Segmentation Using Metropolis Algorithm

Segmentation Identifying objects in a picture by attaching each pixel to a specific segment Pixels are characterized by different parameters: Place, RGB, gradient, etc.

The Metropolis Algorithm Finds global minimum in any function Based on Thermodynamics basic assumption: while not finished if energy isn't changed for the 5th time finish else choose a random point P_new and compare to current point P_current the distance to the new point is proportional to T/Tmax if E(P_new) < E(P_current) switch to the new point if E(P_new) < E(P_best) choose P_new to be P_best choose a random number m if m < exp(-(E(P_new) - E(P_current))/T) if this is the 20'th iteration decrease T by 10 percent

Implementation Energy calculation is based on the sum of min distances of each pixel from its center Distance is in 11-dimensional space: x, y, RGB, gradient Compared against Segmentation Using Clustering Algorithm Comparison based on number of indexing

Poor Results Original Metropolis Clustering

Conclusion - Metropolis Not better than any algorithm Exact centers are less important than converging Calibration of Metropolis is not clear

Conclusion - Segmentation Number of segments is not clear Energy function is not clear Parameters to search are not clear