# DCFS: Continual Test-Time Adaptation via Dual Consistency of Feature and Sample

Source: [DCFS: Continual Test-Time Adaptation via Dual Consistency of Feature and Sample](<https://www.sota2.com/research/paper/68fa7b001e209131cea8628c>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2025-10-26T05:31:45.237Z

Authors: Wenting Yin, Han Sun, Xinru Meng, Ningzhong Liu, Huiyu Zhou

Published: 2025-08-28T07:57:54.000Z

## Abstract

Continual test-time adaptation aims to continuously adapt a pre-trained model to a stream of target domain data without accessing source data. Without access to source domain data, the model focuses solely on the feature characteristics of the target data. Relying exclusively on these features can lead to confusion and introduce learning biases. Currently, many existing methods generate pseudo-labels via model predictions. However, the quality of pseudo-labels cannot be guaranteed and the problem of error accumulation must be solved. To address these challenges, we propose DCFS, a novel CTTA framework that introduces dual-path feature consistency and confidence-aware sample learning. This framework disentangles the whole feature representation of the target data into semantic-related feature and domain-related feature using dual classifiers to learn distinct feature representations. By maintaining consistency between the sub-features and the whole feature, the model can comprehensively capture data features from multiple perspectives. Additionally, to ensure that the whole feature information of the target domain samples is not overlooked, we set a adaptive threshold and calculate a confidence score for each sample to carry out loss weighted self-supervised learning, effectively reducing the noise of pseudo-labels and alleviating the problem of error accumulation. The efficacy of our proposed method is validated through extensive experimentation across various datasets, including CIFAR10-C, CIFAR100-C, and ImageNet-C, demonstrating consistent performance in continual test-time adaptation scenarios.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2508.20516>)

[PDF](<https://arxiv.org/pdf/2508.20516.pdf>)
