Benchmark RadarAI BENCHMARK PROFILE
SteelBench
STEELBENCH evaluates vision-language models on per-worker activity recognition and safety-rule reasoning in industrial CCTV footage. It includes 1,345 clips with dense annotations and a provenance-aware audit protocol.
- Released
- 2026-07-06
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Industrial surveillance poses unique visual and procedural challenges. This benchmark reveals large gaps in VLM performance and shows how annotation provenance can inflate accuracy, guiding reliable deployment.
Motivation
Existing video benchmarks evaluate action recognition on consumer videos, egocentric recordings, or simulated industrial environments.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.