Use a Vision Transformer pre-trained through the DINO protocol to extract features from images. Reduce dimensionality and visualize using t-SNE and cluster the images with HDBSCAN.
Can be run as-is or deployed using a docker image. Developed to cluster large underwater datasets captured by marine robots.
The images can be overlaid on their embeddings to better understand the latent space.
- Build the docker image
docker build -t name_of_image:tag . - In
run_docker.shChange the file path of the mounted volume to your directory containing the images to be clustered. - Run the container:
./run_docker.sh