Diploma Dissertation Thesis - Continual Learning for Object Detection

In recent years, the development of neural networks has led to significant progress in various fields of artificial intelligence, including computer vision. The effectiveness of these algorithms mainly relies on the availability of large volumes of annotated data for training the networks. However, it is known that a model trained on a specific domain experiences rapid performance deterioration when applied to different domains. Recent methods aim to improve the performance of networks, regardless of the domain in which they are used. Test-time adaptation is a subset of domain adaptation and allows a model to update its weights during inference, in order to be able to adapt to new domains. Networks that are trained on new tasks exhibit reduced performance on tasks for which they were previously trained, a phenomenon known as catastrophic forgetting. This has led to the emergence of a new field in machine learning called continual learning. In this thesis, research is conducted to study the problem of continual test-time adaptation for object detection, which aims to adapt the model to a variety of new domains and maintaining enhanced performance both in the new domains and in the initial domain in which it was originally trained. Methods of semi-supervised and self-supervised learning are studied to achieve adaptation, in combination with techniques to eliminate catastrophic forgetting. The experimental results are encouraging and indicate that domain adaptation is feasible and can significantly enhance the performance and stability of deep learning algorithms.

Download Paper View Slides