Data set of 360-degree equirectangular videos, gaze recordings, eye movement (EM) ground-truth and an automatic EM classification algorithm

Ioannis Agtzidis fea59eeded Updated README 2 years ago
em_algorithms 53b320f8aa Initial commit 2 years ago
gaze 53b320f8aa Initial commit 2 years ago
ground_truth 53b320f8aa Initial commit 2 years ago
output_I-S5T_E+H 53b320f8aa Initial commit 2 years ago
output_I-S5T_FOV 53b320f8aa Initial commit 2 years ago
output_I-S5T_combined 53b320f8aa Initial commit 2 years ago
videos 53b320f8aa Initial commit 2 years ago
LICENSE 53b320f8aa Initial commit 2 years ago fea59eeded Updated README 2 years ago
participant_info.csv 66efb649c1 Added participant info 2 years ago

Here we provide 15 360-degree equirectangular videos, together with the eye tracking recordings of 13 subjects and a manually labelled ground-truth subset of the gaze recordings. We also offer information about the manual labelling tool in Section 1.4. Finally we provide a algorithmic implementation of the the process that was followed during manual labelling.


Before starting using the provided algorithms of this repository you should first clone (or download) the matlab_utils repository that offers utilities for mainly handling ARFF files from here. We also need to clone the matlab_360_utils repository that offers utilities for handling 360-degree data from here. Then add the previous folders to the search path of Matlab with the pathtool or addpath commands.

1.1 Video Stimuli

We used as an approximation to naturalistic stimuli 14 Youtube videos from diverse contexts. The videos were published under the Creative Commons license and we give attribution to the original creators by attaching the Youtube IDs at the end of each video clip.

A single synthetic stimulus was created by the authors and comprises of a moving target, which tries to elicit many different kinds of eye motion (i.e. fixations, saccade, SP, head pursuit, VOR, OKN).

Files found in the videos folder.

1.2 Gaze Recordings

In total 13 subjects participated in our study, which amounts to ca. 3.5 hours of eye tracking data. Information about the participants can be found here.

Gaze files are found in the gaze folder.

1.3 Ground Truth

We manually labelled part of the full data set according to the rules presented in our paper. The labelled gaze recordings were split in two non overlapping (subject wise) subsets, where one can be used as training and the other as testing. In total the hand-labelled portion comprises of 2 labelled gaze files per video stimulus and about 16 % of the data.

Manually annotated ground-truth files are found in the ground_truth folder.

1.4 Manual Labelling Interface

The manual labelling interface can be found in the GTA-VI repository. The GTA-VI repository contains the extension that was developed for labelling this data set and enables labelling of 360-degree equirectangular recordings with our two tier method.

1.5 Algorithmic Implementation

We provide an algorithmic implementation for eye movement classification based on the definitions that we provide in our paper. If you want to run the algorithms you should clone and add to your Matlab path the directory that contains algorithms for basic eye movement detection (e.g. saccades, blinks, etc.), which can be found here.

The resulting eye movements after applying our algorithms to ground truth files can be found in the relevant files.

The algorithms and their output are found in the em_algorithms, output_I-S5T_combined, output_I-S5T_FOV, output_I-S5T_E+H folders.


For the whole analysis we use the ARFF data format for input and output from the disk. The initial ARFF format was extended as described in Agtzidis et al. (2016) and was further expanded for 360-degree gaze data.

Here the "@RELATION" is set to gaze_360 to distinguish the recordings from plain gaze recordings. We also make use of the "%@METADATA" special comments which describe the field of view of the used headset. Apart from the default metadata width_px, height_px, distance_mm, width_mm, height_mm we also use the extra metadata fov_width_px, fov_width_deg, fov_height_px, fov_height_deg that describe the headset properties.

The "@ATTRIBUTE" x and y represent the gaze coordinates in the equirectangular space (i.e. the head and eye position together). The x_head and y_head attributes represent head coordinates in the equirectangular space. The angle_deg_head attribute is equivalent to the roll principal axis.

2.1 ARFF Example

@RELATION gaze_360

%@METADATA distance_mm 0.00
%@METADATA height_mm 0.00
%@METADATA height_px 1080
%@METADATA width_mm 0.00
%@METADATA width_px 1920

%@METADATA fov_height_deg 100.00
%@METADATA fov_height_px 1440
%@METADATA fov_width_deg 100.00
%@METADATA fov_width_px 1280

@ATTRIBUTE angle_deg_head NUMERIC
@ATTRIBUTE labeller_1 {unassigned,fixation,saccade,SP,noise,VOR,OKN}


2.2 Recovery of HMD Pose

The x_head, y_head, angle_deg_head allow us to recover the headset pose because we use 360-degree equirectangular videos and therefore exist no translations during video presentation. For understanding the process take a look at functions HeadToVideoRot.m and YZXrotation.m in the matlab_360_utils repository.

3. General Information

Author: Ioannis Agtzidis


If you use any part of the provided data set or algorithms, please cite:

title={360-degree Video Gaze Behaviour: A Ground-Truth Data Set and a Classification Algorithm for Eye Movements},
author={Agtzidis, Ioannis and Startsev, Mikhail and Dorr, Michael},
booktitle={Proceedings of the 27th ACM International Conference on Multimedia (MM ’19)},