multiclass object recognition with sparse, localized features.pdfVIP

Multiclass Object Recognition with Sparse, Localized Features Jim Mutch and David G. Lowe Department of Computer Science University of British Columbia #201 - 2366 Main Mall Vancouver, B.C., Canada, V6T 1Z4 mutch,lowe@cs.ubc.ca Abstract ing the sparsity of basis functions [7, 14] (equivalent to re- ducing the capacity of the classifier) plays an important role We apply a biologically inspired model of visual object in improving generalization performance. Similarly, within recognition to the multiclass object categorization pr oblem. computational neuroscience, it has been found that adding a Our model modifies that of Serre, Wolf, and Poggio. As in sparsity constraint is critical for learning biologically plau- that work, we fir st apply Gaborfilter s at all positions and sible models from the statistics of natural images [20]. For scales;featur e complexity andposition/scale invariance are object class recognition, one way we have found to increase then built up by alternating template matching and max sparsity is to use a lateral inhibition model that eliminates pooling operations. We refine the approach in several bi- weaker responses that disagree with the locally dominant ologically plausible ways, using simple versions of sparsi- ones. We further enhance this approach by matching only fication and lateral inhibition. We demonstrate the value of the dominant orientation at each position within a feature retaining some position and