Continual Test-Time Adaptation (TTA)

Continual Test-Time Adaptation - Domain Adaptation 🧠🔄

Dissertation Thesis Project
Supervisor: Professor Ioannis Pratikakis
This project introduces a continual test-time adaptation method designed to handle domain shifts in real-world scenarios by adapting models on the fly without access to source data or retraining (Online Adaptation).

🎯 Project Objectives

• Project focuses on object detection task

• Address Real-world Domain Shifts
  • Develop techniques to adapt models to changing domains at test time
  • Utilize unsupervised and self-supervised signals during inference
  • Enhance robustness of models in dynamic environments
• Evaluation & Benchmarks
  • Evaluate on various benchmarks to demonstrate improvements (Shift, Cityscapes, Kitti, Clad, Coco)
  • Compare against existing test-time adaptation methods, such as the Mean Teacher framework

🚀 Key Contributions & Outcomes

• Designed a novel test-time adaptation approach specifically tailored for continual test-time adaptation scenarios
• Achieved significant improvements over no adaptation and the Mean Teacher framework
• Proposed and evaluated on four benchmark datasets, simulating continual domain shifts
• Published experimental setups and results can foster further research in the field

📄 Thesis & Publications

• Dissertation Thesis — Publication / Slides
• Mean Teacher with Stochastic Restoration and Contrastive Learning for Domain Adaptive Object Detection — Publication / Slides

🖥️ Code & Implementation

Explore the full implementation on GitHub.