Window mark off by up to 5.4 in and gas meter by 4.8 in against the tape #30

Activity

1. AidenJohnston

   added

   bugSomething is broken or behaving incorrectly.Something is broken or behaving incorrectly.

   needs-infoWaiting on the author for details needed to review.Waiting on the author for details needed to review.

   size:M30-99 effective changed lines.30-99 effective changed lines.

   on Sep 26, 2026
2. AidenJohnston commented on Sep 26, 2026

   

   AidenJohnston

   on Sep 26, 2026

   CollaboratorAuthor

   More actions

   I checked the tap path at the t3/ios-mvf head (5c8c739) against run 1. The existing data can't prove a bug: the run 1 scan and tape readings aren't in the repo, and run 2 (#46) didn't read tape distances aloud. So there's no PR. Below are what I ruled out, the likely causes ranked, and what the next run needs to record to tell them apart.

   How a mark is placed. A feature tap doesn't raycast against anything real. markFeaturePoint maps the view point to a sensor pixel, casts a ray from the last sampled camera pose, and intersects it with the analytic wall plane. That plane goes through the meter tap's hit and faces that hit's normal. The mark's position along the wall is then the perpendicular projection onto the wall line from the meter tap point. Only the meter tap raycasts, against a detected plane first and ARKit's estimated plane otherwise (L31–38).

   Ruled out

   • Rotated image with unrotated intrinsics. CameraProjection.imagePixel(forViewPoint:) undoes the 90° clockwise, aspect-fill display correctly. The intrinsics and image size both come from the landscape ARFrame.camera. CameraFrame.ray(throughPixel:) is the exact inverse of pixel(of:).
   • Tap coordinates. The tap is read in global coordinates, and the view size comes from a full-screen CameraSizeReader, so they match.
   • A fix since run 1. None of the tap path has changed since the run 1 build (657ab28).
   • Scale error or a meter tap off the centre line. Either one moves both window edges the same way. Run 1 has the left edge at +5.4 in and the right edge at −0.5 in, so the edges moved in opposite directions and the window came out 5.9 in narrower. The right edge landing within half an inch suggests any offset shared by both edges is small.

   Likely causes, most likely first

   1. Parallax against the analytic plane. Anything that isn't on that plane moves sideways by its distance from the plane times the ray's sideways slope. That's a pull toward where the phone stood, or a push away from it.
      • Gas meter. The export already assumes a tapped point object stands 0.3 m (about 12 in) off the wall. A tap on its face lands on the wall plane behind it, shifted away from the phone. A 12 in deep object viewed 22° off square gives 12 × tan 22° ≈ 4.8 in, the size of the error. That makes this the top suspect for the gas meter.
      • Window. The pattern fits a wall plane that sits in front of the real wall at the window, or is turned a few degrees. That happens if the meter tap's estimated plane landed on the meter's face, or if its normal is off. A plane in front of the wall pulls both edges toward the phone. With the phone near the right edge, the left edge moves right a lot and the right edge barely moves, which is exactly what run 1 shows. At 7 ft, a 3° turn puts the plane about 4.4 in off.
   2. A bad first corner tap. The left edge was the first tap, and one sample can't separate aim from geometry.
   3. Tracking drift or anchor corrections. When the meter anchor moves more than 2 cm, refreshMeterFromAnchor moves the wall origin. The tapped world points stay put, though, so a correction after a tap shifts that whole feature. That moves both corners of one feature together, so it can't shrink the window, but it could shift the gas meter. The evals (feat(evals): real-data evals for AR drift, wall reconstruction, coverage and the field test #12) put a modern iPhone's position error at a p90 of 8.6 in after 10 ft.
   4. A stale pose. Taps use the last sampled frame, every 6th camera frame (about 0.1 s). That's only about 1 in at 5 ft, unless the phone is panning during the tap.

   What to record in the next field test

   • Share the scan off the phone. The export's meter plane source (detected or estimated) says whether the meter tap used a real wall plane.
   • Say, per tap, where you stand. Give the distance out from the wall and the distance left or right of the meter. Take each feature twice: once square on, and once from about 45° to the side.
     • If it's parallax, the error changes sign with the side you stand on and grows with the angle.
     • If it's a bad tap, the error is random.
     • If it's drift, both taps of one feature shift together, and the shift grows with walk time.
   • Tape the depths. Measure how far the gas meter stand-in's face stands off the wall, and how far the window glass and trim are set in or out. Then the parallax can be computed and compared.
   • Say what each window corner tap is on. Name it aloud: the trim's outer edge or the glass. Tape to the same edge.
   • Tap the meter again at the end of the walk. Also re-tap one window corner. The difference from the first taps measures drift directly.

   If the scan shows that the error flips with the side you stand on, the fix is to raycast feature taps against the mesh or estimated surface and project the hit onto the wall line, instead of intersecting the analytic plane.
3. SamGu-NRX

   added a commit that references this issue on Oct 15, 2026

   Add meter and panel reads (WS-08) (#30)

   ...

   f78ac97