HOW WE MEASURE.
CGX scores camouflage by pairing a published, MIT-licensed object detector with a calibrated human-detection game on real backgrounds. The pairing — not either signal alone — is what the patent defends and what the methodology below describes.
DGNet-S — A PUBLISHED OBJECT-DETECTION MODEL AS THE MACHINE EVALUATOR.
CGX uses DGNet-S — a Camouflaged Object Detection (COD) model published in the peer-reviewed literature and released under an MIT license — as the machine half of the effectiveness measurement. For every (camo × background × placement) trial, DGNet-S produces a segmentation mask that estimates where the camo is in the composite image. We score the segmentation against the ground-truth mask (the placement is known by construction because we put the camo there) and read out an Intersection- over-Union (IoU) value per trial. Higher IoU means the model found the camo; lower IoU means the camo hid.
The choice of a published, MIT-licensed model is deliberate. The intellectual property is in the fusion, the calibration, and the substrate — not in a proprietary detector that customers would have to take on faith. Any reviewer can pull DGNet-S from the published codebase and reproduce our machine-side scores.
- Reproducible by any reviewer with the same checkpoint.
- MIT license — no per-eval royalty surface.
- Peer-reviewed performance on standard COD benchmarks.
- Swappable as the field advances.
OVER 100,000 HUMAN-DETECTION TRIALS — AND COUNTING.
The human half of the score comes from a detection game. Real players are shown a real background with a camo pattern composited onto it. We measure how long they take to click the pattern, how many miss-clicks they make first, and whether they give up. Every trial captures the mask position, click coordinates, miss-click coordinates, and elapsed time — append-only.
CGX's human-trial substrate has been collected continuously since 2022. It carries over 100,000 trials with full per-trial mask-position and miss-click telemetry — a corpus that grows every week the consumer instrument is live.
The substrate itself is the moat. We never expose aggregate effectiveness data on a public surface. Per-cell rankings are returned only through audited admin paths to paying B2B customers.
- Time-to-detection (ms)
- Click coordinates (x, y)
- Miss-click coordinates + count
- Did-not-find / timeout outcome
- Mask position (ground truth by construction)
PER-PAIR IoU + HUMAN TIME — JOINED ON A NAMED TEST POOL.
For each (camo × background × placement) trial, CGX records two parallel signals: the AI detector's IoU score against the ground-truth mask, and a human player's time-to-detection on the same composite. Because both signals key off the exact same physical scene, they join cleanly into a per-pair row of evidence. Each row is reproducible: the background, placement, and camo are all reconstructable from the row's lineage tag — same query, same answer, months later.
We group rows into named, versioned test pools. A pool is a set of trials keyed to a target region, vantage distance, lighting condition, and season. The same pool can be run against any number of candidate camos — competitor patterns, in-house patterns, or proposed designs. Because the pool is fixed, the comparison is apples-to-apples.
Effectiveness is read out per-cell, not as a universal ranking. No pattern wins everywhere. Patterns that win in Pacific arid coastal scrub will lose in Upper Midwest boreal woodland — and we say so, with the rows to back it.
- AI detector IoU (mean + p10 / p90)
- Human time-to-detection (mean + distribution)
- Cell tag (region × season × vantage × lighting)
- Rule-set version
- Lineage tag for byte-identical replay
REPRODUCIBLE COMPARISONS AGAINST NAMED COMPETITOR PATTERNS.
CGX runs structured head-to-head comparisons. A test pool is fixed for a target cell. A CGX-generated or CGX-evaluated pattern is scored against a named commercial competitor pattern (Veil, Mossy Oak, First Lite, Sitka, KUIU, Realtree — whichever is the operative incumbent for that cell) on the same pool, with the same model checkpoint and the same rule set. The winner is reported per-cell with a confidence interval, locked to a rule-set version so the same row can be re-verified months later from the lineage tag alone.
Results are deliberately reported per-cell rather than as an overall victory. A pattern that wins in Pacific arid coastal scrub may lose in Upper Midwest boreal woodland. The reproducible row is the unit of evidence; a B2B report aggregates many rows with full per-cell provenance.
Concrete per-cell numbers — including the IoU and human-time deltas against named competitor patterns — appear in signed B2B reports, not on this public page. The data moat is the business; we describe the mechanism here and license the answer through audited paths.
- Same pool for both candidates
- Same model checkpoint + rule set
- Per-cell winner + CI, not global claim
- Lineage tag for re-verification
- Released only via audited admin paths
The combined human + AI effectiveness measurement, fused into a calibrated score and ranked per camo × background × region × season, is the inventive concept the patent defends. The ingredients — object-detection models, geo-tagged backgrounds, click-trial games — each have prior art. The end-to-end calibrated combination is what's ours.
← Back to CGX