Possible SDM training improvements

[patricksnape]

Having looked at other people's implementations of SDMs, we could try the following:

• Customize the 'perturbation' generation allowing things other than bounding boxes. For example, dlib uses a linear combination of one or more shapes from the training set rather than just the ground truth.
• Add more regressors. For example, CFSS uses an averaged linear regressor computed from a subset of the data.
• Generate perturbations at each cascade. Rather than relying on the convergence properties of the algorithm, we would regenerate the perturbations from the current standard deviation of the errors at each cascade to ensure more varied learning for the regressor.