Method and System for Training a Landmark Detector using Multiple Instance Learning

An apparatus for training a landmark detector, comprising:

• means for receiving training data including a plurality of training bags including:
• a plurality of positive training bags, each including a plurality of positively annotated instances;
• and a plurality of negative training bags, each including at least one negatively annotated instance;
• means for initializing a classification function by training a first weak classifier based on the positive training bags and the negative training bags;
• means for evaluating all training instances using the classification function;
• means for calculating a cost value gradient based on spatial context information of each instance in each positive bag evaluated by the classification function;
• means for calculating a gradient value associated with each weak classifier based on the cost value gradients;
• means for selecting a weak classifier of the plurality of remaining weak classifiers having a lowest associated gradient value;
• means for determining a weighting parameter associated with the selected weak classifier;
• and
• means for adding the selected weak classifier to the classification function.

A non-transitory computer readable medium encoded with computer executable instructions for training a landmark detector, the computer executable instructions defining steps comprising:

• receiving training data including a plurality of training bags including:
• a plurality of positive training bags, each including a plurality of positively annotated instances;
• and a plurality of negative training bags, each including at least one negatively annotated instance;
• initializing a classification function by training a first weak classifier based on the positive training bags and the negative training bags;
• evaluating all training instances using the classification function;
• for each of a plurality of remaining weak classifiers:
• calculating a cost value gradient based on spatial context information of each instance in each positive bag evaluated by the classification function;
• calculating a gradient value associated with each weak classifier based on the cost value gradients;
• selecting a weak classifier of the plurality of remaining weak classifiers having a lowest associated gradient value;
• determining a weighting parameter associated with the selected weak classifier;
• and
• adding the selected weak classifier to the classification function.