“Real-time” appears on almost every Vision AI datasheet, yet it rarely comes with a definition. For operators, the only number that matters is how long it takes from something happening in front of a camera to a human or system being able to act on it.
Three numbers, three meanings
- Frame rate tells you how often the model looks, not how quickly it answers.
- Inference time measures the model alone, excluding decode, tracking, and rules.
- End-to-end latency is the one to specify in contracts: frame captured to alert delivered.
Design for the worst second, not the average
Load spikes, network jitter, and many simultaneous events are when latency matters most. Pipelines should degrade gracefully, prioritising critical analytics and shedding non-essential work under pressure.
All insights3Cubicles Team