computer example of schiele and pentland’s probabilistic object recognition and localization...

15
Computer Example of Computer Example of Schiele and Schiele and Pentland’s Pentland’s Probabilistic Probabilistic Object Recognition and Object Recognition and Localization Localization Marshall Tappen Marshall Tappen 6.899 6.899

Upload: gervais-clark

Post on 02-Jan-2016

219 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

Computer Example of Computer Example of Schiele and Pentland’s Schiele and Pentland’s

Probabilistic Object Probabilistic Object Recognition and Recognition and

LocalizationLocalizationMarshall TappenMarshall Tappen

6.8996.899

Page 2: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

Implementation Details You Implementation Details You Need to Think AboutNeed to Think About

► How many cells per dimension of the How many cells per dimension of the histogram?histogram? It appears that the authors use around 4 or 5 cells It appears that the authors use around 4 or 5 cells

per histogramper histogram► I used 6I used 6

► Do all dimensions have the same number of Do all dimensions have the same number of cells? (Yes)cells? (Yes)

► How do you allocate the cells?How do you allocate the cells? I set the range encompassed by the histogram to I set the range encompassed by the histogram to

cover the full range of possible valuescover the full range of possible values►You could set it to encompass 97% of the valuesYou could set it to encompass 97% of the values

Page 3: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

CommentsComments

►Histograms are big!Histograms are big! 720 images (20 objects, 36 views) takes 720 images (20 objects, 36 views) takes

268 MB of space268 MB of space►Smaller histograms decrease space needed Smaller histograms decrease space needed

dramaticallydramatically

►Executing convolutions to take Executing convolutions to take measurements is by far the slowest partmeasurements is by far the slowest part

►Algorithm is very simple to implementAlgorithm is very simple to implement Getting a good training set is the hard partGetting a good training set is the hard part

Page 4: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

Testing ProcedureTesting Procedure► I used the Columbia Object Image I used the Columbia Object Image

Library (COIL-20) DatabaseLibrary (COIL-20) Database►Grayscale Images of 20 ObjectsGrayscale Images of 20 Objects

72 Images of each object, taken at 72 Images of each object, taken at different poses different poses

Page 5: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

Testing ProcedureTesting Procedure►For each image, I used one half of For each image, I used one half of

the poses as the training setthe poses as the training set►The other half of the poses were the The other half of the poses were the

test settest set

Page 6: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

How Many Bins in Each How Many Bins in Each Histogram?Histogram?

►My first question was how many bins My first question was how many bins should be in each dimension of the should be in each dimension of the histogram.histogram.

►My second question was how many My second question was how many measurements I need to identify an measurements I need to identify an object.object.

Page 7: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

How Many Bins in Each How Many Bins in Each Histogram?Histogram?

Page 8: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

How Many Bins in Each How Many Bins in Each Histogram?Histogram?

Page 9: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

How Many Views Do We How Many Views Do We Need?Need?

►My next question was how many views My next question was how many views are needed.are needed.

►For this test, I chose a portion of the For this test, I chose a portion of the test set and tested it while changing test set and tested it while changing the number of poses in the training setthe number of poses in the training set (I was using 75 measurements and 6 bins (I was using 75 measurements and 6 bins

per dimension)per dimension)

Page 10: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

How Many Views Do We How Many Views Do We Need?Need?

Page 11: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

How well does it handle varying How well does it handle varying object size?object size?

► I wanted to know how robust the I wanted to know how robust the algorithm is to changing object sizesalgorithm is to changing object sizes

► I tested this by using rescaled versions I tested this by using rescaled versions of the test set as input imagesof the test set as input images

Page 12: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

How well does it handle How well does it handle varying object size?varying object size?

Page 13: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

What about new lighting?What about new lighting?

► I want to see how robust the algorithm is to I want to see how robust the algorithm is to changing illumination.changing illumination.

► I faked a new illumination by masking the input I faked a new illumination by masking the input imagesimages

► I masked the test imagesI masked the test images

with a horizontal gradientwith a horizontal gradient► I varied the slope of I varied the slope of

the gradient the gradient

(i.e the beginning value (i.e the beginning value

of the gradient)of the gradient)

Page 14: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

What about new lighting?What about new lighting?

Page 15: Computer Example of Schiele and Pentland’s Probabilistic Object Recognition and Localization Marshall Tappen 6.899

CommentsComments

►The algorithm is surprisingly robustThe algorithm is surprisingly robust► It could really benefit from doing the It could really benefit from doing the

convolutions in hardwareconvolutions in hardware► If you want to use it, ask your advisor If you want to use it, ask your advisor

to buy more RAMto buy more RAM