# DELTA: degradation-free fully test-time adaptation

**Authors:** Bowen Zhao, Chen Chen, Shu-Tao Xia  
**arXiv:** [2301.13018](https://arxiv.org/abs/2301.13018)  
**Date:** 2023-01-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.13018v1/

## Key points

- The paper proposes a solution to adapt pre-trained models to test streams during real-time inference when the test distribution differs from the training distribution.
- Prevalent adaptation methodologies like test-time batch normalization (BN) and self-learning have two unfavorable defects: inaccurate estimates resulting from normalization statistics in test-time BN being completely affected by the currently received test samples, and biased parameter updates towards some dominant classes during test-time adaptation.
- Previous approaches work well in certain scenarios while showing performance degradation in others due to their faults.
- The proposed solution called DELTA for Degradation-freE fuLly Test-time Adaptation consists of two components: Test-time Batch Renormalization (TBR), introduced to improve the estimated normalization statistics; and Dynamic Online re-weighTing (DOT), designed to address class bias within optimization.
- The study investigates various test-time adaptation methods on three commonly used datasets with four scenarios and a newly introduced real-world dataset. DELTA can help deal with all scenarios simultaneously, leading to state-of-the art (SOTA) performance.
- Defects can be exacerbated in more complicated test environments such as time dependent or class imbalanced data, highlighting the importance of developing robust solutions for fully test time adaptation.

## Summary

The paper titled "DELTA: degradation-free fully test-time adaptation" proposes a solution to the problem of adapting a pre-trained model to the test stream during real-time inference, which is urgently required when the test distribution differs from the training distribution. While several efforts have been devoted to improving adaptation performance, the authors find that two unfavorable defects are concealed in prevalent adaptation methodologies like test-time batch normalization (BN) and self-learning. The first defect is related to inaccurate estimates resulting from normalization statistics in test-time BN being completely affected by the currently received test samples. The second defect pertains to biased parameter updates towards some dominant classes during test-time adaptation. The authors observe that previous approaches work well in certain scenarios while showing performance degradation in others due to their faults. In this paper, they provide a plug-in solution called DELTA for Degradation-freE fuLly Test-time Adaptation, which consists of two components: Test-time Batch Renormalization (TBR), introduced to improve the estimated normalization statistics; and Dynamic Online re-weighTing (DOT), designed to address class bias within optimization. The study investigates various test-time adaptation methods on three commonly used datasets with four scenarios and a newly introduced real-world dataset. DELTA can help deal with all scenarios simultaneously, leading to state-of-the art (SOTA) performance. The authors further note that defects can be exacerbated in more complicated test environments such as time dependent or class imbalanced data, highlighting the importance of developing robust solutions for fully test time adaptation. Overall, this paper provides valuable insights into improving adaptation performance and presents an effective solution for addressing existing faults in prevalent methodologies.