SIFR

SIFR is a framework for domain generalized semantic segmentation (DGSS) that takes spectral modeling as a unified principle, simultaneously regulating the interaction of heterogeneous prior features and the refinement process of the dense decoder.

In the feature fusion stage, SIFR adopts spectrum-guided invariant fusion, performing spectral analysis before fusion to filter out response components sensitive to domain changes while preserving the complementary parts of semantic and structural information.

SIFR does not modify the CLIP and DINOv2 base models themselves; instead, it performs spectral prior screening at the feature interaction level.

In the decoding stage, SIFR introduces a normalization strategy based on low-frequency statistics, together with an adaptive frequency selection mechanism, to stabilize multi-scale decoder features.

SIFR achieves consistent improvements over previous state-of-the-art methods on three DGSS benchmark scenarios: synthetic-to-real, real-to-real, and adverse weather.

SIFR is designed so that the model can simultaneously improve domain robustness and segmentation accuracy under severe distribution shift, whereas in previous approaches these two often trade off against each other.

Spectral analysis relies on the prior assumption about "which frequency bands correspond to domain-sensitive information," and the paper has not yet sufficiently discussed whether this assumption generally holds across different datasets and degradation types.

The feasibility of the additional computational overhead introduced by SIFR's dual-stream architecture in real-time segmentation scenarios needs more data to support it.

SIFR currently performs spectral filtering at the fusion level; if it could allocate the contribution weights of the two base models by frequency band at a finer granularity, it might unlock greater potential.

SIFR's contribution lies in elevating frequency-domain modeling into a unified principle running through both fusion and decoding, rather than inventing a new base model or loss function.