Brown Deep Learning Day Fall 2022

Project Checkin 1

Title: Mimicking Natural Eye Gaze
Ji Won Chung (jchung97)
Anita de Mello Koch (ademello)
Arthur Chen (kchen157)
Skye Thompson (rthomp12)

Introduction
With the rise of portable, accessible technology the use of augmented reality has increased. The idea of virtual assistants and generated images that react to the user are now more possible than ever before. However, we still cannot simulate natural eye gaze, instead creating virtual assistants that give off a feeling of uncanny valley. We hope to start moving towards simulating natural eye gaze by first predicting the rest of the eye trajectory given some starting trajectory.

Related Works
There are several works that use LSTM and CNN networks on eye gaze data. Work by Sodoké et al focuses on learning how to filter out noise that is often found in eye gaze predictions using a deep convolutional LSTM. Koochaki et al. try to predict the intent of a user who uses an eye-based interface on computers using CNNs and LSTMs. Both of these works provide a starting point for our project. Additionally, we wish to incorporate natural blinking. For this to be included in the model it is important for us to extract blink intervals from the original dataset. There are works for this on github, for example https://github.com/pathak-ashutosh/Eye-blink-detection.

Data
We will be using the webgazer dataset (https://webgazer.cs.brown.edu/data/). This data was collected using 51 participants in an eye-tracking study. The data includes the user input data (mouse and cursor logs), screen recordings, webcam video of the participants face and eye-gaze locations predicted by a Tobii Pro X3-120 eye tracker. We want to augment this data by also including the participant's blink information which requires us extracting this data from the webcam footage and augmenting the eye-gaze locations.

Methodology
We will be attempting a CNN and LSTM architecture. Eye gaze is dependent on the previous trajectory and so requires a model that has memory. Similarly, there have been previous works that show that these architectures are well suited to this type of problem.

Metrics
We will be predicting the future trajectory based on the previous trajectory. As such we can train on the eye-gaze data from webgazer by splitting the trajectories to form the input and the desired output. As such we can use accuracy as our success metric.

Ethics
Our data contains webcam footage of the participants and so must be carefully handled. Additionally we have only 51 participants so it is likely that the demographics of the participants are unbalanced. Additionally, we intend to use this data to mimic natural eye gaze however what people look at and interact with their eyes can be cultural. As such we should be careful to not become to biased to any one group or culture.
This problem is well suited to deep learning because large amounts of data can be easily collected and existing methods that do not use learning are not able to simulate natural eye gaze. This implies that the problem is difficult to simulate with a simple model which could be improved by learning a model.

Division of labor
Ji Won Chung - LSTM
Anita de Mello Koch - CNN
Arthur Chen - Blink extraction
Skye Thompson - data augmentation

Sodoké, K., Nkambou, R., Dufresne, A. and Tanoubi, I., 2020. Toward a deep convolutional LSTM for eye gaze spatiotemporal data sequence classification. In EDM.
Koochaki, F. and Najafizadeh, L., 2019, July. Eye gaze-based early intent prediction utilizing cnn-lstm. In 2019 41st Annual international conference of the IEEE engineering in medicine and biology society (EMBC) (pp. 1310-1313). IEEE.