Pre-computed embeddings for the paper "Label-efficient underwater image classification with logistic regression on frozen foundation model embeddings"

These embeddings were generated during the evaluation of DINOv3 (facebook/dinov3-vitb16-pretrain-lvd1689m) on the AQUA20 benchmark as foundation for linear separability of marine image classification in the paper 'Label-efficient underwater image classification with logistic regression on frozen foundation model embeddings'. Contained in this repository are: CLS tokens and L2-averaged Patch Mean tokens for the resolutions of 224, 384 and 512 for AQUA20 CLS tokens for the sea animals dataset used in the paper CLS tokens are stored unnormalized; Patch Mean tokens are L2-normalized. The downstream scripts normalize CLS tokens before use. Each .npz file contains two arrays: cls_tokens and patch_mean. Filenames: {IMAGE_SIZE}_official_train_embeddings.npz {IMAGE_SIZE}_official_test_embeddings.npz To use the embeddings: Place the files for your chosen resolution into {IMAGE_SIZE}_label_efficient_classification/cache/embeddings_npy/. Create the directory if it doesn't exist. Example for IMAGE_SIZE = 384: 384_label_efficient_classification/ cache/ embeddings_npy/ 384_official_train_embeddings.npz 384_official_test_embeddings.npz Then run the experiment notebooks from the companion repository: https://github.com/MiaSpugnetta/label-efficient-underwater-image-clf.