Divyanshu-Kumar19/aapdasetu-damage-dataset

AapdaSetu Building Damage Dataset

Ground-level photos of disaster-affected buildings, labelled by damage severity. Used to train the ResNet50 damage classifier Divyanshu-Kumar19/aapdasetu-damage-assessment (98.36% test accuracy with 5-crop TTA).

Part of the AapdaSetu project (Smart India Hackathon).

Classes

Class	Count	Description
MINOR	846	Hairline cracks, broken windows, superficial damage — structure liveable
MAJOR	525	Partial wall/roof collapse, severe damage — main structure still exists
DESTROYED	1,029	Complete collapse, rubble only

Layout

raw/
├── MINOR/        (846 images)
├── MAJOR/        (525 images)
└── DESTROYED/    (1,029 images)

splits.json records the class-stratified split actually used for training: 75% train / 15% val / 10% test per class, with a perceptual-hash leak check verifying no image (or near-duplicate) appears in more than one split.

Usage

from torchvision.datasets import ImageFolder

dataset = ImageFolder("raw")   # class_to_idx is alphabetical:
                               # DESTROYED=0, MAJOR=1, MINOR=2

Or download with the HF CLI:

hf download Divyanshu-Kumar19/aapdasetu-damage-dataset --local-dir ./dataset

Files

File	Description
raw/{MINOR,MAJOR,DESTROYED}/*.jpg	Labelled source images
splits.json	Train/val/test split counts per class
DATASET_GUIDE.md	Guide on collecting and preparing your own photos

Related

• Model + checkpoint: Divyanshu-Kumar19/aapdasetu-damage-assessment
• Service code: ai-service/ in the SIH repo (divyanshu branch)

License

MIT