Advances in Domain Generalization and Adaptation

The field of domain generalization and adaptation is moving towards leveraging powerful zero-shot capabilities of models like CLIP to enhance model robustness across diverse environments. Research is focused on developing innovative techniques to improve domain generalization and adaptation, including the use of adaptive thresholding, feature filtering, and spectral adaptive information propagation. Noteworthy papers include:

  • CLIPXpert, which proposes a novel SF-OSDA approach that integrates adaptive thresholding and unknown class feature filtering, outperforming state-of-the-art methods.
  • FrogDogNet, which introduces a novel prompt learning framework that integrates Fourier frequency filtering and self-attention to improve RS scene classification and domain generalization, consistently outperforming state-of-the-art prompt learning methods.

Sources

Revisiting CLIP for SF-OSDA: Unleashing Zero-Shot Potential with Adaptive Threshold and Training-Free Feature Filtering

CLIP-Powered Domain Generalization and Domain Adaptation: A Comprehensive Survey

AI for the Open-World: the Learning Principles

Invariant Learning with Annotation-free Environments

FrogDogNet: Fourier frequency Retained visual prompt Output Guidance for Domain Generalization of CLIP in Remote Sensing

SAIP-Net: Enhancing Remote Sensing Image Segmentation via Spectral Adaptive Information Propagation

Energy-Based Pseudo-Label Refining for Source-free Domain Adaptation

Built with on top of