Experiment 3
Supervised learning and classification
In experiment 3, we learn the template from positive training images. Then we use the learned template
to score the positive and negative testing images, and obtain the ROC curve. In both the training and
testing images, the bounding boxes are given.
(1) data, codes, and readme for classification comparison
Experiment 3.1. Results obtained by fitting the active
basis model with sigmoid transformation.
The saturation level is 6. The 43 images are 127*85. The number of basis elements
40.
eps
Experiment 3.1. The solid curve is ROC of active basis
model with sigmoid transformation.
The dashed curve is ROC of the
adaboost method where the weak classifiers are based on
thresholding
the filter responses in SUM1 maps.
eps
Experiment 3.1. Left: The template learned by the active
basis model.
Middle: Learned by the adaboost method.
Right: Learned from
tiny training images.
eps1
2
3
ROC report: threshold = .941, thresholdCORR = .971, whiten = .965, whitenCORR = .974,
sigmoid6 = .977, bi-basis = .982, adaptive = .983, tiny = .957, single = .926; adaboost on SUM1 = .936,
adaboost on MAX1 = .943.
(2) another example
Experiment 3.2. The 31 images are 150*120. The
number of basis elements is 40.
eps
Experiment 3.2. Comparison between sigmoid and adaboost.
eps
ROC report: threshold = .981, thresholdCORR = .985, whiten = .982, whitenCORR = .984,
sigmoid6 = .987, bi-basis = .988, adaptive = .989, tiny = .956, single = .980; adaboost on SUM1 = .937,
adaboost on MAX1 = .957.
(3) a third example
Experiment 3.3. The 33 images are 150*100. The
number of basis elements is 40.
eps
sigmoid = .999; adaboost on MAX1 = .994.
Back to active basis homepage