ReliefLens

ReliefLens is a local-first image triage prototype built around apple/MobileCLIP2-S0. It turns a folder of field photos into a ranked CSV, JSONL, and HTML dashboard showing likely needs, hazards, damage, supplies, and follow-up actions.

The core idea: in floods, fires, storms, housing inspections, mutual-aid intake, or community health work, people often have messy photo dumps before they have clean forms. ReliefLens uses zero-shot image/text matching to make those photos searchable and actionable without sending them to a hosted vision API.

Why this model

apple/MobileCLIP2-S0 is small, current, and underused. Apple lists it as 11.4M image parameters plus 63.4M text parameters, which makes it a good fit for high-volume local transformation work.

This project is intentionally Apple-model-first:

• Default model: apple/MobileCLIP2-S0
• Official implementation: apple/ml-mobileclip
• Upstream contribution: apple/ml-mobileclip#11
• Apple-specific notes: APPLE_MOBILECLIP.md

What it produces

• triage.csv: one row per image with top category, confidence, action, and per-category scores.
• triage.jsonl: machine-readable records for downstream workflows.
• dashboard.html: a static, shareable review page with severity counts, category filters, search, and top-match evidence.
• embeddings.npz: optional cached image embeddings for re-labeling without re-encoding images.

Install

Create a Python environment, then install:

pip install -r requirements.txt

The model card for MobileCLIP2 says to install Apple's ml-mobileclip package. If it is not available through pip in your environment, install from Apple's GitHub repo, then rerun ReliefLens:

pip install git+https://github.com/apple/ml-mobileclip.git

Run

python relieflens.py scan `
  --images C:\path\to\photos `
  --out C:\path\to\relieflens-output `
  --model apple/MobileCLIP2-S0

For a faster dry run that validates parsing and output generation without loading the model:

python relieflens.py demo --out C:\path\to\relieflens-demo

Customize the mission

Edit taxonomy.yaml. Each category has prompts and an action. ReliefLens embeds the prompts, compares them to each image, and reports the best match.

Good taxonomies are concrete:

• Prefer a washed out road or bridge over infrastructure.
• Prefer standing flood water inside a home over water issue.
• Add local needs: insulin, oxygen tanks, wheelchair ramps, school meals, damaged documents, mold, downed power lines, blocked exits.

Validate

python -m unittest discover -s tests
python relieflens.py demo --out relieflens-output

Open relieflens-output/dashboard.html to review the static dashboard.

github-actions-test.example.yml contains a ready-to-copy GitHub Actions workflow. It is kept as an example because some OAuth tokens cannot create active workflow files without the extra workflow scope.

Safety notes

ReliefLens is a triage aid, not an emergency decision-maker. Treat its output as an attention-routing layer. Human review should decide actual priority, especially for medical, legal, housing, insurance, or emergency response use.

See IMPACT.md for deployment ideas and the roadmap.

Citation

ReliefLens depends on Apple's MobileCLIP work. If this project is useful, cite the MobileCLIP2 and MobileCLIP papers listed in Apple's ml-mobileclip README.