Cepheid
A cross-geometry motion completion method that decouples residual structure from physical scale, learns a dimensionless prior, and combines it with amplitude estimation to achieve a unified motion completion model without requiring skeletons or cross-mesh correspondences.
Timeline (last 90 days)
Cepheid decouples the residual structure from physical scale, learns a dimensionless prior, estimates amplitude through observables, and adjusts with a lightweight correction fitted only on training sources, achieving cross-family motion completion.
Cepheid's unified model surpasses harmonic extension on all evaluated geometric families, including previously unseen cloth and legged bodies, and does not require skeletons, registration, or cross-mesh correspondences.
The same Cepheid prior can be directly applied, without retraining, to animal videos for point tracking and to scenarios of robotic grasping of soft objects, demonstrating zero-shot transfer capability.
Cepheid's dimensionless prior divides the residual by its own root-mean-square amplitude, bringing the output near unit scale, so that the residual spatial distribution patterns of different families fall within the same numerical range.
Cepheid does not assume that meshes have the same topological structure, and can handle meshes of different resolutions and different connectivity relations.
Cepheid's amplitude estimation module uses only the observed sparse displacements and regresses a scale factor through a lightweight network, and the correction module contains only a small number of parameters, avoiding overfitting.
At test time Cepheid's prior is fixed, and both amplitude estimation and correction are forward computations that require no iterative optimization, making the model very efficient.
Cepheid is currently still limited to offline evaluation, and in the future its robustness in real video streams and dynamic scenes needs to be verified; the accuracy of amplitude estimation depends on observation quality, and may degrade when sparse observations are noisy.
Cepheid is expected to reduce the cost of motion capture and animation production, for example by using a small number of markers to drive the cloth or soft bodies of different characters, without needing to train a separate model for each character.
The Cepheid project page is https://su-terry.github.io/cepheid-website/
All Facts (10)
The Cepheid project page is https://su-terry.github.io/cepheid-website/
50%UnverifiedCepheid's unified model surpasses harmonic extension on all evaluated geometric families, including previously unseen cloth and legged bodies, and does not require skeletons, registration, or cross-mesh correspondences.
50%UnverifiedThe same Cepheid prior can be directly applied, without retraining, to animal videos for point tracking and to scenarios of robotic grasping of soft objects, demonstrating zero-shot transfer capability.
50%UnverifiedCepheid's dimensionless prior divides the residual by its own root-mean-square amplitude, bringing the output near unit scale, so that the residual spatial distribution patterns of different families fall within the same numerical range.
50%UnverifiedCepheid does not assume that meshes have the same topological structure, and can handle meshes of different resolutions and different connectivity relations.
50%UnverifiedCepheid's amplitude estimation module uses only the observed sparse displacements and regresses a scale factor through a lightweight network, and the correction module contains only a small number of parameters, avoiding overfitting.
50%UnverifiedAt test time Cepheid's prior is fixed, and both amplitude estimation and correction are forward computations that require no iterative optimization, making the model very efficient.
50%UnverifiedCepheid is currently still limited to offline evaluation, and in the future its robustness in real video streams and dynamic scenes needs to be verified; the accuracy of amplitude estimation depends on observation quality, and may degrade when sparse observations are noisy.
50%UnverifiedCepheid decouples the residual structure from physical scale, learns a dimensionless prior, estimates amplitude through observables, and adjusts with a lightweight correction fitted only on training sources, achieving cross-family motion completion.
50%UnverifiedCepheid is expected to reduce the cost of motion capture and animation production, for example by using a small number of markers to drive the cloth or soft bodies of different characters, without needing to train a separate model for each character.
50%