This repository contains the code and models from the paper V. Stojnic, V. Risojevic, "Self-Supervised Learning of Remote Sensing Scene Representations Using Contrastive Multiview Coding", In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2021.

This code is buit on top of the official implementation of the Contrastive Multiview Coding by Y. Tian.

To run the code please create Anaconda environment using dependancies defined in cmc_rssr.yml.

Pre-trained CMC models trained in our papaer are available here.

Trained linear classifiers are available here.

Finetuned CMC models for our downstream tasks are available here.

We don't provide pre-trained CMC models on the whole ImageNet as they are available on the official CMC implementation.

Pre-trained supervised models are used from torchvision model zoo.

Given linear classifier and finetuned models can give slightly different results from the ones in the paper, but main conclusions still hold for these models too.

To run the training of the CMC model use the following python script.

python train_CMC.py --data_folder PATH_TO_FOLDER_CONTAINING_THE_IMAGES_OR_LMDB_FOLDER_FOR_MS_IMAGES --image_list PATH_TO_FILE_WITH_IMAGE_LIST --model MODEL_NAME --model_path PATH_TO_DIRECTORY_TO_SAVE_THE_MODEL --batch_size BATCH_SIZE [--multispectral --pca --ben]

Use the --multispectral flag if you are training on multispectral data and if you want PCA based views use the --pca flag.

Use the --resize_image_aug flag if you are training on BigEarthNet RGB dataset to resize original images to 256x256 pixels.

Other command line arguments are also possible. Please read the script and parser in util.py.

To run the feature extraction use the following python script.

python extract_features.py --data_folder PATH_TO_FOLDER_CONTAINING_THE_IMAGES_OR_LMDB_FOLDER_FOR_MS_IMAGES --image_list PATH_TO_FILE_WITH_IMAGE_LIST --model MODEL_NAME --resume PATH_TO_TRAINED_CMC_MODEL --features_path PATH_FOR_FEATURE_SAVING [--multilabel_targets PATH_TO_JSON_FILE_WITH_MULTILABEL_TARGETS --multispectral --pca --multispectral_dataset DATASET_NAME]

Use --multilabel_targets if you are extracting features for a multilabel RGB dataset.

Use the --multispectral flag if you are extracting features on multispectral data and if you want PCA based views use the --pca flag. For multispectral datasets it is necessary to supply --multispectral_dataset DATASET_NAME dataset name BigEarthNet or So2Sat as these datasets need different preprocessing methods.

To run the linear classifier use the following script.

python linear_classifier.py --train_data_path PATH_TO_TRAIN_FEATURES --val_data_path PATH_TO_VAL_FEATURES --batch_size BATCH_SIZE --epochs NUM_OF_EPOCHS --learning_rate LEARNING_RATE --lr_decay_epochs DECAY_EPOCHS --lr_decay_rate DECAY_RATE --weight_decay WEIGHT_DECAY --resume PATH_TO_MODEL_SAVE_OR_RESUME [--evaluate]

Use --evaluate flag for evaluation of the trained linear classifier.

To run the finetuining use the following python script.

python finetuning.py --data_folder PATH_TO_FOLDER_CONTAINING_THE_IMAGES_OR_LMDB_FOLDER_FOR_MS_IMAGES --train_image_list PATH_TO_FILE_WITH_TRAIN_IMAGE_LIST --val_image_list PATH_TO_FILE_WITH_VAL_IMAGE_LIST --model MODEL_NAME --model_path PATH_TO_TRAINED_CMC_MODEL --weight_decay WEIGHT_DECAY --epochs NUM_OF_EPOCHS --batch_size BATCH_SIZE --lr_decay_epochs DECAY_EPOCHS --lr_decay_rate DECAY_RATE [--save_path PATH_TO_MODEL_SAVE --learning_rate LEARNING_RATE --resume PATH_TO_MODEL_RESUME --multilabel_targets PATH_TO_JSON_FILE_WITH_MULTILABEL_TARGETS --multispectral --pca --multispectral_dataset DATASET_NAME --evaluate]

Use --multilabel_targets if you are extracting features for a multilabel RGB dataset.

Use the --multispectral flag if you are extracting features on multispectral data and if you want PCA based views use the --pca flag. For multispectral datasets it is necessary to supply --multispectral_dataset DATASET_NAME dataset name BigEarthNet or So2Sat as these datasets need different preprocessing methods.

Use --evaluate flag for evaluation of the trained linear classifier.

Additional python scripts that can be used to create different dataset formats suitable for this implementation are available in helper directory.

Dataset splits used in this paper can be found in data_splits directory.

Multilabel targets for every image in BigEarthNet and MLRSNet datasets used in this paper can be found in JSON files in image_target_mapping directory.

@InProceedings{Stojnic_2021_CVPR_Workshops,
    author = {Stojnic, Vladan and Risojevic, Vladimir},
    title = {Self-Supervised Learning of Remote Sensing Scene Representations Using Contrastive Multiview Coding},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month = {June},
    year = {2021}
}