Computer Vision

Stereo baseline

ActivePublished: 28 October 2026Updated: 28 October 2026Published

Key innovation

Reduces 3D depth perception to a single measurable geometric parameter — the distance between the two cameras' optical centers — which linearly scales measured depth and sets the range-versus-accuracy trade-off.

How it works

The two cameras capture the same scene from slightly different positions separated by the baseline B. After calibration and rectification, corresponding points lie on the same image rows and their horizontal shift is the disparity d. From similar-triangle geometry, Z = f·B/d — depth is inversely proportional to disparity and directly proportional to the product of focal length and baseline. Depth error grows with the square of distance and shrinks with the baseline: ΔZ ≈ (Z²/(f·B))·Δd, where Δd is the disparity-estimation uncertainty. Increasing the baseline (or focal length) therefore improves accuracy at long range but simultaneously raises the minimum measurable distance and enlarges regions visible to only one camera (occlusions).

Problem solved

Passive stereo depth estimation must turn two 2D images into a 3D distance. The stereo baseline controls how sensitive that mapping is and over what range it works — too small a baseline yields noisy, coarse depth at distance, while too large a baseline makes near objects unmeasurable. Choosing the baseline addresses the problem of matching depth range and accuracy to a specific application.

Implementation

Implementation pitfalls

Too large a baseline — near-field dead zoneHigh

A large baseline increases the minimum measurable distance and occlusion zones; near objects produce disparity exceeding the search range and get no depth estimate.

Fix:Size the baseline to the closest required distance or increase the disparity search range.

Too small a baseline — poor depth at rangeHigh

A small baseline yields small disparity for distant objects, making depth noisy and heavily quantized.

Fix:Increase the baseline or focal length for the required long range.

Miscalibrated baseline — linear depth scale errorHigh

Because Z is proportional to B, an error in the measured baseline directly produces a proportional scale error in all distances.

Fix:Perform precise extrinsic calibration and periodically verify the rig geometry.

Hyperparameters (configurable axes)

Baseline distance (B)Critical

Metric distance between the cameras' optical centers. Linearly scales measured depth and sets the range-versus-accuracy trade-off.

~75 mmTypical baseline of a consumer stereo-camera module (e.g. OAK-D).

dziesiątki cm – metryLarge baselines used for long-range outdoor/automotive stereo.

Focal length (f)High

Focal length in pixels; together with the baseline it forms the numerator of the depth equation Z = f·B/d.

Disparity search rangeHigh

Maximum disparity considered by the matching algorithm; bounds the minimum measurable distance, which grows with the baseline.