Can there be several evaluation baselines?

Yes. A simple reference and the current production method can answer different comparison questions.

Is baseline the same as benchmark?

A benchmark is an evaluation setup or reference task; a baseline is a method or result used within a comparison, although everyday usage can overlap.

Does the stereo formula apply to every raw camera pair without preparation?

No. The simple form assumes appropriate calibration and rectified geometry with consistent units.

AI Glossary / Graphs and symbolic AI / Beginner

Baseline

Definition

Baseline has two important AI-related meanings in the supplied sources. In model evaluation, it is a reference method or result used for comparison. In stereo vision, it is the separation between the two camera positions used to estimate depth. The intended meaning must be stated.

What it means

An evaluation baseline answers, “What would happen using a simpler or established approach?” It could be a majority-class classifier, a seasonal forecast, an existing production system or a carefully reproduced earlier method. A new model's score becomes more informative when compared with that reference under equivalent conditions.

A stereo baseline is a physical geometric quantity, usually expressed as a distance. The two views of a scene differ because the cameras occupy different positions. Combined with camera calibration and matched image points, that separation helps recover depth.

These senses are unrelated despite sharing a word. Saying that a team “increased the baseline” could mean changing camera spacing, while “beating the baseline” usually means improving a measured result. Documentation should avoid leaving the reader to guess.

How it works

For evaluation, select a reference appropriate to the task and measure it on the same relevant data, metric and operating conditions as the candidate. Keep training and evaluation information properly separated. A deliberately weak comparison can exaggerate the apparent value of a sophisticated model.

For stereo depth, consider a calibrated, rectified pair of ideal cameras. The common relationship is depth Z = fB/d, where B is baseline, f is focal length expressed in compatible image units and d is disparity between matched points. Camera geometry and coordinate conventions matter; the physical baseline alone is insufficient.

Both meanings require careful reporting, but for different reasons. An evaluation result needs a reproducible comparison protocol. A stereo setup needs calibrated dimensions and valid correspondences. More camera separation can increase disparity for a given depth, yet practical visibility and matching difficulties mean that larger is not automatically better.

Three practical examples

A demand-planning team compares its learned forecast with a seasonal reference that predicts the corresponding period from the previous cycle.
A text classifier is evaluated against a simple established method to assess whether added complexity improves the relevant outcome.
A stereo inspection system records camera separation as part of the calibration used to convert image disparity into depth.

Think of it this way

For evaluation, think of a control route used to compare a proposed shortcut: both routes should be assessed for the same journey. For stereo vision, think of viewing a nearby object first with one eye and then the other; the viewpoint separation creates a visible shift.

Neither analogy is complete. A fair model comparison needs more than one journey, and depth recovery requires calibrated geometry rather than simply noticing that the image moved.

Worked example

Two calculations with different meanings

First consider an invented binary dataset with 900 negative cases and 100 positive cases. A classifier that always predicts negative achieves 900/1,000 = 90% accuracy. A new model achieving 92% accuracy improves that metric by 2 percentage points, but the comparison alone does not reveal how well it finds positives. Recall, precision and operational costs may still determine usefulness.

Now consider a stereo pair with baseline B = 0.12 meters and focal length f = 800 pixels. A matched point has disparity d = 24 pixels. Under the stated rectified-camera model, depth is 800 × 0.12 / 24 = 4 meters.

The 90% reference and the 0.12-metre separation are both baselines in their respective contexts. They have different units, purposes and validation requirements; combining them under one unexplained field name would be misleading.

Illustrative case study

Fictional case: Maple Vision develops a stereo-based counting system. Its project dashboard contains a field called “baseline,” used by one team for camera spacing and another for the existing counting model's performance.

A review uncovers that experiment notes mix the two meanings. The team renames the fields “camera separation” and “reference-model score,” records their units and links them to separate calibration and evaluation records. It also discovers that the new model had been compared with an older reference on a different image set and repeats the comparison consistently.

The lesson is that common vocabulary can conceal different technical quantities. Clear naming and matched evaluation conditions are small changes that make results much easier to interpret.

Common misconceptions and mistakes

A baseline is always a trivial model. It can be a strong existing system or prior method.
Beating one baseline proves business value. The metric, uncertainty, costs and deployment conditions still matter.
Accuracy above a majority baseline guarantees useful minority-class performance. The error distribution must be examined.
Stereo depth depends only on camera separation. Focal length, disparity and calibration are also required.
A larger stereo baseline is always preferable. Matching, occlusion and operating distance create tradeoffs.

Business use

In evaluation, a credible baseline makes claims of improvement understandable and helps justify complexity. Establish it early, preserve the comparison conditions and report meaningful absolute as well as relative changes.

In stereo projects, baseline is part of the sensing design and calibration record. Changes to the camera arrangement require appropriate recalibration and validation. For multidisciplinary teams, specify the meaning and unit wherever the term appears. This avoids confusing a comparison standard with a physical parameter that directly affects depth estimates.

Check your understanding

A report says “baseline = 120” without context. Is that enough to interpret the result?

Answer: No. It could be a camera distance, a metric value or something else. The report must identify the quantity, units, measurement method and relevant comparison or geometric assumptions.

Related terms

Benchmarks
Evaluation metric
Disparity
Reference model
benchmark
majority-class classifier
stereo vision
focal length
camera calibration

Learn next

For a familiar prediction task, define one simple reference and one established comparator, then specify a fair evaluation protocol. Separately, practice the stereo depth formula with stated units and inspect how disparity changes with depth. Keep the two meanings in distinct notes or glossary cross-references.

Disparity

Sources

Google Machine Learning Glossary, baseline, supplies the evaluation-reference sense: https://developers.google.com/machine-learning/glossary#baseline

AIMA glossary, baseline, supplies the binocular-stereo sense: https://github.com/aimacode/aima-glossary/blob/master/glossary.md?plain=1

OpenCV, Depth Map from Stereo Images, gives the focal-length, baseline and disparity relationship: https://docs.opencv.org/5.0/py_tutorials/py_calib3d/py_depthmap/py_depthmap.html

Previous: Base relation Next: Basic emotions