Conclusion
My training set of 20 was probably 'hard' since the detectors generally performed better across the board on the full 200.
The first modification, the shuriken filters, noticably hindered performance. If I had it to do over again,
I would leave them out and include a set of Difference of Gaussian or other center-surround-like filters which have seen success
in biology and in silicon.
The second modification, the gradient-gradient, also noticably hindered performance. If I were to try this again,
I would include some better weighting scheme such as SIFT and HOG employ. I might even use a proven implementation
for this purpose. I would also only consider the masks in the orientation of the actual image gradient at each point,
since that is the orientation of the edge that one is proposing. This seems like it would give better values for ALL
of the gradient values (bg, tg, and gg), but who knows.
The third modification, the single SVM, in general attained a high precision but poor overall recall.
The fourth modification, the eccentric ensemble SVM, gave a definite overall boost to performance, flat out beating the Canny-baseline
for the first half of the curve, and then crossing at a common point. I attribute the crossing point
to the weakness of the underlying features, which incorporated the shuriken filters and the gradient gradient.
I spent no time whatsoever tuning the underlying features... By the time that I realized that I could cache the time-consuming
calculations and tweak the parameters of the combination of gradients without spending 20 hours re-crunching, I really only
had time to implement and tweak the SVM portion on the old features I kept.
So, if I were to do this over again, I would use proven filters such as center-surround, cache the time consuming calculations
so that I could play with them efficiently, attempt to consider masks only of the appropriate orientation, and tweak the weighting combination
of {canny, bg, tg} for optimal values. I would then take these superior features and play the SVM game with them, which I
predict would give an even better boost.