A Model-based Approach for Rigid Object Recognition

Most object recognition systems require large databases of real images for classifier training. To collect real images for this purpose is a difficult and expensive process. This paper introduces a unified framework based on the creation and use of synthetic images for training various classifiers to achieve recognition of real-world objects. A 3D model of the object (i.e. trolley in this case) is constructed from a minimum of two photographs. The constructed 3D model is used to automatically generate the relevant synthetic images that are subsequently used to train the Adaboost and support vector machine-based recognition systems. Experimental results obtained are very encouraging suggesting that synthetically generated images generated by our approach can augment the real training samples used in current recognition systems