Underwater instance segmentation models are typically trained and evaluated on a single dataset, leaving cross-dataset generalization unstudied. This project measures the cross-dataset domain gap between TrashCan (7,212 deep-sea ROV images, 22 classes) and SeaClear (8,610 shallow-water images, 40 classes) using Mask R-CNN with a COCO-pretrained ResNet-50 FPN backbone. Models trained on one dataset fail catastrophically on the other despite overlapping object categories. We show this gap is visual, not semantic: silhouette analysis of backbone features reveals strong dataset clustering even at the per-class level. To mitigate the gap, we pool both datasets into common category spaces via a generic coarsening hierarchy and show that a single pooled model recovers or exceeds in-domain performance on both test sets. We also re-split TrashCan using frame-chunking to fix data leakage in the original split. All results are explorable via an interactive Streamlit dashboard.