Exploring the Underwater World Segmentation without Extra Training

The introduction of AquaOV255, a large-scale underwater segmentation dataset with 255 categories and over 20,000 images, represents a significant leap in marine biodiversity monitoring and ecological assessment. Existing models have largely been limited to terrestrial environments, creating a gap that AquaOV255 aims to fill. This dataset is complemented by the establishment of UOVSBench, the first underwater OV segmentation benchmark, which integrates AquaOV255 with other datasets for comprehensive evaluation. The Earth2Ocean framework further enhances this initiative by enabling effective segmentation in underwater domains without the need for additional training, leveraging terrestrial vision-language models. Extensive experiments on UOVSBench have demonstrated significant performance improvements with Earth2Ocean, underscoring its potential impact on marine research and conservation efforts.