Investigate: small-marker pinball-demo.jpg not detected under downsampling (works in WebARKitLib-rs & jsartoolkitNFT)

[kalwalt]

Summary

With detection-side pyramid downsampling restored (#44 part A, #52), the original pinball-demo.jpg — a small marker in a large frame — fails to acquire: detection runs on the pyrDown'd frame and the marker falls below the detector threshold (No keypoints detected!). A marker-fills-frame image (pinball-demo-big.jpg, 1920×1440) tracks fine at the same pyrLevel.

The same pinball.jpg reference + pinball-demo.jpg scene does track in:

• WebARKitLib-rs (Rust port), and
• the jsartoolkitNFT Node.js version.

So the marker is detectable in principle; something in this C++/WASM detection path (or its parameters) is weaker for small markers under downsampling.

Hypotheses to investigate

• Detector type / parameters differ from the Rust and NFT paths (TEBLID feature count TEBLID_MAX_FEATURES, FAST/AKAZE thresholds, featureImageMinSize).
• Pyramid level chosen here is too aggressive for a small marker — the marker's effective size after pyrDown is below the descriptor's minimum.
• Reference-image preprocessing / scale handling differs between the implementations.
• nn_match_ratio / minNumMatches gating differences.

Suggested next steps

1. Compare detector config (type, max features, thresholds, featureImageMinSize) across WebARKit (C++/WASM), WebARKitLib-rs, and jsartoolkitNFT.
2. Log keypoint counts on the downsampled detectionFrame for pinball-demo.jpg at each pyrLevel; confirm whether the marker simply yields too few keypoints post-pyrDown.
3. Evaluate whether featureImageMinSize should be larger (less aggressive downsampling) or whether small-marker scenes warrant a different floor.

Context

Filed as a follow-up to #44 (part A landed in #52). Not a regression in #52 — full-resolution detection on pinball-demo.jpg also exercises this; downsampling just makes it more visible.