You quote 127° on the sensor spec, mount the stereo camera, measure the depth output and find barely 80° of usable view. Nothing is broken. Rectification keeps only the pixels both cameras can see and correlate, so the depth FOV is the pair's overlap, not the FOV of either sensor. This page gives the loss to expect for wide-angle and narrow lenses, the alpha setting that trades crop against edge quality, a worked sizing example for a pallet lane, and a measurement method built on feature-rich targets that returns the real angle instead of an artefact of a blank wall. By the end you can size the mount from the depth FOV and know which spec line to ask for before you order.
Design to the rectified depth FOV, not the sensor FOV
Take a representative wide-angle stereo head: two wide mono sensors on a fixed baseline with an IR pattern projector. Its published figures give 127° horizontal and 79.5° vertical for the sensors, and 106°/97°/70° diagonal/horizontal/vertical for the rectified depth image — a real spec sheet, used here as the running example. The depth numbers are what your application receives: 97° against 127° is a loss of 30°, roughly a quarter of the horizontal angle; 70° against 79.5° gives up 9.5° vertically. A mount computed from the sensor line alone ends up wrong by exactly that gap.
How much will rectification take from your camera? Distortion grows with FOV, and the crop grows with distortion, so treat the loss as roughly proportional to angle. A 127° lens gives up about a quarter; a mild 60° stereo lens may lose only a few degrees. Ask any vendor for the rectified depth FOV, not the sensor FOV. If only the sensor line exists, measure your unit once with the method further down before you drill mounting holes.
Where the degrees actually go
Rectification re-projects both images as if each came from an ideal, distortion-free camera with aligned scanlines, because the correlator needs matching features to sit on the same row. Straightening a wide lens bends its borders back hardest, and the output is cropped to what both straightened images share. Wider lens, more distortion, bigger cut — that is the whole mechanism.
Second, matching needs consent from both cameras. A pixel keeps depth only where the other sensor also sees the same surface, so after alignment the depth image is the intersection of the two views. This is why the depth stream covers less than the RGB stream on the same head: the colour sensor is a separate eye with its own spec, and stereo coverage comes out smaller than either.
Third, the baseline itself. Both cameras must see a point from their own positions, and at short range the rig's separation hides the extreme edges from one of the pair, so usable depth narrows further as working distance falls. The same geometry produces occlusion bands when depth is aligned to a colour image. Give close-range jobs more standoff than the FOV arithmetic alone suggests.
The projector has no say in any of this. A pattern cone quoted at 78°/61° (±7%) says where the projector paints texture, not where the cameras can match: aim the pair at a textured scene and valid depth appears outside the pattern, while a blank wall inside it can return nothing. The projector is a texture source, not a field stop. Under bright ambient light the pattern washes out, and depth shrinks to wherever real scene texture lives.
The alpha setting: degrees back, edge quality paid
Rectification lets you choose how hard the crop bites. The stereo node's alpha setting moves between two extremes: at alpha 0 the output is the largest rectangle in which every pixel is valid — the smallest but cleanest view; at alpha 1 the full undistorted image survives — degrees recovered at the borders, paid for in edge pixels that were resampled hardest and that some SDKs mark invalid. Vendors quote the rectified depth FOV at the conservative default, so read their number as the safe figure.
Worked example: covering a 3.0 m pallet lane with a 97° depth FOV
Inputs, each with its source:
- Rectified depth horizontal FOV: 97° — the running example head's rectified-depth spec.
- Lane width to cover with depth: 3.00 m — the dimensioning station's requirement.
- Standoff Z: measured from the camera's optical centre, not the bracket face; where Z is measured from changes the answer, as in choosing a stereo world origin.
The rule, in words: a camera with horizontal FOV θ covers a width W = 2·Z·tan(θ/2) on a plane at distance Z; inverted, the standoff a lane needs is Z = W / (2·tan(θ/2)).
Arithmetic. Depth coverage at a 1.00 m test distance: 2 × 1.00 m × tan(48.5°) = 2.26 m. Sensor-spec coverage at the same distance: 2 × 1.00 m × tan(63.5°) = 4.01 m — the number that tempts a low mount. Standoff that actually covers 3.00 m with the depth FOV: 3.00 m ÷ (2 × 1.130) = 1.33 m. The mount the sensor line would have suggested: 3.00 m ÷ (2 × 2.006) = 0.75 m, where the depth FOV spans only 2 × 0.75 m × 1.130 = 1.70 m and the outer 0.65 m per side of the lane returns no depth. Vertical check: the 70° depth FOV spans 2 × 1.33 m × tan(35°) = 1.86 m, which brackets a 1.8 m pallet only with the optical axis at pallet mid-height.
Decision: hang the camera at 1.47 m — 1.33 m plus a 10% margin — where coverage is 2 × 1.47 m × 1.130 = 3.32 m against the 3.30 m requirement, and re-check far-edge depth error, because stereo depth noise grows with the square of range at fixed lens and baseline. Threshold for any layout: if depth coverage at your standoff is under the lane width plus 10%, add standoff or change the measurement device.
Measuring the depth FOV so the number is real
A spec line settles arguments; a tape-and-tags measurement settles them faster, because the correlator needs texture and a plain wall under-reports — one wide head read 80° against its 97° rectified spec until targets with features went up.
- Tape AprilTag or QR markers across a plane at a measured distance Z — 1.00 m is convenient — spanning more than the expected angle on both axes, and record each marker's position.
- Set the stereo node's alpha to the value production will run.
- Capture a single depth frame or point cloud and note which markers return valid depth.
- Take the outermost valid markers per axis and read W as the taped distance between them; FOV = 2·atan(W/(2Z)).
- Repeat at two or three distances from 0.5 to 3.0 m. The angle should hold steady past about 0.5 m; a smaller figure close in is the baseline parallax from earlier doing what it always does.
Three traps cause nearly every wrong measurement. A uniform target under-reads, as above. The conversion can halve the answer: FOV is the full angle, and a factor-of-two slip in the width-to-angle step halves the report. And reading FOV out of the SDK without checking its source: where the calibration FOV call accepts a useSpecTranslation flag, set it false so the figure comes from your device's calibration rather than the nominal design. If the calibration itself is stale, re-run stereo target capture and stream setup first.
What to buy when the depth FOV still misses the scene
Standoff is the cheapest fix and the most expensive at the far edge: depth noise grows with the square of range for a fixed lens and baseline, so doubling height quadruples far-edge error. If the job is truly long-range, match the baseline to the range or select a wide-baseline camera for depth mapping from the start.
Otherwise change the geometry of the measurement. The 3D sensors range splits into laser profilers, structured-light heads and ToF families, and each answers the FOV problem differently. A laser profiler projects a calibrated light line and rebuilds one calibrated profile after another as parts pass, so there is no stereo overlap to lose and no crop at the edges.
- Type
- Laser profiler
- Type
- Laser profiler
For dense close-range surfaces, a projected structured-light camera such as the MV-DPS200P-02 trades working distance for point density. For zone and volume work, where degree-level edges matter less than coverage, a ToF sensing camera such as the MV-DT01WMU fills the role; the 2D-versus-3D choice for conveyor pick and place walks through matching the sensor family to the job in detail.
- Type
- Stereo / structured light
- Type
- Depth / ToF camera
And if the task collapses to 2D — presence, label, gauge reading — skip the depth stage altogether: a mono area camera with a wide lens delivers its full sensor FOV with no rectification loss. Browse area scan cameras or the camera finder, and size the lens with the focal-length calculator below; the 3D-versus-2D decision guide covers where that line sits.
Checks before you order
Run this list before the bracket is machined:
- Ask for the rectified depth FOV in diagonal/horizontal/vertical; the depth image obeys that line, not the sensor line.
- Compute your required angle as FOV = 2·atan(W/(2Z)) from scene width and standoff, then demand a 10% margin.
- Check the vertical axis too: target height and lane length live there.
- Record the alpha at which the FOV is quoted, and run production at that alpha.
- For work under about 0.5 m, ask for the minimum range as well; near-field parallax narrows the usable view beyond the spec.
- Read FOV from calibration (
useSpecTranslationfalse where the flag exists) and confirm once with a 1.00 m tag measurement.
More depth-image coverage, calibration and alignment fixes sit on the 3D vision topic page. If the layout is unusual, send lane width and standoff through the contact page or start at the camera step of the configurator, and we will check the mount height against the depth FOV with you.
Questions
Does the IR projector limit the depth FOV?
No. The projector only paints texture that helps the two cameras correlate; matching happens wherever both cameras see features. A pattern cone quoted at 78°/61° does not cap depth at those angles — textured scenes give valid depth outside the pattern, and a blank wall inside it can return nothing. The ±7% tolerance on the cone is an optical figure and changes none of this.
How many degrees does rectification typically cost?
It scales with lens distortion, so wide pairs lose the most. The worked example's head gives up 30° horizontally (127° to 97°, about 24%) and 9.5° vertically (79.5° to 70°, about 12%), with the diagonal landing between. A mild 60° stereo lens loses only a few degrees. Buy on the rectified-depth line, never the sensor line.
Why does a depth FOV measurement on a plain wall read far too small?
Two usual causes. A uniform wall gives the correlator nothing to lock onto, so valid depth shrinks to wherever stray texture or the projector pattern lands; re-measure with AprilTag or QR targets at a taped distance. Then check the conversion: FOV is the full angle, 2·atan(W/(2Z)), and a factor-of-two slip there can halve the result.
Why do the borders of my rectified depth image show holes or black pixels?
Those pixels fall outside the overlap both cameras share after undistortion, so no disparity exists and the SDK leaves them invalid. Raising the alpha setting stretches the image over more of the sensor and pushes the invalid band outward, but recovered border pixels were resampled hardest and carry the largest depth error. Keep targets of interest inside the valid rectangle.
Does the depth FOV shrink when the camera works close to the target?
Yes. Both cameras must see each point from their own position, and the baseline hides the extreme edges from one of the pair at short range, so usable depth narrows further as distance falls. Ask the vendor for the minimum range as well as the FOV, add standoff over the arithmetic minimum, and expect the effect to settle past roughly 0.5 m.
How do I read the FOV my calibrated camera actually has?
From the calibration file, not the marketing table. Most stereo SDKs expose a FOV call on the calibration data; where that call accepts a `useSpecTranslation` flag, set it false so the figure is computed from your device's calibrated intrinsics and translation rather than the nominal design. Cross-check once with a tag measurement at a known distance and treat disagreements of more than a few degrees as a calibration problem.