Experiment 8
Geometric transformation of template
In experiment 8, we transform a template by dilation, rotation, and changing
the aspect ratio.
(0) Code with rotation and scaling in detection
==> Number of training images = 12; Number of elements = 50; Length of Gabor = 17 pixels; Local normalization or not = 1; Range of displacement = 4 pixels; Subsample rate = 1 pixel; Image height and width = 100 by 100 pixels
==> Number of testing images = 129; Range of rotation = 4 times pi/15;
For each testing image, the search is over a range of 10
resolutions, from 10% to 100% of the original image.
(0) Code with change of aspect ratio and scale of template in detection
==> Number of training images = 9; Number of elements = 40; Length of Gabor = 17 pixels; Range of displacement = 4 pixels; Subsample rate = 1 pixel; Image height and width = 82 by 80 pixels
==> Number of testing images = 7
Negative experience in Experiment 8. We encountered some difficulty with the bicycle
template. When the viewing distance is close, the size of one wheel can be larger
than the size of the other wheel, so a single scale factor does not give very good
fit. Also, the frontal wheel may turn to a different direction than the back wheel.
The above difficulty suggests that we should better split the bicycle template into
two part-templates, and each part-template has its own geometric transformation. We
study the composition of multiple part-templates in Experiment 7.
(1)
data, codes, and readme for geometric transformation (March 2009)
The codes in this experiment are not optimized for speed.
(1.1)
Code that pools q() from 2 large natural images (May 2009)
Experiment 8.1. The 27 images are 180*180. Number of elements is 60.
eps
Experiment 8.1. The number of elements is 60. The image size is 252 $\times$ 320.
The scale factor is 1.4. The rotation is 1 $\times$ pi/15. The aspect factor is 0.9
eps
Experiment 8.1. The image size is 200 $\times$ 250. The scale factor is 1.4.
The rotation is 1 $\times$ pi/15. The aspect factor is 1.
eps
Experiment 8.1. The image size is 248 $\times$ 232. The scale factor is 1.2.
The rotation is -1 $\times$ pi/15. The aspect factor is 0.6.
eps
(1.2)
Code with local normalization of filter responses (September 2009)
Experiment 8.1. Location normalization with a window whose half size is 20x20.
The allowed shift in location is up to 3 pixels.
eps
Experiment 8.1. Location normalization with a window whose half size is 20x20.
The allowed shift in location is up to 3 pixels.
eps
(2)
another example (March 2009)
(2.1)
Code that pools q() from 2 large natural images (May 2009)
Experiment 8.1. The 30 images are 150*120. Number of elements is 40.
eps
Experiment 8.2. The number of elements is 40. The image size is 166 $\times$ 202.
The scale factor is 1.2 The rotation is 0. The aspect factor is 0.8
eps
Experiment 8.2.
The image size is 192 $\times$ 144. The scale factor is 1. The rotation is 4 $\times$ $pi/15$.
The aspect factor is 1.4.
eps
(2.2)
Code with local normalization of filter responses (September 2009)
Experiment 8.1. Location normalization with a window whose half size is 20x20.
The allowed shift in location is up to 3 pixels.
eps
Experiment 8.1. Location normalization with a window whose half size is 20x20.
The allowed shift in location is up to 3 pixels.
eps
Experiment 8.1. Location normalization with a window whose half size is 20x20.
The allowed shift in location is up to 3 pixels.
eps
Back to active basis homepage