kNN Results

Experiment Setup: The 400 odd examples where used as positive examples and the image was run on the image marked up by Dr. Marc. 13248 examples generated by rotating converged synapses where taken as positive examples. Randomly selected other patches (14400) where selected as negative examples. k = 5 Nearest Neighbor classifier was used as the classifier. The test data was the drifted SIFT points generated by the SIFT algorithm.

Results: There where totally 98.94% of SIFT points identified as synapses. That is whopping 49040 (of 49567) synapses in place of 65 identified :(.

Modifications for the next experiment:

1. Ground truth data: Mr. Marc's dataset will be used as the training dataset and first dataset as testing dataset as per Antonio's recommendation.
   
2. Convergence of centroids: Need to understand if the centroid is converging to the darkest or largest dark patch
3. Pick tougher negative examples: The negative examples selected last time where random locations. This time it will be the drifted SIFT point farther than (2 * diskSize) distance.
4. Performance enhancement: We need C Code for distance calculation and interface with Matlab.