Researchers have developed a new method called Semantic-Edge Response Decoding (SERD) to improve crack segmentation for infrastructure inspection. This technique leverages the internal semantic responses of the SAM3 foundation model, which better capture continuous crack evidence than its final mask outputs. SERD interprets these responses as a crack-likelihood field, refines them with an edge prior, and generates masks without requiring any task-specific annotations or fine-tuning. Experiments show SERD significantly outperforms native SAM3 and other zero-shot segmentation methods, achieving a 4.63-point improvement in Crack IoU. AI
IMPACT Improves zero-shot segmentation capabilities for thin and fragmented objects, potentially enhancing infrastructure inspection and health assessment.
RANK_REASON This is a research paper detailing a new method for image segmentation.
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- Gotit.pub
- Hugging Face
- SAM3
- ScienceCast
- Semantic-Edge Response Decoding
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →