2.2 · Coordinate Frames and Camera Geometry
A numerically valid pose can be physically wrong when frames, handedness, transform direction, calibration, or image preprocessing differ. Geometry must be tested as part of the data pipeline.
- Reason about rigid transforms and frame direction.
- Connect camera pixels to physical rays and coordinates.
- Design round-trip checks that expose frame and calibration errors.
Rigid transformations
A transform combines rotation and translation from one named frame to another. Composition order matters. Every transform should record source frame, target frame, convention, units, and timestamp.
Round-trip verification
Transform a known point into another frame and apply the inverse. Recovery of the original point tests both ordering and frame labels. Use physical fixtures or calibration targets—not only synthetic unit tests.
Camera geometry
Image coordinates depend on intrinsics, distortion, extrinsics, resize, and crop. When pixels are transformed, effective camera intrinsics must be updated. Old calibration paired with new pixels creates systematic physical error.
Frame labels and calibration are learned-system inputs, even when they never appear as tensors.
