Reproduce “Semantic-Aware Domain Generalized Segmentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, Wen Li
This research introduces Semantic-Aware Normalization (SAN) and Semantic-Aware Whitening (SAW) modules designed to improve Domain Generalized Semantic Segmentation (DGSS). The proposed framework achieves state-of-the-art performance on various unseen target domains by learning features that are both domain-invariant and semantically discriminative, notably without significant degradation of source domain performance.
Key insights
- Global normalization and whitening methods, while aiming for domain invariance, can inadvertently collapse distinct semantic features, leading to loss of discriminability and inaccurate segmentation boundaries.
- Integrating category-level semantic information into feature alignment processes (normalization and whitening) allows models to achieve both effective style elimination and robust feature discrimination simultaneously.
- By promoting intra-category compactness and inter-category separability, models can generalize to unseen domains without compromising performance on the source domain.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.