Skip to main content
ResearchProjects

AI

AI models can lose accuracy when the images they encounter differ from their training data, for example because of noise, blur, or changing environments. Continual test-time adaptation allows models to adjust to new data without requiring additional labels.
This research investigates a fundamental design choice: which parts of an image should be masked during adaptation? By systematically comparing spatial and frequency-based masking, the study aims to identify approaches that support stable and reliable AI performance over time.

01 Sep 2025 - 31 Aug 2027

EU HORIZON-MSCA-2024-PF-01

About the project

  • Background

    The study finds that spatial patch masking generally provides more stable adaptation for vision transformers on standard image-corruption benchmarks. Frequency masking can lead to substantial performance degradation in these settings.

    However, there is no single best approach for every model or task. The differences are smaller for convolutional neural networks, while frequency masking can be competitive for some fine-grained recognition tasks using larger vision transformers.

    These results highlight the importance of matching the masking technique to the model architecture and application.

  • Project goals

    The project aims to improve the reliability and adaptability of artificial intelligence (AI) models when operating in changing real-world environments. It focuses on understanding how different image masking techniques influence the performance and stability of computer vision models during continual test-time adaptation.

Participants

  • NMBU participants

  • External participants