01
Evaluating Vision-Language Models Underwater: A Benchmark Study Toward Domain-Specific VLM Design
Aditya Tejwani
Abstract
Vision-language models (VLMs) such as CLIP [1] perform well on land-based imagery but struggle in underwater environments due to color distortion, turbidity, and noise [2]. A benchmark is performed to assess CLIP ViT-B/32, CLIP ViTL/14, and DINOv2 performance on the AQUA20 dataset [3] of 6,559 testing images across 20 marine species. CLIP achieves 42.99% and 54.08% accuracy, while DINOv2 [4] reaches 92.45% using only visual features. Two main limitations are identified: semantic ambiguity between labels and degradation of visual features under poor imaging conditions. These results suggest that language-supervised models are poorly suited for underwater recognition without domain-specific training.
Download paper ↓