Home | Search | Help  
Home Page Università di Genova

Technical Report Details

Date 1-6-2007
Number DISI-TR-07-01
Title A regularized approach to feature selection for face detection
Authors A. Destrero, C. De Mol, F. Odone, A. Verri
Bibtex Entry
E-mail odone@disi.unige.it
Abstract In this paper we present a trainable method for selecting features from an overcomplete dictionary of measurements. The starting point is a thresholded version of the Landweber algorithm for providing a sparse solution to a linear system of equations. We consider the problem of face detection as a case study and adopt rectangular features as an initial representation for allowing straightforward comparisons with existing techniques. For computational efficiency and memory requirements, instead of implementing the full optimization scheme on tenths of thousands of features, we propose to solve a number of smaller size optimization problems obtained by randomly subsampling the feature vector. Then, we compile a list of the selected features by choosing all the features which have been selected each time they have been sampled. Strong empirical evidence of the similarity of the obtained solution with respect to the one which can be obtained by implementing the full optimization scheme is provided. The obtained set of features is still highly redundant, so we apply the same feature selection algorithm in a second selection stage on the list resulting from the first stage, and obtain a final list of features. This list is far more effective for detection purposes than the list obtained by enforcing sparse solutions of similar cardinality through appropriate tuning of the regularization parameter in the first stage. Experimental results of an optimized version of the method obtained on benchmarks and newly acquired face images and image sequences indicate that this method is a serious competitor of other feature selection schemes recently popularized in computer vision for dealing with problems of real time object detection.
Back to Technical Reports