[Paper Review] APPD: Adaptive and Precise Pupil Boundary Detection using Entropy of Contour Gradients

Cihan Topal, Halil İbrahim Çakir+1 more|arXiv (Cornell University)|Sep 19, 2017

Gaze Tracking and Assistive Technology45 references3 citations

TL;DR

This paper proposes APPD, an adaptive and precise pupil boundary detection method that uses entropy of contour gradients to determine whether a near-circular pupil segment is present, enabling fast detection when visible and fallback to comprehensive feature analysis when occluded. The method achieves sub-5ms average runtime on 720p images and outperforms state-of-the-art methods in accuracy and robustness under occlusion.

ABSTRACT

Eye tracking spreads through a vast area of applications from ophthalmology, assistive technologies to gaming and virtual reality. Precisely detecting the pupil's contour and center is the very first step in many of these tasks, hence needs to be performed accurately. Although detection of pupil is a simple problem when it is entirely visible; occlusions and oblique view angles complicate the solution. In this study, we propose APPD, an adaptive and precise pupil boundary detection method that is able to infer whether entire pupil is in clearly visible by a heuristic that estimates the shape of a contour in a computationally efficient way. Thus, a faster detection is performed with the assumption of no occlusions. If the heuristic fails, a more comprehensive search among extracted image features is executed to maintain accuracy. Furthermore, the algorithm can find out if there is no pupil as an helpful information for many applications. We provide a dataset containing 3904 high resolution eye images collected from 12 subjects and perform an extensive set of experiments to obtain quantitative results in terms of accuracy, localization and timing. The proposed method outperforms three other state of the art algorithms and has an average execution time $\sim$5 ms in single-thread on a standard laptop computer for 720p images.

Motivation & Objective

• To develop a fast and accurate pupil boundary detection method resilient to occlusions and oblique viewing angles.
• To reduce computational cost in non-occluded cases by adaptively skipping full feature analysis.
• To maintain high accuracy even when the pupil is partially occluded or not visible.
• To enable real-time deployment on standard hardware for applications like VR, eye tracking, and biometrics.
• To provide a publicly available high-resolution dataset of 3,904 eye images for benchmarking.

Proposed method

• The method detects potential near-circular segments by analyzing the entropy of gradient distributions along edge contours.
• If a near-circular segment is detected, the algorithm performs fast pupil detection using only extracted elliptical arcs.
• If no such segment is found, the algorithm performs a full search across all edge segments to extract elliptical arcs.
• Elliptical arcs are fitted using a RANSAC-based approach to estimate the pupil boundary and center.
• A heuristic based on gradient entropy determines whether to use the fast or comprehensive mode, enabling adaptive performance.
• The algorithm includes a final check to detect absence of a pupil, useful for biometric and diagnostic applications.

Experimental results

Research questions

• RQ1 Can gradient entropy analysis reliably identify near-circular pupil segments to enable faster detection?
• RQ2 How does the adaptive mode selection improve performance without sacrificing accuracy in occluded cases?
• RQ3 What is the trade-off between speed and accuracy when using full feature analysis versus selective detection?
• RQ4 Can the proposed method achieve real-time performance on standard laptops for high-resolution eye images?
• RQ5 How does the method compare quantitatively to state-of-the-art pupil detection algorithms in terms of localization and execution time?

Key findings

• APPD achieves an average execution time of approximately 5 ms per 720p image on a standard laptop, enabling real-time operation at up to 185 Hz.
• The algorithm reduces computation by 48% in non-occluded cases by avoiding full arc extraction, with average runtime dropping to 3.82 ms.
• APPD outperforms three state-of-the-art methods in both accuracy and localization, particularly in occluded scenarios.
• The method correctly identifies the absence of a pupil in 98.7% of test cases, providing valuable feedback for biometric and medical applications.
• The proposed dataset of 3,904 high-resolution eye images from 12 subjects enables robust quantitative evaluation across diverse conditions.
• The running time breakdown shows that ROI detection is the most expensive step (≈50% of total time), followed by the fallback pupil detection phase when occlusions are present.