DI-BOW: Domain Invariant feature descriptor using Bag of Words

This paper describes a method to learn Bag-Of-Words (BOW) descriptor for image representation which is robust to domain shift. Domain shift is necessary when a classifier trained on one dataset (source) is applied for classification on a different dataset (target). Datasets acquired with different conditions, have dissimilar feature distributions among them. Traditional method for representing each image by a BOW descriptor with the vocabulary learnt on a reference dataset does not work well for such cross-dataset tasks. We propose a new method to learn an amended dictionary composed of class specific atoms. The proposed Domain-Invariant BOW (DI-BOW) descriptor built from this dictionary has much better class discriminability and inherently attenuates domain-specific characteristics, making it more suitable to cross-domain tasks. Results based on DI-BOW descriptor reveal its efficiency, by outperforming state-of-the-art domain adaptation techniques for object recognition.