TLDR: Researchers developed a new dataset, MobilePoG, and a “dynamic calibration” method for mobile phone gaze estimation. This method, where users move their phone during calibration, significantly improves accuracy and robustness to head movements, overcoming limitations of traditional static calibration. Their findings emphasize that diverse head poses during calibration are crucial for effective gaze tracking.
Gaze estimation, which predicts where a person is looking, is a rapidly evolving field in computer vision. It holds immense potential for various applications, from medical diagnosis and human-computer interaction to virtual/augmented reality and assisted driving. Modern appearance-based methods, powered by deep learning, can predict gaze coordinates on a screen using just a front-facing camera on devices like mobile phones, eliminating the need for specialized hardware.
However, despite significant advancements, these gaze estimation models face a persistent challenge: they often struggle to generalize across different individuals. This is primarily due to unique personal differences, such as variations in eye anatomy. To overcome this, personalized calibration is typically required, where the model is adapted to a specific user before deployment.
The conventional approach to personalized calibration involves users fixating on a few designated points while their facial images are recorded. The problem is that these calibrated models are highly sensitive to head pose variations. If a user’s head moves differently during actual use compared to during calibration, the accuracy of gaze estimation can drop significantly. This sensitivity stems from the static nature of traditional calibration, where users are asked to maintain a fixed head pose.
To address this critical limitation, researchers Yujie Zhao, Jiabei Zeng, and Shiguang Shan investigated the key factors influencing calibrated estimators and explored strategies for more robust calibration. Their work is detailed in their research paper, “Pose-Robust Calibration Strategy for Point-of-Gaze Estimation on Mobile Phones,” which can be found here: RESEARCH_PAPER_URL.
Introducing MobilePoG: A New Benchmark Dataset
A major hurdle in studying calibration strategies was the lack of suitable datasets. Existing mobile gaze datasets often lacked precise control over head pose during data collection, making it difficult to simulate diverse calibration scenarios. To fill this gap, the researchers constructed a new benchmark dataset called MobilePoG. This dataset includes facial images from 32 individuals, captured while they focused on specific points under two conditions: either fixed head poses or continuously changing head poses.
MobilePoG is divided into two subsets: Static-MobilePoG, where participants maintain one of 12 fixed head pose configurations, and Dynamic-MobilePoG, where participants move and rotate their phones while fixating on points, naturally introducing continuous head pose variations. This unique design allows for systematic analysis of how different calibration points and head poses influence estimation accuracy.
Key Findings: The Importance of Pose Diversity
Experiments conducted on the Static-MobilePoG dataset revealed a crucial insight: introducing a wider range of head poses during calibration significantly improves the estimator’s ability to handle pose variation. In contrast, simply increasing the number of calibration points while maintaining a fixed head pose offered only marginal improvements, and in some cases, even led to overfitting.
The analysis showed that models calibrated with monotonous head poses failed to generalize well to varying head poses encountered in real-world scenarios. This highlighted that head pose diversity is a principal factor driving personalized calibration performance and should be prioritized in practical calibration strategies.
The Proposed Solution: Dynamic Calibration Strategy
Building on these findings, the researchers proposed a novel dynamic calibration strategy. Instead of asking users to remain static, this strategy instructs users to fixate on calibration points while smoothly moving and rotating their mobile phones. This simple yet effective approach naturally introduces diverse head pose variations during the calibration process.
The Dynamic-MobilePoG dataset was specifically designed to simulate and evaluate this new strategy. Experimental results demonstrated that the dynamic calibration strategy consistently outperformed conventional static strategies across various models and calibration algorithms. It not only significantly enhanced the robustness of the calibrated model to varying head poses but also led to more stable performance during the calibration process, making it more practical for real-world applications.
Also Read:
- Advancing Facial Expression Recognition with Personalized Feature Translation
- AI Generates Realistic Driver Data to Improve Distraction Detection
Conclusion
The research successfully identified that head pose diversity in calibration samples is the main bottleneck for personalized point-of-gaze estimation on mobile phones. By introducing the MobilePoG dataset and proposing a user-friendly dynamic calibration strategy, the authors have provided a robust solution that significantly improves the accuracy and generalization of gaze estimators, paving the way for more reliable and practical gaze-based applications on mobile devices.
