Comparison of Appearance-Based Methods for Generic Object Recognition

—In this paper, we present a large-scale examination of different appearance-based, segmentation-free classification methods for their usability in generic object recognition. Generic object recognition is a method to handle the objects never seen before in classification by a hierarchical approach with a coarse-to-fine graduation. Unknown objects are only classified into coarse categories and rejected to be assigned to classes that are too specific. Comparison of PPCA, NN, and KPCA approaches is made on the basis of their recognition rate. The global generic recognition rate is computed for the best method, and its robustness according to dif-ferent types of noise is examined. Our experiments show that the PCA-based method with nearest neighbor classification provides, in general, the best recognition rates, whereas the models based on principal component analysis outperform the other methods in computation time and model size. Gaussian KPCA models can be used for generic object recognition by varying the variance of the Gaussian kernel