False-Positive Stress Test · $1,500
Find out what fools your detector before your customers do.
- Fixed price
- $1,500
- Turnaround
- 5 days
- On request
- NDA
- Weights kept
- 0
01What you get
A report you can act on, not a score.
- Precision and recall per scenario and per class
- Every false positive and missed detection, with the frame
- Failure clusters ranked by how often they would fire in production
- A short fix list: which data would close each gap
- A 30-minute call to walk through the results
Report excerpt · illustrative
| Scenario | FP rate | Fix |
|---|---|---|
| Umbrella, folded | high | hard negatives |
| Phone, raised | medium | hard negatives |
| Night, IR | medium | low-light set |
| Crowd, occluded | low | — |
The real report has per-class numbers and the frames behind every row.
02Process
Five working days, start to finish.
- Day 0
Short call: classes, cameras, where the alarms go wrong.
- Day 1
You send weights or run our set and return predictions.
- Day 2–4
We run every scenario, cluster the failures and write the fix list.
- Day 5
Report and walkthrough call.
Most teams follow up with a custom pack that targets the failures we found.
03Questions
Common questions
Do you need our model weights?
No. You can send weights (ONNX, PyTorch, TensorRT) for us to run, or we send you the test set under NDA and you return the predictions in a standard format. Either way the report is the same.
Which models can you test?
Any object detector that outputs boxes and classes — YOLO variants, RT-DETR, Faster R-CNN or a vendor model you can export predictions from. Classification and VLM-based alarms can be tested on request.
What happens to our model and results?
Weights are deleted after the test and results are only shared with you. We sign your NDA if you have one.
What if the model passes?
Then you have an independent report that says so, with the edge cases it was tested on — useful for customers and procurement.
Book