30
Multisensory Perception Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications Moving Forward Multisensory Perception Alex Teichman August 26, 2009

Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

MultisensoryPerception

Alex Teichman

August 26, 2009

Page 2: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

The Object Recognition Problem

How much does this look like a hand?

Page 3: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

The Object Recognition Problem

Where is the coke can?

Page 4: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

The Object Recognition Problem

What is the object in the box?

Page 5: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Some Difficulties Among Many

Labels are painful to acquire.

Generalization to data not in your training set can bearbitrarily hard.

Descriptors are generally suited to a particular task.

Page 6: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Some Difficulties Among Many

Labels are painful to acquire.

Generalization to data not in your training set can bearbitrarily hard.

Descriptors are generally suited to a particular task.

Page 7: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Some Difficulties Among Many

Labels are painful to acquire.

Generalization to data not in your training set can bearbitrarily hard.

Descriptors are generally suited to a particular task.

Page 8: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Boosting - Intuition

Page 9: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Boosting - Intuition

Page 10: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Boosting - Math

Training examples (ym,~xm) for all m = 1..M, ym ∈ {−1,+1}.

minH

∑m

exp(−ymH(~xm)) (1)

H(~x) =∑

t

ht(~x) (2)

ht(~x) =

{at if dist(~x , ~xt) < θt

0 otherwise(3)

Stagewise optimization:minht+1

∑m exp (−ym (H(~xm) + ht+1(~xm)))

Page 11: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Boosting - Math

Training examples (ym,~xm) for all m = 1..M, ym ∈ {−1,+1}.

minH

∑m

exp(−ymH(~xm)) (1)

H(~x) =∑

t

ht(~x) (2)

ht(~x) =

{at if dist(~x , ~xt) < θt

0 otherwise(3)

Stagewise optimization:minht+1

∑m exp (−ym (H(~xm) + ht+1(~xm)))

Page 12: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Boosting - Math

Training examples (ym,~xm) for all m = 1..M, ym ∈ {−1,+1}.

minH

∑m

exp(−ymH(~xm)) (1)

H(~x) =∑

t

ht(~x) (2)

ht(~x) =

{at if dist(~x , ~xt) < θt

0 otherwise(3)

Stagewise optimization:minht+1

∑m exp (−ym (H(~xm) + ht+1(~xm)))

Page 13: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Boosting - Math

Training examples (ym,~xm) for all m = 1..M, ym ∈ {−1,+1}.

minH

∑m

exp(−ymH(~xm)) (1)

H(~x) =∑

t

ht(~x) (2)

ht(~x) =

{at if dist(~x , ~xt) < θt

0 otherwise(3)

Stagewise optimization:minht+1

∑m exp (−ym (H(~xm) + ht+1(~xm)))

Page 14: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Image Descriptors Library

Goal: Plug-and-play image descriptors.

Page 15: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Image Descriptors Library

See descriptors 2d package

Page 16: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Hand Detection

Histogram of Oriented Gradient descriptors (3 differentparameter settings)

Superpixel Color Histogram descriptors (4 differentsuperpixel sizes)

1 hour to train.

Collaboration with Romain and Alex S.

Page 17: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Hand Detection

Histogram of Oriented Gradient descriptors (3 differentparameter settings)

Superpixel Color Histogram descriptors (4 differentsuperpixel sizes)

1 hour to train.

Collaboration with Romain and Alex S.

Page 18: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Hand Detection

Histogram of Oriented Gradient descriptors (3 differentparameter settings)

Superpixel Color Histogram descriptors (4 differentsuperpixel sizes)

1 hour to train.

Collaboration with Romain and Alex S.

Page 19: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Hand Detection

Histogram of Oriented Gradient descriptors (3 differentparameter settings)

Superpixel Color Histogram descriptors (4 differentsuperpixel sizes)

1 hour to train.

Collaboration with Romain and Alex S.

Page 20: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications
Page 21: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications
Page 22: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Video Label Propagation

Label a video given just a few frame labels.

Page 23: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Mechanical Turk Feedback Loop

Implementing the feedback loop of the hands training methodologyon mechanical turk with a bottles dataset.

Page 24: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Online Learning

mina

∑m

exp(−ymH(~xm))

Can solve this for all weak classifier responsessimultaneously.Constant space for any number of training examples.

Maintain the gradient and Hessian as a stream of labels comesinto the classifier:

g =1

M

∑m

−ymIm

H =1

M

∑m

ImITm

Page 25: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Moving Forward

Labels are painful to acquire.

Put classifier predictions on Mechanical Turk, getcorrections, feed back into boosting.

Generalization to data not in your training set can bearbitrarily hard.

Put classifier predictions on Mechanical Turk, getcorrections, feed back into boosting.

Descriptors are generally suited to a particular task.

Use a large library of descriptors and allow boosting tochoose which are most useful through the optimization.

Page 26: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Moving Forward

Labels are painful to acquire.

Put classifier predictions on Mechanical Turk, getcorrections, feed back into boosting.

Generalization to data not in your training set can bearbitrarily hard.

Put classifier predictions on Mechanical Turk, getcorrections, feed back into boosting.

Descriptors are generally suited to a particular task.

Use a large library of descriptors and allow boosting tochoose which are most useful through the optimization.

Page 27: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Moving Forward

Labels are painful to acquire.

Put classifier predictions on Mechanical Turk, getcorrections, feed back into boosting.

Generalization to data not in your training set can bearbitrarily hard.

Put classifier predictions on Mechanical Turk, getcorrections, feed back into boosting.

Descriptors are generally suited to a particular task.

Use a large library of descriptors and allow boosting tochoose which are most useful through the optimization.

Page 28: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Moving Forward

Labels are painful to acquire.

Put classifier predictions on Mechanical Turk, getcorrections, feed back into boosting.

Generalization to data not in your training set can bearbitrarily hard.

Put classifier predictions on Mechanical Turk, getcorrections, feed back into boosting.

Descriptors are generally suited to a particular task.

Use a large library of descriptors and allow boosting tochoose which are most useful through the optimization.

Page 29: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Moving Forward

Labels are painful to acquire.

Put classifier predictions on Mechanical Turk, getcorrections, feed back into boosting.

Generalization to data not in your training set can bearbitrarily hard.

Put classifier predictions on Mechanical Turk, getcorrections, feed back into boosting.

Descriptors are generally suited to a particular task.

Use a large library of descriptors and allow boosting tochoose which are most useful through the optimization.

Page 30: Multisensory Perception - Robot Operating Systemros.org/presentations/2009-08_Alex-Teichman_2d...Alex Teichman The Object Recognition Problem Boosting Image Descriptors Library Applications

MultisensoryPerception

AlexTeichman

The ObjectRecognitionProblem

Boosting

ImageDescriptorsLibrary

Applications

MovingForward

Thanks!