Reproduce “CoRe-SAM3: Conditional Semantic--Visual Reconciliation for SAM3 Crack 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
Shipeng Liu, Liang Zhao, Dengfeng Chen
CoRe-SAM3 adapts the SAM3 vision foundation model for crack segmentation by reconciling its internal semantic and visual representations in a conditional, parameter-efficient manner. The framework improves average Crack IoU by 8.13 percentage points across five datasets, utilizing only 18.914 K trainable parameters.
Key insights
- SAM3's internal semantic representation is the primary carrier of crack-related information, while its native visual representation is most effective when conditionally applied for localized error correction rather than direct prediction.
- Directly combining semantic and visual features for error prediction is less effective than leveraging visual evidence conditionally to refine an existing semantic belief, suggesting an asymmetric role for these internal representations.
- Significant performance improvements in crack segmentation can be achieved with a minimal increase in trainable parameters (0.0023% of SAM3's total) by designing an adaptation that precisely reconciles internal representations rather than adding extensive external capacity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.