Aayush Dhakal*, Adeel Ahmad Subash Khanal, Srikumar Sastry, Hannah Kerner Nathan Jacobs

The repository is the official implementation of Sat2Cap [CVPRW, EarthVision 2024, Best Paper Award]. Sat2Cap model solves the mapping problem in a zero-shot approach. Instead of predicting pre-defined attributes for a satellite image, Sat2Cap attempts to learn the text associated with a given location.

You can use the run_geo.sh script to train the Sat2Cap model. All the necessary hyperparameters can be set in the bash script.

Once you have the trained model use the generate_map_embedding.py file under evaluations to generate Sat2Cap embeddings for all images of interest. Use merge_embeddings.py to add location and temporal input to the generated embeddings. Finally, the get_similarity.py file generates similarity values for a given prompt. These similarity values can then be used to create zero-shot maps.

@inproceedings{dhakal2024sat2cap,
  title={Sat2cap: Mapping fine-grained textual descriptions from satellite images},
  author={Dhakal, Aayush and Ahmad, Adeel and Khanal, Subash and Sastry, Srikumar and Kerner, Hannah and Jacobs, Nathan},
  booktitle={IEEE/ISPRS Workshop: Large Scale Computer Vision for Remote Sensing (EARTHVISION)},
  pages={533--542},
  year={2024}
}

Check out our lab website for other interesting works on geospatial understanding and mapping:

• Multi-Modal Vision Research Lab (MVRL) - Link
• Related Works from MVRL - Link