whuhxb / caco

official code for CVPR 2023 paper 'Change-Aware Sampling and Contrastive Learning for Satellite Images'

Change-Aware Sampling and Contrastive Learning for Satellite Images

This is the implementation for CVPR 2023 paper Change-Aware Sampling and Contrastive Learning for Satellite Images.

Authors: Utkarsh Mall, Bharath Hariharan, Kavita Bala

Installation

This codebase is built with and tested with python3.9, and torch 1.7.1. We highly recommend creating environment using anaconda.

conda create --name cacoenv python=3.9.13 -y After activating the environment:

conda activate cacoenv

Install the following packages:

conda install cudatoolkit=11.0 -y

pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 -f https://download.pytorch.org/whl/torch_stable.html

pip install matplotlib pytorch-lightning==1.1.8 pytorch-lightning-bolts==0.3.0 scikit-learn rasterio lmdb pandas

pip installjupyter progressbar opencv-python

Optionally also install

pip install wandb gym

If running into error due to some library mismatch, please refer to requirements.txt for the correct versions.

Quick inference

Use the notebook for quick inference of on our pretrained model.

cd notebooks python -m jupyter notebook --ip 0.0.0.0 --no-browser

The notebook also provides instructions on how to download pretrained models.

The pretrained models can be found at this link

Training

In order to retrain the model, we need to download the CACo data and then run training. A sample training procedure is provided in sample_training.sh

bash sample_training.sh

The data can be found at this link Refer to the sample_training.sh for downloading and placing data.

To train a resnet-18 model with 10k data and caco loss run the following

python3 main_pretrain.py --data_dir ../data/clean_10k_geography/ --base_encoder resnet18 --batch_size 256 --data_mode caco --max_epochs 1000 --schedule 600 800 -d gereric_description

A few more things will be released soon.

• Code to download custom dataset using CACo sampling.
• Setup for evaluating on other datasets, BigEarthNet, DynmamicEarthNet, OSCD, and Change Events.
• 1m training set.

Acknowledgements

This repository follows the stucture from SeCo and uses same versions of libraries used.