EIDSeg: A Pixel-Level Semantic Segmentation Dataset for Post-Earthquake Damage Assessment from Social Media Images

EIDSeg is a large-scale post-earthquake infrastructure damage segmentation dataset collected from nine major earthquakes (2008–2023).
This repository provides the raw dataset in CVAT XML format, along with the corresponding images organized by split.
It is intended to be used together with our official codebase for parsing XML annotations and training segmentation models. See our github repo for more detail.

📥 Downloading the Dataset

You can download the dataset using any of the methods below.

🔹 1. Using huggingface_hub

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="HuiliHuang/EIDSeg",
    repo_type="dataset",
    local_dir="EIDSeg"
)

📁 Data Layout

The code expects CVAT-style XML annotations and images arranged like:

data/
├── train/
│   ├── train.xml
│   └── images/
│       └── default/
│           ├── 0001.jpg
│           ├── 0002.png
│           └── ...
└── val/
    ├── val.xml
    └── images/
        └── default/
            ├── 1001.jpg
            └── ...

Annotations (CVAT XML):

<annotations>
  <image name="0001.jpg" ...>
    <polygon label="D_Building" points="x1,y1;x2,y2;..." />
    <polygon label="UD_Road"    points="..." />
    ...
  </image>
</annotations>

Class mapping (6 classes):

0: UD_Building
1: D_Building
2: Debris
3: UD_Road
4: D_Road
5: void (Background / Undesignated)

Benchmark Results

Semantic Segmentation Benchmark of EIDSeg

Model	Backbone	Pre-train	Input	mIoU (%)	FWIoU (%)	PA (%)	FLOPs (G)	Params (M)
DeepLabV3+	ResNet-101	Cityscapes	512²	67.1	68.2	86.0	79.29	58.76
SegFormer	MiT-B5	Cityscapes	512²	74.4	75.2	86.9	110.16	84.60
Mask2Former-S	Swin-S	Cityscapes	512²	76.1	77.1	87.7	93.21	81.42
Mask2Former-L	Swin-L	Cityscapes	512²	77.4	78.4	88.7	250.54	215.45
BEiT-B	ViT-B	ADE20K	640²	78.7	79.6	89.8	1823.53	441.09
BEiT-L	ViT-L	ADE20K	640²	79.0	79.8	89.9	3182.73	311.62
OneFormer	Swin-L	Cityscapes	512²	79.8	80.2	89.8	1042.14	218.77
EoMT	ViT-L	Cityscapes	1024²	80.8	80.9	90.3	1341.85	319.02

Class-wise IoU and mIoU (%) for each model on EIDSeg

Model	UD_Building	D_Building	Debris	UD_Road	D_Road	mIoU (%)
DeepLabV3+	34.5	65.4	77.3	75.7	73.7	67.1
SegFormer	54.9	73.5	82.3	79.9	79.4	74.4
Mask2Former-S	58.9	76.7	83.8	80.2	80.1	76.1
Mask2Former-L	63.5	76.9	84.9	82.0	80.9	77.4
BEiT-B	66.0	76.7	85.1	82.3	78.7	78.7
BEiT-L	66.4	77.9	85.1	82.6	78.7	79.0
OneFormer	68.7	79.7	85.0	84.1	79.9	79.8
EoMT	70.1	80.0	84.6	82.0	87.3	80.8

Contact

Huili Huang - [email protected]; [email protected]

Please ⭐ if you find it useful so that I find the motivation to keep improving this. Thanks

Citation

If you find this work or the EIDSeg dataset useful in your research, please consider citing our paper. Your citation helps support and encourage future development of this project.

@article{huang2025eidseg,
  title   = {EIDSeg: Post-Earthquake Infrastructure Damage Segmentation Dataset},
  author  = {Huili Huang and Chengeng Liu and Danrong Zhang and Shail Patel and Anastasiya Masalava and Sagar Sadak and Parisa Babolhavaeji and Weihong Low and Max Mahdi Roozbahani and J.~David Frost},
  journal = {arXiv preprint arXiv:https://arxiv.org/abs/2511.06456},
  year    = {2025}
}