Imitation Learning for Generating Human Eye Movements from Natural Image

No Thumbnail Available

URL

Journal Title

Journal ISSN

Volume Title

Perustieteiden korkeakoulu | Master's thesis

Date

2022-10-17

Department

Major/Subject

Computer Science

Mcode

SCI3042

Degree programme

Master’s Programme in Computer, Communication and Information Sciences

Language

en

Pages

37 + 0

Series

Abstract

Generating human eye movements is important for understanding human attention mechanisms. Previous works in eye movement synthesis cannot generate eye movements for given visual stimuli. And scanpath prediction methods focus on generating a certain type of eye movements, fixation, for given images. Only a sequence of fixations cannot well represent the process of exploration by the human eye. To model human gaze behaviour better, in this paper, we introduce a new task generating human eye movements from visual stimuli. The goal of this task is to generate high-frequency eye movements similar to those captured by eye trackers for input images. To solve this task, I proposed an imitation learning based two-stage framework. The framework translated the task to generating eye movements from person-specific fixation maps. We proposed an eye movement inverse reinforcement learning environment to model the eye movements in person-specific fixation maps. The results of IQ-Learn agents trained in the proposed environment on MIT1003 dataset show a high similarity to human eye movements from a velocity perspective. In conclusion, our work makes fundamental contributions to generating human eye movements from natural images via imitation learning methods.

Description

Supervisor

Oulasvirta, Antti

Thesis advisor

Bulling, Andreas
Bâce, Mihai

Keywords

eye tracking, generating eye movements, imitation learning, inverse reinforcement learning, behavioral modeling

Other note

Citation