Core Concepts
The core message of this paper is to enhance the generalizability of Mamba-like state space models (SSMs) towards unseen domains by proposing a novel framework named DGMamba. DGMamba comprises two key modules: Hidden State Suppressing (HSS) to mitigate the detrimental effect of domain-specific information in hidden states, and Semantic-aware Patch Refining (SPR) to encourage the model to focus more on the object rather than the context.
Abstract
The paper proposes a novel framework named DGMamba to enhance the generalizability of Mamba-like state space models (SSMs) towards unseen domains.
The key highlights are:
-
Hidden State Suppressing (HSS): This module is introduced to mitigate the detrimental effect of domain-specific information contained in hidden states by selectively suppressing the corresponding hidden states during output prediction.
-
Semantic-aware Patch Refining (SPR):
- Prior-Free Scanning (PFS): This module randomly shuffles the context patches within images to break the spurious correlations caused by the fixed scanning strategies and provide a more flexible and effective 2D scanning mechanism for Mamba.
- Domain Context Interchange (DCI): This module substitutes the context patches of images with those from different domains, introducing local texture noise and regularizing the model on the combination of mismatched context and object.
The proposed DGMamba achieves state-of-the-art generalization performance on four commonly used domain generalization benchmarks, demonstrating its effectiveness in boosting the generalizability of SSM-based models towards unseen domains.
Stats
The paper does not provide any specific numerical data or statistics. The key results are presented in the form of performance comparisons on various domain generalization benchmarks.