Unsupervised domain adaptation on object recognition

Abstract

Together with the development of deep neural networks, artificial intelligence is getting unprecedented accuracies on various tasks, including Computer Vision, Natural Language Processing, etc. Accuracies on certain datasets have improved more than 50% in less than ten years. Yet these numbers are acquired in similar datasets, which means the training data and the testing data are from the same domain. When dealing with datasets that fell in another domain, the well-trained model might not be able to output a satisfactory result. In this situation, domain adaptation is proposed as a solution to deal with the differences between data domains. Besides, previous models are not as helpful when dealing with unlabeled datasets, which is unavoidable in real-world situations. With the goal to solve this problem in the area of object recognition, this project wants to create a novel solution for Unsupervised Domain Adaptation on Object Recognition.

In this project, different solutions of unsupervised domain adaptation are evaluated, and their performances are studied, their corresponding advantages and disadvantages are discussed. Then, a new model using a transformer as its backbone is proposed, with the help of pseudo labeling and the Cross Attention mechanism. By using this model, we could not only increase the accuracies of the domain adaptation tasks but also reduce the needed time and resources to train and inference.