Measure object size from a camera image

A detected object has an image width, but that width does not establish its physical dimensions. The same part occupies different numbers of pixels when its distance, orientation or imaging geometry changes. For inspection engineers, the first decision is therefore not which detector to use, but which geometric assumptions the station can maintain. You can then choose a fixed scale, planar calibration, telecentric imaging or a depth-aware method, and verify the dimensions against independent references.

Choose the geometry before the measurement tool

To measure object size from a camera image, you need a known relationship between image coordinates and physical coordinates. A calibrated reference can establish that relationship, but only within the conditions represented by the calibration.

Start by identifying the surface containing the dimensions you need. A label lying flat on a conveyor presents a different problem from a cylinder whose top edge changes height.

The practical starting point for a flat inspection is the guide to measuring planar distances with a calibrated camera. It explains the coordinate mapping that connects visible features to a measurement plane.

Separate detection from dimensional measurement

An object detector typically reports a class, confidence and image coordinates. Those coordinates identify where to inspect; they do not establish a physical scale.

A bounding box also may not represent the engineering dimension you want. An axis-aligned box grows when a rectangular part rotates, even when the part itself remains unchanged. Shadows, occlusion and inconsistent training annotations introduce further differences between the box and the physical boundary.

Use detection to locate a region, then measure suitable features inside it. Depending on the part, those features might be fitted straight edges, circle boundaries or specific reference marks. A segmentation outline can help, but its boundary still needs validation against the required dimensional accuracy.

Keep coordinates consistent throughout the pipeline. If detection uses a resized or padded image, reverse that transformation before applying calibration from the original image. The same requirement applies to cropped regions: restore their coordinate offsets before mapping points into physical space.

Establish when a fixed scale is valid

A fixed scale works when the relevant surface stays at a fixed distance and orientation, and scale variation across the measurement region is acceptably small. Lens distortion must either be corrected or demonstrated to be negligible for the required accuracy.

Measure a known reference in the same plane as the inspected features. For a simple perpendicular-view estimate:

Suppose a reference spanning 100.00 mm measures 2000 pixels. The scale is 100.00 / 2000 = 0.05000 mm per pixel. An object spanning 760 pixels therefore measures 760 × 0.05000 = 38.00 mm.

That arithmetic does not prove the station is accurate to the displayed decimal places. Reference uncertainty, edge localization, distortion and height variation still contribute to the error.

Check reference lengths near the center and corners rather than calibrating from one central span alone. For the optical relationship between sensor sampling and magnification, use the explanation of pixel scale calculated from magnification.

Correct perspective on a flat measurement plane

An angled camera produces position-dependent scale across a plane. A single millimeters-per-pixel value cannot generally convert every horizontal and vertical image distance correctly.

For a planar target, a homography maps image points to coordinates on the target plane. Establish the mapping from known, well-distributed reference points, usually after correcting lens distortion. Use more than the mathematical minimum when practical, and reserve additional points for independent verification.

Transform the measurement endpoints into plane coordinates, then calculate their separation:

Do not convert an image-space bounding-box width using a scale measured somewhere else in the image. Mapping the relevant endpoints addresses perspective more directly and avoids confusing projected width with the intended feature length.

The guide to length calibration using image markers covers the reference geometry behind this approach. Remember that a plane mapping does not become a three-dimensional measurement merely because its calibration fits well.

Quantify the error caused by object height

With a conventional lens, a feature moving toward the camera usually appears larger. Calibrating at the conveyor surface and measuring a raised top face therefore introduces a predictable scale error.

For an ideal pinhole model, let D be the calibration-plane distance from the effective projection center. Let h be the feature height toward the camera:

This is an estimate that omits lens distortion and other optical effects. D is not automatically the working distance measured from the front of a lens.

For example, take D = 500 mm and h = 10 mm. The scale factor becomes 500 / 490 = 1.0204. A true width of 50 mm would be reported as approximately 51.02 mm, an error of about 1.02 mm or 2.04%.

That error can dominate any benefit from more pixels or subpixel edge fitting. If height is known and the camera model is valid, geometry can sometimes compensate. If height is unknown, use a fixture, measure depth, add another view or reconsider the optics.

Surface tilt introduces another limitation: the visible projection of a sloped feature is not its full surface length. Neither a fixed scale nor a planar mapping recovers an unknown out-of-plane slope.

Use telecentric optics for the right reason

Telecentric lenses reduce magnification changes as an object moves along the optical axis within the lens's useful operating range. They are valuable when projected width or diameter must remain stable despite limited height variation.

Telecentric imaging does not remove every measurement error. Finite telecentricity, distortion, defocus, illumination changes and part tilt still matter. It also does not reveal hidden surfaces or recover arbitrary three-dimensional dimensions.

Use the guide on when telecentric optics are appropriate for gauging to decide whether the geometry justifies that approach. Then validate both image sharpness and dimensional stability through the expected height range, following the distinction explained in telecentric focus range and measurement accuracy.

A sharply focused image alone is not proof that magnification is sufficiently stable for your tolerance.

Budget sampling separately from accuracy

For a perpendicular view with approximately uniform scale, field width divided by horizontal pixel count gives nominal sampling. This helps size the camera, but it is not a complete accuracy specification.

Suppose the required field width is 120 mm and the selected sampling target is 0.05 mm per pixel:

The sampling target must leave room for edge uncertainty and other error sources. As an illustrative assumption, if each endpoint can shift by 0.2 pixel, the worst-case separation error from those shifts is 0.4 pixel. At 0.05 mm per pixel, that contribution is 0.4 × 0.05 = 0.02 mm.

This example assumes the endpoint behavior has actually been established. It does not claim that any edge algorithm routinely achieves that value. Correlated errors, asymmetric illumination and uncertain physical boundaries can produce bias that averaging does not remove.

The discussion of camera resolution and measurement accuracy in motion explains why pixel count must be considered alongside blur and acquisition conditions.

Choose camera and lens candidates without assuming accuracy

Select area scan cameras by the required pixel dimensions, shutter behavior, triggering and acquisition needs. The camera finder helps narrow those choices, but catalog resolution alone does not establish dimensional performance.

The MV-CU020-90GC is a global-shutter, C-mount camera candidate. Application suitability still depends on sampling, exposure and the complete optical setup.

For machine vision lenses, check sensor coverage, mount, field coverage, working distance and usable image quality. Browse the lens finder only after defining the measurement region and available space.

The LS-FFL-0679 is a C-mount machine vision lens candidate. A conventional lens can suit a constrained measurement plane when the resulting errors fit the application budget.

The LS-TCL-1390 is an object-side telecentric lens candidate. Consider it when reduced sensitivity to axial position is relevant to the measurement.

These parts illustrate component choices, not a validated measuring system. Confirm the complete combination against the required field, detail, height range and tolerance before committing to hardware.

Keep lighting stable during evaluation. An edge threshold that moves with glare or exposure can change the reported dimension even when the mechanical geometry remains fixed.

Validate dimensions against independent references

Calibration estimates the mapping; validation tests whether it is adequate for the inspection. Do not use only the same reference points that fitted the mapping to judge its performance.

Use a short acceptance checklist covering the operating envelope:

• Reference quality: use known dimensions with uncertainty comfortably below the measurement requirement. Record the reference and its condition.
• Field position: check the center, corners and any region where parts will actually appear. Look for location-dependent bias.
• Size range: test short and long dimensions. A mapping can appear satisfactory on one span while failing elsewhere.
• Height range: evaluate the lowest and highest permitted feature planes. Include fixture variation rather than testing only the nominal position.
• Orientation: rotate representative parts through expected angles. Confirm that the algorithm measures the intended dimension, not an enclosing box.
• Repeatability: capture, for example, 30 stationary images, then repeat after removing and replacing the part. Keep imaging variation separate from positioning variation.
• Motion and lighting: repeat at operating speed and under expected illumination changes. A stationary bench check does not validate moving inspection.
• Configuration control: record focus, aperture, camera settings, image dimensions and calibration version. Revalidate after changes that affect geometry or edge appearance.

Report signed error against the reference as well as repeatability. A tight cluster of readings can still be consistently wrong.

For an illustrative maximum-error budget of 0.20 mm, you might allocate 0.10 mm to residual geometry, 0.05 mm to edge extraction and 0.05 mm to remaining contributions. Those allocations sum to 0.20 mm; they are design limits to verify, not measured performance or a statistical uncertainty calculation.

Recognize when one calibrated image is insufficient

The usual fixed-camera answer fails when the part can change distance unpredictably, rotate freely in three dimensions or hide the features defining its true dimension. Adding a more capable detector does not supply the missing geometry.

A known-size reference can help only when its geometric relationship to the measured feature is established. A marker on the table does not automatically calibrate the top of a tall object. Likewise, estimating range from an assumed object size cannot independently verify that same size.

Choose the simplest constraint that makes the measurement observable: flatten the measurement plane, control height, add depth information or introduce another calibrated view. Define which physical dimension is being measured before choosing the algorithm.

For a fixed inspection station, start the lens configuration process with your field dimensions, tolerance and expected height variation. If the geometry remains uncertain, send those details through applications support before selecting components.