can you see it? annotating image regions based on users' gaze information

Can you see it? Annotating Image Regions based on Users' Gaze Information

Ansgar Scherp, Tina Walber, Steffen Staab

Technical University of ViennaOctober 2012

A. Scherp, T. Walber, S. Staab – Identifying Objects in Images Slide 2 of 40

Benefiting of Eye Tracking

Information for Image Region Annotation

Idea


Eye-tracking Hardware

X60


Saccade Fixation

Recorded Data


Scenario: Image Tagging

Find specific objects in images Analyzing the user’s gaze path

sidewalk

car

store

tree

people

girl


Investigation in 3 Steps

Gaze + Manual Regions

Interactive Tagging Application

Gaze + Automatic Segments

3

2

1


1st Step

1.Best fixation measure to find the correct image region given a specific tag?

2. Can we differentiate two regions in the same image?


3 Steps Conducted by Users

Look at red blinking dot Decide whether tag can be seen (“y” or “n”)


Dataset LabelM community images

Manually drawn polygons Regions annotated with tags

182.657 images (August 2010)

High-quality segmentation and annotation

Used as ground truth

http://labelme.csail.mit.edu/Release3.0/


Dataset (continued)


Experiment Images and Tags Randomly selected images from LabelMe Each image: at least two regions, 1000p x 700p

Created three sets of 51 images each Assigned a tag to each image

Keep subjects concentrated during experiment

Tags are either “true” or “false” “true” object described by tag can be seen “false” object cannot be seen on the image


Subjects & Experiment System 30 subjects

21 male, 9 female (age: 22-45, Ø=28.7) Undergrads (10), PhD (17), office clerks (3)

Experiment system Simple web page in Internet Explorer Standard notebook, resolution 1680x1050 Tobii X60 eye-tracker (60 Hz, 0.5° accuracy)


Conducting the Experiment Each user looked at 51 tag-image-pairs First tag-image-pair dismissed

94.6% correct answers Roughly equal for true/false tags ~2.8s avg. until decision (true), ~3.8s avg. (false)

Users felt comfortable during the experiment (avg.: 4.4, SD: 0.75)

Eyetracker did not much influence comfort


Pre-processing of Eye-tracking Data Obtained 799 gaze paths from 30 users where

Image has “true” tag assigned Users gave correct answers

Fixation extraction Tobii Studio’s velocity & distance thresholds Fixation: focus on particular point on screen

One fixation inside or near the correct region 656 gaze paths fulfill this requirement (82%)


Analysis of Gaze Fixations (1) Applied 13 fixation measures on the 656 paths (2 new, 7 standard Tobii , 4 literature)

Fixation measure: function on users’ gaze paths Calculated for each image region, over all users

viewing the same tag-image-pair


Considered Fixation MeasuresNr Name Favorite region r Origin

1 firstFixation No. of fixations before 1st on r Tobii

2 secondFixation No. of fixations before 2nd on r [13]

3 fixationsAfter No. of fixations after last on r [4]

4 fixationsBeforeDecision

fixationsAfter, but before decision New

5 fixationsAfterDecision

fixationsBeforeDecision and after New

6 fixationDuration Total duration of all fixations on r Tobii

7 firstFixationDuration

Duration of first fixation on r Tobii

8 lastFixationDuration

Duration of last fixation on r [11]

9 fixationCount Number of fixations on r Tobii

10 maxVisitDuration Max time first fixation until outside r Tobii

11 meanVisitDuration Mean time first fixation until outside r Tobii

12 visitCount No. of fixations until outside r Tobii

13 saccLength Saccade length, before fixation on r [6]


Analysis of Gaze Fixations (2)

For every image region (b) the fixation measure is calculated over all gaze paths (c)

Results are summed up per region Regions ordered according to fixation measure If favorite region (d) and tag (a) match, result is

true positive (tp), otherwise false positive (fp)


Precision per Fixation Measure

meanVisitDuration

Sum

of t

p an

d fp

ass

ignm

ents

Fixation measures

P

fixationsBeforeDecision

lastFixationDuration

fixationDuration


Adding Boundaries and Weights Take eye-tracker inaccuracies into account Extension of region boundaries by 13 pixels

Larger regions more likely to be fixated Give weight to regions < 5% of image size

lastFixationDuration increases to P = 0.65


Weighted Measure Function

Measure function fm(r) on region r with m=1…13

Relative region size: sr

Threshold when weighting is applied: T

Maximum weighting value: M


Weighted Measure Function


Examples: Tag-Region-Assignments


Comparison with Baselines

Naïve baseline: largest region r is favorite Salience baseline: Itti et al., TPAMI, 20(11), Nov 1998 Random baseline: randomly select favorite r

Gaze / Gaze* significantly better (all tests: p < 0.0015) Least significant result X2=(1,N=124)=10.723

P


Effect of Gaze Path Aggregation

+46%

+7%

Aggregation of precision P for Gaze*

P

# of gaze paths used


Research Questions



lastFixationDuration with precision of 65%


Experiment Images and Tags Randomly selected images from LabelMe Images contained at least two tagged regions Organized in three sets of 51 images each

Assigned a tag to each image

Tags are either “true” or “false”

Two of the image sets share the same images Thus, these images have two tags each


Differentiate Two Objects Use first and second tag set to identify different objects in the same images

16 images (of our 51) have two “true” tags 6 images had two correct regions identified

Proportion of 38%

Average precision for single object is 63%

Correct tag assignment for two images: 40%


Correctly Differentiated Objects


Research Questions



lastFixationDuration with precision of 65%

Accuracy of 38%






3

2

1


=

car + +

For 63% of the images, we can identify the correct region.

T. Walber, A. Scherp, and S. Staab: Identifying Objects in Images from Analyzing the Users' Gaze Movements for Provided Tags, MMM, Klagenfurt,Austria, 2012.

car

So far …


=

car + +

car

Automatic segmentation LabelMe segments only used as ground truth

T. Walber, A. Scherp, and S. Staab: Can you see it? Two Novel Eye-Tracking-Based Measures for Assigning Tags to Image Regions, MMM, Huangshan, China, 2013.

Now:


2nd Step: New Measure Automatic segmentation measure Berkeley Segmentation Data Set and Benchmarks 500 (BSDS500)

Berkley‘s bPb-owt-ucm algorithm Segmentation on different hierarchy levels Combination of contour detection and segmentation

Oriented Watershed Transform and Ultrametric Contour Map

P. Arbeléz, M. Maire, C. Fowlkes, and J. Malik. Contour detection andhierarachical image segmentation. IEEE TPAMI, 33(5):898–916, May 2011.


Segmentation Example Segmentations with different k = 0 … 0.4


Automatic Segments + Gaze Conducted same computations as before But on the automatically extracted segments


Results for different k’s: P/R/FP P

Eye-tracking-based automatic segmentation measure

Golden sections rule baseline


Baseline: Golden Sections Rule

a + b / a = a / b


Best Precision & Best F-measure

Eye-tracking-based automatic segmentation measure significantly outperforms golden sections baseline

Also shown: eye-tracking-based heatmap measure (no automatic segmentation)






3

2

1


3rd Step: Interactive Application

► tree_car ; house ; girl


APPENDIX


Influence of Red Dot

First 5 fixations, over all subjects and all images


Experiment Data Cleaning Manually replaced images with

a) Tags that are incomprehensible, require expert-knowledge, or nonsense

b) Tag refers to multiple regions, but not all are drawn into the image (e.g., bicycle)

c) Obstructed objects (bicycle behind a car)

d) “False”-tag actually refers to a visible part of the image and thus were “true” tags


How to Compute P/R? Rfav is calculated from

Automatic segmentation measure Baseline measure

can you see it? annotating image regions based on users' gaze information

Technology

images slide

experiment images

selected images

specific objects

tina walber

image tags

image region annotation

image region b