Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DATTA: Domain-Adversarial Test-Time Adaptation for Cross-Domain WiFi-Based Human Activity Recognition

Julian Strohmayer, Rafael Sterzinger, Matthias Wödlinger, Martin Kampel

TL;DR

This work proposes Domain-Adversarial Test-Time Adaptation (DATTA), a novel framework combining domain-adversarial training, test-time adaptation, and weight resetting to facilitate adaptation to unseen target domains and to prevent catastrophic forgetting.

Abstract

Cross-domain generalization is an open problem in WiFi-based sensing due to variations in environments, devices, and subjects, causing domain shifts in channel state information. To address this, we propose Domain-Adversarial Test-Time Adaptation (DATTA), a novel framework combining domain-adversarial training (DAT), test-time adaptation (TTA), and weight resetting to facilitate adaptation to unseen target domains and to prevent catastrophic forgetting. DATTA is integrated into a lightweight, flexible architecture optimized for speed. We conduct a comprehensive evaluation of DATTA, including an ablation study on all key components using publicly available data, and verify its suitability for real-time applications such as human activity recognition. When combining a SotA video-based variant of TTA with WiFi-based DAT and comparing it to DATTA, our method achieves an 8.1% higher F1-Score. The PyTorch implementation of DATTA is publicly available at: https://github.com/StrohmayerJ/DATTA.

DATTA: Domain-Adversarial Test-Time Adaptation for Cross-Domain WiFi-Based Human Activity Recognition

TL;DR

This work proposes Domain-Adversarial Test-Time Adaptation (DATTA), a novel framework combining domain-adversarial training, test-time adaptation, and weight resetting to facilitate adaptation to unseen target domains and to prevent catastrophic forgetting.

Abstract

Cross-domain generalization is an open problem in WiFi-based sensing due to variations in environments, devices, and subjects, causing domain shifts in channel state information. To address this, we propose Domain-Adversarial Test-Time Adaptation (DATTA), a novel framework combining domain-adversarial training (DAT), test-time adaptation (TTA), and weight resetting to facilitate adaptation to unseen target domains and to prevent catastrophic forgetting. DATTA is integrated into a lightweight, flexible architecture optimized for speed. We conduct a comprehensive evaluation of DATTA, including an ablation study on all key components using publicly available data, and verify its suitability for real-time applications such as human activity recognition. When combining a SotA video-based variant of TTA with WiFi-based DAT and comparing it to DATTA, our method achieves an 8.1% higher F1-Score. The PyTorch implementation of DATTA is publicly available at: https://github.com/StrohmayerJ/DATTA.

Paper Structure

This paper contains 31 sections, 11 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Contributions
  3. Related Work
  4. Domain-Adversarial Training
  5. Test-Time Training
  6. Test-Time Adaptation
  7. Domain-Adversarial Test-Time Adaptation
  8. Domain-Adversarial Training
  9. Model Input/Output
  10. Augmentation Module
  11. Feature Extractor
  12. Activity Recognizer
  13. Domain Discriminator
  14. Domain-Adversarial Loss
  15. Test-Time Adaptation
  16. ...and 16 more sections

Figures (3)

  • Figure 1: Overview of the domain-adversarial training approach used in DATTA.
  • Figure 2: Widar3.0 recording setup featuring a single transmitter TX and six receivers RX$_{1-6}$. A-E indicate the positions at which activities (hand gestures) are carried out.
  • Figure 3: TTA performance across continuous test domain sequences. From top to bottom: (1) ascending domain order (D0 to D8), (2) descending domain order (D8 to D0), and (3) alternating domain order with prolonged domains D0 and D2. Depicted are F1-Scores computed with a rolling window of 100 samples for DATTA models with weight resetting ($W_{\text{DATTA+R}}$, blue), without weight resetting ($W_{\text{DATTA}}$, green), and the baseline DAT model without TTA ($W_{\text{DAT}}$, black).