Experiments 1 and 2
Supervised learning and detection
In experiment 1, we learn the template from images with given bounding boxes.
In experiment 2, we use the learned template to detect and sketch the object.
(0) Prototype code
The above code is simple and convenient for development purpose. But it is not
optimized for speed.
(0.1) matching pursuit
Earlier code for matching pursuit. Linear additive representation and matching
pursuit are the prototype machinery for the present work.
(1) Data and code
for Experiments 1.1 and 2.1
(1.1) code
with whitening
(1.2) code
with sigmoid transformation
Experiment 1.1. The 37 training images are 82*164. The first block displays the
learned active basis consisting of 60 elements. Each element is symbolized by a bar.
The rest of the blocks display the observed images and the corresponding deformed
active bases. The images are displayed in the descending order of the log-likelihood
ratio, which scores the template matching.
eps
Experiment 2.1. Left: Testing image. The recognition algorithm is run on 10 resolutions,
from 270*360 to 432*536.
eps
Right: Superposed with sketch of the 60 elements of the deformed active basis at the
optimal resolution 378*504. The bounding box of the template is 82*164.
eps
Experiment 2.1. Left: MAX2 scores at resolutions 1 to 10.
eps
Right: SUM2 map at the optimal resolution 7.
eps
Experiment 2.1. Negative square root of SUM1 maps at the
optimal resolution. There are 15 orientations.
eps
Experiment 2.1. Negative square root of MAX1 maps at the
optimal resolution. There are 15 orientations.
eps
(1.3)
code with plotting sum-max maps
(2)
Data and code for Experiments 1.2 and 2.2
Experiment 1.2. The 27 images are 180*180. Number of elements is 60.
eps
Experiment 2.2. Left: Testing image. The algorithm is run on 10 resolutions, from
320*240 to 500*375.
eps
Right: Superposed with sketch of 60 elements of the deformed active basis at the
optimal resolution 400*300. The bounding box of the template is 180*180.
eps
Experiment 2.2. Left: MAX2 scores at resolutions 1 to 10.
eps
Right: SUM2 map at the optimal resolution 5.
eps
(3)
Data and code for Experiments 1.3 and 2.3
Experiment 1.3. The 15 images are 179*112. Number of elements is 50.
eps
Experiment 2.3. Testing image.
eps
Experiment 2.3. Superposed sketch of 50 elements of the deformed active basis
at each of the 10 resolutions, from 192
*220 to 322*368. The bounding box is 179*112.
eps1
2
3
4
5
6
7
8
9
10
(3.1)
another example
eps
(4)
Data and code for Experiments 1.4
Experiment 1.4. The 12 images are 120*167. Number of elements is 50.
eps
(5)
Data and code for Experiments 1.5
Experiment 1.5. The 9 images are 122*120. Number of elements is 50.
eps
(6)
Data and code for Experiments 1.6
Experiment 1.6. The 11 images are 133*140. Number of elements is 50.
eps
(7)
Data and code for Experiments 1.7
Experiment 1.7. The 11 images are 105*84. Number of elements is 30.
eps
Back to active basis homepage