EG-SIF: Improving Appearance Based Gaze Estimation using Self Improving Features
Submission Type: Full Paper
Keywords: Eyegaze, Adabins, Appearance Based, Image Quality, Multitask
TL;DR: EG-SIF segregates images based on their quality, generates a pair of good and adverse images, and applies multitask training with image enhancement .
Abstract: Accurate gaze estimation is integral to a myriad of applications, from augmented reality
to non-verbal communication analysis. However, the performance of gaze estimation models
is often compromised by adverse conditions such as poor lighting, artifacts, low-resolution
imagery, etc. To counter these challenges, we introduce the eye gaze estimation with self-
improving features (EG-SIF) method, a novel approach that enhances model robustness
and performance in suboptimal conditions. The EG-SIF method innovatively segregates eye
images by quality, synthesizing pairs of high-quality and corresponding degraded images.
It leverages a multitask training paradigm that emphasizes image enhancement through
reconstruction from impaired versions. This strategy is not only pioneering in the realm
of data segregation based on image quality but also introduces a transformative multitask
framework that integrates image enhancement as an auxiliary task. We implement adaptive
binning and mixed regression with intermediate supervision to refine capability of our model
further. Empirical evidence demonstrates that our EG-SIF method significantly reduces the
angular error in gaze estimation on challenging datasets such as MPIIGaze, improving from
4.64◦ to 4.53◦, and on RTGene, from 7.44◦ to 7.41◦, thereby setting a new benchmark in the
field. Our contributions lay the foundation for future eye appearance based gaze estimation
models that can operate reliably despite the presence of image quality adversities.
Submission Number: 28
Loading