Fig. 5
From: Digital reality: a model-based approach to supervised learning from synthetic data
Different sampling strategies to generate additional training data. a) Consider a parametric model with two parameters. Every point in the parameter space (white crosses) corresponds to a concrete instance of a scenario (simulation ready scene). b) As a first sampling strategy, class balance can be achieved by generating the same number of instances in the parts of the parameter space that corresponds to every output class. In the example, six scenarios are generated for each class a, b, and c. After training and running initial classification experiments, it becomes clear that the system has difficulties differentiating certain instances of class b and c (black crosses). c) A straight forward sampling strategy is to generate additional samples in the classes b and c using uniform random sampling. d) A more controlled approach is to use some version of importance sampling to generate additional samples close to known misclassified samples. This requires bookkeeping of the parameters for each training data point. e) Additional training data can also be generated in pairs. Hereby, two similar data points are generated that differ in one parameter only but fall in the different classes