Evaluating SAM2 Automatic Mask Generation for Petrographic Thin Sections: Effects of Parameter Tuning

Abstract

Optical thin-section segmentation remains difficult because petrographic images combine fine textures, strong local variability, and often ambiguous phase boundaries. This study evaluates the effect of automatic mask-generator parameter tuning for a fixed pretrained SAM2 model in a fully automatic zero-shot setting. The experiments were conducted on six petrographic thin sections, with five slides used for cross-slide parameter evaluation and one additional slide reserved for a final hold-out comparison. Predictions were evaluated at the mask level using prediction-dependent references derived from fixed semantic phase maps, with IoU, Dice, purity, and a weighted score. Within the tested parameter space and Optuna budget, the best configuration reached an objective value of 0.3121 during parameter selection. On the held-out thin section, the corrected tuned configuration produced only small descriptive changes relative to the default configuration. Because the reference is constructed from each generated proposal and entirely missed reference regions are not directly penalized, the reported scores characterize automatic mask proposals under the specified evaluation procedure rather than complete phase coverage or mineral identification accuracy. Under these experimental conditions, parameter tuning produced only modest gains, while spatial agreement with the associated references remained limited.