Set up defect detection

You do not need training data or a data scientist. You describe what you are looking for in plain words, and the cameras start finding it.

Before you start

• At least one registered device in a device group
• A clear idea of the defect you want caught — “solder bridge”, “missing clip”

1. Choose the group

Open Captures and pick your device group from the selector in the top right, then open the Config tab.

Everything you set here applies to every camera in that group.

2. Describe what to look for

Turn Detection Enabled on and type what you are looking for into Text Queries, separated by commas:

solder bridge, missing component, bent pin

This is the part people expect to be harder than it is. The model reads your description and finds matching objects — there is no dataset to collect and no model to train.

Use the words your team already uses on the floor. If an operator would say “bent pin”, write that.

3. Set how strict it is

Setting	What it does	Start with
Threshold	How confident the model must be before it counts something. Higher means fewer, surer detections	0.2, then tune
Frame Interval	Run detection every N frames. Higher means less load on the device	10
Similarity Threshold	How alike two detections must be to count as the same thing	Leave as is

Getting the threshold right is trial and error, and that is expected. Too low and every shadow is a defect; too high and real ones slip through. Do not agonise over it here — step 5 is how you settle it properly.

4. Turn on tracking if you are counting

Setting	Turn on when
Object Tracking	You are counting things. Without it, one defect sitting in view for ten frames counts ten times
SAM Segmentation	You need the exact shape of a defect, not just a box around it
Denoise	The feed is grainy — poor light, cheap optics, vibration

Tracking is the one most people need and forget. If your counts look absurdly high, this is why.

Changes apply immediately. There is no restart, and no need to schedule a window — the devices pick up the new settings and carry on.

5. Prove it before you trust it

Open the Sandbox tab, upload a recorded video from the line, and run it with the settings you just chose.

You get a processing history you can compare across runs, which is how you answer the real question: would this setting have caught last month’s escape?

Do this before you rely on a threshold in production. It costs ten minutes and it is the difference between a number you picked and a number you can defend.

Next steps

• Get alerted when detections spike, instead of watching the feed
• Build a dashboard to see the trend
• Inspection if you are comparing against a known-good reference rather than describing defects