Artiom Myaskouvskey (CS, Technion)
We describe a new part based detection algorithm.
The proposed algorithm uses an empirical model,
based on data extracted from the test image to bound
the probability that a model candidate arises from a noncategory-related image.
The decision is adaptive and does not rely on parameters optimized for possibly unrelated non-category images.
Finally, the decision process provides an bound on the number of false alarms in the image.
We explain the a-contrario method, derive the empirical model, and specify a principled detection procedure
for simple models and for more complex ones.
Experiments show the validity of the predictions and their utility for setting detection parameters.
*Msc research under the supervision of Prof. Micha Lindenbaum