Core Concepts
Proposing a method for semi-supervised domain generalization by leveraging feature-based conformity and semantics alignment to address key challenges.
Abstract
The article addresses the challenge of semi-supervised domain generalization (SSDG) by proposing a method that leverages feature-based conformity and semantics alignment. Existing methods struggle with exploiting unlabeled data, leading to poor performance in SSDG settings. The proposed approach aims to align posterior distributions from the feature space with pseudo-labels from the model's output space. By introducing a feature-based conformity technique and a semantics alignment loss, the method enhances model performance in SSDG settings. The plug-and-play nature of the approach allows seamless integration with different SSL-based SSDG baselines without additional parameters. Experimental results across challenging DG benchmarks demonstrate consistent and notable gains in two different SSDG settings.
Stats
Extensive experimental results across five challenging DG benchmarks.
Consistent and notable gains in two different SSDG settings.
Quotes
"Our method is plug-and-play and can be readily integrated with different SSL-based SSDG baselines without introducing any additional parameters."
"Extensive experimental results suggest that our method provides consistent and notable gains in two different SSDG settings."