Learning Image Feature Extraction: Modeling, Tracking and Predicting Human Performance

In this paper we consider how basic image feature extraction can be posed in terms of the development of a class of Machine Learning algorithms which are capable of tracking and predicting how humans perform tasks such as contour extraction and shape boundary tracking. In particular, we consider how both Recursive Modular Neural Networks (RMNN) and Hidden Markov Models (HMM) can provide reasonably robust models for such tasks. Finally, we investigate how well they can predict human performance and so provide a reasonable basis for the development of more efficient and reliable human-machine annotation systems. Examples in sketching and cartography are discussed.