|M.Sc Student||Myaskouvskey Artiom|
|Subject||Using A Contrario Methology for Part-Based Object|
|Department||Department of Computer Science||Supervisor||Professor Michael Lindenbaum|
|Full Thesis text|
Our work is a variation of a known part based detection algorithm, which differs mostly in the way the parameters are set. Using the a contrario methodology, the detection thresholds are fixed only by controlling the false detection rate (false alarms). False alarm prediction is not based on training data but is done online using only data from the particular (test) image at hand and some simple statistical assumptions. A contrario based detection has several advantages. First, the detection parameters are tuned adaptively to the image at hand and not by trying to optimize average performance over some training image set, which may be nonrepresentative for specific images. Second, any detection decision made by the a contrario based algorithm is accompanied by a reliability estimate (i.e., confidence). For many methods, no such confidence measure is available. Third, the algorithmic process becomes very flexible. There is a uniform criterion for all decisions, even if made with a different number of parts or with multiple models.
The main contributions of our work are the evaluation of this new statistical model for a contrario and the empirical validation of the a contrario criteria on real-world images. We show that applying a contrario methodology allows us to predict the false detection rate and therefore makes the algorithm more stable with respect to it. We propose several variations of the a contrario detection algorithm that differ by how they predict the false alarm rate (with or without considering the correlations between the model parts), whether they rely on one or many models per category, and how they detect the optimal number of parts in the model: automatically or using a predefined number.
We tested our approach on the Caltech-4 dataset and showed that we are able to approximate false alarm rate. We tested our algorithm on the Ponce texture database and showed that the model distance, using a contrario, is more robust than the standard (L1 metric) distance, when training and test image distributions differ.