Font Size: 

Object categories from small samples with highperformance is a challenging problem, where machine learningtools can in general provide very few guarantees. Exploiting priorknowledge may be useful to reproduce the human capabilityof recognizing objects even from only one single view. Thispaper presents an SVM-based model adaptation algorithm ableto select and weight appropriately prior knowledge coming fromdifferent categories. The method relies on the solution of a convexoptimization problem which ensures to have the minimal leaveone-out error on the training set. Experiments on a subset of theCaltech-256 database show that the proposed method producesbetter results than all choosing one single prior model, andtransferring from all previous experience in a flat uninformativeway.

Algorithm