A Simple Test-time Adaptation Method for Source-free Domain Generalization

Vaasudev Narayanan, Sai Srinivas Kancheti, Sriranjani Ramakrishnan, Vineeth N. Balasubramanian

Keywords: Source-free Domain Generalization, Test-time Adaptation

TL;DR: A simple method for solving source-free domain generalization by leveraging unlabeled test-time data.

Abstract: In this paper, we tackle source-free domain generalization (SFDG), where the objective is to perform well on an unseen target domain using only models trained on source domains, without assuming any access to labeled source images. We propose an effective, yet simple method for solving SFDG by using $\textit{unlabeled}$ target data available only during inference to give a dynamic, adaptive prediction at the batch-level. Specifically, during test-time, we (1) pass the test batch through each source model, (2) select as pseudo-label the class with the highest average probability score, (3) minimize cross-entropy loss for each model using the pseudo-label and finally (4) forward pass through the adapted models and predict the class with the highest average probability. We compare our test-time pseudo-labeling method \textit{\name}, with a wide variety of baselines and outperform them on average accuracy across four benchmark DG datasets, namely PACS, OfficeHome, VLCS and TerraIncognita.

Submission Number: 34