aapdasetu-damage-assessment

AapdaSetu Damage Assessment — ResNet50

Post-disaster building damage classifier for the AapdaSetu platform. Given a photo of a disaster-affected building, the model classifies damage severity into one of three grades and drives automated compensation assessment in the AapdaSetu relief pipeline.

Part of the AapdaSetu project (Smart India Hackathon).

Live Deployment

The full assessment service (this model + EXIF authenticity checks + pHash duplicate detection + compensation estimation) is deployed and publicly accessible:

🌐 https://aapdasetu-damage-api.onrender.com

• Interactive demo UI: / · Swagger API docs: /docs · Health check: /health
• Main endpoint: POST /api/assess-damage (multipart: photo, claimed_lat, claimed_lng, property_type, disaster_cutoff)
• Production inference runs on ONNX Runtime using best.onnx from this repo (torch-free, fits Render's 512 MB free tier). Free-tier instances sleep after ~15 min idle, so the first request after a pause cold-starts in a few seconds.

Model Details

Training Recipe

• Dataset: 2,400 post-disaster building images — MINOR (846), MAJOR (525), DESTROYED (1,029), split 75 / 15 / 10 per class with a leak check (image hashes verified disjoint across splits — see leak_check.json). Dataset hosted at Divyanshu-Kumar19/aapdasetu-damage-dataset.
• Two-phase training:
  1. Backbone frozen — only the new FC head trains
  2. Full fine-tuning of the entire network
• Evaluation: 5-crop Test-Time Augmentation (TTA) on the held-out test set.

Evaluation Results

Held-out test set (244 images), with TTA:

Full metrics: eval_report.json · Confusion matrix: confusion_matrix.png · ROC curves: roc_curves.png · Training curves: training_curves.png · Misclassified samples: misclassified.png · Training history: history.json · Split leak audit: leak_check.json

Quick Start — ONNX (recommended, CPU-friendly)

import json
import onnxruntime as ort
import numpy as np
from PIL import Image
from torchvision import transforms

mapping = json.load(open("class_mapping.json"))          # {"DESTROYED": 0, "MAJOR": 1, "MINOR": 2}
idx_to_class = {int(v): k for k, v in mapping.items()}

transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

img = Image.open("building.jpg").convert("RGB")
x = transform(img).unsqueeze(0).numpy()

sess = ort.InferenceSession("best.onnx", providers=["CPUExecutionProvider"])
probs = sess.run(["scores"], {"image": x})[0][0]          # softmax already baked in

print("Damage grade:", idx_to_class[int(probs.argmax())])
print("All scores:", {idx_to_class[i]: round(float(p), 4) for i, p in enumerate(probs)})

Quick Start — PyTorch (best.pt)

import torch
import torch.nn as nn
from torchvision import models, transforms
from PIL import Image

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# 1. Build the exact architecture used during training
model = models.resnet50(weights=None)
model.fc = nn.Sequential(
    nn.Dropout(p=0.0),                    # training-only; inert at inference
    nn.Linear(model.fc.in_features, 3),
)

# 2. Load the checkpoint
ckpt = torch.load("best.pt", map_location=device, weights_only=True)
model.load_state_dict(ckpt["model_state"])
model.to(device).eval()

# 3. IMPORTANT — use the checkpoint's index->class mapping.
#    ImageFolder trained classes in ALPHABETICAL order
#    (DESTROYED=0, MAJOR=1, MINOR=2), NOT ["MINOR", "MAJOR", "DESTROYED"].
idx_to_class = {int(v): k for k, v in ckpt["class_to_idx"].items()}

# 4. Preprocess exactly like training validation
transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

# 5. Predict
img = Image.open("building.jpg").convert("RGB")
x = transform(img).unsqueeze(0).to(device)
with torch.no_grad():
    probs = torch.softmax(model(x), dim=1)[0]

scores = {idx_to_class[i]: round(float(p), 4) for i, p in enumerate(probs)}
print("Damage grade:", idx_to_class[int(probs.argmax())])
print("All scores:", scores)

Download the checkpoints with:

hf download Divyanshu-Kumar19/aapdasetu-damage-assessment --local-dir ./model

Intended Use

• Triage of post-disaster building damage photos in the AapdaSetu claim pipeline (maps grade → compensation under NDRF/SDRF norms).
• Not a substitute for a certified structural engineer's assessment.
• Out of scope: images that are not disaster-affected buildings, and safety-critical decisions without human review.

Files in This Repository

License

MIT