3D Human Body Modeling From Partial Views for Faithful Avatars
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating avatars fail to faithfully represent the 3D morphology and surface topology of an individual's body, relying on parametric models or partial recognition of facial features, which limits the accuracy of the human body model reconstruction.
Innovation Solution
A method and system using a trained neural network to generate a 3D human body model directly from partial 3D images, bypassing the need for pose or joint coefficients, by applying a supervised regression model on three-dimensional meshes generated from structured, ordered, and connected partial views.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a skeleton-like body model with parametric points of interest is used to represent the human body, then the kinematics of the avatar can be faithful to the individual's movements, but the 3D representation of the human body model does not faithfully represent the individual's morphology or surface topology
Solution Approach 1:
The method segments the human body into multiple partial views captured from different positions and angles. Each partial view is processed independently by the neural network to generate corresponding 3D mesh segments, which are then assembled into a complete full-body model. This segmentation approach allows the system to capture detailed morphology information from multiple perspectives while maintaining computational efficiency.
Solution Approach 2:
The invention transitions from traditional 2D image processing to 3D mesh generation. The neural network takes 2D partial views as input and directly outputs 3D mesh representations with accurate morphology and surface topology. This dimensional transformation enables faithful reconstruction of the individual's body shape while maintaining real-time processing capabilities.
2Measurement precision
If facial feature recognition based on points of interest is used to generate an individual's avatar, then particular zones such as the nose or eye contour can be recognized, but the method cannot enable the entire human body to be faithfully reconstructed
Solution Approach 1:
The neural network is designed with universal applicability to process any body part rather than being specialized for facial features only. The same network architecture and processing pipeline can handle facial regions, limbs, torso, and other body parts uniformly. This multi-functionality is achieved by training the network on diverse partial views covering the entire body, enabling faithful reconstruction of the complete human body while maintaining high precision for specific features.
Solution Approach 2:
The body is divided into multiple partial views that can be captured from different positions. Each partial view may contain specific features of interest (such as facial features in one view and limbs in another), and the neural network processes each segment independently before assembling them into a complete model. This segmentation strategy enables both detailed feature recognition and comprehensive full-body coverage.
3Adaptability or versatility
If a body model extracted from a data library is used, then the model can be customizable, but it does not specifically represent an individual's body model, resulting in an avatar that borrows specificities without faithfully representing the individual's morphology
Solution Approach 1:
The system enables the individual to serve as their own source of morphological data. By capturing partial views of the individual's actual body and processing them through the neural network, the system generates a 3D mesh that faithfully represents the individual's unique morphology. This self-service approach eliminates the need to borrow data from generic libraries, allowing the avatar to accurately reflect the individual's specific physical characteristics while maintaining customizability.
Data Source
AI summary
A method for generating a human body model includes acquiring at least one partial view of an individual defining a 3D image from at least one optical device arranged in a position of space, and generating a human body model by applying a trained neural network to generate a 3D model of the individual body from input 3D images, the training of the neural network being performed from a set of data describing partial views of the human body.