Accurate Pupil Center Detection in Off-the-Shelf Eye Tracking Systems Using Convolutional Neural Networks

Abstract

Remote eye tracking technology has suffered an increasing growth in recent years due to its applicability in many research areas. In this paper, a video-oculography method based on convolutional neural networks (CNNs) for pupil center detection over webcam images is proposed. As the first contribution of this work and in order to train the model, a pupil center manual labeling procedure of a facial landmark dataset has been performed. The model has been tested over both real and synthetic databases and outperforms state-of-the-art methods, achieving pupil center estimation errors below the size of a constricted pupil in more than 95% of the images, while reducing computing time by a 8 factor. Results show the importance of use high quality training data and well-known architectures to achieve an outstanding performance.

A YOLO Detector Providing Fast and Accurate Pupil Center Estimation using Regions Surrounding a Pupil

“…From the results, the winner detector was the one that was trained with α = 3 and β = 1.8125 datasets at the checkpoint of 6,000 because this detector gave rise to the best accuracy of detection and almost all the best relative errors. 3 compares the relative error for pupil center estimation on the GI4E dataset between the winner detector against the previous approaches, which produced the relative error emax ≤ 0.025 greater than or equal to 79.50% [23]. It is noticeable that all previous approaches employ DNN to create the detectors.…”

“…Because of its shallow network architecture, this configuration has demonstrated real-time detection performance even without GPU acceleration on the computer platform [26]. It should be noted that the state-of-the-art approach presented by Larumbe-Bergera et al [23] makes use of the Resnet-50 backbone with the integration of five additional layers. Layers are more than double the number of YOLOv3 in the tiny configuration.…”

“…To compare our approach to the state-of-the-art one proposed by Larumbe-Bergera et al [23], all the datasets used in their publication were also used in our experiments. Briefly, these datasets include:…”

“…Accuracy is calculated as a percentage of images for which this error falls below certain thresholds. State-of-the-art approaches, especially the one benchmarked by our approach [23], are commonly compared using emax ≤ 0.025(2.5%), emax ≤ 0.050(5%) and emax ≤ 0.100(10%).…”

“…The approach achieved high accuracy and speed, with a precision error of less than 10 mm and a 5 𝑚𝑠 processing time on a limited proprietary test dataset. Larumbe-Bergera et al [23] provided the research community with an improved version of the facial landmark dataset. They also proposed a novel pupil center estimation approach that extends the Resnet-50 backbone by incorporating a pooling layer and four additional fully connected layers to detect eye landmarks used to estimate a pupil.…”