Bare Body Shape Estimation from Concealed Scans Using Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for estimating body shapes from dressed scans are computationally expensive, sensitive to initialization, and lack feature details due to non-rigid cloth deformations, making it difficult to obtain accurate undressed body shapes from dressed scans.
Innovation Solution
A method is developed to generate a training dataset of concealed and corresponding unveiled body shapes using computer simulation, incorporating noise and pose variations, and utilize a neural network with encoder-decoder architecture to predict undressed body shapes from dressed scans.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If statistical human body models are used to fit dressed body scans through constrained optimization, then body shape estimation can be performed, but the method becomes computationally expensive and sensitive to initialization parameters
Solution Approach 1:
The patent replaces the traditional mechanical optimization-based fitting process with a neural network-based deep learning system. The neural network is trained on paired dressed and undressed body data to directly predict undressed body shapes from dressed scans, eliminating the need for iterative constrained optimization and statistical model fitting, thereby achieving both high accuracy and computational efficiency
Solution Approach 2:
The patent performs preliminary training of the neural network on a large dataset of paired dressed and undressed body scans before deployment. This pre-training phase captures the complex mapping between dressed and undressed body geometries, allowing the system to make rapid predictions without requiring computationally expensive optimization during actual body shape estimation
2Manufacturing precision
If optimization parameters are increased to obtain detailed body features, then manufacturing precision improves, but device complexity and computational cost increase
Solution Approach 1:
The patent replaces complex iterative optimization processes with a single-pass neural network inference. The neural network architecture is designed to directly output detailed body features without requiring multiple optimization steps, reducing both computational complexity and sensitivity to initialization while maintaining high precision in body feature reconstruction
3Measurement precision
If real dressed and undressed body scans are collected for training, then training data accuracy improves, but data collection becomes extremely expensive and time-consuming
Solution Approach 1:
The patent uses virtual reality technology to create virtual bodies with known undressed geometries and generates corresponding dressed body scans through virtual clothing simulation. This copying approach creates synthetic training datasets that accurately represent the dressed-undressed body relationship without requiring actual physical scanning of people in both states, dramatically reducing data collection time and cost while maintaining training data quality
Data Source
AI summary
Methods are disclosed for generating a training dataset of concealed shapes and corresponding unveiled shapes of a body for training a neural network. These methods may include generating with the aid of computing means a first dataset comprising a plurality of first surface representations representative of a plurality of bare shapes of a plurality of bodies. The plurality of bare shapes are concealed virtually by means of a computer implemented program in order to obtain a plurality of simulated concealed shapes of the plurality of bodies. The plurality of simulated concealed shapes are applied to a scanning simulator, the scanning simulator generating a second dataset comprising a plurality of second surface representations representative of the plurality of simulated concealed shapes.