DatasetOps Vision Lab

Local-first image dataset audit dashboard with a Python computer vision engine.

DatasetOps Vision Lab audits image-classification datasets before modeling. It reads local folders, calculates transparent readiness metrics, generates latest-report.json, and visualizes the result in a premium Next.js dashboard.

Features

• Python Audit Engine for local/offline image dataset scans
• Folder parsing for train/val/test/<label> and unsplit <label> datasets
• Class distribution and imbalance detection
• Resolution, blur, brightness, contrast, and corrupt image checks
• SHA-256 exact duplicate and train/val/test leakage detection
• Optional confusion matrix CSV parser
• Deterministic weighted risk score and quality score
• Evidence-based recommendations: problem -> evidence -> action
• Next.js dashboard for importing latest-report.json
• Browser Fast Scan for quick folder/class checks

Limitations

• No semantic object understanding.
• No near-duplicate detection in v1.
• No training or classification in v1.
• No guarantee test performance improves; this tool audits readiness risk only.

Structure

apps/web/        Next.js dashboard
engine/python/   Python audit engine
reports/         Local report outputs, gitignored

Setup

pnpm install
python -m pip install pytest jsonschema opencv-python pillow numpy

For local engine commands from this repo, set Python path if the package is not installed:

$env:PYTHONPATH="engine/python/src"

Run Python Audit

Run this first when you need duplicate, leakage, corrupt image, blur, brightness, and contrast evidence:

pnpm engine:scan -- --path ./dataset --out ./reports/latest-report.json

Outputs:

• latest-report.json
• dataset-audit.csv
• bad-images.csv
• duplicates.csv
• leakage.csv
• recommendations.md

Run Dashboard

pnpm dev

Open http://localhost:3000, then import reports/latest-report.json.

Browser Fast Scan is available for quick folder/class checks, but it does not run deep audit metrics. Leakage needs a train/val/test structure so the engine has split boundaries to compare.

Quality Checks

pnpm lint
pnpm typecheck
pnpm test
pnpm engine:test
pnpm build
pnpm test:e2e

Deployment

Deploy apps/web as a standalone Vercel project named datasetops-vision-lab. No environment variables are required.