Platform
Generating the night shift you cannot film
Storms, glare and the hour before dawn are the conditions that break a model and the hardest to capture on site. How to ask for them instead.
AI in the real worldKey takeaways
- The conditions that break a model are the ones you cannot schedule a shoot for.
- Describe the scene in operational language, not in rendering language.
- Vary one condition at a time when testing, and everything at once when training.
- Generated conditions are for coverage; the held back real week still decides.
Every dataset is quietly a daylight dataset. Not by choice, by availability. Somebody went out with a camera during working hours, in the weather that happened, and the model learned that the world is well lit and dry.
The hours that break things
- The twenty minutes either side of dawn, when auto exposure gives up.
- Wet tarmac under sodium lighting, where everything has a second copy of itself.
- Low winter sun straight into a fixed camera.
- Fog, which removes contrast rather than light and fools thresholds tuned for night.
You cannot book those. You can wait eighteen months for them to coincide with someone being on site with a plan, which is the same as not having them.
How to ask for them
Brief a generator the way you would brief a crew: the place, the hour, the weather, what the people in frame are doing and what they are wearing. Rendering vocabulary produces rendering results; operational vocabulary produces the scene you were actually picturing.
Ask for the loading bay at 05:40 in heavy rain with two people unloading and one vest missing, not for 'low light, high noise, realistic'.
One variable at a time, when testing
For training, vary everything; breadth is the point. For testing, hold the scene still and change one condition, or a failure tells you the model is worse at night without telling you whether it is the light, the rain or the reflections that did it.
Frequently asked
Does generated night footage look like real night footage?
Closely enough to train on, once the optics and compression are matched. Whether it is close enough for your site is a question the held back real week answers, not the renderer.
How many conditions are worth generating?
Start with the four that your incident log says you miss things in. A condition matrix nobody derived from real failures is a large set that improves nothing.
Keep reading
All articles
Deployment
What the first week on site actually looks like
Day by day, from a spreadsheet of camera credentials to the first alert somebody acted on, including the parts that never make it into a rollout plan.
Read article
Field notes
Writing a handover the next shift will read
An incident log is only worth the ninety seconds the next operator gives it. A short template, and the three fields that earn their place.
Read articleField notes
What a false positive actually costs you
Every alert nobody trusts is a shift spent scrubbing footage. A short accounting of what a noisy model does to a security team, and the two numbers worth watching.
Read article
