Lab · Geometry
Moving the camera or moving the object?
A point can acquire new coordinates without moving in the world. Follow four landmarks through a change of frame, then inspect what happens when the camera transform is used in the wrong direction.
01 The world
A, B, C and D are the same four vertices in every view. The outlined frustum marks the camera’s field of view; the axes stay fixed in the world.
Object0° yaw 0.00 m shift
Camera0° yaw 0.00 m shift
Try this: move the camera right. Which coordinates remain unchanged?
02 The camera image
360 × 240 pxLet C (the camera pose, not landmark C) map camera coordinates to the world, and p be a world point. The image uses C−1p.
All four landmarks in view.
03 The same landmarks
| Point | World | Camera |
|---|---|---|
| A | -0.75-0.65-2.50 | -0.75-0.65-3.50 |
| B | 0.90-0.65-2.75 | 0.90-0.65-3.75 |
| C | -0.25-0.65-3.80 | -0.25-0.65-4.80 |
| D | 0.100.95-3.10 | 0.100.95-4.10 |
The object has yaw 0° and a world-x shift of 0.00 m. Changing its pose changes both its world and camera coordinates; the camera stays fixed. Yaw turns the object about its own center.
The object has yaw 0° and a world-x shift of 0.00 m. Changing its pose changes both its world and camera coordinates; the camera stays fixed. Yaw turns the object about its own center.
Conventions and the calculation
World axes are right-handed, with +y up. The camera’s local +x points right, +y up, and −z forward. Positive yaw follows the right-hand rule about +y. The object turns about (0, 0, −3) m; the camera turns in place, initially at (0, 0, 1) m. Each mode retains its own pose.
The camera pose C is a 4 × 4 homogeneous matrix: its upper blocks are [R | t] and its last row is (0, 0, 0, 1). Here R is the 3 × 3 rotation and t is the camera’s world position. Matrix multiplication appends a final 1 to each point. In three coordinates, the inverse gives RT(p − t). For camera coordinates (x, y, z), depth d = −z, and image coordinates are u = cx + fx/d and v = cy − fy/d.
Focal length f = 220 px; principal point (cx, cy) = (180, 120) px. Image u runs right and v down. All vertices are labeled, including those a solid face would occlude. Points outside the image are reported; points on or behind the camera are not projected. No lens distortion, physics or measured imagery is modeled.
Let T be a rigid motion expressed in the world frame. Moving the camera from C to TCgives (TC)−1p = C−1T−1p. The same camera coordinates result from holding the camera fixed and moving the object by the full inverse motion T−1, including its translation.
Blender coordinates convert as (x, y, z) → (x, z, −y). Read the coordinates and conventions or the asset provenance.
For a standard treatment of camera coordinates and the pinhole model, see Szeliski, R. (2022). Computer Vision: Algorithms and Applications, 2nd ed. Springer. Chapter 2, Image formation. szeliski.org/Book.
Next
The same question in a transformer: Rotary position embedding turns each query and key by an angle set by its position. If both move by the same number of places, does their attention score change? It is the check you just answered, for a rotation instead of a shift.