Tatiana Tommasi

Over the last years, several authors have signaled that state of the art categorization methods fail to perform well when trained and tested on data from different databases. The general consensus in the literature is that this issue, known as domain adaptation and/or dataset bias, is due to a distribution mismatch between data collections.

But what if broadly adopted features and classifiers would be part of the problem, rather than good ingredients for its solution?

Since the seminal work of Boiman et al., the Naive Bayes Nearest Neighbor (NBNN) method has challenged (i) the vector quantization step in Bag of Words (BOW) descriptors, that allows to have a compact feature representation to the expenses of its informative content, and (ii) the computation of image-to-image distances, that enables to use kernel based classification methods, but that does not generalize much beyond the labelled images.

Even though the domain adaptation/dataset bias problem is clearly at its core a generalization problem, the almost totality of approaches presented so far use image-to-image learning algorithms on top of BOW representations.