Synthetic Human Keypoint Detection for Robust Model Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The generation of training images for neural networks used in body keypoint detection is tedious, costly, and limited, leading to networks that cannot generalize to out-of-distribution images, and manual ground truth labeling is time-consuming and error-prone, with potential legal and privacy issues.

Innovation Solution

A method for generating synthetic training data using parametric models of human bodies, randomly varying parameters to create diverse virtual humans with annotated keypoints, backgrounds, and lighting conditions, which can be used to train data-driven models for robust keypoint detection.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If real patient images are used for training neural networks, then the training data is authentic and representative of real clinical scenarios, but the process is time-consuming and expensive

Engineering Contradiction:
Improvetraining data authenticityVSAvoiddata generation time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent creates synthetic copies of real patient images using generative adversarial networks (GANs). The GAN generates synthetic images that replicate the statistical properties and visual characteristics of real clinical images without using actual patient data. This allows the training dataset to be populated with authentic-looking images that preserve the reliability needed for training while eliminating the time-consuming process of collecting and annotating real patient images.

Inventive Principle:
Principle #26Copying

2Measurement precision

If manual ground truth labeling is performed on real images, then accurate annotations are obtained, but the process is time-consuming and error-prone

Engineering Contradiction:
Improveannotation accuracyVSAvoidlabeling time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The GAN-based synthesis process automatically generates both the training images and their corresponding ground truth annotations simultaneously. The synthetic images come with automatically generated keypoint labels and segmentation masks, eliminating the need for manual annotation while maintaining high precision. The synthetic ground truth is generated through the same GAN process that creates the images, ensuring consistency between images and annotations without human error.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The system performs self-annotation through the GAN training process. The generator network learns to produce images with embedded ground truth information, and the discriminator network helps refine this information. This self-service annotation mechanism eliminates the need for external manual labeling while maintaining high annotation accuracy, as the system generates its own training data with perfect ground truth labels.

Inventive Principle:
Principle #25Self-service

3Adaptability or versatility

If diverse training data is collected from multiple sources, then the model generalizes better to out-of-distribution images, but the data collection effort and cost increase

Engineering Contradiction:
Improvemodel generalizationVSAvoiddata collection efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The GAN framework allows for systematic variation of image parameters such as lighting conditions, patient positioning, imaging angles, and anatomical variations. By adjusting these parameters during training, the model learns to generalize across diverse conditions without requiring separate data collection campaigns. The synthetic data generation process can efficiently explore the parameter space to create diverse training examples that improve model adaptability.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The GAN-based synthetic data generation system serves multiple functions simultaneously: it creates diverse training images, generates corresponding ground truth annotations, ensures patient privacy by using synthetic rather than real data, and can be adapted to different imaging modalities and clinical scenarios. This multi-functional approach improves productivity by replacing multiple separate data collection and processing workflows with a single unified system.

Inventive Principle:
Principle #6Universality (Multi-functionality)

4Quantity of substance

If real patient data is used for training, then the training process is comprehensive, but privacy and legal concerns arise

Engineering Contradiction:
Improvetraining data volumeVSAvoidprivacy risks
Core Design Contradiction:
Quantity of substanceVSObject-affected harmful factors

Solution Approach 1:

The patent replaces real patient images with synthetic copies generated by GANs. These synthetic images maintain the statistical properties, visual characteristics, and anatomical variations of real patient data, allowing comprehensive training with large volumes of data. However, since they are synthetic rather than real, they eliminate privacy and legal concerns associated with using actual patient information. The copying process preserves the useful characteristics of real data while removing the harmful privacy risks.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4586215A1Accurate body keypoint detection using purely synthetic data for workflow support
Publication Date: 2025.07.16 KONINKLIJKE PHILIPS NV
  • EP4586215A1 patent drawingFigure 1A~1F
  • EP4586215A1 patent drawingFigure 2
  • EP4586215A1 patent drawingFigure 3

AI summary

The present invention relates to body keypoint detection. A synthetic training data generation method is proposed to automatically create synthetic (i.e. artificial) training data. With the proposed approach, it is possible to create a huge variety of different virtual humans with different clothes as well as different environment conditions. The synthetic training data can then be used to train neural networks that work on real data. This may create more robust and more reliable neural networks.