Facial Joint Alignment for Robust Head-Top Estimation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional 3D skeleton recognition technologies struggle to accurately specify the position of a person's head, particularly in images affected by appearance, hair disorder, or occlusion, leading to inaccurate performance evaluation.

Innovation Solution

A computer executes processing to specify positions of facial joints using a machine learning model, estimating the head position by aligning facial joint positions through rotation, translation, and scale parameters, minimizing error using least squares methods.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of time

If machine learning models are trained using only correctly annotated training data, then model training time is reduced, but training accuracy deteriorates because the model cannot learn from incorrectly annotated data

Engineering Contradiction:
Improvemodel training timeVSAvoidtraining accuracy
Core Design Contradiction:
Loss of timeVSMeasurement precision

Solution Approach 1:

The patent segments the training data into two distinct groups: correctly annotated data and incorrectly annotated data. This segmentation allows the model to learn from both types of data separately, preventing the degradation of training accuracy while maintaining efficient training time. The segmentation is achieved through the training data generation unit that creates multiple versions of training data with different annotation qualities.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the parameter of data quality by introducing incorrectly annotated training data with controlled error rates. This parameter change allows the model to adapt and learn robust features that are invariant to annotation errors, thereby improving training accuracy without significantly increasing training time. The degree of incorrect annotation can be adjusted as a parameter to optimize the balance between training time and accuracy.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If incorrectly annotated training data is used, then training accuracy improves, but model reliability deteriorates due to learning from erroneous annotations

Engineering Contradiction:
Improvetraining accuracyVSAvoidmodel reliability
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent segments training data into correct and incorrect annotation groups, allowing the model to learn from both while maintaining reliability. The segmentation enables controlled exposure to errors without compromising overall model trustworthiness, as the correct annotations provide a reliable foundation for learning.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies partial incorrect annotation rather than complete corruption of training data. By introducing only a portion of incorrect annotations mixed with correct ones, the model learns to handle errors without being overwhelmed by them, maintaining reliability while still improving accuracy through exposure to annotation variations.

Inventive Principle:
Principle #16Partial or excessive action

3Stability of the object's composition

If the degree of incorrect annotation is increased, then robustness of the model improves, but training time increases due to additional processing requirements

Engineering Contradiction:
Improvemodel robustnessVSAvoidtraining time
Core Design Contradiction:
Stability of the object's compositionVSLoss of time

Solution Approach 1:

The patent uses parameter changes by adjusting the degree of incorrect annotation as a controllable variable. This allows optimization of the balance between robustness and training time - higher degrees of incorrect annotation improve robustness but increase training time, while lower degrees maintain faster training with adequate robustness. The parameter can be tuned based on specific application requirements.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent applies partial incorrect annotation rather than complete corruption, which provides sufficient robustness improvement without the full training time cost of extensive error injection. This partial approach achieves diminishing returns optimization, where moderate levels of incorrect annotation provide most of the robustness benefit with minimal training time penalty.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4417273B1Estimation program, estimation method, and information processing device
Publication Date: 2026.05.20 FUJITSU LTD
  • EP4417273B1 patent drawingFigure 1
  • EP4417273B1 patent drawingFigure 2
  • EP4417273B1 patent drawingFigure 3

AI summary

An information processing device specifies positions of a plurality of joints included in a face of a player by inputting an image in which a head of the player is in a predetermined state to a machine learning model. The information processing device estimates a position of a top of the head of the player using each of the positions of the plurality of joints.