Computer vision without the glue code

FGCV was founded by CV engineers who had each rebuilt the same fragile pipeline at three companies: a labeling tool here, a training script there, an inference server held together with cron jobs, and no versioned connection between any of them. Every model incident began the same way — 'which dataset was this trained on?' The platform exists to end that sentence as a question.

What the platform does

FGCV is one versioned pipeline from data to production:

• Dataset management: labeled datasets are versioned artifacts, linked to every model trained on them, with drift monitoring that tells you when your data stops resembling your reality.
• Training and evaluation: reproducible runs with task-aware eval suites — detection, segmentation, tracking — scored before anything deploys.
• GPU-efficient inference: batched, quantized serving that holds latency targets on commodity hardware, with monitoring wired to the same lineage as training.
• An open-core SDK with a public roadmap, so the community can see — and vote on — what ships next.

Who runs on it

Teams that cannot afford a missed defect, a model drift incident, or a six-month integration project: industrial inspection lines, medical imaging groups, autonomous systems, and the enterprise CV platforms built on top of us. Enterprise deployments run with full audit trails, because in regulated environments 'the model changed' needs a paper trail.

Computer vision earns its keep in production, not in notebooks. FGCV is the operating system for that last mile — built by practitioners, versioned end to end, and boring in exactly the ways you want.