Training-Free Semantic-Edge Response Decoding of SAM3 for Cross-Domain Infrastructure Crack Segmentation

Shipeng Liu, Zhanping Song, Liang Zhao, Dengfeng Chen

Cross-project crack segmentation is hindered by variations in materials, imaging conditions, crack morphology, and background interference. Text-promptable foundation models reduce task-specific training, but SAM3's final region proposals may suppress, truncate, or distort weak and fragmented cracks. We propose \textbf{S}emantic-\textbf{E}dge \textbf{R}esponse \textbf{D}ecoding (\textbf{SERD}), a training-free method that replaces the proposal interface with the internal language-conditioned response. SERD normalizes this dense response, calibrates it using a fixed Sobel structural prior, and applies a single threshold to generate the crack mask. SAM3 remains frozen, with no target-domain annotation, auxiliary prompting network, or test-time optimization. When CamCrack789 is used only for threshold selection, SERD achieves 58.00\% average Crack IoU on five unseen datasets, compared with 54.33\% for native SAM3. Across six rotated source-domain settings, SERD obtains 60.23\% mean target-domain IoU and 67.18\% Boundary F1, exceeding SAM3 by 3.27 and 2.70 percentage points. Regional, structural, foreground-area, robustness, and latency analyses show that direct response decoding recovers substantial crack evidence lost during proposal formation, while structural calibration improves precision and boundary localization. The results demonstrate that internal-response decoding provides a simple and transferable interface for cross-domain infrastructure crack segmentation. \textit{Code is available at: \href{https://github.com/xauat-liushipeng/SERD}{GitHub}}.