Figure 4 – Total loss curves for the four training datasets. Figure 5 – Average Precision (AP) by IoU threshold, object size, and validation dataset. IoU values of 50, 75, and 50-95 correspond to the IoU thresholds. The first three bar groups include all object sizes. 3.2.2 Validation dataset The COCO evaluator was used for the validation dataset. The validation metric is Average Precision (AP), computed over varying IoU thresholds and reported according to object COCO scale. Figure 5 shows four colored bars for each validation dataset, grouped by IoU threshold and object size (s-small, m-medium, and l-large). The XPL-10X dataset exhibits the highest values, whereas PPL-10X yields the lowest performance. No substantial differences are observed between the datasets with different augmentation levels (3X and 10X). Notably, larger objects achieve higher precision values, with the highest AP reaching 35% for the XPL-10X dataset and the lowest AP reaching 19.9% for the PPL-10X dataset. 3.2.3 Evaluation dataset Model inferences were analyzed using an evaluation dataset of 20 images. Figure 6 presents examples of segmentations generated by the four models. The first row shows the groundtruth mask and segmentations of the images. On the left, Figure 6 exhibits a gabbro with regularto-good segmentation across all models. However, mineral overlap and misclassification are observed: in the 3X and 10X models under PPL conditions, quartz is incorrectly identified, and pyroxenes are misclassified as olivine. On the right, it shows a quartz diorite with low-quality segmentation, where few minerals are segmented. Overlapping labels on the same object are also observed, along with multiple identification errors, particularly for biotite and amphibole.
RkJQdWJsaXNoZXIy MTM0Mzk2