# industrial-vlm-temporal-grounding

A research benchmark and training project for small vision-language models that spot factory procedure errors in first-person video by identifying exactly when a step starts and ends, aimed at real-time deviation warnings before an incident happens.

industrial-vlm-temporal-grounding develops small vision-language models that watch first-person factory video and identify temporal grounding: the start and end timestamps of a specific worker action described in a sentence, or that no matching action occurred. The goal is a low-latency, on-device system that tracks which procedure steps a worker has completed and flags a dangerous or out-of-order step before it leads to an accident, equipment damage or a quality failure, without sending sensitive video to an external API.

The project benchmarks several models of four billion parameters or fewer on twenty short, manually annotated first-person factory videos, scoring each by temporal intersection-over-union against human-labeled ground truth rather than trusting machine-generated annotations. Results are reported as a diagnostic comparison on this development set rather than as unseen-test accuracy, and the repository's stated long-term aim is accumulating precise, site-specific procedure definitions that a small model can be fine-tuned on repeatedly, rather than relying on a larger general-purpose model.

Website: https://github.com/shure-dev/small-vlm-sop-check