Experiment 1 Results

We completed our initial experiment to segment an image with a modified version of k-means by using the histograms for each label from the manual segmentation.

 Here are the results for one run, we have colored the segments according to the highest probable label.

The following table shows the final probabilities assigned to each cluster. 

					Label
Cluster	Pixel Count	1	2	3	4	5	6	7
A	4359	0.9854*	0.0023	0.0003	0.0063	0.0034	0.0002	0.002
B	2589	0.0002	0.0001	0.0002	0	0	0.9994	0
C	2758	0.0039	0.9592	0.0077	0.0003	0.0284	0.0001	0.0003
D	1465	0.3841*	0.2337	0.2917	0.0045	0.0522	0.0282	0.0056
E	220	0.1431*	0.1429	0.1429	0.1428	0.1429	0.143	0.1425
F	1114	0.0178	0.001	0.0019	0.5741	0.0181	0	0.387
G	264	0.0004	0.0125	0.8674	0.0002	0.1183	0.0006	0.0007

 Some interesting observations:
* Clusters A, D and E were assigned the same label
* Clusters E and G have very low pixel counts
* Labels 5 and 7 were not assigned to any cluster.

Need to have some mechanism to deal with clusters that have very low pixel counts


 Here are some more runs comparing our modified K-means with normal k-means. On average, our modified version does segment the image closer to the target segmentation, but there is plenty of room for improvement.

At this point we will start designing the parse tree for slightly more complicated indoor scenes, as well as a bigram model to exploit spatial and contextual relationships between objects. We will also design a mechanism to iteratively modify the value of k by merging and dividing clusters.