turkergaze: crowdsourcing saliency with webcam based eye...
TRANSCRIPT
TurkerGaze: Crowdsourcing Saliency with Webcam based Eye TrackingPingmei Xu Krista A. Ehinger Yinda Zhang Adam Finkelstein Sanjeev R. Kulkarni Jianxiong Xiao
Motivations
System Overview
Our Approach
http://isun.cs.princeton.eduVISI N GROUP
1. Traditional eye tracking requires special hardware. 2. Collecting data is expensive, tedious, slow. 3. Data cannot be collected from crowd sourcing platform, e.g. AMT.
Data Collection
Dataset & Benchmark1. Small datasets limit the potential for training data intensive algorithm2. Benchmark can be easily overfit.
1. We propose a webcam-based gaze tracking system, which can collect large-scale data from Amazon Mechanical Turk2. We apply a carefully designed gaming protocol to ensure the quality of the collected data.
1. Efficiency: Real-time performance in web browser.2. Uncontrolled environment: Head movement & Lack of attention.3. Privacy issue: Cannot streaming recorded video.
Challenging
Gaming Protocol
System Performance
Crowdsourcing vs. Lab Collected
0 256 512 768 1024
192
384
576
768
Screen WidthSc
reen
Hei
ght
Error in pixelError in pixel
0 100
Time (ms)0 500 1000 1500 2000 2500
X
200
300
400
500
600
700
800
Time (ms)0 500 1000 1500 2000 2500
Y
200
300
400
500
600EyelinkOurs
0
0.5
1
head fla
gtex
tclo
ck
paint
ingscr
eenpic
ture
signbo
ttle
flowers
desk
lamp
cush
ion glass
perso
nva
serai
ling
stree
tlight
pillow
hous
e
doub
le do
or
sconc
e
buildi
ngsmirro
rpa
ne door
book
window
baske
t
buildi
ng
armch
air
5%10%20%40%60%80%
head flag text clock painting screen picture sign bottle flowers desk lamp cushion glass person vase
railing streetlight pillow house double door sconce buildings mirror pane door book window basket building armchair
~20 21~30 31~40 41~50 51~50 60~0%
10%
20%
30%
40%
0%
47% male female
Age-Gender Distribution
Asian Hispanic White Native Black Other Unknown0%
10%
20%
30%
40%
50%
0%
56%
0.18
0.08
0.56
0.02 0.02 0.02
0.12
Race DistributionGeographic Distribution
Mechanical Turk Worker
Fixations from Judd et al. (2009) Fixations from AMTurk workers
Calibration Train Regression Model Prediction
1. Leverage the impact of these well known games to attract more workers.2. Encourage workers to work more carefully.3. Integrate the quality control to the experiment naturally.
Gaze prediction accuracy evaluated by a 33-point chart
Comparison with professional eye trackerRed: Eyelink Green: Our system
Fixation detection results. Our meanshift based algorithm on low sampling rate data provides satisfactory estimations. Red: ground truth, Yellow: estimation.
Statistics of basic information for workers from AMTurk
Distribution of fixations collected in the lab and on AMTurk
20 40 60 800.5
0.6
0.7
0.8
0.9
1
JuddAMTurk
20 40 60 80
20 40 60 80
Fixation prediction with 14 Judd subjects’ fixations
iSUN Database
Video Saliency Panoramic Saliency
Example of the saliency data obtained by our system. Object saliency statistics. Overlapping with top percents of saliency map.
Eye tracking saliency on SUN database: 9000 images (expected to label 20608 images)1. Saliency data on scene images. 2. Pixelwise semantic labelling + Scene category