M2SODAI: Multi-Modal Ship and Floating Matter Detection Image Dataset With RGB and Hyperspectral Image Sensors
Jonggyu Jang, Sangwoo Oh, Youjin Kim, Dongmin Seo, Youngchol Choi, Hyun Jong Yang
Conference on Neural Information Processing (NeurIPS) 2023
Paper link: proceedings.neurips.cc
![image](https://private-user-images.githubusercontent.com/88477912/313915840-fb42288e-1662-469d-a72d-6e3ed46fc394.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.u6-54hPtmB6unce6XC4Q7qhOjBnXU8uwomOxl9VqY6g)
- 2024/05: Splitted Dataset (with corresponding trained weights) Will Be Provided.
- 2024/04: OneDrive Link. For people who cannot access Google Drive, I will prepare OneDrive Link. Due to my storage limit, only preprocessed data will be accessible.
- 2024/04: Add Image Registration History. Some people are interested in our image registration procedure (
rgb
andhsi
data). Now, the overall procedure is available here. It is very time-consuming and labor-intensive job(I worked all night for about 4-5 days doing only this registration procedure from scratch.). So, If you mail me before starting, I can give you tip/advice I learned doing this.- Due to the policy of the research institute, we cannot provide raw data of RGB/HSI sensors. If you have plan for building another dataset using image registration, I can help you for free.
- 2024/04: Correct this repo. We correct this repo. Thanks to
Another-0
andXiaodian Zhang
. We sincerely apologize for being later than the promised time, and we truly appreciate your continued interest in our work. The previous code is wrong; hence, please use the current dataset/code. - 2024/02: Recognize mistakes in our repo. Now, we recognize that some files are missing/wrong when we upload the source code. We will fix this issue by Uploading preprocessing code, as well as processed data and Replica of the trained weights.
- Anaconda
conda create -n Maritime python=3.7
- Pytorch
pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio===0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
- Other extra dependencies
pip install spectral matplotlib scikit-image fire openmim
pip install mmcv-full==1.3.17 -f https://download.openmmlab.com/mmcv/dist/cu110/torch1.7/index.html
# for instaboost
pip install instaboostfast
# for panoptic segmentation
pip install git+https://github.com/cocodataset/panopticapi.git
# for LVIS dataset
pip install git+https://github.com/lvis-dataset/lvis-api.git
# for albumentations
pip install -r requirements/albu.txt
pip install -r requirements/build.txt
pip install -v -e . # or "python setup.py develop"
Download
To encourage related research, we will provide datasets upon your request. Please fill out this form. We ask for your information only to ensure that the dataset is used for non-commercial purposes. I will give you access authority within 1~2 days. If not, please remind me again.
Do not provide this dataset to third parties or post it publicly anywhere.
Download the processed data via the above link.
.M2SODAI
├── configs
├── data_tools
├── ...
├── data
│ ├── train
│ ├── test
│ ├── val
│ ├── train_coco
│ ├── test_coco
│ ├── val_coco
│ ├── pca_mean_std.mat
│ ├── model.pkl
│ ├── ...
│ └── label.txt
- (Skip) run data_tools/lableme2coco.py
python data_tools/labelme2coco.py data/train data/train_coco --label data/label.txt
python data_tools/labelme2coco.py data/test data/test_coco --label data/label.txt
python data_tools/labelme2coco.py data/val data/val_coco --label data/label.txt
- (Skip) make
pca_mean_std.mat
and IPCA.
python tools/IPCA_data.py
- Training
python tools/train.py {config_file}
ex) R50 configuration file is in configs/faster_rcnn/faster_rcnn_r50_rgb_hsi.py
ex) python tools/train.py configs/faster_rcnn_faster_rcnn_r50_rgb_hsi.py
- Evaluation
python tools/test.py {config_file} {ckpt_file} --eval bbox
- R-50-RGB-HSI
- mAP: 43.7
- Model (GDrive)
- Model (OneDrive)
- R-50-RGB
- mAP: 37.7
- Model (GDrive)
- Model (OneDrive)
- R-50-HSI (The mAP is enhanced while correcting our code)
- mAP: 13.2
- Model (GDrive)
- Model (OneDrive)
- One Drive Dataset/Trained Weights Link.
- Thank you for your suggestion,
Yangfan Li
.
- Thank you for your suggestion,
- (In processing) Split images by smaller ones (make 11GB GPU work). Due date: end of April.
One challenge is that most of the target objects are cut out doing this. Is there anyone who has an idea to solve this?- Thank you for your contribution,
Xiaodian Zhang
.
- This work is forked from MMdetection Repository https://github.com/open-mmlab/mmdetection
- A useful instruction of Faster R-CNN: https://www.lablab.top/post/how-does-faster-r-cnn-work-part-ii/
@inproceedings{NEURIPS2023_a8757b88,
author = {Jang, Jonggyu and Oh, Sangwoo and Kim, Youjin and Seo, Dongmin and Choi, Youngchol and Yang, Hyun Jong},
booktitle = {Advances in Neural Information Processing Systems},
editor = {A. Oh and T. Neumann and A. Globerson and K. Saenko and M. Hardt and S. Levine},
pages = {53831--53843},
publisher = {Curran Associates, Inc.},
title = {M\^{}\lbrace 2\rbrace SODAI: Multi-Modal Maritime Object Detection Dataset With RGB and Hyperspectral Image Sensors},
url = {https://proceedings.neurips.cc/paper_files/paper/2023/file/a8757b889350a3782b384a3ec0dfbae9-Paper-Datasets_and_Benchmarks.pdf},
volume = {36},
year = {2023},