Taiwan SAR-based Ship and Weather Dataset (TSSWD) is the latest SAR ship dataset combined with weather data. Link to the buoy stations map. Link to the research paper.
Here's why you should use our dataset:
- Dataset build up with buoy weather data, i.e. every image has various weather data including wave height, wind speed ... etc. So not only the dataset can be use in Ship detection, it also got the potential to do sea weather prediction through SAR images!
- We only use Sentinel-1 as our image source, no mixed-source satellite, which is more common practice in Remote Sensing Field.
- The ship size in our dataset is smaller than previous dataset, which means it's more challenging 😄
We now only provide Mask-RCNN-X152 and YOLOV5 model's weight due to the storage limitation.
If you need other model weight, please contact us.
Satellite | Mode | Images | Ships | Size | Resolution |
---|---|---|---|---|---|
Sentinel-1 | IW | 2904 | 8255 | 800x800 | 10 m |
Annotation Type | Training | Validation | Testing | Weights | Dataset |
COCO / VOC / YOLOV4 / YOLOV5 | 2324 | 290 | 290 | Click Me | Click Me |
The Object Detection / Instance Segmentation Models are list bellow. Bold represent Instance Segmentation.
Training and testing params set in log file.
Mask R-CNN use Detectron2 platform.
YOLOV5 and CenterMask use their official github repo.
All model trained by GTX 1060 use the ppdet platform
Models | GPU | Notes |
---|---|---|
Mask R-CNN, YOLOV5, CenterMask | Nvidia Tesla V100 32G Nvida RTX 2080Ti 12G |
Provide by NTUCE-NCREE AI Research Center |
Rest of them | Nvidia GTX 1060 6G | Home PC |
Model | Bbox mAP | Bbox mAP50 | Seg mAP | Seg mAP50 | log |
---|---|---|---|---|---|
Mask R-CNN-r101 | 49.475 | 90.672 | 43.753 | 90.581 | log |
Mask R-CNN-x101 | 48.732 | 90.835 | 45.247 | 92.047 | log |
Mask R-CNN-x152 | 51.211 | 93.061 | 46.512 | 93.04 | log |
CenterMask-V99 | 47.024 | 86.21 | 43.726 | 87.736 | log |
CenterMask-V57 | 47.168 | 85.929 | 43.468 | 86.861 | log |
YOLO V5-l | 54.8 | 94.9 | x | x | log |
YOLO V5-l6 | 55.4 | 93.7 | x | x | log |
YOLO V5-x | 55.5 | 94.8 | x | x | log |
YOLO V5-x6 | 52.9 | 95.1 | x | x | log |
PPYOLO | 49 | 91.9 | x | x | log |
HRNET | 49 | 91 | x | x | log |
TOOD | 42.9 | 79.3 | x | x | log |
Cascade R-CNN | 50 | 90.5 | x | x | log |
Sparse R-CNN | 44.9 | 86.3 | x | x | log |
Res2Net | 49.6 | 91.1 | x | x | log |
FCOS | 43.6 | 80.3 | x | x | log |
Distributed under the MIT License. See LICENSE
for more information.
If our work is helpful to your research, use this bibtex to cite this repository:
@misc{TSSWD,
title={TSSWD: Taiwan SAR-based Ship Detection and Weather Dataset},
author={Shang-Fong, Yang and YaLun, Tsai},
year={2023},
publisher={Github},
journal={GitHub repository},
howpublished={\url{https://github.com/GMfatcat/TSSWD}}
}
use this citation if the article is published (not yet):
@article{TSSWD,
title={TSSWD: Taiwan SAR-based Ship Detection and Weather Dataset},
author={Shang-Fong, Yang and YaLun, Tsai},
keywords={SAR,Ship Detection,Object Detection,Instance Segmentation,Buoy}
year={2023},
publisher={},
journal={},
volume={},
number={},
pages={},
}
Email - r10521801@ntu.edu.tw
Email - yalunstsai@ntu.edu.tw
LAB Website: EORSLAB