Finding objects with hypothesis testing
Emanuele Franceschi, Francesca Odone, Fabrizio Smeraldi and Alessandro Verri
In: Learning for Adaptable Visual Systems, 22 Aug 2004, Cambridge, UK.
We present a trainable method for detecting objects
in images from positive examples, based on hypothesis testing.
During training a large number of image features is computed and the empirical probability distribution of each measurement is estimated from the available examples.
Through a two--step feature selection method we obtain
a subset of N discriminative and pairwise independent features.
At run time, a hypothesis test is performed for each feature at
a fixed level of significance. The null hypothesis is, in each case, the presence of the object. An object is detected if at least M of the N tests are passed. The overall significance level depends on M as well as on the level of the single tests. We report experiments on face detection, using the CBCL-MIT database for training and validation, and images randomly downloaded from the Web for testing.
The image measurements we use for these experiments include grey level values, integral measurements, and ranklets. Comparisons with whole face detectors
indicate that the method is able to generalize from positive
examples only and reaches state-of-the-art recognition rates.
|EPrint Type:||Conference or Workshop Item (Paper)|
|Project Keyword:||Project Keyword UNSPECIFIED|
|Deposited By:||Alessandro Verri|
|Deposited On:||23 November 2004|