Actual recording · Synthetic English narration and captions · Real images · VisA
THE PROBLEM
An inspection alert needs reference data, separate evaluation and review of false positives.
THE WORKFLOW
A ResNet18 patch memory is fitted on 60 normal VisA PCB1 images. Another 20 calibrate the threshold; 80 separate images evaluate the result. The interface displays the heat map and preserves human corrections.
01Normal references
02Separate calibration
03Held-out image evaluation
04Map, review and export
RECORDED EVIDENCE
Results, with context.
Recorded checks on the stated dataset
Check
Cases
Held-out images correctly classified
76 of 80
Dataset
line-inspector-v1
Completed cases
80/80
Median / p95 per case
0.112 s / 1.549 s
Evaluation record (UTC)
2026-10-09
Source: local pipeline execution on the identified inputs. CPU; no GPU. HTTP timings refer to the interface samples; for reconstruction they measure loading recorded geometry, not a new reconstruction. These are not customer outcomes or a guarantee on other data.
An inspection pilot with a data audit, normal references, acceptance checks and integration.
What we would define first
Part types, defects, capture, variation, false-positive/false-negative costs and your own acceptance dataset.
Demonstration limits
Small CC BY 4.0 VisA PCB1 subset. Correctly classified 76/80 images: 4 false positives and 0 false negatives. This is neither the full VisA evaluation nor a factory performance guarantee. The map is not certified defect segmentation.
Recorded demo
Recorded demo