Abstract

---

Test-time adaptation (TTA) aims to address distributional shifts between training and testing data using only unlabeled test data streams for continual model adaptation. However, most TTA methods assume benign test streams, while test samples could be unexpectedly diverse in the wild. For instance, an unseen object or noise could appear in autonomous driving. This leads to a new threat to existing TTA algorithms; we found that prior TTA algorithms suffer from those non-interest test samples as they blindly adapt to incoming samples. To address this problem, we present Screening-out Test-Time Adaptation (SoTTA), a novel TTA algorithm that is robust to non-interest samples. The key enabler of SoTTA is two-fold: (i) input-wise robustness via high-confidence uniform-class sampling that effectively filters out the impact of non-interest samples and (ii) parameter-wise robustness via entropy-sharpness minimization that improves the robustness of model parameters against large gradients from non-interest samples. Our evaluation with standard TTA benchmarks with various non-interest scenarios shows that our method outperforms state-of-the-art TTA methods under the presence of non-interest samples and achieves comparable accuracy to those methods without non-interest samples.

Publications

---

SoTTA: Robust Test-Time Adaptation on Noisy Data Streams
PDF Code Video/Poster OpenReview

People

---