This repo contains the code for the paper "Extending Scene-to-Patch Models: Multi-Resolution Multiple Instance Learning for Earth Observation".
We use Bonfire for the backend MIL functionality.

Below we break down each of the directories in this repo:

Configuration for model parameters.

Contains the trained model files. Five repeats per configuration.

Interpretability outputs and other figures that are used in the paper

Raw results for our experiments: scene-level RSME and MAE, patch-level mIoU, and pixel-level mIoU.

Contains our executable scripts. These are the entry points to our experiments, and should be run from the root of the repo.

Contains our project-specific code. This includes the dataset, high-level model implementations (low-level model code is implemented in Bonfire), and interpretability studies.