acceptodds
Desk rejected at ICLR 2027

Directional Balance: A General Design Principle for Binary Fringe Projection

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

High-speed structured-light imaging often relies on binary-defocusing fringe projection, in which binary patterns are optically low-pass filtered into approximately sinusoidal fringes. Under shallow defocus, error diffusion is the preferred way to preserve sinusoidal fidelity, but its kernel is usually selected by hand or by black-box search, so the direction-dependent phase bias is not explicitly controlled. Rather than correcting this bias at the fringe patterns, we analyze the diffusion kernel itself and introduce a normalized balance coefficient that measures the asymmetry between forward and backward error propagation. The dominant directional phase bias vanishes when this coefficient is zero. This condition yields the Guideline-Based Optimization (GBO) kernels for horizontal and vertical fringes with no additional run-time overhead and no change to the downstream reconstruction process. Unlike sequence-level inter-fringe compensation, our approach does not reorganize the phase-shifting sequence. On real measurements, GBO reduces AE by 17.84% and RMSE by 22.45% relative to the second-best baseline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Binary Fringe Projection, Directional Balance, Diffusion Kernel, Phase Error, High-Speed 3D Reconstruction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_w4f6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies error diffusion kernel design for binary defocused fringe projection. It introduces a directional balance coefficient to describe the asymmetry of error propagation along the phase gradient direction and uses this coefficient to design the GBO kernel. The method reduces directional phase bias by adjusting the diffusion kernel used during encoding, without changing the subsequent reconstruction pipeline. The authors evaluate it through simulations, real-system experiments, and additional three-step phase-shifting and dynamic examples in the appendix.

Critical strengths
  1. The proposed design principle is interpretable. The authors connect the directional balance coefficient to the first moment of the diffusion kernel along the phase gradient direction. This gives the kernel parameters a clear mathematical basis and helps explain why the direction of error propagation affects phase estimation.

  2. The paper includes real hardware experiments comparing several methods on the same projector and industrial camera system. The test objects include planar, spherical, and more complex surfaces. The appendix further reports point cloud errors in millimeters, connecting the results to practical 3D measurement accuracy.

  3. The method is relatively easy to integrate into existing systems. It only requires replacing the diffusion kernel used during encoding, while retaining the existing phase-shifting and phase-unwrapping procedures. It also does not require training an additional reconstruction network, which makes it practical.

Critical weaknesses
  1. The core theoretical derivation relies on strong approximation assumptions, including the slowly varying quantization residual and the neglect of certain contributions in the vertical direction. However, error diffusion inherently produces high-frequency, two-dimensional residuals. Without an analysis of the approximation errors, it is difficult to determine whether the derived balance condition remains valid under conditions such as weak defocus or small spatial periods.

  2. The evidence supporting generalization to real-world visual scenarios remains limited. The extensive evaluations across different spatial periods and defocus levels are primarily based on Gaussian simulations, while the real-world experiments are restricted to a single hardware setup and orthogonal stripe patterns. Furthermore, the dynamic experiments lack comparisons between different binary encoding methods at the same acquisition speed, making it difficult to assess the method's generalizability across different devices, arbitrary pattern orientations, and dynamic scenes.

Questions for authors

See weakness

2 comments
AuthorsOpenReview ↗

Theory and Modeling

We thank you for your constructive comments on the theory, experiments, and generalization. We have conducted further quantitative analyses and experiments to address the concerns raised; our responses follow below.

(1) Theoretical Validity

The theoretical approximation in this work is based on previous studies and the simulation analysis in the paper. The original binarization residual does contain high-frequency, two-dimensional spatial components. However, it should be noted that the fundamental mechanism of binary defocused fringe projection is to suppress high-frequency quantization errors through the projector's optical low-pass filtering. Therefore, sufficient defocusing is a necessary condition for binary defocused fringe projection. The multi-period and multi-defocus simulations in Fig. 7 further show that shorter fringe periods require less defocusing, while GBO maintains good performance even at small fringe periods.

As shown by the Fourier decomposition in Fig. 4(b), the main phase error of the FS algorithm originates from the fundamental sinusoidal error component, whose amplitude is approximately 26.7 times that of the largest higher-order harmonic shown. Furthermore, we further analyzed the influence of the residual two-dimensional error after filtering on phase recovery. Specifically, we removed the fundamental Fourier component obtained from the decomposition in Fig. 4(b) from the defocused fringes in Fig. 4(a), retained the remaining two-dimensional residual, and calculated the resulting phase error using the following equation: Δϕ(x,y)=2NB∑n=1Nsin⁡(ϕ+2πnN)ΔIn(x,y).\Delta\phi(x,y)=\frac{2}{NB}\sum_{n=1}^{N}\sin\left(\phi+2\pi\frac{n}{N}\right)\Delta I_n(x,y). The phase error induced by the remaining two-dimensional residual is approximately 0.00673 rad, which is only 11.67% of the phase error induced by the fundamental component. This ratio further decreases as the defocus level increases. This indicates that the fundamental error component along the xx-direction is the dominant factor.

These error characteristics are also consistent with the relevant analyses in Chen et al. (2026) and You et al. (2022b). Therefore, our theoretical analysis focuses on the low-frequency error components that remain after optical filtering and have a major influence on the final phase recovery. We also acknowledge the need for further investigation of higher-order phase errors and will address them in future work.

(2) Gaussian and Fringe Orientation

Gaussian Defocus

This work uses a Gaussian PSF to approximate the optical defocus effect of the projector. This modeling approach has been widely adopted in related studies. For example, You et al. (2022b), Shan et al. (2023), and Chen et al. (2026) all used a Gaussian PSF to describe the projector's optical defocusing process. The effectiveness of existing research results ensures the reliability of this assumption.

Furthermore, simulation experiments are also widely used in existing studies to analyze and optimize binary projection methods. This is because the error of this method originates from the residual intensity error after binary fringes pass through the projector's optical filtering; it is an inherent error on the projection side, rather than an error introduced by a specific object or acquisition scene. Therefore, simulation analysis targeting this error mechanism can provide a direct theoretical reference for practical measurement. On this basis, Fig. 7 further presents systematic simulation results under different fringe periods and defocus levels, verifying GBO's error suppression capability and applicability over a wide parameter range.

Fringe Orientation

In real measurement, horizontal and vertical fringes are the two simplest and most common configurations in FPP. Their phase variation directions align with the image-coordinate axes used in camera–projector calibration, facilitating direct mapping between phase and projector pixels. Other orientations generally require additional coordinate transformations or phase mapping. Therefore, horizontal and vertical fringes better meet the needs of high-speed measurement for algorithmic simplicity and computational efficiency. Moreover, most of the comparison methods in the experiments in this paper are also designed for horizontal fringes. Therefore, this paper mainly validates the directional balance criterion for these two standard orientations. This criterion is constructed based on the relationship between the first moment of the diffusion kernel and the fringe phase gradient direction. Its simple constraint provides a theoretical basis for extension to other orientations. We will analyze other orientations in future work.

AuthorsOpenReview ↗

Experimental Generalization

(3) Generalization

Physical Experiments

We tested GBO on another FPP system under different hardware and complex scenes. The new system uses an acA1600-20um camera with a resolution of 1626×12361626\times1236 pixels. We also redesigned the projector–camera geometry for the new system. With the new system, we measured two objects with complex geometric features: the curved dinosaur model has obvious angular undulations, multiple sharp edges, and severe self-occlusion; the building-block dog model has multiple building-block connection surfaces, protruding structures, and severe occlusion.

Moreover, we captured a multi-jointed puppet twice. The puppet has multiple human-like joints and self-occlusion near the joints. In the first capture, we covered the background with a black cloth; in the second, we left the background unisolated to evaluate the effect of background interference. The point clouds reconstructed from the corresponding 8-bit fringe patterns were used as references. The above experimental results are as follows:

Table 1: MAE/RMSE of reconstructed point clouds (mm)

ObjectFSSOKMKHTFFPCGBO
Dinosaur0.315/0.3220.121/0.1310.192/0.2170.071/0.6130.059/0.0760.049/0.064
Dog0.188/0.1920.065/0.0750.109/0.1280.034/0.0460.049/0.0600.031/0.042
Puppet (with background isolation)0.622/0.6350.216/0.2390.341/0.3890.264/3.1800.141/0.1730.095/0.123
Puppet (without background isolation)0.585/0.5970.200/0.2290.335/0.3880.304/3.5470.169/0.1990.093/0.125

In the first two experiments, GBO achieved the lowest MAE and RMSE. Compared with FPC, GBO reduced the average MAE and RMSE in the two experiments by 26.78% and 22.05%, respectively. GBO therefore maintained strong error suppression despite changes in camera resolution and projector–camera geometry, even for objects with complex edges and occlusion.

In the repeated-acquisition experiments, after the background changed from isolated to non-isolated, the MAE and RMSE of the second-best FPC increased by 19.75% and 15.07%, respectively; in contrast, the MAE of GBO changed by -1.37%, and its RMSE increased by only 1.26%. GBO achieved the lowest MAE and RMSE in both captures. Compared with FPC, the average MAE and RMSE in the two captures were reduced by 38.60% and 33.10%, respectively. This shows that GBO still has good measurement stability when background interference changes.

Dynamic Experiments

In real dynamic measurement, the object's motion trajectory, initial position, and final state are difficult to repeat completely across different experiments; therefore, the measurement results of different methods are easily affected by differences in motion states. Moreover, under the current high-speed measurement conditions, 8-bit sinusoidal fringes cannot provide a synchronized reference at the same speed as binary methods owing to projection-speed limits.

To achieve fairer and repeatable comparisons among different binary encoding methods, we further built a Blender-based dynamic simulation system. This system can precisely control the object's motion trajectory, motion speed, and measurement timing and can obtain the corresponding true geometric information. The projector resolution was set to 1000×10001000\times1000 pixels, the camera resolution was set to 1250×10801250\times1080 pixels, and the projector-camera geometry configuration was also adjusted. The experiment used the actual projector's maximum projection frame rate of 4225 frames/s, and the average object motion speed was approximately 73.25 mm/s. All methods were compared under identical motion trajectories, speeds, projection frame rates, defocus configurations, and system parameters.

The three dynamic scenes are as follows: a watch with a transparent glass surface and localized reflection interference; a lion-head model with obvious hair texture, rich details, and self-occlusion; and a cup with severe specular reflection and obvious interreflection.

The experimental results are as follows:

Table 2: MAE/RMSE of dynamic experiments (mm)

ObjectFSSOKMKHTFFPCGBO
Watch0.112/0.1170.042/0.0500.082/0.1050.047/0.4580.033/0.0420.026/0.034
Lion0.099/0.1060.037/0.0480.065/0.0790.046/0.6350.036/0.0460.028/0.038
Cup0.105/0.1690.038/0.0660.075/0.1360.309/2.2590.030/0.1360.023/0.056

GBO achieved the lowest MAE and RMSE in all three dynamic experiments. Compared with FPC, the average MAE and RMSE in the three dynamic experiments were reduced by 22.35% and 31.86%, respectively. These results demonstrate that GBO provides stable error suppression and maintains good performance in scenes involving transparent media, complex textures, self-occlusion, and specular reflection.

We again thank you for the constructive comments. We hope that the above response can fully address your concerns and help improve your assessment of our work.

Reviewer_nVutOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how the error-diffusion kernel used to binarize sinusoidal fringes causes direction-dependent phase errors in binary-defocusing fringe projection profilometry. The authors find that forward and backward diffusion bias the recovered phase in opposite directions. They propose a balance coefficient to measure this asymmetry and argue that the main bias disappears when the two are balanced.

Based on this rule, they design kernels for horizontal and vertical fringes (GBO). On a real projector–camera system, GBO gives lower phase and point-cloud errors than five existing methods under both 3-step and 4-step phase shifting, with no extra run-time cost.

Critical strengths
  1. Simple, interpretable design rule. The balance condition has a clear physical meaning (the kernel's centroid along the phase gradient) and is easy to apply.
  2. No run-time cost. Only the offline kernel changes. The phase-shifting sequence and reconstruction are untouched, and the method works for both 3-step and 4-step phase shifting.
  3. Promising results. GBO achieves the lowest errors on almost every object across all experiments and avoids the large edge errors seen with HTF.
Critical weaknesses
  1. Limited and self-collected evaluation is my main concern. All real-world comparisons rely on only 9 static objects captured by the authors: 4 in Tables 2–3, 3 in Table 4 and 2 in Table 5. Each object appears to be captured once, on a single setup with one fringe period and one defocus level. No public data are used, the data are not released, and no error bars are given, although the margins are small (e.g. 0.030 vs. 0.037 rad on Block). The headline improvements come from the horizontal-fringe experiment alone, averaged over four objects against different second-best methods. Moreover, only the 4-tap Floyd–Steinberg layout is tested, so the claim of "a general design principle" is not supported. A larger and more diverse set covering different materials, geometries, fringe periods and defocus levels is needed, along with repeated captures and a data release.
  2. Limited comparisons. Cancelling the directional bias of error diffusion is well studied in the halftoning literature, for example through frequency analysis of the kernel and serpentine scanning, but neither is discussed. Serpentine-scan Floyd–Steinberg and other kernel-optimization methods would be natural baselines. Chen et al. (2026), which combines a Floyd–Steinberg kernel with its mirror image, addresses the same forward–backward asymmetry. It is cited in the related work but not discussed in relation to the balance rule. Since it is very recent, a full comparison may not be expected, but the paper should explain how the two approaches differ. In addition, the vertical-fringe comparison is unfair: the baselines use kernels designed for horizontal fringes, while GBO uses a rotated kernel.
  3. Weak theory and unclear kernel design. The derivation treats the quantization residual as slowly varying, although it is high-frequency by construction. The recommended weight for the downward direction (about 0.23) comes from where a fitted polynomial reaches zero. The final kernel then uses a different value (about 0.43) without explaining how it was chosen. The paper also says this weight is found by iterative optimization, while the conclusion states that no iterative optimization is needed.

Minor

  1. Please remove strong expressions that the evidence does not support, such as "a dominant active-vision approach" (line 37), "critical limitation", "must satisfy" and "highly suitable".
  2. Please define the fringe orientation. Describing "horizontal fringes" as varying along x is opposite to the usual convention.
  3. The symbol η is used both for the balance coefficient and for the normalized kernel weights.
  4. Phase errors are given in pixels in Table 1 but in radians in Tables 2–3.
Questions for authors
  1. How exactly was the GBO kernel chosen?
  2. How does GBO compare with serpentine-scan Floyd–Steinberg, and how does it differ from Chen et al. (2026)?
  3. Does the advantage hold over repeated captures and on a larger, more varied set of objects?
3 comments
AuthorsOpenReview ↗

Kernel Design and Comparisons

We sincerely thank you for your careful evaluation and constructive comments. We would like to clarify and address the concerns as follows.

(1) GBO Kernel Selection and Generality

The core contribution of GBO is the directional balance criterion (η=0\eta=0), rather than identifying a particular set of kernel parameters. All diffusion kernels satisfying this criterion fall within the GBO design framework. As shown in Fig. 4(c), satisfying the balance condition can significantly reduce phase errors.

We would like to clarify that we did not obtain the GBO kernel through our own parameter tuning. As described in Section 4, we directly adopted the diffusion kernel proposed by Zhou et al. (2020), which happens to satisfy the proposed directional balance criterion, without re-optimizing its parameters. Therefore, this kernel was not specifically optimized for either the simulations or the real measurement data in this paper. Its favorable performance across different objects further supports the effectiveness of the proposed criterion.

Its downward diffusion weight is η3≈0.43\eta_3\approx0.43, which differs from the reference value of 0.23 obtained through fitting. 0.23 is merely a reference value fitted from a limited set of simulation data, rather than a strictly optimal solution. The only essential constraint of GBO is η=0\eta=0, while the remaining parameters can be flexibly selected or further optimized according to practical requirements. We will clarify this distinction in the revised manuscript.

Furthermore, we directly modified the classical FS kernel [7, 3, 5, 1] into a balanced GBOFS kernel [7, 8, 5, 1] (η3≈0.238\eta_3\approx0.238), which also achieves favorable performance (Table 1). This further demonstrates that the effectiveness of the directional balance criterion does not depend on a particular set of kernel parameters.

(2) Comparison with Serpentine FS and Chen et al. (2026)

We thank you for pointing out the omission of the serpentine scanning method. In response, we conducted additional simulations comparing serpentine FS with two GBO kernels across different fringe periods. All methods use the same Gaussian defocus parameter (σ=2\sigma=2) to ensure sufficient defocusing of binary fringes at larger periods:

Table 1: RMSE comparison under different fringe periods (pixel)

Period (px)FSSerpentine FSGBOFSGBO
120.16210.03240.02860.0289
150.17180.03210.02600.0276
160.16360.03340.02580.0282
200.18240.03650.02760.0312
240.19310.04460.03850.0369
320.19780.05440.04570.0442
480.20850.07730.06430.0658
640.21870.10210.08760.0887

Both balanced kernels consistently outperform serpentine FS across all eight fringe periods, achieving average RMSE reductions of 17.32% and 14.87%, respectively. These results further validate the effectiveness of the balance criterion and demonstrate that the performance improvement does not depend on a particular combination of kernel parameters.

Regarding Chen et al. (2026), their method achieves error compensation by combining forward- and reverse-direction binarized patterns, and its specific implementation requires four-step phase shifting, making it inapplicable to three-step phase shifting. In contrast, GBO directly constrains the error diffusion kernel to suppress directional bias during the encoding of each fringe pattern, without relying on a specific number of phase-shifting steps. The three-step phase-shifting experiments in the supplementary material further validate this advantage.

(3) Pre-compensation Property of GBO

This work addresses the residual intensity error that remains after binary fringes are projected and optically defocused by the projector. This error originates from the projection process rather than being introduced by a particular object or acquisition scene. Therefore, simulations targeting this error mechanism provide direct theoretical guidance for practical measurements, as is also the basis of many existing studies (e.g., You et al., 2022b; Chen et al., 2026).

By designing the diffusion kernel, GBO suppresses this inherent error before fringe acquisition. Its compensation mechanism does not depend on measurement data from specific objects or acquisition scenes, nor does it require kernel re-optimization for different scenarios. This is an important advantage of the proposed pre-compensation approach.

We acknowledge that challenging conditions introduce additional measurement errors. However, these errors have different physical origins from the binary encoding errors. Extensive research has addressed these additional challenges, and GBO, as a pre-compensation method, can be integrated with existing techniques to further improve 3D measurement performance.

AuthorsOpenReview ↗

Experimental Generalization

(4) Experimental Results and Generalization

The GBO kernel used in this work was not tuned for the test objects, yet it consistently achieves improvements across objects with different geometric characteristics and in both three-step and four-step phase-shifting experiments. This indicates that its effectiveness does not depend on specific measurement objects or phase-shifting configurations. Fig. 7 further evaluates GBO under various fringe periods and defocus levels. Since the error primarily originates from the projection process, its characteristics and the effectiveness of the correction can be effectively analyzed through simulation. These results demonstrate the effectiveness of the directional balance criterion under different operating conditions.

To further address your concerns regarding experimental scale, hardware generalization, and repeated measurements, we additionally established two FPP systems and conducted multiple experiments.

First, we replaced the camera with an acA1600-20um camera with a resolution of 1626×12361626\times1236 pixels and redesigned the projector-camera geometric configuration. Two objects were measured using this system: a Dinosaur model featuring pronounced geometric variations, multiple sharp edges, and severe self-occlusion, and a block-built Dog model featuring multiple connecting surfaces, protruding structures, and severe occlusion. Moreover, we performed two measurements of the same multi-jointed Puppet. Black cloth was used to isolate the background in the first measurement, whereas no black cloth was used in the second, allowing us to investigate the effect of background interference on the measurement results.

Second, we developed a Blender-based dynamic simulation system to precisely control object trajectories, velocities, and measurement timing, while providing the corresponding ground-truth geometry. The system uses a projector resolution of 1000×10001000\times1000 pixels and a camera resolution of 1250×10801250\times1080 pixels. The acquisition frame rate matches the maximum frame rate of the actual projector, reaching 4225 frames/s, with an average object velocity of approximately 73.25 mm/s. All methods are compared under identical motion trajectories, velocities, projection frame rates, defocus configurations, and system parameters. The three dynamic scenes include a watch with a transparent glass surface and local reflection interference, a lion head model with rich fur details and self-occlusion, and a cup with severe specular reflection and pronounced interreflection.

For real-world measurements, the corresponding point clouds reconstructed from 8-bit fringe patterns are used as references, while the ground-truth geometry of the Blender scenes is used as the reference for dynamic simulations. The experimental results are presented below.

Table 2. MAE/RMSE of the generalization experiments (mm)

ObjectFSSOKMKHTFFPCGBO
Dinosaur0.315/0.3220.121/0.1310.192/0.2170.071/0.6130.059/0.0760.049/0.064
Dog0.188/0.1920.065/0.0750.109/0.1280.034/0.0460.049/0.0600.031/0.042
Puppet (isolated background)0.622/0.6350.216/0.2390.341/0.3890.264/3.1800.141/0.1730.095/0.123
Puppet (non-isolated background)0.585/0.5970.200/0.2290.335/0.3880.304/3.5470.169/0.1990.093/0.125
Watch0.112/0.1170.042/0.0500.082/0.1050.047/0.4580.033/0.0420.026/0.034
Lion0.099/0.1060.037/0.0480.065/0.0790.046/0.6350.036/0.0460.028/0.038
Cup0.105/0.1690.038/0.0660.075/0.1360.309/2.2590.030/0.1360.023/0.056

As shown in the table, GBO achieves the lowest MAE and RMSE in all seven experiments. Compared with FPC, GBO reduces the average MAE and RMSE across the seven experiments by 28.29% and 29.59%, respectively.

Notably, in the repeated acquisition experiments, when the background changes from isolated to non-isolated, the MAE and RMSE of FPC increase by 19.75% and 15.07%, respectively, whereas the MAE of GBO decreases by only 1.37% and its RMSE increases by only 1.26%, demonstrating good measurement stability. The dynamic simulation experiments further demonstrate that GBO maintains consistent error suppression under different conditions, including transparent media, complex textures, self-occlusion, and specular reflection.

These additional experiments further validate the effectiveness and generalization capability of GBO across different hardware configurations, complex objects, background conditions, and dynamic scenes. We will incorporate the corresponding experimental details into the revised manuscript and make the relevant data publicly available upon acceptance to improve reproducibility.

AuthorsOpenReview ↗

Theoretical and Clarifications

(5) Fairness of the Vertical Fringe Experiments

We would like to clarify that the purpose of the vertical fringe experiments in Section 4.2 is not to repeat the performance comparison among different methods, but to verify that the directional balance condition must be aligned with the phase variation direction of the fringes.

If the fringe pattern and diffusion kernel are rotated simultaneously by 90°, the resulting binarization process is essentially a rotated reproduction of the horizontal fringe experiment in Section 4.1 under the corresponding coordinate transformation. Such a comparison cannot reveal the relationship between the kernel orientation and the fringe direction.

Therefore, we deliberately included GBO' with the unrotated kernel and GBO with the rotated kernel for comparison. The clear difference in their phase errors demonstrates the importance of aligning the diffusion kernel orientation with the fringe phase variation direction. Moreover, the other kernel methods exhibit similar phase-error behavior to GBO' under the same orientation mismatch, further confirming the necessity of this alignment. We will further clarify the purpose of this experiment in the revised manuscript.

(6) Theoretical Approximation

We acknowledge that the original binarization residual contains high-frequency components. However, the fundamental mechanism of binary defocusing projection is to suppress high-frequency quantization errors through the optical low-pass filtering of the projector. The Fourier decomposition in Fig. 4(b) shows that the directional phase bias of FS is primarily attributed to the fundamental sinusoidal error component, whose amplitude is more than 26 times that of the largest higher-order harmonic shown. Based on this physical mechanism, we adopt a low-frequency approximation in the theoretical derivation (Supplementary Material). We focus on the dominant error components remaining after optical filtering and analyze their relationship with the diffusion kernel parameters. We will further elaborate on this point in the revised manuscript.

(7) Other Issues

We thank you for pointing out the presentation and formatting issues. We will revise overly strong statements that are not sufficiently supported by evidence, clarify the definition of fringe orientation, distinguish the notations for the directional balance coefficient and normalized diffusion weights, and further specify the definitions and units of phase errors in different tables.

Once again, we sincerely thank you for your constructive comments. We hope that our clarifications and additional experiments adequately address your concerns and merit an improved rating.

Reviewer_4o4wOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper proposes a method for structured-light 3D reconstruction. Binary-defocusing fringe projection methods use binary patterns that are optically low-pass filtered into approximately sinusoidal fringes. To improve the approximation at moderate defocus, dithering algorithms shape the binary pattern before projection. Among them, error diffusion is commonly used, and the authors introduce a normalized balance coefficient that measures the asymmetry between forward and backward error propagation for the diffusion kernel. On experimental results, they achieve a marginal improvement over existing methods.

Critical strengths

The two contributions the authors claim for their work are adequate, that is:

  1. They qualitatively analyze the four error-diffusion coefficients and show that they contribute differently to the recovered phase error. This per-parameter analysis reveals that the forward and backward components dominate the directional bias.

This is confirmed analytically and with empirical results.

  1. They introduce a normalized directional balance coefficient for error-diffusion kernels and show that the first-order phase bias vanishes when the forward and backward propagation strengths are balanced.

This is a sound modification of the error-diffusion kernel that is validated in simulation and experimentally.

Critical weaknesses

The paper lacks technical scope and experimental margins:

  1. Scope: While the contribution to error-diffusion is a sound one, it is a marginal tweak of a niche method. Patterned illumination 3D scanning itself is niche for ICLR (as opposed to specialized venues such as ICCP) and the proposed method is hardly a learning method (although backpropagation is used). The proposed method introduces marginally improved forward model, which is calibrated using gradient descent. As such, it fall under calibration approaches in my view.

  2. Marginal Results: The experimental findings show very marginal improvement compared to existing approaches. The atuhors report essentially on-par performance with FPC on nominal scenes while not reporting complex scenes (self-occlusions, refractive objects) at all.

Questions for authors

How does your method perform on complex scenes including heavy self-occlusions, sun-light and refractive object with almost no lambertian term.

3 comments
AuthorsOpenReview ↗

Performance and Experiments

We thank you for evaluating our work. To address your concerns regarding improvement magnitude, complex scenes, methodological nature, and scope, we have conducted several additional experiments and further clarified the design principle and positioning of GBO.

(1) Performance Improvement

Among the nine main experiments reported in the paper, including four horizontal-fringe phase-measurement experiments, three point-cloud reconstruction experiments, and two three-step phase-shifting experiments, GBO reduces MAE/RMSE by an average of 79.50%/68.65% relative to FS and 24.56%/20.41% relative to FPC. These results show that the performance improvements achieved by GBO are not merely marginal numerical gains. To further validate the generalization capability of the method, we added seven experiments covering different hardware configurations, objects with complex geometries, background interference, and dynamic measurement scenarios.

For physical measurements, we added an FPP system using an acA1600-20um camera with a resolution of 1626×12361626\times1236 pixels, and redesigned the projector-camera geometric configuration. Using this system, we measured two objects: a Dinosaur model with pronounced geometric variations, multiple sharp edges, and severe self-occlusion, and a block-built Dog model with multiple connecting surfaces, protruding structures, and severe occlusion. In addition, we captured the same multi-jointed Puppet twice, using isolated and non-isolated background configurations, respectively, to examine the effect of background changes on the measurement results.

For dynamic measurements, we built a Blender-based simulation system that allows precise control of the object trajectory, speed, and measurement timing, while providing the corresponding ground-truth geometry. The system uses a projector resolution of 1000×10001000\times1000 pixels and a camera resolution of 1250×10801250\times1080 pixels. The projection frame rate matches the maximum frame rate of the actual projector, at 4225 frames/s, and the average object speed is approximately 73.25 mm/s. All binary encoding methods were compared under the same object trajectory, speed, projection frame rate, defocus configuration, and system parameters. The three dynamic scenes are a Watch with a transparent glass surface and localized reflection interference, a Lion with rich fur details and self-occlusion, and a Cup with severe specular reflection and pronounced interreflection.

For the physical measurements, the point clouds reconstructed from the corresponding 8-bit fringe patterns were used as references. For the dynamic simulations, the ground-truth geometry of the Blender scenes was used as the reference. The additional experimental results are presented below:

Table 1. MAE/RMSE in the additional experiments (mm)

Test objectFSSOKMKHTFFPCGBO
Dinosaur0.315/0.3220.121/0.1310.192/0.2170.071/0.6130.059/0.0760.049/0.064
Dog0.188/0.1920.065/0.0750.109/0.1280.034/0.0460.049/0.0600.031/0.042
Puppet (isolated background)0.622/0.6350.216/0.2390.341/0.3890.264/3.1800.141/0.1730.095/0.123
Puppet (non-isolated background)0.585/0.5970.200/0.2290.335/0.3880.304/3.5470.169/0.1990.093/0.125
Watch0.112/0.1170.042/0.0500.082/0.1050.047/0.4580.033/0.0420.026/0.034
Lion0.099/0.1060.037/0.0480.065/0.0790.046/0.6350.036/0.0460.028/0.038
Cup0.105/0.1690.038/0.0660.075/0.1360.309/2.2590.030/0.1360.023/0.056

GBO achieves the lowest MAE and RMSE in all seven experiments. Averaging the improvement ratios over FPC across the seven experiments, the MAE and RMSE are reduced by 28.29% and 29.59%, respectively. Together with the results reported in the paper, these experiments show consistent improvements across objects, hardware configurations, and dynamic conditions.

AuthorsOpenReview ↗

Complex Scenes and Method Nature

(2) Complex scenes

For scenes with severe self-occlusion, the additional physical measurements of the Dinosaur, Dog, and Puppet, as well as the dynamic simulation of the Lion, all contain evident self-occlusion or regions with limited visibility. GBO achieves the lowest MAE and RMSE in all five corresponding experiments. Compared with FPC, the average MAE and RMSE are reduced by 30.70% and 25.85%, respectively. This indicates that the directional balance criterion can still effectively suppress binary encoding errors.

For sunlight illumination, ambient light mainly introduces an additional background intensity component into the phase-shifted images. This DC component can be removed through phase-shifting demodulation. In addition, appropriately increasing the projector intensity can enhance the fringe modulation relative to the ambient-light background; therefore, its impact is limited.

For non-Lambertian conditions, the additional Watch dynamic simulation includes a transparent glass surface and localized reflection interference, while the Cup experiment involves severe specular reflection and pronounced interreflection. Compared with FPC, GBO reduces the MAE and RMSE by 21.21% and 19.05%, respectively, in the Watch experiment, and by 23.33% and 58.82%, respectively, in the Cup experiment. These results show that GBO can still effectively suppress phase errors introduced by binary encoding under complex imaging conditions involving non-Lambertian interference.

It should be emphasized that the core objective of this work is to address directional phase errors caused by binary fringe encoding in high-speed FPP, rather than to solve measurement failures under all extreme imaging conditions. These experiments are intended to verify that the projection-side errors targeted by GBO can still be effectively suppressed when other complex imaging factors are present.

(3) The Nature of GBO

GBO is not an improvement to the forward model calibrated using gradient descent.

First, GBO does not improve the forward model or optimize the diffusion-kernel parameters through gradient descent. Instead, it analyzes the directional propagation characteristics of binarization errors and proposes a directional balance design criterion with a clear physical meaning. The parameter fitting in Section 3.3 is used to analyze the relationship between diffusion-kernel parameters and phase errors, rather than to calibrate the GBO kernel actually used through gradient descent. More importantly, in the Supplementary Material, we independently derive the directional balance condition Mx=0M_x=0 from the first moment of the diffusion kernel, namely, λ1+λ4=λ2\lambda_1+\lambda_4=\lambda_2. This analytical derivation further demonstrates that the directional balance criterion is not obtained through empirical parameter tuning.

Second, the GBO kernel actually used does not rely on parameter optimization. We directly use the existing diffusion kernel proposed by Zhou et al. (2020), which already satisfies the directional balance condition and has not been re-optimized for the simulation data or measurement objects in this paper. In addition, starting from the classical FS kernel [7, 3, 5, 1], we directly construct another balanced kernel, GBOFS [7, 8, 5, 1]. Its second diffusion weight is directly determined by the directional balance condition: λ2=λ1+λ4=7+1=8.\lambda_2=\lambda_1+\lambda_4=7+1=8. This process does not involve gradient descent, parameter search, or calibration based on measurement data. To further verify that the design criterion does not depend on a particular kernel-parameter configuration, we conducted additional simulations of GBO and GBOFS over eight fringe periods. The results are presented in Table 2.

Table 2. Phase RMSE of the balanced kernels (pixel)

Kernel12 px15 px16 px20 px24 px32 px48 px64 px
FS0.16210.17180.16360.18240.19310.19780.20850.2187
GBOFS0.02860.02600.02580.02760.03850.04570.06430.0876
GBO0.02890.02760.02820.03120.03690.04420.06580.0887

Both balanced kernels achieve relatively low phase errors across all eight fringe periods. GBOFS performs better at six periods, while GBO performs better at the other two. The two kernels have different weight configurations, but both satisfy the same directional balance condition and achieve good performance. This further demonstrates that the criterion does not depend on a particular set of parameters.

Therefore, the core contribution of this work is not to calibrate or fine-tune an existing forward model through gradient descent, but to reveal the relationship between diffusion-kernel directionality and phase error, and to propose an analytical constraint that can directly guide the design of different diffusion kernels. GBO is a binary fringe encoding method based on a theoretical criterion, rather than a data-driven forward-model calibration method.

(4) Relevance to the ICLR Research Scope

This work investigates binary fringe encoding errors in high-speed 3D measurement and improves reconstruction accuracy, representing a typical vision application. At submission, we also selected Applications in vision, audio, speech, language, and other modalities as the Primary Area.

We believe that this is consistent with the statement regarding the main research areas in the ICLR Call for Papers. The ICLR official Subject Areas explicitly state:

Subject Areas

We consider a broad range of subject areas including feature learning, metric learning, compositional modeling, structured prediction, reinforcement learning, uncertainty quantification and issues regarding large-scale learning and non-convex optimization, as well as applications in vision, audio, speech, language, music, robotics, games, healthcare, biology, sustainability, economics, ethical considerations in ML, and others.

This indicates that ICLR explicitly welcomes application research in vision and does not restrict its scope to learning methods. Fringe Projection Profilometry (FPP) is a classical active 3D vision technique widely used in high-precision 3D sensing and measurement. The present work falls within the vision-application research areas explicitly listed by ICLR.

Once again, we thank you for the evaluation. We hope that the additional quantitative experiments, validation in complex scenes, and further clarification of the method's design principle and research positioning adequately address your concerns and encourage you to reconsider your rating of the paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.