MUSHROOM VISION / FULL PIXEL RETEST
How small can
the target get?
Compare a whole-frame scan with a scan of four overlapping tiles plus the whole frame. Follow the same target from its original photograph down to less than one pixel wide—and inspect the exact images used.
The saved test could not be loaded. No missing measurement has been replaced with zero.
Loading the measured comparison…
TWO SETS OF EVIDENCE, KEPT SEPARATE
Whole frame or tiled scan?
| Target width in the saved input | Whole-frame matches | Tiled matches | Whole-frame extra boxes | Tiled extra boxes | Control alerts whole / tiled | Time per image whole / tiled |
|---|
“Extra boxes” are all remaining detections after at most one box is assigned to the selected positive reference. They can include duplicate overlaps or objects elsewhere. Incomplete references cannot establish that every extra box is a false alarm. Control alerts come from a separate negative-reference set, whose authority is stated above. A selected region can span a cluster, so detecting one cap may still miss the region-overlap rule. This measures selected-region localization, not species identification or population accuracy.
PIXELS CAN ILLUSTRATE DISTANCE
A camera assumption,
not a flight measurement.
A 5 cm target and a focal length of 1,380 pixels give these ideal geometric widths: target pixels = 1,380 × 0.05 ÷ metres
These distances are illustrative, not measured detection rates. Your camera calibration, target size and angle, blur, lighting and occlusion change the image. A larger cluster spans more pixels at the same distance. Synthetic shrinking does not reproduce all those effects.
INSPECT THE ACTUAL TEST INPUT
See the pixels behind a result.
Choose a source and target width, then inspect either detector’s saved boxes. Each displayed image is checked against its recorded SHA-256 hash. Magnifying a tiny target makes its pixels larger; it does not restore missing detail.
This saved result
Image identity and measured geometry
Bring the difficult cases back to review.
The useful next evidence is a corrected image: a tiny mushroom, a missed cluster or a convincing leaf. Human and AI observations remain separate, and no result on this page activates a new model.