TEST-TIME TRAINING-FREE DOMAIN ADAPTATION

Deploying deep learning models to new environments is very challenging. Domain adaptation (DA) is a promising paradigm to solve the problem by collecting and adapting to unlabeled data in new environments. Though research efforts have led to steady performance improvement over the past decade, DA algorithms are still hard to deploy. To make DA practical, in this paper we study a problem named Test-time Training-Free Domain Adaptation (TTDA), where trained models adapt to a single input without training. By exploiting spatial activation that was previously overlooked and simply averaged out, we propose a simple method based on Feature Statistics Transformation on-the-fly for each test example. The proposed algorithm is tested in the TTDA setting on two standard DA benchmarks. Surprisingly, it surpasses or performs on par with SOTA training DA methods. We envision this training-free paradigm has the potential to bring DA to embedded devices and be of interest to audience of community.