A streamlined Approach to Multimodal Few-Shot Class Incremental Learning for Fine-Grained Datasets (CLIP-M3)

Preprint

We release our three introduced fine-grained datasets below: (Code coming soon)

FGVCAircraft

gdown https://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/archives/fgvc-aircraft-2013b.tar.gz;
tar -xzf fgvc-aircraft-2013b.tar.gz;
    

StanfordCars

pip install kaggle;
kaggle datasets download -d jessicali9530/stanford-cars-dataset -p [FOLDER];
unzip stanford-cars-dataset.zip;
    

iNF200

Our dataset is comprised of the 200 first Fungi classes from iNaturalist. Train and test sets available here. Since the labels for the test set are not available, we used the validation set for evaluation.

'iNF200_dict.pkl' contains:

• Train set meta data
• Test set meta data
• labels to classnames
• classname to label