Skip to content

Latest commit

 

History

History
51 lines (35 loc) · 1.51 KB

File metadata and controls

51 lines (35 loc) · 1.51 KB

Project Context

Product Summary

DatasetOps Vision Lab audits local image-classification datasets before training or test evaluation. It combines a Python computer vision engine with a local Next.js dashboard for transparent readiness scoring and evidence-based recommendations.

Primary User

The primary user is a data science learner or builder preparing image datasets for modeling who needs to catch obvious readiness risks before trusting training or validation results.

Core Workflow

  1. Run the Python engine:

    pnpm engine:scan -- --path ./dataset --out ./reports/latest-report.json
  2. Import latest-report.json into the dashboard.

  3. Review quality score, class distribution, leakage, duplicates, image quality, corrupt files, and recommendations.

  4. Optionally use Browser Fast Scan for quick local folder structure checks.

Architecture

  • engine/python: canonical audit engine and report exporters.
  • apps/web: report viewer and browser Fast Scan dashboard.
  • reports: local output folder, ignored by git except .gitkeep.

Scoring

riskScore = weighted sum, qualityScore = 100 - riskScore.

Weights:

  • Leakage: 30
  • Imbalance: 20
  • Low resolution: 15
  • Duplicates: 10
  • Blur: 10
  • Brightness/contrast: 10
  • Confusion concentration: 5

Non-Goals

  • Training models
  • Classifying image content
  • Semantic object understanding
  • Near-duplicate/perceptual similarity detection
  • Uploading datasets
  • Cloud storage or accounts
  • Guaranteeing model or test performance improvement