DamageTriage-Bench

DamageTriage-Bench is a footprint-conditioned benchmark for per-building damage typing from single post-event aerial images. Its five classes distinguish roof from structural damage and partial from total affected extent:

ID	Class
0	Undamaged
1	Partial Roof Damage
2	Total Roof Damage
3	Partial Structural Damage
4	Total Structural Collapse

Quick statistics

Item	Value
Tiles	7,472 (1024 × 1024 PNG)
Labeled polygon fragments	87,017 before the 30-pixel evaluation filter
Disasters	Hurricane Michael (2018), Hurricane Helene (2024), and the 2025 Los Angeles Palisades/Eaton wildfire complex
Acquisition sub-events	12
Recommended train / validation / test split	5,231 / 1,124 / 1,117 tiles

Files

.
├── images/                       # post-event RGB tiles
├── damage/                       # paired RGB damage-typology masks
├── photo_splits.json             # recommended photo-atomic split
├── stratified_splits.json        # superseded tile-level split
├── class_index.json              # per-class tile listings
└── README.md

images/<tile_id>.png and damage/<tile_id>.png are paired by filename.

Annotation format

Each damage file is a unified RGB polygon mask. Background is black; each non-background color gives the class of a labeled building fragment.

RGB	Class
(0, 0, 0)	Background / ignore
(255, 255, 255)	Undamaged
(0, 255, 83)	Partial Roof Damage
(246, 255, 11)	Total Roof Damage
(255, 138, 18)	Partial Structural Damage
(255, 0, 0)	Total Structural Collapse

Connected components of a class color are treated as labeled fragments. The paper's reported metrics exclude components smaller than 30 pixels.

For damaged buildings, the annotation rubric first distinguishes roof-surface damage from damage extending into structural components. Within each branch, a 50% affected-area threshold separates partial from total damage. The threshold is a pragmatic annotation convention, not a validated operational decision threshold.

Recommended photo-atomic split

Use photo_splits.json for the revised paper results. Within each of the 12 acquisition sub-events, the source photo is the assignment unit: all tiles cut from one source photo remain in the same fold. The split contains 5,231 training, 1,124 validation, and 1,117 test tiles.

This protocol prevents fragments derived from one source photo from crossing folds. It does not reconstruct buildings clipped at tile boundaries, merge repeated fragments within a fold, or rule out the same physical building appearing in overlapping source photos. Reported instance metrics therefore use the visible labeled fragment as the evaluation unit.

stratified_splits.json is retained only to reproduce the originally submitted tile-level experiment. Under that superseded split, fragments of 5,027 audited buildings entered more than one fold; it should not be used for new comparisons.

Validation and test supports for the recommended split

These counts include labeled fragments with area of at least 30 pixels.

Class	Validation	Test
Undamaged	7,580	7,257
Partial Roof Damage	1,246	1,280
Total Roof Damage	121	85
Partial Structural Damage	316	312
Total Structural Collapse	890	921
Total	10,153	9,855

Revised reference results

On the recommended split, the revised paper reports:

Configuration	Validation macro F1	Test macro F1
Damage-TriageFormer, crop mode	0.646	0.616
Damage-TriageFormer, tile mode	0.575	0.579

Both are single-seed results. The crop model assumes the supplied building support and is not a footprint-free detector. Code is provided at https://github.com/YimingXiao98/Damage-TriageFormer.

Limitations and intended use

The benchmark supports research on footprint-conditioned damage typing. It does not validate autonomous habitability, safety, inspection-priority, or resource- allocation decisions. The three events do not establish generalization to a new disaster; leave-one-event-out performance in the paper is substantially lower. Rare roof-damage classes have small support, and verification-style annotation review is not equivalent to blind independent labeling.

Use of the imagery and footprint-derived labels remains subject to the terms of the underlying NOAA Emergency Response Imagery and source footprint layers.

License

CC BY-NC 4.0 for non-commercial research use, subject to the underlying source terms noted above.

Citation

@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}, 
}