Answers are grounded in this paper and cite their sources.
Automatic crack segmentation in open-pit mining is constrained by the scarcity of representative, labeled datasets. Vision foundation models (VFMs) offer transferable visual representations, but their training-free performance on mining imagery and the benefits of synthetic-only adaptation remain insufficiently established. This study evaluates training-free VFM pipelines and investigates whether synthetic supervision improves their transfer to real-world crack imagery. SynthMineCrack is introduced as a diverse open-pit mining crack dataset, supporting model training and hyperparameter selection without real imagery or manual annotations. SAM 3 is adapted using low-rank adaptation (LoRA), layer normalization (LN) tuning, and adapters, individually and jointly, and compared with task-specific models trained on the same synthetic data. Evaluation on both in- and out-of-domain datasets shows that SAM 3 achieves the strongest training-free performance, while joint adaptation of all three operators improves intersection-over-union (IoU) on the mining evaluation dataset by 18.25 percentage points (pp). This configuration achieves a mean IoU of 69.39% across the five datasets, exceeding SegFormer, the strongest task-specific model, by 10.77 pp. Diagnostic analyses associate these gains with more complete crack recovery while the pretrained visual representation remains largely preserved. This configuration updates only 0.326% of the instantiated model’s parameter count and incurs little additional inference overhead. Together, these findings support synthetic-only adaptation as a practical means of improving SAM 3 for real-world crack segmentation while retaining inference requirements close to those of the base model.
No takes yet. Share an insight, caveat, or question.
Roux et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: