Projection Topographic Imaging

Updated 5 July 2026

Projection Topographic Imaging is a family of techniques that employ projection operators—optical, geometric, interferometric, or algebraic—to recover and manipulate surface topography and scene structure.
Methods range from neural projection mapping and orthographic depth extraction to structured-light scanning and interferometric stereovision, each offering tailored solutions for calibrations and measurement challenges.
Recent advances enable joint optimization of geometry, reflectance, and calibration while addressing global illumination effects, projector misalignment, and the trade-offs between topographic and full-range distance measurements.

Projection Topographic Imaging designates a family of imaging, inverse-rendering, and inverse-problem formulations in which a projection operator is the central mechanism for recovering, isolating, or manipulating surface topography and related scene structure. In recent arXiv literature, the term spans differentiable projector-integrated neural reflectance fields for joint topography recovery and projection mapping, orthographic image-space extraction of micro-topography from dense 3D reconstructions, structured-light scanning under global illumination via projection functions, topography-first interferometric stereovision, subspace projection for local electrical impedance tomography, and transformation-driven generation of comparable anatomical projections (Erel et al., 2023, Zeppelzauer et al., 2015, Li et al., 2023, bao et al., 1 Jun 2026, Jääskeläinen et al., 20 Jan 2026, Pojda et al., 15 Jun 2026). The common feature is not a single hardware arrangement, but an explicitly modeled projection—optical, geometric, interferometric, or algebraic—that maps scene structure into a representation in which local height, shape variation, or region-specific change becomes estimable.

1. Scope and formal structure

Across the cited works, Projection Topographic Imaging is best understood as a class of methods rather than a single standardized pipeline. In one line of work, the projector itself is embedded in a neural reflectance field and optimized jointly with geometry, materials, and appearance control. In another, dense 3D measurements are orthographically projected to a 2D depth map and converted into Enhanced Topography Maps for classification. A third line uses oblique sinusoidal illumination and discrete Radon projections of projector–camera light transport. A fourth replaces range triangulation with an interferometric measurement optimized for topography. A fifth treats projection as an orthogonal operation in measurement space that suppresses nuisance subspaces in EIT. A sixth propagates independently transformable anatomical objects directly into projection space under explicit imaging assumptions (Erel et al., 2023, Zeppelzauer et al., 2015, Li et al., 2023, bao et al., 1 Jun 2026, Jääskeläinen et al., 20 Jan 2026, Pojda et al., 15 Jun 2026).

Formulation	Projection mechanism	Representative paper
Neural projection mapping	Differentiable projector inside a neural reflectance field	(Erel et al., 2023)
Image-space topography extraction	Orthographic projection of 3D points to a depth map	(Zeppelzauer et al., 2015)
pPSI structured light	Discrete Radon projection functions of light transport	(Li et al., 2023)
Topographic stereovision	Interferometric projection of parallax-induced phase	(bao et al., 1 Jun 2026)
Local EIT reconstruction	Orthogonal projection away from nuisance Jacobian subspaces	(Jääskeläinen et al., 20 Jan 2026)
Anatomical projection imaging	Explicit transform propagation from scene to detector space	(Pojda et al., 15 Jun 2026)

Mathematically, these formulations differ in what is being projected. In image-space extraction, a point cloud is mapped to $(u,v,h)$ coordinates relative to a support plane. In projector-integrated neural fields, a world point is mapped into projector image coordinates by $P_p = K [R \mid t] P_w$ . In pPSI, a 2D pixel transport image is compressed into a 1D projection function by a discrete Radon transform. In EIT, the projector is $P = I - U_k U_k^\top$ , acting on measurements and Jacobians rather than on rays. In transformation-driven anatomical imaging, the scene is observed through a factorized pipeline $I_{\text{display}} = P(A(S,G,\Theta_A), \Theta_P)$ (Zeppelzauer et al., 2015, Erel et al., 2023, Li et al., 2023, Jääskeläinen et al., 20 Jan 2026, Pojda et al., 15 Jun 2026).

2. Differentiable projector-integrated neural reflectance fields

In "Neural Projection Mapping Using Reflectance Fields" (Erel et al., 2023), Projection Topographic Imaging aims to recover accurate surface topography and reflectance while simultaneously enabling photorealistic projection mapping. The method introduces a projector as a high-resolution, spatially adaptive direct light source embedded in a neural reflectance field. The field consists of three neural networks: a geometry network $N_{\text{geo}}$ predicting local density $\sigma(x)$ and normals $n(x)$ , a material network $N_{\text{mat}}$ predicting dielectric microfacet BRDF parameters, and a transmittance network $N_{\text{tr}}$ predicting line-of-sight transmittance $\tau(x \rightarrow y)$ . Image formation combines projector illumination and a camera co-located light through volumetric rendering,

$P_p = K [R \mid t] P_w$ 0

with

$P_p = K [R \mid t] P_w$ 1

The projector is modeled as an inverse pinhole camera:

$P_p = K [R \mid t] P_w$ 2

The acquisition is intentionally minimal: a fixed, uncalibrated RGB projector (EPSON EH-TW5350, $P_p = K [R \mid t] P_w$ 3), a handheld RGB camera (Point Grey FL3-U3-13S2C-CS, $P_p = K [R \mid t] P_w$ 4), and a co-located white LED in a dark room. Real scenes are captured from 102 viewpoints, each under black flood-fill, white flood-fill, and a random lollipop pattern, for 306 images total. Camera intrinsics and extrinsics are recovered using COLMAP, and foreground masks are generated using U2-Net. The optimization schedule is staged: scene training with black frames using $P_p = K [R \mid t] P_w$ 5, then projector calibration with frozen networks, then joint fine-tuning. This schedule is explicitly described as critical for convergence because it prevents geometry from chasing a misposed projector.

The lollipop patterns are central to calibration. They are described as high-frequency, dense, and non-centrally symmetric, and in ablation they consistently yield robust projector calibration with the lowest $P_p = K [R \mid t] P_w$ 6 and rotation errors. Their strong center pulls the solution toward a single global minimum and avoids symmetric local minima. The practical implication is that topography recovery, projector calibration, and light editing are solved in one inverse-rendering problem rather than in separate per-view calibration stages.

The reported applications include projector compensation at novel viewpoints, multiview optimization of a single projector texture, CDC-based text-to-projection editing, and a see-through "XRAY" mode implemented with a two-pass render that swaps projector and camera in rendering. On a real bunny scene with 14 target images, the reported PSNR values are 17.8 for classical structured-light calibration plus gamma correction plus color compensation, 13.2 for no geometric calibration, 15.8 for geometric-only, 16.7 for CompenNeSt++, and 17.5 for the neural method, with evaluation from a viewpoint that was novel for the method. On synthetic relighting under novel projections and views, average PSNR over 36 novel views is 17.0 for NeRF, 23.5 for NeRF-AA, and 27.8 for the proposed method; training is approximately 3 hours per scene on an RTX A6000 and inference is 2–10 seconds for $P_p = K [R \mid t] P_w$ 7 images. The limitations are equally explicit: the method relies on direct projector illumination dominating the light field, neglects global illumination, assumes a dielectric microfacet BRDF, can underestimate some shadow extents, and remains computationally demanding.

3. Orthographic image-space extraction of measured surface topography

In "Efficient Image-Space Extraction and Representation of 3D Surface Topography" (Zeppelzauer et al., 2015), Projection Topographic Imaging refers to projecting dense 3D surface measurements into a 2D image-space representation that isolates and enhances the surface’s micro-structure for efficient analysis and classification. Here projection is neither structured light nor inverse rendering; it is orthographic flattening of a high-resolution point cloud onto a support plane. Given a support plane defined by a point $P_p = K [R \mid t] P_w$ 8, unit normal $P_p = K [R \mid t] P_w$ 9, and in-plane basis $P = I - U_k U_k^\top$ 0, each 3D point $P = I - U_k U_k^\top$ 1 is mapped by

$P = I - U_k U_k^\top$ 2

and the depth map is $P = I - U_k U_k^\top$ 3.

The central operation is macro-surface removal by Gaussian filtering:

$P = I - U_k U_k^\top$ 4

The residual $P = I - U_k U_k^\top$ 5 is split into valleys and peaks,

$P = I - U_k U_k^\top$ 6

then smoothed and log-compressed to produce Enhanced Topography Maps,

$P = I - U_k U_k^\top$ 7

This pipeline suppresses global curvature, isolates local deviations such as crannies and engravings, and produces 2D maps whose absolute values are intended to carry discriminative signal more effectively than raw gradients.

The reported experiments use four high-resolution 3D rock surface reconstructions totaling $P = I - U_k U_k^\top$ 8 points at less than $P = I - U_k U_k^\top$ 9 mm resolution in $I_{\text{display}} = P(A(S,G,\Theta_A), \Theta_P)$ 0, $I_{\text{display}} = P(A(S,G,\Theta_A), \Theta_P)$ 1, and $I_{\text{display}} = P(A(S,G,\Theta_A), \Theta_P)$ 2. Engravings constitute 16.6% of the data, and classification uses RUSBoost to handle imbalance. The proposed ETM representation is paired with block-wise descriptors including Global Histogram Shape, defined as the first 30 low-frequency DCT coefficients of the block histogram, and Spatial Frequencies, defined as the first $I_{\text{display}} = P(A(S,G,\Theta_A), \Theta_P)$ 3 low-frequency 2D DCT coefficients. Quantitatively, dense 3D descriptors perform weakly, with PFH peaking at $I_{\text{display}} = P(A(S,G,\Theta_A), \Theta_P)$ 4; color reaches a maximum $I_{\text{display}} = P(A(S,G,\Theta_A), \Theta_P)$ 5; the depth gradient map reaches $I_{\text{display}} = P(A(S,G,\Theta_A), \Theta_P)$ 6; the raw depth map reaches $I_{\text{display}} = P(A(S,G,\Theta_A), \Theta_P)$ 7 with HOG; ETM reaches $I_{\text{display}} = P(A(S,G,\Theta_A), \Theta_P)$ 8 with GLCM, $I_{\text{display}} = P(A(S,G,\Theta_A), \Theta_P)$ 9 with GHS, $N_{\text{geo}}$ 0 with SF, and $N_{\text{geo}}$ 1 with combined GHS+SF, with $N_{\text{geo}}$ 2 versus all other features.

The computational profile is also part of the method’s identity. Orthographic projection is $N_{\text{geo}}$ 3 in the number of points, convolution is $N_{\text{geo}}$ 4, and subtraction, cropping, and log compression are $N_{\text{geo}}$ 5. The method is therefore presented as linear in the number of points and pixels, suitable for multi-million-point clouds. Its limitations are geometric rather than statistical: flattening non-developable surfaces to a plane introduces distortion when Gaussian curvature is high, the filtering scale $N_{\text{geo}}$ 6 determines which topographic granularity is emphasized, and reconstruction artifacts near depth discontinuities can survive smoothing.

4. Structured-light transport analysis under global illumination

"Projective Parallel Single-Pixel Imaging: 3D Structured Light Scanning Under Global Illumination" (Li et al., 2023) situates Projection Topographic Imaging within structured-light scanning, but replaces standard direct-path assumptions with an explicit light-transport model. For a projector pattern $N_{\text{geo}}$ 7 and camera pixel $N_{\text{geo}}$ 8, the image formation model is

$N_{\text{geo}}$ 9

where $\sigma(x)$ 0 is the light transport coefficient from projector pixel $\sigma(x)$ 1 to camera pixel $\sigma(x)$ 2. The method reduces the 4D LTC to 1D projection functions by a discrete Radon transform:

$\sigma(x)$ 3

These functions are captured by oblique $\sigma(x)$ 4-step sinusoidal patterns, Fourier-demodulated into $\sigma(x)$ 5, and inverted by a 1D IDFT.

Two algorithmic ideas govern the method. The Local Maximum Constraint states that if the projection line orthogonal to direction $\sigma(x)$ 6 through the direct speckle does not pass through any global-illumination speckles, then the direct speckle’s projected position is a local maximum of $\sigma(x)$ 7. The Local Slice Extension accelerates capture by first estimating a coarse support with low frequencies, then reconstructing only a limited local slice patch, periodically extending it, and masking it by the coarse support. The paper states a perfect reconstruction property for LSE when the maximal projected reception field fully covers the nonzero support of the true projection function.

The practical outcome is a structured-light method that remains effective under inter-reflections and subsurface scattering. For general global illumination, the recommended configuration is four directions $\sigma(x)$ 8 with $\sigma(x)$ 9, $n(x)$ 0, $n(x)$ 1 pixels, and capture ratio $n(x)$ 2, requiring approximately 336 patterns and about 2 seconds of acquisition. Full LTC-based PSI requires approximately 51,000 patterns and about 5 minutes, while quick pPSI uses one direction and approximately 84 patterns in about 0.5 seconds. The paper reports that under inter-reflections, $n(x)$ 3 yields SME $n(x)$ 4 px versus $n(x)$ 5 px at $n(x)$ 6, with the first 40% frequencies carrying 90.36% energy; under subsurface scattering, $n(x)$ 7 yields SME $n(x)$ 8 px versus $n(x)$ 9 px at $N_{\text{mat}}$ 0, with the first 16% frequencies carrying 95.06% energy.

The validation includes several geometry benchmarks. On a V-groove under inter-reflections, plane-fitting RMS is $N_{\text{mat}}$ 1– $N_{\text{mat}}$ 2 mm for the upper plane and $N_{\text{mat}}$ 3– $N_{\text{mat}}$ 4 mm for the lower plane depending on capture ratio. On a translucent polyamide sphere, the fitted diameter is 25.432 mm versus ground truth 25.449 mm, with absolute error 0.017 mm and RMS 0.031 mm. The method is presented as outperforming micro-phase shifting and epipolar-imaging baselines in cases where projected patterns are overlapped, low-modulation, or dominated by global illumination. Its main failure mode is explicit: if severe global illumination overlap occurs along all tested directions, the Local Maximum Constraint can cease to isolate the direct correspondence.

5. Quantum-inspired topographic stereovision

"Quantum-inspired Topographic Stereovision" (bao et al., 1 Jun 2026) reformulates distant stereovision by asserting that shape rather than absolute distance is the relevant observable. The scene is parameterized by a mean range $N_{\text{mat}}$ 5 and topography $N_{\text{mat}}$ 6 such that $N_{\text{mat}}$ 7. Three apertures at $N_{\text{mat}}$ 8, $N_{\text{mat}}$ 9, and $N_{\text{tr}}$ 0 define a stereovision apparatus in which the central path acts as a local oscillator and the side paths are Mach–Zehnder arms. In the long-range regime, the parallax-induced phases are

$N_{\text{tr}}$ 1

Residual disparity is linked to topography by

$N_{\text{tr}}$ 2

The central conceptual claim is an observable–measurement mismatch: the measurement that saturates the quantum Fisher information for absolute distance is not optimal for the distance gradient or topography. This is quantified through information regret,

$N_{\text{tr}}$ 3

with the long-range trade-off

$N_{\text{tr}}$ 4

and $N_{\text{tr}}$ 5 in the long-range limit, implying $N_{\text{tr}}$ 6. Triangulation that optimizes $N_{\text{tr}}$ 7 therefore necessarily regrets topographic information, whereas the proposed topographic interferometer sets $N_{\text{tr}}$ 8 at the expense of distance precision.

The interferometer uses balanced homodyne detection with the central arm as the local oscillator. After stereo regularization of anisotropic efficiencies and 50/50 beam-splitter transformations, the balanced signal at the right detector depends on the disparity and topographic anisotropy terms through

$N_{\text{tr}}$ 9

For Gaussian PSFs, Hermite–Gaussian projections are described as optimal POVMs for geometric moments. In the single-segment setting with $\tau(x \rightarrow y)$ 0 and $\tau(x \rightarrow y)$ 1, the QFI for disparity is

$\tau(x \rightarrow y)$ 2

which yields

$\tau(x \rightarrow y)$ 3

The paper further states that with strong local oscillator and small aperture ratio $\tau(x \rightarrow y)$ 4, the classical FI of the topographic interferometer saturates the QFI, $\tau(x \rightarrow y)$ 5, whereas triangulation performs poorly in the sub-Rayleigh regime.

This formulation is notable because it eliminates stereo matching and does not measure an absolute distance profile. Instead, it uses cross-detector correlations and quadrature measurements to estimate local height and slope, then reconstructs a 2D topographic map by integrating measured gradients with regularization. A plausible implication is that, in this literature, "projection" can refer not only to illumination patterns or coordinate flattening, but also to modal projection of parallax-induced phase gradients onto measurement channels optimized for topographic observables.

6. Projection in inverse problems and transformation-driven anatomical scenes

"Local electrical impedance tomography via projections" (Jääskeläinen et al., 20 Jan 2026) extends the term into inverse problems where the relevant projection is algebraic. The conductivity Jacobian is partitioned into ROI and nuisance blocks, $\tau(x \rightarrow y)$ 6, and dominant nuisance directions are identified from the eigendecomposition of $\tau(x \rightarrow y)$ 7 or, alternatively, a mass-weighted construction. Selecting the first $\tau(x \rightarrow y)$ 8 nuisance eigenvectors gives $\tau(x \rightarrow y)$ 9, and the orthogonal projector

$P_p = K [R \mid t] P_w$ 00

is then applied to both residuals and sensitivities:

$P_p = K [R \mid t] P_w$ 01

The inverse problem is solved with TV regularization and lagged diffusivity,

$P_p = K [R \mid t] P_w$ 02

In water-tank experiments mimicking hemorrhagic stroke with physiological scalp variations, nuisance projection with $P_p = K [R \mid t] P_w$ 03, $P_p = K [R \mid t] P_w$ 04, and $P_p = K [R \mid t] P_w$ 05 removes or strongly suppresses artifacts from conductivity changes outside the ROI. The trade-off is equally explicit: as $P_p = K [R \mid t] P_w$ 06 grows, nuisance suppression improves but useful ROI information is also removed, reducing contrast and effective degrees of freedom.

"Transformation-driven generation of comparable projection images from multimodal anatomical scenes" (Pojda et al., 15 Jun 2026) uses projection in a different sense again: a multimodal scene composed of volumes, segmented structures, surface meshes, dental scans, therapeutic objects, and annotations is propagated directly into projection space through explicit transformations. Each object carries its own rigid transform $P_p = K [R \mid t] P_w$ 07, with evaluation performed on demand by

$P_p = K [R \mid t] P_w$ 08

The forward model is modularized into scene $P_p = K [R \mid t] P_w$ 09, geometry $P_p = K [R \mid t] P_w$ 10, acquisition $P_p = K [R \mid t] P_w$ 11, material interpretation $P_p = K [R \mid t] P_w$ 12, and presentation $P_p = K [R \mid t] P_w$ 13:

$P_p = K [R \mid t] P_w$ 14

For cone-beam geometry, rays are formed from source $P_p = K [R \mid t] P_w$ 15 toward detector pixels, and attenuation is modeled with Beer–Lambert,

$P_p = K [R \mid t] P_w$ 16

or by discrete ray marching and Siddon traversal. The framework supports slab-limited projection by restricting integration either to a ray interval or to a planar slab, which is used in TMJ visualizations to compare 35 mm and 5 mm slabs under identical imaging assumptions.

The emphasis here is methodological comparability rather than full radiographic realism. The paper reports that sampling versus Siddon discrepancy is small, with relative MAE below 1.2% on tested phantoms, and shows that mandibular motion and therapeutic repositioning can be visualized as changes in directly comparable VirtualRTG projections. This suggests a broader interpretation of Projection Topographic Imaging: not only reconstructing shape from measurements, but also constructing reproducible projection-space observations of known, independently transformable anatomy so that anatomy–projection relationships can be studied under controlled assumptions.

A useful contrast is provided by "Computational 3D topographic microscopy from terabytes of data per sample" (Zhou et al., 2023). STARCAM is explicitly described as recovering surface topography without any projected patterns. It uses 54 synchronized micro-cameras in a $P_p = K [R \mid t] P_w$ 17 array, 3-axis translation, 65-slice z-stacks, and a 7D data hypervolume of 224,640 images, about 2.1 TB per sample. The reported system achieves a synthetic lateral field of view greater than 110 cm $P_p = K [R \mid t] P_w$ 18, approximately 6 gigapixels in the final all-in-focus composite, micron-scale lateral resolution, and on gauge blocks a mean absolute error of approximately 5.8 $P_p = K [R \mid t] P_w$ 19m, RMSE of approximately 6.8 $P_p = K [R \mid t] P_w$ 20m, and within-block standard deviation of approximately 10 $P_p = K [R \mid t] P_w$ 21m. STARCAM is therefore adjacent to Projection Topographic Imaging rather than an instance of projected-light PTI: it shows that large-scale topographic recovery can also be accomplished by passive multi-view stereo and shape-from-focus when projected illumination is impractical.

A common misconception is that Projection Topographic Imaging necessarily denotes projector–camera triangulation with active patterns. Across the cited literature, the projection operator may act in illumination space, image space, measurement space, modal interferometric space, or transformation space. Another misconception is that all such methods optimize the same observable. The stereovision work explicitly argues that absolute distance and topography are incompatible observables for optimal measurement; the neural projection-mapping work optimizes joint geometry, materials, transmittance, calibration, and projector texture; the EIT work optimizes locality by suppressing nuisance subspaces; the anatomical framework optimizes comparability rather than full physical realism (bao et al., 1 Jun 2026, Erel et al., 2023, Jääskeläinen et al., 20 Jan 2026, Pojda et al., 15 Jun 2026).

The limitations are correspondingly heterogeneous. Neural projection mapping assumes direct projector illumination dominates and uses a dielectric microfacet BRDF. ETM-based image-space extraction can distort highly curved non-developable surfaces. pPSI depends on at least some projection directions for which the direct speckle is not overlapped by global-illumination speckles and is best suited to static or quasi-static objects. QiTS requires tight phase stability, alignment, and turbulence mitigation, while explicitly trading distance precision for topographic precision. Local EIT via projections requires careful selection of $P_p = K [R \mid t] P_w$ 22, because heavy projection reduces usable information. Transformation-driven anatomical projection imaging deliberately omits scatter, beam hardening, polychromatic spectra, detector noise, and focal blur to preserve controllability and reproducibility (Erel et al., 2023, Zeppelzauer et al., 2015, Li et al., 2023, bao et al., 1 Jun 2026, Jääskeläinen et al., 20 Jan 2026, Pojda et al., 15 Jun 2026).

The future directions named in the cited works are similarly diverse but structurally aligned. They include dynamic scenes and real-time optimization for neural projection mapping, adaptive acquisition patterns for projector calibration, multi-projector and multi-camera setups, richer BRDF and transport models, multi-scale ETM fusion, learned features on ETM, automated nuisance-subspace selection in EIT, hybrid Bayesian and projection-based inverse methods, non-rigid scene transforms for anatomical projection imaging, GPU acceleration for transformation-driven rendering, and extensions of topographic interferometry to remote sensing and astronomy (Erel et al., 2023, Zeppelzauer et al., 2015, Jääskeläinen et al., 20 Jan 2026, bao et al., 1 Jun 2026, Pojda et al., 15 Jun 2026). Taken together, these works indicate that Projection Topographic Imaging is best treated as a unifying methodological category centered on explicitly modeled projections that make topography, local structure, or ROI-specific change observable in otherwise ill-posed sensing problems.