Stereoscopic Profilometry Techniques

• Stereoscopic profilometry is a set of quantitative 3D surface measurement methods that recover depth from multiple views using triangulation, phase correspondence, or normal-field integration.
• It employs fringe projection, phase shifting, and stereo correspondence with precise calibration and rectification to achieve sub-micron depth accuracy in various applications.
• The technique also acts as a correction layer in multimodal imaging, enhancing optical property estimates in systems like laparoscopic and helium microscopy setups.

Stereoscopic profilometry denotes a family of quantitative 3D surface-measurement methods in which depth is recovered from multiple geometrically distinct observations and then mapped into metric shape through triangulation, phase correspondence, or normal-field integration. In fringe-projection implementations, a digital projector and two cameras form a stereo system whose correspondence can be established directly in phase space; in laparoscopic active stereo, binocular depth is used to correct profile-dependent optical-property estimates; in heliometric stereo, multiple detector directions in a scanning helium microscope are inverted to recover local surface normals and then a height field $z=f(x,y)$ (Hu et al., 2018, Song et al., 2024, Vargas et al., 2020, Radic et al., 23 Jan 2025).

1. Geometric foundations

In binocular fringe-projection profilometry, the basic hardware consists of one digital fringe-pattern projector and two cameras, often telecentric when high lateral resolution is required. The projector casts phase-coded sinusoidal fringes onto the object, and each camera views the object from a slightly different angle, forming a binocular stereo pair. Typical geometric parameters include the baseline $B$ between the two camera centers, focal length $f$, pixel pitch $p$, and the principal point $(c_x,c_y)$. The projector can be modeled as an “inverse camera” with its own intrinsic and extrinsic parameters, although explicit projector calibration is not always necessary when phase-based stereo matching is used (Hu et al., 2018).

After stereo rectification, corresponding points lie on the same image row $v$, so a pixel $(u_L,v)$ in the left image matches a point $(u_R,v)$ in the right image, with disparity

$d=u_L-u_R.$

Under a rectified pinhole model, depth is then

$Z=\frac{Bf}{dp}.$

This rectified-row constraint is one of the main simplifications that makes stereoscopic profilometry computationally tractable in dense settings (Hu et al., 2018).

A laparoscopic realization uses the same geometric logic in a more constrained form factor. Song et al. describe a rigid dual-camera laparoscope with a stereo baseline $B$0, angular field-of-view $B$1, a field at $B$2 of $B$3, and an object-pixel size of $B$4. In that system, $B$5 in practice, and low-coherence RGB illumination is used for active stereo depth estimation (Song et al., 2024).

Heliometric stereo generalizes the stereoscopic idea beyond conventional cameras. In a scanning helium microscope, a low-energy neutral helium beam is raster-scanned over the sample, while one or more mass-spectrometer detectors count atoms scattered into known directions $B$6. If scattering is approximately diffuse, the measured signal satisfies Knudsen’s cosine law,

$B$7

where $B$8 is a local albedo factor and $B$9 is the unit surface normal. Here, “stereo” is realized by multiple detector directions rather than multiple image planes (Radic et al., 23 Jan 2025).

2. Calibration, rectification, and distortion modeling

The standard calibration workflow in stereoscopic profilometry begins with intrinsic calibration of the cameras using a chessboard or dot-grid target, followed by estimation of the relative rotation $f$0 and translation $f$1 between the two views. From these, one computes the fundamental matrix $f$2 or the essential matrix

$f$3

Rectification then warps the two images so that epipolar lines become horizontal, reducing dense matching to a 1D search along rows (Hu et al., 2018).

In the stereo laparoscope of Song et al., checkerboard calibration used 69 views of a $f$4 square target with $f$5 squares in MATLAB. The intrinsic matrices are

$f$6

with projection matrices

$f$7

Rectification homographies $f$8 are then computed so that

$f$9

and the transformed fundamental matrix $p$0 enforces horizontal epipolar geometry (Song et al., 2024).

Calibration is a central methodological fault line in fringe-projection profilometry. Vargas et al. distinguish between phase-coordinate mapping (PCM) methods and back-projection stereo-vision (SV) methods: PCM methods are cumbersome to implement because they require precise positioning of the calibration target relative to the FPP system, but they produce highly accurate measurements within the calibration volume; SV methods allow arbitrary positioning of the calibration target, but generally do not achieve the same accuracy level. Their hybrid method leverages the SV calibration approach using a PCM method to achieve higher accuracy, while remaining robust to lens distortions and preserving a simple relation between recovered phase and metric coordinates (Vargas et al., 2020).

The hybrid workflow begins with standard stereo calibration using a B/W checkerboard in arbitrary poses, together with projected fringe patterns to recover absolute phases $p$1 and infer projector-corner coordinates. Bouguet’s camera-calibration toolbox then estimates $p$2, $p$3, $p$4, and distortion coefficients. Lens distortion is modeled on normalized coordinates by radial coefficients $p$5 and tangential coefficients $p$6:

$p$7

After coarse stereo triangulation of planar board positions, each pixel is assigned a cubic phase-to-coordinate mapping,

$p$8

which is stored in a look-up table for runtime reconstruction (Vargas et al., 2020).

3. Fringe projection, phase recovery, and stereo correspondence

In phase-shifting fringe projection, for each fringe set of spatial frequency $p$9 and $(c_x,c_y)$0 phase steps, the camera intensity at pixel $(c_x,c_y)$1 is modeled as

$(c_x,c_y)$2

where $(c_x,c_y)$3 is the background, $(c_x,c_y)$4 is the fringe modulation, and $(c_x,c_y)$5 is the wrapped phase in $(c_x,c_y)$6 (Hu et al., 2018).

If none of the $(c_x,c_y)$7 intensities is saturated, the wrapped phase can be recovered by the standard $(c_x,c_y)$8-step formula

$(c_x,c_y)$9

A special four-step variant often quoted is

$v$0

For shiny surfaces, however, some $v$1 may clip at the camera maximum. The multi-frequency phase-shifting scheme of Zhang et al. addresses this by discarding saturated samples and solving

$v$2

by generalized least squares, provided that at least three intensities remain unsaturated at the pixel. If fewer than three unsaturated samples remain, the pixel is marked unrecoverable at that frequency (Hu et al., 2018).

The same work uses multiple spatial frequencies $v$3, or equivalently periods $v$4. Denser fringes yield higher phase sensitivity but are more vulnerable to defocus and highlight saturation; low-frequency fringes provide a coarse but robust phase scaffold. Sequential temporal unwrapping is performed as

$v$5

with $v$6. To preserve completeness, invalid pixels in the densest fringe map are filled by equivalent phases from coarser frequencies,

$v$7

proceeding from $v$8 downward until holes are filled (Hu et al., 2018).

Once unwrapped phase maps are available in both cameras, correspondence can be established without explicit projector calibration. After epipolar rectification, for each left-image pixel $v$9 with unwrapped phase $(u_L,v)$0, the matching right-image coordinate $(u_L,v)$1 is found on the same row by phase consistency with $(u_L,v)$2. Because phase varies monotonically left-to-right with the projected fringes, the search is 1D. Sub-pixel refinement is then obtained by interpolation in phase:

$(u_L,v)$3

The disparity $(u_L,v)$4 is converted into depth through triangulation; in a projector–camera interpretation, one may also express depth using a phase difference term $(u_L,v)$5 and fringe pitch $(u_L,v)$6 (Hu et al., 2018).

4. Stereo profilometry as a correction layer in optical imaging

Stereoscopic profilometry is not limited to direct shape metrology; it can also serve as a geometric correction stage in multimodal imaging. Song et al. combine active stereo depth estimation with speckle-illumination spatial frequency domain imaging (si-SFDI) in a compact two-camera laparoscope to recover profile-corrected, pixel-level absorption and reduced scattering maps in tissues with complex geometries (Song et al., 2024).

The stereo pipeline uses low-coherence RGB laser flood illumination and fast coarse-to-fine block matching on rectified speckle images to obtain disparity $(u_L,v)$7 and depth

$(u_L,v)$8

Each depth map is back-projected into the camera-1 frame through

$(u_L,v)$9

yielding a dense point cloud. Surface normals are then estimated by local plane fitting, using either PCA or least squares on $(u_R,v)$0 nearest neighbors (Song et al., 2024).

Those normals are used for geometric correction of the measured SFDI intensities. Assuming Lambertian reflectance and a source vector parallel to the camera axis, the local irradiance scales as $(u_R,v)$1, where $(u_R,v)$2. To recover a virtual flat reflectance at the reference distance $(u_R,v)$3 and normal incidence, Song et al. apply

$(u_R,v)$4

and divide measured $(u_R,v)$5 and $(u_R,v)$6 by $(u_R,v)$7 before optical-property lookup (Song et al., 2024).

The si-SFDI acquisition itself uses only two red-channel images under high-coherence speckle illumination: an AC image with the laser-speckle reducer off, and a DC image with the reducer on. A sliding-window Wiener–Khinchin analysis estimates the local power spectral density $(u_R,v)$8, from which the modulation metrics

$(u_R,v)$9

are formed. After correction, the diffuse reflectances are mapped through a precomputed LUT to obtain $d=u_L-u_R.$0. This suggests a broader role for stereoscopic profilometry as an enabling geometric prior for quantitative imaging, rather than solely a source of height maps (Song et al., 2024).

5. Quantitative performance and uncertainty

The uncertainty structure of stereoscopic profilometry depends on the observation model. In multi-frequency phase-shifting profilometry, the variance of wrapped phase noise is reported as

$d=u_L-u_R.$1

where $d=u_L-u_R.$2 is the camera’s additive-noise variance, $d=u_L-u_R.$3 is the step number, $d=u_L-u_R.$4 is fringe frequency in fringes per field of view, and $d=u_L-u_R.$5 is fringe modulation. Propagation through disparity and triangulation yields depth noise $d=u_L-u_R.$6, typically on the order of a micron for microscopic telecentric systems (Hu et al., 2018).

The reported performance figures across representative stereoscopic profilometry systems are summarized below.

Configuration	Reported quantitative result	arXiv id
Microscopic multi-frequency phase stereo	repeatability better than $d=u_L-u_R.$7; angular errors $d=u_L-u_R.$8; plane-fit residuals below $d=u_L-u_R.$9	(Hu et al., 2018)
Hybrid-calibrated FPP	inclined plane RMS $Z=\frac{Bf}{dp}.$0 vs $Z=\frac{Bf}{dp}.$1 for SV; cylinder RMS $Z=\frac{Bf}{dp}.$2 vs $Z=\frac{Bf}{dp}.$3; $Z=\frac{Bf}{dp}.$4 k points in $Z=\frac{Bf}{dp}.$5 vs $Z=\frac{Bf}{dp}.$6	(Vargas et al., 2020)
Stereo laparoscopic si-SFDI	$Z=\frac{Bf}{dp}.$7 height-error per cm reduced from $Z=\frac{Bf}{dp}.$8 to $Z=\frac{Bf}{dp}.$9; $B$00 from $B$01 to $B$02; in vivo finger errors $B$03 for $B$04 and $B$05 for $B$06	(Song et al., 2024)
Heliometric stereo	facet-angle measurements within $B$07 of ideal $B$08; area error $B$09; sphere RMS errors $B$10 and shape errors $B$11; up to $B$12 on a larger-aspect-ratio sphere	(Radic et al., 23 Jan 2025)

Vargas et al. also report stereo-vision reprojection errors of $B$13 for the camera and $B$14 for the projector, together with RMS fitting errors over a calibration volume of approximately $B$15 satisfying

$B$16

These results indicate that calibration strategy can dominate runtime and reconstruction error even when the underlying stereo geometry is unchanged (Vargas et al., 2020).

6. Failure modes, misconceptions, and methodological boundaries

A persistent misconception is that shiny or highly reflective surfaces are intrinsically unrecoverable by fringe-projection profilometry. The microscopic study of shiny metal surfaces shows a more conditional situation: if at least one fringe set provides at least three unsaturated samples per pixel, the saturated measurements can be discarded and the phase can still be recovered; fully saturated blind spots at every frequency cannot be recovered (Hu et al., 2018). A plausible implication is that robustness depends less on reflectivity per se than on the interaction among exposure, defocus, and multi-frequency redundancy.

Another misconception is that explicit projector calibration is always mandatory. In phase-based stereo matching, the unwrapped phase maps from the two cameras already encode the projector geometry implicitly, so the projector need not be explicitly calibrated for correspondence establishment (Hu et al., 2018). Conversely, the calibration literature shows that pure stereo and phase-coordinate mapping are not mutually exclusive frameworks: the hybrid procedure of Vargas et al. combines the flexible acquisition of stereo-vision calibration with a pixel-wise phase-to-XYZ mapping to obtain higher accuracy and lower reconstruction time than a pure stereo-vision model (Vargas et al., 2020).

Several boundary conditions recur across modalities. For the multi-frequency HDR fringe method, success depends on a moderate depth of field; in systems with very large depth of field, high-frequency fringes will not defocus sufficiently to create the unsaturated black bands needed in highlights. Telecentric cameras with Scheimpflug tilts or non-telecentric optics add calibration complexity because tilt must be modeled in the intrinsic matrix or handled through affine rectification (Hu et al., 2018).

In the laparoscopic si-SFDI system, correction quality degrades at steeper angles because normal estimation noise propagates into the Lambertian $B$17 compensation; for a hemisphere, the reported improvement for polar angles below $B$18 deteriorates at steeper angles for this reason (Song et al., 2024). In heliometric stereo, the dominant failure modes are masking, shadowing, and multiple scattering. Masking drives detector counts toward zero when the detector line of sight is occluded; shadowing reduces incident flux when the beam is blocked; multiple scattering violates the simple cosine law. Normal estimation is also sensitive to the conditioning of the detector-direction matrix $B$19, since

$B$20

Accordingly, the literature recommends at least four detectors or sample-rotation angles, well-separated detector geometries, and exclusion of shadowed or masked pixels before solving for normals (Radic et al., 23 Jan 2025).

Potential extensions identified in the literature remain modality-specific. For shiny-surface fringe profilometry, these include adaptive fringe contrast, dynamic-range coded patterns such as inverted fringes and coded HDR exposures, polarization control, and multi-sensor fusion with stereo photometric shading (Hu et al., 2018). For helium-based implementations, annular or ring detectors coaxial with the beam could eliminate masking for one view, provided they are paired with off-axis detectors to recover all normal components, and emerging SHeM designs promise sub-$B$21 lateral resolution (Radic et al., 23 Jan 2025). Together, these developments place stereoscopic profilometry at the intersection of geometric reconstruction, radiometric modeling, and instrument-specific forward physics.