---

title: >- [Paper Note] Modeling and Compensating Phase Error in High-speed 3D Reconstruction description: >- [ECCV 2026][3D Vision][Fringe Projection] Proposes a theoretical model for dithering-induced phase errors in binary defocusing and develops an Error Diffusion Fusion (EDF) scheme leveraging forward and inverse Floyd-Steinberg kernels for inter-fringe self-compensation. tags: - ECCV 2026 - 3D Vision - Structured Light - Binary Defocusing - Error Diffusion - Phase Error Compensation date: 2026-09-19 content_hash: c0a3c29df31bb2d3

Modeling and Compensating Phase Error in High-speed 3D Reconstruction

Conference: ECCV 2026
Paper: ECCV Official
Cached Source: ../paper_cache/ECCV2026/eccv-4008.txt
Area: 3D Vision
Keywords: Structured light 3D reconstruction, binary defocusing profilometry, error diffusion dithering, four-step phase shifting, phase error self-compensation

TL;DR

The paper formulates an explicit theoretical model revealing that error diffusion dithering in binary defocusing introduces systematic phase errors correlated with diffusion direction but invariant to phase-shifting steps and defocusing levels, and proposes an Error Diffusion Fusion (EDF) strategy pairing forward and inverse Floyd-Steinberg kernels in four-step phase shifting to achieve complete inter-fringe error cancellation at zero runtime overhead.

Background & Motivation

Fringe projection profilometry (FPP) is a cornerstone of active optical 3D sensing due to its high spatial resolution, non-contact nature, and sub-pixel accuracy. Standard phase-shifting profilometry operates under the fundamental premise that the target object remains static across the multi-frame projection sequence. However, in high-speed dynamic environments, continuous object motion introduces temporal misalignment among sequential phase-shifted frames, corrupting the phase relationships and causing severe motion-induced motion artifacts and reconstruction distortions. While analytical motion estimation and single-shot Fourier transforms offer partial remedies, elevating the projection frame rate to approximately freeze dynamic scenes remains the most reliable and direct solution.

To overcome the projection speed ceiling of digital light processing (DLP) systems—which typically operate at around 120 Hz in standard 8-bit grayscale sinusoidal mode—projector binary defocusing techniques convert 8-bit continuous fringes into 1-bit binary patterns. By leveraging the optical low-pass filtering property of a slightly defocused projector lens, the binary patterns are smoothed into quasi-sinusoidal profiles. This binary paradigm bypasses non-linear gamma calibration, matches the ultra-fast switching capability of digital micromirror devices (DMDs) reaching several kilohertz, and seamlessly integrates with high-speed micro-LED arrays. Nevertheless, the binarization process inevitably injects high-order harmonics and quantization noise. Among various binary encoding strategies, dithering algorithms like Floyd-Steinberg (FS) error diffusion yield superior sinusoidal approximation under shallow defocus. Yet, prior error diffusion methods ignore the asymmetric spatial coupling between the unidirectional error propagation and the 1D orientation of the fringe pattern, resulting in noticeable systematic phase shifts in the recovered phase.

Existing optimization efforts have predominantly focused on empirically modifying the diffusion kernel layout or applying post-reconstruction constant phase offset adjustments. However, kernel restructuring fails to eliminate directional phase bias, while constant offset compensation assumes asymptotic behavior that fails severely on short-period fringes. This paper attacks the problem from a dual perspective: since unidirectional diffusion creates a distinct directional phase bias, a spatially mirrored inverse diffusion kernel should produce a strictly anti-symmetric error profile. Core idea: theoretically formulate the closed-form phase error of error diffusion dithering to uncover its invariance to shifting steps and defocusing amounts, and design a four-step phase-shifting Error Diffusion Fusion (EDF) strategy where paired forward and inverse Floyd-Steinberg kernels enable seamless algebraic self-cancellation of systematic phase errors during standard phase retrieval.

Method

Overall Architecture

The proposed Error Diffusion Fusion (EDF) framework operates as an offline pre-correction binarization strategy. The complete pipeline comprises three interconnected stages: First, a set of four-step phase-shifted 8-bit sinusoidal fringe patterns is partitioned into two complementary pairs; the first pair (\(I_1, I_2\)) is binarized using the standard forward Floyd-Steinberg (FS) dithering kernel, while the second pair (\(I_3, I_4\)) is binarized using the inverse Floyd-Steinberg (IFS) kernel. Second, the generated binary patterns are projected at high speeds through a defocused optical projector onto the moving or static scene, where the camera synchronously captures the optically filtered, deformed fringe patterns. Third, the phase is retrieved using standard four-step phase-shifting arctangent reconstruction. Because the intensity errors in the two groups carry opposite signs, the algebraic subtractions in the numerator and denominator completely cancel out the error terms, restoring high-precision wrapped phase without any online computational overhead.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: 8-bit four-step phase-shifted fringes<br/>I1, I2, I3, I4"] --> B["Stage 1: Dual-Kernel Binarization<br/>I1, I2 via FS; I3, I4 via inverse IFS"]
    B --> C["Stage 2: High-Speed Defocused Projection<br/>Gaussian optical smoothing yields pseudo-sinusoids"]
    C --> D["Stage 3: Algebraic Inter-Fringe Self-Compensation<br/>Difference operations cancel out β error terms"]
    D --> E["Output: Systematic-error-free 3D point cloud"]

Key Designs

1. Theoretical modeling of error diffusion phase error: establishing step- and defocus-invariant characteristics

Standard dithering algorithms process pixels sequentially via raster scanning, distributing quantization errors to adjacent unprocessed neighbors. Because vertical sinusoidal fringes maintain uniform intensity vertically, error diffusion primarily translates into horizontal intensity perturbations. The recovered fringe intensity can be analytically decomposed into the fundamental sinusoidal component and error components modulated by the forward and backward diffusion parameters \(k_1\) and \(k_2\):

\[I_n = a + \alpha \cos(\phi - \delta_n) \cos\frac{2\pi}{P} + \beta \sin(\phi - \delta_n) \sin\frac{2\pi}{P}\]

where \(P\) denotes the fringe period, and \(\alpha, \beta\) are combined trigonometric coefficients governed by the kernel. By treating the sine component \(\Delta I_n = \beta \sin(\phi - \delta_n) \sin\frac{2\pi}{P}\) as an intensity disturbance and substituting it into the temporal noise sensitivity model \(\frac{\partial \phi}{\partial I_n} = \frac{2}{N\alpha} \sin(\phi + \delta_n) / \cos\frac{2\pi}{P}\), the total phase error simplifies to:

\[\Delta \phi = \frac{\beta}{\alpha} \tan\frac{2\pi}{P}\]

This closed-form formulation reveals fundamental physical properties: the phase error \(\Delta \phi\) is independent of the number of phase-shifting steps \(N\); because further optical defocus acts as a Gaussian point spread function whose frequency response \(H_1(\omega, 0)\) scales both \(\alpha\) and \(\beta\) identically, the error is invariant to the degree of defocusing once defocusing is sufficient; and for large fringe periods (\(P > 24\)), \(\tan(2\pi/P) \approx 2\pi/P\), reducing the horizontal coordinate displacement to an asymptotic constant \(\Delta u = \Delta \phi \cdot \frac{P}{2\pi} \approx \frac{\beta}{\alpha}\).

2. Inverse Floyd-Steinberg (IFS) kernel formulation: generating anti-symmetric error profiles

The standard Floyd-Steinberg kernel assigns fractional error weights of \(7/16\) to the right neighbor, and \(3/16, 5/16, 1/16\) to the bottom-left, bottom, and bottom-right neighbors. This geometric asymmetry imposes a directional bias in the horizontal intensity trajectory. The paper constructs an inverse Floyd-Steinberg (IFS) kernel by horizontally mirroring the error distribution weights: assigning \(7/16\) to the left neighbor (under reversed traversal) or exchanging the effective error propagation coefficients \(k_1\) and \(k_2\).

Under this configuration, the sign of the sinusoidal error component in the recovered fringe intensity is flipped. The resulting phase error becomes strictly anti-symmetric:

\[\Delta \phi_{\text{IFS}} = -\frac{\beta}{\alpha} \tan\frac{2\pi}{P}\]

Cross-sectional 1D error analysis under identical Gaussian defocusing (\(\sigma = 3\)) verifies that the intensity error of IFS is the exact mirror image of standard FS, confirming that the systematic phase deviation can be precisely inverted via directional kernel design.

3. Four-step inter-fringe self-compensation (EDF): zero-overhead algebraic error cancellation

When error diffusion adds interference variables, the captured fringe intensities across four phase-shifting steps (\(\delta_n \in \{0, \pi/2, \pi, 3\pi/2\}\)) expand to five unknowns. Rather than increasing the projection sequence to five steps, the Error Diffusion Fusion (EDF) strategy groups the four fringes: the first two (\(I_1, I_2\)) are binarized via the standard FS kernel, while the last two (\(I_3, I_4\)) are binarized via the inverse IFS kernel. The resulting captured intensities satisfy:

\[\begin{cases} I_1 = a + \beta \sin\phi \sin\frac{2\pi}{P} + \alpha \cos\phi \cos\frac{2\pi}{P} \\ I_2 = a - \beta \cos\phi \sin\frac{2\pi}{P} + \alpha \sin\phi \cos\frac{2\pi}{P} \\ I_3 = a + \beta \sin\phi \sin\frac{2\pi}{P} - \alpha \cos\phi \cos\frac{2\pi}{P} \\ I_4 = a - \beta \cos\phi \sin\frac{2\pi}{P} - \alpha \sin\phi \cos\frac{2\pi}{P} \end{cases}\]

Applying the standard four-step phase calculation formula \(\phi = \arctan\frac{I_2 - I_4}{I_1 - I_3}\) reveals that the unknown interference terms with coefficient \(\beta\) cancel out identically:

\[I_1 - I_3 = 2\alpha \cos\phi \cos\frac{2\pi}{P}, \quad I_2 - I_4 = 2\alpha \sin\phi \cos\frac{2\pi}{P}\]

Dividing the two differences eliminates the shared scale factor \(2\alpha \cos(2\pi/P)\), directly isolating \(\tan\phi\) and retrieving the ground-truth phase without any residual binarization bias. Because all error compensation occurs during offline fringe binarization, online 3D measurement incurs zero computational cost.

Key Experimental Results

The experimental platform incorporates a DLP LightCrafter 4500 projector (resolution \(912 \times 1140\), binary frame rate up to 4225 Hz) and a synchronized Basler acA640-750um camera (\(640 \times 480\), 751 fps). Reconstructed point clouds from standard 8-bit grayscale sinusoidal fringes provide the Ground Truth baseline. Fringe patterns are evaluated at a period of \(P = 16\) pixels under slight defocusing against five methods: standard Floyd-Steinberg (FS), Optimized Kernel (OK), Half-Truncated Fringe (HTF), Fixed Offset Compensation (FOC), and Error Fitting (EF).

Main Results: Smooth Geometric Surface Reconstruction

Evaluated on three representative smooth models (a stepped Block, a Horse model, and a Dog sculpture) across Average Error (AE) and Root Mean Square Error (RMSE).

Object	Metric (mm)	FS [30]	OK [26]	HTF [40]	FOC [41]	EF [12]	EDF (Ours)
Stepped Block	AE / RMSE	0.218 / 0.311	0.099 / 0.158	0.066 / 0.116	0.073 / 0.124	0.108 / 0.186	0.064 / 0.109
Horse Model	AE / RMSE	0.105 / 0.210	0.051 / 0.115	0.041 / 0.191	0.046 / 0.107	0.066 / 0.180	0.041 / 0.095
Dog Sculpture	AE / RMSE	0.164 / 0.248	0.073 / 0.130	0.055 / 0.212	0.061 / 0.116	0.082 / 0.161	0.053 / 0.104

Ablation & Robustness Study: Complex Non-Smooth Objects

Evaluated on six objects exhibiting sharp geometric discontinuities, deep depth steps, and intricate textures (Deer, Caster, Chip, and three Portrait sculptures).

Object	Metric (mm)	FS [30]	OK [26]	HTF [40]	FOC [41]	EDF (Ours)
Deer	AE / RMSE	0.147 / 0.358	0.068 / 0.274	0.052 / 0.596	0.052 / 0.259	0.047 / 0.114
Portrait 1	AE / RMSE	0.133 / 0.236	0.055 / 0.119	0.043 / 0.185	0.050 / 0.110	0.042 / 0.091
Portrait 2	AE / RMSE	0.173 / 0.282	0.073 / 0.170	0.090 / 0.896	0.071 / 0.167	0.057 / 0.155
Caster	AE / RMSE	0.193 / 0.360	0.074 / 0.237	0.126 / 1.825	0.062 / 0.228	0.049 / 0.220
Chip	AE / RMSE	0.104 / 0.199	0.062 / 0.171	0.062 / 0.680	0.065 / 0.135	0.051 / 0.110
Portrait 3	AE / RMSE	0.272 / 0.581	0.142 / 0.664	0.165 / 1.621	0.118 / 0.462	0.107 / 0.273

Additional Computational Overhead Comparison (Table 3 in Original Paper)

Under identical pattern acquisition time (0.009 s): - Pre-correction methods (offline pattern generation stage): FS (0 s), OK (0 s), HTF (0 s), EDF (0 s) - Post-correction methods (runtime phase calculation stage): FOC (0.007 s), EF (0.085 s)

Key Findings

• High geometric fidelity and boundary stability: On smooth surfaces, EDF reduces AE by 72.22% and RMSE by 58.30% compared to standard FS. While HTF achieves low AE, its truncation strategy introduces severe boundary artifacts near depth discontinuities, causing RMSE to surge up to 1.825 mm. In contrast, EDF maintains smooth, artifact-free reconstructions.
• Superiority over heuristic fixed-offset compensation: FOC assumes an asymptotic constant phase shift of 0.19 pixels; at high-frequency fringe periods (\(P = 16\)), the true phase error deviates noticeably from this asymptotic value, causing overcompensation. Across all non-smooth objects, EDF surpasses the second-best approach (FOC/OK) by 13.85% in AE and 23.77% in RMSE.
• High-speed dynamic robustness: In robotic arm dynamic tracking experiments, 8-bit projection suffers from severe motion blur and geometric distortion due to its 120 Hz projection ceiling. EDF leverages binary projection at several kilohertz to freeze the motion, reconstructing pristine 3D point clouds at all trajectory waypoints.

Highlights & Insights

• Insightful directional error modeling: Rather than treating dithering as isotropic quantization noise, the authors formally derive how 1D fringe geometry interacts with directional error diffusion, turning a complex nonlinear process into a compact trigonometric error formula.
• Cost-free algebraic self-cancellation: By embedding forward and inverse diffusion kernels into the symmetric phases of the four-step phase-shifting algorithm, systematic errors cancel out automatically during standard numerator/denominator subtraction, achieving superior accuracy without adding runtime latency.
• Hardware-native industrial compatibility: EDF operates exclusively on pre-generated binary fringe bitmaps. It integrates seamlessly into existing DLP projectors and emerging micro-LED arrays without requiring firmware modifications, GPU accelerators, or post-processing pipelines.

Limitations & Future Work

• Reliance on even-step phase shifting: The algebraic cancellation mechanism requires pairing opposite phase shifts, making it straightforward for four-step (or even-step) sequences but not directly applicable to minimal three-step phase-shifting algorithms.
• Odd-step harmonic risks: The authors discuss a double-binarization strategy (averaging forward and backward kernels within each frame) for three-step FPP, which may introduce additional high-order harmonics under shallow defocusing.
• Residual speckle under near-focused optics: Under extremely minimal defocusing, high-frequency spatial dithering residuals may remain, requiring mild spatial filtering or fine-tuned optical aperture control.

Related Work & Insights

• vs Floyd-Steinberg (FS) [30]: Standard FS provides excellent sinusoidal approximation under defocusing but suffers from systematic horizontal phase drift; EDF preserves FS's simplicity and high fidelity while completely eliminating the directional bias through kernel pairing.
• vs Fixed Offset Compensation (FOC) [41]: FOC applies an empirical post-processing offset correction (0.19 pixels) that overcompensates short-period fringes and adds runtime delay; EDF resolves error cancellation at the pattern design level, remaining accurate across arbitrary periods with zero online overhead.
• vs Half-Truncated Fringe (HTF) [40]: HTF clips negative half-waves to zero to stabilize binary transitions, but severely degrades sharp boundaries and high-curvature surfaces (inflating RMSE); EDF preserves continuous modulation, delivering robust sub-millimeter precision across both smooth and complex geometries.

Rating

• Novelty: ⭐⭐⭐⭐☆ Elegant mathematical derivation of dithering phase errors and clever inter-fringe self-compensation design.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation across smooth, non-smooth, and dynamic objects with rigorous error and latency benchmarks.
• Writing Quality: ⭐⭐⭐⭐⭐ Cohesive theoretical narrative, clear derivations, and clean experimental visualizations.
• Value: ⭐⭐⭐⭐⭐ A highly practical, zero-overhead enhancement for high-speed industrial 3D optical profilometry.

Related Papers

• [CVPR 2026] Color-Encoded Illumination for High-Speed Volumetric Scene Reconstruction
• [CVPR 2026] Clay-to-Stone: Phase-wise 3D Gaussian Splatting for Monocular Articulated Hand-Object Manipulation Modeling
• [ECCV 2026] Gaussians on Fire: High-Frequency Reconstruction of Flames
• [CVPR 2026] LoG3D: Ultra-High-Resolution 3D Shape Modeling via Local-to-Global Partitioning
• [CVPR 2025] Event Fields: Capturing Light Fields at High Speed, Resolution, and Dynamic Range