Details
- Supervisors
- Faculty
- Degree label
- Abstract
- Most machine learning problems assume that past training data follows the same distribution as future test data. However, this assumption often does not hold, such that the so-called source domain, from which a model learns, is distributed differently from the so-called target domain, which contains the data on which predictions are made. Hence, the learning algorithm fails to generalize. Domain adaptation has been proposed to address such cases, and aims to improve the performance of a model on the target domain using the knowledge learned in the source domain. In this work, we use domain adaptation to solve a traffic sign detection task with images taken under different conditions. For this purpose, we apply an adversarial training technique that is able to integrate into a deep learning algorithm and learn a domain-invariant representation of the input. We train it in an unsupervised fashion, such that only the ground-truth labels for the source domain are revealed, while those of the target domain are hidden. We show that the resulting adaptive model is effective in learning a detector that performs better on the target domain than a non-adaptive model trained solely on the source domain.