FAKIR represents the underlying anatomy using a "sphere-mesh." Each bone or limb segment is modeled as a union of spheres centered along a line segment with a varying radius. This creates a pliable envelope that captures the volume and thickness of limbs, allowing the model to adapt to the specific "fleshiness" or "slenderness" of a particular statue.
The FAKIR Algorithm
The core algorithm, Forward And backward Iterative Registration (FAKIR), is inspired by inverse kinematics techniques like FABRIK. Starting from a manually selected anchor point (typically the pelvis), the algorithm processes chains of bones through two main phases:
- Forward Pass: The algorithm iteratively rotates and scales each bone to align with the local point cloud density.
- Backward Pass: Joint positions and sphere radii are refined by minimizing the distance between the model's surface and the input points.
To handle complex poses where limbs may be close to or touching the torso, FAKIR employs Normal-Constrained Projection. By using the orientation of surface normals from the point cloud, the algorithm ensures that points are assigned to the correct bone segment, preventing "sticking" artifacts common in proximity-based methods.
Evaluation and Results
The researchers evaluated FAKIR across diverse datasets, including synthetic bone chains with noise, high-resolution archaeological scans (e.g., Venus de Milo, Dancing Faun), and the TOSCA benchmark for human and animal poses.
Versatility and Efficiency
A key strength of FAKIR is its ability to handle non-humanoid topologies. By simply adjusting the skeleton's connectivity, the same algorithm can fit diverse creatures such as horses, gorillas, or mythological figures like mermaids.
In terms of performance, FAKIR is highly efficient. For a point cloud containing 10,000 points, the initial fitting pass takes approximately 2.5 seconds, with subsequent refinements completing in under a second. This speed makes it suitable for large-scale archival workflows.
Comparative Accuracy
When compared to established methods like Pinocchio, FAKIR demonstrated superior anatomical accuracy, particularly in challenging regions like the neck and shoulders. It also proved robust against missing data, successfully processing statues with broken limbs or significant surface noise.
Discussion and Conclusion
FAKIR's success stems from prioritizing "pliability over priors." By treating limb length and thickness as variables to be optimized on-the-fly, it achieves a level of "kinematic scaling" that rigid human templates cannot match. Furthermore, working directly on raw point clouds eliminates the error-prone step of surface reconstruction.
Despite its advantages, FAKIR currently requires a manual "point-and-click" initialization for the starting joint. Future work aims to automate this step using Principal Component Analysis (PCA) and to incorporate symmetry constraints to better handle heavily damaged artifacts. Nevertheless, FAKIR represents a significant step forward in the structural analysis of cultural heritage, providing a flexible tool for understanding the stylized anatomy of the past.
Summary updated