Dominant Eye Selector for Gaze Estimation Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current gaze estimation systems face limitations in accurately determining the dominant eye and predicting the point-of-regard due to variations in human eye structure, ocular dominance, and environmental factors, requiring complex calibration and being sensitive to image quality and obstructions.
Innovation Solution
A real-time, unsupervised machine-learning based system that uses unlabeled images to automatically determine the dominant eye by extracting eye patches and employing a convolutional neural network to predict the point-of-regard, without the need for explicit calibration or human intervention, and can function effectively in various lighting conditions and occlusions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If appearance-based or model-based gaze estimation methods are used, then gaze position can be estimated, but accuracy is limited due to eye structure variations, ocular dominance, and environmental factors
Solution Approach 1:
The system performs self-calibration by automatically determining ocular dominance and eye-specific parameters from captured images without requiring manual user input or external hardware calibration. The calibration process is autonomous, using the imaging module to capture images and the processor to extract features and determine dominance, eliminating the need for technical expertise or human intervention.
Solution Approach 2:
The system dynamically adjusts gaze estimation parameters based on detected eye-specific characteristics. By determining ocular dominance and eye structure variations, the system modifies the mapping functions and regression parameters to accommodate individual differences, thereby maintaining accuracy across diverse eye structures and environmental conditions.
2Measurement precision
If external hardware and physical calibration are used, then reasonable accuracy can be achieved, but the system requires human intervention and technical expertise
Solution Approach 1:
The system performs self-calibration by automatically determining ocular dominance and eye-specific parameters from captured images without requiring manual user input or external hardware calibration. The calibration process is autonomous, using the imaging module to capture images and the processor to extract features and determine dominance, eliminating the need for technical expertise or human intervention.
Solution Approach 2:
The system extracts only the necessary calibration information directly from standard images captured by the imaging module. By isolating eye regions and extracting relevant features (pupil contours, glint positions, eye corners), the system obtains calibration data without requiring specialized equipment or complex external hardware setups.
3Productivity
If standard gaze estimation systems are used, then gaze can be tracked, but they are sensitive to image quality, occlusions, and environmental conditions
Solution Approach 1:
The system is designed to function with partial eye visibility by adjusting the calibration and estimation process when occlusions are detected. By determining ocular dominance and adapting to the visible eye, the system maintains operational capability even when one eye is partially or fully occluded, ensuring continuous productivity under degraded conditions.
Solution Approach 2:
The system dynamically adapts its calibration parameters and estimation algorithms based on real-time detection of eye conditions. By continuously monitoring image quality and occlusion levels, the system adjusts its approach to maintain reliable gaze tracking despite varying environmental conditions and partial occlusions.
Data Source
AI summary
The disclosure relates to systems, methods, and programs for implicitly determining the relation of the eyes associated with gaze inference, including salient features (e.g., ocular dominance) that are not visible within a digital frame, allowing for real-time determination of the user's dominant eye for the purpose of gaze estimation, and point-of-regard mapping onto a 2D plane, achieved through an end-to-end training of an eye selector agent with domain-expertise knowledge embedded in an unsupervised manner via a convolutional deep neural network training process.