104
8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 1/104 Lecture 6 Scenes and objects 6.870 Object Recognition and Scene Understanding http://people.csail.mit.edu/torralba/courses/6.870/6.870.recognition.htm

MIT6870_ORSU_lecture6: Scenes and objects

  • Upload
    zukun

  • View
    214

  • Download
    0

Embed Size (px)

Citation preview

Page 1: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 1/104

Lecture 6Scenes and objects

6.870 Object Recognition and Scene Understandinghttp://people.csail.mit.edu/torralba/courses/6.870/6.870.recognition.htm

Page 2: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 2/104

Class business

Next Wednesday«

Page 3: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 3/104

Week 2: Objects without scenes

Week 5: Scenes without objects

Week 6: Scenes and objects

Page 4: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 4/104

Why is detection hard?

x

y

1,000,000 images/dayPlus, we want to do this for ~ 1000 objects

10,000 patches/object/image

time

Page 5: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 5/104

Standard approach to scene analysis

1) Object representation based on intrinsic features:

 Local 

 featuresno car 

Classifier 

 p( car | VL )

2) Detection strategy:

Sky

Mountain

Buildings

cars

3) The scene representation

Page 6: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 6/104

Is local information enough?

Page 7: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 7/104

With hundreds of categories

roadtablechair 

keyboardtablecar 

road

If we have 1000 categories (detectors), and each detector produces 1 fa every 10

images, we will have 100 false alarms per image« pretty much garbage«

Page 8: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 8/104

Is local information even enough?

Page 9: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 9/104

Is local information even enough?

Distance

Information

Local featuresContextual features

Page 10: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 10/104

We know there is a keyboard present in this scene even if we cannot see it clearly.

We know there is no keyboard present in this scene

« even if there is one indeed.

The system does not care about the

scene, but we do«

Page 11: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 11/104

The multiple personalities of a blob

Page 12: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 12/104

The multiple personalities of a blob

Page 13: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 13/104

Page 14: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 14/104

Page 15: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 15/104

Page 16: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 16/104

Look-Alikes by Joan Steiner 

Page 17: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 17/104

Look-Alikes by Joan Steiner 

Page 18: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 18/104

Look-Alikes by Joan Steiner 

Page 19: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 19/104

Why is context important?

Changes the interpretation of an object (or its function)

Context defines what an unexpected event is

Page 20: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 20/104

The influence of an object extends beyond its physical boundaries

Page 21: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 21/104

The context challenge

How far can you go withoutusing an object detector?

Page 22: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 22/104

21

What are the hidden objects?

Page 23: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 23/104

What are the hidden objects?

Chance ~ 1/30000

Page 24: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 24/104

Page 25: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 25/104

The importance of context

Cognitive psychology ± Palmer 1975

 ± Biederman 1981

 ± «

Computer vision ± Noton and Stark (1971)

 ± Hanson and Riseman (1978)

 ± Barrow & Tenenbaum (1978) ± Ohta, kanade, Skai (1978)

 ± Haralick (1983)

 ± Strat and Fischler (1991)

 ± Bobick and Pinhanez (1995)

 ± Campbell et al (1997)

Page 26: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 26/104

Biederman 1972

Arrow appeared before or after picture.

Selected object from 4 pictures.

Page 27: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 27/104

Page 28: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 28/104

Page 29: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 29/104

Biederman 1972

Better accuracy with normal scene and

with pre-cue.

C

oherence of surroundings affected objectperception.

But, jumbled pictures had unnatural edge

artifacts.

Page 30: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 30/104

Palmer 1975

Scene preceded object to identify.

Better identification when preceded by a

semantically consistent scene.

Objects seen for 20, 40, 60 or 120 ms.

Page 31: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 31/104

Palmer 

Scenes shown ahead of time for 2 s.

More accurate recognition of consistent

objects than inconsistent objects. Similar looking objects were misnamed,

showing a bias effect.

Page 32: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 32/104

Loftus & Mackworth

Inconsistent objects

fixated earlier and

longer.

Suggested additional

processing of objects

out of context. Similar results found

by Friedman (1979).

Page 33: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 33/104

De Graef et al. 1990

Prior results due to memory task?

Measured eye movements during non-

object search task.

Page 34: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 34/104

Page 35: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 35/104

Object Detection

Biederman et al. 1982, relational violations

Page 36: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 36/104

Page 37: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 37/104

Biederman 1982

Pictures shown for 150

ms.

Objects in appropriate

context were detected

more accurately than

objects in aninappropriate context.

Scene consistency

affects object detection.

Page 38: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 38/104

Objects and Scenes

Biederman¶s violations (1981):

Page 39: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 39/104

Support

[Golconde Rene Magritte]

Page 40: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 40/104

Interposition

[Blank Check Rene Magritte]

Page 41: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 41/104

Page 42: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 42/104

Position, Probability

[P ersonal Values Rene Magritte]

Page 43: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 43/104

Object Consistencies

Biederman et al (1982), DeGraef(1990).

Page 44: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 44/104

Object Consistencies

Examples of inconsistencies

Biederman et al (1982), DeGraef(1990).

Page 45: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 45/104

Contextual cueing

Chun & Jiang, 1998

Page 46: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 46/104

Object priming

Torralba, Sinha, Oliva, VSS 2001

Increasing contextual information

Page 47: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 47/104

Object priming

Torralba, Sinha, Oliva, VSS 2001

bj i i

Page 48: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 48/104

?Car, pedestrian, mailbox, «

Object priming

 p(object | scene)

Torralba, Sinha, Oliva, VSS 2001

Obj i i

Page 49: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 49/104

Object priming

Torralba, Sinha, Oliva, VSS 2001

Page 50: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 50/104

Ex amples of consistent scenes (a), inconsistent scenes (b), and isolated objects and 

backgrounds (c); from Davenport & P otter, 2004

Page 51: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 51/104

But do we really need context?

Page 52: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 52/104

Hollingworth & Henderson

Concerns with object detection studies

 ± Object label could bias results.

 ± Location cue selectively helpful for consistent

objects.

Controlled for false alarm biases with post-

cue and 2AFC.

Failed to find consistency effects.

Page 53: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 53/104

Hollingworth & Henderson

Post-cue

2AFC with object

labels

 ± Both consistent or 

inconsistent.

2AFC with tokendiscrimination.

 ± E.g. sports car or 

sedan.

 

Page 54: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 54/104

Page 55: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 55/104

Page 56: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 56/104

Page 57: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 57/104

Page 58: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 58/104

CONDOR systemStrat and Fischler (1991)

Guzman (S EE ), 1968

Noton and Stark 1971

Hansen & Riseman (VI S ION S ), 1978

Barrow & Tenenbaum 1978

Brooks ( ACRONYM ), 1979

Marr, 1982

Ohta & Kanade, 1978

Yakimovsky & Feldman, 1973

A A f S U d di

Page 59: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 59/104

 An Age of Scene Understanding

Guzman (S EE ), 1968 Noton and Stark 1971

Hansen & Riseman(VI S ION S ), 1978

Barrow & Tenenbaum 1978

Brooks ( ACRONYM ), 1979

Marr, 1982

Ohta & Kanade, 1978

Yakimovsky & Feldman, 1973

[Ohta & Kanade 1978]

C

Page 60: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 60/104

Current approaches

1) Scene to object dependencies

2) Object to object dependencies

L l f

Page 61: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 61/104

Levels of context

Context in low-level vision

Part-based models

Objects relations Long-range connections

Weak constraints

Multimodal

Fix graph structurescan be useful

approximations

C t h

Page 62: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 62/104

Current approaches

1) Scene to object dependencies

2) Object to object dependencies

Page 63: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 63/104

b t this co occ rrence has a hidden

Page 64: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 64/104

« but this co-occurrence has a hidden

common ³cause´: the scene

streetsoffices

It is easier to first recognize the scene, then predict object presence, than

running local object classifiers

Th l d t t f

Page 65: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 65/104

The layered structure of scenes

In a display with multiple targets present, the location of one target constraints the µy¶

coordinate of the remaining targets, but not the µx¶ coordinate.

 Assuming a human observer standing on the ground

Page 66: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 66/104

D t ti f ith t f d t t

Page 67: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 67/104

Detecting faces without a face detector 

Torralba & Sinha, 01; Torralba, 03

Context based vision system for place

Page 68: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 68/104

Context-based vision system for place

and object recognition

Hidden states = location (63 values) Observations = vG

t (80 dimensions)

Transition matrix encodes topology of 

environment Observation model is a mixture of 

Gaussians centered on prototypes (100

views per place)

Office 610 Corridor 6b Corridor 6c Office 617

We use 17 annotated sequences for training

Torralba, Murphy, Freeman and Rubin. ICCV 2003

O bil i

Page 69: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 69/104

Our mobile rig

Torralba, Murphy, Freeman, Rubin. 2003

Page 70: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 70/104

Page 71: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 71/104

Page 72: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 72/104

Page 73: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 73/104

An integrated model of Scenes

Page 74: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 74/104

 An integrated model of Scenes,

Objects, and Parts

Ncar 

S

g

Scene

Scene

gist

features

0

0

1

1

5

5

N

P(Ncar | S = street)

P(Ncar | S = park)

0 5 10 1 50

0 .0 5

0. 1

0 .1 5

0. 2

0 5 1 0 1 50

0. 2

0. 4

0. 6

0. 8

N

Page 75: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 75/104

Global to local

Page 76: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 76/104

Global to local

Use global context to predict objects but there is nomodeling of spatial relationships between objects.

Op1,c1

vp1,c1

OpN,c1

vpN,c1. . .

Op1,c2

vp1,c2

OpN,c2

vpN,c2. . .

Class 1 Class 2

E1 E2

S

c2maxVc1

maxV

X1X2

vg

Keyboards

Murphy, Torralba & Freeman (03)

3d Scene Context

Page 77: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 77/104

3d Scene Context

Image World

Hoiem, Efros, Hebert ICCV 2005

3d Scene Context

Page 78: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 78/104

3d Scene Context

meters

    m    e     t    e

    r    s

Ped

Ped

Car 

Hoiem, Efros, Hebert ICCV 2005

3D City Modeling using Cognitive Loops

Page 79: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 79/104

3D City Modeling using Cognitive Loops

N.C

ornelis, B. Leibe, K.C

ornelis, L.V

an Gool.CV

PR'06

Current approaches

Page 80: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 80/104

Current approaches

1) Scene to object dependencies

2)O

bject to object dependencies

Where should I put the silverware?

Page 81: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 81/104

Where should I put the silverware?

Sampling from the labels

Page 82: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 82/104

Sampling from the labels

Sampling from the labels

Page 83: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 83/104

Sampling from the labels

Cf. Hoiem et al; Hays, Efros. Siggraph 2007

Contextual object relationships

Page 84: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 84/104

Contextual object relationshipsCarbonetto, de Freitas & Barnard (2004) Kumar, Hebert (2005)

Torralba Murphy Freeman (2004)

Fink & Perona (2003)E. Sudderth et al (2005)

Object-Object Relationships

Page 85: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 85/104

Fink & Perona (NIPS 03)Use output of boosting from other objects at previous

iterations as input into boosting for this iteration

Object-Object Relationships

Pixel labeling using MRFs

Page 86: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 86/104

Pixel labeling using MRFs

Enforce consistency between neighboring

labels, and between labels and pixels

Carbonetto, de Freitas & Barnard, ECCV¶04

Beyond nearest-neighbor grids

Page 87: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 87/104

Beyond nearest-neighbor grids

Most MRF/C

RF models assume nearest-neighbor graph topology

This cannot capture long-distance

correlations

Dynamically structured trees

Page 88: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 88/104

Dynamically structured trees

Each node pick its parents(Storkey& Williams, PAMI¶03)

2D SCFGs(Pollak, Siskind, Harper & Bouman IC ASSP¶03)

Object-Object Relationships

Page 89: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 89/104

Object Object Relationships

Use latent variables to induce long distance correlations

between labels in a Conditional Random Field (CRF)

He, Zemel & Carreira-Perpinan (04)

Object-Object Relationships

Page 90: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 90/104

Object-Object Relationships

[Kumar Hebert 2005]

Hierarchical Sharing and Context

Page 91: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 91/104

Hierarchical Sharing and Context

Scenes share objects

Objects share parts

Parts share features

E. Sudderth, A. Torralba, W. T. Freeman, and A. Wilsky.

3d Scene Context

Page 92: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 92/104

3d Scene Context

Image Support Vertical Sky

V-Left V-Center  V-Right V-Porous V-Solid

[Hoiem, Efros, Hebert ICCV 2005]

Object

Surface?

Support?

Detecting difficult objects

Page 93: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 93/104

Detecting difficult objects

Office Maybethere is

a mouse

Start recognizing the scene

Torralba, Murphy, Freeman. NIPS 2004.

Page 94: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 94/104

Detecting difficult objects

Page 95: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 95/104

Detecting difficult objects

Detect first simple objects (reliable detectors) that provide strong

contextual constraints to the target (screen -> keyboard -> mouse)

Torralba, Murphy, Freeman. NIPS 2004.

BRF for screen/keyboard/mouse

Page 96: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 96/104

BRF for screen/keyboard/mouse

Iteration

BRF for screen/keyboard/mouse

Page 97: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 97/104

BRF for screen/keyboard/mouse

Iteration

BRF for screen/keyboard/mouse

Page 98: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 98/104

BRF for screen/keyboard/mouse

Iteration

BRF for screen/keyboard/mouse

Page 99: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 99/104

BRF for screen/keyboard/mouse

Iteration

BRF for screen/keyboard/mouse

Page 100: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 100/104

BRF for screen/keyboard/mouse

Iteration

BRF for car detection: topology

Page 101: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 101/104

BRF for car detection: topology

BRF for car detection: results

Page 102: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 102/104

BRF for car detection: results

A ³car´ out of context is less of a car

Page 103: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 103/104

 A car out of context is less of a car 

Car  Building Road

 b F G  b F G  b F G

From image

From detectors

Thresholded beliefs

Page 104: MIT6870_ORSU_lecture6: Scenes and objects

8/3/2019 MIT6870_ORSU_lecture6: Scenes and objects

http://slidepdf.com/reader/full/mit6870orsulecture6-scenes-and-objects 104/104

Contextor no context