Docs / CAMZA.AI
Measure the workflow. Show the conditions.
An evaluation should explain the footage, queries, labels and environment behind every performance number. This methodology separates search relevance, missed events, nuisance alerts and timing. No Camza accuracy or latency benchmark is published here. The website sample is a predefined demonstration and must not be used as product performance evidence.
UPDATED
Build a representative evaluation set
- Record camera models, image settings, locations and evaluation dates
- Include ordinary shifts, lighting changes, weather and occlusion
- Keep positive examples and negative examples that look similar
- Separate tuning footage from footage used for final evaluation
Define the ground truth
Have people who understand the work label the event and time window. Write labeling rules before scoring. Resolve disagreement explicitly and keep uncertain examples visible instead of forcing every clip into a confident label.
Report distinct measures
| Measure | What to report | Why it matters |
|---|---|---|
| Search relevance | Relevant reviewed results / all reviewed results | Whether returned moments help an investigation |
| Missed events | Known target events not found | What useful activity the workflow overlooks |
| Nuisance alerts | Unhelpful alerts per camera and operating period | The interruptions the team must absorb |
| Timing | Start and end markers, distribution and sample count | The delay the user or recipient experiences |
| Coverage | Cameras, conditions and excluded cases | Where the result can reasonably apply |
Choose explicit timing markers
Distinguish capture-to-detection, detection-to-destination and query-to-visible-result. Report the observation method, network conditions, warm or cold state, sample count and the distribution. A single favorable run is not an operating guarantee.
Website sample disclosure
The website uses illustrative scenes and predefined event metadata. Search results and simulated rules are deterministic sample behavior, not live model inference or a performance test.
A few useful questions.
Can I try this on my own cameras?
Start with a compatibility request. Share camera models, firmware and the event you want to find. An evaluation needs an agreed scope before any camera access is arranged.
Is the website demonstration a product benchmark?
No. The sample demonstration uses illustrative scenes and predefined events. It does not measure accuracy, response time or compatibility on your cameras.
What should we agree before a pilot?
Agree which streams are in scope, who can access the data, where processing runs, retention and deletion, expected outcomes, and the person responsible for reviewing results.
YOUR CAMERAS. YOUR CONTEXT.
Start with one
useful question.
A camera inventory. An event to find.
A clear next step for your team.