OpenEarthMap Land Cover Mapping Few-Shot Challenge
Generalized Few-shot Semantic Segmentation

Challenge proposed by the Geoinformatics Team of RIKEN-AIP and co-organized
with the 3rd L3D-IVU Workshop @ CVPR 2024 Conference

We are excited to be seeing great ideas from across the globe working hard towards a better understanding of our environment. We look forward to having fun in our quest to obtain more accurate semantic segmentation maps for practical applications of few-shot learning in remote sensing image understanding. Let us come together to push the state-of-the-art generalized few-shot semantic segmentation (GFSS) learning methods to promote research in AI for social good.

Get involved! Check out the following links:

• Challenge Webpage https://cliffbb.github.io/OEM-Fewshot-Challenge/
• L3D-IVU Workshop @ CVPR 2024 Conference https://sites.google.com/view/l3divu2024/overview
• Dataset Download https://zenodo.org/records/10828417
• Submission Portal https://codalab.lisn.upsaclay.fr/competitions/17568

Context

This repository contains the baseline model for the OpenEarthMap land cover mapping generalized few-shot semantic segmentation challenge. The motivation is to enable researchers to develop and benchmark learning methods for generalized few-shot semantic segmentation of high-resolution remote sensing imagery. The challenge is in two phases: the development phase is for developing and testing methods on a valset, and the evaluation phase is for tweaking and testing on a testset for final submission.

Dataset

This dataset extends the original 8 semantic classes of the OpenEarthmap benchmark dataset to 15 classes for 5-shot generalized few-shot semantic segmentation (GFSS) task with 4 novel classes and 7 base classes. It consists of only 408 samples from the original OpenEarthMap dataset. The 408 samples are also split into 258 as trainset, 50 as valset, and 100 as testset. The valset is used for the development phase of the challenge, and the testset is for the evaluation phase. Both sets consist of support_set and query_set for GFSS tasks. A detailed description of the dataset can be found here, where it can also be downloaded. Below are examples of novel classes in the support_set (first two columns), and base classes + novel classes in the query_set (last two columns).

Baseline

The PSPNet architecture with EfficientNet-B4 encoder from the Segmentation Models Pytorch GitHub repository is adopted as a baseline network. The network was pretrained using the trainset with the Catalyst library. Then, the state-of-the-art framework called distilled information maximization (DIaM) was adopted to perform the GFSS task. The code in this repository contains only the GFSS portion. As mentioned by the baseline authors, any pretrained model can be used with their framework. The code was adopted from here. To run the code on the valset, simply clone this repository and change your directory into the OEM-Fewshot-Challenge folder which contains the code files. Then from a terminal, run the test.sh script. as:

bash test.sh 

The results of the baseline model on the valset are presented below. To reproduce the results, download the pretrained models from here. Follow the instructions in the Usage section, then run the test.sh script as explained.

Phase	base mIoU	novel mIoU	Average base-novel mIoU	Weighted base mIoU	Weighted novel mIoU	Weighted-Sum base-novel mIoU
valset	29.48	03.18	16.33	11.79	1.91	13.70
testset	30.07	9.39	19.73	12.03	5.36	17.66

The weighted mIoUs are calculated using `0.4:0.6 => base:novel`. These weights are derived from the state-of-the-art results presented in the baseline paper.

Usage

The repository structure consists of a configuration file that can be found in config/; data splits for each set in data/; and all the codes for the GFSS task are in src/. The testing script test.sh is at the root of the repo. The docs folder contains only GitHub page files.

To use the baseline code, you first need to clone the repository and change your directory into the OEM-Fewshot-Challenge folder. Then follow the steps below:

1. Install all the requirements. Python 3.9 was used in our experiments. Install the list of packages in the requirements.txt file using pip install -r requirements.txt.
2. Download the dataset from here into a directory that you set in the config file oem.yaml
3. Download the pretrained weights from here into a directory that you set in the config file oem.yaml
4. In the oem.yaml you need to set only the paths for the dataset and the pretrained weights. The other settings need not be changed to reproduce the results.
5. Test the model by running the test.sh script as mentioned in the Baseline section. The script will use the support_set to adapt and predict the segmentation maps of the query_set. After running the script, the results are provided in a results folder which contains a .txt file of the IoUs and mIoUs, and a preds and targets folder for the predicted and the targets maps, respectively.

You can pretrained your model using the trainset and any simple training scheme of your choice. The baseline paper used the train_base.py script and base learner models of BAM (see the baseline paper for more info).

Citation

For any scientific publication using this data, the following paper should be cited:

@InProceedings{Xia_2023_WACV,
    author    = {Xia, Junshi and Yokoya, Naoto and Adriano, Bruno and Broni-Bediako, Clifford},
    title     = {OpenEarthMap: A Benchmark Dataset for Global High-Resolution Land Cover Mapping},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2023},
    pages     = {6254-6264}
}

Acknowledgements

We are most grateful to the authors of DIaM, Semantic Segmentation PyTorch, and Catalyst from which the baseline code is built on.