Semi-supervised learning on frozen DINOv3 embeddings achieves near-expert underwater species classification with less than 5% labeled data

Researchers use DINOv3 ViT-B frozen embeddings with self-training to classify 20 marine species on the AQUA20 benchmark without fine-tuning

Self-training approach closes most of the performance gap to fully supervised ConvNeXt baseline using fewer than 5% of labeled training data

Method addresses the annotation bottleneck in underwater imagery classification and shows strong generalization across different environmental conditions

At full supervision, performance gap narrows to just a few percentage points, with some species exceeding the supervised baseline accuracy