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

Paper| Extended Version | Dataset on HF|Dataset on DesignSafe

Overview of EIDSeg

The code in this repository implements the data preparation, model training, and evaluation protocols as described in paper EIDSeg: A Pixel-Level Semantic Segmentation Dataset for Post-Earthquake Damage Assessment from Social Media Images in AAAI-AISI 2026. In this repo, we unified training pipeline for the EIDSeg dataset (post-earthquake infrastructure damage segmentation) supporting:

• Hugging Face models: Mask2Former, OneFormer, SegFormer, BEiT, EoMT
• DeepLabV3+ from the VainF repo)

📦 Repository Layout

.
├─ src/
│  ├─ datasets.py            # Dataset Function
│  ├─ models.py              # model factory (HF + DeepLabV3+)
│  ├─ eval_EIDSeg.py         # CLI evaluation entry
│  └─ train_EIDSeg.py        # CLI training entry point (all models)
├─ data/                     # your dataset lives here (see “Data Layout” below)
└─ DeepLabV3Plus-Pytorch/    # external repo (optional; required for DeepLabV3+)

🧰 Installation

Follow the steps below to set up the environment and run the EIDSeg experiments.

1. Clone the Repository

git clone https://github.com/HUILIHUANG413/EIDSeg.git
cd EIDSeg

2. Create a Conda Environment

It is recommended to use Python 3.10+.

conda create -n eidseg python=3.12 -y
conda activate eidseg

3. Install Dependencies

All required packages are listed in requirements.txt.

pip install -r requirements.txt

Optional (for DeepLabV3+):

git clone https://github.com/VainF/DeepLabV3Plus-Pytorch.git
# (optional but convenient)
pip install -e DeepLabV3Plus-Pytorch

📁 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)

▶️ Quick Start

A) Train a Hugging Face model (e.g., EoMT)

python train_EIDSeg.py \
  --train-xml   ../data/train/train.xml \
  --train-imgdir ../data/train/images \
  --val-xml     ../data/val/val.xml \
  --val-imgdir   ../data/val/images \
  --model-name   tue-mps/cityscapes_semantic_eomt_large_1024 \
  --image-size   1024 1024 \
  --epochs 50 --batch-size 1 --lr 1e-5 \
  --augment \
  --run-dir runs/eomt/Plan5_1024_1024

Supported HF models in paper:

• "nvidia/segformer-b5-finetuned-cityscapes-1024-1024"
• "microsoft/beit-base-finetuned-ade-640-640"
• "microsoft/beit-large-finetuned-ade-640-640"
• "shi-labs/oneformer_cityscapes_swin_large"
• "tue-mps/cityscapes_semantic_eomt_large_1024"
• "facebook/mask2former-swin-small-cityscapes-semantic"
• "facebook/mask2former-swin-large-cityscapes-semantic"

B) Train DeepLabV3+ (VainF repo, optional)

1. Make sure the external repo exists at ./DeepLabV3Plus-Pytorch/.
2. (Optional) Download a Cityscapes checkpoint (e.g., best_deeplabv3plus_resnet101_cityscapes_os16.pth.tar) and note the path.

python train_EIDSeg.py \
  --train-xml   ../data/train/train.xml \
  --train-imgdir ./data/train/images \
  --val-xml     ../data/val/val.xml \
  --val-imgdir   ../data/val/images \
  --model-name   deeplabv3plus-resnet101 \
  --deeplab-backbone resnet101 \
  --deeplab-os 16 \
  --deeplab-cityscapes-ckpt DeepLabV3Plus-Pytorch/weights/best_deeplabv3plus_resnet101_cityscapes_os16.pth.tar \
  --image-size 512 512 \
  --epochs 150 --batch-size 2 --lr 1e-5 \
  --augment \
  --run-dir runs/deeplabv3plus/Plan5_512_512

🗂️ Outputs

Inside --run-dir:

run_dir/
├─ training_log.txt
├─ hyperparameters.json
├─ loss_plot.png
└─ weights/
   ├─ best_model.pth
   ├─ final_model.pth
   └─ checkpoint.pth     # contains epoch, optimizer/scheduler states, curves, best_miou

Here are clean, properly formatted Markdown versions of both tables for your README.

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

🙌 Acknowledgments

The research described herein was supported in part by the School of Computing Instruction and the Elizabeth and Bill Higginbotham Professorship at Georgia Tech. This support is gratefully acknowledged. Support for the undergraduate students participating in the project was provided by the US National Science Foundation through the Geotechnical Engineering Program under Grant No. CMMI-1826118. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of NSF.

License & Credit

• DeepLabV3+ implementation courtesy of VainF: https://github.com/VainF/DeepLabV3Plus-Pytorch
• Our Classification dataset EID: https://github.com/HUILIHUANG413/Earthquake-Infrastracture-Damage

Contact

Huili Huang - huilihuang1997@gmail.com; hhuang413@gatech.edu

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}
}

@article{huang2025enhancing_fidelity,
  title   = {Enhancing the Fidelity of Social Media Image Datasets in Earthquake Damage Assessment},
  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 = {Earthquake Spectra},
  volume  = {41},
  number  = {3},
  pages   = {2616-2635},
  year    = {2025},
  doi     = {10.1177/87552930251335649},
  url     = {https://doi.org/10.1177/87552930251335649}
}