File size: 3,375 Bytes
0e4b1c5
cfa4b76
 
 
 
 
 
 
 
 
 
0e4b1c5
cfa4b76
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f61e88c
 
 
 
 
 
 
 
cfa4b76
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
---
license: cc-by-nc-4.0
library_name: pytorch
tags:
  - remote-sensing
  - building-damage-assessment
  - post-disaster
  - damage-typology
  - dinov3
  - disaster-response
pipeline_tag: image-segmentation
---

# Damage-TriageFormer (trained checkpoint)

Trained weights for **Damage-TriageFormer**, a foundation-model framework for
decision-relevant building damage typology from single post-event imagery. Given
a post-event RGB tile and building instance masks (footprints), the model assigns
each building one of five damage-typology classes:

| Class | Name |
|------:|------|
| 0 | Undamaged |
| 1 | Partial Roof Damage |
| 2 | Total Roof Damage |
| 3 | Partial Structural Damage |
| 4 | Total Structural Collapse |

## Files

- `best_model.pth` — the reported checkpoint, selected by validation macro-F1 in the footprint-conditioned setting.
- `config.json` — the exact training configuration used to produce it.

## Architecture

DINOv3 ViT-L/16 backbone (last 4 blocks fine-tuned) → Simple Feature Pyramid
(target stride 8, 128×128 feature map) → mask-pooled instance features → a
two-stage gated damage head (any-damage gate → 4-way damaged-class leaf), with an
auxiliary severity-regression head used during training.

## Training

30 epochs, AdamW (LR 5e-5 heads / 5e-6 backbone), effective batch 32 (2×A100 40GB,
DDP), label smoothing 0.1, EMA (decay 0.9995), long-tailed logit adjustment
(τ=1.0), inverse-square-root class weighting on the leaf head. Trained on
**DamageTriage-Bench** (Hurricane Michael 2018, Hurricane Helene 2024, and the
2025 Los Angeles wildfire complex).

## Results

Macro F1 **0.624** (validation) / **0.619** (held-out test) on the stratified
split; per-class test F1 of 0.91 (Undamaged) and 0.84 (Total Structural Collapse).
Total Roof Damage remains the hardest class (F1 ≈ 0.33).

## Usage

Load these weights with the training/inference code:
[github.com/YimingXiao98/Damage-TriageFormer](https://github.com/YimingXiao98/Damage-TriageFormer)
(MIT). The model expects a 1024×1024 post-event RGB tile and building instance
masks, and returns a 5-class damage-typology probability per building.

## Training data

DamageTriage-Bench: [huggingface.co/datasets/Ymx1025/DamageTriage-Bench](https://huggingface.co/datasets/Ymx1025/DamageTriage-Bench) (CC-BY-NC-4.0).

## Intended use and limitations

Screening-grade decision support for post-disaster triage, not a substitute for
engineering inspection. Footprint-conditioned (assumes building masks are
available). Rare and visually ambiguous roof-damage categories, especially Total
Roof Damage, are the least reliable. Trained on a single seed.

## License

Released under CC-BY-NC-4.0, consistent with the DamageTriage-Bench dataset it was
trained on. Subject to the redistribution terms of the underlying NOAA Emergency
Response Imagery and the source building-footprint layers.

## Citation

```bibtex
@misc{xiao2026damagetriageformerfoundationmodelframeworktypologybased,
      title={Damage-TriageFormer: A Foundation-Model Framework for Typology-Based Building Damage Assessment from Mono-Temporal Imagery}, 
      author={Yiming Xiao and Yu-Hsuan Ho and Sanjay Thasma and Junwei Ma and Ali Mostafavi},
      year={2026},
      eprint={2606.12248},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.12248}, 
}
```