ADR 0010: The IMU-to-lens alignment and frame timing

Status: accepted (2026-10-19), measured on the ONE R

Context

The attitude integrated from the gyro and accelerometer (ADR 0004, 0005) is in the IMU's axes, and the lens poses are in the calibration frame (ADR 0007). Stabilizing a stitched frame chains them (ADR 0006): an output direction goes to the world by the stabilized view, into the body by the inverse attitude, and into the lens frame. That last step needs the rotation from the IMU's axes to the lens frame, which the recording does not state. The frame also has to be stabilized with the attitude of the instant it was captured, so frames and motion must share a clock.

Decision

• CameraRecording.imuToCamera carries the IMU-to-lens rotation, a per-camera-model constant the reader supplies. motion/outputToCamera(imuToCamera, attitude, view) composes the per-frame stitch rotation from it.
• It is measured from the footage, among the 24 axis alignments. An IMU is mounted square to the body, so the rotation is one of the cube's rotations (motion/axisAlignments). With the right one, locking the view to the world makes two frames a couple of seconds apart show the same scene; with a wrong one, the view turns between them. The developer harness ranks the candidates by the correlation between those two locked renders (WEB360_PREVIEW_IMU_SEARCH).
• Measure where the camera turns hard. In steady freefall the camera turns only a few degrees in two seconds, and every candidate looks still. Canopy turns (90 to 130 degrees in two seconds, with the ground in view) separate them: on ONE R footage a quarter turn about X (the IMU's z along the lens frame's y) wins in five of six such windows, and stabilizing with it levels the horizon with the canopy overhead.
• A frame shown t into the video was captured at captureClock + t on the motion clock. The trailer's exposure records step at the video's 29.97 frames per second and include a few pre-roll frames, and captureClock matches the record of the first video frame to a few microseconds. A swept offset between the clocks moved the stillness by at most a frame's worth (20 ms), so none is applied.

Consequences

• A world-locked view holds a correlation near 0.75, not near 1, through canopy turns. The limit is most likely what is close to the camera (the canopy, its lines, the camera's mount), which no rotation keeps still, and rolling shutter, which this measure does not correct (ADR 0011); the alignment itself is unambiguous where the camera turns.
• Another camera model needs its own measurement; the harness search is the tool, and a recording with hard turns over distant scenery is the input.