SIFR fuses CLIP and DINOv2 spectrally for domain-general segmentation
Original title: Spectral Invariant Fusion and Refinement for Domain Generalized Semantic Segmentation
The method filters domain-sensitive responses at fusion and stabilizes multi-scale decoder features before dense prediction.
Researchers propose SIFR, a frequency-aware dual-stream framework for domain-generalized semantic segmentation without target-domain data during training. Instead of directly fusing frozen CLIP and DINOv2 features, it uses spectral-guided invariant fusion and a spectral decoder refinement module based on low-frequency statistics and adaptive frequency selection. The authors report consistent gains over prior methods on synthetic-to-real, real-to-real, and adverse-weather DGSS benchmarks.
Unconfirmed: Reported gains are the authors' benchmark results and are not independently verified here.
- SIFR regulates both prior fusion and dense decoder refinement with spectral modeling.
- Low-frequency normalization and adaptive frequency selection reduce residual domain bias.
- Evaluations cover synthetic-to-real, real-to-real, and adverse-weather settings.
From the original article
1 min read at arxiv.orgDomain generalized semantic segmentation is challenging because it requires accurate dense prediction on unseen domains without any access to target-domain data during training.
By Yansong Bai · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.