Experiment 5

Learning from non-aligned images

In experiment 5, we learn a template from training images that are not aligned.

Experiment 5.a

In experiment 5.a, we assume that the bounding box of the object in the first image is given. This requirement is eliminated in Experiment 5.c.

(1) data, codes, and readme for learning from non-aligned images

Experiment 5.a.1. The bounding box of the first image is given. The size of the bounding box is 136*140. The number of elements in the active basis is 60. eps 1 2 3 4 5

(2) another example

Experiment 5.a.2. The bounding box is 116*116. Number of elements is 50. eps 1 2 3 4 5

(3) another example

Experiment 5.a.3. The bounding box is 115*161. Number of elements is 50. eps 1 2 3 4 5

(4) another example

Experiment 5.a.4. The bounding box is 94*99. Number of elements is 30. eps 1 2 3 4 5

(5) another example

Experiment 5.a.5. The bounding box is 94*138. Number of elements is 60. eps 1 2 3 4 5

(6) another example

Experiment 5.a.6. The bounding box is 134*148. Number of elements is 60. eps 1 2 3 4 5

Experiment 5.b

In Experiment 5.b, we learn a template from the first image (on the left) with no activity and without any given bounding box. Then we restore the activity and deform the learned template to sketch the second image (on the right). We scan the template over 7 resolutions of the second image, from .7 to 1.3 times the size of the original image. This aligns the first image to the second image. In this experiment, we use active correlation for learning and recognition.

(1) data, codes, and readme for image alignment

Experiment 5.a.1. The number of elements is 50. eps1 eps2

Experiment 5.a.2. The number of elements is 50. eps1 eps2

(2) another example

Experiment 5.a.3. The number of elements is 50. eps1 eps2

Experiment 5.a.4. The number of elements is 100. eps1 eps2

(3) another example

Experiment 5.a.5. The number of elements is 80. eps1 eps2

Experiment 5.c

In experiment 5.c, we do not assume that the bounding box of the object in the first image is given. We simply start from the template learned from the whole image of the first example.

data, codes, and readme for learning from non-aligned images

Experiment 5.c.1. The first image is the starting template. The second image is the learned template. The number of elements in the active basis is 60. eps0 eps 1 2 3 4 5

Experiment 5.c.2. The first image is the starting template. The second image is the learned template. The number of elements in the active basis is 60. eps0 eps 1 2 3 4 5

Experiment 5.c.3. The first image is the starting template. The second image is the learned template. The number of elements in the active basis is 30. The allowed activity in location is 3. The sub-sampling rate is 1. eps0 eps 1 2 3 4 5

Experiment 5.c.4. The first image is the starting template. The second image is the learned template. The number of elements in the active basis is 50. eps0 eps 1 2 3 4 5

Experiment 5.c.5. The first image is the starting template. The second image is the learned template. The number of elements in the active basis is 50. eps0 eps 1 2 3 4 5

Experiment 5.c.6. The first image is the starting template. The second image is the learned template. The numberof elements in the active basis is 60. eps0 eps 1 2 3 4 5

another example

Experiment 5.c.7. The first image is the starting template. The second image is the learned template. The numberof elements in the active basis is 60. eps0 eps 1 2 3 4 5 4 5

another example

Experiment 5.c.8. The first plot is the starting template, learned from the first image with no activity, and with no given bounding box. The remaining plots display the learned templates in the first 10 iterations. The images are rescaled to .5 times the original sizes. The number of elements in the active basis is 60. eps0 1 2 3 4 5 6 7 8 9 10

Experiment 5.c.8. We use the first 20 images of the Weizmann horse images. The above plots display the superposed deformed templates in the last iteration. The Weizmann horse images have big deformations. So it is more reasonable to learn part templates first and then compose them into a recursive active basis. We do not claim that a single layer active basis can model articulate objects. eps1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

another example