Deep learning has achieved impressive performance in discriminative computer vision tasks, including applications in medical imaging and natural scene analysis. However, these models are highly sensitive to domain shifts changes in visual distribution between training and testing or the emergence of unseen classes at test time, which severely degrade performance. Previous solutions require labeled data from multiple domains or access to source data for adaptation, both of which are impractical in many real-world settings.

This thesis aims to tackle this generalization problem through test-time strategies involving updates of model parameters or input data during inference using only incoming data streams and the model' s own predictions, without access to the original training data. These data streams may exhibit varying unknown distributions, posing significant challenges to robust adaptation. The thesis builds upon recent contributions in test-time adaptation (TTA) and addresses increasingly realistic scenarios where test data may follow shifting, recurring, or temporally correlated and open-set distributions.

The first contribution presents a data-centric approach that transforms test images to bring them closer to the source distribution. This method achieves strong results for simple visual shifts, such as changes in brightness or contrast, but shows its limits when facing more complex variations, motivating the exploration of more robust model-centric strategies. The second contribution identifies that in the presence of open set samples, noisy feature representations in test data drive generalization failures; we overcome this by distilling robust embeddings from a pre-trained self supervised models to guide source free adaptation. Recognizing the limited generalization of these approaches in highly dynamic environments, the work transitions to online TTA. In this setting, the thesis highlights the critical role of normalization layers in the performance degradation of state-of-the-art TTA under temporally correlated data streams and proposes a lightweight, easily integrable test-time normalization technique. Finally, a multi-model strategy is introduced to handle rapidly changing visual domains in dynamic environments.