



























Study with the several resources on Docsity
Earn points by helping other students or get them with a premium plan
Prepare for your exams
Study with the several resources on Docsity
Earn points to download
Earn points by helping other students or get them with a premium plan
Peekaboom is a game developed by luis von ahn, ruoran liu, and manuel blum for locating objects in images. The game involves processing images and guessing their labels, with features like passes, hints, pings, and a bonus round to provide examples for machine learning. Peekaboom also discusses the use of captcha for security, focusing on image recognition problems and hard ai problems as captcha challenges.
Typology: Study notes
1 / 35
This page cannot be seen from the preview
Don't miss anything!




























PeekaboomPeekaboom:
A Game for :
A Game for
Locating Objects in ImagesLocating Objects in ImagesB
L
Ah
R
L
d M
l Bl
By Luis von Ahn, Ruoran Liu, and Manuel Blum
ESP GameESP GameESP GameESP Game
y
Only attaches labels to random images
y
g
from the internet
y
Does not specify location of image
oes
ot spec y ocat o
o
age
◦
Thus, insufficient for training computer visionalgorithms
g
y
Given label can we identify where label
y
Given label, can we identify where labelactually is?
Sample Boom ImageSample Boom ImageSample Boom ImageSample Boom Image
Sample Peek ImageSample Peek ImageSample Peek ImageSample Peek Image
BUSH
Data that is collected for trainingData that is collected for trainingData that is collected for trainingData that is collected for training
y
Which pixels are necessary to guess the word?
y
From hints: what is the relation between wordand image?
y
Data from pings: which pixels are inside theobject?
y
What are the most relevant aspects of theobject?El
f
d
y
El
imination of poor image-word pairs
Dealing with single player gamesDealing with single player gamesDealing with single player gamesDealing with single player games
y
Bot can easily simulate a Boom from
y
previously saved human clicks
y
S
imulating Peek is much harder S
u at
g
ee
s
uc
a
e
Object Bounding BoxesObject Bounding BoxesObject Bounding BoxesObject Bounding Boxes
y
We can use Peekaboom results to establish tags for keywords in images
y
Example below: eyes and nose
ResultsResultsResultsResults
y
Is this an effective way to collect data?
y
Yes!
y
Game is enjoyable
y
Game is enjoyable^ ◦
Each person played average of 158.7 images ◦
That’s 72 96 minutes per person in one ◦
That s 72.96 minutes per person in onemonth!^ x
User review:
“This game is like crack. I’ve been
User review: This game is like crack. I ve beenPeekaboom-free for 32 hours…”
DiscussionDiscussionDiscussionDiscussion
y
What are some weaknesses of Peekaboom?
y
Ca
n you think of any other applications of
Ca
you t
o a y ot e
app cat o s o
Peekaboom?
Using HardUsing Hard
Using HardUsing Hard
Problems for SecurityProblems for SecurityBy Luis von Ahn Manuel Blum Nicholas J HopperBy Luis von Ahn, Manuel Blum, Nicholas J. Hopper,John Langford
More ExamplesMore ExamplesMore ExamplesMore Examples
More ExamplesMore ExamplesMore ExamplesMore Examples
What is a CAPTCHA?What is a CAPTCHA?What is a CAPTCHA?What is a CAPTCHA?
y
What kind of AI problems do we want to
p
use?
Hard problems
.^
a
p ob e
s
◦
Hard w.r.t. complexity
◦
AI community agrees it
’s hard
AI community agrees it s hard
Useful problems
What is a CAPTCHA?What is a CAPTCHA?What is a CAPTCHA?What is a CAPTCHA?
y
The fact that CAPTCHA is based on a hard AI problem gives us a win-winsituation^ ◦
Either someone figures out a bot to solveCAPTCHAs (solving a hard, useful AI problem) ◦
Or there is a way to tell humans from computers, which is useful for security