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.
## 1. CONTENT
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](https://web.gin.g-node.org/ioannis.agtzidis/matlab_utils). We also
need to clone the *matlab_360_utils* repository that offers utilities for handling 360-degree
data from [here](https://web.gin.g-node.org/ioannis.agtzidis/matlab_360_utils). 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](https://web.gin.g-node.org/ioannis.agtzidis/360_em_dataset/src/master/participant_info.csv).
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](https://web.gin.g-node.org/ioannis.agtzidis/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](https://gin.g-node.org/ioannis.agtzidis/360_em_algorithms).
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.
## 2. DATA FORMAT
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 time INTEGER
@ATTRIBUTE x NUMERIC
@ATTRIBUTE y NUMERIC
@ATTRIBUTE confidence NUMERIC
@ATTRIBUTE x_head NUMERIC
@ATTRIBUTE y_head NUMERIC
@ATTRIBUTE angle_deg_head NUMERIC
@ATTRIBUTE labeller_1 {unassigned,fixation,saccade,SP,noise,VOR,OKN}
@DATA
0,960.00,540.00,1.00,960.00,540.00,1.22,fixation
5000,959.00,539.00,1.00,959.00,539.00,1.23,fixation
13000,959.00,539.00,1.00,959.00,539.00,1.23,fixation
18000,959.00,539.00,1.00,959.00,539.00,1.23,fixation
29000,959.00,539.00,1.00,959.00,539.00,1.24,fixation
34000,959.00,539.00,1.00,959.00,539.00,1.24,fixation
45000,959.00,539.00,1.00,959.00,539.00,1.24,fixation
49000,959.00,539.00,1.00,959.00,539.00,1.24,fixation
61000,959.00,539.00,1.00,959.00,539.00,1.24,fixation
66000,959.00,539.00,1.00,959.00,539.00,1.24,fixation
77000,959.00,539.00,1.00,959.00,540.00,1.24,fixation
82000,959.00,539.00,1.00,959.00,540.00,1.24,fixation
94000,959.00,539.00,1.00,960.00,540.00,1.24,fixation
99000,959.00,539.00,1.00,960.00,540.00,1.24,fixation
110000,959.00,539.00,1.00,960.00,540.00,1.25,fixation
114000,959.00,539.00,1.00,960.00,540.00,1.25,fixation
125000,958.00,538.00,1.00,960.00,540.00,1.26,saccade
129000,956.00,537.00,1.00,960.00,540.00,1.27,saccade
141000,948.00,530.00,1.00,960.00,540.00,1.28,saccade
```
### 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](https://web.gin.g-node.org/ioannis.agtzidis/matlab_360_utils/src/master/HeadToVideoRot.m) and
[YZXrotation.m](https://web.gin.g-node.org/ioannis.agtzidis/matlab_360_utils/src/master/YZXrotation.m)
in the
[matlab_360_utils](https://web.gin.g-node.org/ioannis.agtzidis/matlab_360_utils) repository.
## 3. General Information
Author: Ioannis Agtzidis
Contact: ioannis.agtzidis@tum.de
If you use any part of the provided data set or algorithms, please cite:
> @inproceedings{agtzidis2019dataset,
> 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)},
> year={2019},
> organization={ACM},
> doi={10.1145/3343031.3350947}
> }