VisionCrew Inspector

v2 · 10-step fail-safe pipeline · OpenCV 5 · runs locally

Pipeline steps (human verifier is the last step)

Original vs flagged

Original image (640x480 frame)
Original
Annotated image

Why it was flagged

    v1 vs v2

    Per-step timing

    Recall by failure type (caught = not auto-passed)

    The fixed-camera detectors do not apply to these photos (other cameras), so v2 routes all of them to a person. That is the fail-safe: it never auto-passes a possible spill, and it also escalates every safe photo. No real-world accuracy is claimed.

    Every photo · expected vs got

    PhotoWhatExpectedGotResultmsLicense

    Batch inspect

    Drop images here, or

    Each file runs through the full pipeline (up to 60 per batch). Nothing is stored.

    Summary

    #ImageVerdictFlagsWhyms
    No images yet.
    Nothing to review yet. Inspect an image first.

    Ten steps, the person always last

    Fail-safe rules

    PASSOnly when the image is usable, from the calibrated camera, every policy rule passes and both spill detectors agree the floor is dry.
    FAILA hazard confirmed by two detectors (v1 spill ≥ 0.80 and the recall sweep).
    ESCALATEDExit blocked (obstruction ≥ 0.80): a person is alerted now.
    NEEDS HUMANAnything uncertain: bad quality, unknown camera, detectors disagree, a finding in the 0.65–0.80 band, or a weak reading v1 dropped.

    Never auto-passes a possible spill. The cost is more escalations (see the Real photos tab and evaluation/V2-EVIDENCE.json). The v1 page is still at /v1.