Documentation

# Set up and verify video analysis

## Choose where images are analyzed

Open Watch settings to see the choices available in your build. Choose local analysis when images must stay on the monitoring computer. For a remote service, confirm that your organization approves sending site images there, then enter the connection details supplied by its administrator.

## Adjust frequency and image detail

1.  In Settings → Camera AI engine, note the current image detail and analysis interval. Change one setting at a time.
2.  Choose Standard or High image detail. Find the analysis interval under Analysis frequency and AI model connection. A shorter interval requests more frequent analyses.
3.  Repeat the same scene on the same camera. Compare result timestamps, content, and update frequency. Restore the previous setting if results become slower or unstable.

For video without analyses, start with [troubleshooting](https://watch.tensordynamics.org/docs/faq?lang=en). The hardware checks and short observation below require the [desktop assistant](https://watch.tensordynamics.org/docs/setup-assistant?lang=en); ordinary web chat can guide you through manual checks.

[Set up local analysis](https://watch.tensordynamics.org/docs/engine-local?lang=en) · [Connect an approved service](https://watch.tensordynamics.org/docs/engine-remote?lang=en)

## Test an analysis

Connect a camera, enable Monitoring in its settings, and give it a task such as “Describe whether someone enters the doorway.” Have a colleague walk through. Check for a fresh timestamp and matching description in that camera’s log. If you have live video but no result, check the mission, monitoring state, and analysis settings, then repeat the same test.

[Adjust a camera mission](https://watch.tensordynamics.org/docs/missions?lang=en)

## Check hardware and workload requirements

Once the desktop assistant is installed and connected, ask it to check the OS version, CPU, memory, free disk space, GPU, and graphics memory. Unavailable specifications remain unknown. Check separately whether Watch uses an available accelerator. Apple unified memory is shared by the system and GPU, so count it only once.

Then establish camera count, resolution, video FPS, analysis interval and required response time. Eight cameras analyzed every five seconds need an average 1.6 completed analyses per second. Video frame rate and analysis frequency are separate workload measurements.

## What TOPS means for camera capacity

```
Theoretical TOPS = 2 × MAC units × clock GHz ÷ 1000
```

This formula counts a multiply-accumulate as two operations. It requires the MAC unit count, clock speed, and precision. CPU/GPU core counts cannot be substituted. Compare vendor numbers only at matching precision and dense/sparse conditions. TOPS is theoretical peak; memory, bandwidth, decoding, thermals, network and the site workload still affect camera capacity. [TOPS definition and formula](https://www.qualcomm.com/news/onq/2024/04/a-guide-to-ai-tops-and-npu-performance-metrics) · [Precision and sparsity](https://developer.nvidia.com/blog/accelerating-inference-with-sparsity-using-ampere-and-tensorrt)

## Measure performance per camera

Start the cameras you want to test, then ask the assistant to observe for 30–45 seconds. Review completed analyses per second, processing P50/P95, the age of the latest analyzed frame, waiting and replaced samples, and CPU and memory use. Measure API round-trip time, analysis time, and capture-to-notification latency separately. The check observes existing activity without starting cameras or adding test load.

```
Provisional cameras = floor(measured completed analyses/sec × analysis interval/sec × (1 − reserve))
```

The estimate reserves 25% headroom by default and is capped at the number of cameras tested. It applies only to the same resolution, FPS, missions, and analysis interval. Resolve insufficient or stale data, counter resets, growing waits, or cameras that stop updating before relying on the estimate. Use the result as an initial trial limit, then test sustained operation.

## Increase load and verify site operation

Increase load in stages such as 1, 2, 4 and 8 cameras up to the requirement. Follow each short check with 5–10 minutes of representative load. Before handover, test a representative operating period, including reconnections, errors, memory growth, temperature-related slowdowns, and notification delivery. Record hardware, Watch version, camera settings, measurement duration and the first failing check. Detection accuracy and physical camera-to-notification latency need separate tests.

[

Next

On-device analysis

](https://watch.tensordynamics.org/docs/engine-local?lang=en)

---
Source: https://watch.tensordynamics.org/docs/inference
More Watch guides: https://watch.tensordynamics.org/llms.txt