Feature keypoints double-scaled for frames larger than featureImageMinSize (marker mis-localized at ~2x position)

[kalwalt]

Summary

WebARKitTracker mis-localizes the marker for any frame larger than featureImageMinSize (640x480): the homography and solvePnP pose place tracked content at roughly 2x the marker's true position. Typical 640x480 webcams are unaffected, which is why this has gone unnoticed.

Root cause

WebARKitTrackerImpl::initialize() computes _featureDetectScaleFactor from an image-pyramid downsample level (_featureDetectPyrLevel), intended to scale keypoints detected on a downsampled image back up to full-frame coordinates.

But the downsampling is disabled — the pyrDown block in resetTracking() is commented out and extractFeatures() runs on the full-resolution frame. The matched frame keypoints (already full-frame) are then still multiplied by the factor in MatchFeatures():

finalMatched1[i].pt.x *= _featureDetectScaleFactor[0];
finalMatched1[i].pt.y *= _featureDetectScaleFactor[1];

With featureImageMinSize = 640x480:

Frame	pyrLevel	factor	effect
640x480	0	1.0	OK (hidden)
2000x1500	1	2.0	every keypoint doubled
1920x1080	1	2.0	broken

Doubled keypoints -> homography fit to doubled points (degenerate/exploded bbox) -> solvePnP fed doubled image points -> wrong translation -> content rendered at ~2x position.

Evidence

Centroid of matched reference vs frame keypoints on a 2000x1500 frame: ref(1034,908) -> frame(1816,1250), while the marker physically sits at ~(908,625). (1816,1250) = 2 x (908,625).

Fix

While detection runs full-res, the factor must be identity:

_featureDetectScaleFactor = cv::Vec2f(1.0f, 1.0f);

(Keep the pyramid-derived computation commented out; restore it together with the pyrDown detection path if downsampled detection is re-enabled.)

Follow-up

The disabled downsampling block (// This if cond. doesn't works as expected in artoolkitx ...) and the related _featureDetectScaleFactor machinery should be reconciled with the ArtoolkitX OCVT reference: either restore downsampled detection (perf) with a consistent rescale on both the match side and createFeatureMask, or remove the now-vestigial scale path entirely.