[DATASET] Golden Dataset Curation & Benchmark Ground Truth

[hdenizkaraman]

🎯 Task Description

Curate and annotate a high-quality Golden Dataset (150–300 annotated video samples) representing industrial safety environments to serve as the ground truth benchmark for the project's evaluation metrics.

🛠️ Affected Domain / Package

• apps/dashboard (Next.js Live Monitoring Panel)
• apps/identity (Next.js Live Monitoring Panel)
• apps/gateway (Rust Axum API Gateway & Realtime SSE)
• apps/stream (Video Ingestion & Dynamic Frame Sampling)
• apps/ai (Agentic Reasoning & Decision Logic)
• packages/optics (Image & Media Processing Toolkit)
• packages/database (SurrealDB, Qdrant)
• platform/ or tools/ (Docker, Infrastructure & Benchmarks)
• Other (please specify): ``

---

📋 Task Checklist

• Source and slice 150–300 targeted video clips across all distribution categories.
• Generate structured Ground Truth JSON files matching TEKNOFEST requirements (timestamps, events, severity, ground-truth actions).
• Set up metric evaluation scripts (Event Detection Recall, False Alarm Rate, Incident Classification Accuracy).

---

📊 Dataset Distribution Breakdown

• Clear Incident / Work Accidents (30%): Unambiguous incidents (e.g., forklift rollover, worker falls, machinery entrapment).
• Edge Cases & Near-Misses (15%): Hard-to-distinguish edge cases (e.g., worker bending down to pick up tools vs. falling, stumbling without injury).
• Normal Operations / False-Positive Control (30%): Standard workflow without safety violations or accidents.
• Environmental Challenges (15%): Adverse visual conditions (low resolution, heavy fog/smoke, night vision, poor factory lighting).
• Visual Artifacts & Reflections (5%): Tricky optical artifacts (e.g., forklift appearing on an ambient CCTV/TV screen reflection, mirror surfaces).

---

🔗 Related Resources / Docs

• Internal Docs: documents/dataset/golden.md