Fix CV->GL pose conversion in updateTrackable()

[kalwalt]

Summary

Stop applying a partial CV→GL conversion in WebARKitPatternTrackingInfo::updateTrackable(). Today the rotation columns Y and Z are negated but the translation isn't — an inconsistent half-conversion that leaves consumers with a pose where translation Z stays positive. Under a standard OpenGL projection (which expects objects in front to have negative Z in view space), tracked objects clip behind the camera and never render.

After this change, getPoseMatrix2() / trans carry the raw OpenCV pose (rotation unmodified, translation kept at the existing 0.001 * 1.64 scale). The CV→GL handedness flip (negate Y and Z rows including translation) is then applied uniformly downstream by arglCameraViewRHf, producing a mathematically clean diag(1, -1, -1, 1) * pose modelview that places the tracked object in front of the GL camera.

Both downstream paths benefit:

• JS side: WebARKitController.process_raw → arglCameraViewRHf(pose, ...) → correct matrixGL_RH.
• C++ side: WebARKitManager::getGLViewMatrix() → arglCameraViewRHf(getPoseMatrix2(), ...) → correct modelview.

Surfaced and verified downstream by webarkit/webarkit-testing#31 / #32 (static-image Teblid example): with this fix the AR overlay is no longer frustum-clipped behind the camera.

Breaking change note

This changes the runtime semantics of pose / trans for any consumer that relied on the previous partially-converted shape. Consumers rendering with OpenGL/three.js should now use matrixGL_RH (or getGLViewMatrix() in C++) as the GL-ready modelview, and treat pose as the raw OpenCV camera pose ([R | t], +Z forward).

Diff

A single 4-line hunk in WebARKit/WebARKitPattern.cpp (drop two - signs, plus a comment explaining the convention).

Test plan

• Existing gtests still pass (no projection-matrix value asserted against this code path).
• Downstream: rebuild WASM, load examples/threejs_teblid_static_image_ES6_example.html from Add static image example with Teblid tracker (#30) webarkit-testing#32 (after the companion PR), confirm sphere appears anchored to the marker rather than clipped behind the camera.

Related

• Resolves the rendering side of Tracked pose renders behind the camera (OpenCV vs OpenGL convention); matrixGL_RH throws #34 (pose convention).
• A separate projection X-mirror remains, tracked in Asymmetric -2*f_x/screenWidth in cameraProjectionMatrix — origin of X-mirror unclear #35.
• Downstream PR consuming this: webarkit/webarkit-testing (will be posted as a follow-up).

🤖 Generated with Claude Code