Eye-Gaze Data Capture Layout for Edge and Corner Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional eye-gaze tracking systems face accuracy degradation at screen edges and corners due to insufficient training data and non-uniform distribution, leading to less accurate predictions, particularly when users interact with common interface elements like buttons.
Innovation Solution
A grid-based approach with uniformly distributed regions and a non-overlapping scan path is used to capture eye-gaze data, ensuring more data points at screen boundaries, and multiple cameras are employed to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If evenly distributed eye-gaze training data is collected across the screen, then the calibration process covers the entire screen area, but the prediction accuracy degrades at screen edges and corners due to insufficient neighboring data points
Solution Approach 1:
The patent applies local quality by creating non-uniform data sampling density across different screen regions. Specifically, the scan path spends more time and captures more data points at screen edges and corners where accuracy degradation occurs, while using fewer points in the screen center. This localized adjustment of data quality resolves the contradiction by concentrating training resources where they are most needed to improve prediction accuracy at critical boundary regions.
Solution Approach 2:
The patent uses preliminary action by performing a preliminary analysis of the screen layout to identify regions with insufficient training data coverage, particularly edges and corners. Based on this analysis, the scan path is pre-configured to prioritize these under-sampled regions. This preliminary preparation ensures that the main calibration process automatically generates adequate training data density at problem areas without requiring real-time adjustments during data collection.
2Measurement precision
If more training data is collected to improve prediction accuracy at screen boundaries, then the accuracy of eye-gaze prediction improves, but the time required for calibration increases causing stress on the operator
Solution Approach 1:
The patent applies dynamics by making the scan path adaptive and flexible rather than static. The scan path dynamically adjusts its trajectory, speed, and data sampling rate based on real-time feedback from preliminary analysis and actual data collection quality. This dynamic approach allows the system to collect sufficient training data at edge and corner regions while maintaining overall calibration time within acceptable limits, resolving the contradiction between accuracy improvement and time consumption.
Solution Approach 2:
The patent uses partial action by focusing calibration efforts selectively on only the most critical regions (edges and corners) rather than uniformly increasing data collection across the entire screen. This targeted approach achieves the necessary improvement in prediction accuracy at problem areas without requiring excessive calibration time, as full-screen uniform oversampling would be unnecessarily time-consuming.
3Productivity
If a non-overlapping scan path is used to cover the whole screen, then the data collection time is minimized and each gaze point is captured once, but the uniform distribution of training data across regions becomes harder to achieve
Solution Approach 1:
The patent applies asymmetry by designing an intentionally non-uniform scan path that does not treat all screen regions equally. Instead of symmetric uniform sampling, the scan path creates asymmetric data distribution with higher density at edges and corners and lower density in the center. This asymmetric approach efficiently balances the contradiction by achieving adequate coverage of the entire screen through a single pass while concentrating resources where they provide maximum benefit to prediction accuracy.
Data Source
AI summary
Systems and methods are provided for collecting eye-gaze data for training an eye-gaze prediction model. The collecting includes selecting a scan path passing through a series of regions of a grid on a screen of a computing device, moving a symbol as an eye-gaze target along the scan path, and receiving facial images at eye-gaze points. The eye-gaze points are uniformly distributed within the respective regions. Areas of the regions that are adjacent to edges and corners of the screen are smaller than other regions. The difference in areas shifts centers of the regions toward the edges, density of data closer to the edges. The scan path passes through locations in proximity to the edges and corners of the screen for capturing more eye-gaze points in the proximity. The methods interactively enhance variations of facial images by displaying instructions to the user to make specific actions associated with the face.