pith. sign in

arxiv: 2606.17093 · v1 · pith:73FHK5WRnew · submitted 2026-06-13 · 💻 cs.LG · eess.IV

Diagnosing and Repairing Shape-Prior Shortcuts in Long-Range Single-Shot Fringe Projection Profilometry

Adam Haroon , Anush Lakshman , Cody Fleming , Beiwen Li This is my paper

Pith reviewed 2026-06-27 04:51 UTC · model grok-4.3

classification 💻 cs.LG eess.IV
keywords single-shot fringe projection profilometrylong-range FPPshape priorsmechanistic interpretabilityconformal uncertainty quantificationphase to depth calibrationneural network shortcutswrapped phase
0
0 comments X

The pith

UNet baselines for long-range single-shot fringe projection profilometry solve the task using object-boundary shape priors rather than fringe-phase decoding.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that standard UNet models for long-range single-shot FPP at distances beyond 1 m achieve low error by learning object shape priors from boundaries instead of decoding the actual fringe phase information present in the image. Three independent diagnostic probes—linear probing, Grad-CAM visualizations, and out-of-distribution flat-plane tests—all converge on the same physical failure locus. The repair replaces direct depth regression with an architecture that first outputs wrapped phase and then applies a fixed differentiable calibration layer to convert phase to depth, thereby removing the shape-prior solution from the hypothesis space. This change reduces object mean absolute error from 14.54 mm to 4.46 mm on a 15,600-image photorealistic synthetic benchmark. A comparable physics-informed loss penalty applied to a depth-regressing network yields no improvement, isolating the architectural constraint as the operative factor.

Core claim

On the 15,600-image benchmark the baseline UNet reaches 14.54 mm object MAE by relying on object-boundary shape priors. PhiCalNet, which outputs wrapped phase rather than depth and applies a fixed differentiable calibration layer mapping phase to depth, reduces object MAE 3.3 times to 4.46 mm. The residual error is carried by the 0.103 percent of pixels at the +/-pi wrap discontinuity. Pixel-wise conformal uncertainty quantification confirms the diagnosis: rejecting the top 5 percent of object pixels by snapshot disagreement cuts PhiCalNet RMSE by 64 percent (20.6 mm to 7.4 mm) versus only 3.5 percent for the baseline.

What carries the argument

PhiCalNet architecture that outputs wrapped phase rather than depth and applies a fixed differentiable calibration layer mapping phase to depth, removing shape-prior solutions from the hypothesis space.

If this is right

  • The shape-prior shortcut is the dominant failure mode in baseline models rather than an artifact of the loss function.
  • Architectural removal of the shortcut outperforms soft physics-informed loss penalties on depth regression.
  • Mechanistic interpretability and conformal uncertainty quantification converge on the same failure locus at object boundaries.
  • Residual error concentrates at the +/-pi phase wrap discontinuities.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the synthetic benchmark statistics match real hardware, similar shape-prior shortcuts may appear in other single-shot 3D reconstruction tasks that are ill-posed without multi-frame information.
  • Explicit handling of wrap discontinuities could further reduce the remaining 4.46 mm error.
  • The same fixed-calibration architectural pattern could be applied to other inverse problems where direct regression allows networks to bypass the intended physical decoding step.

Load-bearing premise

The photorealistic synthetic benchmark with 50 objects at 1.5-2.1 m faithfully captures the dominant failure modes and noise statistics of real long-range fringe projection hardware.

What would settle it

Capture real long-range fringe images with the same hardware geometry, run both the baseline UNet and PhiCalNet on them, and check whether the 3.3 times error reduction holds and whether the three diagnostic probes still isolate object-boundary shape priors as the dominant mechanism.

Figures

Figures reproduced from arXiv: 2606.17093 by Adam Haroon, Anush Lakshman, Beiwen Li, Cody Fleming.

Figure 1
Figure 1. Figure 1: Virtual camera-projector calibration setup with a pinhole camera model, rectangular light-source projector, calibration board, and matte background plane. All objects in FPP-ML-Bench use consistent matte ma￾terial properties (roughness=0.95, specular=0.15, AO-to￾diffuse=0.95) representative of typical structured light scan￾ning. (Zapico, Meana, Cuesta and Mateos, 2023; Ou, Xu, Gan, He, Li, Qu, Zhang and Ca… view at source ↗
Figure 2
Figure 2. Figure 2: Depth map visualizations for three normalization strategies on the same object (wooden boards). From left to right: raw depth (0–2026 mm), global normalized depth (0–2.026 m), and individual normalized depth ([0, 1] range mapped to 1561–2026 mm). All use the same underlying depth data, differing only in normalization. This strategy reduces the numerical range to approxi￾mately 0–2 m, which may improve nume… view at source ↗
Figure 3
Figure 3. Figure 3: Overall error distributions (MAE and RMSE) across 30 test samples for three normalization strategies. Top: Raw depth shows high variance with mean MAE 35.20 mm. Middle: Global normalized depth reduces error to 14.81 mm mean MAE. Bottom: Individual normalized depth achieves 2.30 mm mean MAE with tight distribution, demonstrating superior and consistent performance. A. Haroon et al.: Preprint submitted to El… view at source ↗
Figure 4
Figure 4. Figure 4: Single-shot depth reconstruction for the magazine stack object comparing three normalization strategies. From top to bottom: raw depth (426.50 mm object MAE), global normalized depth (52.77 mm object MAE), individual normalized depth (19.17 mm object MAE). Each row shows, from left to right: input fringe pattern, ground truth depth, predicted depth, and absolute error map (clipped at 95th percentile for vi… view at source ↗
Figure 5
Figure 5. Figure 5: Overall error distributions for background-removed fringe inputs. Compare with [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Impact of background fringe removal across three normalization strategies for the same magazine stack object. From top to bottom: raw depth (479.93 mm object MAE), global normalized depth (323.03 mm object MAE), individual normalized depth (66.59 mm object MAE). Background removal degrades all strategies: raw 1.1×, global 6.1×, individual 3.5× worse. Note severe boundary artifacts and inconsistent depth pr… view at source ↗
Figure 7
Figure 7. Figure 7: Overall error distributions for single-orientation multi-frame training conditions. Top: horizontal only (1.40 mm mean MAE). Bottom: vertical only (1.26 mm mean MAE). Both single-orientation conditions produce tight, low-error distributions comparable to each other. 2. L1 loss: Mean absolute error computed over all pixels, which is less sensitive to outliers than RMSE: L1 = 1 𝐻𝑊 ∑ 𝑊 𝑢=1 ∑ 𝐻 𝑣=1 | | | 𝐷̂(𝑢… view at source ↗
Figure 8
Figure 8. Figure 8: Overall error distributions for mixed multi-frame training conditions. Top: all sinusoidal (2.70 mm mean MAE). Bottom: all frames (2.14 mm mean MAE). Mixing fringe orientations or including non-sinusoidal patterns increases both error magnitude and variance compared to single-orientation training ( [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Overall error distributions for loss function comparison (part 1). Top: RMSE baseline (2.30 mm MAE). Middle: L1 (2.03 mm MAE). Bottom: Masked RMSE (122.38 mm MAE), exhibiting catastrophic scale drift with errors orders of magnitude larger than unmasked losses. A. Haroon et al.: Preprint submitted to Elsevier Page 16 of 44 [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Overall error distributions for loss function comparison (part 2). Top: Masked L1 (147.28 mm MAE), also exhibiting catastrophic scale drift. Middle: Hybrid L1 𝛼 = 0.7 (3.31 mm MAE, optimal), balancing object focus with scale stability. Bottom: Hybrid L1 𝛼 = 0.9 (4.66 mm MAE), slightly worse due to stronger masking weight. A. Haroon et al.: Preprint submitted to Elsevier Page 17 of 44 [PITH_FULL_IMAGE:fig… view at source ↗
Figure 11
Figure 11. Figure 11: Single-shot depth reconstruction for the power drill object under three representative loss functions. Top: RMSE baseline (16.70 mm object MAE). Middle: Hybrid L1 with 𝛼 = 0.7 (12.21 mm object MAE, 27% improvement, best overall). Bottom: Masked RMSE (23.93 mm object MAE) shows scale drift artifacts in background despite lower nominal object error. prevents pathological solutions while the dominant masked … view at source ↗
Figure 12
Figure 12. Figure 12: Overall error distributions for architecture comparison (part 1). Top: UNet (3.31 mm overall MAE, best). Bottom: TransUNet (3.56 mm overall MAE). Both trained with the optimal configuration (individual normalization + Hybrid L1 𝛼 = 0.7) [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Overall error distributions for architecture comparison (part 2). Top: ResUNet (3.73 mm overall MAE). Bottom: Pix2Pix (5.48 mm overall MAE, worst). UNet ( [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Single-shot depth reconstruction for the container bottle object across four architectures. From top to bottom: UNet (21.65 mm object MAE, 27.08 mm object RMSE), TransUNet (15.85 mm object MAE, 23.14 mm object RMSE), ResUNet (40.73 mm object MAE, 46.89 mm object RMSE), Pix2Pix (39.27 mm object MAE, 51.16 mm object RMSE, worst performance). All models capture coarse geometry but fail on fine-scale accuracy… view at source ↗
Figure 15
Figure 15. Figure 15: Linear probing analysis comparing how easily geometric (edges) versus physical (depth) information can be decoded from UNet activations at each layer. Edges (green) are consistently easier to predict than depth values (red), with an average ratio of 2.8× across all layers. 4.7.2. GradCAM Attention Analysis While linear probing reveals what information is en￾coded in network activations, it does not direct… view at source ↗
Figure 16
Figure 16. Figure 16: GradCAM attention analysis showing correlation of GradCAM heatmaps with edge maps versus fringe patterns across layers (30 test samples). The network attends more strongly to geometric boundaries than fringe intensity patterns, with an average edge/fringe correlation ratio of 1.28. (a) Input (b) enc3 (c) enc4 (d) bottleneck (e) dec1 (f) dec3 (g) dec4 [PITH_FULL_IMAGE:figures/full_fig_p023_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Example GradCAM visualization showing network attention across layers for a test sample. The network progressively focuses on object boundaries, with strongest attention at edges in enc4 and dec4 layers [PITH_FULL_IMAGE:figures/full_fig_p023_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Flat plane out-of-distribution test. Left: Input fringe image of a flat plane captured at 1.8 m, within the trained depth range. Right: Network prediction showing a non-uniform depth map clustered near zero (mean = 0.088) with an extreme range [−1.19, 4.38] far outside the valid [0, 1] training range, indicating the network fails to recognize the flat surface and defaults to background-level predictions a… view at source ↗
Figure 19
Figure 19. Figure 19: GradCAM analysis on the flat plane test. Encoder layers (enc3, enc4) activate selectively on the fringes inside the rectangular front plane, ignoring those on the surrounding background. Decoder layers show minimal spatial attention. The network uses the rectangular boundary to define a region of interest but cannot translate the fringe pattern within it into a uniform depth estimate. form: Proposition 1 … view at source ↗
Figure 20
Figure 20. Figure 20: PhiCalNet architecture. Top row (learning): the input fringe 𝐼 is fed to a trainable UNet backbone (31 M parameters; encoder–bottleneck–decoder with a final 1×1 convolution) producing a two-channel field (𝑠, 𝑐) = (sin 𝜙, cos 𝜙). Bottom row (physics): a fixed phase head maps (𝑠, 𝑐) to the wrapped phase 𝜙̂ via 𝐿2 unit-circle projection and atan2; 𝜙̂ and the auxiliary oracle fringe order 𝑘gt then feed a fixe… view at source ↗
Figure 21
Figure 21. Figure 21: PhiCalNet predictions on the clean test sample container_bottle_A180 (object MAE 1.63 mm, RMSE 2.28 mm). Shallow depth variation keeps the object within a few fringe bands. The depth error panel is clipped at the 95th object percentile for visibility; the sin 𝜙 and cos 𝜙 error panels use a fixed symmetric range in [−1,+1] so the near-zero bulk appears near-white and any wrap-boundary sign flips stand out … view at source ↗
Figure 22
Figure 22. Figure 22: PhiCalNet predictions on the failure test sample power_drill_A0 (object MAE 13.18 mm, RMSE 46.28 mm). Complex multi-level geometry makes the object surface cross many ±𝜋 wrap lines, and the resulting wrap-boundary sign errors accumulate into the horizontal stripes seen in the depth error map. Colorbar conventions follow [PITH_FULL_IMAGE:figures/full_fig_p030_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: PhiCalNet linear probing analysis across all four targets. Phase (sin 𝜙, cos 𝜙) is the most decodable target everywhere except the bottleneck and is best decoded at enc3. Depth is the least decodable target by an order of magnitude, despite being the network’s deployment output. The bottleneck spike for the phase targets is a resolution artifact (bottleneck activations live at 7.5 × 7.5 and the probe targ… view at source ↗
Figure 24
Figure 24. Figure 24: PhiCalNet GradCAM correlation summary in the two modes. Phase mode (top) gives an average edge/fringe ratio of 1.06, with fringe correlation matching or exceeding edge correlation at the layers that drive the output (enc4, dec4). Depth mode (bottom) gives an average ratio of 1.54, edge-favored, because gradients flowing back through the analytical calibration layer inherit the region/extent sensitivity of… view at source ↗
Figure 25
Figure 25. Figure 25: PhiCalNet GradCAM visualization for test sample container_bottle_A60 in the two modes. Phase mode (top row, gradient target = ̂𝑠 + ̂𝑐) and depth mode (bottom row, gradient target = calibrated depth) are produced from identical activations and weights; the difference between rows is attributable to the gradient target alone. Phase-mode attention covers the fringes within and around the object body; depth-m… view at source ↗
Figure 26
Figure 26. Figure 26: PhiCalNet flat-plane out-of-distribution test. (a) Input fringe image of a flat plane captured at 1.8 m, within the trained depth range. (b, c) Predicted sin 𝜙 and cos 𝜙 recover the fringe pattern everywhere fringes are visible, including over the background the network never saw at train time. The quarter-period offset between their bands is the expected sin/cos quadrature, confirming the outputs do not … view at source ↗
Figure 27
Figure 27. Figure 27: PhiCalNet GradCAM analysis on the flat-plane input, phase mode. Encoder layers (enc3, enc4) and decoder layers (dec1, dec3, dec4) both overlay attention on the horizontal fringe pattern across the full image, including the gray background fringes that lie outside the training distribution. Compare with the UNet baseline ( [PITH_FULL_IMAGE:figures/full_fig_p037_27.png] view at source ↗
read the original abstract

Learning-based single-shot fringe projection profilometry (FPP) has been studied mostly at close range. The long-range regime (standoff beyond 1 m) remains largely unaddressed: inverse-square intensity falloff lowers fringe signal-to-noise ratio and degrades physical ground truth, the single-shot problem is ill-posed because fringe-order information is absent from one image, and these architectures have not been studied mechanistically. We present a diagnose-repair-verify study using mechanistic interpretability (MI) and conformal uncertainty quantification (UQ) as convergent diagnostics: they agree on one physical failure locus, driving and verifying an architectural repair. On a photorealistic synthetic benchmark (15,600 fringe images, 50 objects at 1.5-2.1 m), a best UNet baseline reaches 14.54 mm object mean absolute error (MAE). Three probes (linear probing, Grad-CAM, flat-plane out-of-distribution test) converge: the baseline solves the task via object-boundary shape priors rather than fringe-phase decoding. We repair this with PhiCalNet, which outputs wrapped phase rather than depth and applies a fixed differentiable calibration layer mapping phase to depth, removing the shape-prior solution from the hypothesis space architecturally rather than by a loss penalty. A physics-informed loss that enforces the same physics as a soft penalty on a depth-regressing network yields no measurable gain, isolating the architecture as the operative factor. PhiCalNet reduces object MAE 3.3x to 4.46 mm; the residual is carried by 0.103% of pixels at the +/-pi wrap discontinuity. Pixel-wise conformal UQ confirms the diagnosis: rejecting the top 5% of object pixels by snapshot disagreement cuts PhiCalNet RMSE by 64% (20.6->7.4 mm) versus 3.5% for the baseline. MI and UQ converge on the same failure locus.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that in long-range single-shot fringe projection profilometry, standard UNet baselines solve the ill-posed task via object-boundary shape priors rather than fringe-phase decoding. This is diagnosed on a 15,600-image photorealistic synthetic benchmark (50 objects at 1.5-2.1 m) using convergent mechanistic interpretability probes (linear probing, Grad-CAM, flat-plane OOD test) and conformal UQ. The authors introduce PhiCalNet, which regresses wrapped phase and applies a fixed differentiable calibration layer to depth, architecturally excluding the shape-prior solution; a physics-informed loss baseline on depth regression yields no gain. PhiCalNet reduces object MAE from 14.54 mm to 4.46 mm (3.3×), with UQ filtering the top 5% pixels cutting RMSE by 64%.

Significance. If the central diagnosis holds, the work supplies a concrete example of shortcut learning in a physics-constrained vision task and shows that removing a solution from the hypothesis space via architecture can outperform soft loss penalties. Strengths include the internal control (physics-informed loss), agreement between MI and UQ on the failure locus, and the quantitative UQ result. The long-range FPP setting is underexplored; the approach could generalize to other single-shot phase problems where physical consistency must be enforced without fitting parameters.

major comments (2)
  1. [Benchmark description and § on experiments] Benchmark and experimental sections: All diagnosis, repair, and verification (including the three probes and UQ filtering) are performed exclusively on the photorealistic synthetic renderer with no reported real-camera long-range FPP images or hardware validation. This is load-bearing for the claim that object-boundary shape priors are the operative failure mode, because the probes could respond to renderer-specific cues (perfect edges, idealized inverse-square falloff, absent sensor noise/speckle) rather than physical long-range phenomena.
  2. [PhiCalNet architecture section] PhiCalNet description and ablation: The fixed differentiable calibration layer is presented as removing the shape-prior solution from the hypothesis space, yet the manuscript does not report whether this layer is derived from the same synthetic renderer parameters or from independent calibration; if the former, the architectural constraint may still embed renderer-specific assumptions.
minor comments (2)
  1. [Data generation paragraph] The abstract states 15,600 images but the full text should expand the object selection criteria, rendering parameter ranges, and exact noise model to support reproducibility claims.
  2. [Methods] Notation for wrapped phase output and the calibration mapping could be formalized with an equation to clarify that no learned parameters remain in the phase-to-depth step.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback and for recognizing the potential significance of the work. We address each major comment below with clarifications and proposed revisions.

read point-by-point responses

  1. Referee: [Benchmark description and § on experiments] Benchmark and experimental sections: All diagnosis, repair, and verification (including the three probes and UQ filtering) are performed exclusively on the photorealistic synthetic renderer with no reported real-camera long-range FPP images or hardware validation. This is load-bearing for the claim that object-boundary shape priors are the operative failure mode, because the probes could respond to renderer-specific cues (perfect edges, idealized inverse-square falloff, absent sensor noise/speckle) rather than physical long-range phenomena.

    Authors: We agree this is a substantive limitation. The synthetic renderer was constructed to model the dominant physical effects of the long-range regime (inverse-square falloff, reduced SNR, single-shot ambiguity), and the three MI probes plus conformal UQ converge on the same failure mode. Nevertheless, renderer-specific artifacts cannot be fully ruled out without real hardware data. We will add an explicit limitations subsection discussing the domain gap and the need for future physical validation; no real-camera images are available in the current study. revision: partial

  2. Referee: [PhiCalNet architecture section] PhiCalNet description and ablation: The fixed differentiable calibration layer is presented as removing the shape-prior solution from the hypothesis space, yet the manuscript does not report whether this layer is derived from the same synthetic renderer parameters or from independent calibration; if the former, the architectural constraint may still embed renderer-specific assumptions.

    Authors: The calibration layer implements the standard phase-to-depth mapping using the known camera/projector intrinsics and baseline geometry employed by the renderer; these parameters are obtained from the same geometric calibration procedure that would be performed on physical hardware and are independent of object shape or texture. We will revise the architecture section to state this explicitly and to note that the layer contains no learned parameters or object-specific information. revision: yes

standing simulated objections not resolved
  • No real-camera long-range FPP images or hardware validation are available in the current study.

Circularity Check

0 steps flagged

No circularity: derivation relies on empirical diagnostics and architectural constraint, not self-referential reduction

full rationale

The paper's core argument—that the UNet baseline exploits object-boundary shape priors (diagnosed via linear probing, Grad-CAM, and flat-plane OOD tests) and that PhiCalNet removes this via an architectural change to wrapped-phase output plus fixed calibration layer—is supported by an internal control (physics-informed loss yields no gain) and conformal UQ verification on the same benchmark. No equations, fitted parameters, or self-citations are shown to reduce the 3.3× MAE improvement to a tautology or prior result by construction. The derivation chain is self-contained against the reported synthetic benchmark and interpretability probes.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the assumption that the synthetic data generator produces fringe images whose only reliable cue is phase (not object geometry) and that the fixed calibration layer implements the true physical mapping without learnable parameters. No free parameters, axioms, or invented entities are explicitly introduced beyond standard supervised learning assumptions.

pith-pipeline@v0.9.1-grok · 5901 in / 1462 out tokens · 30635 ms · 2026-06-27T04:51:27.327467+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

76 extracted references · 3 canonical work pages · 2 internal anchors

  1. [1]

    2016 , publisher=

    High-Speed 3D Imaging with Digital Fringe Projection Techniques , author=. 2016 , publisher=

    2016

  2. [2]

    Advances in optics and photonics , volume=

    Structured-light 3D surface imaging: a tutorial , author=. Advances in optics and photonics , volume=. 2011 , publisher=

    2011

  3. [3]

    Wiley Encyclopedia of Electrical and Electronics Engineering , pages=

    Structured Light Techniques and Applications , author=. Wiley Encyclopedia of Electrical and Electronics Engineering , pages=. 2016 , publisher=

    2016

  4. [4]

    Optics and lasers in engineering , volume=

    Recent progresses on real-time 3D shape measurement using digital fringe projection techniques , author=. Optics and lasers in engineering , volume=. 2010 , publisher=

    2010

  5. [5]

    Scientific reports , volume=

    Temporal phase unwrapping using deep learning , author=. Scientific reports , volume=. 2019 , publisher=

    2019

  6. [6]

    Optics express , volume=

    Deep neural networks for single shot structured light profilometry , author=. Optics express , volume=. 2019 , publisher=

    2019

  7. [7]

    Light: Science & Applications , volume=

    Deep learning in optical metrology: a review , author=. Light: Science & Applications , volume=. 2022 , publisher=

    2022

  8. [8]

    Advanced Photonics Nexus , number=

    Kaiqiang Wang and Qian Kemao and Jianglei Di and Jianlin Zhao , title=. Advanced Photonics Nexus , number=. 2022 , doi=

    2022

  9. [9]

    2019 , author=

    Fringe pattern denoising based on deep learning , journal=. 2019 , author=

    2019

  10. [10]

    Optical Engineering , volume=

    Hformer: Hybrid convolutional neural network transformer network for fringe order prediction in phase unwrapping of fringe projection , author=. Optical Engineering , volume=. 2022 , publisher=

    2022

  11. [11]

    Optics Express , volume=

    Single-shot fringe projection profilometry based on deep learning and computer graphics , author=. Optics Express , volume=. 2021 , publisher=

    2021

  12. [12]

    Sensors , volume=

    Single-shot 3D shape reconstruction using structured light and deep convolutional neural networks , author=. Sensors , volume=. 2020 , publisher=

    2020

  13. [13]

    Optics Express , volume=

    Deep-learning-enabled single-shot fringe projection profilometry based on inner shifting-phase encoding , author=. Optics Express , volume=. 2025 , publisher=

    2025

  14. [14]

    Measurement , volume=

    End-to-end single-shot composite color FPP network for multiple separated objects reconstruction , author=. Measurement , volume=. 2025 , publisher=

    2025

  15. [15]

    2021 , publisher=

    Synthetic data for deep learning , author=. 2021 , publisher=

    2021

  16. [16]

    Trends in cognitive sciences , volume=

    Next-generation deep learning based on simulators and synthetic data , author=. Trends in cognitive sciences , volume=. 2022 , publisher=

    2022

  17. [17]

    Zheng and S

    Y. Zheng and S. Wang and Q. Li and B. Li , title=. Optics Express , volume=. 2020 , publisher=

    2020

  18. [18]

    Optical Review , volume=

    Fringe projection profilometry system verification for 3D shape measurement using virtual space of game engine , author=. Optical Review , volume=. 2021 , publisher=

    2021

  19. [19]

    IEEE Access , volume=

    Measurement Simulation System of Fringe Projection Profilometry Based on Ray Tracing , author=. IEEE Access , volume=. 2023 , publisher=

    2023

  20. [20]

    Optical Review , pages=

    Deep-learning-assisted single-shot 3D shape and color measurement using color fringe projection profilometry , author=. Optical Review , pages=. 2025 , publisher=

    2025

  21. [21]

    VIRTUS-FPP: Virtual Sensor Modeling for Fringe Projection Profilometry in NVIDIA Isaac Sim , year=

    Haroon, Adam and Lakshman, Anush and Balasubramaniam, Badrinath and Li, Beiwen , journal=. VIRTUS-FPP: Virtual Sensor Modeling for Fringe Projection Profilometry in NVIDIA Isaac Sim , year=

  22. [22]

    High-resolution real-time 360

    Qian, Jiaming and Feng, Shijie and Xu, Mingzhu and Tao, Tianyang and Shang, Yuhao and Chen, Qian and Zuo, Chao , journal=. High-resolution real-time 360. 2021 , publisher=

    2021

  23. [23]

    Optical Metrology and Inspection for Industrial Applications IV , volume=

    Three-dimensional surface inspection for semiconductor components with fringe projection profilometry , author=. Optical Metrology and Inspection for Industrial Applications IV , volume=. 2016 , organization=

    2016

  24. [24]

    Precision Engineering , volume=

    Machine learning enhanced high dynamic range fringe projection profilometry for in-situ layer-wise surface topography measurement during LPBF additive manufacturing , author=. Precision Engineering , volume=. 2023 , publisher=

    2023

  25. [25]

    Measurement , volume=

    A systematic study and framework of fringe projection profilometry with improved measurement performance for in-situ LPBF process monitoring , author=. Measurement , volume=. 2022 , publisher=

    2022

  26. [26]

    Dimensional Optical Metrology and Inspection for Practical Applications XIII , volume=

    Autonomous robotic 3D scanning for smart factory planning , author=. Dimensional Optical Metrology and Inspection for Practical Applications XIII , volume=. 2024 , organization=

    2024

  27. [27]

    Robotics and Computer-Integrated Manufacturing , volume=

    Robotic measurement system based on cooperative optical profiler integrating fringe projection with photometric stereo for highly reflective workpiece , author=. Robotics and Computer-Integrated Manufacturing , volume=. 2024 , publisher=

    2024

  28. [28]

    What Is Isaac Sim? , year=

  29. [29]

    ACM transactions on graphics (TOG) , volume=

    Optix: a general purpose ray tracing engine , author=. ACM transactions on graphics (TOG) , volume=. 2010 , publisher=

    2010

  30. [30]

    2025 , howpublished=

    Universal Scene Description (. 2025 , howpublished=

    2025

  31. [31]

    2025 , howpublished=

    2025

  32. [32]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Close the Sim2real Gap via Physically-based Structured Light Synthetic Data Simulation , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization=

    2024

  33. [33]

    2015 , publisher=

    Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas , booktitle=. 2015 , publisher=

    2015

  34. [34]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Image-to-Image Translation with Conditional Adversarial Networks , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  35. [35]

    High-Resolution Image Synthesis and Semantic Manipulation with Conditional

    Wang, Ting-Chun and Liu, Ming-Yu and Zhu, Jun-Yan and Tao, Andrew and Kautz, Jan and Catanzaro, Bryan , booktitle=. High-Resolution Image Synthesis and Semantic Manipulation with Conditional

  36. [36]

    Frontiers of Architectural Research , year=

    A dual-aspect evaluation framework for architectural-like plan generation via pix2pix series algorithms , author=. Frontiers of Architectural Research , year=

  37. [37]

    International Manufacturing Science and Engineering Conference , volume=

    Single Shot 3D Shape Measurement of Non-Volatile Data Storage Devices , author=. International Manufacturing Science and Engineering Conference , volume=. 2023 , organization=

    2023

  38. [38]

    2009 , organization=

    Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li , booktitle=. 2009 , organization=

    2009

  39. [39]

    Microsoft

    Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Doll. Microsoft. Computer Vision--ECCV 2014: 13th European Conference , pages=. 2014 , organization=

    2014

  40. [40]

    The International Journal of Robotics Research , volume=

    Yale-CMU-Berkeley dataset for robotic manipulation research , author=. The International Journal of Robotics Research , volume=. 2017 , publisher=

    2017

  41. [41]

    2025 , howpublished =

    2025

  42. [42]

    2003 , publisher=

    Multiple view geometry in computer vision , author=. 2003 , publisher=

    2003

  43. [43]

    Optics and Lasers in Engineering , volume=

    Status, challenges, and future perspectives of fringe projection profilometry , author=. Optics and Lasers in Engineering , volume=. 2020 , publisher=

    2020

  44. [44]

    Pattern recognition , volume=

    Pattern codification strategies in structured light systems , author=. Pattern recognition , volume=. 2004 , publisher=

    2004

  45. [45]

    Applied optics , volume=

    Three-dimensional vision based on a combination of Gray-code and phase-shift light projection: analysis and compensation of the systematic errors , author=. Applied optics , volume=. 1999 , publisher=

    1999

  46. [46]

    Optics and lasers in engineering , volume=

    Calibration of fringe projection profilometry: A comparative review , author=. Optics and lasers in engineering , volume=. 2021 , publisher=

    2021

  47. [47]

    Materials , volume=

    Zapico, Pablo and Meana, Victor and Cuesta, Eduardo and Mateos, Sabino , title=. Materials , volume=. 2023 , doi=

    2023

  48. [48]

    Micromachines , volume=

    Ou, Jia and Xu, Tingfa and Gan, Xiaochuan and He, Xuejun and Li, Yan and Qu, Jiansu and Zhang, Wei and Cai, Cunliang , title=. Micromachines , volume=. 2022 , doi=

    2022

  49. [49]

    Optical Metrology and Inspection for Industrial Applications III , volume=

    Experimental study for the influence of surface characteristics on the fringe patterns , author=. Optical Metrology and Inspection for Industrial Applications III , volume=. 2014 , organization=

    2014

  50. [50]

    2017 IEEE/RSJ international conference on intelligent robots and systems (IROS) , pages=

    Domain randomization for transferring deep neural networks from simulation to the real world , author=. 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS) , pages=. 2017 , organization=

    2017

  51. [51]

    TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation

    Transunet: Transformers make strong encoders for medical image segmentation , author=. arXiv preprint arXiv:2102.04306 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  52. [52]

    Distill , volume=

    Zoom In: An Introduction to Circuits , author=. Distill , volume=. 2020 , doi=

    2020

  53. [53]

    Understanding intermediate layers using linear classifier probes

    Understanding intermediate layers using linear classifier probes , author=. arXiv preprint arXiv:1610.01644 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  54. [54]

    and Cogswell, Michael and Das, Abhishek and Vedantam, Ramakrishna and Parikh, Devi and Batra, Dhruv , booktitle=

    Selvaraju, Ramprasaath R. and Cogswell, Michael and Das, Abhishek and Vedantam, Ramakrishna and Parikh, Devi and Batra, Dhruv , booktitle=. Grad-

  55. [55]

    Nature Machine Intelligence , volume=

    Shortcut learning in deep neural networks , author=. Nature Machine Intelligence , volume=. 2020 , publisher=

    2020

  56. [56]

    Journal of Computational physics , volume=

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , author=. Journal of Computational physics , volume=. 2019 , publisher=

    2019

  57. [57]

    Nature Reviews Physics , volume=

    Physics-informed machine learning , author=. Nature Reviews Physics , volume=. 2021 , publisher=

    2021

  58. [58]

    Advances in neural information processing systems , volume=

    Characterizing possible failure modes in physics-informed neural networks , author=. Advances in neural information processing systems , volume=

  59. [59]

    Computer Methods in Applied Mechanics and Engineering , volume=

    Respecting causality for training physics-informed neural networks , author=. Computer Methods in Applied Mechanics and Engineering , volume=. 2024 , publisher=

    2024

  60. [60]

    Radio science , volume=

    Satellite radar interferometry: Two-dimensional phase unwrapping , author=. Radio science , volume=. 1988 , publisher=

    1988

  61. [61]

    Optics and lasers in engineering , volume=

    Temporal phase unwrapping algorithms for fringe projection profilometry: A comparative review , author=. Optics and lasers in engineering , volume=. 2016 , publisher=

    2016

  62. [62]

    Foundations and Trends in Machine Learning , volume=

    Conformal Prediction: A Gentle Introduction , author=. Foundations and Trends in Machine Learning , volume=. 2023 , doi=

    2023

  63. [63]

    and Weinberger, Kilian Q

    Huang, Gao and Li, Yixuan and Pleiss, Geoff and Liu, Zhuang and Hopcroft, John E. and Weinberger, Kilian Q. , booktitle=. Snapshot Ensembles: Train 1, Get

  64. [64]

    arXiv preprint arXiv:2601.02572 , year=

    Heteroscedastic Snapshot-Ensemble Uncertainty Quantification for Single-Shot Fringe Projection Profilometry , author=. arXiv preprint arXiv:2601.02572 , year=

    work page arXiv

  65. [65]

    What Uncertainties Do We Need in

    Kendall, Alex and Gal, Yarin , booktitle=. What Uncertainties Do We Need in

  66. [66]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  67. [67]

    2005 , publisher=

    Algorithmic Learning in a Random World , author=. 2005 , publisher=

    2005

  68. [68]

    Information Fusion , volume=

    A Review of Uncertainty Quantification in Deep Learning: Techniques, Applications and Challenges , author=. Information Fusion , volume=. 2021 , publisher=

    2021

  69. [69]

    Advanced photonics , volume=

    Fringe pattern analysis using deep learning , author=. Advanced photonics , volume=. 2019 , publisher=

    2019

  70. [70]

    Applied optics , volume=

    Real-time 3D shape measurement using 3LCD projection and deep machine learning , author=. Applied optics , volume=. 2019 , publisher=

    2019

  71. [71]

    Journal of Imaging , volume=

    Deep learning for single-shot structured light profilometry: A comprehensive dataset and performance analysis , author=. Journal of Imaging , volume=. 2024 , publisher=

    2024

  72. [72]

    Light: Science & Applications , volume=

    Modeling the measurement precision of fringe projection profilometry , author=. Light: Science & Applications , volume=. 2023 , publisher=

    2023

  73. [73]

    International Journal of Medical Informatics , volume=

    Explainability and uncertainty: Two sides of the same coin for enhancing the interpretability of deep learning models in healthcare , author=. International Journal of Medical Informatics , volume=. 2025 , publisher=

    2025

  74. [74]

    Sciurus: Shared circuits for interpretable uncertainty representations in language models , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2025

  75. [75]

    and Adam Haroon and Beiwen Li , title =

    Anush Lakshman S. and Adam Haroon and Beiwen Li , title =. Photonic Instrumentation Engineering XIII , editor =. 2026 , doi =

    2026

  76. [76]

    Journal of biomedical optics , volume=

    Deep-learning-based endoscopic single-shot fringe projection profilometry , author=. Journal of biomedical optics , volume=. 2025 , publisher=

    2025