Does relative translation always have physical units?

No. Some two-view estimates provide only a direction up to an unknown scale.

Is the inverse translation just the negative vector?

Under X2=RX1+t, it is −R transpose times t, together with the inverse rotation.

Can bad matches affect pose?

Yes. Robust estimation helps, but it does not guarantee correctness under every scene or calibration condition.

AI Glossary / Computer vision and multimodal AI / Advanced

Relative pose

Definition

Relative pose describes the rotation and translation relating two coordinate frames, such as two camera views. It expresses how coordinates in one frame transform into the other under a stated convention. In calibrated two-view geometry, image correspondences can support estimation of relative camera rotation and translation direction, but ordinary monocular observations do not determine absolute translation scale by themselves.

What it means

A pose is always relative to a reference frame, even when a system calls one frame global. Relative pose focuses on the relationship between two frames rather than their separate positions in an external map. The direction of the transformation matters: a transform from camera one to camera two is not numerically identical to its inverse. Translation units and scale must also be clear. A returned unit translation vector conveys direction, not a measured one-meter movement.

How it works

Using one convention, a point X1 in the first frame maps to X2=RX1+t in the second, where R is a rotation matrix and t is a translation vector. The inverse uses R transpose and translation −R transpose times t. Camera estimation can use calibrated correspondences and an essential matrix, followed by checks that reconstructed points lie in front of the cameras. Robust estimation helps handle incorrect matches. Degenerate or weak geometry, calibration error and insufficient parallax can limit estimation. Additional information is needed when metric scale is required.

Three practical examples

A stereo-processing team uses a known relative camera pose to relate measurements from two calibrated viewpoints.
A visual-odometry system estimates inter-frame motion while distinguishing scale-free monocular translation from metric motion supplied by additional constraints.
A developer verifies coordinate transformations by mapping known points forward and back before using them in triangulation or visualization.

Think of it this way

Imagine giving directions between two rooms, each with its own compass orientation and coordinate origin. You need both an orientation change and an offset to translate a location description from one room's system into the other's. Reversing the directions requires reversing the complete relationship, not merely negating every number independently. The analogy explains frame conventions. It does not establish physical distance if the available observations only reveal the direction of movement without a calibrated scale.

Worked example

In an original planar illustration embedded in three dimensions, let R rotate 90 degrees about the z axis and let t=(1,0,0). Point X1=(2,0,0) rotates to (0,2,0), then translates to X2=(1,2,0). To recover X1, first subtract t, obtaining (0,2,0), then apply the inverse rotation, obtaining (2,0,0). Negating t alone while retaining the forward rotation would not implement the inverse transform. This example assumes a known metric translation and a stated coordinate convention; it does not demonstrate recovering metric scale from two monocular images.

Illustrative case study

A fictional inspection-robot team combines relative poses from two software components. The estimated trajectory appears to turn the wrong way because one component returns camera-to-camera coordinates in the opposite direction from the other. The team checks point transformations, reverses the relation correctly and separately verifies translation scale. It then tests the full trajectory on representative motion. This scenario illustrates convention and scale problems that can resemble poor estimation. It does not claim that a correct coordinate conversion eliminates drift, calibration error or ambiguous image correspondences.

Common misconceptions and mistakes

Relative pose is not simply the difference between two pixel positions. Rotation and translation relate coordinate frames in space. Another mistake is interpreting an essential-matrix translation direction as a metric displacement. Transform inversion also requires rotating the translation appropriately, not only changing its sign. Pure rotation and very small baselines can make translation or depth estimation difficult. Finally, a mathematically valid rotation matrix does not prove that the estimated pose matches the physical scene; correspondence quality and model assumptions still require evaluation.

Business use

Relative pose supports robotics, stereo vision, reconstruction and augmented-reality systems. Teams should document frame names, transformation direction, units and scale assumptions at every interface. Check known correspondences and round-trip coordinate conversions before diagnosing model performance. Validate under relevant motion and scene conditions, including weak texture and limited parallax. When measurements require physical distances, establish the source of scale explicitly. A pose estimate can be useful for visualization while remaining insufficiently calibrated or stable for a more demanding measurement task.

Check your understanding

Where did the example point end up?

Answer: At (1,2,0).

What two steps recovered it?

Answer: Subtract the translation and apply the inverse rotation.

Why should a unit translation vector not be read as one meter?

Answer: Its normalization can express direction without establishing metric scale.

Related terms

Essential matrix
Epipolar geometry
Camera pose
rigid transformation
triangulation
visual odometry

Learn next

Write down two named coordinate frames and transform a few known points between them. Verify the inverse algebraically and numerically. Then examine calibrated two-view estimation and identify which additional information would be needed to recover translation in physical units.

Sources

OpenCV, Camera Calibration and 3D Reconstruction, recoverPose reference: https://docs.opencv.org/4.10.0/d9/d0c/group__calib3d.html The rotated-point calculation, room-coordinate analogy and inspection-robot case are original educational illustrations. The entry states its transform convention and distinguishes known metric pose from scale-ambiguous monocular estimation.

Previous: Relative frequency view Next: Relative risk or risk ratio