Resources
Monocular versus stereo versus RGB-D egocentric data
Choose camera observations and depth deliverables by the model task, calibration evidence and actual output files.
Published 2026-10-27 · 6 min read
| Option | Core deliverable | Evidence to request |
|---|---|---|
| Monocular | One image stream | Placement, settings, visibility and original timeline |
| Stereo | Paired observations | Intrinsics, inter-camera transform and exposure timing |
| RGB-D | Color plus depth | Depth method, units, validity and color alignment |
Monocular means one camera view. Stereo uses two views with a known geometric relationship to support correspondence and triangulation. RGB-D means color plus depth output; it does not identify a single depth-sensing technology. A stereo system can itself produce RGB-D data.
Choose the delivered observations your model needs rather than treating these labels as a quality ladder. A relevant monocular sequence may fit an observation task, while a geometry-dependent task needs more evidence than a second video file.
Monocular: useful visual context, with explicit geometry limits
A single wearable camera can record the objects, hands and sequence of an activity. State camera placement, field of view, native settings, timestamps and whether original frames survive editing. One view can miss contact or action hidden behind an object.
A depth model can infer a depth map from monocular video, but identify that output as model-derived. It is not an additional sensor measurement. Do not describe an estimated hand trajectory or inferred scale as measured ground truth without a suitable reference and validation.
Stereo: a pair plus calibration and timing
Ask for left/right camera IDs, intrinsics, lens distortion, the transform between cameras, units and calibration applicability. OpenCV documents these geometry components. Multiple independently moving cameras are not one fixed calibrated stereo rig simply because both recordings exist.
The views also need an established timing relationship. Inspect moving objects in a representative pilot, frame gaps and timing drift. Geometry alone does not make exposures simultaneous. A factory-calibrated product still needs documentation of the actual settings and exported streams.
RGB-D: inspect how depth was generated
Request the depth method, units, invalid-value convention, confidence or validity data and the color-to-depth relationship. Stereo-derived depth, structured-light depth and time-of-flight depth have different pipelines; evaluate the actual device and operating conditions.
Stereolabs documents a stereo pipeline producing color images, depth and confidence outputs. This is a concrete example of stereo and RGB-D overlapping, not a performance guarantee for arbitrary devices. Preserve the original images and processing configuration when those are part of the agreed deliverable.
Compare failure modes on the target activity
For mono, review visibility and whether the intended task needs unavailable geometry. For stereo, inspect overlapping coverage, calibration applicability and timing. For RGB-D, inspect depth validity around hands, surfaces and working distances relevant to the task.
Set acceptance with the buyer using representative clips. Do not prescribe a universal baseline, resolution, range or depth-error threshold. A capture label cannot replace a pilot that loads in the intended workflow and exposes missing or unreliable signals.
Price the complete scope, not the camera label
Separate unique activity hours from stream-hours and from the time spent generating depth or pose. Ask which files, calibration, logs and labels are included, and whether processing or review is separately charged. No universal stereo or RGB-D price multiplier is established here.
Owners can describe the material they already have. Collection teams should describe capability before buying equipment. Fiund confirms source fit, availability and commercial scope after review; this comparison does not imply that every configuration is currently available.
Frequently asked questions
Are stereo and RGB-D mutually exclusive?
No. A stereo pipeline can output color and depth. Describe both the capture method and delivered representations.
Is estimated monocular depth measured sensor depth?
No. Mark it as a derived estimate and describe the model, scale assumptions and evaluation.
Does more cameras always mean more useful data?
No. Extra views need relevant coverage, documented geometry where required, timing and a usable delivery package.