Modular Body-Part Skeleton Modeling for Noisy Motion Reconstruction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for modeling virtual character skeletons in animation struggle with better motion reconstruction in noisy contexts while maintaining a smaller model size, as they often consider the skeleton as a whole object, leading to unwanted artifacts and inefficiencies.

Innovation Solution

The proposed method splits the virtual character's skeleton into smaller sub-graphs, trains each part independently, and uses a prioritized concatenation strategy to reconstruct the full skeleton, allowing for modular and flexible motion modeling.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If the skeleton is modeled as a whole object, then the model structure is simple, but the motion reconstruction quality deteriorates in noisy contexts and produces unwanted artifacts

Engineering Contradiction:
Improvemotion reconstruction qualityVSAvoidmodel structure complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The skeleton is divided into multiple independent body part models (e.g., upper body, lower body, arms, legs) that can be trained and processed separately. Each sub-graph represents a specific body part with its own latent space, allowing for specialized motion reconstruction that handles noise better while maintaining overall coherence through the hierarchical structure.

Inventive Principle:
Principle #1Segmentation

2Productivity

If the skeleton is modeled as a whole object, then the implementation is easier, but the model size increases and efficiency decreases

Engineering Contradiction:
Improvemodel efficiencyVSAvoidmodel size
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

By segmenting the skeleton into modular body part sub-graphs, each with its own compact latent space representation, the overall model size is reduced compared to a monolithic model. The modular structure allows for more efficient processing and training, as each sub-model can be optimized independently.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a hierarchical dimension to the model architecture, with body part sub-graphs organized in a tree structure rooted at the pelvis. This dimensional organization allows the model to process motion data more efficiently by operating at multiple levels of abstraction, from individual body parts to the full skeleton.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Adaptability or versatility

If the skeleton is modeled as a whole object, then the model is less flexible, but the implementation is more straightforward

Engineering Contradiction:
Improveflexibility in generating or modifying body partsVSAvoidmodel architecture complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The modular body part sub-graph architecture enables independent manipulation, generation, and modification of specific body parts. Each sub-model can be trained separately and combined in different configurations, providing high flexibility for creating variations or fixing issues in specific body parts without affecting the entire model.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The hierarchical tree structure with the pelvis as root allows for dynamic reconfiguration of body parts. The modular design enables the model to adapt to different animation scenarios by activating or deactivating specific sub-graphs, and the latent spaces can be manipulated dynamically to generate diverse motion patterns.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4661402A1Body-parts animation modeling
Publication Date: 2025.12.10 INTERDIGITAL CE PATENT HOLDINGS SAS
  • EP4661402A1 patent drawingFigure 1
  • EP4661402A1 patent drawingFigure 2
  • EP4661402A1 patent drawingFigure 3

AI summary

Methods and apparatus are provided for training a body-parts model that is split into portions and then training a network of corresponding parts entirely independent from its neighbor. In one embodiment, training is performed on different parts of a model with different parts of input data, representing human movement. In another embodiment, specific body parts are independent and weighted based on a loss specific to each body part. In other embodiments, the trained model is encoded and subsequently decoded. The decoded data is concatenated to build a total human skeletal model.