Role of Coastal Characteristics and Environmental Context in Segmentation Performance

Determine whether characteristics specific to coastal semantic classes or coastal environmental context contribute to their weaker vision-language segmentation performance relative to conventional classes.

Background

The paper evaluates seven vision-LLMs on a densely annotated coastal dataset from Oahu, Hawaii. Although coastal classes generally receive lower text-to-mask segmentation scores than landscape and conventional classes, the authors examine whether this difference can be explained by class size. Spearman correlations between class spatial extent and segmentation performance are weak, inconsistent in sign and magnitude, and statistically insignificant across all evaluated models. The authors therefore leave unresolved whether properties unique to coastal classes or the surrounding environmental context are responsible for their weaker segmentation performance.