The following are the results of Divya's classification algorithm (I pulled these images from her webpage) and my classification algorithm on five different test images.

In these pictures red denotes "Not Interesting" and blue and green "Somewhat Interesting" or "Interesting".

I ran 10 experiments with the probabilistic majority voting algorithm. Each of those are described in the following table. The word "balanced" is important only for pictures 1, 2, and 5, where the there is a small number (1-2) pixels of the minority class in a partition with the rest of the majority class. Though the minority class is technically present in that partition, there should probably be some kind of threshold. Those experiments which were "balanced" do not count 1-2 class c pixels in a parition to be capable of voting on class c.

(1)	Random Subspaces, 1000 Trees
(2)	Random Forests, 1000 Trees
(3)	Majority Bagging, Random Subspaces, 1000 Trees
(4)	Majority Bagging, Random Forests, 1000 Trees
(5)	Neural Network, 10 hidden, 500 epochs
(6)	(Balanced) Random Subspaces, 1000 Trees
(7)	(Balanced) Random Forests, 1000 Trees
(8)	(Balanced) Majority Bagging, Random Subspaces, 1000 Trees
(9)	(Balanced) Majority Bagging, Random Forests, 1000 Trees
(10)	(Balanced) Neural Network, 10 hidden, 500 epochs
(11)	(Balanced) Combined Regions, Random Subspaces, 1000 Trees
(12)	(Balanced) Combined Regions, Random Forests, 1000 Trees
(13)	(Balanced) Combined Regions, Neural Network, 10 hidden, 500 epochs
(14)	(Balanced) Combined Regions, Neural Network, 10 hidden, 10 epochs
(15)	(Balanced) All Regions Voting, Random Forests, 1000 Trees
(16)	(Balanced) All Regions Voting, Neural Network, 10 hidden, 500 epochs
(17)	Baseline C4.5 unpruned trained on all of the data on all of the pictures
(18)	Baseline C4.5 pruned trained on all of the data on all of the pictures
(19)	Baseline RF trained on all of the data on all of the pictures, 1000 trees
(20)	Baseline NN trained on all of the data on all of the pictures, 10/500

Ground Truth

Divya's Algorithm

Prob. Maj. (1)

Prob. Maj. (2)

Prob. Maj. (3)

Prob. Maj. (4)

Prob. Maj. (5)

Prob. Maj. (6)

Prob. Maj. (7)

Prob. Maj. (8)

Prob. Maj. (9)

Prob. Maj. (10)

Prob. Maj. (11)

Prob. Maj. (12)

Prob. Maj. (13)

Prob. Maj. (14)

Prob. Maj. (15)

Prob. Maj. (16)

Baseline C45-un (17)

Baseline C45-pr (18)

Baseline RF (19)

Baseline NN (20)

Here is an analysis of how the probabilistic majority vote algorithm is doing for each of the images. This ROC looks at NotInteresting vs. Interesting+SomewhatInteresting as a function of how many of the 40 classifiers trained on the 40 partitions vote for the minority class. The classifier used was a neural network, as in (16) above.