GlanceWriter: Writing Text by Glancing Over Letters with Gaze

Title of the Paper

GlanceWriter: Writing Text by Glancing Over Letters with Gaze

Paper Information

• Research Area: Human-Computer Interaction, Eye-Tracking Technology, Contactless Text Input Methods
• Keywords: Eye-tracking, Text Input, Gesture Input, No Gaze Fixation, Dynamic Decoding

Research Background and Problem Statement

• Identified Issues or Challenges: Current eye-tracking-based text input methods suffer from slow input speed and eye fatigue. Traditional methods often require users to fixate on individual keys or perform additional operations to mark the start and end positions, which results in time-consuming, inefficient interactions and induces eye strain.
• Importance of the Problem: Providing an efficient and comfortable hands-free input solution can enhance human-computer interaction experiences in scenarios such as virtual reality, augmented reality, and for users with physical disabilities.
• Research Motivation and Related Work:
  • Traditional gaze-based input and gesture input methods face challenges in input efficiency and reliability.
  • EyeSwipe uses a "reverse crossing" mechanism to mark the start and end positions, but its operational steps are complex and time-consuming.
  • Commercial systems (e.g., Tobii Dynavox's Communicator 5) employ unknown decoding mechanisms but still face limitations in speed and error rates.
  • To address these limitations, the study proposes a method that eliminates the need for gaze fixation or additional marking of path start and end points, achieving a smoother input experience through probabilistic decoding.

Solution

• Proposed Method or Solution:

  • GlanceWriter is an eye-tracking-based text input method that does not require gaze fixation or "reverse crossing." Users can input words by quickly glancing over characters on a keyboard.
  • It employs a dynamic decoding algorithm that analyzes dynamic characteristics of gaze paths (e.g., speed, stability, distance from keyboard center) to probabilistically determine input characters.
  • A Trie (prefix tree) is used to store the vocabulary, combined with a language model to optimize word suggestion candidates.

• Innovations:

  • The innovative probabilistic decoding algorithm eliminates the need for explicit marking of input path start and end points.
  • A scoring mechanism based on gaze stability and path distance is designed to improve decoding accuracy.
  • Provides non-crossing and non-fixation path cancellation functionality, reducing instability issues in gaze-based input.

• Implementation Steps and Techniques:

  • Divide the keyboard area into regions corresponding to characters, and record user gaze paths through eye-tracking sampling (100Hz).
  • Evaluate character input probabilities using Gaussian functions and stability calculations.
  • Quickly match user gaze paths to the vocabulary using a Trie structure, and enhance prediction accuracy with a language model.
  • Sort candidate words after decoding and confirm the final selection through gaze fixation.

Research Outcomes

• Specific Results:

  • In terms of input speed, GlanceWriter improved by 67% compared to EyeSwipe (from 6.49 WPM to 10.89 WPM) and by 28.7% compared to the commercial system Communicator 5 (from 7.41 WPM to 9.54 WPM).
  • Error rates significantly decreased: GlanceWriter achieved a 2.71% error rate compared to EyeSwipe's 6.85%, and also improved over Communicator 5 (from 16.32% to 12.89%).
  • Users reported significantly better subjective experiences compared to other input methods, including lower mental load and higher comfort and learnability.

• Advantages:

  • Eliminates the need for complex reverse crossing or gaze fixation operations, significantly enhancing the fluidity and efficiency of the input experience.
  • The decoding algorithm demonstrates strong robustness and adaptability to unstable gaze inputs.

• Experimental or Evaluation Results:

  • Two experiments compared GlanceWriter with EyeSwipe and Tobii Communicator 5, validating its performance advantages.
  • User feedback indicated that GlanceWriter outperformed other methods in subjective evaluations of mental load, physical load, comfort, and learning difficulty.

• Limitations and Future Directions:

  • The current decoding algorithm relies on a vocabulary and cannot directly handle out-of-vocabulary (OOV) words.
  • Further research and implementation are needed for OOV word input mechanisms.
  • The introduction of language models significantly impacts performance; future work could explore more advanced models such as GPT-3.
  • Consider optimizing the decoding performance for edge characters by addressing the influence of keyboard region positions on input accuracy.