Appearance-Based Gaze Estimation With Deep Learning: A Review and Benchmark

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Human gaze is vital for understanding focus and intent.
• Deep learning has advanced appearance-based gaze estimation.
• Lack of standardized guidelines hinders deep learning gaze estimation algorithm development due to comparison inconsistencies.

Purpose of the Study:

• To systematically review deep learning-based appearance-based gaze estimation methods.
• To establish fair comparison metrics and address pre-processing/post-processing variations.
• To provide a comprehensive benchmark and development guidelines for future research.

Main Methods:

• Surveying deep learning gaze estimation algorithms across the pipeline: feature extraction, model design, calibration, and platforms.
• Summarizing pre-processing and post-processing techniques for fair performance comparison (e.g., face/eye detection, data rectification, 2D/3D gaze conversion).
• Establishing a benchmark including dataset characterization and source code for typical algorithms.

Main Results:

• A systematic review of current deep learning gaze estimation techniques is presented.
• Standardized methods for pre-processing and post-processing are summarized to enable fair comparisons.
• A comprehensive benchmark with public datasets and source code is provided.

Conclusions:

• This work offers a valuable reference for developing deep learning-based gaze estimation methods.
• It serves as a guideline to standardize and advance future research in gaze estimation.
• The benchmark and reviewed methods aim to improve the reliability and comparability of gaze estimation algorithms.