Classifying rock thin sections usually relies on petrographers to label each slide, and those labels are used to meta-train few-shot classifiers. That identification step is expensive and often unavailable for new collections. The arXiv preprint proposes a cross-polarization pairing method to reduce this dependency, aiming for label-free few-shot classification.
The idea is to exploit paired images taken under different polarization conditions, letting the model learn from mineral optical properties without requiring an expert to annotate every new sample. This could make classifiers easier to deploy on collections that lack prior petrographic labeling.
The abstract does not include experimental results, so the method's accuracy relative to conventional labeled meta-training is not yet quantified. The authors frame cross-polarization pairing as a step toward removing the labeling bottleneck, but further evaluation is needed to confirm its practical benefit.